text
stringlengths 2.85k
2.55M
| label
class label 11
classes |
|---|---|
Representations of language in a model of visually grounded speech signal
Grzegorz Chrupała
Lieke Gelderloos
Tilburg University
Tilburg University
[email protected] [email protected]
arXiv:1702.01991v3 [cs.CL] 30 Jun 2017
Abstract
We present a visually grounded model of
speech perception which projects spoken
utterances and images to a joint semantic space. We use a multi-layer recurrent
highway network to model the temporal
nature of spoken speech, and show that it
learns to extract both form and meaningbased linguistic knowledge from the input
signal. We carry out an in-depth analysis of the representations used by different
components of the trained model and show
that encoding of semantic aspects tends
to become richer as we go up the hierarchy of layers, whereas encoding of formrelated aspects of the language input tends
to initially increase and then plateau or decrease.
1
Introduction
Speech recognition is one of the success stories
of language technology. It works remarkably well
in a range of practical settings. However, this
success relies on the use of very heavy supervision where the machine is fed thousands of hours
of painstakingly transcribed audio speech signal.
Humans are able to learn to recognize and understand speech from notably weaker and noisier supervision: they manage to learn to extract structure and meaning from speech by simply being exposed to utterances situated and grounded in their
daily sensory experience. Modeling and emulating this remarkable skill has been the goal of numerous studies; however in the overwhelming majority of cases researchers used severely simplified
settings where either the language input or the extralinguistic sensory input, or both, are small scale
and symbolically represented. Section 2 provides
a brief overview of this research.
Afra Alishahi
Tilburg University
[email protected]
More recently several lines of work have moved
towards more realistic inputs while modeling or
emulating language acquisition in a grounded setting. Gelderloos and Chrupała (2016) use the
image captioning dataset MS COCO (Lin et al.,
2014) to mimic the setting of grounded language
learning: the sensory input consists of images of
natural scenes, while the language input are phonetically transcribed descriptions of these scenes.
The use of such moderately large and low-level
data allows the authors to train a multi-layer recurrent neural network model, and to explore the
nature and localization of the emerging hierarchy
of linguistic representations learned in the process.
Furthermore, in a series of recent studies Harwath
and Glass (2015); Harwath et al. (2016); Harwath
and Glass (2017) use image captioning datasets
to model learning to understand spoken language
from visual context with convolutional neural network models. Finally, there is a small but growing body of work dedicated to elucidating the nature of representations learned by neural networks
from language data (see Section 2.2 for a brief
overview). In the current work we build on these
three strands of research and contribute the following advances:
• We use a multi-layer gated recurrent neural
network to properly model the temporal nature of speech signal and substantially improve performance compared to the convolutional architecture from Harwath and Glass
(2015);
• We carry out an in-depth analysis of the representations used by different components of
the trained model and correlate them to representations learned by a text-based model and
to human patterns of judgment on linguistic
stimuli. This analysis is especially novel for
a model with speech signal as input.
The general pattern of findings in our analysis is
as follows: The model learns to extract from the
acoustic input both form-related and semanticsrelated information, and encodes it in the activations of the hidden layers. Encoding of semantic
aspects tends to become richer as we go up the hierarchy of layers. Meanwhile, encoding of formrelated aspects of the language input, such as utterance length or the presence of specific words,
tends to initially increase and then decay.
We release the code for our models
and analyses as open source, available at
https://github.com/gchrupala/visually-groundedspeech. We also release a dataset of synthetically
spoken image captions based on MS COCO, available at https://doi.org/10.5281/zenodo.400926.
2
Related work
Children learn to recognize and assign meaning
to words from continuous perceptual data in extremely noisy context. While there have been
many computational studies of human word meaning acquisition, they typically make strong simplifying assumptions about the nature of the input. Often language input is given in the form
of word symbols, and the context consists of a
set of symbols representing possible referents (e.g.
Siskind, 1996; Frank et al., 2007; Fazly et al.,
2010). In contrast, several studies presented models that learn from sensory rather than symbolic input, which is rich with regards to the signal itself,
but very limited in scale and variation (e.g. Roy
and Pentland, 2002; Yu and Ballard, 2004; Lazaridou et al., 2016).
2.1 Multimodal language acquisition
Chrupała et al. (2015) introduce a model that
learns to predict the visual context from image
captions. The model is trained on image-caption
pairs from MSCOCO (Lin et al., 2014), capturing both rich visual input as well as larger scale
input, but the language input still consists of word
symbols. Gelderloos and Chrupała (2016) propose
a similar architecture that instead takes phonemelevel transcriptions as language input, thereby incorporating the word segmentation problem into
the learning task. In this work, we introduce an architecture that learns from continuous speech and
images directly.
This work is related to research on visual
grounding of language. The field is large and
growing, with most work dedicated to the ground-
ing of written text, particularly in image captioning tasks (see Bernardi et al. (2016) for an
overview). However, learning to ground language
to visual information is also interesting from an
automatic speech recognition point of view. Potentially, ASR systems could be trained from naturally co-occurring visual context information,
without the need for extensive manual annotation – a particularly promising prospect for speech
recognition in low-resource languages. There
have been several attempts along these lines. Synnaeve et al. (2014) present a method of learning
to recognize spoken words in isolation from cooccurrence with image fragments. Harwath and
Glass (2015) present a model that learns to map
pre-segmented spoken words in sequence to aspects of the visual context, while in Harwath and
Glass (2017) the model also learns to recognize
words in the unsegmented signal.
Most closely related to our work is that of Harwath et al. (2016), as it presents an architecture
that learns to project images and unsegmented
spoken captions to the same embedding space.
The sentence representation is obtained by feeding the spectrogram to a convolutional network.
The architecture is trained on crowd-sourced spoken captions for images from the Places dataset
(Zhou et al., 2014), and evaluated on image search
and caption retrieval. Unfortunately this dataset is
not currently available and we were thus unable to
directly compare the performance of our model to
Harwath et al. (2016). We do compare to Harwath
and Glass (2015) which was tested on a public
dataset. We make different architectural choices,
as our models are based on recurrent highway networks (Zilly et al., 2016). As in human cognition,
speech is processed incrementally. This also allows our architecture to integrate information sequentially from speech of arbitrary duration.
2.2
Analysis of neural representations
While analysis of neural methods in NLP is often limited to evaluation of the performance on
the training task, recently methods have been introduced to peek inside the black box and explore
what it is that enables the model to perform the
task. One approach is to look at the contribution
of specific parts of the input, or specific units in the
model, to final representations or decisions. Kádár
et al. (2016) propose omission scores, a method to
estimate the contribution of input tokens to the fi-
nal representation by removing them from the input and comparing the resulting representations to
the ones generated by the original input. In a similar approach, Li et al. (2016) study the contribution of individual input tokens as well as hidden
units and word embedding dimensions by erasing
them from the representation and analyzing how
this affects the model.
Miao et al. (2016) and Tang et al. (2016) use visualization techniques for fine-grained analysis of
GRU and LSTM models for ASR. Visualization
of input and forget gate states allows Miao et al.
(2016) to make informed adaptations to gated recurrent architectures, resulting in more efficiently
trainable models. Tang et al. (2016) visualize
qualitative differences between LSTM- and GRUbased architectures, regarding the encoding of information, as well as how it is processed through
time.
We specifically study linguistic properties of the
information encoded in the trained model. Adi
et al. (2016) introduce prediction tasks to analyze information encoded in sentence embeddings
about word order, sentence length, and the presence of individual words. We use related techniques to explore encoding of aspects of form and
meaning within components of our stacked architecture.
3
Models
We use a multi-layer, gated recurrent neural network (RHN) to model the temporal nature of
speech signal. Recurrent neural networks are designed for modeling sequential data, and gated
variants (GRUs, LSTMs) are widely used with
speech and text in both cognitive modeling and
engineering contexts. RHNs are a simple generalization of GRU networks such that the transform
between time points can consist of several steps.
Our multimodal model projects spoken utterances and images to a joint semantic space. The
idea of projecting different modalities to a shared
semantic space via a pair of encoders has been
used in work on language and vision (among them
Vendrov et al. (2015)). The core idea is to encourage inputs representing the same meaning in
different modalities to end up nearby, while maintaining a distance from unrelated inputs.
The model consists of two parts: an utterance
encoder, and an image encoder. The utterance encoder starts from MFCC speech features, while
the image encoder starts from features extracted
with a VGG-16 pre-trained on ImageNet. Our loss
function attempts to make the cosine distance between encodings of matching utterances and images less than the distance between encodings of
mismatching utterance/image pairs, by a margin:
X X
u0
u,i
+
X
i0
max[0, α + d(u, i) − d(u0 , i)]
! (1)
max[0, α + d(u, i) − d(u, i0 )]
where d(u, i) is the cosine distance between the
encoded utterance u and encoded image i. Here
(u, i) is the matching utterance-image pair, u0
ranges over utterances not describing i and i0
ranges over images not described by u.
The image encoder enci is a simple linear projection, followed by normalization to unit L2
norm:
enci (i) = unit(Ai + b)
(2)
where unit(x) = (xT xx)0.5 and with (A, b) as
learned parameters. The utterance encoder encu
consists of a 1-dimensional convolutional layer of
length s, size d and stride z, whose output feeds
into a Recurrent Highway Network with k layers and L microsteps, whose output in turn goes
through an attention-like lookback operator, and
finally L2 normalization:
encu (u) = unit(Attn(RHNk,L (Convs,d,z (u))))
(3)
The main function of the convolutional layer
Convs,d,z is to subsample the input along the temporal dimension. We use a 1-dimensional convolution with full border mode padding. The attention operator simply computes a weighted sum of
the RHN activation at all timesteps:
X
Attn(x) =
α t xt
(4)
t
where the weights αt are determined by learned
parameters U and W, and passed through the
timewise softmax function:
exp(U tanh(Wxt ))
αt = P
t0 exp(U tanh(Wxt0 ))
(5)
The main component of the utterance encoder is a
recurrent network, specifically a Recurrent Highway Network (Zilly et al., 2016). The idea behind
RHN is to increase the depth of the transform between timesteps, or the recurrence depth. Otherwise they are a type of gated recurrent networks.
The transition from timestep t − 1 to t is then defined as:
(L)
(L)
rhn(xt , st−1 ) = st
(6)
(l)
where xt stands for input at time t, and st denotes the state at time t at recurrence layer l, with
L being the top layer of recurrence. Furthermore,
(l)
(l)
(l)
(l−1)
(l)
st = ht
tt + st
1 − tt
(7)
where
(l)
ht
(l)
tt
is elementwise multiplication, and
(l−1)
(8)
= tanh I[l = 1]WH xt + UHl st
(9)
= σ I[l = 1]WT xt + UTl s(l−1)
Here I is the indicator function: input is only included in the computation for the first layer of recurrence l = 1. By applying the rhn function repeatedly, an RHN layer maps a sequence of inputs
to a sequence of states:
RHN(X, s0 )
(L)
(10)
= rhn(xn , . . . , rhn(x2 , rhn(x1 , s0 )))
Two or more RHN layers can be composed into a
stack:
(L)
(L)
RHN2 (RHN1 (X, s1 0 ), s2 0 ),
(11)
(l)
where sn t stands for the state vector of layer n of
the stack, at layer l of recurrence, at time t. In our
version of the Stacked RHN architecture we use
residualized layers:
RHNres (X, s0 ) = RHN(X, s0 ) + X
(12)
This formulation tends to ease optimization in
multi-layer models (cf. He et al., 2015; Oord et al.,
2016).
In addition to the speech model described
above, we also define a comparable text model.
As it takes a sequence of words as input, we replace the convolutional layer with a word embedding lookup table. We found the text model did
not benefit from the use of the attention mechanism, and thus the sentence embedding is simply
the L2-normalized activation vector of the topmost
layer, at the last timestep.
4
Experiments
Our main goal is to analyze the emerging representations from different components of the model
and to examine the linguistic knowledge they encode. For this purpose, we employ a number of
tasks that cover the spectrum from fully formbased to fully semantic.
In Section 4.2 we assess the effectiveness of our
architecture by evaluating it on the task of ranking images given an utterance. Sections 4.3 to 4.6
present our analyses. In Sections 4.3 and 4.4 we
define auxiliary tasks to investigate to what extent
the network encodes information about the surface
form of an utterance from the speech input. In Section 4.5 and 4.6 we focus on where semantic information is encoded in the model. In the analyses,
we use the following features:
Utterance embeddings: the weighted sum of the
unit activations on the last layer, as calculated
by Equation (3).
Average unit activations: hidden layer activations averaged over time and L2-normalized
for each hidden layer.
Average input vectors: the MFCC vectors averaged over time. We use this feature to examine how much information can be extracted
from the input signal only.
4.1
Data
For the experiments reported in the remainder of
the paper we use two datasets of images with spoken captions.
4.1.1
Flickr8K
The Flickr8k Audio Caption Corpus was constructed by having crowdsource workers read
aloud the captions in the original Flickr8K corpus (Hodosh et al., 2013). For details of the
data collection procedure refer to Harwath and
Glass (2015). The datasets consist of 8,000 images, each image with five descriptions. One
thousand images are held out for validation, and
another one thousand for the final test set. We
use the splits provided by (Karpathy and Fei-Fei,
2015). The image features come from the final
fully connect layer of VGG-16 (Simonyan and
Zisserman, 2014) pre-trained on Imagenet (Russakovsky et al., 2014).
We generate the input signal as follows: we extract 12-dimensional mel-frequency cepstral coefficients (MFCC) plus log of the total energy. We
then compute and add first order and second order
differences (deltas) for a total of 37 dimensions.
We use 25 milisecond windows, sampled every 10
miliseconds.1
4.1.2
Synthetically spoken COCO
We generated synthetic speech for the captions
in the MS COCO dataset (Lin et al., 2014) via
the Google Text-to-Speech API.2 The audio and
the corresponding MFCC features are released as
Chrupała et al. (2017)3 . This TTS system we used
produces high-quality realistic-sounding speech.
It is nevertheless much simpler than real human
speech as it uses a single voice, and lacks tempo
variation or ambient noise. The data consists of
over 300,000 images, each with five spoken captions. Five thousand images each are held out for
validation and test. We use the splits and image
features provided by Vendrov et al. (2015).4 The
image features also come from the VGG-16 network, but are averages of feature vectors for ten
crops of each image. For the MS COCO captions
we extracted only plain MFCC and total energy
features, and did not add deltas in order to keep
the amount of computation manageable given the
size of the dataset.
4.2 Image retrieval
We evaluate our model on the task of ranking images given a spoken utterance, such that highly
ranked images contain scenes described by the utterance. The performance on this task on validation data is also used to choose the best variant
of the model architecture and to tune the hyperparameters. We compare the speech models to models trained on written sentences split into words.
The best settings found for the four models were
the following:
Flickr8K Text RHN 300-dimensional word embeddings, 1 hidden layer with 1024 dimensions, 1 microstep, initial learning rate 0.001.
Flick8K Speech RHN convolutional layer with
length 6, size 64, stride 2, 4 hidden layers
with 1024 dimensions, 2 microsteps, atten1
We noticed that for a number of utterances the audio signal was very long: on inspection it turned out that most of
these involved failure to switch off the microphone on the
part of the workers, and the audio contained ambient noise or
unrelated speech. We thus trucated all audio for this dataset
at 10,000 miliseconds.
2
Available at https://github.com/pndurette/gTTS.
3
Available at https://doi.org/10.5281/zenodo.400926.
4
See https://github.com/ivendrov/order-embedding.
tion MLP with 128 hidden units, initial learning rate 0.0002
COCO Text RHN 300-dimensional word embeddings, 1 hidden layer with 1024 dimensions, 1 microstep, initial learning rate 0.001
COCO Speech RHN convolutional layer with
length 6, size 64, stride 3, 5 hidden layers
with 512 dimensions, 2 microsteps, attention
MLP with 512 hidden units, initial learning
rate 0.0002
All models were optimized with Adam
(Kingma and Ba, 2014) with early stopping: we
kept the parameters for the epoch which showed
the best recall@10 on validation data.
Model
Speech RHN4,2
Spectr. CNN
Text RHN1,1
R@1
0.055
0.127
R@5
0.163
0.364
R@10
0.253
0.179
0.494
r̃
48
11
Table 1: Image retrieval performance on Flickr8K.
R@N stands for recall at N; r̃ stands for median
rank of the correct image.
Model
Speech RHN5,2
Text RHN1,1
R@1
0.111
0.169
R@5
0.310
0.421
R@10
0.444
0.565
r̃
13
8
Table 2:
Image retrieval performance on
MS COCO. R@N stands for recall at N; r̃ stands
for median rank of the correct image.
Table 1 shows the results for the human speech
from the Flickr8K dataset. The Speech RHN
model scores substantially higher than model of
Harwath and Glass (2015) on the same data. However the large gap between its perfomance and the
scores of the text model suggests that Flickr8K
is rather small for the speech task. In Table 2
we present the results on the dataset of synthetic
speech from MS COCO. Here the text model is
still better, but the gap is much smaller than for
Flickr8K. We attribute this to the much larger size
of dataset, and to the less noisy and less variable
synthetic speech.
While the MS COCO text model is overall better than the speech model, there are cases where
it outperforms the text model. We listed the top
hundred cases where the ratio of the ranks of the
correct image according to the two models was the
smallest, as well as another hundred cases where
it was the largest. Manual inspection did not turn
up any obvious patterns for the cases of text being better than speech. For the cases where speech
outperformed text, two patterns stood out: (i) sentences with spelling mistakes, (ii) unusually long
sentences. For example for the sentence a yellow
Figure 1: Images returned for utterance a yellow
and white birtd is in flight by the text (left) and
speech (right) models.
and white birtd is in flight the text model misses
the misspelled word birtd and returns an irrelevant
image, while the speech model seems robust to
some degree of variation in pronunciation and returns the target image at rank 1 (see Figure 1). In
an attempt to quantify this effect we counted the
number of unique words with training set frequencies below 5 in the top 100 utterances with lowest
and highest rank ratio: for the utterances where
text was better there were 16 such words; for utterances where speech was better there were 28,
among them misspellings such as streeet, scears
(for skiers), contryside, scull, birtd, devise.
The distribution of utterance lengths in Figure 2 confirms pattern (ii): the set of 100 sentences where speech beats text by a large margin
are longer on average and there are extremely long
outliers among them. One of them is the 36-word-
better
text
speech
●
10
20
●
●
30
Length
●
40
Figure 2: Length distribution for sentences where
one model performs much better than the other.
long utterance depicted in Figure 3, with ranks 470
and 2 for text and speech respectively. We suspect
that the speech model’s attention mechanism enables it to cherry pick key fragments of such monster utterances, while the text model lacking this
mechanism may struggle. Figure 3 shows the plot
of the attention weights for this utterance from the
speech model.
4.3
Predicting utterance length
Our first auxiliary task is to predict the length of
the utterance, using the features explained at the
beginning of Section 4. Since the length of an utterance directly corresponds to how long it takes to
articulate, we also use the number of time steps5 as
a feature and expect it to provide the upper bound
for our task, especially for synthetic speech. We
use a Ridge Regression model for predicting utterance length using each set of features. The model
is trained on 80% of the sentences in the validation
set, and tested on the remaining 20%. For all features regularization penalty α = 1.0 gave the best
results.
Figure 4 shows the results for this task on human speech from Flickr8K and synthetic speech
from COCO. With the exception of the average input vectors for Flickr8K, all features can explain
a high proportion of variance in the predicted utterance length. The pattern observed for the two
datasets is slightly different: due to the systematic
conversion of words to synthetic speech in COCO,
using the number of time steps for this dataset
yields the highest R2 . However, this feature is not
as informative for predicting the utterance length
in Flickr8K due to noise and variation in human
speech, and is in fact outperformed by some of the
features extracted from the model. Also, the input
vectors from COCO are much more informative
than Flickr8K due to larger quantity and simpler
structure of the speech signal. However, in both
datasets the best (non-ceiling) performance is obtained by using average unit activations from the
hidden layers (layer 2 for COCO, and layers 3 and
4 for Flickr8K). These features outperform utterance embeddings, which are optimized according
to the visual grounding objective of the model and
most probably learn to ignore the superficial characteristics of the utterance that do not contribute to
matching the corresponding image.
Note that the performance on COCO plateaus
after the second layer, which might suggest that
form-based knowledge is learned by lower layers.
Since Flickr8K is much smaller in size, the stabilising happens later in layer 3.
5
This is approximately
duration in milliseconds
.
10×stride
Figure 3: Attention weight distribution for a long utterance.
Figure 4: R2 values for predicting utterance length
for Flickr8K and COCO. Layers 1–5 represent
(normalized) average unit activation, whereas the
first (#0) and last point represent average input
vectors and utterance embeddings, respectively.
4.4 Predicting word presence
Results from the previous experiment suggest that
our model acquires information about higher level
building blocks (words) in the continuous speech
signal. Here we explore whether it can detect the
presence or absence of individual words in an utterance. We formulate detecting a word in an utterance as a binary classification task, for which
we use a multi-layer perceptron with a single hidden layer of size 1024, optimized by Adam. The
input to the model is a concatenation of the feature vector representing an utterance and the one
representing a target word. We again use utterance embeddings, average unit activations on each
layer, and average input vectors as features, and
represent each target word as a vector of MFCC
features extracted from the audio signal synthetically produced for that word.
For each utterance in the validation set, we randomly pick one positive and one negative target
(i.e., one word that does and one that does not appear in the utterance) that is not a stop word. To
balance the probability of a word being positive
or negative, we use each positive target as a negative target for another utterance in the validation
set. The MLP model is trained on the positive and
negative examples corresponding to 80% of the utterances in the validation set of each dataset, and
evaluated on the remaining 20%.
Figure 5 shows the mean accuracy of the MLP
on Flickr8K and COCO. All results using features
extracted from the model are above chance (0.5),
with the average unit activations of the hidden layers yielding the best results (0.65 for Flickr8K on
layer 3, and 0.79 for COCO on layer 4). These
numbers show that the speech model infers reliable information about word-level blocks from
the low-level audio features it receives as input.
The observed trend is similar to the previous task:
average unit activations on the higher-level hidden layers are more informative for this task than
the utterance embeddings, but the performance
plateaus before the topmost layer.
Figure 5: Mean accuracy values for predicting the
presence of a word in an utterance for Flickr8K
and COCO. Layers 1–5 represent the (normalized)
average unit activations, whereas the first (#0) and
last point represent average input vectors and utterance embeddings, respectively.
4.5
Sentence similarity
Next we explore to what extent the model’s representations correspond to those of humans. We
employ the Sentences Involving Compositional
Knowledge (SICK) dataset (Marelli et al., 2014).
SICK consists of image descriptions taken from
1.00
RER
0.75
0.50
0.25
Figure 6: Pearson’s r of cosine similarities of averaged input MFCCs and COCO Speech RHN hidden layer activation vectors and embeddings of
sentence pairs with relatedness scores from SICK,
cosine similarity of COCO Text RHN embeddings, and edit similarity.
Flickr8K and video captions from the SemEval
2012 STS MSRVideo Description data set (STS)
(Agirre et al., 2012). Captions were paired at random, as well as modified to obtain semantically
similar and contrasting counterparts, and the resulting pairs were rated for semantic similarity.
For all sentence pairs in SICK, we generate
synthetic spoken sentences and feed them to the
COCO Speech RHN, and calculate the cosine similarity between the averaged MFCC input vectors,
the averaged hidden layer activation vectors, and
the sentence embeddings. Z-score transformation
was applied before calculating the cosine similarities. We then correlate these cosine similarities
with
• semantic relatedness according to human ratings
• cosine similarities according to z-score transformed embeddings from COCO Text RHN
• edit similarities, a measure of how similar the sentences are in form, specifically, 1−normalized Levenshtein distance
over character sequences
Figure 6 shows a boxplot over 10,000 bootstrap
samples for all correlations. We observe that (i)
correlation with edit similarity initially increases,
then decreases; (ii) correlation with human relatedness scores and text model embeddings increases until layer 4, but decreases for hidden layer
5. The initially increasing and then decreasing correlation with edit similarity is consistent with the
findings that information about form is encoded
by lower layers. The overall growing correlation
with both human semantic similarity ratings and
0.00
0
2
4
6
layer
words
peaking/peeking
great/grate
mantle/mantel
peer/pier
tale/tail
wit/whit
weight/wait
isle/aisle
sight/site
pic/pick
sun/son
wears/wares
pause/paws
tied/tide
ware/wear
sales/sails
boarder/border
plane/plain
lapse/laps
rose/rows
stares/stairs
log(mincount)
4
5
6
seen/scene
plains/planes
see/sea
main/mane
rains/reins
tea/tee
stair/stare
waist/waste
hole/whole
suite/sweet
pairs/pears
cole/coal
sale/sail
7
Figure 7: Disambiguation performance per layer.
Points #0 and #6 (connected via dotted lines) represent the input vectors and utterance embeddings,
respectively. The black line shows the overall
mean RER.
the COCO Text RHN indicate that higher layers
learn to represent semantic knowledge. We were
somewhat surprised by the pattern for the correlation with human ratings and the Text model similarities which drops for layer 5. We suspect it may
be caused by the model at this point in the layer
hierarchy being strongly tuned to the specifics of
the COCO dataset. To test this, we checked the
correlations with COCO Text embeddings on validation sentences from the COCO dataset instead
of SICK. These increased monotonically, in support of our conjecture.
4.6
Homonym disambiguation
Next we simulate the task of distinguishing between pairs of homonyms, i.e. words with the
same acoustic form but different meaning. We
group the words in the union of the training and
validation data of the COCO dataset by their phonetic transcription. We then pick pairs of words
which have the same pronunciation but different
spelling, for example suite/sweet. We impose the
following conditions: (a) both forms appear more
than 20 times, (b) the two forms have different
meaning (i.e. they are not simply variant spellings
like theater/theatre), (c) neither form is a function word, and (d) the more frequent form constitutes less than 95% of the occurrences. This
gives us 34 word pairs. For each pair we generate a binary classification task by taking all the utterances where either form appears, using average
input vectors, utterance embeddings, and average
unit activations as features. Instances for all feature sets are normalized to unit L2 norm.
For each task and feature set we run stratified 10-fold cross validation using Logistic Regression to predict which of the two words the
utterance contains. Figure 7 shows, for each
pair, the relative error reduction of each feature
set with respect to the majority baseline. There
is substantial variation across word pairs, but
overall the task becomes easier as the features
come from higher layers in the network. Some
forms can be disambiguated with very high accuracy (e.g. sale/sail, cole/coal, pairs/pears), while
some others cannot be distinguished at all (peaking/peeking, great/grate, mantle/mantel). We examined the sentences containing the failing forms,
and found out that almost all occurrences of peaking and mantle were misspellings of peeking and
mantel, which explains the impossibility of disambiguating these cases.
5
Conclusion
We present a multi-layer recurrent highway network model of language acquisition from visually
grounded speech signal. Through detailed analysis we uncover how information in the input signal is transformed as it flows through the network:
formal aspects of language such as word identities
that not directly present in the input are discovered
and encoded low in the layer hierarchy, while semantic information is most strongly expressed in
the topmost layers.
Going forward we would like to compare the
representations learned by our model to the brain
activity of people listening to speech in order to
determine to what extent the patterns we found
correspond to localized processing in the human
cortex. This will hopefully lead to a better understanding of language learning and processing by
both artificial and neural networks.
Acknowledgements
We would like to thank David Harwath for making the Flickr8k Audio Caption Corpus publicly
available.
References
Yossi Adi, Einat Kermany, Yonatan Belinkov, Ofer
Lavi, and Yoav Goldberg. 2016. Fine-grained analysis of sentence embeddings using auxiliary prediction tasks. arXiv preprint arXiv:1608.04207 .
Eneko Agirre, Mona Diab, Daniel Cer, and Aitor
Gonzalez-Agirre. 2012. Semeval-2012 task 6: A pilot on semantic textual similarity. In Proceedings of
the First Joint Conference on Lexical and Computational Semantics. Association for Computational
Linguistics, volume 2, pages 385–393.
Raffaella Bernardi, Ruket Cakici, Desmond Elliott,
Aykut Erdem, Erkut Erdem, Nazli Ikizler-Cinbis,
Frank Keller, Adrian Muscat, and Barbara Plank.
2016. Automatic description generation from images: A survey of models, datasets, and evaluation
measures. arXiv preprint arXiv:1601.03896 .
Grzegorz Chrupała, Akos Kádár, and Afra Alishahi.
2015. Learning language through pictures. In Proceedings of the 53rd Annual Meeting of the Association for Computational Linguistics.
Grzegorz Chrupała, Lieke Gelderloos, and Afra
Alishahi. 2017.
Synthetically spoken COCO.
https://doi.org/10.5281/zenodo.400926.
Afsaneh Fazly, Afra Alishahi, and Suzanne Stevenson. 2010. A probabilistic computational model of
cross-situational word learning. Cognitive Science:
A Multidisciplinary Journal 34(6):1017–1063.
Michael C. Frank, Noah D. Goodman, and Joshua B.
Tenenbaum. 2007. A Bayesian framework for crosssituational word-learning. In Advances in Neural Information Processing Systems. volume 20.
Lieke Gelderloos and Grzegorz Chrupała. 2016. From
phonemes to images: levels of representation in
a recurrent neural model of visually-grounded language learning. In Proceedings of COLING 2016,
the 26th International Conference on Computational
Linguistics: Technical Papers.
David Harwath and James Glass. 2015. Deep multimodal semantic embeddings for speech and images.
In IEEE Automatic Speech Recognition and Understanding Workshop.
David Harwath and James R Glass. 2017. Learning word-like units from joint audio-visual analysis.
arXiv preprint arXiv:1701.07481 .
David Harwath, Antonio Torralba, and James Glass.
2016. Unsupervised learning of spoken language
with visual context. In Advances in Neural Information Processing Systems. pages 1858–1866.
Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian
Sun. 2015. Deep residual learning for image recognition. arXiv:1512.03385 .
Micah Hodosh, Peter Young, and Julia Hockenmaier.
2013. Framing image description as a ranking task:
Data, models and evaluation metrics. Journal of Artificial Intelligence Research 47:853–899.
Ákos Kádár, Grzegorz Chrupała, and Afra Alishahi.
2016.
Representation of linguistic form and
function in recurrent neural networks.
CoRR
abs/1602.08952.
Andrej Karpathy and Li Fei-Fei. 2015. Deep visualsemantic alignments for generating image descriptions. In Proceedings of the IEEE Conference on
Computer Vision and Pattern Recognition. pages
3128–3137.
Diederik P. Kingma and Jimmy Ba. 2014.
A method for stochastic optimization.
abs/1412.6980.
Adam:
CoRR
Angeliki Lazaridou, Grzegorz Chrupała, Raquel
Fernández, and Marco Baroni. 2016. Multimodal
semantic learning from child-directed input. In
The 15th Annual Conference of the North American
Chapter of the Association for Computational Linguistics: Human Language Technologies.
Jiwei Li, Will Monroe, and Dan Jurafsky. 2016. Understanding neural networks through representation
erasure. arXiv preprint arXiv:1612.08220 .
Tsung-Yi Lin, Michael Maire, Serge Belongie, James
Hays, Pietro Perona, Deva Ramanan, Piotr Dollár,
and C Lawrence Zitnick. 2014. Microsoft coco:
Common objects in context. In Computer Vision–
ECCV 2014, Springer, pages 740–755.
Marco Marelli, Stefano Menini, Marco Baroni, Luisa
Bentivogli, Raffaella Bernardi, and Roberto Zamparelli. 2014. A sick cure for the evaluation of compositional distributional semantic models. In LREC.
pages 216–223.
Yajie Miao, Jinyu Li, Yongqiang Wang, Shi-Xiong
Zhang, and Yifan Gong. 2016. Simplifying long
short-term memory acoustic models for fast training
and decoding. In IEEE International Conference on
Acoustics, Speech and Signal Processing (ICASSP).
IEEE, pages 2284–2288.
Aaron van den Oord, Nal Kalchbrenner, and Koray
Kavukcuoglu. 2016. Pixel recurrent neural networks. arXiv preprint arXiv:1601.06759 .
Deb K Roy and Alex P Pentland. 2002. Learning words
from sights and sounds: a computational model.
Cognitive Science 26(1):113 – 146.
Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause,
Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein,
Alexander C. Berg, and Li Fei-Fei. 2014. ImageNet
Large Scale Visual Recognition Challenge.
Karen Simonyan and Andrew Zisserman. 2014. Very
deep convolutional networks for large-scale image
recognition. CoRR abs/1409.1556.
Jeffrey M. Siskind. 1996. A computational study of
cross-situational techniques for learning word-tomeaning mappings. Cognition 61(1-2):39–91.
Gabriel Synnaeve, Maarten Versteegh, and Emmanuel
Dupoux. 2014. Learning words from images and
speech. In NIPS Workshop on Learning Semantics,
Montreal, Canada.
Zhiyuan Tang, Ying Shi, Dong Wang, Yang Feng,
and Shiyue Zhang. 2016. Memory visualization for
gated recurrent neural networks in speech recognition. arXiv preprint arXiv:1609.08789 .
Ivan Vendrov, Ryan Kiros, Sanja Fidler, and Raquel
Urtasun. 2015. Order-embeddings of images and
language. arXiv preprint arXiv:1511.06361 .
Chen Yu and Dana H Ballard. 2004. A multimodal
learning interface for grounding spoken language in
sensory perceptions. ACM Transactions on Applied
Perception (TAP) 1(1):57–80.
Bolei Zhou, Agata Lapedriza, Jianxiong Xiao, Antonio Torralba, and Aude Oliva. 2014. Learning deep
features for scene recognition using places database.
In Z. Ghahramani, M. Welling, C. Cortes, N. D.
Lawrence, and K. Q. Weinberger, editors, Advances
in Neural Information Processing Systems 27, Curran Associates, Inc., pages 487–495.
Julian Georg Zilly, Rupesh Kumar Srivastava,
Jan Koutnı́k, and Jürgen Schmidhuber. 2016.
Recurrent highway networks.
arXiv preprint
arXiv:1607.03474 .
| 2cs.AI
|
1
Multi-Cell-Aware Opportunistic Random Access
for Machine-Type Communications
Huifa Lin, Member, IEEE, and Won-Yong Shin, Senior Member, IEEE
arXiv:1708.02861v1 [cs.IT] 9 Aug 2017
Abstract
Due to the difficulty of coordination in multi-cell random access, it is a practical challenge how to achieve the optimal
throughput with decentralized transmission. In this paper, we propose a decentralized multi-cell-aware opportunistic random
access (MA-ORA) protocol that achieves the optimal throughput scaling in an ultra-dense K -cell random access network
with one access point (AP) and N users in each cell, which is suited for machine-type communications. Unlike opportunistic
scheduling for cellular multiple access where users are selected by base stations, under our MA-ORA protocol, each user
opportunistically transmits with a predefined physical layer data rate in a decentralized manner if the desired signal power to
the serving AP is sufficiently large and the generating interference leakage power to the other APs is sufficiently small (i.e., two
threshold conditions are fulfilled). As a main result, it is proved that the aggregate throughput scales as K
(1−ǫ) log(snr log N )
e
K−1
in a high signal-to-noise ratio (SNR) regime if N scales faster than snr 1−δ for small constants ǫ, δ > 0. Our analytical result
is validated by computer simulations. In addition, numerical evaluation confirms that under a practical setting, the proposed
MA-ORA protocol outperforms conventional opportunistic random access protocols in terms of throughput.
Index Terms
Decentralized transmission, inter-cell interference, machine-type communications (MTC), multi-cell-aware opportunistic
random access (MA-ORA), multiuser diversity, throughput scaling.
✦
1
I NTRODUCTION
Recently, random access in wireless communications is receiving growing attention due to the fast development of machinetype communications (MTC) and Internet of Things networks with the necessity of a relatively low protocol overhead and
high spectral efficiency [1], [2]. For several decades, various random access protocols have been developed based on
ALOHA and its variation with carrier sensing [3]. In MTC, transmission activity of vast devices tends to be irregular and
unpredictable with short packets [1]. Under the assumption of such a traffic pattern, state-of-the-art media access control
(MAC) protocols such as carrier-sense multiple access with collision avoidance (CSMA/CA) are not suitable because of a
great volume of protocol overheads [4], and rather uncoordinated random access protocols that incur much less protocol
overheads (e.g., slotted ALOHA) become favorable. However, the major problem of slotted ALOHA is its low MAC layer
efficiency. To solve this problem, there have already been research efforts on improving the MAC throughput by introducing
coded slotted ALOHA [5] and cooperative slotted ALOHA for multiple base station systems [6].
Furthermore, along with the explosive growth of users and their generated data packets in MTC, there is a trend of
densely deploying access points (APs), e.g., network densification in ultra-dense networks (UDN) [7]. Thus, it is crucial
to fully understand the nature of random access networks that consist of multiple cells sharing the same frequency band,
so called multi-cell random access networks.1 In such networks, besides the intra-cell collision (simultaneous transmission
from multiple users in the same cell), transmission without coordinated scheduling among APs will cause interference to
other-cell APs, which may lead to a failure of packet decoding at the receivers. Hence, inter-cell interference should be
carefully managed in multi-cell random access networks. In this paper, we address a challenging and fundamental issue of
multi-cell random access for MTC.
•
•
H. Lin is with the Communications & Networking Laboratory, Dankook University, Yongin 16890, Republic of Korea.
Email: [email protected].
W.-Y. Shin is with the Department of Computer Science and Engineering, Dankook University, Yongin 16890, Republic of Korea.
E-mail: [email protected]
1. Here, we use the term “cell” to denote the domain of an AP and the associated users.
2
1.1 Related Work
On the one hand, there are extensive studies on handling interference management of cellular networks with multiple base
stations [8], [9]. While it has been elusive to find the optimal strategy with respect to the Shannon-theoretic capacity in
multiuser cellular networks, interference alignment (IA) was recently proposed for fundamentally solving the interference
problem when there are multiple communication pairs [10]. It was shown that IA can asymptotically achieve the optimal
degrees of freedom, which are equal to K
2 , in the K -user interference channel with time-varying coefficients. Subsequent
work showed that interference management schemes based on IA can be well applicable to various communication
scenarios [11], [12], [13], including interfering multiple access networks [14], [15], [16]. In addition to the multiple access
scenarios in which collisions can be avoided, it is of significant importance how to manage interference in random access.
For multi-cell random access networks, several studies were carried out to manage interference by performing IA [17],
[18], [19] or successive interference cancellation (SIC) [6], [20]. In [21], [22], decentralized power allocation approaches were
introduced by means of interference mitigation for random access with capabilities of multi-packet reception and SIC at
the receiver.
On the other hand, there have been a great deal of studies on the usefulness of fading in single-cell broadcast channels
by exploiting the multiuser diversity gain as the number of users is sufficiently large, where opportunistic scheduling
[23], opportunistic beamforming [24], and random beamforming [25] were employed. Moreover, scenarios obtaining the
multiuser diversity gain were studied in multi-cell environments. In particular, opportunism can be utilized in multi-cell
broadcast networks through a simple extension of [25]. More recently, for multi-cell multiple access networks, the optimal
throughput scaling was analyzed by showing that the full multiuser diversity gain can be achieved by a distributed user
scheduling strategy in each cell, provided that scheduling criteria are properly determined and the number of users in each
cell is larger than a certain level [26]. Besides the aforementioned multiple access scenarios, the benefits of opportunistic
transmission can also be exploited in random access networks. The idea of single-cell-aware opportunistic random access
(SA-ORA) (also termed channel-aware slotted ALOHA in the literature) was proposed for slotted ALOHA random access
networks with a single AP [27], [28]. By assuming that channel state information (CSI) can be acquired at the transmitters,
the SA-ORA protocols in [27], [28] were shown to achieve the multiuser diversity gain without any centralized scheduling.
This idea was further extended to slotted ALOHA random access networks with imperfect CSI [29], a scenario with
discontinuous channel measurements [30], carrier sense multiple access networks [31], and multichannel wireless networks
[32], [33]. Nevertheless, all the protocols in [27], [28], [29], [30], [31], [32], [33] deal only with the single-AP problem, and
thus cannot be straightforwardly applied to multi-cell random access networks where the inter-cell interference exists.
1.2 Motivation and Contributions
In this paper, we consider an ultra-dense K -cell slotted ALOHA random access network, consisting of one AP and N users in
each cell, which is suited for MTC. We then introduce a decentralized multi-cell-aware opportunistic random access (MA-ORA)
protocol that achieves the optimal throughput scaling by effectively exploiting the multiuser diversity gain in the network
model. Precisely, not only a multiplexing gain of K
e but also a power gain of log log N can be achieved in the K -cell slotted
ALOHA random access network under consideration. First of all, it is worthwhile to address the fundamental differences
between our MA-ORA protocol and the two aforementioned different types of opportunistic transmission protocols:
•
•
Unlike opportunistic scheduling in [23], [24], [25], [26] for cellular multiple access environments where base stations
select users based on feedback information, in our MA-ORA protocol designed for random access, both the intra-cell
collision and inter-cell interference are mitigated solely by users’ opportunistic transmission in a decentralized manner.
The SA-ORA protocol in [27] was shown to achieve the multiuser diversity gain (i.e., the power gain) for single-AP
slotted ALOHA random access, but its extension to multi-cell random access is not straightforward due to the existence
of inter-cell interference. Moreover, in addition to the power gain, it remains open in the literature how to provide the
K -fold increase in the multiplexing gain via properly mitigating the inter-cell interference in the K -cell slotted ALOHA
random access network.
In our ultra-dense multi-cell random access network, users in each cell contend for the same channel at random without
centralized coordination from the serving AP. Consequently, without carefully designing a random access protocol, it is
impossible to entirely avoid the inter-cell interference as well as the intra-cell collision, which may result in a failure of
packet decoding at the receivers. Under our network model, each user needs to determine whether to transmit or not
by itself, without any centralized coordination from the serving AP. The nature of such random access imposes another
3
difficulty on the protocol design. We thus aim to respond to these challenges by introducing the MA-ORA protocol.
To do so, we first assume that uplink channel gains to multiple APs are available at the transmitters by exploiting the
uplink/downlink reciprocity in time-division duplex (TDD) mode. We utilize this partial CSI at the transmitter (CSIT) (but
not the global CSIT) to design our protocol. In the initialization phase, two thresholds and a physical layer (PHY) data
rate are computed offline and are broadcast over the network as system parameters. Thereafter, each user in a cell first
estimates the uplink channel gains via the downlink channel in each time slot. Then, each user determines whether both
1) the channel gain to the serving AP is higher than one threshold and 2) the total inter-cell interference leakage generated
by this user to the other APs is lower than another threshold. Users opportunistically transmit with the PHY data rate if
the above two conditions are fulfilled. By virtue of such opportunistic transmission, when N is large in our ultra-dense
random access setup, we are able to successfully decode the desired packets sent from multiple users in different cells with
high probability, while guaranteeing the optimal throughput scaling. Note that during the communication phase, no control
signaling from the APs is required, i.e., all the users independently perform opportunistic transmission. To the best of our
knowledge, multi-cell random access in a PHY perspective has not been well investigated before in the literature.
Our main results are twofold and summarized as follows.
•
•
In the ultra-dense K -cell slotted ALOHA random access network, it is shown that the aggregate throughput achieved
by the proposed MA-ORA protocol scales as K
e (1 − ǫ) log(snr log N ) in a high signal-to-noise ratio (SNR) regime,
K−1
provided that N scales faster than snr 1−δ for an arbitrarily small constant ǫ > 0 and a constant 0 < δ < 1. This reveals
that even without any centralized scheduling from the APs, the proposed protocol is able to achieve the full multiuser
diversity gain, while obtaining the multiplexing gain of 1e in a cell, which is the best we can hope for under slotted
ALOHA-type protocols [3].
Our analytical result is validated by numerically evaluating the aggregate throughput through Monte-Carlo simulations. We evaluate the throughput in finite SNR (or N ) regimes. Under a practical setting, it is also shown that our
MA-ORA protocol with a slight modification outperforms the conventional SA-ORA protocol in terms of throughput
for almost all realistic SNR regimes. In addition, to check the robustness of our MA-ORA protocol in the presence of
channel uncertainty, we perform simulations under the assumption of imperfect partial CSIT. It is examined that the
MA-ORA protocol with the imperfect partial CSIT achieves comparable performance on the aggregate throughput to
the case with the perfect partial CSIT if the amount of uncertainty is below a tolerable level and still outperforms the
SA-ORA protocol.
Some interference management protocols for multi-cell random access that employ IA and/or the AP cooperation [18],
[20] showed implementation successes based on industrial standards such as IEEE 802.11. Our MA-ORA protocol also
sheds important insights into a simple implementation of multi-cell random access, since neither the dimension expansion
nor the AP cooperation is required.
1.3 Organization and Notations
Section 2 presents system and channel models. Section 3 describes the proposed MA-ORA protocol. Section 4 shows how
to asymptotically achieve the optimal aggregate throughput scaling and the corresponding user scaling law. Section 5
provides numerical results to validate our analysis. Section 6 summarizes the paper with some concluding remarks.
Throughout the paper, C denotes the field of complex numbers. We use the following asymptotic notation: g(x) =
Ω(f (x)) means that there exist constants C and c such that g(x) ≤ Cf (x) for all x > c.
2
S YSTEM AND C HANNEL M ODELS
As illustrated in Fig. 1, we consider an ultra-dense multi-cell random access network consisting of K ≥ 1 APs using
the same frequency band, where N users are served in each cell and N is sufficiently large. We assume that there is no
cooperation among the APs for decoding, i.e., each AP attempts to decode the received packets from the belonging users
independently.2 All the users and APs are equipped with a single antenna. A slotted ALOHA-type protocol is adopted
and we assume perfect slot-level synchronization not only between the users and the serving AP but also among the APs.
We assume fully-loaded traffic such that each user has a non-empty queue of packets to transmit, similarly as in [34]. For
each user, a head-of-line packet is transmitted with probability p at random, regardless of the number of retransmissions,
2. Unlike [6], we do not utilize decoding cooperation among APs but our developed idea can be extended to to another framework allowing
cooperation among APs as future work.
4
User i
in cell j
AP j
Cell k
AP k
Fig. 1. The system model of an ultra-dense K -cell random access network with one AP and N users in each cell.
i.e., each packet is assumed to be the same for all retransmission states. We adopt a modified signal-to-interference-plusnoise ratio (SINR) capture model, where each AP is able to decode the received packet if its SINR exceeds a given decoding
threshold, while treating the inter-cell interference (the interfering signals from the other-cell users) as noise. The concurrent
intra-cell transmission such that two or more users in the same cell simultaneously transmit causes collision, and thus the
corresponding receiver (AP) fails to decode any packet. That is, in order to simplify system modeling and protocol design,
we do not adopt multi-packet reception and multiuser detection studied in [35], [36].
Let hij→k denote the fading channel coefficient from the i-th user in the j -th cell to the k -th AP for i ∈ {1, · · · , N } and
j, k ∈ {1, · · · , K}, where hij→k ∈ C is modeled by an independent and identically distributed (i.i.d.) complex Gaussian
random variable. In the UDN scenario where the APs and users are densely located, the large-scale path-loss components
from users to APs are assumed to be almost identical and thus are omitted. We assume the partial CSIT such that the users
can acquire the uplink channel gains to multiple APs. For instance, the channel gain from the i-th user in the j -th cell to the
i
k -th AP, denoted by gj→k
= |hij→k |2 , is available at the i-th user. The partial CSIT can be acquired via the downlink channel
by exploiting the uplink/downlink reciprocity in TDD mode. Practical CSIT acquisition methods have also been introduced
for various multi-cell multi-antenna systems [37], [38] and wireless local area networks [39], [40]. In the previous seminal
literature on SA-ORA with one AP deployment [27], [28], the partial CSIT and the corresponding distribution information
were assumed to be available. Recently, the CSIT acquisition process was described in more detail for the multichannel
SA-ORA [32]. Inspired by [27], [28], [32], a channel gain acquisition process for our model deploying multiple APs (i.e., our
multi-cell random access model) is described as follows. Under the slotted ALOHA protocol, each AP broadcasts (0, 1, e)
feedback to inform the belonging users of the reception status after each time slot via the downlink channel, where 0 means
that no packet is received (idle); 1 means that only one packet is received (successful transmission); and e indicates that two
or more packets are transmitted simultaneously (collision). In our multi-cell random access scenario, each AP broadcasts
the feedback message in an orthogonal mini-time slot, which requires a small amount of coordination among the APs.
By exploiting the uplink/downlink channel reciprocity, each user is capable of estimating the channel gains to multiple
APs through the received feedback messages. That is, it is possible for each user to perform multi-cell-aware channel gain
estimation. It is worth noting that since only the amplitude information of the CSIT (but not the phase information) is
required in our protocol, the length of feedback messages can be greatly shortened by using quantization.
We consider a quasi-static fading model, i.e., the channel coefficients are constant during one time slot and vary
independently in the next time slot. The received signal yk ∈ C at the k -th AP is given by
yk =
nk
X
π(u ) π(uk )
hk→kk xk
uk =1
|
π(u )
{z
desired signal
}
+
nj
K X
X
j=1
j6=k
|
π(u ) π(uj )
hj→kj xj
+zk ,
(1)
uj =1
{z
inter-cell interference
}
where xk k is the transmitted signal from the π(uk )-th user in the k -th cell and the binomial random integer nk ∼ B(N, p)
is the number of transmitting users in the k -th cell. The received signal is corrupted by the i.i.d. complex additive white
Gaussian noise (AWGN)
zk ∈C with zero-mean and the variance N0 . For each transmission, there is an average transmit
power constraint E
π(uk ) 2
xk
≤ PTX . We define the average SNR at each receiver as snr ,
PTX
N0 .
5
Users (Transmitters)
APs (Receivers)
Initialization phase:
Broadcast
over the network
2. Receive three parameters
broadcasted by APs
1. Obtain optimal
offline
.
.
.
3. Estimate uplink channel
gains and transmit
a packet opportunistically
Data packet
Feedback
message
4. Receive and
decode the packet
.
.
.
Fig. 2. The initialization phase of the proposed MA-ORA protocol.
3
M ULTI -C ELL -AWARE O PPORTUNISTIC R ANDOM ACCESS (MA-ORA)
In this section, we describe the entire procedure of our MA-ORA protocol including the selection of system parameters.
Under the proposed MA-ORA protocol, users opportunistically transmit with a properly selected PHY data rate R if
both the channel gain to the serving AP exceeds a given threshold ΦG and the sum of channel gains to the other APs is
below another given threshold ΦI . Compared to the conventional SA-ORA protocols that aim at enhancing the desired
signal power by only using ΦG , in our ultra-dense multi-cell setup, confining the inter-cell interference leakage to a given
sufficiently low level is more crucial to achieve the multiplexing gain. Hence, provided with the channel gains to other-cell
APs, we introduce another threshold ΦI , used to control the inter-cell interference leakage by exploiting the opportunism.
Our MA-ORA protocol operates in a decentralized manner without any additional control signaling from the APs after the
initialization phase, and hence fits well into ultra-dense random access networks. The detailed description of the protocol
is elaborated on in the following subsections.
3.1 The Overall Procedure
In this subsection, we describe the overall procedure of the proposed MA-ORA protocol. As illustrated in Fig. 2, in the
initialization phase, the APs broadcast two thresholds ΦG and ΦI , as well as the PHY data rate R for opportunistic
transmission in our ultra-dense multi-cell random access network. During the data communication phase, as shown in Fig.
3, it is examined that the maximum MAC throughput of the conventional slotted ALOHA protocol deploying one AP is
achieved at the transmission probability p = N1 for large N (see, e.g., the MAC throughput performance for N = 100 in
Fig. 3).3 Similarly, in our MA-ORA protocol, the transmission probability p is also set to N1 to avoid excessive intra-cell
collisions or idle time slots, thus enabling us to find a relationship between ΦG and ΦI (to be discussed in Section 3.2).
In each time slot, each user first estimates the uplink channel gains by using the feedback messages sent from the APs.
Then, for i ∈ {1, · · · , N } and j ∈ {1, · · · , K} (i.e., for all the users), the i-th user in the j -th cell compares the channel gains
with the given two thresholds to examine whether the following two inequalities are fulfilled:
i
gj→j
≥ ΦG
and
K
X
i
gj→k
≤ ΦI ,
(2)
(3)
k=1
k6=j
where (2) indicates a “good” channel condition to the serving AP that leads to a large desired signal power and (3) means
that the inter-cell interference leakage generated by this user is well confined due to a “weak” channel condition to the
3. The MAC throughput here is defined as the average number of successful transmissions per time slot, where only one user transmits in a
cell, and thus is given by N p(1 − p)N−1 .
6
Fig. 3. The MAC throughput versus transmission probability p, where N = 100 and the conventional slotted ALOHA with one AP deployment is
assumed.
other APs. In each cell, users satisfying both (2) and (3) transmit with the PHY data rate R (to be selected in Section 3.3),
while the other users keep idle in this time slot. Each AP then receives and decodes the desired packet by treating all the
interference as noise. By virtue of such opportunistic transmission, when N is sufficiently large in our ultra-dense random
access network, it is possible for the desired packets that are simultaneously sent from multiple users in different cells to
be successfully decoded at the receivers with high probability.
Note that our MA-ORA protocol operates for general K values. As a special case, for K = 1, each user just checks
whether the condition (2) is fulfilled or not, which corresponds to the conventional SA-ORA protocol.
3.2 The Selection of Two Thresholds
In this subsection, we show how to select the thresholds ΦG and ΦI . First, according to the opportunistic transmission
conditions (2) and (3), the probability that each user accesses the channel is expressed as
K
X
i
i
≥ ΦG ,
≤ ΦI
gj→k
p = Pr gj→j
k=1
k6=j
i
= Pr gj→j
K
X
i
≥ ΦG Pr
gj→k
≤ ΦI ,
k=1
k6=j
where the second equality holds since the channel gains to different APs are independent of each other. From the fact that
p is set to N1 , we have
K
X
1
i
i
Pr gj→j
≥ ΦG Pr
gj→k
≤ ΦI = .
N
k=1
k6=j
Then, the relationship between ΦG and ΦI is given by
ΦG = FG−1 1 − (FI (ΦI )p−1 )−1
= FG−1 1 − (FI (ΦI )N )−1 ,
i
where FG and FI denote the cumulative distribution functions (CDFs) of gj→j
and
PK
i
gj→k
, respectively.
In our MA-ORA protocol, the threshold ΦI is set to snr−1 so that the optimal throughput scaling (i.e., the multiplexing
gain of 1e in a cell and the power gain) can be achieved for our slotted ALOHA random access as snr increases (to be
proven in Section 4). More precisely, due to the fact that the SINR at the receivers is lower-bounded by
snr · ΦG
,
1 + snr · ΦI
k=1
k6=j
(4)
7
the term snr · ΦI in (4) should scale as a positive constant independent of snr in order to achieve the multiplexing gain of
1
e in a cell under the slotted ALOHA-type protocol. In consequence, the two thresholds are given by
(
ΦI = snr−1
ΦG = FG−1 1 − (FI (snr−1 )N )−1 .
3.3 The Selection of PHY Data Rate
In this subsection, we show how to optimally select the PHY data rate R in terms of maximizing the throughput scaling
law. We start from computing the resulting successful decoding probability. At the receiver, even if an AP receives only
one packet from one of the belonging users, this packet may still be corrupted by the noise and the inter-cell interference.
Thus, it is required that the received SINR of the desired packet exceeds a certain decoding threshold given by 2R − 1. The
successful decoding probability ps is then expressed as
i
PTX gj→j
R
ps = Pr
>
2
−
1
,
nk
K P
P
π(u)
N0 +
PTX gk→j
k=1
k6=j
(5)
u=0
where the binomial random variable nk ∼ B(N, p) is the number of simultaneously transmitting users in the k -th cell and
π(u) denotes the index of transmitting users in each cell. Using the successful decoding probability in (5), the throughput
(j)
at the j -th AP for j ∈ {1, · · · , K}, denoted by Rsum , is given by
(j)
Rsum
= N p(1 − p)N −1 ·R · ps ,
|
{z
}
MAC throughput
where N p(1 − p)N −1 is the MAC throughput and R is the target PHY data rate. From the fact that p =
throughput of the K -cell random access network is given by
Rsum =
K
X
1
N,
the aggregate
(j)
Rsum
j=1
1 N −1
· R · ps
=K 1−
N
1 N −1
ΦG
R
≥K 1−
· R · Pr
>
2
−
1
,
N
snr−1 + ñΦI
(6)
where ñ ∼ B((K − 1)N, p) is a binomial random variable, representing the total number of interfering signals from the
PK
other cells, and is given by ñ = k=1 nk . Here, the inequality comes from (2) and (3).
k6=j
Now, we focus on computing a lower bound on the successful decoding probability, denoted by p̃s , shown below:
ΦG
R
>
2
−
1
p̃s = Pr
snr−1 + ñΦI
(7)
(K−1)N
X
ΦG
R
Pr
=
> 2 − 1 Pr(ñ = i).
snr−1 + iΦI
i=0
Let us consider an integer ν ∈ {0, 1, · · · , (K − 1)N }. If R is set to a value such that
ΦG
ΦG
< 2R − 1 ≤
,
−1
+ (ν + 1)ΦI
snr + νΦI
ΦG
R
then the probability Pr snr−1
≥
2
−
1
is given by 1 and 0 for i ∈ {0, 1, · · · , ν} and i ∈ {ν + 1, · · · , (K − 1)N },
+iΦI
respectively. Based on this observation, the entire feasible range of 2R − 1 (i.e., the decoding threshold) can be divided into
the following (K − 1)N + 1 sub-ranges:
ΦG
,··· ,
0,
snr−1 + (K − 1)N ΦI
ΦG
ΦG
ΦG
,
,∞
.
,··· ,
snr−1 + (ν + 1)ΦI snr−1 + νΦI
snr−1
snr−1
8
ΦG
( snr
−1 , ∞),
In particular, for R ∈
wehave p̃s = 0, which is thus neglected in our work.
ΦG
R
(7) is an indicator function of R, we set R to the maximum
Using the fact that the term Pr snr−1
+iΦI ≥ 2 − 1 in
i
ΦG
R
G
value under the condition that 2 − 1 lies in each sub-range snr−1 +(ν+1)Φ
, snr−1Φ+νΦ
, which is given by
I
I
R = log2 1 +
ΦG
snr−1 + νΦI
.
(8)
Here, the parameter ν ∈ {0, 1, · · · , (K − 1)N } in (8) can be interpreted as a tolerable number of interfering signals from the
other-cell users. That is, a relatively high successful decoding probability can be guaranteed even when the desired signal
at the receiver is interfered by ν signals caused by the other-cell users. Note that we can improve p̃s for each transmitted
packet at the cost of a lower R (corresponding to a larger ν ). In other words, we can trade a lower PHY data rate for a
higher successful decoding probability (refer to (7)).
Remark 1: It is worth noting that the optimal PHY data rate can be found by solving the aggregate throughput
maximization problem as follows: R̂ = arg max Rsum (R). However, as an alternative approach, it is sufficient to set R
R
to (8) in the sense of guaranteeing the optimal throughput scaling because i) selecting a proper finite ν leads to a successful
decoding probability ps approaching 1, which plays a crucial role in showing the optimality of our protocol, and ii) the
expression in (8) is analytically tractable and thus enables us to derive a closed-form expression of the aggregate throughput
scaling law. Detailed analytical discussions are addressed in the next section.
4
A NALYSIS OF AGGREGATE T HROUGHPUT S CALING L AW
In this section, we show that the proposed MA-ORA protocol asymptotically achieves the optimal throughput scaling,
i.e., K
e (1 − ǫ) log(snr log N ) for an arbitrarily small constant ǫ > 0, in the ultra-dense K -cell slotted ALOHA random
access network, provided that N scales faster than a certain level. To be specific, we first provide some preliminaries
including a simplification of p̃s and a tractable lower bound on FI (x). We then present our main result by analyzing
the aggregate throughput scaling under a certain user scaling condition. In addition, we show an upper bound on the
aggregate throughput scaling and compare the proposed MA-ORA protocol with the existing opportunistic scheduling
approach for the interfering multiple access channel (IMAC).
4.1 Preliminaries
In this subsection, we introduce two important lemmas to present our main result. To obtain the first lemma, we start
from revisiting the lower bound on the aggregate throughput in (6). To simplify this lower bound, we derive an explicit
i
expression of ΦG as a function of ΦI . Due to the fact that the channel gain gj→j
follows the exponential distribution whose
−x
CDF is given by FG (x) = 1 − e , it is possible to establish the following relationship between ΦG and ΦI :
ΦG = FG−1 1 − (FI (ΦI )p−1 )−1
(9)
= ln (FI (ΦI )N ) .
By using (9) in (6), we have
ln (FI (ΦI )N )
1 N −1
R
>
2
−
1
.
Rsum ≥K 1 −
·R · Pr
N
snr−1 + ñΦI
|
|
{z
}
{z
}
MAC throughput
(10)
p̃s
For analytical convenience, the following lemma is presented since the resulting form of p̃s is still not analytically
tractable.
Lemma 1. When the PHY data rate R is selected from the discrete set in (8), we have
p̃s = I1− N1 ((K − 1)N − ν, ν + 1) ,
(11)
where Ix (y, z) is the regularized incomplete beta function.
Proof:
As shown in (8),the PHY data rate R is a function of the integer ν ∈ {0, 1, · · · , (K − 1)N }. We recall that the term
ΦG
R
Pr snr−1
+iΦI ≥ 2 − 1 in (7) is given by 1 and 0 for i ∈ {0, · · · , ν} and i ∈ {ν + 1 · · · , (K − 1)N }, respectively, if R is
set to a value such that
ΦG
ΦG
< 2R − 1 ≤
.
snr−1 + (ν + 1)ΦI
snr−1 + νΦI
9
Fig. 4. The function p̃s = I1−
1
N
((K − 1)N − ν, ν + 1) versus ν for K = {2, 3, 4}.
Thus, by dividing the whole range of i into two sub-ranges, (7) can be expressed as
ν
X
p̃s =
i=0
ν
X
=
(K−1)N
1 · Pr(ñ = i) +
X
0 · Pr(ñ = i)
i=ν+1
Pr(ñ = i)
i=0
= Pr(ñ ≤ ν),
which is the CDF of the binomial random variable ñ ∼ B((K − 1)N, N1 ). It is known that this CDF can be expressed as the
regularized incomplete beta function [41], which is given by
p̃s = I1− N1 ((K − 1)N − ν, ν + 1) .
This completes the proof of this lemma.
From this lemma, we are able to equivalently transform the expression of p̃s in (10) to the well-known regularized
incomplete beta function. In Fig. 4, the function I1− 1 ((K − 1)N − ν, ν + 1) versus ν is plotted for various K = {2, 3, 4}.
N
This function tends to get steeply increased for small ν and then to gradually approach one. That is, it is monotonically
increasing with ν . For example, for K = 2, it is numerically found that the pairs of (p̃s , ν) are given by (0.9, 3), (0.99, 5),
and (0.999, 7). Based on this observation, it is possible to select a proper finite value of ν that makes p̃s approach almost
one.
Using (8), (10), and (11), the aggregate throughput is now lower-bounded by
1 N −1
ln (FI (ΦI )N )
Rsum ≥K 1 −
· log2 1 +
N
snr−1 + νΦI
|
{z
}
(12)
PHY data rate (R)
· I1− N1 ((K − 1)N − ν, ν + 1) .
However, it is still not easy to find a closed form expression of (12) due to the complicated form of FI (ΦI ). This
motivates us to introduce the following lemma.
Lemma 2. For any 0 ≤ x < 2, the CDF of
K
P
k=1
k6=j
i
gj→k
, FI (x), is lower-bounded by
FI (x) ≥ c1 x(K−1) ,
where c1 =
e−1 2−(K−1)
(K−1)Γ(K−1)
is a constant independent of N and SNR. Here, Γ(x) = (x − 1)! is the Gamma function.
Proof: We refer to [26, Lemma 2] for the proof.
10
Using this lemma leads to a more tractable lower bound on the aggregate throughput, which will be analyzed in the
next subsection.
4.2 Aggregate Throughput Scaling and User Scaling Laws
As our main result, we are now ready to establish the following theorem, which presents the aggregate throughput scaling
law achieved by the proposed MA-ORA protocol.
Theorem 1. Consider the MA-ORA protocol in the ultra-dense K -cell slotted ALOHA random access network. Suppose that ΦI =
snr−1 . Then, the MA-ORA protocol achieves an aggregate throughput scaling of
with high probability in the high SNR regime if
K
(1 − ǫ) log(snr log N )
e
(13)
K−1
N = Ω snr 1−δ ,
(14)
where ǫ > 0 is an arbitrarily small constant and 0 < δ < 1 is a certain constant.
Proof: Using ΦI = snr−1 , (12) can be rewritten as
1 N −1
snr · ln(FI (snr−1 )N )
Rsum ≥K 1 −
· log2 1 +
N
ν +1
(15)
· I1− N1 ((K − 1)N − ν, ν + 1) .
We then select a finite value of ν that makes the function I1− 1 ((K − 1)N − ν, ν + 1) approach almost one by following
N
the equation below:
(16)
I1− 1 ((K − 1)N − ν ∗ , ν ∗ + 1) = 1 − ǫ,
N
4
where ǫ > 0 is an arbitrarily small constant. Since the function I1− 1 ((K − 1)N − ν, ν + 1) is monotonically increasing
N
with ν as shown in Fig. 4, we can find finite ν ∗ (K, ǫ) by taking the inverse of the regularized incomplete beta function in
(16).
By substituting (16) and ν = ν ∗ (K, ǫ) into (15), we have
snr · ln(FI (snr−1 )N )
1 N −1
· log2 1 +
Rsum ≥K 1 −
N
ν ∗ (K, ǫ) + 1
· (1 − ǫ).
N −1
From the fact that 1 − N1
is monotonically decreasing with increasing N and lim 1 −
N →∞
aggregate throughput is lower-bounded by
snr · ln(FI (snr−1 )N )
K
· (1 − ǫ)
· log2 1 +
Rsum ≥
e
ν ∗ (K, ǫ) + 1
!
K−1
snr · ln(c1 snr−1
N)
K
· (1 − ǫ),
· log2 1 +
≥
e
ν ∗ (K, ǫ) + 1
1 N −1
N
=
1
e,
the
where the second inequality holds owing to Lemma 2. In order to achieve the logarithmic gain (i.e., the power gain with
K−1
increasing N ), it is required that c1 snr−1
N ≥ N δ for 0 < δ < 1, which finally yields
K−1
N = Ω snr 1−δ .
(17)
In consequence, under the user scaling condition in (17), the aggregate throughput can be lower-bounded by
K
δ · snr ln(N )
Rsum ≥ (1 − ǫ) · log2 1 + ∗
,
e
ν (K, ǫ) + 1
(18)
∗
which scales as K
e (1 − ǫ) log(snr log N ) since δ and ν (K, ǫ) are some constants independent of N . This completes the
proof of the theorem.
4. Note that ǫ is a given design parameter and ν ∗ can be expressed as a function of given ǫ and K .
11
K
e
Theorem 1 indicates that the proposed MA-ORA protocol can achieve not only the near
multiplexing gain
(corresponding to the pre-log term in (13)) but also the power gain of log log N in our multi-cell random access network.
To obtain such gains, the inter-cell interference leakage generated by a certain user needs to be well confined by setting
ΦI = snr−1 . This implies that a larger N is necessary to confine the total interference to a lower level by virtue of the
multiuser diversity gain for a higher snr. More precisely, in order to achieve the optimal throughput scaling, the number of
per-cell users, N , is required to scale with snr as shown in (14). Based on the analytical result in Theorem 1, the following
interesting observations are provided with respect to the parameters K , δ , and ǫ.
Remark 2: According to (13) and (14), it is found that the aggregate throughput scaling linearly increases with K at the
cost of more stringent user scaling condition (note that N needs to exponentially increase with K for given snr). In ultradense random access networks with many users, our MA-ORA protocol can improve the multiplexing gain by deploying
more APs, but more per-cell users are required to guarantee this improvement.
Remark 3: We now turn to the effect of δ on the aggregate throughput performance. Although increasing δ leads to
the aggregate throughput increment (refer to (18)), it does not fundamentally change the aggregate throughput scaling
law. However, increasing δ yields a more stringent user scaling condition. Thus, it is sufficient to select a proper value of
0 < δ < 1 according to the given network condition.
Remark 4: The pre-log term 1 − ǫ corresponds to the successful decoding probability, which can be interpreted as a
penalty of random access without any coordination. This penalty cannot be totally resolved because there is always a nonzero probability of unsuccessful decoding caused by the excessive inter-cell interference generated by random transmission
(refer to (16)). This is a distinctive phenomenon of multi-cell random access, compared to the IMAC scenario where each
base station performs user selection.
To increase the successful decoding probability, we can decrease ǫ by selecting a larger ν ∗ according to (11). However,
this results in a lower PHY data rate R in (8) that affects the resulting aggregate throughput. Thus, we need to carefully
balance this trade-off between ǫ and ν ∗ when selecting system parameters in practice. Nevertheless, it is sufficient to
assume a finite ν ∗ leading to an arbitrarily small ǫ > 0 to analyze our throughput scaling result.
4.3 Discussions
In this subsection, we first present an upper bound on the aggregate throughput scaling that matches our analytically
achievable result. Then, we compare the proposed MA-ORA protocol with the opportunistic scheduling protocol for the
IMAC [26].
An upper bound on the aggregate throughput scaling: An aggregate throughput scaling of K
e log(snr log N ) can be obtained
by a genie-aided removal of all the inter-cell interference, which is explained as follows. Consider an ideal scenario for
the multi-cell random access network in which concurrent interfering signals sent from other cells are completely canceled
out. This corresponds to a system consisting of K interference-free parallel SA-ORA networks. Since the throughput scales
as 1e log(snr log N ) in each cell of such SA-ORA networks [27], the aggregate throughput scales as K
e log(snr log N ). This
corresponds to an upper bound on the aggregate throughput scaling for our K -cell random access network, which matches
the achievable aggregate throughput scaling in (13) to within a factor of ǫ > 0.
Comparison with the opportunistic scheduling protocol for the IMAC: The aggregate throughput scaling of K log(snr log N )
is achieved by the opportunistic scheduling protocol for the IMAC, where collisions can be avoided by adopting user
selection at each base station [26]. By comparing our MA-ORA protocol with this scheduling protocol, we observe some
interesting consistency as follows: 1) user scaling conditions required by the two protocols are exactly the same and 2)
under this user scaling condition, after removing the intra-cell contention loss factor (note that the MAC throughput in a
cell is given by 1 and 1e in multiple access and slotted ALOHA random access, respectively), these two protocols achieve
the same throughput scaling to within an arbitrary small ǫ > 0 gap. Hence, essential similarities are revealed by applying
opportunism to both multiple access and random access networks.
In contrast to the observed similarities, there are fundamental differences between the two protocols. Under the
opportunistic scheduling protocol in [26], users in each cell who fulfill the given channel conditions send transmission
requests to the base station. Then, each base station randomly selects one user and informs the user of the PHY data rate
for transmission via the downlink channel. On the other hand, under our MA-ORA protocol, without controlling from the
APs, each user determines whether to transmit by comparing its own channel gains with the two thresholds. Besides, the
PHY data rate does not need to be sent from the APs since it is already broadcast in the initialization phase and keeps the
same afterwards. These features make our MA-ORA protocol suitable for uncoordinated random access networks.
12
Fig. 5. The aggregate throughput versus SNR, where the proposed MA-ORA protocol is employed as N scales according to the user scaling
condition.
5
N UMERICAL E VALUATION
In this section, we validate the proposed MA-ORA protocol by intensively performing numerical evaluation through
Monte-Carlo simulations, where the channels in (1) are randomly generated 1 × 105 times for each system parameter.
First, our analytical result is validated by comparing the aggregate throughput through numerical evaluation with
the corresponding theoretical one. Performance of the proposed MA-ORA protocol is also compared with that of the
conventional SA-ORA and slotted ALOHA protocols in practical simulation environments. In addition, performance of
both the MA-ORA and SA-ORA protocols is evaluated under the assumption of imperfect partial CSIT.
5.1 Validation of Analytical Results
In this subsection, we validate our analytical result in an ultra-dense multi-cell random access setup by evaluating the
aggregate throughput [bps/Hz] of the proposed MA-ORA protocol in Sections 3 and 4. Suppose that we set ǫ = 0.01
and δ = 0.1. Then, the parameter ν can be found according to the relationship between ǫ and ν in (16). We assume that
ΦI = snr−1 and ΦG = ln(FI (ΦI )N ) (see (9)). Based on these parameters, the PHY data rate R can be computed from (8).
In Fig. 5, we evaluate the aggregate throughput achieved by the MA-ORA protocol versus snr in dB scale for K = {2, 3}.
K−1
The parameter N is set to a different scalable value according to snr, i.e., N = snr 1−δ in (14). In the figure, the dotted lines
are also plotted from the theoretical result in Theorem 1 with a proper bias to check the slopes of K
e (1 − ǫ) log(snr log N )
for K = {2, 3}. One can see that the slopes of the simulated curves coincide with the theoretical ones in the high SNR
regime. This numerical results are sufficient to guarantee our achievability (i.e., the throughput scaling under the given
user scaling law) in Section 4.
Now, in Fig. 6, we evaluate the aggregate throughput versus N , where K = 2 and snr = {10, 15, 20, 25}dB.5 We adopt
the aforementioned parameter setting except for δ that is used to specify the user scaling condition in (14). One can see that
all the curves tend to increase at most logarithmically with N owing to the multiuser diversity gain, which corresponds to
the throughput scaling in (13). It is also observed that for large N , increasing the snr leads to superior performance on the
aggregate throughput due to the power gain. As illustrated in the figure, it is worth noting that this superior throughput
performance can be achieved when N is sufficiently large (or the user scaling law in (14) is fulfilled).
5.2 Performance Evaluation in a Practical Setting
In this subsection, to further ascertain the efficacy of our MA-ORA protocol, performance on the aggregate throughput is
evaluated in finite SNR (or N ) regimes. Under this practical setting, instead of using the original MA-ORA protocol, we
5. Even if it seems unrealistic to have a great number of users even in ultra-dense random access models, the wide range of parameter N is
taken into account to precisely see some trends of curves varying with N .
13
Fig. 6. The aggregate throughput versus N , where the proposed MA-ORA protocol is employed for K = 2.
K\N
2
3
4
2
3
4
2
3
4
50
100
snr = 10dB
(1.40, 2.32) (1.70, 3.64)
(1.70, 2.14) (1.90, 2.45)
(1.90, 1.54) (2.10, 1.60)
snr = 20dB
(0.90, 4.40) (0.80, 4.88)
(1.70, 2.33) (1.60, 2.77)
(1.70, 1.70) (1.90, 1.89)
snr = 30dB
(0.60, 4.65) (0.65, 5.70)
(1.40, 2.42) (1.45, 3.05)
(1.90, 1.54) (2.00, 1.89)
200
(2.20, 3.91)
(2.30, 2.72)
(2.20, 2.02)
(1.10, 5.61)
(1.30, 3.18)
(2.20, 2.09)
(0.70, 6.57)
(1.50, 3.16)
(1.80, 2.13)
TABLE 1
The lookup table of (Φ∗G , R∗ ) according to various K , N , and snr.
slightly modify the MA-ORA protocol in Sections 3 and 4 by numerically finding the optimal parameters Φ∗G and R∗ in
terms of maximizing the resulting aggregate throughput Rsum (ΦG , R),6 which is given by
(Φ∗G , R∗ ) = argmax Rsum (ΦG , R).
(19)
ΦG ,R
The corresponding Φ∗I can be found by using Φ∗G in (9). The optimal values of (Φ∗G , R∗ ) can be found via exhaustive search
for given parameter configuration including K, N , and snr. The optimal values of (Φ∗G , R∗ ) are summarized in Table 1.
From the table, some insightful observations are made as follows. For given snr and N , the optimal PHY data rate R∗ tends
to decrease with K . This is because more inter-cell interference is generated for larger K , thus leading to a lower SINR at
the receivers and the resulting lower R∗ . Moreover, for given snr and K , R∗ tends to increase with N . This follows from
the fact that for larger N , the inter-cell interference can be better mitigated with the help of the multiuser diversity gain,
thereby leading to a higher SINR at the receivers. This enables us to adopt a higher R∗ .
In Fig. 7, the aggregate throughput of the proposed MA-ORA protocol versus snr in dB scale is plotted for K = 2 and
N = 100. As baseline schemes, performance of the SA-ORA and slotted ALOHA protocols is also
in the figure.
shown
ΦG PTX
Under the SA-ORA protocol, each user opportunistically transmits with the PHY data rate of log2 1 + N0
if its uplink
channel gain exceeds ΦG [27]. This protocol can be treated as a special case of our MA-ORA protocol with K = 1 and thus
leads to worse performance as K ≥ 2. For the slotted ALOHA protocol, since both details of the PHY layer and the effects
of fading are neglected in the protocol design phase, we adopt the PHY data rate of log2 (1 + snr) for fair comparison. It
is observed that the MA-ORA protocol outperforms the slotted ALOHA and SA-ORA protocols in the low and moderate
SNR regimes, and then gets saturated to a certain value in the high SNR regime. This throughput saturation comes from
6. Note that the MA-ORA protocol needs to be slighted modified when it is evaluated in practical settings since it is inherently designed for
asymptotically achieving the optimal throughput scaling.
14
Fig. 7. The aggregate throughput versus snr, where the proposed MA-ORA protocol (K = 2) as well as the conventional SA-ORA and slotted
ALOHA protocols are employed for N = 100.
Fig. 8. The aggregate throughput versus snr, where the proposed MA-ORA protocol is employed for K = {2, 3, 4} and N = 100.
TABLE 2
The lookup table of (Rsum , snr) according to various K and N , where snr denotes a crossover where two curves coincide.
K\N
2
3
4
50
100
150
200
(2.83, 17dB)
(1.91, 9dB)
(1.69, 7dB)
(3.59, 22dB)
(2.35, 12dB)
(2.08, 10dB)
(4.05, 26dB)
(2.68, 14dB)
(2.28, 11dB)
(4.31, 28dB)
(2.88, 16dB)
(2.45, 12dB)
the fact that the multiplexing and multiuser diversity gains are not fully achieved due to the limited N (which is less than
the one required by our user scaling condition) in the network suffering from the severe inter-cell interference.
Figure 8 illustrates the aggregate throughput of the MA-ORA protocol versus snr in dB scale for K = {2, 3, 4} and
N = 100. It is seen that when K becomes large, the MA-ORA protocol achieves superior aggregate throughput in the low
SNR regime, but gets saturated earlier due to a more stringent user scaling condition (note that N needs to exponentially
increase with K for given snr). The curve for K = 2 achieves inferior performance to the other curves for K = {3, 4} in
the low SNR regime, but tends to increase steadily with snr and then get saturated at a relatively high snr point. Thus,
it is worthwhile to investigate a crossover where two curves meet when the aggregate throughput of the MA-ORA and
SA-ORA protocols versus snr is plotted. The crossover snr and the resulting aggregate throughput Rsum (i.e., (Rsum , snr))
are summarized in Table 2 according to various K and N .
15
Fig. 9. The aggregate throughput versus σ2 under the imperfect partial CSIT assumption, where the proposed MA-ORA (K = 2) and SA-ORA
protocols are employed for snr = 10dB and N = 100.
5.3 Performance Evaluation Under the Assumption of Imperfect Partial CSIT
The numerical results in the previous subsections were obtained based on the assumption of perfect partial CSIT (i.e.,
perfect channel gain acquisition). However, it is hardly possible to acquire the perfect partial CSIT due to the channel
estimation or feedback error in practice. To check the robustness of our MA-ORA protocol in the presence of channel
i
uncertainty, we perform simulations by assuming the imperfect channel gain ĝj→k
= |ĥij→k |2 at each transmitter, where
i
i
i
i
ĥj→k = hj→k + ∆hj→k . Here, the error component, ∆hj→k , is modeled as an i.i.d. and complex Gaussian random variable
with zero-mean and the variance σ 2 [42]. In the MA-ORA protocol, we use the optimal parameters Φ∗G and R∗ found based
on (19).
As illustrated in Fig. 9, it is obvious to see that the aggregate throughput of both MA-ORA and SA-ORA protocols
gets degraded as σ 2 increases. It is also observed that in the presence of the error component, our MA-ORA protocol still
outperforms the SA-ORA protocol and achieves acceptable performance if the error variance σ 2 is below a tolerable level
(e.g., σ 2 < 10−2 ). Moreover, when σ 2 is large (e.g., σ 2 > 10−1 ), the performance gap between the MA-ORA and SA-ORA
protocols tends to be wider, showing that our MA-ORA protocol is less sensitive to the channel gain inaccuracy.
6
C ONCLUDING R EMARKS
The MA-ORA protocol operating in a decentralized manner, appropriate for MTC, was proposed for the ultra-dense K cell slotted ALOHA random access network, where no centralized coordination from the serving APs is required. The
aggregate throughput scaling achieved by the proposed protocol was then analyzed. As our main result, it was proved that
the MA-ORA protocol asymptotically achieves the aggregate throughput scaling of K
e (1−ǫ) log(snr log N ) in our multi-cell
K−1
random access network, provided that N scales faster than snr 1−δ for small constants ǫ > 0 and 0 < δ < 1. Extensive
computer simulations were also performed to validate the MA-ORA protocol and its analytical result—the throughput
scaling and user scaling laws were confirmed numerically; the superiority of our protocol over baseline schemes was
shown in a practical setting; and the robustness of our protocol was investigated when imperfect partial CSIT is assumed.
Our random access framework would shed important insights for intelligently solving intra-cell collision and inter-cell
interference problems.
Future research directions include extensions to networks with multiple antennas, opportunistic random access with
user fairness, and cooperative slotted ALOHA systems.
ACKNOWLEDGMENTS
This research was supported by the Basic Science Research Program through the National Research Foundation of Korea
(NRF) funded by the Ministry of Education (2017R1D1A1A09000835) and by the Ministry of Science, ICT & Future Planning
(MSIP) (2015R1A2A1A15054248). The material in this paper has been presented in part at the IEEE International Conference
on Communications 2017 [43] and the IEEE International Symposium on Information Theory 2017 [44]. Won-Yong Shin is
the corresponding author.
16
R EFERENCES
[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]
A. Rajandekar and B. Sikdar, “A survey of MAC layer issues and protocols for machine-to-machine communications,” IEEE Internet Things
J., vol. 2, no. 2, pp. 175–186, Jan. 2015.
C. Bockelmann, N. Pratas, H. Nikopour, K. Au, T. Svensson, C. Stefanovic, P. Popovski, and A. Dekorsy, “Massive machine-type
communications in 5G: physical and MAC-layer solutions,” IEEE Commun. Mag., vol. 54, no. 9, pp. 59–65, Sep. 2016.
D. Bertsekas and R. Gallager, Data networks (2nd Ed.). New Jersey, USA: Prentice-Hall, Inc., 1992.
G. Durisi, T. Koch, and P. Popovski, “Toward massive, ultrareliable, and low-latency wireless communication with short packets,” Proc. IEEE,
vol. 104, no. 9, pp. 1711–1726, Sep. 2016.
E. Paolini, G. Liva, and M. Chiani, “Coded slotted ALOHA: A graph-based method for uncoordinated multiple access,” IEEE Trans. Inf.
Theory, vol. 61, no. 12, pp. 6815–6832, Dec. 2015.
D. Jakovetić, D. Bajović, D. Vukobratović, and V. Crnojević, “Cooperative slotted ALOHA for multi-base station systems,” IEEE Trans.
Commun., vol. 63, no. 4, pp. 1443–1456, Feb. 2015.
S. Chen, R. Ma, H.-H. Chen, H. Zhang, W. Meng, and J. Liu, “Machine-to-machine communications in ultra-dense networks–a survey,” IEEE
Commun. Surveys Tuts., May 2017.
A. D. Wyner, “Shannon-theoretic approach to a Gaussian cellular multiple-access channel,” IEEE Trans. Inf. Theory, vol. 40, no. 6, pp. 1713–
1727, Aug. 1994.
O. Somekh and S. Shamai, “Shannon-theoretic approach to a Gaussian cellular multiple-access channel with fading,” IEEE Trans. Inf. Theory,
vol. 46, no. 4, pp. 1401–1425, Aug. 2002.
V. R. Cadambe and S. A. Jafar, “Interference alignment and degrees of freedom of the K -user interference channel,” IEEE Trans. Inf. Theory,
vol. 54, no. 8, pp. 3425–3441, Jul. 2008.
T. Gou and S. A. Jafar, “Degrees of freedom of the K user M × N MIMO interference channel,” IEEE Trans. Inf. Theory, vol. 56, no. 12, pp.
6040–6057, Nov. 2010.
V. R. Cadambe and S. A. Jafar, “Degrees of freedom of wireless X networks,” in Proc. IEEE Int. Symp. Inf. Theory (ISIT), Toronto, Canada, Jul.
2008, pp. 1268–1272.
H. J. Yang, W.-Y. Shin, B. C. Jung, C. Suh, and A. Paulraj, “Opportunistic downlink interference alignment for multi-cell MIMO networks,”
IEEE Trans. Wireless Commun., vol. 16, no. 3, pp. 1533–1548, Mar. 2017.
C. Suh and D. Tse, “Interference alignment for cellular networks,” in Proc. 46th Annu. Allerton Conf., Monticello, IL, Sep. 2008, pp. 1037–1044.
B. C. Jung and W.-Y. Shin, “Opportunistic interference alignment for interference-limited cellular TDD uplink,” IEEE Commun. Lett, vol. 15,
no. 2, pp. 148–150, Feb. 2011.
B. C. Jung, D. Park, and W.-Y. Shin, “Opportunistic interference mitigation achieves optimal degrees-of-freedom in wireless multi-cell uplink
networks,” IEEE Trans. Commun., vol. 60, no. 7, pp. 1935–1944, Jul. 2012.
S. Gollakota, S. D. Perli, and D. Katabi, “Interference alignment and cancellation,” ACM SIGCOMM Computer Communication Review, vol. 39,
no. 4, pp. 159–170, Aug. 2009.
K. C.-J. Lin, S. Gollakota, and D. Katabi, “Random access heterogeneous MIMO networks,” ACM SIGCOMM Computer Communication Review,
vol. 41, no. 4, pp. 146–157, Oct. 2011.
H. Jin, S.-W. Jeon, and B. C. Jung, “Opportunistic interference alignment for random access networks,” IEEE Trans. Veh. Technol., vol. 64,
no. 12, pp. 5947–5954, Jan. 2015.
W. Zhou, T. Bansal, P. Sinha, and K. Srinivasan, “BBN: Throughput scaling in dense enterprise WLANs with bind beamforming and nulling,”
in Proc. ACM 20th Annu. Int. Conf. on Mobile Computing and Networking, Maui, HI, Sep. 2014, pp. 165–176.
H. Lin, K. Ishibashi, W.-Y. Shin, and T. Fujii, “A simple random access scheme with multilevel power allocation,” IEEE Commun. Lett., vol. 19,
no. 12, pp. 2118–2121, Dec. 2015.
——, “Decentralized power allocation for secondary random access in cognitive radio networks with successive interference cancellation,”
in Proc. of IEEE Int. Conf. Commun. (ICC), Kuala Lumper, Malaysia, May 2016, pp. 1–6.
R. Knopp and P. A. Humblet, “Information capacity and power control in single-cell multiuser communications,” in Proc. IEEE Int. Conf.
Commun. (ICC), Seattle, WA, Jun. 1995, pp. 331–335.
P. Viswanath, D. N. C. Tse, and R. Laroia, “Opportunistic beamforming using dumb antennas,” IEEE Trans. Inf. Theory, vol. 48, no. 6, pp.
1277–1294, Aug. 2002.
M. Sharif and B. Hassibi, “On the capacity of MIMO broadcast channels with partial side information,” IEEE Trans. Inf. Theory, vol. 51, no. 2,
pp. 506–522, Jan. 2005.
W.-Y. Shin, D. Park, and B. C. Jung, “Can one achieve multiuser diversity in uplink multi-cell networks?” IEEE Trans. Commun., vol. 60,
no. 12, pp. 3535–3540, Dec. 2012.
X. Qin and R. A. Berry, “Distributed approaches for exploiting multiuser diversity in wireless networks,” IEEE Trans. Inf. Theory, vol. 52,
no. 2, pp. 392–413, Jan. 2006.
S. Adireddy and L. Tong, “Exploiting decentralized channel state information for random access,” IEEE Trans. Inf. Theory, vol. 51, no. 2, pp.
537–561, Jan. 2005.
S.-H. Wang and Y.-W. P. Hong, “Transmission control with imperfect CSI in channel-aware slotted ALOHA networks,” IEEE Trans. Wireless
Commun., vol. 8, no. 10, pp. 5214–5224, Oct. 2009.
H. Moon, “Channel-adaptive random access with discontinuous channel measurements,” IEEE J. Sel. Areas Commun., vol. 34, no. 5, pp.
1704–1712, Apr. 2016.
G. Miao, Y. Li, and A. Swami, “Channel-aware distributed medium access control,” IEEE/ACM Trans. Netw., vol. 20, no. 4, pp. 1290–1303,
Dec. 2012.
Z. Khanian, M. Rasti, F. Salek, and E. Hossain, “A distributed opportunistic MAC protocol for multichannel wireless networks,” IEEE Trans.
Wireless Commun., vol. 15, no. 6, pp. 1–1, Mar. 2016.
17
[33] Y. Liu and M. Liu, “To stay or to switch: Multiuser multi-channel dynamic access,” IEEE Trans. Mobile Comput., vol. 14, no. 4, pp. 858–871,
Apr. 2015.
[34] Y. Li and L. Dai, “Maximum sum rate of slotted Aloha with capture,” IEEE Trans. Commun., vol. 64, no. 2, pp. 690–705, Feb. 2016.
[35] Y. J. Zhang, P. X. Zheng, and S. C. Liew, “How does multiple-packet reception capability scale the performance of wireless local area
networks?” IEEE Trans. Mobile Comput., vol. 8, no. 7, pp. 923–935, Jul. 2009.
[36] Y. H. Bae, B. D. Choi, and A. S. Alfa, “Achieving maximum throughput in random access protocols with multipacket reception,” IEEE Trans.
Mobile Comput., vol. 13, no. 3, pp. 497–511, Mar. 2014.
[37] P. Komulainen, A. Tölli, and M. Juntti, “Effective CSI signaling and decentralized beam coordination in TDD multi-cell MIMO systems,”
IEEE Trans. Sig. Process., vol. 61, no. 9, pp. 2204–2218, May 2013.
[38] R. Rogalin, O. Y. Bursalioglu, H. Papadopoulos, G. Caire, A. F. Molisch, A. Michaloliakos, V. Balan, and K. Psounis, “Scalable synchronization
and reciprocity calibration for distributed multiuser MIMO,” IEEE Trans. Wireless Commun., vol. 13, no. 4, pp. 1815–1831, Apr. 2014.
[39] “Local and metropolitan area networksspecific requirements; part 11: wireless LAN medium access control (MAC) and physical layer (PHY)
specifications; amendment 4: enhancements for very high throughput for operation in bands below 6 GHz,” IEEE P802.11ac/D3.1, 2012.
[40] R. Liao, B. Bellalta, M. Oliver, and Z. Niu, “MU-MIMO MAC protocols for wireless local area networks: A survey,” IEEE Commun. Surveys
Tuts., vol. 18, no. 1, pp. 162–183, 2014.
[41] M. Abramowitz and I. A. Stegun, Handbook of Mathematical Functions. New York, NY, USA: Dover, 1966.
[42] X. Zhang, D. P. Palomar, and B. Ottersten, “Statistically robust design of linear MIMO transceivers,” IEEE Trans. Sig. Process., vol. 56, no. 8,
pp. 3678–3689, Aug. 2008.
[43] H. Lin and W.-Y. Shin, “Achieving the optimal throughput scaling in multi-cell random access networks,” in Proc. of IEEE Int. Conf. Commun.
(ICC), Paris, France, May 2017.
[44] ——, “Multi-cell aware opportunistic random access,” in Proc. of IEEE Int. Symp. Inf. Theory (ISIT), Aachen, Germany, Jun. 2017.
| 7cs.IT
|
arXiv:1610.02672v1 [math.GR] 9 Oct 2016
Internal and external duality in abstract polytopes
Gabe Cunningham
University of Massachusetts Boston
and
Mark Mixer
Wentworth Institute of Technology
October 11, 2016
Abstract
We define an abstract regular polytope to be internally self-dual if its selfduality can be realized as one of its symmetries. This property has many
interesting implications on the structure of the polytope, which we present
here. Then, we construct many examples of internally self-dual polytopes. In
particular, we show that there are internally self-dual regular polyhedra of
each type {p, p} for p ≥ 3 and that there are both infinitely many internally
self-dual and infinitely many externally self-dual polyhedra of type {p, p} for p
even. We also show that there are internally self-dual polytopes in each rank,
including a new family of polytopes that we construct here.
Key Words: abstract polytope, self-dual polytope, regular polytope
AMS Subject Classification (2010): 20B25, 52B15, 51M20, 52B05
1
Introduction
Whenever a polytope is invariant under a transformation such as duality, we often
describe this as an “external” symmetry of the polytope. In the context of abstract
regular polytopes, self-duality manifests as an automorphism of the symmetry group
of the polytope. For example, the n-simplex is self-dual, and this fact is reflected
by an automorphism of its symmetry group Sn+1 . Since the symmetric group Sn+1
has no nontrivial outer automorphisms (unless n + 1 = 6), we see that in general the
1
self-duality of the simplex must somehow correspond to an ordinary rank-preserving
symmetry. What does this mean geometrically? What other polytopes have this
property that self-duality yields an inner automorphism?
In this paper, we define a regular self-dual abstract polytope to be internally
self-dual if the group automorphism induced by its self-duality is an inner automorphism. Otherwise, a regular self-dual polytope is externally self-dual. Our search
for internally self-dual regular polytopes started with the atlas of regular polytopes
with up to 2000 flags [2]. Using a computer algebra system, we found the following
results on the number of regular polytopes (up to isomorphism and duality).
Rank
3
4
5
Total
3571
2016
258
Self-dual
242
156
15
Internally self-dual
54
2
1
Externally self-dual
188
154
14
Table 1: Data on dualities of small regular polytopes
The internally self-dual polyhedra we found include examples of type {p, p} for
3 ≤ p ≤ 12 and p = 15. In rank 4, the only two examples are the simplex {3, 3, 3}
and a toroid of type {4, 3, 4}. In rank 5, the only example is the simplex {3, 3, 3, 3}.
The data suggest that there are many internally self-dual polyhedra (although
many more externally self-dual). Indeed, we will show that there are internally selfdual polyhedra of type {p, p} for each p ≥ 3. The data in ranks 4 and 5 seem less
promising as far as the existence of internally self-dual polytopes. We will show,
however, that there are several families of internally self-dual polytopes in every
rank. Furthermore, due to the ubiquity of symmetric groups, it seems likely that
in every rank, there are many polytopes that are internally self-dual for the same
reason as the simplex.
In addition to the existence results mentioned above, we explore the consequences
of a self-dual polytope being either externally or internally self-dual.
The paper is organized as follows. In section 2 we briefly outline the necessary
definitions and background on abstract polytopes. Section 3 consists of basic results following from the definition of an internally self-dual polytope. In section 4,
more structural results about internally self-dual polytopes are examined. Section 5
includes existence results regarding both self-dual regular polyhedra and self-dual
polytopes of higher ranks. Finally, in section 6, we highlight a few open questions
and variants of the problems considered here.
2
2
Background
Most of our background, along with many more details can be found in from [9].
An abstract n-polytope P is a ranked partially ordered set of faces with four defining
properties. First, P has a least face F−1 of rank −1, and a greatest face Fn of rank n;
here if an F ∈ P has rank(F ) = i, then F is called an i-face. Faces of rank 0, 1, and
n−1 are called vertices, edges, and facets, respectively. Second, each maximal totally
ordered subset of P contains n + 2 faces. These maximal totally ordered subsets of
P are called flags. Third, if F < G with rank(F ) = i − 1 and rank(G) = i + 1, then
there are exactly two i-faces H with F < H < G. Finally, P is strongly connected,
which is defined as follows. For any two faces F and G of P with F < G, we call
G/F := {H | H ∈ P, F ≤ H ≤ G} a section of P. If P is a partially ordered
set satisfying the first two properties, then P is connected if either n ≤ 1, or n ≥ 2
and for any two faces F and G of P (other than F−1 and Fn ) there is a sequence
of faces F = H0 , H1 , . . . , Hk−1 , Hk = G, not containing F−1 and Fn , such that Hi
and Hi−1 are comparable for i = 1, . . . , k. We say that P is strongly connected if
each section of P (including P itself) is connected. Due to the relationship between
abstract polytopes and incidence geometries, if two faces F and G are comparable
in the partial order, then it is said that F is incident on G.
Two flags of an n-polytope P are said to be adjacent if they differ by exactly one
face. If Φ is a flag of P, the third defining property of an abstract polytope tells us
that for i = 0, 1, . . . , n − 1 there is exactly one flag that differs from Φ in its i-face.
This flag is denoted Φi and is i-adjacent to Φ. We extend this notation recursively
and let Φi1 ···ik denote the flag (Φi1 ···ik−1 )ik . If w is a finite sequence (i1 , . . . , ik ), then
Φw denotes Φi1 ···ik . Note that Φii = Φ for each i, and Φij = Φji if |i − j| > 1. An
n-polytope (n ≥ 2) is called equivelar if for each i = 1, 2, ..., d − 1, there is an integer
pi so that every section G/F defined by an (i − 2)-face F incident on an (i + 1)-face
G is a pi -gon. If P is equivelar we say that it has (Schläfli ) type {p1 , p2 , . . . , pn−1 }.
The automorphism group of an n-polytope P (consisting of the order-preserving
bijections from P to itself) is denoted by Γ(P). For any flag Φ, any finite sequence
w, and any automorphism ϕ, we have Φw ϕ = (Φϕ)w . An n-polytope P is called
regular if its automorphism group Γ(P) has exactly one orbit when acting on its
flags of P, or equivalently, if for some base flag Φ = {F−1 , F0 , F1 , . . . , Fn } and each
i = 0, 1, . . . , n−1 there exists a unique involutory automorphism ρi ∈ Γ(P) such that
Φρi = Φi . For a regular n-polytope P, its group Γ(P) is generated by the involutions
ρ0 , . . . , ρn−1 described above. These generators satisfy the relations
(ρi ρj )pij = (0 ≤ i, j ≤ n − 1),
3
(1)
where pii = 1 for all i, 2 ≤ pji = pij if j = i − 1, and
pij = 2 for |i − j| ≥ 2.
(2)
Any group hρ0 , . . . , ρn−1 i satisfying properties 1 and 2 is called a string group
generated by involutions. Moreover, Γ(P) has the following intersection property:
hρi | i ∈ Ii ∩ hρi | i ∈ Ji = hρi | i ∈ I ∩ Ji for I, J ⊆ {0, . . . , n − 1}.
(3)
Any string group generated by involutions that has this intersection property is called
a string C-group. The group Γ(P) of an abstract regular polytope P is a string Cgroup. Conversely, it is known (see [9, Sec. 2E]) that an abstract regular n-polytope
can be constructed uniquely from any string C-group. This correspondence between
string C-groups and automorphism groups of abstract regular polytopes allows us to
talk simulatneously about the combinatorics of the flags of abstract regular polytopes
as well as the structure of their groups.
Let Γ = hρ0 , . . . , ρn−1 i be a string group generated by involutions acting as a
permutation group on a set {1, . . . , k}. We can define the permutation representation
graph X as the r-edge-labeled multigraph with k vertices, and with a single i-edge
{a, b} whenever aρi = b with a < b. Note that, since each of the generators is an
involution, the edges in our graph are not directed. When Γ is a string C-group, the
multigraph X is called a CPR graph, as defined in [11].
If P and Q are regular polytopes, then we say that P covers Q if there is a welldefined surjective homomorphism from Γ(P) to Γ(Q) that respects the canonical
generators. In other words, if Γ(P) = hρ0 , . . . , ρn−1 i and Γ(Q) = hρ00 , . . . , ρ0n−1 i, then
P covers Q if there is a homomorphism that sends each ρi to ρ0i .
Suppose Γ = hρ0 , . . . , ρn−1 i is a string group generated by involutions, and Λ is a
string C-group such that Γ covers Λ. If the covering is one-to-one on the subgroup
hρ0 , . . . , ρn−2 i, then the quotient criterion says that Γ is itself a string C-group [9,
Thm. 2E17].
Given regular polytopes P and Q, the mix of their automorphism groups, denoted
Γ(P) Γ(Q), is the subgroup of the direct product Γ(P) × Γ(Q) that is generated by
the elements (ρi , ρ0i ). This group is the minimal string group generated by involutions
that covers both Γ(P) and Γ(Q) (where again, we only consider homomorphisms that
respect the generators). Using the procedure in [9, Sec. 2E], we can build a poset
from this, which we call the mix of P and Q, denoted P Q. This definition naturally
extends to any family of polytopes (even an infinite family); see [10, Section 5] for
more details.
The comix of Γ(P) with Γ(Q), denoted Γ(P) Γ(Q), is the largest string group
generated by involutions that is covered by both Γ(P) and Γ(Q) [3]. A presentation
4
for Γ(P) Γ(Q) can be obtained from that of Γ(P) by adding all of the relations of
Γ(Q), rewriting the relations to use the generators of Γ(P) instead.
The dual of a poset P is the poset P ∗ with the same underlying set as P, but with
the partial order reversed. Clearly, the dual of a polytope is itself a polytope. If P ∗ ∼
=
P, then P is said to be self-dual. A duality d : P → P is an order-reversing bijection.
If w = (i1 , . . . , ik ), then we define w∗ to be the sequence (n − i1 − 1, . . . , n − ik − 1).
For any flag Φ, finite sequence w, and duality d, we have Φw d = (Φd)w∗ .
If P is a self-dual regular n-polytope, then there is a group automorphism of Γ(P)
that sends each ρi to ρn−i−1 . If ϕ ∈ Γ(P), then we will denote by ϕ∗ the image of
ϕ under this automorphism. In particular, if ϕ = ρi1 · · · ρik , then we define ϕ∗ to be
ρn−i1 −1 · · · ρn−ik −1 . Thus, if Φ is the base flag of P and Φϕ = Φw , then Φϕ∗ = Φw∗ .
3
3.1
Internal self-duality
Basic notions
As just noted, if P is a self-dual regular n-polytope, then there is a group automorphism of Γ(P) that sends each ρi to ρn−i−1 . When this automorphism is inner, there
is a polytope symmetry α ∈ Γ(P) such that αρi = ρn−i−1 α for 0 ≤ i ≤ n − 1. We
use this idea to motivate the following definitions.
Definition 3.1. Suppose P is a regular self-dual polytope. An automorphism α ∈
Γ(P) is a dualizing automorphism (or simply dualizing) if αρi = ρn−i−1 α for 0 ≤
i ≤ n−1 (equivalently, if αϕ = ϕ∗ α for every ϕ ∈ Γ(P)). A regular self-dual polytope
is internally self-dual if it has a dualizing automorphism. Otherwise, it is externally
self-dual.
Depending on the automorphism group of a polytope, we can sometimes determine internal self-duality without knowing anything deep about the polytope’s
structure.
Proposition 3.2. Suppose P is a regular self-dual polytope, and that Γ(P) is a
symmetric group. Then P is internally self-dual unless Γ(P) ∼
= S6 and P has type
{6, 6} or {4, 4, 4}.
Proof. For k 6= 6, the symmetric group Sk has no nontrivial outer automorphisms,
and so a self-dual polytope with this automorphism group must be internally selfdual. Up to isomorphism, there are 11 regular polytopes with automorphism group
S6 ; eight of them are not self-dual, one of them (the 5-simplex) is internally selfdual, and the remaining two (denoted {6,6}*720a and {4, 4, 4}*720) are externally
self dual [8].
5
It would perhaps be possible to find other abstract results of this type, but they
could never tell the whole story. After all, it is possible for the automorphism that
sends each ρi to ρn−i−1 to be inner, even if Γ(P) has nontrivial outer automorphisms.
So let us shift our focus away from the abstract groups.
What does it mean geometrically to say that P is internally self-dual? Let Φ be
the base flag of P, and let Ψ = Φα for some dualizing automorphism α. Let ϕ ∈ Γ(P)
and suppose that Φϕ = Φw . Then
Ψϕ = Φαϕ
= Φϕ∗ α
∗
= Φw α
= (Φα)w
∗
∗
= Ψw .
In other words, every automorphism acts on Ψ dually to how it acts on Φ.
Definition 3.3. We say that flags Φ and Ψ are dual (to each other) if every auto∗
morphism acts dually on Φ and Ψ. That is, if Φϕ = Φw , then Ψϕ = Ψw .
Note that if P is regular and Ψ is dual to Φ, then in particular Ψρi = Ψn−i−1 .
Proposition 3.4. A regular polytope P is internally self-dual if and only if its base
flag Φ has a dual flag Ψ.
Proof. The discussion preceding Definition 3.3 shows that if P is internally selfdual, then Φ has a dual flag. Conversely, suppose that Φ has a dual flag Ψ. Since
P is regular, there is an automorphism α ∈ Γ(P) such that Ψ = Φα. Then for
0 ≤ i ≤ n − 1,
Φαρi α−1 = Ψρi α−1
= Ψn−i−1 α−1
= (Ψα−1 )n−i−1
= Φn−i−1 .
So αρi α−1 acts on Φ the same way that ρn−i−1 does, and since the automorphism
group acts regularly on the flags, it follows that αρi α−1 = ρn−i−1 for each i. Thus α
is a dualizing automorphism, and P is internally self-dual.
6
Proposition 3.4 provides an intuitive way to determine whether a regular polytope
is internally self-dual: try to find a flag that is dual to the base flag. Let us consider
some simple examples. Let P = {p}, with 3 ≤ p < ∞. Fix a vertex and an edge
in the base flag Φ. In order for Ψ to be dual to Φ, we need for Ψρ0 = Ψ1 , which in
particular means that ρ0 fixes the vertex of Ψ. Now, whenever p is even or infinite,
the reflection ρ0 does not fix any vertices, and so P must be externally self dual (See
Figure 1). When p is odd, then there is a unique vertex v fixed by ρ0 . Furthermore,
there is an edge incident to v that is fixed by ρ1 . The flag Ψ consisting of this vertex
and edge is dual to Φ, and so in this case, P is internally self-dual. Indeed, when
p is odd, then the automorphism (ρ0 ρ1 )(p−1)/2 ρ0 is dualizing. The following result is
then clear.
Proposition 3.5. The p-gon P = {p} is internally self-dual if and only if p is odd.
Figure 1: Adjacent flags and dual flags in a pentagon and a square
Next consider P = {3, 3}, the simplex. Since Γ(P) ∼
= S4 , which has no nontrivial
outer automorphisms, it follows that P is internally self-dual. Nevertheless, let us see
what the dual to the base flag looks like. Consider the labeled simplex in Figure 2,
with vertices {1, 2, 3, 4}, edges {a, b, c, d, e, f }, and facets {L, F, R, D}. Let us pick
the triple Φ = (1, a, L) as the base flag.
To find the dual of the base flag of this simplex consider the action of each of
the distinguished generators of the automorphism group on the vertices, edges, and
facets. We summarize the actions in Table 2.
Since ρ1 and ρ2 both fix the base vertex, we need for their duals, ρ1 and ρ0 , to
fix the vertex of the dual flag. The only possibility is vertex 3. Next, since ρ0 and ρ2
7
Figure 2: A simplex with labeled faces
ρ0
ρ1
ρ2
Vertices
Edges
(1, 2)(3)(4) (a)(b, d)(c, e)(f )
(1)(2, 4)(3) (a, c)(b)(d, f )(e)
(1)(2)(3, 4) (a)(b, c)(d, e)(f )
Facets
(L)(F, R)(D)
(L)(F, D)(R)
(L, D)(F )(R)
Table 2: Symmetries of the labeled simplex
fix the base edge, we need for their duals to fix the edge of the dual flag. The only
edges fixed by both ρ0 and ρ2 are edges a and f , and the only one of those incident
on vertex 3 is f . Finally, since ρ0 and ρ1 fix the base facet, we need for ρ2 and ρ1
to fix the facet of the dual flag, and so the only possibility is R. So the flag that is
dual to (1, a, L) is (3, f, R) (shown in Figure 3).
The process just described is a good illustration of the general process of finding
a dual flag. Let us now describe that process. Suppose P is a regular, self-dual
polytope, with base flag Φ and Γ(P) = hρ0 , . . . , ρn−1 i. A flag Ψ will be dual to Φ if
and only if, for 0 ≤ i ≤ n − 1, the i-face of Ψ is fixed by hρj | j 6= n − i − 1i. So to find
Ψ, we start by looking for a vertex F0 that is fixed by hρ0 , . . . , ρn−2 i. If no such vertex
exists, then Φ does not have a dual flag. Otherwise, once we have found F0 , we now
need an edge F1 that is incident to F0 and fixed by hρ0 , . . . , ρn−3 , ρn−1 i. Since ρn−1
does not fix F0 , the only way that it will fix F1 is if it interchanges the endpoints.
8
Figure 3: A base and dual flag of a simplex
Thus, F1 must be incident on the vertices F0 and F0 ρn−1 . Similar reasoning then
shows that F2 must be incident on all of the edges in F1 hρn−2 , ρn−1 i. Continuing in
this way, we arrive at the following algorithm:
Algorithm 3.6. Suppose P is a regular n-polytope with base flag Φ and Γ(P) =
hρ0 , . . . , ρn−1 i. We pick a flag Ψ = (F0 , F1 , . . . , Fn−1 ) as follows.
1. Set F0 to be a vertex of P that is fixed by hρ0 , . . . , ρn−2 i.
2. Once Fi is chosen, pick Fi+1 to be an (i + 1)-face that is incident to every i-face
in Fi hρn−i−1 , . . . , ρn−1 i.
3. Repeat Step 2 until Fn−1 is chosen.
For simplicity, let us assume for now that P is vertex-describable, meaning that
each face of P is uniquely determined by its vertex-set.
Proposition 3.7. Suppose P is a regular, self-dual, vertex-describable polytope.
Then P is internally self-dual if and only if the procedure described in Algorithm 3.6
finishes successfully.
Proof. If P is internally self-dual, then its base flag Φ has a dual flag Ψ. Let Ψ =
(F0 , . . . , Fn−1 ). Then for each i, the face Fi is fixed by hρj | j 6= n − i − 1i. In
9
particular, since Fi < Fi+1 , it follows that each ϕ ∈ hρn−i−1 , . . . , ρn−1 i, we have
Fi ϕ < Fi+1 ϕ = Fi+1 . Thus, once we have chosen Fi , the face Fi+1 satisfies the
desired properties, and we can continue the process until we have produced Ψ.
Conversely, suppose that the algorithm produced a sequence of faces (F0 , . . . , Fn−1 ).
To show that P is internally self-dual, it suffices to show that ρj fixes every face except for Fn−j−1 . By construction, the vertex-set of each face Fi is F0 hρn−i , . . . , ρn−1 i.
For j < n − i − 1, the automorphism ρj commutes with hρn−i , . . . , ρn−1 i and fixes F0 ,
and so
F0 hρn−i , . . . , ρn−1 iρj = F0 ρj hρn−i , . . . , ρn−1 i = F0 hρn−i , . . . , ρn−1 i.
For j > n − i − 1, we have
F0 hρn−i , . . . , ρn−1 iρj . = F0 hρn−i , . . . , ρn−1 i.
So for j 6= n − i − 1, the automorphism ρj fixes the vertex set of Fi . Since P is
vertex-describable, it follows that ρj fixes Fi itself. Thus ρj fixes every face except
(possibly) for Fn−j−1 . But it is clear that ρj cannot also fix Fn−j−1 , because the only
automorphism that fixes any flag is the identity. This shows that Ψ = (F0 , . . . , Fn−1 )
is dual to Φ.
Corollary 3.8. Let P be a self-dual regular polytope with Γ(P) = hρ0 , . . . , ρn−1 i. If
no vertices are fixed by the subgroup hρ0 , . . . , ρn−2 i, then P is externally self-dual.
3.2
Properties of dualizing automorphisms
Now let us return to the algebraic point of view to determine some properties of
dualizing automorphisms.
Proposition 3.9. If α ∈ Γ(P) is dualizing, then α = α∗ .
Proof. If α is dualizing, then for all ϕ ∈ Γ(P), we have that αϕ = ϕ∗ α. Taking
ϕ = α yields the desired result.
Proposition 3.10. If α ∈ Γ(P) is dualizing, then α is not in hρ0 , . . . , ρn−2 i or in
hρ1 , . . . , ρn−1 i.
Proof. Since α is dualizing, ρn−1 = α−1 ρ0 α. If α ∈ hρ0 , . . . , ρn−2 i, then this gives us
that ρn−1 is in hρ0 , . . . , ρn−2 i, which violates the intersection condition (Equation 3).
Similarly, if α ∈ hρ1 , . . . , ρn−1 i, then the equation ρ0 = α−1 ρn−1 α shows that ρ0 is in
hρ1 , . . . , ρn−1 i, which again violates the intersection condition.
10
The following properties are straightforward to verify.
Proposition 3.11. Let P be a self-dual regular polytope.
(a) If α and β are dualizing automorphisms, then αβ is central, and αβ = βα.
(b) If ϕ is central and α is dualizing, then ϕα is dualizing.
(c) If α is dualizing, then any even power of α is central, and any odd power of α
is dualizing. In particular, α has even order.
(d) If α is dualizing, then α−1 is dualizing.
Proposition 3.12. Let P be an internally self-dual regular polytope such that Γ(P)
has a finite center. Then the number of dualizing automorphisms of Γ(P) is equal to
the order of the center of Γ(P).
Proof. Proposition 3.11 implies that the central and dualizing automorphisms of
Γ(P) together form a group in which the center of Γ(P) has index 2. The result
follows immediately.
3.3
Internal self-duality of nonregular polytopes
Generalizing internal self-duality to nonregular polytopes is not entirely straightforward. When P is a self-dual regular polytope, then Γ(P) always has an automorphism (inner or outer) that reflects this self-duality. This is not the case for general
polytopes.
One promising way to generalize internal self-duality is using the notion of dual
flags. Indeed, Definition 3.3 does not require the polytope to be regular, and makes
sense even for nonregular polytopes. Let us determine some of the simple consequences of this definition.
Proposition 3.13. Suppose that Ψ is dual to Φ. Then Ψn−i−1 is dual to Φi .
Proof. Let ϕ ∈ Γ(P). We need to show that ϕ acts dually on Φi and Ψn−i−1 . Suppose
∗
that Φϕ = Φw , from which it follows that Ψϕ = Ψw . Then:
(Φi )ϕ = (Φϕ)i = Φwi = (Φi )iwi ,
whereas
(Ψn−i−1 ϕ) = (Ψϕ)n−i−1 = Ψw
∗ (n−i−1)
and the claim follows.
11
= (Ψn−i−1 )(n−i−1)w
∗ (n−i−1)
,
Since polytopes are flag-connected, the following is an immediate consequence.
Corollary 3.14. If P has one flag that has a dual, then every flag has a dual.
Thus we see that the existence of dual flags is in fact a global property, not a
local one. Here are some further consequences of the definition of dual flags.
Proposition 3.15. Let P be a polytope, and let Φ and Ψ be flags of P.
(a) If Φ and Ψ are dual, then P is self-dual, and there is a duality d : P → P that
takes Φ to Ψ.
(b) Φ is dual to Ψ if and only if Ψ = Φd for some duality d that commutes with
every automorphism of P.
(c) Φ is dual to Ψ if and only if Ψ = Φd for some duality d that is central in the
extended automorphism group of P.
Proof. (a) We attempt to define the duality d by Φd = Ψ and then extend it by
∗
Φw d = Ψw . To check that this is well-defined, suppose that Φw = Φ; we then
∗
need to show that Ψw = Ψ. Indeed, if Φw = Φ, then taking ϕ to be the
identity we have that Φϕ = Φw , whence Ψ = Ψϕ = Ψw∗ , with the last equality
following since Φ and Ψ are dual.
(b) Suppose Φ and Ψ are dual. Then the first part shows that Ψ = Φd for some
duality d. Now, let ϕ be an automorphism, and suppose Φϕ = Φw . We get:
∗
∗
Φϕd = (Φw )d = (Φd)w = Ψw = Ψϕ = Φdϕ.
Thus the automorphisms d−1 ϕd and ϕ both act the same way on Φ, and so they
are equal. So ϕd = dϕ, and d commutes with every automorphism. Conversely,
if d commutes with every automorphism, then
∗
∗
Ψϕ = Φdϕ = Φϕd = (Φw )d = (Φd)w = Ψw ,
and Ψ is dual to Φ.
(c) In light of the previous part, all that remains to show is that, if a duality
commutes with every automorphism of P, then it also commutes with every
duality. If d2 is a duality, then d2 d−1 is an automorphism, and thus:
dd2 d−1 = d(d2 d−1 ) = (d2 d−1 )d = d2 ,
and so d commutes with d2
12
We see that dual flags have several nice properties, even in the nonregular case.
We could define a polytope to be internally self-dual if every flag has a dual (equivalently, if any single flag has a dual). It is not entirely clear if this is the “right”
definition. In any case, we do not pursue the nonregular case any further here.
4
4.1
Properties of internally self-dual polytopes
Basic structural results
Internally self-dual polytopes have a number of structural restrictions. Many of them
are consequences of the following simple property.
Proposition 4.1. If P is an internally self-dual regular polytope, then any regular
polytope covered by P is also internally self-dual.
Proof. Since P is internally self-dual, there is an automorphism α ∈ Γ(P) such that
αρi = ρn−i−1 α for 0 ≤ i ≤ n − 1. If P covers Q, then the image of α in Γ(Q) has the
same property, and so Q is internally self-dual.
Proposition 4.1 makes it difficult to construct internally self-dual polytopes with
mixing, since the mix of any two polytopes must cover them both. The next proposition characterizes which pairs of polytopes can be mixed to create an internally
self-dual polytope.
Proposition 4.2. Let P and Q be regular n-polytopes such that P Q is polytopal.
Then P Q is internally self-dual if and only if
(a) P and Q are internally self-dual, and
(b) There are dualizing automorphisms α ∈ Γ(P) and β ∈ Γ(Q) such that the
images of α and β in Γ(P) Γ(Q) coincide.
Proof. Let Γ(P) = hρ0 , . . . , ρn−1 i and Γ(Q) = hρ00 , . . . , ρ0n−1 i, and let Γ(P) Γ(Q) =
hσ0 , . . . , σn−1 i, where σi = (ρi , ρ0i ). Let ϕ = (α, β) ∈ Γ(P) Γ(Q). Then ϕ is
dualizing if and only if α and β are both dualizing, since ϕσi = σn−i−1 ϕ if and
only if (αρi , βρ0i ) = (ρn−i−1 α, ρ0n−i−1 β). Therefore, P Q is internally self-dual if
and only if there are dualizing automorphisms α ∈ Γ(P) and β ∈ Γ(Q) such that
(α, β) ∈ Γ(P) Γ(Q). By [3, Prop. 3.7], this occurs if and only if the images of α
and β in Γ(P) Γ(Q) coincide.
13
Corollary 4.3. If P and Q are internally self-dual regular n-polytopes with “the
same” dualizing element (i.e., the image of the same word in the free group on
ρ0 , . . . , ρn−1 ), then P Q is internally self-dual.
If we have a presentation for Γ(P), it is often simple to show that P is not
internally self-dual. Indeed, because of Proposition 4.1, all we need to do is find
some non-self-dual quotient of Γ(P). For example, if P is internally self-dual, then
adding a relation that forces ρ0 = must also force ρn−1 = . Here are some similar
results that are easily applied.
Proposition 4.4. If P is an internally self-dual regular n-polytope, then in the
abelianization of Γ(P), the image of each ρi is equal to the image of ρn−i−1 .
Proof. By Proposition 4.1, the abelianization of Γ(P) must have a dualizing automorphism. Since such an automorphism also must commute with the image of every
ρi , it follows that the images of ρi and ρn−i−1 must coincide.
Suppose P is a regular n-polytope, with m-faces isomorphic to K. We say that
a regular n-polytope has the Flat Amalgamation Property (FAP) with respect to its
m-faces if adding the relations ρi = to Γ(P) for i ≥ m yields a presentation for
Γ(K) (rather than a proper quotient).
Proposition 4.5. If P has the FAP with respect to its m-faces for some m with
1 ≤ m ≤ n − 1, then P is not internally self-dual.
Proof. If P is internally self-dual, then adding the relation ρn−1 = to Γ(P) must
force ρ0 = , and this precludes P from having the FAP with respect to its m-faces
for any m ≥ 1.
Corollary 4.6. If P is internally self-dual and e is the unique 1-polytope, then
Γ(P e) ∼
= Γ(P) × C2 .
Proof. The group Γ(P e) is either isomorphic to Γ(P) or Γ(P) × C2 . From Proposition 4.5 and [9, Thm. 7A11] we know that Γ(P e) ∼
6= Γ(P).
Corollary 4.7. If P is internally self-dual and e is the unique 1-polytope, then
(P e)∗ e is externally self-dual, with automorphism group Γ(P) × C22 .
Proof. In general, (P Q)∗ ∼
= P ∗ Q∗ , and so (P e)∗ e is self-dual. Since P is
self-dual and Γ(P e) ∼
= Γ(P) × C2 , the group Γ(P e∗ ) is isomorphic to Γ(P) × C2 as
well. Thus Γ((P e)∗ e) ∼
= Γ(P) × C22 . Furthermore, if you mix (P e)∗ e with an
edge again, then the automorphism group does not change. Thus, by Corollary 4.6,
(P e)∗ e must be externally self-dual.
14
Corollary 4.8. If P is an internally self-dual polyhedron of type {p, p}, then (P
e)∗ e is an externally self-dual polyhedron of type {q, q}, where q = lcm(p, 2).
Proof. Let Γ(P) = hρ0 , ρ1 , ρ2 i and let Γ(e) = hλ0 i. For 0 ≤ i ≤ 2, let σi = (ρi , λi ) ∈
Γ(P) × Γ(e), where we take λ1 = λ2 = . Then (σ0 σ1 )p = ((ρ0 ρ1 )p , λp0 ) = (, λp0 ). If p
is even, then this gives us (, ), and so σ0 σ1 has order p. Otherwise σ0 σ1 has order
2p. So P e is of type {q, p}, and by taking the dual and mixing with e again, we
get a polyhedron of type {q, q}.
In some sense, Corollary 4.7 says that externally self-dual polytopes are at least
as common as internally self-dual polytopes.
4.2
Internal self-duality of universal polytopes
A natural place to start looking for internally self-dual polytopes is the universal
polytopes {p1 , . . . , pn−1 } whose automorphism groups are string Coxeter groups. Let
us start with those polytopes with a 2 in their Schläfli symbol. Recall that a polytope
is flat if every vertex is incident on every facet.
Proposition 4.9. There are no flat, regular, internally self-dual polytopes.
Proof. Suppose P is a flat regular polytope, with Γ(P) = hρ0 , . . . , ρn−1 i. The stabilizer of the base facet is Γn−1 = hρ0 , . . . , ρn−2 i, which acts transitively on the vertices
of the base facet. Since P is flat, it follows that Γn−1 acts transitively on all the
vertices of P. In particular, Γn−1 does not fix any vertices, and thus Corollary 3.8
implies that P is not internally self-dual.
Corollary 4.10. If P is a regular internally self-dual polytope of type {p1 , . . . , pn−1 },
then each pi is at least 3.
Proof. Proposition 2B16 in [9] proves that if some pi is 2, then P is flat.
Next, we can rule out infinite polytopes.
Proposition 4.11. If P is a self-dual regular polytope such that Γ(P) is an infinite
string Coxeter group, then P is externally self-dual.
Proof. Lemma 2.14 in [6] proves that an infinite Coxeter group with no finite irreducible components has no nontrivial inner automorphisms that realize any graph
automorphisms of the Coxeter diagram. Since self-duality is a graph automorphism,
it follows that P cannot be internally self-dual.
15
We can now easily cover the remaining self dual string Coxeter groups.
Theorem 4.12. The only internally self-dual regular polytopes such that Γ(P) is a
string Coxeter group are simplices and p-gons with p odd.
Proof. In light of Corollary 4.10 and Proposition 4.11, the only possibilities left to
consider are simplices, polygons, and the 24-cell {3, 4, 3}. Propositions 3.2 and 3.5
establish the claim for simplices and p-gons. Using a computer algebra system, we
can verify that {3, 4, 3} is not internally self-dual.
4.3
Restrictions on the automorphism group
Based on the data in Table 1 and the restrictions from the previous section, it
seems that regular internally self-dual polytopes could be relatively rare, especially
in ranks other than three. Before exploring some existence results about internally
self-dual polytopes, let us discuss a few natural questions that arise while looking for
examples. First, one might want to know if the existence of internally or externally
self-dual polytopes with a certain group Γ as an automorphism group might depend
on the abstract class of Γ. In a simple way, the answer to this question is yes,
in that if the automorphism group of Γ is isomorphic to Γ itself, then every group
automorphism is inner, and there can be no external dualities of the polytope (as seen
in Proposition 3.2). Otherwise, it seems that the abstract structure of the group does
not provide insight into whether self-dual regular polytopes will be either externally
or internally self-dual. In particular, there are both internally and externally selfdual regular polytopes with simple groups as their automorphism groups; examples
of which can be easily found by considering alternating groups, for example.
Second, while looking for internally self-dual polytopes of higher rank, a natural
question is whether they are built from only internally self-dual polytopes of lower
rank. For example, must the medial sections of an internally self-dual polytope
be internally self-dual themselves? (The medial section of a polytope is a section
F/v where F is a facet and v is a vertex.) Consider the unique self-dual regular
polytope of type of rank four with an alternating group acting on nine points as its
automorphism group (see Figure 4 of [4]). This is easily seen to be internally selfdual, as the duality is realized as an even permutation on the nine points. However,
this polytope is of type {5, 6, 5}, and so its medial sections, being hexagons, are not
internally self-dual; see Proposition 3.5.
Finally, Proposition 4.1 says that if a regular polytope is internally self-dual,
then the only regular polytopes that it covers are also internally self-dual. This is a
stringent requirement, so one might hope that the converse would be true. However,
16
there are externally self-dual polytopes that only cover internally-self dual polytopes.
For example, the unique polyhedron P of type {5, 5} with 320 flags is externally selfdual, and it double-covers the unique polyhedron Q of type {5, 5} with 160 flags,
which is internally self-dual. Furthermore, every quotient of the former (polyhedral
or not) filters through the latter, since the kernel of the covering from P to Q was
the unique minimal normal subgroup of Γ(P). So P only covers internally self-dual
polyhedra, despite being externally self-dual itself.
5
Examples of internally self-dual polytopes
In this section we will prove the existence of internally and externally self-dual polytopes with various characteristics. We mainly focus on rank 3, but higher rank
polytopes are also constructed.
5.1
Rank three
First we construct a few families of internally self-dual regular polyhedra. Our main
result is the following.
Theorem 5.1. For each p ≥ 3 (including p = ∞), there is an internally self-dual
regular polyhedron of type {p, p}, and for each p ≥ 4 (including p = ∞), there is an
externally self-dual regular polyhedron of type {p, p}. Furthermore, if p is even, then
there are infinitely many internally and externally self-dual regular polyhedra of type
{p, p}.
We will focus on constructing internally self-dual regular polyhedra; then Theorem 4.12 and Corollary 4.8 will take care of the rest. First we will show that there
is an internally self-dual polyhedron of type {p, p} for each p ≥ 3. The data from [2]
provides examples for 3 ≤ p ≤ 12. We will construct a family that covers p ≥ 7. We
start with a simple lemma.
Lemma 5.2. Suppose π1 and π2 are distinct permutations that act cyclically on n
points and that π1d1 = π2d2 for some d1 and d2 . Suppose that for some positive integer
k, there is a unique point i such that iπ1j = iπ2j for all j with 1 ≤ j ≤ k. Then
d1 = d2 = 0 (mod n).
17
Proof. Since π1d1 = π2d2 , it follows that π2 and π1d1 commute. Then, for 1 ≤ j ≤ k,
(iπ1d1 )π2j = (iπ2j )π1d1
= (iπ1j )π1d1
= (iπ1d1 )π1j .
That is, π1j and π2j act the same way on (iπ1d1 ) for 1 ≤ j ≤ k. By assumption, i was
the only point such that π1j and π2j act the same way on that point for 1 ≤ j ≤ k.
It follows that π1d1 (which is equal to π2d2 ) fixes i, which implies that d1 = d2 =
0 (mod n).
Theorem 5.3. For each p ≥ 7, there is an internally self-dual polyhedron of type
{p, p} such that (ρ0 ρ2 ρ1 )6 is dualizing.
Proof. We will construct a permutation representation graph, and then show that it
is a CPR graph. If p is odd, then consider the following permutation representation
graph:
4
1
5
2
0 2
0
1
6
0
3
1
7
0
2
8
1
9
p−2
0
2
p−1
1
p
0
2
p
2
2
1
If p is even, then then instead consider the following.
4
1
5
2
0 2
0
1
6
0
3
1
1
7
0
2
8
1
9
p−2
1
p−1
2
2
In both cases, it is easy to see that the group is a string group generated by involutions. To verify this, we only need to notice that the subgraph induced by edges
of labels 0 and 2, consists of connected components that are either isolated vertices,
double edges, or squares with alternating labels.
18
If p is odd, then
ρ0 ρ2 ρ1 = (1, 7, 6)(2, 4, 5, 3)(8, 9)(10, 11) · · · (p − 1, p),
and if p is even, then
ρ0 ρ2 ρ1 = (1, 7, 6)(2, 4, 5, 3)(8, 9)(10, 11) · · · (p − 2, p − 1).
In both cases, it follows that (ρ0 ρ2 ρ1 )6 = (2, 5)(3, 4). It is simple to show that this
is in fact a dualizing automorphism. In other words, (ρ0 ρ2 ρ1 )6 ρi acts the same on
every vertex as ρ2−i (ρ0 ρ2 ρ1 )6 .
It remains to show that hρ0 , ρ1 , ρ2 i is a string C-group. Following [9, Proposition
2E16], it suffices to show that hρ0 , ρ1 i ∩ hρ1 , ρ2 i = hρ1 i.
Let ϕ be in the intersection, ϕ ∈
/ hρ1 i. Without loss of generality, we may assume
that ϕ = (ρ0 ρ1 )d1 for some d1 (since if ϕ is odd, then ϕρ1 is also in the intersection
and can be written like that). Now, ϕ is also in hρ1 , ρ2 i. If ϕ = ρ1 (ρ2 ρ1 )d2 for some
d2 , then ϕ2 = , and thus (ρ0 ρ1 )2d1 = . If p is odd, then this only happens if ϕ
is itself the identity, contrary to our assumption. So in that case we may assume
that ϕ = (ρ2 ρ1 )d2 for some d2 . If p is even, then in principle, it is possible that
ϕ = (ρ0 ρ1 )p/2 , and so it could happen that ϕ = ρ1 (ρ2 ρ1 )d2 .
Suppose p is odd, p ≥ 9. Then
ρ0 ρ1 = (1, 3, 5, 7, 9, . . . , p, p − 1, p − 3, . . . , 6, 4, 2)
and
ρ2 ρ1 = (1, 4, 2, 7, 9, . . . , p, p − 1, p − 3, . . . , 6, 3, 5).
We note that these cycles act the same way on 3, on 4, and on 7 through p. Indeed,
7 is the start of a unique longest sequence of points on which ρ0 ρ1 and ρ2 ρ1 act, and
so we can apply Lemma 5.2 with i = 7 and k = p − 6. It follows that there is no
nontrivial equation of the form (ρ0 ρ1 )d1 = (ρ2 ρ1 )d2 .
Now, suppose p is even. Then
ρ0 ρ1 = (1, 3, 5, 7, 9, . . . , p − 1, p, p − 2, p − 4, . . . , 6, 4, 2)
and
ρ2 ρ1 = (1, 4, 2, 7, 9, . . . , p − 1, p, p − 2, p − 4, . . . , 6, 3, 5).
As in the odd case, (ρ0 ρ1 )d1 cannot equal (ρ2 ρ1 )d2 , by Lemma 5.2 with i = 7 and
k = p − 6. We still need to rule out the case (ρ0 ρ1 )p/2 = ρ1 (ρ2 ρ1 )d2 . Note that
(ρ0 ρ1 )p/2 always sends 1 to p. In order for ρ1 (ρ2 ρ1 )d2 to do the same thing, we would
19
need d2 = p/2 as well. But then (ρ0 ρ1 )p/2 sends 3 to p − 2, whereas ρ1 (ρ2 ρ1 )p/2 sends
3 to p − 3. So that rules out this case, proving that the intersection condition holds.
The remaining cases where p = 7, 8 can be verified using a computer algebra
system.
Using the polyhedra built in Theorem 5.3 as a base, we can construct an infinite
polyhedron that is internally self-dual.
Theorem 5.4. Consider a family of internally self-dual regular polyhedra {Pi }∞
i=1 .
(i) (i) (i)
Let Γ(Pi ) = hρ0 , ρ1 , ρ2 i. Suppose that there is a finite sequence j1 , . . . , jm such
(i)
(i)
that ρj1 · · · ρjm is dualizing in each Γ(Pi ). Then P = P1 P2 · · · is an infinite
internally self-dual regular polyhedron.
(1)
(2)
(1)
(2)
Proof. It is clear that P is infinite. Let ρ0 = (ρ0 , ρ0 , . . .), ρ1 = (ρ1 , ρ1 , . . .),
(1) (2)
ρ2 = (ρ2 , ρ2 , . . .). To show that P is a polyhedron, we need to show that hρ0 , ρ1 , ρ2 i
satisfies the intersection condition. In particular, by Proposition 2E16 of [9], the only
intersection that is nontrivial to check is hρ0 , ρ1 i ∩ hρ1 , ρ2 i. Suppose ϕ ∈ hρ0 , ρ1 i ∩
hρ1 , ρ2 i. For each i, let ϕi be the natural projection of ϕ in Γ(Pi ), where we send each
(i)
(i)
ρj to ρj . Since Pi is a polyhedron, each ϕi is either or ρ1 . Now, since hρ0 , ρ1 i is
dihedral, the automorphism ϕ is either even or odd. Furthermore, its projection ϕi
(i) (i)
into the dihedral group hρ0 , ρ1 i must have the same parity as ϕ itself. Therefore,
(i)
every ϕi must have the same parity, and so either ϕi = for every i, or ϕi = ρ1
for every i. In the first case, ϕ = , and in the second, ϕ = ρ1 . This proves the
intersection condition.
Finally, it is clear that ρj1 · · · ρjm is dualizing in Γ(P), and so P is internally
self-dual.
Corollary 5.5. There is an infinite internally self-dual regular polyhedron of type
{∞, ∞}, with dualizing automorphism (ρ0 ρ2 ρ1 )6 .
Proof. Apply the construction of Theorem 5.4 to the polyhedra in Theorem 5.3.
The infinite polyhedron of Corollary 5.5 is a little difficult to work with; we have
neither a permutation representation nor a presentation for the automorphism group.
With this example, however, we can now build a simpler example.
Corollary 5.6. Let Γ be the quotient of [∞, ∞] by the three relations
(ρ0 ρ2 ρ1 )6 ρi = ρ2−i (ρ0 ρ2 ρ1 )6 , where 0 ≤ i ≤ 2.
Then Γ is the automorphism group of an infinite internally self-dual regular polyhedron of type {∞, ∞}.
20
Proof. First, note that Γ covers the automorphism group of any polyhedron with
dualizing automorphism (ρ0 ρ2 ρ1 )6 . Therefore, Γ covers the automorphism group of
the polyhedron in Corollary 5.5. Then the quotient criterion (see [9, Thm. 2E17])
shows that this is a polyhedron, since it covers the polyhedron in Corollary 5.5
without any collapse of the facet subgroup hρ0 , ρ1 i.
Now let us prove that there are infinitely many internally self-dual polyhedra of
type {p, p} when p ≥ 4 and p is even. We first cover the case p = 4.
Proposition 5.7. The polyhedron {4, 4}(s,0) is internally self-dual if and only if s is
odd.
Proof. Let P = {4, 4}(s,0) , and let us identify the vertices of P with (Z/sZ)2 . Let us
choose the base flag to consist of the origin, the edge from the origin to (1, 0), and
the square [0, 1] × [0, 1]. Then ρ0 sends each vertex (x, y) to (1 − x, y), ρ1 sends (x, y)
to (y, x), and ρ2 sends (x, y) to (x, −y). If s is even, then ρ0 does not fix any vertex,
and so by Corollary 3.8, P is externally self-dual. If s is odd, say s = 2k − 1, then
the unique vertex fixed by hρ0 , ρ1 i is (k, k). Continuing with Algorithm 3.6, we want
an edge that contains (k, k) and (k, k)ρ2 = (k, −k) = (k, k + 1). Finally, we want a
square that contains that edge and its images under hρ1 , ρ2 i, which consists of the 4
edges bounding the square with corners (k, k) and (k + 1, k + 1). Thus, the flag that
is dual to the base flag consists of the vertex (k, k), the edge to (k, k + 1), and the
square that also includes (k + 1, k + 1). See Figure 4.
We now construct a family of examples to cover the remaining cases.
Theorem 5.8. For each even p ≥ 6, there are infinitely many internally self-dual
polyhedra of type {p, p}.
Proof. For each even p ≥ 6, we will construct a family of permutation representation
graphs, and show that each one is the CPR graph of a distinct internally self-dual
polyhedron of type {p, p}. First, consider the following permutation representation
graph G:
.
p−3
2
1
1
2
0
3
1
4
0
p−4
p−1
0
2
p−2
21
1
p
2p−8
2
2p−7
1
2p−6
Figure 4: Base and dual flags in the regular polyhedron {4, 4}(5,0)
Let ρ0 , ρ1 , ρ2 be the permutations induced by edges of the appropriate label. We
have
ρ0 = (2, 3)(4, 5) · · · (p − 6, p − 5)(p − 4, p − 2)(p − 3, p − 1),
ρ1 = (1, 2)(3, 4) · · · (p − 5, p − 4)(p − 1, p)(p + 1, p + 2) · · · (2p − 7, 2p − 6),
ρ2 = (p − 4, p − 3)(p − 2, p − 1)(p, p + 1)(p + 2, p + 3) · · · (2p − 8, 2p − 7).
It is again clear that hρ0 , ρ1 , ρ2 i is a string group generated by involutions. Next, note
that ρ0 ρ2 ρ1 interchanges p − 2 with p − 3, while cyclically permuting the remaining
2p−8 points. Now consider σ = (ρ0 ρ2 ρ1 )p−4 . This interchanges 1 with 2p−6 and thus
1(ρ0 ρ2 ρ1 )j with (2p − 6)(ρ0 ρ2 ρ1 )j , for each j. Since the action of (ρ0 ρ2 ρ1 ) on vertices
on the left is the mirror of the action on the right, it follows that σ interchanges i
with 2p − 5 − i for 1 ≤ i ≤ p − 4, while fixing p − 2 and p − 3.
We will now build a larger graph X using the one above as a building block. Let
k be an odd positive integer, and take N := k(p − 4) copies of the above graph,
labeled G1 , G2 , . . . , GN , and arrange them cyclically. Let us use (i, j) to mean vertex
i in Gj (and where j is considered modulo N if necessary). We connect the graphs
Gj by adding edges labeled 1 from (p − 2, j) to (p − 3, j + 1) By Theorem 4.5 in [11],
this is the CPR graph of a polyhedron,. Furthermore, if we erase the edges labeled
2, then the connected components either have 2 vertices or p vertices. The latter
consists of the first p − 4 vertices, then the bottom of a diamond, then the top of the
next diamond, the right of that diamond, and one more vertex. The same happens
if we erase the edges labeled 0, and so we get a polyhedron of type {p, p}.
22
Let us now redefine ρ0 , ρ1 , and ρ2 as the permutations induced by edges of X ,
and let σ = (ρ0 ρ2 ρ1 )p−4 as before. The new σ acts in exactly the same way on every
vertex in every copy of the original CPR graph except for the top and bottom of the
diamonds. Indeed, σ takes (p − 3, j) to (p − 3, j − p + 4) and it takes (p − 2, j) to
(p − 2, j + p − 4). Then the order of σ is 2k, since σ to any odd power interchanges
every (1, j) with (2p−6, j), and σ k is the smallest power of σ that fixes every (p−3, j)
and (p − 2, j).
We claim that σ k is dualizing. To prove that, we need to show that ρi σ k = σ k ρ2−i
for i = 0, 1, 2. That is clearly true for all vertices other than the tops and bottoms
of diamonds, because σ k acts as a reflection through the middle of the diagram, and
this reflection also dualizes every label. Checking that ρi σk and σ k ρ2−i act the same
on the top and bottom of every diamond is then easy.
Finally, we claim that the constructed polyhedra are distinct for each k. For
this, it suffices to show that in each polyhedron, k is the smallest positive integer
such that σ k is dualizing. (In principle, if σ k is dualizing, it might also be true that
σ m is dualizing for some m dividing k.) In order for a power of σ to dualize most
vertices, it must be odd, since σ 2 fixes every vertex other than the tops and bottoms
of diamonds. So consider σ m for some odd m < k. The permutation σ m ρ1 sends
(p − 3, 1) to (p − 2, m(p − 4)), whereas ρ1 σ m sends (p − 3, 1) to (p − 2, N − m(p − 4)) =
(p − 2, (k − m)(p − 4)). In order for these two points to be the same, we need
k − m = m, so that k = 2m. But k is odd, so this is impossible. So we get infinitely
many internally self-dual polyhedra of type {p, p}.
We can now prove our main result.
Proof of Theorem 5.1. The data from [2], combined with Theorem 5.3 and Corollary 5.5, show that there are internally self-dual polyhedra of type {p, p} for all
3 ≤ p ≤ ∞. Theorem 4.12 then shows that there are externally self-dual polyhedra
of type {p, p} for 4 ≤ p ≤ ∞. Proposition 5.7 and Theorem 5.8 show that there are
infinitely many internally self-dual polyhedra of type {p, p} for p even, and combining with Corollary 4.8, we get infinitely many externally self-dual polyhedra of type
{p, p} for p even as well.
5.2
Higher ranks
Now that the rank three case is well established, let us consider internally self-dual
regular polytopes of higher ranks. We will start by showing that there are infinitely
many internally self-dual polytopes in every rank. By Theorem 4.12, we already
23
know that the n-simplex is internally self-dual. It is instructive to actually show this
constructively, using Algorithm 3.6.
Consider the representation of the regular n-simplex P as the convex hull of
the points e1 , e2 , . . . , en+1 in Rn+1 . Each i-face of P is the convex hull of i + 1 of
the vertices, and each flag of P can be associated to an ordering of the vertices
(ei1 , ei2 , . . . , ein+1 ), where the i-face of the flag is the convex hull of the first i + 1
vertices.
Let us set the base flag Φ to be the flag corresponding to the ordering (e1 , e2 , . . . , en+1 ).
Then each automorphism ρi acts by switching coordinates i+1 and i+2, corresponding to a reflection in the plane xi+1 = xi+2 . In order to find a flag dual to Φ, we need
to first find a vertex that is fixed by hρ0 , . . . , ρn−2 i. The only such vertex is en+1 .
Next, we need an edge that is incident to en+1 and fixed by hρ0 , . . . , ρn−3 i × hρn−1 i.
Since ρn−1 interchanges en and en+1 and it must fix the edge, it follows that the edge
must be incident on en . Continuing in this manner, it is easy to see that the dual
flag must be (en+1 , en , . . . , e1 ).
Now let us find the dualizing automorphism of Γ(P). We can identify Γ(P) with
the symmetric group on n + 1 points, where ρi is the transposition (i + 1, i + 2). We
noted above that the dual of the base flag simply reversed the order of the vertices.
So the dualizing automorphism of Γ(P) can be written as
(ρ0 ρ1 · · · ρn−1 )(ρ0 ρ1 · · · ρn−2 ) · · · (ρ0 ρ1 )(ρ0 ),
as this “bubble-sorts” the list (1, 2, . . . , n + 1) into its reverse.
Here is another example of high-rank internally self-dual polytopes. That they
are internally self-dual follows from Proposition 3.2.
Proposition 5.9. For each n ≥ 5 the regular n-polytope with group Sn+3 described
in [5, Proposition 4.10], is internally self dual.
1
0
1
2
3
n−4 n−3 n−2 n−1 n−2
4
The CPR graph of these polytopes is shown above. Each of this polytopes is obtained from a simplex, by first taking the Petrie contraction and then dualizing and
taking another Petrie contraction. (The Petrie contraction of a string group generated by involutions hρ0 , ρ1 , ρ2 , . . . , ρn−1 i is the group generated by hρ1 , ρ0 ρ2 , ρ3 , . . . , ρn−1 i.)
The cubic toroids (described in [9, Section 6D]) provide an infinite family of
internally self-dual polytopes with automorphism groups other than the symmetric
group.
24
Theorem 5.10. The regular n + 1-polytope {4, 3n−2 , 4}(s,0n−1 ) is internally self-dual
if and only if s is odd.
Proof. Let us take the vertex set of P to be (Z/sZ)n . For 0 ≤ i ≤ n, let Gi be the
i-face of P containing vertices 0, e1 , e1 + e2 , . . . , e1 + e2 + · · · + ei , where {ei }ni=1 is
the standard basis. Let Φ = (G0 , . . . , Gn ) be our base flag. Then the generators of
Γ(P) can be described geometrically as follows (taken from [9, Eq. 6D2]). ρ0 sends
(x1 , . . . , xn ) to (1 − x1 , x2 , . . . , xn ), ρn sends (x1 , . . . , xn ) to (x1 , . . . , xn−1 , −xn ), and
for 1 ≤ i ≤ n − 1, ρi interchanges xi and xi+1 .
We now try to build a flag that is dual to Φ, using Algorithm 3.6. First, we need
to find a vertex that is fixed by hρ0 , . . . , ρn−1 i. In order for (x1 , . . . , xn ) to be fixed
by ρ0 , we need for x1 ≡ 1 − x1 (mod s); in other words, 2x1 ≡ 1 (mod s). That has
a solution if and only if s is odd, so that already establishes that when s is even, the
polytope P is not internally self-dual. On the other hand, if s is odd, then we can take
x1 = (s + 1)/2 as a solution. Now, in order for ρ1 , . . . , ρn−1 to also fix this vertex, we
need all of the coordinates to be the same. So we pick F0 = ((s + 1)/2, . . . , (s + 1)/2).
Next we need an edge incident on F0 and F0 ρn . The latter is simply ((s +
1)/2, . . . , (s + 1)/2, (s − 1)/2), which is indeed adjacent to F0 , and that gives us our
edge F1 . To pick F2 , we look at the orbit G1 hρn−1 , ρn i; this gives us the square
whose 4 vertices are ((s + 1)/2, . . . , (s + 1)/2, ±(s + 1)/2, ±(s + 1)/2). In general,
we take Fi to be the i-face such that its vertices are obtained from F0 by any combination of sign changes in the last i coordinates. Then it is clear that Fi is fixed by
hρ0 , . . . , ρn−i−1 , ρn−i+1 , . . . , ρn i, and thus we have a dual flag to Φ.
To show that there are examples of internally self-dual polytopes other than the
toroids and simplices (and self-dual petrie contracted relatives) in each rank r ≥ 5,
we give the following family of examples and prove that they yield string C-groups
which are internally self-dual. (There are other examples in rank 4 as well, such as
the polytope of type {5, 6, 5} that was mentioned at the end of Section 4.3.)
Lemma 5.11. For each n ≥ 4, the following permutation representation graph is
the CPR graph of a regular n-polytope with automorphism group Γ isomorphic to the
symmetric group on n + 3 points.
2
0
1
2
3
4
n−1
0
2
Proof. For n ≥ 6 this is shown in [5, Proposition 4.8]. The remaining cases can either
be checked by hand or using a computer algebra system.
25
Lemma 5.12. For each n ≥ 6, the following permutation representation graph is
the CPR graph of a regular n-polytope with automorphism group Γ isomorphic to the
symmetric group on n + 4 points. Furthermore, the element (ρn−2 ρn−1 ρn−2 ρn−3 ρn−2 )
is a five cycle with the last five points in its support.
2
1
0
2
3
4
n−3 n−2 n−1 n−2
0
2
Proof. For n ≥ 7 the fact that this is a CPR graph is shown in [5, Proposition
4.14]. The remaining case can either be checked by hand or using a computer algebra
system. The structure of (ρn−2 ρn−1 ρn−2 ρn−3 ρn−2 ) can easily be checked by hand.
Theorem 5.13. For each n ≥ 5, the following permutation representation graph is
the CPR graph of an internally self-dual regular n-polytope with automorphism group
Γ acting on n + 5 points.
3
2
1
4
0
5
2
6
3
7
n−4
n−1
n
n−2
n−3
n−3
n+1
n+2
n+3
n−1
0
2
2
1
n−1
n+5
n−3
n+4
Proof. First we notice that these graphs have slightly different structure depending
on whether n is odd or even. The two smallest cases seen below, of ranks 5 and 6,
are verified to yield string C-groups using a computer algebra system. In the rest of
the proof we will assume that n ≥ 7.
3
2
1
4
0
0
2
2
1
5
2
6
3
7
2
8
4
4
10
3
2
2
0
9
1
4
0
2
2
1
5
2
6
3
7
4
8
3
11
9
5
5
3
10
By design, Γ is self-dual, since interchanging every label i with n − 1 − i is
a symmetry of the graph, corresponding to a reflection through a line that goes
through the middle edge (if n is odd) or through the middle node (if n is even). Let
us show that Γ is internally self-dual.
First, when n is odd and n = 2k + 1, the permutation π1 = ρk interchanges
the two nodes that are incident on the middle edge, while fixing everything else.
Then, setting π2 = (ρk−1 ρk+1 )π1 (ρk+1 ρk−1 ), we get that π2 interchanges the next two
nodes from the center, while fixing everything else. Continuing this way, we can find
permutations π1 , . . . , πk+3 that each interchange a node with its dual while fixing
everything else. The product of all of these will be a dualizing automorphism in Γ.
26
When n is even and n = 2k, the middle node is incident to edges labeled k − 1
and k. The permutation π1 = ρk−1 ρk ρk−1 is easily seen to interchange the two
nodes that are incident on the middle node, while fixing everything else. Then,
setting π2 = (ρk−2 ρk+1 )π1 (ρk+1 ρk−2 ), we get that π2 interchanges the two nodes at a
distance of 2 from the middle node, while fixing everything else. As in the odd case,
we can continue to define permutations π1 , . . . , πk+2 that each interchange a node
with its dual, and the product of all of these will be a dualizing automorphism in Γ.
Now we need to show that each Γ is a string C-group. It is clear that it is a
string group generated by involutions, so we only need to show that it satisfies the
intersection condition. We will do this utilizing [9, Proposition 2E16] by showing
that its facet group and vertex figure group are string C-groups, and that their
intersection is what is needed.
We will take advantage of the fact that Γ is self-dual by design, so the structure
of each of its parabolic subgroups is the same as the “dual” subgroup. For any
subset S of {0, 1, . . . , n − 1}, we define ΓS = hρi | i 6∈ Si. The structure of ΓS
is determined by the structure of the given permutation representation graph after
we delete all edges with labels in S. For example, the group Γ0,1,n−2,n−1 yields a
permutation representation graph with a chain with labels 2 through n − 3, along
with four isolated components; two labeled 2 and two labeled n − 3. It follows that
this is the group obtained from a simplex by mixing with an edge, dualizing, and
mixing with another edge. It is a string C-group which is isomorphic to 2 × 2 × Sn−3 .
The groups Γ0,1,n−1 and Γ0,n−2,n−1 are maximal parabolic subgroups of the group
obtained by mixing a string C-group from Lemma 5.11 with an edge. Both of these
groups are thus string C-groups, and can be shown to be isomorphic to 2 × 2 × Sn−1 .
Similarly, the groups Γ0,1,2 and Γn−3,n−2,n−1 are both described by Lemma 5.11. They
each are string C-groups isomorphic to Sn . Each group Γ0,1 and Γn−2,n−1 is the mix
of a group from Lemma 5.11 with an edge, and can be shown to be isomorphic to
2 × Sn+1 .
Now, to show that Γ0,n−1 is a string C-group, we use the fact that both Γ0,1,n−1
and Γ0,n−2,n−1 are string C-groups acting on three orbits, and as a symmetric group
on the largest orbit. The intersection of these two groups cannot be any larger than
the direct product of two groups of order two with the symmetric group acting on
the points shared by the large orbits of both groups. This is exactly the structure of
Γ0,1,n−2,n−1 , so Γ0,n−1 is a string C-group.
Similarly, we show that Γ0 is a string C-group (the result then follows for Γn−1 ,
by self-duality). Since both Γ0,1 and Γ0,n−1 are string C-groups, we just need to show
that their intersection is Γ0,1,n−1 . This is true following the same logic as above,
analyzing the orbits of each subgroup.
27
Also, following Lemma 5.12, Γ0 is a symmetric group extended by a single transposition. Furthermore, since element (ρ2 ρ1 ρ2 ρ3 ρ2 )5 fixes all the nodes of the connected component of the graph of Γ0 , and interchanges the nodes of the isolated
edge labeled 2, it follows that this single transposition is in the group Γ0 . Thus
Γ0 ∼
= 2 × Sn+3 .
Finally, Γ is a string C-group since both Γ0 and Γn−1 are string C-groups, and
Γ0,1 ∼
= 2 × 2 × Sn+1 is maximal in Γ0 ∼
= 2 × Sn+3 .
6
Related problems and open questions
Some problems on the existence of internally self-dual polytopes remain open. Here
are perhaps the most fundamental.
Problem 1. For each odd p ≥ 5, describe an infinite family of internally self-dual
regular polyhedra of type {p, p}, or prove that there are only finitely many internally
self-dual regular polyhedra of that type.
Problem 2. Determine the values of p and q such that there is an internally self-dual
regular 4-polytope of type {p, q, p}.
Problem 3. Determine whether each self-dual (n − 2)-polytope occurs as the medial
section of an internally self-dual regular n-polytope.
To our knowledge, these problems are open even if we consider all self-dual regular
polytopes, rather than just the internally self-dual ones.
To what extent does the theory we have developed apply to transformations
other than duality? For example, the Petrie dual of a polyhedron P, denoted P π ,
is obtained from P by interchanging the roles of its facets and its Petrie polygons
(see [9, Sec. 7B]). If P π ∼
= P, then we say that P is self-Petrie. If P is regular, with
Γ(P) = hρ0 , ρ1 , ρ2 i, then P is self-Petrie if and only if there is a group automorphism
of Γ(P) that sends ρ0 to ρ0 ρ2 , while fixing ρ1 and ρ2 . We can then say that P is
internally self-Petrie if this automorphism is inner.
Working with internally self-Petrie polyhedra is not substantially different from
working with internally self-dual polyhedra, due to the following result.
Proposition 6.1. A regular polyhedron P is internally self-Petrie if and only if
(P ∗ )π is internally self-dual.
28
Proof. Suppose that P is (internally or externally) self-Petrie, and let Γ(P) =
hρ0 , ρ1 , ρ2 i. Then there is some π ∈ Aut(Γ(P)) such that ρ0 π = ρ0 ρ2 , ρ1 π = ρ1 ,
and ρ2 π = ρ2 . Now, Γ(P ∗ ) = hρ2 , ρ1 , ρ0 i, and then Γ((P ∗ )π ) = hρ0 ρ2 , ρ1 , ρ0 i =:
hλ0 , λ1 , λ2 i. Then it is easy to show that λi π = λ2−i for i = 0, 1, 2. It follows that
(P ∗ )π is self-dual. Furthermore, since Γ(P) ∼
= Γ((P ∗ )π ) as abstract groups and π
induces the self-Petriality of the former and the self-duality of the latter, it follows
that π is either inner for both groups or outer for both. The result then follows.
Perhaps there are other transformations of polytopes where it would make sense
to discuss internal versus external invariance under that transformation.
7
Acknowledgements
The computations done in this paper were verified independently using GAP [7] and
Magma [1].
References
[1] Wieb Bosma, John Cannon, and Catherine Playoust, The Magma algebra system. I. The user language, J. Symbolic Comput. 24 (1997), no. 3-4, 235–265,
Computational algebra and number theory (London, 1993). MR MR1484478
[2] Marston Conder, Lists of all regular polytopes with up to 2000 flags,
http://www.math.auckland.ac.nz/∼conder.
[3] Gabe Cunningham, Variance groups and the structure of mixed polytopes, Rigidity and Symmetry (Robert Connelly, Asia Ivic Weiss, and Walter Whiteley,
eds.), Fields Institute Communications, vol. 70, Springer New York, 2014,
pp. 97–116 (English).
[4] Maria Elisa Fernandes, Dimitri Leemans, and Mark Mixer, Polytopes of high
rank for the alternating groups, J. Combin. Theory Ser. A 119 (2012), no. 1,
42–56. MR 2844081
[5]
, An extension of the classification of high rank regular polytopes, (2015).
[6] William N Franzsen, Automorphisms of coxeter groups, Ph.D. thesis, University
of Sydney, 2001.
29
[7] The GAP Group, GAP – Groups, Algorithms, and Programming, Version
4.4.12, 2008.
[8] Michael I. Hartley, An atlas of small regular abstract polytopes, Periodica Mathematica Hungarica 53 (2006), 149–156, 10.1007/s10998-006-0028-x.
[9] Peter McMullen and Egon Schulte, Abstract regular polytopes, Encyclopedia of
Mathematics and its Applications, vol. 92, Cambridge University Press, Cambridge, 2002. MR 1965665 (2004a:52020)
[10] Barry Monson, Daniel Pellicer, and Gordon Williams, Mixing and monodromy
of abstract polytopes, Transactions of the American Mathematical Society 366
(2014), no. 5, 2651–2681.
[11] Daniel Pellicer, CPR graphs and regular polytopes, European Journal of Combinatorics 29 (2008), no. 1, 59–71.
30
| 4math.GR
|
1
Energy-Efficient Resource Allocation for
Elastic Optical Networks using Convex
arXiv:1705.06891v1 [cs.IT] 19 May 2017
Optimization
Mohammad Hadi, Member, IEEE, and Mohammad Reza Pakravan, Member, IEEE
Abstract
We propose a two-stage algorithm for energy-efficient resource allocation constrained to QoS and
physical requirements in OFDM-based EONs. The first stage deals with routing, grooming and traffic
ordering and aims at minimizing amplifier power consumption and number of active transponders. We
provide a heuristic procedure which yields an acceptable solution for the complex ILP formulation of the
routing and grooming. In the second stage, we optimize transponder configuration including spectrum
and transmit power parameters to minimize transponder power consumption. We show how QoS and
transponder power consumption are represented by convex expressions and use the results to formulate a
convex problem for configuring transponders in which transmit optical power is an optimization variable.
Simulation results demonstrate that the power consumption is reduced by 9% when the proposed routing
and grooming algorithm is applied to European Cost239 network with aggregate traffic 60 Tbps. It is
shown that our convex formulation for transponder parameter assignment is considerably faster than
its MINLP counterpart and its ability to optimize transmit optical power improves transponder power
consumption by 8% for aggregate traffic 60 Tbps. Furthermore, we investigate the effect of adaptive
modulation assignment and transponder capacity on inherent tradeoff between network CAPEX and
OPEX.
Index Terms
May 22, 2017
DRAFT
2
Convex optimization, Green communication, Elastic optical networks, Resource allocation, Quality
of service.
I. I NTRODUCTION
Energy consumption of equipments in telecommunication networks has been recognized as
one of major contributors in overall future energy consumption. Traditional fixed grid optical
networks are not adaptive and therefore, they can not efficiently use system resources such as
transmit optical power and spectrum to minimize network energy consumption. Elastic Optical
Networks (EON) can provide an energy-efficient network configuration by adapting the allocated
resources to the communication demands and physical conditions. It has been shown that the
cost and energy savings of migration to an energy-efficient EON outweigh the added cost of the
equipments [1]. Higher energy efficiency of Orthogonal Frequency Division Multiplex (OFDM)based optical networks has been demonstrated in [2] while the outstanding features such as
scalability, compatibaility and adaptation which nominate OFDM as the enabling technology of
EONs have been reported in [3].
Many variants of optimization formulations and algorithms have been proposed for resource
assignment in optical networks [4]–[7]. Optimized Routing and Wavelength Assignment (RWA)
in nonlinear Wavelength Division Multiplex (WDM) networks has been studied in [8]–[11]
which is not applicable to EON with variable assignment of spectrum width and optical carrier
frequencies. Some of the proposed solutions for resource allocation in EONs do not consider
the Quality of Service (QoS) requirements in their optimization analysis [7], [12], [13] which
is a necessary parameter for practical network planning. Some of the researchers have proposed
a general nonlinear expression to include QoS requirement as a fundamental constraint of the
resource assignment optimization problem [14]–[17] but several of them have focused on energyefficiency [18], [19]. Khodakarami et al. have proposed a Mixed-Integer NonLinear Program
(MINLP) for energy-efficient resource allocation in EONs which is decomposed into three subDRAFT
May 22, 2017
3
problems with heuristic solutions [19]. Their general formulation selects the best configuration
for transponders and grooming switches such that the total network power consumption is
minimized. Although they consider QoS requirement, they do not take into account the transmit
optical power as an optimization variable. Traffic grooming is employed to reuse active physical
components in the network and thereby increase the energy efficiency [19], [20]. However, there
is no comprehensive work that simultaneously considers traffic grooming along with adaptive
assignment of spectrum and transmit power variables and investigates their effect on different
network elements.
A convex optimization is a type of mathematical optimization problems characterized by
convex objective and constraint functions. Because of their desirable properties, convex optimization problems can be solved with a variety of efficient algorithms. There are software
packages that can be used to solve large scale convex optimization problems [21], [22]. Using
Geometric Programming (GP) convexification techniques, communication optimization problems
can be converted to convex form to allow us to use fast and efficient convex optimizer software
packages [17], [23].
The introduction of OFDM-based EONs, as a promising solution for developing green core
optical networks, has raised new networking challenges and increased resource provisioning
complexity. This mandates designing efficient and fast algorithms to utilize network adaptation
capabilities for reducing power consumption. In this paper, we consider the energy-efficient resource allocation in OFDM-based EONs and following the approach of [15], [19], we decompose
it into two interconnected optimization sub-problems 1) Routing, Grooming and Ordering (RGO)
2) Transponder Parameter Assignment (TPA). RGO is responsible for routing the traffic along
the best possible path while minimizing the number of assigned lightpaths and used transponders
by aggregating traffic partitions into a traffic segment closer to the information bit rate of the
optical transponders. It also defines the orders of traffic routes on each link. In TPA, transponder
variables such as transmit optical power, modulation level, coding rate, number of sub-carriers
May 22, 2017
DRAFT
4
and location of optical carrier frequency are optimally configured such that total transponder
power consumption is minimized, QoS requirements are guaranteed and physical constraints
are satisfied. We formulate RGO as an Integer Linear Program (ILP) and provide a heuristic
approach which yields its near-optimum solution in a reasonable time. Long and high volume
traffic requests are more susceptible to transmission impairments and have considerable power
consumption and spectrum usage. In RGO, we mainly focus on such long and high volume
traffic requests and groom them over shorter lightpaths to reduce their high amount of power
consumption and spectrum usage. TPA is conventionally formulated as a complex MINLP which
is an NP-hard problem. A key contribution of this paper is that we have proposed a method
to covert the complex MINLP to a convex problem which can be solved with readily available
convex optimization algorithms [22]. In the proposed convex formulation, joint optimization
of power and spectrum variables are addressed. We consider Optical Signal to Noise Ratio
(OSNR) as an indicator of QoS and show how the required levels of QoS can be expressed by
a convex inequality constraint. We also provide a convex expression for power consumption of
transponders. Simulation results show that the proposed convex formulation provides the same
results as its MINLP counterpart but in a much shorter time. Simulation results also demonstrate
that optimizing transmit optical power considerably improves the power consumption of different
network elements compared to the fix assignment of the transmit optical power. As another result,
we demonstrate that it is not required to have a complex transponder supporting high number
of different modulation levels and coding rates. We also show that how transponder capacity
affects the expected power saving of our proposed algorithm for RGO. Furthermore, we reveal the
inherent tradeoff between CAPital EXpenditure (CAPEX) and OPerational EXpenditure (OPEX)
as the transponder capacity increases and show that increasing transponder capacity increases
CAPEX while results in lower amount of OPEX.
The rest of the paper is organized as follows. System model is introduced in Section II. In
Sections III, we propose our two-stage resource allocation algorithm. Section IV is devoted to
DRAFT
May 22, 2017
5
RGO sub-problem while in Section V, we show how TPA sub-problem is formulated as a convex
optimization problem. Simulation results are included in VI. Finally, we conclude the paper in
Section VII.
Notation. Optimization variables are shown in lower case letters. Bold letters are used to
denote vectors and sets. We use calligraphic font to show constants. Functions are also shown
by capital letters and their arguments are given by lower indexes. Optical fiber characterizing
parameters are shown by their typical notations.
II. S YSTEM M ODEL
Consider an optical network and assume that V and L denote its sets of optical nodes and
directional optical fiber links. Optical fiber l begins at node B(l), ends at node E(l) and has
length Ll . The optical fiber bandwidth B is assumed to be gridless. Q is the set of traffic requests.
Each traffic request originates from source node S(q), terminates at destination node D(q) and
has volume of Rq bit per second. As shown in Fig. 3, each optical node has a bank of transmit
and receive transponders and an electronic grooming switch. Input/output traffic requests and
added/dropped traffics can be merged in grooming switch to reduce the unused capacity of
transponders and number of active transponders [19]. There are at most T pairs of transmit and
receive transponders in each bank which are indexed by T = {1, · · · , T }. f(v,t,q) shows the
part of qth traffic request that may be added at tth transmit transponder of node v, i.e. transmit
transponder (v, t). Similarly, c(v,t,q) denotes the part of qth traffic request that may be dropped
at receive transponder (v, t). x(v,t,l) is a binary variable that shows if link l is used for routing
the traffics added by transmit transponder (v, t). Furthermore, binary variable d(v,t,v0 ,t0 ) shows
that if the added traffic of transmit transponder (v, t) is dropped at receive transponder (v 0 , t0 ).
Each transmit transponder is given a contiguous spectrum bandwidth ∆(v,t) around optical carrier
frequency ω(v,t) . The assigned contiguous spectrum bandwidth includes 2b(v,t) OFDM sub-carriers
with sub-carrier space of F so, ∆(v,t) = 2b(v,t) F. To facilitate optical switching and remove the
May 22, 2017
DRAFT
6
TABLE I: Available modulation levels and coding rates along with their corresponding minimum required OSNR
Θ(c, r) to achieve a pre-FEC BER values of 1 × 10−4 .
r/c
2/3
3/4
8/9
1
1.5
1.7
3.6
2
2.3
2.9
4.6
3
5.9
7.8
12.9
4
9.1
12.0
20.9
5
17.4
24.0
42.7
6
28.8
40.7
75.8
high cost of spectrum conversion [24], we assume that the assigned spectrum bandwidth to each
transponder is continuous over its routed path. The added traffic of transponder (v, t) passes
N(v,t) equal-length fiber spans along its routed path and has N(v,t,v0 ,t0 ) shared fiber spans with
transmit transponder (v 0 , t0 ) 6= (v, t). Each fiber span has fixed length Lspn and is equipped
with its own optical amplifier to compensate for the attenuation. Hl shows the set of transmit
transponders that share optical fiber l on their allocated routes. There are pre-defined modulation
levels c and coding rates r where each pair of (c, r) requires minimum OSNR Θ(c, r) to get a
pre-Forward-Error-Correction (FEC) Bit-Error-Rate (BER) value of 1 × 10−4 , as shown in Tab. I
[15]. The optical transponder (v, t) is given a modulation level c(v,t) and a coding rate r(v,t) and
fills its assigned optical bandwidth ∆(v,t) with optical power p(v,t) . Transponders have maximum
information bit rate C and fill both of the polarizations with the same power. The configuration
of each receive transponder is the same as its corresponding transmit transponder. According to
filtering and switching constraints in optical nodes, we consider a guard band G between any
two adjacent spectrum bandwidths of a link.
Transponders, grooming switches and optical amplifiers are the main sources of network power
consumption [12], [19] so, total network power consumption is given by:
X
v,t
X(v,t) +
X
v
G(v) +
X
A(l)
(1)
l
where X(v,t) is total power consumption of each linked pair of transmit and receive transponders,
G(v) is power consumption of grooming switch v and Al shows the total power consumption of
DRAFT
May 22, 2017
7
IP Router
Groom Switch
Fiber
Fiber
RX Transponder
TX Transponder
Drop
Add
Optical Switch
Fiber
Fiber
Optical Node
Fig. 1: Optical node block diagram which has a bank of transmit and receive transponders and an electronic
grooming switch.
optical amplifiers over link l. Considering the architecture of Fig. 2, X(v,t) can be calculated as
follows:
X(v,t) = Ptxb +
Penc Pf f t b(v,t)
Pdec
Pf f t b(v,t)
+
2
b(v,t) + Prxb +
+
2
b(v,t) + 2b(v,t) Pdsp
r(v,t)
2
r(v,t)
2
(2)
where Ptxb and Prxb are transmit and receive transponder bias terms (which includes the power
consumption of bit interleaver, laser, A/D converter, D/A converter, Mach–Zehnder modulator,
serial/parallel, parallel/serial), Penc and Pdec are the scaling coefficient of encoder and decoder
power consumptions, Pf f t denotes the power consumption for a two point FFT operation and
Pdsp is the power consumption scaling coefficient of the receiver DSP operations. The electronic
grooming switch consumes Egrm per each switched bit [19]. Considering the total groomed traffic
at node v, we can write [19]:
G(v) = Egrm
X
X
[f(v,t,q) + c(v,t,q) ] − Egrm
Rq [δ(v − S(q)) + δ(v − D(q))]
t,q
(3)
q
where δ() is discrete impulse function. Each optical amplifier also needs Pamp input power,
May 22, 2017
DRAFT
To Fiber
From Fiber
Iy
Qy
Polarization Diversity
Optical Hybrid
Qx
A/D A/D A/D A/D
Synchronization
Serial/Parallel
FFT
Ix
Equalization
Symbol Demapping
Parallel/Serial
Decoder
Output Stream
Sy
Polarization Beam Splitter
Sx
Oscilator
IQ Modulator
Laser
Qx
D/A
Parallel/Serial
Prefix/ Pilot
IFFT
Symbol Mapping
Serial/Parallel
Encoder
Input Stream
Ix
D/A
8
Fig. 2: Block diagram of a pair of OFDM-based transmit and receive transponders.
therefore the total power consumption of optical amplifiers over an active link is [19]:
A(l)
ih
i
h L
X
l
+ 1 1 − δ(
x(v,t,l) )
= Pamp
Lspn
v,t
(4)
III. R ESOURCE A LLOCATION P ROBLEM
To have a green EON, we need a resource allocation algorithm to determine the values of
system model variables such that the network consumes the minimum power while physical
constrains are satisfied and desired levels of QoS (or equivalently desired levels of OSNR) are
guaranteed. In general, such a problem is modeled as an optimization problem and it is shown
to be NP-hard [15]. Therefore, we propose a two-stage algorithm in which the complex resource
allocation problem is decomposed into two sub-problems which are named RGO and TPA. The
pseudo code of this algorithm is shown in Alg. 1. Solving RGO sub-problem, the connection
requests are routed, the grooming switches are configured and the order of traffics on links are
determined. Transponder parameters such as transmit optical power, modulation level, coding
rate, number of sub-carriers and optical carrier frequency are optimized during TPA sub-problem.
DRAFT
May 22, 2017
9
The power consumption of optical amplifiers and grooming switches are respectively related to
the traffic routes and configuration of the grooming switches which are mainly addressed in RGO
sub-problem. On the other hand, total transponder power consumption are minimized by solving
TPA sub-problem. Usually the search for a near optimal solution involves iterations between
these two sub-problems. To save this iteration time, it is of great interest to hold the run time
of each sub-problem at its minimum value.
In the coming sections, we formulate RGO problem as an ILP and then provide a heuristic
solution for it. We also introduce an MINLP formulation for TPA sub-problem and show how
this complex formulation can be converted to a convex optimization problem.
IV. ROUTING , G ROOMING AND O RDERING S UB -P ROBLEM
In this section, an ILP is formulated for routing and traffic grooming and then a heuristic
approach is proposed to solve it.
A. ILP Formulation
Generally, traffic grooming aims at minimization of active transponders to save transponder
power consumption while shortest path routing is usually used to minimize the length of routes
and consequently the number of used optical amplifiers [19]. Routing and traffic grooming can
May 22, 2017
DRAFT
10
Algorithm 1 Resource Allocation Algorithm
Input: network topology, traffic matrix, link parameters, noise and nonlinearity parameters, transponder parameters
Output: routes, groomed traffics, number of subcarriers, optical carrier frequencies, modulation levels, coding rates,
transmit optical powers
Stage 1: Routing, Grooming, Ordering
for all connection requests q do
while Rq > C do
add a new full load transmit transponder at node S(q);
add a new full load receive transponder at node D(q);
use shortest path for the added transponder pair;
Rq ← Rq − C;
end while
end for
sort traffics q = 1, · · · ,|Q| decreasingly according to the product of shortest path length and rate;
for all traffics q do
MSPL ← ∞;
MATC ← 0;
for all groom scenarios on qth traffic shortest path do
MSPLt ← MSPL of the current path decomposition;
MATCt ← MATC of the current path decomposition;
if MATCt < Rq then
continue;
end if
if MSPLt < MSPL then
MSPL ← MSPLt ;
MATC ← MATCt ;
end if
end for
groom qth request to the scenario having MATC and MSPL;
update sub-paths traffic volumes;
remove the groomed path;
end for
assign transmit and receive transponder pairs to the remaining traffic requests;
route the remaining connection requests over their shortest path;
order all the assigned paths according to the product of shortest path length and rate;
Stage 2: Transponder Parameter Assignment
31: do
use convex TPA to optimize transponder parameters;
32:
33:
I ← 0;
34:
do
35:
for all transmit transponders (v, t) do
for all spectral efficiency values c of Tab. I do
36:
37:
if if c(v,t) falls in I-neighborhood of c then
38:
round and fix c(v,t) to c;
39:
break;
40:
end if
41:
end for
42:
for all coding rate values r of Tab. I do
43:
if if r(v,t) falls in I-neighborhood of r then
44:
round and fix r(v,t) to r;
45:
break;
46:
end if
47:
end for
48:
end for
49:
I ← I +J;
50:
while there is no valid value in c and r
51:
remove fixed values of c and r;
52: while there is an inacceptable value in c and r
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:
DRAFT
May 22, 2017
11
be formulated as the following ILP:
X
min
x,d,f ,c
f(v,t,q) + δ(v − D(q))Rq =
t∈T
X
d(v,t,v0 ,t0 ) + K2
v,v 0 ∈V,t,t0 ∈T
v∈V,l∈L,t∈T
X
s.t.
X
x(v,t,l) Ll + K1
X
X
(C −
v∈V,t∈T
X
f(v,t,q) )
(5)
q∈Q
c(v,t,q) + δ(v − S(q))Rq , ∀v ∈ V, ∀q ∈ Q
(6)
t∈T
f(v,t,q) 6 C, ∀v ∈ V, ∀t ∈ T
(7)
c(v,t,q) 6 C, ∀v ∈ V, ∀t ∈ T
(8)
q∈Q
X
q∈Q
c(v0 ,t0 ,q) − f(v,t,q) 6 K(1 − d(v,t,v0 ,t0 ) ), ∀v, v 0 ∈ V, ∀t, t0 ∈ T, ∀q ∈ Q
X
f(v,t,q) 6 K
q∈Q
X
l∈L:E(l)=v 0
X
(9)
x(v,t,l) , ∀v ∈ V, ∀t ∈ T
(10)
X
(11)
l∈L:B(l)=v
x(v,t,l) 6
X
l∈L:B(l)=v 0
x(v,t,l) +
d(v,t,v0 ,t0 ) 6 1, ∀v, v 0 ∈ V, ∀t ∈ T
t0 ∈T
where K is a large positive number. Each transmit transponder is assigned a route to its destination
so, minimization of used link lengths automatically results in simultaneous shortest path routing
and minimum number of active transponders. This is the main cost of the goal function. The
second term of the goal function is a penalty term to prevent unnecessary add/drop of the
traffic paths in grooming switches. Furthermore, we add another penalty term to incur a cost for
transponders with unused remaining capacity and therefore, force the transponders to be fully
utilized. Constraint (6) is KCL law for added, dropped, input, output and groomed traffic of
qth connection request at node v. Constraints (7) and (8) hold the information bit rate of the
transmit and receive transponders below their capacity. Constraint (9) states that all the traffics
of a transmit transponder should be dropped at its assigned receive transponder. Constraint (10)
specifies that if a transponder has no allocated route, it should be inactive. Finally, constraint
(11) is the optical switching constraint which states that if a traffic arrives to a node, it should
be dropped to a receive transponder or switched to a neighbor node. This optimization problem
May 22, 2017
DRAFT
12
has 2|V||T||Q| +|V||T||L| +|V|2 |T|2 integer variables and its number of constraints has the
order of |V|2 |T|2 |Q|. Although the problem is ILP, its number of variables and constraints is
very large for practical values of |V|, |T| and |Q|. Therefore, a heuristic approach is required to
provide a fast-achievable near-optimum solution for routing and traffic grooming problem.
B. Heuristic Solution
Long and high volume traffic requests are more susceptible to amplifier noise and nonlinear
interference effects and need low level modulations, low coding rates, high transmit optical
powers and therefore consume a lot of power. If such traffics are groomed to short traffic
segments, each segment can be routed using high modulation levels and coding rates and
consequently consumes lower spectrum and power. Merging long and high volume traffics to
existing short traffic requests also reduces the number of required transponders and assigned
links by utilizing the unused remaining capacity of the transponders which is our goal in ILP
of (5)-(11). It is conventionally stated that if the path length is reduced by a factor of 0.5,
modulation level can be doubled [4]. This in turn reduces the assigned spectrum width by a
factor of 0.5 and results in lower transponder power consumption according to (2). Consider
the sample scenario of Fig. 3 in which transmit transponders are shown by rectangles below
the nodes. For each transponder, the used capacity is highlighted by dark color and the number
below each transponder shows its destination node. Assume we should route an input traffic with
rate
C
2
from source 1 to destination 4 over its shortest path which is 1 → 2 → 3 → 4. There are
four grooming scenarios over this path which are summerized in Tab. II. Minimum Available
Transponder Capacity (MATC) is defined as the remaining capacity of the transmit transponder
that has the maximum load over sub-paths of the grooming scenario. For example in scenario
3, we should use transponder 3 of node 1 and transponder 4 of node 3 which both of them
have remaining capacity of C2 , therefore MATC = C2 . Maximum Sub Path Length (MSPL) is the
maximum length of the sub-paths along the grooming scenario. For example in scenario 3, the
DRAFT
May 22, 2017
13
1
1500
2 3 4
2
3 4
2000
3
4
1000
4
Fig. 3: An example for traffic grooming.
TABLE II: Grooming scenarios for example of Fig. 3
Scenario Number
1
2
.
3
4
Grooming Scenario
1→4
1→2→4
1→3→4
1→2→3→4
MATC
C
0
C
2
0
MSPL
4500
3000
3500
2000
sub-path 1 → 3 has the maximum length which is 3500. As the value of MSPL decreases, there
are more chances to partition the path to shortest sub-paths and save power. As reported in Tab.
II, 1 → 3 → 4 has the minimum MSPL and its MATC is enough to accommodate the required
load and therefore, it is our chosen grooming scenario. Note that at worst case the traffic is
routed over its shortest path without grooming to sub-paths.
During RGO stage of Alg. 1, traffic volumes are partitioned such that no traffic volume is
greater than transponder capacity C. Then, we choose partitioned traffic volumes with full load
of C and route them over their shortest paths. The remaining traffic volumes are descendingly
sorted according to the product of traffic volume and shortest path length. Next, we groom the
sorted traffics using the procedure explained above, as written in Alg. 1. Finally, we use the
metric of rate-length product to order traffics over links as explained in [17].
V. T RANSPONDER PARAMETER A SSIGNMENT S UB -P ROBLEM
In this section, we provide an MINLP formulation for TPA and then show how it can be
converted to a convex optimization problem.
May 22, 2017
DRAFT
14
A. Nonlinear Formulation
A MINLP formulation for TPA is as follows [15]:
X
min
c,b,r,p,ω
X(v,t)
Ψ(v,t) > Θ(v,t) ,
s.t.
(12)
v∈V,t∈T
∀v ∈ V, ∀t ∈ T
∆Υl,j+1
∆Υl,j
+ G 6 ωΥl,j+1 −
, ∀l ∈ L, j = 1, · · · ,|Hl | − 1
2
2
∆(v,t)
6 ω(v,t) 6 B −
, ∀v ∈ V, ∀t ∈ T
2
ωΥl,j +
∆(v,t)
2
X
f(v,t,q) 6 2r(v,t) c(v,t) ∆(v,t) ,
∀v ∈ V, ∀t ∈ T
(13)
(14)
(15)
(16)
q
where c, b, r, p and ω are variable vectors of transponder configuration parameters i.e. modulation level, number of sub-carriers, coding rate, transmit optical power and central frequency. The
goal is to minimize the total transponder power consumption where X(v,t) is obtained using (2).
Constraint (13) is the QoS constraint that forces Ψ(v,t) , the OSNR of transponder (v, t), to be
greater than its required minimum threshold Θ(v,t) . Ψ(v,t) is a nonlinear function of b, p and ω
[16], [17] while the value of Θ(v,t) are picked up from Tab. I. Constraint (14) is nonoverlappingguard constraint that prevents two transmit transponders share a same frequency spectrum. Υl,j
is a function that shows which transponder occupies j-th spectrum bandwidth on link l and
its values are determined during ordering part of RGO sub-problem. Constraint (15) holds all
assigned central frequencies within the acceptable range of the fiber spectrum and finaly the
P
last constraint guarantees that the transponder can convey the input traffic rate q f(v,t,q) . In
fact, this MINLP is an extension of the formulation of [19] in which transmit optical powers
also contribute to the optimization and QoS constraint is specified by the general and nonlinear
expressions of [15].
DRAFT
May 22, 2017
15
80
#(c,r)
60
40
20
0
0.9
0.85
6
0.8
5
4
0.75
3
0.7
r
2
1
c
Fig. 4: 2D curve fitting for (c, r, Θ) data samples given in Tab. I.
B. Convexification Technique
We first provide a generalized posynomial expression [21] for the optimization and then
define a variable change to convexify the problem. A posynomial expression for OSNR has
been proposed in [17]:
Ψ(v,t) =
ζN(v,t) ∆(v,t) + κ1 ςp(v,t)
p(v,t)
P
p2(v0 ,t0 ) N(v,t,v0 ,t0 )
v 0 ∈V,t0 ∈T
(v 0 ,t0 )6=(v,t)
∆(v0 ,t0 ) d(v,t,v0 ,t0 )
where κ1 = 0.4343, ζ = (eαLspn − 1)hνnsp , ς =
3γ 2
2απ|β2 |
, ∀v ∈ V, ∀t ∈ T (17)
+ ςιN(v,t) p3(v,t)
and ι =
π 2|β2 |
.
2α
nsp is the spontaneous
emission factor, ν is the light frequency, h is Planck’s constant, α is attenuation coefficient, β2
is dispersion factor and γ is nonlinear constant. Furthermore, d(v,t,v0 ,t0 ) is the distance between
carrier frequencies ω(v,t) and ω(v0 ,t0 ) and equals to d(v,t,v0 ,t0 ) = ω(v,t) − ω(v0 ,t0 ) . Similar to [17],
May 22, 2017
DRAFT
16
we use a posynomial curve fitting for date samples of Tab. I:
Θ(c, r) ≈ rκ2 (1 + κ3 c)κ4 , 1 6 c 6 6, 0.6 6 r 6 1
(18)
where κ2 = 3.37, κ3 = 0.21, κ4 = 5.73. Fig. 4 shows data samples and the posynomial expression
for data curve fitting.
Following the same approach as [17], we arrive at this new representation of the optimization
problem:
min
c,b,r,p,ω,t,d
X
X
X(v,t) + K3
d−1
(v,t,v 0 ,t0 )
(19)
v,v 0 ∈V,t,t0 ∈T
(v,t)6=(v 0 ,t0 )
v∈V,t∈T
N(v,t,v0 ,t0 ) 6=0
s.t.
h
κ2
b(v,t)
4
r(v,t)
tκ(v,t)
ζFN(v,t) p−1
+ ςιN(v,t) p2(v,t) + κ1 ςF −1
(v,t) 2
X
p2(v0 ,t0 ) N(v,t,v0 ,t0 ) 2−b(v0 ,t0 ) d−1
(v,t,v 0 ,t0 )
i
v 0 ∈V,t0 ∈T
(v 0 ,t0 )6=(v,t)
6 1,
∀v ∈ V, ∀t ∈ T
(20)
−1
−1
−1
−1
ωΥl,j ωΥ
+ 0.5F2bΥl,j ωΥ
+ GωΥ
+ 0.5F2bΥl,j+1 ωΥ
6 1,
l,j+1
l,j+1
l,j+1
l,j+1
∀l ∈ L, j = 1, · · · ,|Hl | − 1
(21)
0.5F2b(v,t) B −1 + ω(v,t) B −1 6 1,
∀v ∈ V, ∀t ∈ T
−1
0.5F2b(v,t) ω(v,t)
6 1,
∀v ∈ V, ∀t ∈ T
−1 −1 −b(v,t)
0.5F −1 r(v,t)
c(v,t) 2
X
f(v,t,q) 6 1,
(22)
(23)
∀v ∈ V, ∀t ∈ T
(24)
q
−1
t−1
(v,t) + κ3 c(v,t) t(v,t) 6 1,
∀v ∈ V, ∀t ∈ T
(25)
−1
−1
d(Υl,i ,Υl,j ) ωΥ
+ ωΥl,j ωΥ
6 1, ∀l ∈ L, j = 1, · · · ,|Hl | − 1, i = j + 1, · · · ,|Hl |
l,i
l,i
(26)
−1
−1
d(Υl,i ,Υl,j ) ωΥ
+ ωΥl,i ωΥ
6 1,
l,j
l,j
(27)
∀l ∈ L, j = 2, · · · ,|Hl | , i = 1, · · · , j − 1
Ignoring constraints (25), (26), (27) and the penalty term of the goal function (19), the above
formulation is equivalent GP of the previous MINLP in which expressions (17) and (18) have
DRAFT
May 22, 2017
17
been used for QoS constraint (20). Constraints (25) and (26) and the penalty term are added to
guarantee the implicit equality of d(v,t,v0 ,t0 ) = ω(v,t) − ω(v0 ,t0 ) [17]. Constraint (25) is also needed
to convert the generalized posynomial QoS constraint to a valid GP one, as explained in [21].
Now, consider the following variable change:
x = eX : x ∈ R>0 −→ X ∈ R,
x = X : x ∈ R>0 −→ X ∈ R>0 ,
∀x ∈ c ∪ r ∪ p ∪ ω ∪ t ∪ d
∀x ∈ b
(28)
If one applies this variable change, the new representation of the problem is convex. As an
example for the goal function we have:
X
[Ptxb + Prxb + (Pdec + Penc )e−r(v,t) + b(v,t) 2b(v,t) Pf f t + 2b(v,t) Pdsp ] + K
v∈V,t∈T
X
e−d(v,t,v0 ,t0 )
v,v 0 ∈V,t,t0 ∈T
(v,t)6=(v 0 ,t0 )
N(v,t,v0 ,t0 ) 6=0
(29)
Clearly, e−d(v,t,v0 ,t0 ) and e−r(v,t) are convex over R while 2b(v,t) and b(v,t) 2b(v,t) are convex over
R>0 . Consequently, function (29) which is a nonnegative weighted sum of convex functions is
also convex. The same statement can be applied to show the convexity of the constraints under
variable change of (28) (for some constraints, we need to apply an extra log to both side of the
inequality to have a convex constraint).
The original MINLP has 5|V||T| variables and at most 4|V||T| +|V||T||L| constraints while
its convex representation requires 6|V||T|+|V|2 |T|2 variables and at most 5|V||T|+3|V||T||L|.
Also the new convex representation of the problem has more variables and constraints than its
MINLP counterpart but it can be quickly solved using fast convex optimization algorithms. In
the second stage of Alg. 1, a relaxed continuous version of the proposed mixed-integer convex
formulation is iteratively optimized in a local loop [21]. At each epoch, the continuous convex
optimization is solved and obtained values for relaxed integer variables (here, c(v,t) ’s and r(v,t) ’s)
May 22, 2017
DRAFT
18
TABLE III: Simulation constant parameters.
|β2 |
fs2 /m
20393
α
dB/km
0.22
L
km
80
ν
THz
193.55
nsp
1.58
γ
1/W/km
1.3
G
GHz
20
B
THz
2
J
0.1
Ptxb
W
16
Prxb
W
20
Penc
W
0.2
Pdec
W
3
Pf f t
mW
4
Pdsp
mW
10
Egrm
pJ/bit
400
Pamp
W
12
are rounded by precision I. If none of the rounded variables is valid based on Tab. I, the precision
is increased by J and the rounding is again applied. This process continues until at least one
valid element is found for relaxed integer variables. Then, we fix the acceptable rounded values
and solve the relaxed continuous convex problem again. The local loop continues until all the
integer variables have valid and acceptable values. Note that the number of iterations in the
heuristic algorithm is at most equal to the number of integer variables (here, at most 2|V||T|).
The number of iterations is usually lower than its maximum. Furthermore, a simpler problem
should be solved as the number of iteration increases because of fixing some of the integer
variables during each loop.
VI. N UMERICAL R ESULTS
In this section, we use simulation results to demonstrate the performance of the proposed
formulations and algorithms. The European Cost239 optical network is considered with the
topology of Fig. 5 and the normalized traffic matrix given in Tab. IV. One can scale the
normalized traffic values of Tab. IV to get 2017 west Europe aggregate traffic value of 16.8
EByte/month [19], [25]. The optical link parameters and other simulation constants are reported
in Tab. III. We use MATLAB software for programming and simulation [26]. YALMIP [27]
and CVX [28] optimization applications are also used for modeling and solving the proposed
optimization problems. Simulations run over a computer equipped with Intel Core i7-472HQ
CPU and 8 GB of RAM.
Fig. 6 compares total power consumption versus aggregate network traffic with and without
traffic grooming procedure of Alg. 1. Clearly, traffic grooming reduces the consumed power and
DRAFT
May 22, 2017
19
1
390
1310
760
660
4
740
3
550
210
2
340
390 1090
390
5
220
450
7
660
730
300
6
400
565
350
8
600
9
320
730
10
930
820
320
820
11
Fig. 5: European Cost239 optical network with 11 optical nodes and 26 bi-directional links. Digits on each link
shows its length in km.
TABLE IV: European Cost239 optical network normalized traffic matrix.
Node
01
02
03
04
05
06
07
08
09
10
11
01
00
01
01
03
01
01
01
35
01
01
01
02
01
00
05
14
40
01
01
10
03
02
03
03
01
05
00
16
24
01
01
05
03
01
02
04
03
14
16
00
06
02
02
21
81
09
09
05
01
40
24
06
00
01
11
06
11
01
02
06
01
01
01
02
01
00
01
01
01
01
01
07
01
01
01
02
11
01
00
01
01
01
01
08
35
10
05
21
06
01
01
00
06
02
05
09
01
03
03
81
11
01
01
06
00
51
06
10
01
02
01
09
01
01
01
02
51
00
81
11
01
03
02
09
02
01
01
05
06
81
00
the amount of saved power is improved for higher aggregate traffic values where more traffic
grooming scenarios are available. As an example, the power consumption is reduced by a factor
of 9% when aggregate traffic is about 60 Tbps.
Alg. 1 partitions the input traffic streams according to the capacity of transponders C and uses
May 22, 2017
DRAFT
20
40
Without Tra/c Grooming
With Tra/c Grooming
35
Total Power Consumption (kW)
30
25
20
15
10
5
0
5
10
15
20
25
30
35
40
45
50
55
Aggregate Tra/c (Tbps)
Fig. 6: Total power consumption versus aggregate traffic with and without applying Alg. 1 for traffic grooming.
the proposed traffic grooming procedure to fill transponders with remaining unused capacity.
We define Transponder Utilization Ratio (TUR) as the ratio of the aggregate traffic to the total
capacity offered by all transponders. TUR = 1 means that all the transponders work at full
capacity. Traffic Grooming Ratio (TGR) is defined as the ratio of the number of unloaded
transponders by traffic grooming to the number of unfilled transponders when no traffic grooming
is applied. TGR shows how traffic grooming reduces the number of required transponders.
Fig. 7 shows TUR and TGR versus transponder capacity C. As C increases, there are more
transponders with unused remaining capacity that can be used for traffic grooming and therefore,
TGR is improved. On the other hand, we need more traffic volumes to use full capacity of the
transponders when C is increased and therefore, TUR goes down. Higher values of TUR reduces
the required CAPEX for network deployment while higher values of TGR reduces the power
consumption and consequently reduces the OPEX. Fig. 7 shows that there is a tradeoff between
DRAFT
May 22, 2017
21
1
Transponder Utilization Ratio
Tra/c Grooming Ratio
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
100
200
300
400
500
600
700
800
900
1000
Transponder Capacity (Gbps)
Fig. 7: Transponder utilization and traffic grooming ratios versus transponder capacity C.
TUR and TGR or equivalently CAPEX and OPEX for higher values of C.
Fig. 8 shows TGR versus aggregate network traffic for different values of transponder capacity.
As the aggregate traffic increases, more traffics are groomed and we save more power due to
unloaded transponders. The improvement is more for transponders with higher capacity. As an
example, when the aggregate traffic is 67 Tbps and C increases from 200 Gbps to 400 Gbps,
TGR is improved by a factor of 1.51.
Fig. 9 shows how our proposed TPA convex formulation runs faster than its MINLP counterpart. We use MINLP formulations of (12)-(16) as a test bench in which QoS constraint is replaced
by the nonlinear and general expression of [15]. Results show that the convex formulation is
much faster than the MINLP formulation. As an example, it can be about 20 times faster at
aggregate traffic of 18 Tbps. Simulation results also show that for higher values of aggregate
traffic, the nonlinear optimizer package may fail to provide a feasible solution while the convex
May 22, 2017
DRAFT
22
0.35
C = 200 Gbps
C = 400 Gbps
C = 600 Gbps
0.3
Tra/c Grooming Ratio
0.25
0.2
0.15
0.1
0.05
0
10
20
30
40
50
60
70
80
90
Aggregate Tra/c (Tbps)
Fig. 8: Traffic grooming ratio versus aggregate network traffic for different values of transponder capacity.
optimizer software still yields a feasible solution in a reasonable run time.
The effect of adaptive assignment of coding rate and modulation level on power consumption
has been investigated in literature [19]. We also analyze this issue and evaluate the complexity of
transponders. Fig. 10 shows total power consumption of different network elements in terms of
the number of available modulation formats. The power consumption values are normalized to
their corresponding values for the scenario with two available modulation formats c = 1, 2. An
interesting observation in Fig. 10 is that increasing the number of modulation levels decreases
the power consumption of various network elements with the DSP power consumption showing
the highest level of power reduction. Clearly, as the number of modulation levels increases,
the amount of power consumption decreases but there is no considerable gain if the number of
modulation levels be more than 6. On the other hand, increasing the number of modulation levels
in a transponder may have an important effect on the complexity of the receiver architecture
DRAFT
May 22, 2017
23
104
Mixed Integer Nonlinear Formulation
Convex Formulation
Run Time (s)
103
102
101
100
4
6
8
10
12
14
16
18
Aggregate Tra/c (Tbps)
Fig. 9: Run time in terms of aggregate traffic for MINLP and convex formulations of transponder parameter
assignment problem.
and its signal processing requirements. Therefore, one should choose the number of modulation
levels in a transponder to have a balance between complexity (CAPEX) and power consumption
(OPEX).
The effect of adaptive assignment of coding rate and modulation level on power consumption has been studied in the literature. However, to the best of our knowledge, there is no
comprehensive investigation for the effect of adaptive transmit optical power assignment on
power consumption of different network elements. The total power consumption of different
network elements in terms of aggregate traffic with and without adaptive transmit optical power
assignment has been reported in Fig. 11. We have used the proposed expression of [29] for fixed
assignment of transmit optical power. Clearly, for all the elements, the total power consumption
is approximately a linear function of aggregate traffic but the slope of the lines are lower when
May 22, 2017
DRAFT
24
1
0.9
0.8
0.7
0.6
0.5
0.4
0.3
Normalized
Normalized
Normalized
Normalized
0.2
Total
Total
Total
Total
Encoder+Decoder Power Consumption
Transmit+Receive Transponder Power Consumption
IFFT/FFT Power Consumption
DSP Power Consumption
0.1
1
2
3
4
5
6
7
8
Number of Available QAM Modulations
Fig. 10: Normalized total power consumption of different network elements in terms of the number of available
modulation formats.
transmit optical powers are adaptively assigned. As an example, adaptive transmit optical power
assignment improves total transponder power consumption by a factor of 8% for aggregate traffic
of 60 Tbps.
VII. C ONCLUSION
Energy-efficient resource allocation and quality of service provisioning are two fundamental
problems of green elastic optical networks that are conventionally formulated using complex
mixed-integer nonlinear optimization problems. To address the complexity of the problem and
provide a fast-achievable near-optimum solution, we propose a two-stage resource allocation
algorithm in which the main problem is decomposed into two inter-connected sub-problems
named routing, grooming and ordering, and transponder parameter assignment. We introduce
a heuristic approach for routing, grooming and ordering sub-problem that determines trafDRAFT
May 22, 2017
25
40
Total
Total
Total
Total
Total
Total
Total
Total
35
30
Encoder+Decoder Power Consumption (Adaptive Transmit Power)
Encoder+Decoder Power Consumption (Fixed Transmit Power)
DSP Power Consumption (Adaptive Transmit Power)
DSP Power Consumption (Fixed Transmit Power)
FFT+IFFT Power Consumption (Adaptive Transmit Power)
FFT+IFFT Power Consumption (Fixed Transmit Power)
Transmit+Receive Transponder Power Consumption (Adaptive Transmit Power)
Transmit+Receive Transponder Power Consumption (Fixed Transmit Power)
kW
25
20
15
10
5
0
5
10
15
20
25
30
35
40
45
50
55
Aggregate Tra/c (Tbps)
Fig. 11: Total power consumption of different network elements in terms of aggregate traffic with and without
adaptive transmit optical power assignment.
fic routes, active transponders and grooming switch configuration to reduce amplifier power
consumption and number of active transponders. To have an efficient transponder parameter
assignment, we formulate a convex optimization problem in which joint optimization of power
and spectrum resources constrained to physical and quality of service requirements is addressed.
We use geometric convexification techniques to provide convex expressions for transponder
power consumption and optical signal to noise ratio as a metric of quality of service. The
performance of the proposed resource allocation procedure is validated through simulation results.
We demonstrate the improved performance of our presented routing and grooming procedure in
terms of network power consumption, transponder utilization ratio and traffic grooming ratio.
Furthermore, numerical outcomes show that the proposed convex formulation for configuring
the transponders yields the same results in a considerable lower run time compared to its
mixed-integer nonlinear counterpart. The improved effect of the joint optimization of power and
May 22, 2017
DRAFT
26
spectrum variables on power consumption of different network elements is also demonstrated by
simulation. According to the results, adaptive assignment of transmit optical power considerably
reduces total transponder power consumption. Furthermore, we analyze the inherent trade-off
between CAPEX and OPEX as the capacity and/or design complexity of the transponders change.
R EFERENCES
[1] E. Palkopoulou, M. Angelou, D. Klonidis, K. Christodoulopoulos, A. Klekamp, F. Buchali, E. Varvarigos, and I. Tomkos,
“Quantifying spectrum, cost, and energy efficiency in fixed-grid and flex-grid networks [invited],” Journal of Optical
Communications and Networking, vol. 4, no. 11, pp. B42–B51, 2012.
[2] A. Nag, T. Wang, and B. Mukherjee, “On spectrum-efficient green optical backbone networks,” in Global Telecommunications Conference (GLOBECOM 2011), 2011 IEEE.
IEEE, 2011, pp. 1–5.
[3] J. Armstrong, “Ofdm for optical communications,” Journal of lightwave technology, vol. 27, no. 3, pp. 189–204, 2009.
[4] B. C. Chatterjee, N. Sarma, and E. Oki, “Routing and spectrum allocation in elastic optical networks: A tutorial,” IEEE
Communications Surveys & Tutorials, vol. 17, no. 3, pp. 1776–1800, 2015.
[5] F. S. Abkenar and A. G. Rahbar, “Study and analysis of routing and spectrum allocation (rsa) and routing, modulation
and spectrum allocation (rmsa) algorithms in elastic optical networks (eons),” Optical Switching and Networking, vol. 23,
pp. 5–39, 2017.
[6] N. Sambo, G. Meloni, F. Cugini, A. D’Errico, L. Potı̀, P. Iovanna, and P. Castoldi, “Routing code and spectrum assignment
(rcsa) in elastic optical networks,” Journal of Lightwave Technology, vol. 33, no. 24, pp. 5114–5121, 2015.
[7] M. Hadi and M. R. Pakravan, “Improved routing and spectrum assignment formulations for optical ofdm networks,” pp.
1–6, 2016.
[8] I. Roberts, J. M. Kahn, and D. Boertjes, “Convex channel power optimization in nonlinear wdm systems using gaussian
noise model,” Journal of Lightwave Technology, vol. 34, no. 13, pp. 3212–3222, 2016.
[9] D. J. Ives, P. Bayvel, and S. J. Savory, “Physical layer transmitter and routing optimization to maximize the traffic throughput
of a nonlinear optical mesh network,” in Optical Network Design and Modeling, 2014 International Conference on. IEEE,
2014, pp. 168–173.
[10] ——, “Adapting transmitter power and modulation format to improve optical network performance utilizing the gaussian
noise model of nonlinear impairments,” Journal of Lightwave Technology, vol. 32, no. 21, pp. 3485–3494, 2014.
[11] ——, “Routing, modulation, spectrum and launch power assignment to maximize the traffic throughput of a nonlinear
optical mesh network,” Photonic Network Communications, vol. 29, no. 3, pp. 244–256, 2015.
[12] A. Fallahpour, H. Beyranvand, and J. A. Salehi, “Energy-efficient manycast routing and spectrum assignment in elastic
optical networks for cloud computing environment,” Journal of Lightwave Technology, vol. 33, no. 19, pp. 4008–4018,
2015.
DRAFT
May 22, 2017
27
[13] É. Archambault, N. Alloune, M. Furdek, Z. Xu, C. Tremblay, A. Muhammad, J. Chen, L. Wosinska, P. Littlewood, and
M. P. Bélanger, “Routing and spectrum assignment in elastic filterless optical networks,” IEEE/ACM Transactions on
Networking, vol. 24, no. 6, pp. 3578–3592, 2016.
[14] J. Zhao, H. Wymeersch, and E. Agrell, “Nonlinear impairment-aware static resource allocation in elastic optical networks,”
Journal of Lightwave Technology, vol. 33, no. 22, pp. 4554–4564, 2015.
[15] L. Yan, E. Agrell, H. Wymeersch, and M. Brandt-Pearce, “Resource allocation for flexible-grid optical networks with
nonlinear channel model [invited],” Journal of Optical Communications and Networking, vol. 7, no. 11, pp. B101–B108,
2015.
[16] L. Yan, E. Agrell, M. N. Dharmaweera, and H. WymeerschP, “Joint assignment of power, routing, and spectrum in static
flexible-grid networks,” Journal of Lightwave Technology, no. 99, 2017.
[17] M. Hadi and M. R. Pakravan, “Resource allocation for elastic optical networks using geometric optimization,” arXiv
preprint arXiv:1705.01191, 2017.
[18] H. Khodakarami, B. S. G. Pillai, and W. Shieh, “Quality of service provisioning and energy minimized scheduling in
software defined flexible optical networks,” Journal of Optical Communications and Networking, vol. 8, no. 2, pp. 118–
128, 2016.
[19] H. Khodakarami, B. S. G. Pillai, B. Sedighi, and W. Shieh, “Flexible optical networks: An energy efficiency perspective,”
Journal of Lightwave Technology, vol. 32, no. 21, pp. 3356–3367, 2014.
[20] M. M. Hasan, F. Farahmand, J. P. Jue, and J. J. Rodrigues, “A study of energy-aware traffic grooming in optical networks:
Static and dynamic cases,” IEEE Systems Journal, vol. 7, no. 1, pp. 161–173, 2013.
[21] S. Boyd, S.-J. Kim, L. Vandenberghe, and A. Hassibi, “A tutorial on geometric programming,” Optimization and
engineering, vol. 8, no. 1, p. 67, 2007.
[22] S. Boyd and L. Vandenberghe, Convex optimization.
Cambridge university press, 2004.
[23] M. Chiang et al., “Geometric programming for communication systems,” Foundations and Trends R in Communications
and Information Theory, vol. 2, no. 1–2, pp. 1–154, 2005.
[24] M. Hadi and M. R. Pakravan, “Spectrum-convertible bvwxc placement in ofdm-based elastic optical networks,” IEEE
Photonics Journal, no. 99, pp. 1–7, 2017.
[25] A. Muhammad, G. Zervas, N. Amaya, D. Simeonidou, and R. Forchheimer, “Introducing flexible and synthetic optical
networking: Planning and operation based on network function programmable roadms,” Journal of Optical Communications
and Networking, vol. 6, no. 7, pp. 635–648, 2014.
[26] MathWorks Inc., “Matlab,” https://www.mathworks.com, 2017.
[27] J. Lofberg, “Yalmip: A toolbox for modeling and optimization in matlab,” in Computer Aided Control Systems Design,
2004 IEEE International Symposium on, 2005, pp. 284–289.
[28] M. Grant, S. Boyd, and Y. Ye, “Cvx: Matlab software for disciplined convex programming,” 2008.
May 22, 2017
DRAFT
28
[29] G. Gao, X. Chen, and W. Shieh, “Analytical expressions for nonlinear transmission performance of coherent optical ofdm
systems with frequency guard band,” Journal of Lightwave Technology, vol. 30, no. 15, pp. 2447–2454, 2012.
DRAFT
May 22, 2017
| 7cs.IT
|
Statistical physics of inference: Thresholds and algorithms
Lenka Zdeborová1,∗ , and Florent Krzakala2,∗
arXiv:1511.02476v5 [cond-mat.stat-mech] 22 Jan 2018
2
1
Institut de Physique Théorique,
CNRS, CEA, Université Paris-Saclay,
F-91191, Gif-sur-Yvette, France
Laboratoire de Physique Statistique, CNRS, PSL Universités
Ecole Normale Supérieure. Sorbonne Universités
Université Pierre & Marie Curie, 75005, Paris, France.
∗
[email protected] and [email protected]
Many questions of fundamental interest in today’s science can be formulated as inference problems:
Some partial, or noisy, observations are performed over a set of variables and the goal is to recover,
or infer, the values of the variables based on the indirect information contained in the measurements.
For such problems, the central scientific questions are: Under what conditions is the information
contained in the measurements sufficient for a satisfactory inference to be possible? What are the
most efficient algorithms for this task? A growing body of work has shown that often we can
understand and locate these fundamental barriers by thinking of them as phase transitions in the
sense of statistical physics. Moreover, it turned out that we can use the gained physical insight
to develop new promising algorithms. The connection between inference and statistical physics
is currently witnessing an impressive renaissance and we review here the current state-of-the-art,
with a pedagogical focus on the Ising model which, formulated as an inference problem, we call
the planted spin glass. In terms of applications we review two classes of problems: (i) inference of
clusters on graphs and networks, with community detection as a special case and (ii) estimating a
signal from its noisy linear measurements, with compressed sensing as a case of sparse estimation.
Our goal is to provide a pedagogical review for researchers in physics and other fields interested in
this fascinating topic.
2
CONTENTS
A. Organization of the manuscript
B. What is statistical inference?
1. Terminology and simple examples
2. Teacher-student scenario
3. Bayes optimal versus mismatched
4. Estimators
5. The high-dimensional limit
C. Statistical physics and inference
1. Basic dictionary
2. The teacher-student scenario defines the planted ensemble
3. Inference via Monte Carlo and variational mean-field
4. Useful statistical physics concepts
5
6
6
8
10
10
12
12
13
13
14
15
II. Planted spin glass as a paradigm of statistical inference
A. Definition of the planted spin glass
B. Nishimori’s mapping via gauge transformation
C. Bayes-optimality on the Nishimori line
D. Parameter mismatch and learning
E. Relations between planted, quenched and annealed disorders
1. The quenched and annealed average
2. Properties of the planted ensemble
3. Quiet planting
F. Replica symmetry breaking
G. No replica symmetry breaking in Bayes-optimal inference.
H. Applications of planting for studies of glasses
1. Equilibration for free in glasses
2. Following states and analyzing slow annealing
3. Melting and glass transition
18
18
20
21
22
23
23
24
25
25
26
27
27
28
29
III. Phase diagram from the cavity method
A. Random graphs
B. Belief propagation and the cavity method
C. Properties of the randomly-quenched ensemble
D. Phase diagram of the planted spin glass
E. Systems with first order phase transitions
F. Comments on the phase transitions in various problems
30
30
31
33
35
36
38
IV. From physics insight to new algorithmic ideas
A. Non-backtracking spectral method
B. On computational hardness of the hard phase
C. Spatial coupling
D. The spatially coupled Curie-Weiss model
E. A brief history of time indices in the TAP equations
40
40
41
42
43
44
V. Clustering of networks and community detection
A. Biased overview of community detection
B. Asymptotically exact analysis of the SBM
C. Examples of phase transitions in the SBM
D. Spectral redemption for clustering sparse networks
E. More on the non-backtracking operator
F. The deceptiveness of the variational Bayes method
G. Real networks are not generated by the SBM
47
47
48
50
52
54
55
55
VI. Linear estimation and compressed sensing
A. Some examples of linear estimation
B. Compressed sensing
57
57
58
3
C. Approximate message passing
1. Relaxed Belief propagation
2. From r-BP to G-AMP
3. The potential
4. Note on the convergence and parameter learning
5. Examples of priors and outputs
D. Replica free energy, state evolution and phase transitions
1. State Evolution
2. Free energy
3. Phase transitions
E. Spatial coupling
F. Non-random matrices and non-separable priors
1. Structured priors
2. Orthogonal matrices
3. Reconstruction in discrete tomography
G. An application in physics: Optics in complex media
H. Beyond linear estimation: matrix factorization
1. Learning random sparse dictionaries
2. Low rank decomposition
VII. Perspectives
59
60
62
63
64
64
65
65
67
67
68
69
69
70
70
71
71
72
72
74
Acknowledgements
76
Fundings
76
References
76
4
I.
INTRODUCTION
Our goal in this review is to describe recent developments in a rapidly evolving field at the interface between
statistical inference and statistical physics. Statistical physics was developed to derive macroscopic properties of
material from microscopic physical laws. Inference aims to discover structure in data. On the first look, these goals
seem quite different. Yet, it was the very same man, Pierre Simon, Marquis de Laplace (1749–1827), who did one of
the first sophisticated derivations of the gas laws within the caloric theory, and who also created the field of statistical
inference [208]. This suggests there may be a link after all. In fact, the methods and tools of statistical physics are
designed to describe large assemblies of small elements - such as atoms or molecules. These atoms can quite readily
be replaced by other elementary constituents - bits, nodes, agents, neurons, data points. This simple fact, and the
richness and the power of the methods invented by physicists over the last century, is at the roots of the connection,
and the source of many fascinating work in the last decades.
There is a number of books discussing this interdisciplinary connection. Without claim of exhaustivity, see for
instance [86, 111, 181, 217, 238]. Our focus here is to present some more recent contributions that were obtained
using methodology that originated in the field of spin glasses [214]. In particular the replica and cavity methods,
with the related message passing algorithms, are at the roots of our toolbox. We focus on the analogy between phase
transitions in thermodynamics, and the threshold phenomena that arise in the ability to infer something useful from
data. The most interesting outcome of these considerations are deep consequences for the possibility of development
of new algorithms, and their relations to computational feasibility. This focus is summarized in the title by the words:
“threshold and algorithms”.
As we shall see, the natural setup in which the connection between inference and the statistical physics of disordered
systems is particularly transparent is the so-called teacher-student scenario. In a nutshell, the teacher-student scenario
can be described as follows: The teacher uses some ground truth information and a probabilistic model to generate
data that he then hands to the student
who is supposed to recover the ground truth as well as he can only from the knowledge of the data and the model.
This formulation of inference problems is particularly popular in neural networks [86, 281, 310], where the teacher
uses a rule to generate output, and the student is then trying to learn the rule to reproduce the same output. The
advantage of this setting is that it is easily amenable to the Bayesian approach and strategies that recover the ground
truth in an optimal way can be analyzed.
Let us make several comments: First we do not make this Bayesian choice because we think the Bayesian approach
is more correct than other lines of thought common in statistics. We do it because it is in this setting that the
connection between statistical inference and statistical physics becomes the most transparent. Our main point is to
explain and illustrate the insight that can be obtained from this connection in terms of phase transitions. In a sense
the strategy followed in this review is closer to the approach in information theory where we know the properties
of the noise that corrupted the signal than to traditional approach in statistics where we aim to assume as little as
possible.
Second, we focus on the teacher-student scenario and consequently we do not analyze problems where one is given
data or observations and has absolutely no idea of the underlying system that generated them. Nor do we discuss
model selection where one needs to judge from data which model is more suitable. There is a range of textbooks on
statistics where such problems are studied [13, 122, 198]. In theoretical physics it is common to study abstract models
that display interesting behavior and use gained knowledge to generalize about what is happening in the real world.
Third, while the subject is certainly algorithmic in nature, we shall concern ourselves primarily, as the title shows,
with phase transitions and threshold phenomena, where the connection with statistical physics is clear. This means
that while we certainly are interested in analyzing and discussing algorithms, and ask if it is possible for them to
reach the phase transition boundaries, we do not review, for any given problem, the vast amount of literature used to
actually solve it in practice. Our goal, instead, is (in a typical statistical physics attitude) to reach some “universal”
understanding of the problem.
The typical reader we have in mind has some knowledge of statistical physics, knows e.g. about the Ising model,
phase transition and free energy. We have in mind a reader who is curious to know how methods and concepts from
statistical physics such as phase transitions translate into the context of machine learning and inference. Perhaps the
most important message we want to pass in this review is that many calculations that physics uses to understand the
world can be translated into algorithms that machine learning uses to understand data. We also have in mind readers
from statistics and computer science who might want to learn more about the statistical physics approach. While we
do explain some basic physics concepts, such readers could further read a short introduction to statistical physics, as
for instance the appendix B (“some physics”) in the book of David McKay [198], or the Chap II in [217].
5
A.
Organization of the manuscript
The present review is organized as follows.
In section I B we introduce the basic concepts of statistical inference and the terminology. Most of this section is
standard knowledge in statistics and the main purpose of this section is to set the notations and present the concepts
to readers from physics who are not particularly familiar with probability and statistics. The only part of sec. I B that
is perhaps not standard, but certainly crucial for the present discussion, is the idealized teacher-student setting of
Sec. I B 2. Traditional statistics would typically not assume knowledge of the model that generated the data and large
part of statistics is precisely concerned with determining a good model. The main focus of this review is, however, on
the connections to statistical physics. In the teacher-student setting, the analysis of exact phase diagrams and phase
transitions is possible and this is our primary goal.
In Sec. I C we present some basic concepts of statistical physics that non-physics readers might not be familiar
with. The emphasis is given on formulation of the teacher-student setting as a planted statistical ensemble – this
planted ensemble is not standard in physics but we do believe it is a really fundamental concept, that has become
popular in the last decades, and that introducing it greatly elucidates the deep connection between properties of
high-dimensional inference problems, optimization problem, and disordered systems.
Having introduced the basic concepts we then choose a bottom-up approach and present in sec. II a prototypical
inference problem that we call the planted spin glass. We discuss how the physics of this inference problem can
be deduced using the existing knowledge from the physics of spin glasses. The main highlight of this section is that
Bayes-optimal inference can be translated into thermodynamics on the Nishimori line in the underlying planted model.
The physical properties of the Nishimori line are well understood for a wide range of models, this connection hence
provides a unifying picture for a large class of inference problems. Consequently, concepts discussed in section II
bring a lot of insight into many inference problems other than the planted spin glass. We conclude section II with a
discussion on how the concept of planting turns out to be innovative in study of glasses in sec. II H. All the results of
this section are presented in a way that should be understandable without going into any calculations. The purpose
of this is twofold: Readers who are not familiar with the replica and the cavity calculations can still understand
the reasoning and the main results. Readers who are familiar with the calculations can appreciate the reasoning,
connections and consequences without having to separate them from the technicalities.
Section III, however, is a more analytical chapter, that should be read with a pencil so that one can redo the
computations. It summarizes the basic analytical technique based on belief propagation and the cavity method that
we can use to analyse the properties and phase diagrams of high-dimensional inference problems. We do not go into
many details and the related methods are presented in greater depths in other existing literature and textbooks,
e.g. [217]. Readers familiar with these techniques can skip this section, except perhaps for its last part III F where we
present the zoology of the main phase transitions that are encountered in inference problems.
Armed with the statistical physics’ understanding of inference from sections II, III, we discuss related recent
algorithmic ideas in Sec. IV. The various algorithmic ideas are discussed again on the example of the planted spin
glass model or its simple variants. This section can be read separately from the previous ones by readers interested
mainly in the algorithmic consequences stemming from physics.
The planted spin glass problem is a simple example. To illustrate how to generalize from it to other, more practically
relevant, inference problems we present in section V the problem of clustering of networks with its recent development,
and in section VI the generalized linear estimation problem with the compressed sensing as its special case. The two
examples of applications in sections V, VI are chosen with respect to author’s own work. Each of these last two
sections can be seen as a concise introduction into these practically relevant problems and is therefore mainly for
readers that are interested in various applications of physics concepts or who want to specifically start learning about
clustering or compressed sensing.
Finally, a word of caution about the subjects that we have chosen to cover (and more importantly not to cover).
Historically, one of the most prominent example of statistical inference studied in physics are the error correcting codes.
Their studies in physics were pioneered by Sourlas [290] and the results are reviewed nicely e.g. in [217, 225, 238].
We do not discuss error correction in this review, except for a couple of interesting and/or historical connections.
Another very prominent problem that can be formulated as statistical inference and that has been extensively
studied by tools of statistical physics are neural networks. The book [238] covers very nicely the early results. A very
detailed presentation, with an emphasis on teacher-student protocols, can be found in the book [86]. The reviews
[281, 310] are also very good references. While we do not focus on this application that is rather well documented,
we will see in section VI that the problem of perceptron is a special case of the generalized linear estimation setting.
The two applications, network clustering and compressed sensing, that we chose to present in this review have
beed studied only more recently (in the last decade) and there are interesting results stemming from the connection
between inference and physics that are not (yet) well covered in other reviews.
6
B.
What is statistical inference?
Inference is the act or process of deriving logical conclusions from premises known or assumed to be true. Statistical
inference is the process of deducing properties of an underlying distribution by analysis of data. A way to think about
inference problems, that is particularly relevant in the context of today’s data deluge (often referred to as big data), are
procedures to extract useful information from large amounts of data. As the boundaries between fields are breaking,
nowadays inference problems appear literally everywhere, in machine learning, artificial intelligence, signal processing,
quantitative biology, medicine, neuroscience and many others.
A more formal way to set an inference problem is the following: One considers a set of variables x = {xi }i=1,...,N
on which one is able to perform some partial observations or measurements y = {yµ }µ=1,...,M . The goal is to deduce,
or infer, the values of the variables x based on the perhaps indirect and noisy information contained in the observed
data y. To give a couple of examples, the variables may correspond for instance to image pixels, an acoustic signal
or to an assignment of data points into clusters, while the corresponding observations would correspond for instance
to a blurred photograph, a recording of a song performed on a platform while the train was arriving, or positions of
points in space.
For inference problems the two main scientific questions of interest are:
• The question of sufficient information – Under what conditions is the information contained in the observations
sufficient for satisfactory recovery of the variables?
• The question of computational efficiency – Can the inference be done in an algorithmically efficient way? What
are the optimal algorithms for this task?
These two questions are of very different nature. The first one is a question that belongs to the realm of statistics
and information theory. The second one, instead, is in the realm of computer science and computational complexity.
Before the era of big data, statistics was often not concerned with computational tractability, because the number of
parameters to be estimated was usually small. Nowadays the tradeoff between statistical and computational efficiency
has changed, making it most reasonable to study the two above questions together.
We will see that these two questions have a very clear interpretation in terms of phase transitions in physics. More
specifically, the first one is connected to a genuine phase transition while the second one is connected to the type
of the transition (first or second order), and metastability associated with first order phase transitions is related to
algorithmic hardness.
1.
Terminology and simple examples
In this review, we will follow the approach of Bayesian inference. Developed by Bayes and Laplace, it concentrates
on conditional probabilities. The probability of having an event A conditioned on the event B is denoted P (A|B) and
defined as
P (A|B) =
P (A, B)
,
P (A)
(1)
P
where P (A, B) is the joint probability of both events A and B, and P (A) = B P (A, B) is the probability of event A
(the sum is taken over all the possible realizations of event B). When events A, B are taken from a continuous domain
the P (A) denotes probability distribution function, whereas P (A)dA denotes the probability of observing an outcome
in the interval (A, A + dA). For discrete variables the probability of one event, and the probability distribution are
interchangeable. Note that in this review we discuss both models with discrete variables (sec. II or V) and with
continuous variables (sec. VI), the concepts that we build hold for both these cases.
The central point of the approach is the so-called Bayes formula that follows directly from the definition of the
conditional probability (1)
P (x|y) =
P (y|x)
P (x) .
P (y)
(2)
There is a precise terminology for all the terms entering this equation, so we shall start by discussing the Bayesian
inference vernacular:
• P (x|y) is called the posterior probability distribution for x given all the data y.
• P (y|x) is called the likelihood.
7
• P (y), also denoted as a normalisation Z(y), is called the evidence.
• P (x) is called the prior probability distribution for x. It reflects our belief of what x should be before we
received any data.
Example of decaying particles. Our first example is borrowed from Ref. [198]. Unstable particles are emitted
from a source and decay at a distance y from the source. Unfortunately, we cannot build a perfect infinite length
detector, and decay events can only be observed if they occur in a window extending from distances ymin = 1 cm to
ymax = 20 cm from the source.
We run the experiment for some time, and we observe M decays at locations y = {y1 , ..., yM }. We also know that
we should expect an exponential probability distribution with characteristic length x for the event. Writing things in
the inference language, y are the observations and x is the unknown signal (of dimension N = 1). The question we
would like to answer is: Given the observations, how to estimate x?
In this particular example physics tells us exactly what is the probability distribution of a particle decaying at
distance y given the characteristic length x. In terms of Bayesian probability this means that we know exactly the
model that created the observations. This is of course not true in general, but when it is true we can set up the
so-called Bayes-optimal inference: The likelihood for one observation is
(
1
e−y/x if 1 < y < 20 ,
(3)
P (y|x) = xM(x)
0
otherwise ,
R 20
with a normalisation M(x) = 1 dz x1 e−z/x = e−1/x − e−20/x . The observations are independent, and therefore in
this case their joint probability distribution is simply a product of the single observation distributions. From eq. (3)
we can now extract P (x|y) using the Bayes’ formula and get
P (x|y) =
P (x)
e−
M
Z(y) [xM(x)]
PM
i=1
yi /x
.
(4)
Looking at this equation, we see that we have P (x) on the right hand side. If we do not have any additional information
on x, we assume than the prior distribution P (x) is a constant that just enters in the x-independent normalization,
denoting then Z̃(y) = Z(y)/P (x), we obtain
P (x|y) =
1
M
Z̃(y) [xM(x)]
e−
PM
i=1
yi /x
.
(5)
This is the final answer from Bayesian statistics. It contains all the information that we have on x in this approach.
For a dataset consisting of several points, for instance the six points y = 1.5, 2, 3, 4, 5, 12 cm, one can compute the
most probable value of x to be around x̂ = 3.7 cm. Note that this is considerably different from the mere average of
the points which is y = 4.6 cm. The reader can convince herself (for instance performing simulation and comparing
to the true correct value) that the presented approach gives a way better result that the naive average.
Example of linear regression. Most physicists are used to fit a theoretical model to experimental data. Consider
for instance that we measure the height of an object falling due to Newton’s gravity, we assume no initial speed and
no friction. We expect that the height h(t) of the object at time t will follow h(t, g, h0 ) = h0 − gt2 /2, where g is the
acceleration of gravity. We can do an experiment and measure heights of the object yi for many values of ti , with
i = 1, . . . , M . We expect, if we plot yi against t2i /2, to get a straight line that will give us the value of the acceleration
g.
Due to small but real measurement errors, the measured points will not lie exactly on a line. Used to such a
situation, a physicist would in this case perform a linear regression to minimize the so-called sum of squares of errors.
The “best” value of g, or to use the statistics language the estimator ĝ of g, would simply be the one that minimizes
PM
the cost function i=1 [h(ti , g, h0 ) − yi ]2 .
In order to get more familiarity with the concepts of inference we discuss how this familiar procedure can be cast
into the Bayesian language. In general we may not know much about the nature of the measurement noise. In such
cases it is often reasonable to expect that the noise is a Gaussian random variable of zero mean and variance ∆. Under
this assumption the likelihood P (yi |g, ∆) of a single data point is a Gaussian with mean h(ti , g, h0 ) and variance ∆].
We assume that the noise for different measurements is independent which leads to the total likelihood
P (y|g, h0 , ∆) =
M
Y
1
(2π∆)
M
2
i=1
e−
[h(ti ,g,h0 )−yi ]2
2∆
(6)
8
As in the previous example we assume a uniform prior distribution on the acceleration constant P (g), initial height
P (h0 ) and the variance P (∆), the posterior probability distribution is then given by
P (∆, g, h0 |y) =
M
Y
1
Z(y)(2π∆)
M
2
e−
[h(ti ,g,h0 )−yi ]2
2∆
(7)
i=1
PM
Looking for the most probable value of g indeed amounts to minimizing the sum of squares
i=1 [h(ti , g, h0 ) −
yi ]2 , which is the standard solution. Note, however, that the assumption of Gaussian noise independent for every
measurement might not be justified in general. Also, depending on the precise setting, there might be better prior
distributions for the parameters g, h0 , ∆. For more advanced discussion of classical linear regression see for instance
[309]. In section VI of this review we will discuss a high-dimensional version of linear regression, when the number of
parameters to be estimated is comparable to the number of observations M .
Let us make some important comments on the above considerations, which have been at the roots of the Bayesian
approach to inference from the very start. First, we notice that probabilities are used here to quantify degrees of
belief. To avoid possible confusions, it must thus be emphasized that the true x in the real world is not a random
variable, and the fact that Bayesians use a probability distribution P (x|y) does not mean that they think of the
world as stochastically changing its nature between the states described by the different hypotheses. The notation of
probabilities is used here to represent the beliefs about the mutually exclusive hypotheses (here, values of x), of which
only one is actually true. The fact that probabilities can denote degrees of belief is at the heart of Bayesian inference.
The reader interested in these considerations is referred to the very influential work of Bruno de Finetti [70].
Another subtle point is how to choose the prior distribution P (x). In fact, discussions around this point have been
rather controversial in the beginning of the history of statistical inference, and even now there are many schools of
thought. It is not the goal of this review to present these discussions and debates. In the spirit of a reductionism
that is usually done in physics, we introduce in the next section the so-called teacher-student scenario in which we
know what the model and the prior distribution really are, so that the use of the Bayes formula is straightforward
and fully justified. Majority of this review is then set within this teacher-student framework. This can be seen as a
simplified model of the generic situation. As we will see the picture of what is happening is interestingly rich already
in this framework. A first step towards the generic situation of unknown models that generated the data and unknown
prior probabilities is the case when the teacher hides some parameters of the model or/and of the prior distribution
from the student. Analyzing this case provides insight into the generic case, but still within the realm of a solvable
simplified teacher-student framework.
In sec. I C 2 we then argue that the teacher-student scenario for a generic inference problem can be seen as a
so-called planted ensemble of the corresponding statistical physics model. This is a known but not widely used way
of presenting inference in statistical physics, but the authors feel that it is an extremely useful and insightful way. It
allows to extend the intuition gathered over several decades in studies of disordered systems such as glasses and spin
glasses into many currently studied inference problems that otherwise appear highly non-trivial. One of the goals
of the present review is to bring this insight forward and establish the way of thinking about inference problems as
planted statistical ensembles.
2.
Teacher-student scenario
The teacher-student scenario is defined as follows:
• Teacher: In a first step the teacher generates a realization of variables/parameters x∗ from some probability
distribution Ptp (x), where ’tp’ stands for teacher’s prior. In a second step he uses these ground truth values
x∗ to generate the data y from some statistical model characterized by a likelihood of y given x∗ , denoted as
Ptm (y|x∗ ), where ’tm’ stands for teacher’s model. Finally, the teacher hands the data y to the student together
with some information about the distributions Ptp (x) and Ptm (y|x).
• Student: The goal of the student is to infer as precisely as possible (and tractably, when computational complexity is included into considerations) the original values x∗ of the variables from the information provided by
the teacher, i.e. from the data y and the available information about the distributions Ptp (x) and Ptm (y|x).
The variables with a ∗ (e.g. x∗ ) denote the ground truth that was used by the teacher, but that is not available to
the student. In the present manuscript we will consider cases where both the x and y are high-dimensional vectors,
and where the teacher’s prior Ptp (x) and model Ptm (y|x) are parameterized by some scalar (or low-dimensional vector)
values θ∗ and ∆∗ , respectively. This is a case relevant to many real-world problems where we assume existence of many
9
latent variables (collected into the vector x), and we assume that the model and prior distribution is parametrized by
a small number of unknown parameters.
We will in particular be interested in inference problems where the teacher’s prior distribution is separable, i.e. can
be written as a product over components
Ptp (x) =
N
Y
Ptp (xi ) .
(8)
i=1
Further, in this review, we are interested in cases where also the observations {yµ }µ=1,...,M are independent and each
of them depends on a subset of variables denoted ∂µ. The likelihood of the teacher’s model is then written as
Ptm (y|x) =
M
Y
µ=1
Ptm (yµ |{xi }i∈∂µ ) .
(9)
Given the information available to the student, she then follows the strategy of Bayesian statistics. The prior
information she has from the teacher is denoted as P (x), where x are the variables to be inferred. The statistical
model is represented by its likelihood P (y|x). Again we assume that both these can be written in a product form as
above. She then considers Bayes’ theorem (2) and writes the posterior probability which contains all the information
available to her
P (x|y) =
N
M
Y
1 Y
P (xi )
P (yµ |{xi }i∈∂µ ) .
Z(y) i=1
µ=1
(10)
A very useful quantity to define that derives directly from the posterior distribution is the marginal probability of
one variable
Z
Y
µi (xi ) = P (x|y)
dxj .
(11)
j6=i
Examples:
• Let us set the problem of simple linear regression that we used as an example in section I B 1 in the teacherstudent setting. Consider the teacher generated a number g ∗ from some probability distribution Ptp (g), this
g ∗ is the ground-truth value we aim to find back as precisely as possible from the noisy measurements. Then
the teacher generated heights yi independently from a Gaussian probability distribution of mean −g ∗ t2i /2 and
variance ∆∗ , where ti are the known measurement times, i = 1, . . . , N . He then handed the heights y and the
distribution Ptp (g), Ptm (yi |g) to the student. Set in this way the Bayesian solution presented in sec. I B 1 has
no ambiguity in terms of assuming the noise was Gaussian and independent, nor in terms of the choice of the
prior distribution for g.
• Historically the teacher-student setting was introduced for the problem of perceptron, called model B in [103].
In perceptron the aim is to learn weights Ji , i = 1, . . . , N in such a way that the scalar product with each of P
patterns ξiµ , µ = 1, . . . , P is constraint to be either smaller (yµ = −1) of larger (yµ = 1) than some threshold κ.
Mathematically stated one requires for all µ = 1, . . . , M
N
X
yµ = sign(
Ji ξiµ − κ) .
(12)
i=1
This perceptron setting models classification of patterns into two classes. One way to set a solvable version
of the model is to consider both ξiµ and yµ to be independent identically distributed (iid) random variables.
The goal is then to find a set of weights Ji such that the constraints (12) are satisfied. Another way to set
the problem that corresponds to the teacher-student scenario is to consider some teacher weights Ji∗ , generated
from some probability distribution, e.g. ±1 uniformly at random. The teacher perceptron then generates the
PN
output as yµ = sign( i=1 Ji∗ ξiµ − κ), where ξiµ are some known iid patterns. The goal is then to recover the
teacher weights Ji∗ from the knowledge of ξiµ and yµ . A very comprehensive presentation and solution of the
perceptron that uses this teacher-student terminology appears in [238].
10
3.
Bayes optimal versus mismatched
Throughout this manuscript we will be distinguishing two main cases depending on what information about the
distributions Ptp (x) and Ptm (y|x) the teacher gave to the student
• Bayes optimal case: In this case the teacher hands to the student the full and correct form of both the prior
distribution Ptp (x) and the likelihood Ptm (y|x).
• Mismatching prior and/or model: In this case the teacher hands no or only partial information about the prior
distribution Ptp (x) and/or the statistical model Ptm (y|x). In the present manuscript we will assume that the
teacher handed the correct functional form of the prior distribution and of the statistical model, but not the
values of the parameters θ and ∆, respectively.
Let us now consider the following exercise. Take x∗ to be the ground truth values of the variables generated by the
teacher and take x, x1 , x2 to be three independent samples from the posterior probability distribution P (x|y). We
then consider some function f (a, b) of two configurations of the variables a, b. Consider the following two expectations
Z
E [f (x1 , x2 )] = f (x1 , x2 )P (y)P (x1 |y)P (x2 |y)dx1 dx2 dy ,
(13)
Z
Z
E [f (x∗ , x)] = f (x∗ , x)P (x∗ , x)dx dx∗ = f (x∗ , x)P (x∗ , x, y)dx dx∗ dy
Z
= f (x∗ , x)P (x|y, x∗ )Ptm (y|x∗ )Ptp (x∗ )dx dx∗ dy .
(14)
where we used the Bayes formula. We further observe that P (x|y, x∗ ) = P (x|y) because x is independent of x∗ when
conditioned on y. Remarkably, in the Bayes optimal case, i.e. when P (x) = Ptp (x) and P (y|x) = Ptm (y|x), we then
obtain
E [f (x1 , x2 )] = E [f (x∗ , x)] ,
Bayes optimal :
(15)
meaning that under expectations there is no statistical difference between the ground truth assignment of variables
x∗ and an assignment sampled uniformly at random from the posterior probability distribution. This is a simple yet
immensely important property that will lead to numerous simplifications in the Bayes optimal case and it will be used
in several places of this manuscript, mostly under the name Nishimori condition.
In the case of mismatching prior or model the equality (15) typically does not hold.
4.
Estimators
What is the optimal estimator x̂ for the variables x? The answer naturally depends on what the quantity we aim
to optimize is. The most commonly considered estimators are:
Maximum a posteriori (MAP) estimator. MAP simply maximizes the posterior distribution and is given by
x̂MAP = argmax P (x|y) .
(16)
x
The main disadvantage of the MAP estimator is the lack of confidence interval and related overfitting issues.
Minimum mean squared error (MMSE). Ideally we would like to minimize the squared error between x̂ and x∗
SE(x̂, x∗ ) =
N
1 X
(x̂i − x∗i )2 .
N i=1
(17)
In general, however, we do not know the ground truth values of the variables x∗ . The best we can do within Bayesian
statistics is to assume that x∗ is distributed according to the posterior probability distribution. We then want to find
an estimator x̂ that minimizes the squared error averaged over the posterior
1
MSE(x̂) =
N
Z
P (x|y)
N
X
(x̂i − xi )2 dx .
i=1
(18)
11
By a simple derivative with respect to x̂i we see that the minimum of this quadratic function is achieved when
x̂MMSE
=
i
Z
Z
P (x|y)xi dx =
µi (xi ) xi dxi ,
(19)
where the right hand side is actually the mean of the marginal on the i-th variable defined by eq. (11).
The value of the MMSE is then computed as the MSE (18) evaluated at x̂i (19).
Maximum mean overlap (MMO). In cases where the support of the variables is discrete and the variables represent
indices of types rather than continuous values, the mean-squared error is not very meaningful. In that case we would
rather count how many of the N positions obtained the correct type. We define the overlap O as the fraction of
variables for which the estimator x̂ agrees with the ground truth
O(x̂, x∗ ) =
1 X
δx∗i ,x̂i .
N i
(20)
As in the case of MMSE, when we do not have the knowledge of the ground truth, the best Bayesian estimate of the
overlap is its average over the posterior distribution
MO(x̂) =
1
N
Z
P (x|y)
X
δxi ,x̂i dx .
(21)
i
The mean overlap MO maximized over the estimator x̂ leads to the maximum mean overlap estimator
x̂MMO
= argmax µi (xi ) ,
i
(22)
xi
where µ(xi ) is the marginal probability of variable i having type xi , defined by eq. (11). We should note that the
MMO is not an estimator very commonly considered in statistics, but given that statistical physics often deals with
discrete variables, and anticipating the close relation, the MMO will turn out instrumental.
Let us compare the MAP estimator with the ones based on marginals. In most cases the MAP estimator is easier
to approach algorithmically (from the computational complexity point of view optimization is generically easier than
enumeration). Moreover, in many traditional settings, where the number of samples is much higher than the number of
variables, there is usually very little difference between the two since the marginal probabilities are basically peaked
around the value corresponding to the MAP estimator. For instance, let us consider the teacher-student scenario
with unknown parameters θ of the teacher’s prior, or ∆ of the teacher’s model. In situations where the number of
samples M is large, but the parameters θ and ∆ are scalar (or low-dimensional), the posterior probability of these
parameters is so closely concentrated around its expectation that there is no practical difference between the MAP
and the marginals-based estimators for θ and ∆. When it comes to estimating the variables x then in the setting of
high-dimensional statistics (see Sec. I B 5), where the number of samples M is comparable to the number of variables
N the MAP estimator for x will generically provide crude overfitting and we will prefer estimators based on marginals.
Another reason to favor marginalization based estimators is because they come directly with a notion of statistical
significance of the estimate. Yet another reason is that from the perspective of spin glass theory there might even be
computational advantages (ground states might be glassy, whereas Bayes-optimal computations are not, see sec II G).
Let us now investigate the MMSE (and the MMO) estimator in the Bayes optimal case in the view of equality (15).
In section I B 3 we concluded that in the Bayes optimal case (i.e. when the teacher handed to the student the precise
form of the prior distribution and of the model likelihood) and in expectation, a random sample x from the posterior
distribution can be replaced by the ground truth assignment x∗ without changing the values of the quantity under
consideration. It hence follows that
Bayes optimal :
MMSE = E[SE(x̂, x∗ )] .
(23)
In words, the MMSE computed solely using the posterior distribution is on average equal to the squared error between
the MMSE estimator and the ground truth value of the variables x∗ . This is a very nice property. An analogous
identity holds between the MMO and the average of the overlap between the MMO estimator and the ground truth
assignment
Bayes optimal :
MMO = E[O(x̂, x∗ )] .
(24)
12
5.
The high-dimensional limit
Let us denote by N the number of variables x = {xi }i=1,...,N (referred to as dimensionality in statistics), and by M
the number of observations (or samples) y = {yi }i=1,...,M . Traditionally, statistical theory would consider the case of
a finite number N of variables/parameters to estimate and a diverging number of samples M . Think for instance of
the two examples (decaying particles and linear regression) from sec. I B 1.
In today’s data deluge more than ever, it is crucial to extract as much information as possible from available data.
From the information theoretic perspective, we would expect that useful information is extractable even when the
number of variables (dimensionality) N is comparably large (or even somewhat larger) than the number of observations
(samples) M . In that case, separating the useful information about x from noise is much more challenging. This is
referred to as the high-dimensional statistical theory. It is, in particular, in this high-dimensional framework, where
the amount of data to analyze is huge, that we need to pay attention not only to solvability but also to computational
efficiency, keeping in mind both questions from section I B.
FIG. 1. Schematic representation of a typical high-dimensional inference problem.
Typically, we will consider the limit where α = M/N is fixed, whereas N → ∞. In this limit the following scenario,
illustrated schematically in Fig. 1, very often applies: For low values of α < αc , successful inference of the variables
is not possible for any algorithm, the corresponding observed information is simply insufficient. For high values of
α > αs , computationally efficient inference algorithms do exist. In an intermediate regime αc < α < αs , successful
inference is in principal possible but algorithmically considerably harder than above αs . In many settings, the values
αc and αs can be defined in a precise mathematical way as values at which some measure of performance (or its
first derivative) presents a discontinuity in the limit of large size. In such a case, we talk about thresholds or phase
transitions (see section I C 4 for a more precise meaning of the second). In many problems the intermediate hard
phase is missing and αc = αs . On the other hand, settings and examples where the hard phase exists (αc < αs ) are
theoretically very interesting and challenging many researchers to prove some formal results about this hardness. For
a general inference problem, the values of αc and αs are not known, and neither are efficient algorithms that would
work down to αs . The state-of-the-art results usually only give a lower bound on αc , and an algorithmic upper bound
on αs .
The objective of a large part of the statistical physicists interested in these directions and included in this manuscript
could be summarized as follows (see also the schema Fig. 1): Determine the threshold αc between the inferenceimpossible and inference-possible region, and the threshold αs between the inference-hard and inference-easy region.
And develop efficient algorithms that would be able to infer the variables successfully with as little information
contained in the observations as corresponds to αs .
C.
Statistical physics and inference
The connection between statistical physics and statistical inference has a long history. Arguably, the most influential
early works are due to E. T. Jaynes, starting with his Statistical mechanics and information theory in 1957 [141] where
he develops the maximum entropy principle, thus building upon the pioneering work of C. Shannon [283], who named
the amount of information as entropy inspired by mathematically analogous formulas in thermodynamics. Jaynes
actually suggests that statistical mechanics does not have to be regarded as a physical theory dependent on its
assumptions, but simply as a special case of a more general theory of Bayesian inference. An inspiring reference is a
collection of lecture notes Statistical Physics to Statistical Inference and Back from a meeting held in Cargèse in 1992
[111]. The goal of that meeting was precisely to revise the foundational connections between the two fields. This close
connection leads to a continued cross-fertilization between the two fields. Important methodological developments in
13
one field find numerous applications in the other, and a new applicational area in one field profits greatly from the
methodology more common to the other.
The present manuscript is mainly concerned with recent applications of methods from spin glasses and glasses
[214] to inference problems and associated algorithms. This direction was perhaps most influenced by the work on
simulated annealing by Kirkpatrick, Gelatt, and Vecchi [164]. Connections between spin glasses and error correcting
codes [290], or neural networks such as the perceptron [281, 310] belong among the early very influential works. This
line of research is relatively well explored and illustrated on a number of problems as seen from classical books such
as the one by H. Nishimori Statistical Physics of Spin Glasses and Information Processing An Introduction [238], or
more recently by M. Mézard and A. Montanari, Information, Physics, and Computation [217].
Data processing is, however, a very fast developing field with many emerging yet far reaching applications. On top
of that, there was an important development in the theory of spin glasses on sparse random graphs in the early 2000s
[211, 212]. This manuscript aims at exposing some of this recent development.
1.
Basic dictionary
Let us first establish a dictionary between statistical inference and statistical physics. The posterior probability
(10) can be written as
)
( M
N
X
X
1
1
P (x|y) =
log P (xi ) =
log P (yµ |{xi }i∈∂µ ) + β
exp β
exp −βH(x, y) .
Z(y)
Z(y)
µ=1
i=1
(25)
We have used the exponential form to note that it takes the form of the Boltzmann probability distribution on variables
(discrete or continuous) with Z being the partition function, log P (xi ) the local magnetic field, and log P (yµ |{xi }i∈∂µ )
being the interaction between variables, H(x, y) being the Hamiltonian. We also introduced the auxiliary parameter
β, that was previously β = 1, and plays the role of the inverse temperature. The whole exponent divided by the
inverse temperature is then minus the Hamiltonian H of the physical system.
Note that the distribution (25) encompasses a wide range of models, the variables x can be discrete (e.g. spins)
or continuous, the interactions can depend on few or many variables (e.g. in the planted spin glass of section II the
interactions are pair-wise, and in the linear estimation of sec. VI each of the interactions involves all of the variables).
From a physics point of view the marginal probabilities (11) are nothing else than local magnetizations obtained by
summing the Boltzmann distribution over all but one variable. The MMSE (or the MMO) estimator is then computed
using these local magnetizations (at β = 1). On the other hand the MAP estimator can be viewed as the ground
state of the physical system, i.e. the minimum of the Hamiltonian. Another way to think about the MAP estimator
is as the zero temperature limit, β → ∞, of MMSE estimator.
2.
The teacher-student scenario defines the planted ensemble
Looking at the Boltzmann distribution (25) it could represent quite a generic statistical physics model where the
observations y play the role of a quenched disorder. Keeping in mind that we want to study Bayesian inference in
the teacher-student scenario of sec. I B 2 we have to emphasize that the disorder y is generated in a very particular
way (by the teacher) related to the form of the Boltzmann distribution itself. This becomes particularly striking in
the Bayes-optimal setting in which the teacher handed to the student all the information about how he generated the
data y except the realization x∗ .
In the rest of this manuscript we call models created in the teacher-student scenario of sec. I B 2 the planted models.
We call a disorder y that results from the teacher-student scenario the planted disorder. We contrast it with the
most commonly considered case of randomly-quenched disorder where the component yµ are independent identically
distributed variables. In the planted disorder the components yµ are not independent, they are correlated because
they all depend on the ground-truth x∗ that the teacher used to generate y. We call x∗ the planted configuration.
This is the same difference between solving a random system of linear equations y = Fx (which will have no solution
if F is a M × N matrix and M > N ) and solving one where the left hand side was generated as y = Fx∗ , instead of
just being random. In the second case, we have planted a solution x∗ in the problem.
A generic strategy to define the planted statistical ensemble is to take an inference problem in the teacher-student
setting. The Bayes-optimal posterior probability distribution of that inference problem then corresponds to the
Boltzmann distribution of the planted model at temperature T = 1/β = 1. What makes the planted model particular
is the form of the quenched disorder that is not independently random but instead correlated in a very particular way.
14
Perturbing a random ensemble by hiding a solution in it is often a natural thing to do. For instance, protein folding
is often modeled, in statistical physics, by a random energy model (to take into account its glassiness) in which one
special configuration (the so-called native configuration) is artificially created with a very low energy [46]. Thinking
of Bayesian inference in terms of the planted ensemble of statistical physics models is implicitly done in a number of
works. For instance the pioneering articles of Sourlas on finite-temperature decoding in error correcting codes [290–
292] and the relation to the Nishimori line, summarized excellently in [238], were very influential. In mathematics,
planting was introduced by Jerrum [142] as a model to study hardness of finding a large clique planted in a graph, and
also by Jerrum and Sorkin [143] as a model to study graph bisection. Planting was used in many mathematical works
either as a useful model or as a proof technique. It is the use of planting in proofs for constraint satisfaction problems
[4] that brought the authors’ attention to this concept. The authors find that making the connection between planting
and inference very explicit brings a lot of insight into a generic inference problem and for this reason we build the
present review around this connection.
To clarify further what is exactly the planted ensemble we give several examples and counter-examples:
• Planted coloring. Graph coloring is a famous mathematical problem where nodes of a graph are assigned colors
with a constraint that two connected nodes should not have the same color. Coloring of random graphs is a
constraint satisfaction problem, well studied in statistical physics [234, 320]. In planted coloring, the teacher
first generates a random assignment of colors to nodes x∗ , and then he generates edges of the the graph at
random but in a way that two nodes of the same color are not connected. The edges of the graph play the role
of the planted disorder y. The graph (i.e. its set of edges) is then handed to the student, but not the planted
configuration x∗ . The role of the student is to find a configuration as correlated to the planted one as possible.
Physics-study of the planted coloring appeared in [182], and we will see that it is a special case of the stochastic
block model to which we devote section V.
• Planted perceptron. In sec. I B 2 we presented the perceptron as the problem for which the teacher-student
terminology was established. Formulated as a planted problem, the set of weights J∗ the teacher-perceptron
uses is the planted configuration. The patterns ξiµ are in the physics works considered as iid, they play the role
of the ordinary randomly-quenched disorder. The teacher generates the outputs yµ using the patterns and the
planted weights J∗ (12). The output y then plays the role of the planted disorder. The planted weights J∗ are
then hidden from the student and her goal is to recover them as precisely as possible. The associated posterior
probability distribution is the Boltzmann measure of the planted perceptron. On the other hand, when asking
how many random patterns a perceptron can classify (the so-called capacity [103]), one is studying a non-planted
problem.
• Hopfield model [135] is a well known model studied in statistical physics. The goal in the Hopfield model is to
construct an Ising spin model for which a certain pre-defined set of randomly generated patterns ξiµ ∈ {±1},
i = 1, . . . , N and µ = 1, . . . , M , serve as configurations attracting local dynamics. It is, in fact, a model of
associative memory where, when starting close to one of the patterns, the dynamics would converge back to it.
PM
The Ising model studied by Hopfield has Hebb-like interactions Jij = µ=1 ξiµ ξjµ .
We could study the inference problem for recovering the patterns ξiµ from the knowledge on the Jij . For a single
pattern, M = 1, the posterior distribution of this inference problem would indeed correspond to the Hopfield
model, i.e. Ising spin model with interactions Jij . However, for more than one pattern M > 1 the posterior
of the inference problem will be a function of N M variables, and will be very different from the Boltzmann
distribution of the Hopfield model (which is just a function of only N spins). This shows that inferring or
remembering are very different tasks. Clearly, the associative memory studied in the Hopfield model, where one
start the dynamics close to one of the patterns, is different from inference of patterns from the Jij . The Hopfield
model is not a planted model in the sense to which we adhere in this review.
Since this is important we reiterate that the planted model is a statistical physics model where the Boltzmann
distribution is at the same time the posterior probability distribution of the corresponding inference problem. There
are many models considered in the literature where the quenched disorder is created using some auxiliary variables,
but the Boltzmann distribution is not the posterior probability for inference of these auxiliary variables. Then we do
not talk about a planted model, an example is given above using the Hopfield model with more than one pattern.
3.
Inference via Monte Carlo and variational mean-field
When it comes to solving an inference problem, i.e. evaluating the corresponding MMO of MMSE estimators, see
sec. I B 4, there are many algorithmic methods studied and used in computer science, statistics and machine learning.
15
For nice and comprehensive textbooks see e.g. [198, 309]. Two of the classical inference algorithms are particularly
interesting in the present context because they originate in statistical physics. These are the Gibbs sampling of
Monte Carlo Markov chain (MCMC) based methods [104, 105, 235] and the variational mean-field inference methods
[146, 244, 251, 308]. Both these methods and their applications are well covered in the existing literature and in this
review we will not focus on them. We only describe in the present section their basic concepts and ideas and relation
to the message-passing-based algorithms on which we focus in the manuscript.
The development of Monte Carlo goes back to the well know paper of Metropolis et al. [209] and was widely used
since in physics and other scientific fields. The Metropolis-Hastings algorithm [125, 209], most commonly used in
physics, is based on the comparison of energy cost in one state of a system and a small perturbation of the state
(typically the value of one variable is changed). When the cost decreases the change is accepted, when the cost
increases the change is accepted with a probability exponentially proportional to the cost difference, p = e−β∆E ,
where β is a inverse-temperature-like parameter. Another perturbation is introduced and the procedure is repeated.
Gibbs sampling [104, 105], known as heat-bath in physics, is a variant on the Metropolis-Hastings algorithm where
given a fixed state of all but one variable, the probability distribution of the one variable is evaluated conditioning on
the fixed state of the other variables. The value of the variables is then sampled from this probability distribution.
The process is then repeated. Both Metropolis-Hastings and Gibbs sampling are designed to sample the Boltzmann
probability measure. But they can also be famously turned into a generic purpose optimization algorithms called
simulated annealing [164].
The advantage of Monte Carlo based methods it that if iterated for sufficiently long time they provide the exact
answer for a very generic and broad class of systems (only ergodicity and balance condition are required). The
disadvantage is that the time needed to obtain the exact result can be as large as exponential in the size of the system
and it is in general very hard to decide what is the running time necessary for satisfactory precision. Nevertheless,
Monte Carlo based methods are very useful in many cases, in particular when the domain of variables is discrete, and
throughout this manuscript we will be comparing to them.
Tracing the precise origins of the use of variational mean-field for inference from data is a little harder. Ideas related
to variational mean-field were used in statistical physics since Gibbs (together with Maxwell and Boltzmann) created
the field, and are certainly present in the works of Van der Waals [303] and Pierre Weiss [313] arround the 1900’s.
The modern view on variational mean-field for inference and learning is very nicely summarized in [146, 198].
The basic idea of variational inference is to approximate the posterior probability distribution P (x|y) (25) by
another probability distribution Q(x|y) such that (a) for Q(x|y) the Bayesian estimators (MMO or MMSE) can be
evaluated tractably, and (b) the Kullback-Leibler divergence
DKL (Q||P ) =
X
x
Q(x|y) log
Q(x|y)
P (x|y)
(26)
between P and Q is the smallest possible. The variational inference methods are very fast and in many settings and
practical problems they provide precise approximations. However, in some settings they provide crude approximations
and sometimes misleading results (see an example in sec. V F).
The present manuscript focuses on message passing inference algorithms, such as belief propagation (BP). In physics
the origins of these methods trace back to the Bethe approximation [34], the work of Peierls [250] and the ThoulessAnderson-Palmer equations [300]. Interestingly, these equations can be seen as “improved” variational mean-field
equations, and this is at the roots of many perturbation approaches (such as Plefka/Yedidia-Georges [107, 253] or
Kikuchi [163]).
In information theory belief propagation was invented by Gallager [100] and in Bayesian inference by Pearl [249]. In
an excellent article Yedidia, Freeman and Weiss [318] relate belief propagation and the variational mean field methods.
However, strictly speaking belief propagation is not a variational method, because the related Bethe free energy does
not have to be larger than the true free energy. Compared to variational mean field methods belief propagation is
as fast, and generally more precise. An important property on which we focus in this manuscript, is that on a class
of models, that can be called mean-field spin glasses, belief propagation method gives asymptotically exact results
(whereas variational mean field does not) and it is amenable to exact analysis that enables detailed understanding of
many systems of interest.
4.
Useful statistical physics concepts
Traditional physical systems are composed of a number of spins or molecules that is comparable to the Avogadro
number NA ≈ 1023 . It is hence almost always relevant to investigate the limit of large system size that is in statistical
physics referred to as the thermodynamic limit.
16
Self-averaging. In the analogy between inference and statistical physics of disordered systems, the observations y
play the role of the so-called quenched disorder in the Hamiltonian in eq. (25). Generically, quantities of interest, e.g.
the magnetization of variable number 10, depend on the realization of this quenched disorder.
Statistical physics, however, mostly focuses on quantities that have the so-called self-averaging property, i.e. in
the thermodynamic limit their value does not depend on the realization of the disorder y (but only on its statistical
properties). More formally, a quantity A(y) is self-averaging if for every > 0
lim Prob {|A(y) − E [A(y)] | > } = 0 .
N →∞
(27)
One of the most important quantities that we expect to be self-averaging is the free energy density
f (β, y, N ) = −
1
log Z(y)
Nβ
(28)
that is assumed to have a thermodynamic limit
f (β) = lim f (β, y, N )
N →∞
(29)
that does not depend on the realization of y.
Among the quantities discussed previously in this manuscript, the ones for which the self-averaging property is
particularly interesting are the MMSE and the MMO. Under self-averaging, the equalities (23-24) hold not only on
average, but also when the averages on the r.h.s. are removed. This means that the typical squared distance between
the ground truth and the MMSE (MMO) estimator is equal, in the thermodynamic limit, to the one evaluated without
the knowledge of the ground truth. Clearly, focusing on the analysis of self-averaging quantities greatly simplifies the
description of the behavior in the thermodynamic limit and enables a large part of the results obtained in statistical
physics.
It should be noted, however, that from a rigorous point of view, proving that the above quantities indeed are
self-averaging poses a non-trivial technical challenge. This is a recurring discrepancy between theoretical physics and
rigorous approaches, in physics we make assumptions that are correct and enable progress and understanding of the
problems, but that are not always easy to prove.
Phase transitions. Another concept borrowed from physics that will turn out very influential in the analysis of
high-dimensional inference are phase transitions. Intuitively, a phase transition is an abrupt change in behavior as a
parameter (commonly the temperature) is tuned. Since not everything that is termed phase transition in computer
science or engineering literature is a genuine phase transition from a physics point of view, it is useful to make some
reminders.
First of all, true phase transitions do not exist at finite system size. In mathematical physics a phase transition is
defined as a non-analyticity in the free energy density f (β) (29). Since the finite N free energy density is a logarithm
of a sum of exponentials, it is always an analytical function. Only when the limit N → ∞ is taken can the nonanalyticities appear. Therefore when we talk about a phase transition, we should always make sure that it indeed
corresponds to a non-analyticity and not to a smooth (even though perhaps very rapid) change.
Secondly, a traditional phase transition is always associated to either a critical slowing down or to metastability.
According to the Ehrenfest’s classification of phase transitions, there are two main types
• 1st order phase transitions are associated to a discontinuity in the first derivative of the free energy. The first
derivative of the free energy is usually related to some measurable quantity that consequently has a discontinuity.
Since a number of quantities, including e.g. the MMSE or the magnetization, are related to this first derivative,
they also display a discontinuity at a first order phase transition. The mechanism behind a 1st order phase
transition is a competition between two stable phases of the system. For the sake of vocabulary let us call the two
phases “paramagnet” and “ferromagnet”. Around the phase transition there is a region of phase coexistence
in which the properties of the system are ruled by properties of both the phases. In systems that cannot
be embedded into a (finite dimensional) Euclidean space we can define sharp boundaries of this parameterregion, called the spinodal points. The phase with lower free energy (i.e. higher log-likelihood) is said to be
thermodynamically stable, the other one is called metastable. Metastability has profound consequence on the
behavior of dynamics and algorithms. We will come back to this in detail.
• 2nd order phase transitions are associated to a discontinuity in the second derivative of the free energy. This type
of phase transition is associated to a so-called critical slowing down, meaning that dynamical processes become
increasingly slow next to the phase transition. Diverging timescales are closely linked to diverging length-scales
and diverging strength of fluctuations. These are described by the theory of criticality and critical exponents
that ruled statistical physics in the second half of last century.
17
Genuine phase transitions are quite peculiar creatures, and this is largely why they fascinate many physicists.
To make a distinction in vocabulary in physics we use cross-over for a more generic rapid change in behavior as a
parameter is tuned. We will keep the term phase transition or threshold for what is described above. Going back to
Figure 1, a physicist’s mind becomes very intrigued when for a given inference problem we find that, as the amount
of information in the measurements is increased, there exist genuine phase transitions. For a physicist, a phase
transition is something intrinsic, so we immediately expect a profound meaning and non-trivial consequences for the
system under investigation.
18
II.
PLANTED SPIN GLASS AS A PARADIGM OF STATISTICAL INFERENCE
As biology has its drosophila, or computational complexity its K-SAT problems, in statistical physics the canonical
example is the Ising model
H(S) = −
X
(ij)∈E
Jij Si Sj −
N
X
hi Si ,
(30)
i=1
where S = {Si }i=1,...,N and the spins are Si ∈ {−1, 1}, E is the set of interacting pairs, Jij is the interaction strength
between spins and hi is the magnetic field. Statistical physics is the study of the Boltzmann probability measure on
spin configurations
µ(S) =
1
e−βH(S) .
Z(β, J, h)
(31)
In this section we study an inference problem, that we call the planted spin glass, for which, in the teacher-student
setting of sec. I B 2, the posterior probability distribution is the Boltzmann distribution of the Ising model. The
purpose of this exercise is simply that we will be able to use a large part of the knowledge existing in the literature on
spin glasses to describe in detail the properties of this inference problem. Moreover, the physical picture that arises
in this section generalizes to other inference problems in the teacher-student setting. Once we understood in detail
the planted spin glass problem, we can basically deduce without calculation what will qualitatively happen in a wide
range of other inference problems. Whereas the problem presented in this section may appear as a very special and
relatively simple case, all the concepts that we illustrate on this example are very generic and apply to many other
highly non-trivial inference problems. Sections V and VI then give two recently studied examples.
The Ising model with the Hamiltonian (30) is presented in numerous textbooks. The most commonly considered
geometries of interactions are finite dimensional lattices, fully connected graphs, random graphs, or for instance social
networks. Interactions Jij are chosen such that Jij ≥ 0 in the ferromagnetic case (most commonly Jij = J for all
(ij) ∈ E) while in the antiferromagnetic case Jij ≤ 0. In the random field Ising model the fields hi are random. Perhaps
the most intriguing case of an Ising model is the spin glass, in which the interactions Jij are taken independently at
random from some distribution including both negative and positive numbers. In most of the situations considered
in physics literature, the interaction couplings J are chosen independently of any specific configuration S.
A.
Definition of the planted spin glass
To motivate the planted spin glass we consider the following problem: We have a (large) number N of people and
a pack of cards with the number +1 or −1 on each of them. There is about the same number of each. We deal one
card to every person. In the next step we choose pairs of people and ask them whether they have the same card or
not. Their instructions are to roll a dice and according to the result answer truthfully with probability ρ and falsely
with probability (1 − ρ). We collect the answers into a matrix Jij , Jij = 1 if the pair (ij) answered that they had the
same card, Jij = −1 if the answer was that they did not have the same card, and Jij = 0 if we did not ask this pair.
We want to model this system, and one of the questions we have is whether and how we can estimate which person
has which card.
Thinking back about the teacher-student scenario of sec. I B 2 the teacher distributes the cards, i.e. generates a
latent variable Si∗ = ±1, 1 ≤ i ≤ N for each of the nodes/people independently at random. We then select the graph
of observations. Let us assume that the pairs to whom we ask questions are chosen at random, and that we ask M of
them. The set of queried pairs E can then be represented by an Erdős-Rényi (ER) random graph [87, 88] on N nodes
having average degree c = 2M/N , as defined in section III A. Graphs created in this manner are at the core of many
results in this review. The teacher collects the answers into a matrix Jij = ±1, (ij) ∈ E taken from the following
distribution:
P (Jij |Si∗ , Sj∗ ) = ρδ(Jij − Si∗ Sj∗ ) + (1 − ρ)δ(Jij + Si∗ Sj∗ ) ,
(32)
where ρ is a bias of the dice (i.e. the probability of getting the correct answer), and δ(·) is the Dirac delta function.
The latent values Si∗ are then hidden from the student and the goal is to infer from the values of J as precisely as
possible which node had which card. Depending on the situation, the teacher did or did not reveal the value ρ he
used. For the moment we will assume ρ is known.
Let us look at the simplest cases: In the noiseless case, ρ = 1, each connected component of the graph has only
two possible configurations (related by global flip symmetry). When the graph is densely connected, and hence has
19
a giant component, one can therefore easily recover the two groups. When ρ = 1/2, it is as if we chose the couplings
independently at random just as in the most commonly considered spin glass model. The values J do not contain
any information about the latent variables Si∗ , and recovery is clearly impossible. What is happening in between, for
1/2 < ρ < 1? (The behavior for 0 < ρ < 1/2 is then deduced by symmetry.)
The above model is simple to state, but its solution is very interesting, and rather non-trivial. This problem
appeared in several different contexts in the literature, and is called censored block model in [2, 3, 274]. From a
rigorous point of view this problem is still a subject of current research. In the physics literature a closely related
model is known as the Mattis model [207]. Here, we will call the above problem the planted spin-glass model to
differentiate it from the standard spin-glass model, where we chose all the interactions J randomly and uncorrelated.
It should be noted, that “spin glass” here is used as a name for the model, it does not necessarily imply that this
model must exhibit glassy behavior. The phase diagram on the planted spin glass is discussed in detail in Sec. III D.
We stress that the main goal in inference is to retrieve the planted configuration, whereas the usual goal in physics of
spin glasses is to understand the glassy behavior of the model. These two goals are of course related in some aspects,
but also complementary in others.
More formally defined, to create an instance, that is a triple (S∗ , G, J), of the planted spin glass we proceed as
follows
• We consider that the hidden assignment Si∗ , called the planted configuration of spins, is chosen uniformly at
random from all the 2N different possibilities.
• We then choose a graph of interactions G = (V, E), but not yet the associated values of Jij , independently of
the planted configuration S∗ .
• For each edge (ij) ∈ E, the value of Jij is chosen to be either +1 or −1 with a probability taken from (32). The
signs of the couplings then carry information about the planted configuration.
So far the relation with the Ising model may not be very clear. To see it very explicitly we write eq. (32) differently.
Without losing generality, we can define a parameter β ∗ by
∗
eβ
ρ=
and
2 cosh β ∗
∗
e−β
1−ρ=
with
2 cosh β ∗
1
β = log
2
∗
ρ
1−ρ
,
(33)
so that eq. (32) reads
∗
P (Jij |Si∗ , Sj∗ ) =
∗
∗
eβ Jij Si Sj
.
2 cosh β ∗
(34)
Since we want to know the values of the spins S given the knowledge of the couplings J, we write the Bayes formula
Q
β ∗ Jij Si Sj
P (J|S)P (S)
(ij)∈E e
P (S|J) =
= N
P (J)
2 (2 cosh β ∗ )M P (J)
(35)
where we used (34). The posterior probability distribution has to be normalized, therefore
P (J) =
Z(β ∗ , J)
2N (2 cosh β ∗ )M
,
(36)
where Z(β ∗ , J) is the partition function of the Ising model (31) at inverse temperature β ∗ (in zero magnetic field). In
other words, the posterior distribution is simply the standard Boltzmann measure of the Ising model at a particular
inverse temperature β ∗ . In general we shall consider this Ising model at a generic inverse temperature β 6= β ∗ . This
is because a priori we do not know the proportion of errors. In the teacher-student setting, the teacher may not have
handed to the student the value β ∗ .
The special case β = β ∗ corresponds to the Bayes optimal inference as discussed in section I B 3. One of the most
physically important properties of the Bayes-optimal case is that the planted configuration can be exchanged (under
averages, or in the thermodynamic limit) for a random configuration taken uniformly from the posterior probability
measure, which is in physics simply called an equilibrium configuration. All equilibrium configurations with high
probability share their macroscopic properties, e.g. their energy, magnetization, various overlaps etc. This is directly
implied by the Nishimori condition (15) derived in Sec. I B 3 for the Bayes-optimal inference. This property will turn
out to be crucial in the present context.
20
As discussed in Sec. I B 4, since we are dealing with a problem with discrete variables (spins), it is desirable to
minimize the number of mismatches in the estimation of the planted assignment. We hence aim to evaluate the
maximum mean overlap (MMO) estimator (22) which in the case of an Ising model is simply
Ŝi = sign (mi ) ,
(37)
where mi is the equilibrium value of the magnetization of node i in the Ising model. We reduced the inference of the
planted configuration to the problem of computing the magnetizations in a disordered Ising model as a function of
temperature.
An off-the-shelf method for this that every physicist has in her handbag is a Monte-Carlo simulation, for instance
with the Metropolis algorithm. In Sec. III we shall see that for the mean-field setting (i.e. the underlying graph
being random or fully connected) there exists an exact solution of this problem via the cavity method. However,
before turning to the mean-field setting we want to discuss concepts that are more generic and we will hence keep
Monte-Carlo in our mind as a simple and universal method to evaluate phase diagrams of Hamiltonians with discrete
degrees of freedom.
B.
Nishimori’s mapping via gauge transformation
For the planted spin glass there is an alternative way that provides understanding of its phase diagram and quite
interesting additional insights [238–240]. One takes advantage of the so-called gauge transformation
Si Si∗ → S̃i ,
Jij Si∗ Sj∗ → J˜ij .
(38)
If this transformation is applied, the Ising Hamiltonian is conserved since all Si = ±1. In the new variables the
planted configuration becomes ferromagnetic: S̃i∗ = (Si∗ )2 = 1. What happened to the couplings? The energy of each
edge in the planted configuration has not been changed by the gauge transform. Since the frustrated interactions
were chosen independently at random in the planting, after the gauge trasformation we end up with a standard spin
glass with a fraction
∗
ρ=
eβ
2 cosh β ∗
(39)
of positive Jij = +1 couplings and the rest −1, where the −1’s are chosen independently at random.
After the gauge transformation, the planted assignment has been transformed to the ferromagnetic one where all
spins are positive. The overlap between the planted and another randomly chosen equilibrium configuration becomes
simply the magnetization. The system is now a standard spin glass with iid interactions, albeit with a ferromagnetic
bias ρ. The question of the identification of the hidden assignment is now simply mapped to the question of finding the
ferromagnetic “all spins up” configuration. This is a step that from the physics point of view simplifies the discussion.
It should be stressed, however, that the theory of Bayes-optimal inference, that is the main subject of the present
manuscript, applies to a class of planted models for which a similar gauge trasformation does not always exist.
The phase diagram of the spin glass models, with a ferromagnetic bias ρ, versus a temperature 1/β, has been
determined and one has simply to look it up in the literature [56, 57, 189]. The non-zero value of magnetization in
the ferromagnetically biased model then translates into the ability to be able to infer the planted configuration better
than by a random guessing. The value of the local magnetization of spin i is then related to the MMO estimator via
(37). In a remarkable contribution, Nishimori realized that the physics on the line ρ = eβ /(2 cosh β) is particularly
simple and quite different from a generic β. This line is now called the Nishimori line and corresponds to nothing
else but the condition for Bayes optimality β = β ∗ in the planted model.
Consequently, a number of interesting properties, that do not hold in general, hold on the Nishimori line. These
are called Nishimori conditions. In general, Nishimori conditions are properties that hold in the Bayes optimal case,
such as (23) and (24). Another useful Nishimori identity concerns the average energy E = hHi/N (30). For more
Nishimori conditions see [185, 238]. Using again the property that the planted configuration can replace an equilibrium
configuration we obtain immediately for the planted spin glass on Erdős-Rényi random graph (as defined in section
III A) of average degree c
c
E = − tanh β ∗ .
2
(40)
In Fig. 2 we give the phase diagram of the spin glass at temperature T = 1/β with ferromagnetic bias ρ. As an
example we use the spin glass on a random graph in which case an exact solution exists (see sec. III). However, on
21
a generic geometry the phase diagram could be obtained using Monte-Carlo simulation and would be qualitatively
similar. Due to the Nishimori mapping this is also the phase diagram of the planted spin glass, where the ferromagnetic
phase is understood as correlated to the planted configuration. The phenomenology of a rather generic Bayesian
inference problem can be directly read off from this phase diagram.
2
P
P
1.5
T=1/β
F
Tc
1
SG
0.5
F
M
Tc
TF
Nishimori line
0
0.5
0.6
0.7
0.8
ρc
0.9
1
β*
ρ=e /2cosh β*
FIG. 2. Phase diagram of the planted spin glass as a function of temperature T = 1/β and density ρ of edge interactions that
agree with product of spins in the planted configuration. This phase diagram is for random regular graphs of degree c = 3.
The Nishimori line, ρ = eβ /(2 cosh β) (red online), corresponds to the Bayes optimal inference. There is the spin glass (SG)
phase where the system is in a glassy state, the ferromagnetic (F) phase, the paramagnetic (P) phase and the mixed (M) one
where the system is both SG and F. The boundary of the mixed phase (black dashed) is only a guide for the eye between the
multi-critical point and the zero-temperature point from [58]. The same for the boundary of the SG phase (black dotted) that
is obtained from the inexact 1RSB approximation and is hardly distinguishable from the vertical line (black full).
C.
Bayes-optimality on the Nishimori line
First, let us assume that we know (or that we learn from data, see next section) the true value of β ∗ (or equivalently
ρ) for which the couplings J have been generated. Then, our consideration is restricted to the Nishimori line. If the
temperature is large enough, i.e. β ∗ is low enough, we are in the paramagnetic phase, and we find that all marginals
(magnetizations) are zero. It would thus be difficult to find their signs. In fact, this simply means that, in the
paramagnetic phase, inference of the hidden configuration is impossible. There is just not enough information to say
anything about the hidden/planted assignment.
As the temperature is lowered, i.e. β ∗ is increased, the situation changes. If we used β ∗ corresponding to an inverse
temperature above a certain critical value βc , then a Monte-Carlo simulation would rapidly discover the ferromagnetic
phase where local magnetizations are (mostly) positive. Applying rule (37) would give at this point a decent overlap
with the true assignment (which is, after the gauge transformation, all +1).
In terms of inference of the planted configuration these two phases have the following interpretation
• Undetectable: For β ∗ < βc , ρ < ρc , the planted spin glass system is in the paramagnetic state. Meaning that the
set of observed variables J does not contain any information about the planted configuration and its inference
is hence information-theoretically impossible.
• Detectable: For β ∗ > βc , ρ > ρc , the planted spin glass is in a ferromagnetic phase where equilibrium configurations are all correlated with the planted one and inference is possible, and algorithmically tractable e.g. via
Monte-Carlo sampling.
For the planted spin glass of a random graph, the critical temperature Tc = 1/βc where one goes from impossible
inference to easy inference can be taken from the literature [56, 57, 299] (we will show how to compute it for this case
22
in Sec. III) and in the case of random Erdős-Rényi graphs (as defined in section III A) with average degree c it reads
−1
1
TcER (c) = atanh √
,
c
(41)
−1
1
.
TcREG (c) = atanh √
c−1
(42)
while for regular random graphs it reads
On an Erdős-Rényi random graph (as defined in section III A), this indicates that it is only possible to infer something
about the planted assignment if the number of edges in the graph cN/2 is such that
c > cdetect =
1
.
(2ρ − 1)2
(43)
If this is not satisfied, the system is in the paramagnetic phase and nothing can be said about the planted configuration.
If, on the other hand, (43) is satisfied, then the magnetization aligns partly with the planted configuration. This is
a very generic situation in Bayes-optimal inference when we encounter a second order phase transition. We go from
a phase where inference is impossible, to a phase where it is rather straightforward. Note that this is not only a
statistical physics prediction, but one that can be proven rigorously. The impossibility result has been shown in
[126, 191] and the possibility one (if (43) is satisfied) in [274].
D.
Parameter mismatch and learning
Let us now discuss what happens if one does not know the value β ∗ . One approach that seems reasonable is to use
β → ∞, so that we look for the most probable assignment (this would correspond to the MAP estimator) which is the
ground state of the corresponding Ising model. We observe from the phase diagram in Fig. 2 that for ρ < ρc , at T = 0
the system is in the spin glass phase. In terms of Monte-Carlo, the presence of such a phase manifests itself via a
drastic growth of the equilibration time. Computing the corresponding magnetizations becomes very hard. Moreover,
even in the undetectable phase, ρ < ρc , the zero-temperature local magnetizations will not be zero. This is a sign of
overfitting that the MAP estimator suffers from.
In the detectable phase, ρ > ρc , the ferromagnetic phase extends only up to a temperature that for an Erdős-Rényi
random graph corresponds to a certain
1
βF = atanh
.
(44)
c tanh β ∗
Hence if we chose too low value of β as a starting point we could miss the detectability phase. At low temperatures
we might again encounter a spin glass phase that may or may not be correlated to the planted configuration. In either
case, such a phase causes equilibration problems and should better be avoided. Fortunately, there is no such glassy
phase on the Nishimori line as we shall explain in sec. II G.
The above considerations highlight the necessity of learning parameters (in this case β ∗ ). Within the theory
described above there is one particularly natural strategy for learning such parameters, closely related to the well
known algorithm of expectation maximization [74]. Let us illustrate this parameter learning on the example of the
planted Ising model where we assume that the parameter β ∗ is not known.
In Bayesian inference, we always aim to write a posterior probability distribution of what we do not know conditioned
to what we do know. In the present case we need P (β|J). Using the Bayes formula we obtain
P (β|J) =
X
P (S, β|J) =
S
Z(β, J)
P (β)
.
2N P (J) (2 cosh β)M
(45)
We notice that as long as the hyper-prior P (β) is independent of N , in the thermodynamic limit this probability
distribution P (β|J) converges to a Dirac delta function on the maximum of the second fraction on the right hand
side. This maximum is achieved at
*
+
1 X
cN
E=−
Jij Si Sj = −
tanh β .
(46)
N
2
(ij)∈E
23
The corresponding iterative algorithm for learning β is as follows. We start with some initial choice of β and compute
the average energy to update the value of β according to (46) until a fixed point is reached. Depending on the initial
value of β, and on the algorithm used to estimate the average energy, in some cases we may obtain several different
fixed points of (46). In those cases we choose the fixed point that is the global maximum of (45).
Notice now a very useful property, the condition for stationarity of (45) is the same as the Nishimori condition
(40) and indeed β = β ∗ is a fixed point of (46). This strategy is very reasonable, in the Bayes-optimal case all the
parameters are the ground truth ones and a number of Nishimori conditions hold only in this case. Hence in order to
learn the right values of parameters we iteratively impose the Nishimori conditions to hold.
Finally, let us note that in the above planted Ising model the underlying phase transition is of second order, therefore
there is no particular computational issue. Inference is either impossible, or possible and tractable. The hard phase
outlined in section (I B 5) is absent. The picture is richer in problems where in the random case there is a discontinuous
phase transition. Examples of such a case are described in detail in Sec. III E.
E.
Relations between planted, quenched and annealed disorders
From the early works, it sometimes seems necessary to have a gauge transformation (and consequently a special
form of the Hamiltonian) to be able to speak about the Nishimori conditions or other properties of the Nishimori line
[106, 239, 240]. Importantly this is not the case. The properties of the Nishimori line are more general and apply to
every case where it makes sense to talk about Bayes optimal inference. In fact, a more general notion of the Nishimori
line can be identified with the notion of Bayes-optimality. A nice article on this connection was given by Iba [137].
In this section, we shall discuss properties of the planted disorder, and its relation with the more commonly considered - the randomly-quenched and the annealed disorders. This clarifies the generic connection between statistical
physics of spin glasses and Bayesian inference. We stress again that none of this requires a gauge symmetry of the
underlying Hamiltonian.
Let us consider back the Hamiltonian of the Ising system (30). For a given system, e.g. the magnet on the fridge
in your house, the set of couplings Jij is a given fixed finite set of numbers. In the planted spin glass above, this set
was the values of interactions J we observed.
In statistical physics, where in the vast majority of cases we can take advantage of self-averaging, see sec. I C 4,
we often do not think of a given set of numbers Jij but instead about a probability distribution over such sets. For
instance randomly-quenched disorder is generated following
P (Jij ) =
1
1
δ(Jij − 1) + δ(Jij + 1)
2
2
(47)
for all (ij) ∈ E, where E are the edges of the graph of interactions. In this way we do not have a given Hamiltonian,
but rather its probability of occurrence, or equivalently an ensemble where instances appear proportionally to this
probability. Properties of the system then have to be averaged over the disorder (47). This view is adopted in
statistical physics because the self-averaging is assumed to hold in most physical systems of interest and the theory
is then easier to develop. This notion of averaging over disorder is so well engraved into a physicist’s mind that going
away from this framework and starting to use the related (replica and cavity) calculations for a single realization of
the disorder has lead to a major paradigm shift and to the development of revolutionary algorithms such as the survey
propagation [215]. In this section we do, however, want to stay in the mindset of averaging over the disorder and
explain how to fit in the notion of the planted ensemble, i.e. ensemble of instances where we average over the disorder
that was planted.
1.
The quenched and annealed average
How to take the average over disorder of Jij meaningfully is a problem that was solved a long time ago by Edwards
in his work on spin glasses and vulcanization (see the wonderful book Stealing the gold [109]). If one takes a large
enough system, he suggested, then the system becomes self-averaging: all extensive thermodynamic quantities have
the same values (in densities) for almost all realizations of the Hamiltonian. Therefore, one needs to compute the
average free energy
F
1
fquenched =
= lim −
[log Z] ,
(48)
N →∞
N
βN
where here [·] denotes the average over the disorder. This is called the quenched average, and the corresponding
computations are the quenched computations. In fact, the self-averaging hypothesis for the free energy has now been
24
proven rigorously in many cases, in particular for all lattices in finite dimension [312] and for some mean-field models
[115]. Edwards did not stop here and also suggested (and gave credit to Mark Kac for the original idea) a way to
compute the average of the logarithm of Z, known today as the replica trick, using the identity
Zn − 1
.
n→0
n
log Z = lim
(49)
The idea here is that while averaging the logarithm of Z is usually a difficult task, taking the average of Z n might
be feasible for any integer value of n. Performing a (risky) analytic continuation to n = 0, one might compute the
averaged free energy over the disorder as
1
[Z n ] − 1
lim
.
N →∞ N β n→0
n
fquenched = − lim
(50)
The computation of the quenched free energy is in general still very difficult and remains a problem at the core of
studies in statistical physics of disordered systems. It is much easier to consider the so-called annealed average, one
simply averages the partition sum and only then takes the logarithm
fannealed = −
1
log [Z].
Nβ
(51)
This is a very different computation, and of course there is no reason for it to be equal to the quenched computation.
For the example of the Ising model with binary uniformly distributed couplings (47) we obtain (independently of
the dimension or geometry)
Ising
− βfannealed
= log 2 +
c
log(cosh β) ,
2
(52)
where c is the average coordination number (degree) of the interaction graphs.
It is important to notice that the annealed average is wrong if one wants to do physics. The point is that the free
energy is an extensive quantity, so√ that the free energy per variable should be a quantity of O(1) with fluctuations
going in most situations as ∝ (1/ N ) (the exponent can be different, but the idea is that fluctuations are going to
zero as N → ∞). The partition sum Z, however, is exponentially large in N , and so its fluctuations can be huge.
The average can be easily dominated by rare, but large, fluctuations.
Consider for instance the situation where Z is exp (−βN ) with probability 1/N and exp (−2βN ) with probability
1 − 1/N . With high probability, if one picks up a large system, its free energy should be fquenched = 2, however in the
annealed computation one finds
1
1
[Z] =
exp (−βN ) + 1 −
exp (−2βN )
(53)
N
N
and to the leading order, the annealed free energy turns out to be fannealed = 1.
One should not throw away the annealed computation right away, as we shall see, it might be a good approximation
in some cases. Moreover, it turns out to be very convenient to prove theorems. Indeed, since the logarithm is a
concave function, the average of the logarithm is always smaller or equal to the logarithm of the average, so that
fannealed ≤ fquenched .
(54)
This is in fact a crucial property in the demonstrations of many results in the physics of disordered systems, and in
computer science this is the inequality behind the “first moment method” [227].
Furthermore, there is a reason why, in physics, one should sometimes consider the annealed average instead of the
quenched one. When the disorder is changing quickly in time, on timescales similar to those of configuration changes,
then we indeed need to average both over configurations and disorder and the annealed average is the correct physical
one. This is actually the origin of the name annealed and quenched averages.
2.
Properties of the planted ensemble
In sec. II A we described the planted spin glass, where we first considered the planted configuration S∗ and a graph
G = (V, E), and then we considered the probability over interactions P (J|S∗ ) eq. (34). The ensemble of instances
(S∗ , G, J) generated by the planting, i.e. by the three steps of Sec. II A, is called the planted ensemble.
From a physics point of view the two most important properties of the planted ensemble, the two golden rules to
remember, are:
25
• The planted configuration is an equilibrium configuration of the Hamiltonian derived from the posterior distribution. By this we mean that the planted configuration has exactly the same macroscopic properties (e.g. its
energy or magnetization) as any other configuration that is uniformly sampled from the posterior distribution.
Indeed, when we view planting as an inference problem, we can write the corresponding posterior probability distribution as in eq. (35). In section I B 3 we derived that in the Bayes optimal inference the planted
configuration behaves exactly in the same way as a configuration sampled from the posterior.
Given that generating an equilibrium configuration from a statistical physics model is in general a difficult task
(this is, after all, why the Monte-Carlo method has been invented!) this is certainly a very impressive property.
It actually means that, by creating a planted instance, we are generating at the same time an equilibrium
configuration for the problem, something that would normally require a long monte-carlo simulation.
• The realization of the disorder of the planted problem is not chosen uniformly, as in the randomly-quenched
ensemble (47), but instead each planted instance appears with a probability proportional to its partition sum.
This can be seen from eq. (36) where the probability distribution over J in the planted ensemble is related to
the partition function of the planted system Z(J). A more explicit formula can be obtained using the annealed
partition function (52)
Pplanted (J) =
Z(J)
,
ΛZannealed
(55)
where Λ is the number of all possible realizations of the disorder, Λ = 2M for the Ising model. Note the 1/Λ can
also be interpreted as the probability to generate a given set of couplings in the randomly-quenched ensemble
where the couplings are chosen uniformly at random.
The average over the planted ensemble might seem related to the annealed average because for instance the planted
energy (40) is always equal to the annealed average energy, which is easily derived from the annealed free energy
(52). However, the free energy of the planted ensemble is in general different from the annealed free energy (52), as
illustrated e.g. in Fig. 3.
3.
Quiet planting
Let us now imagine we have found a system where the (quenched) free energy is exactly equal to the annealed
free energy. Then according to eq. (55) we see that for such a system Pplanted (J) = 1/Λ = Pquenched (J), meaning
that generating instances from the planted ensemble is the same thing as generating from the randomly-quenched
ensemble. Such planting is denoted as quiet. In fact, we do not even need the free energies to be exactly equal, it is
sufficient that the free energy densities are the same. This is because atypical instances are usually exponentially rare
and hence any difference in the free energies that does not show up in the leading order will not generate atypical
instances.
It so happens that in mean-field systems the (self-averaging) free energies of paramagnets are indeed equal to the
annealed free energies. The equilibrium in these systems can hence be studied using the planted ensemble and this is
greatly advantageous as we will see in section II H 1.
The idea of quiet planting comes from a rigorous work [4] where the above notion was proven rigorously. In
mathematics this property is formalized and the two ensembles termed contiguous [4, 231]. The paper that pioneered
the usage of quiet planting in physics, and that also established the name “quiet”, is Hiding quiet solutions in random
constraint satisfaction problems [182].
F.
Replica symmetry breaking
The replica and cavity method, that are among the main assets that statistical physics offers for analysis of
inference problems, were originally developed to understand behavior of glasses and spin glasses via the so-called
replica symmetry breaking (RSB). We will explain in the next section, that for Bayes-optimal inference RSB is not
needed. But to appreciate fully the results exposed in this manuscript it is useful to have a very basic idea of what
RSB is and what it implies.
RSB is usually explained together with the related (relatively involved) calculations, see e.g. [217, 238]. In this
section we try to convey the basic ideas behind RSB without going into the computations. We rely on the descriptions
of what would happen to a Monte Carlo simulation in a system that exhibits various versions of RSB.
26
0.25
0.2
splanted
0.15
s planted−cluster
0.1
s quenched
s annealed
0.05
Σ
0
12.6
12.8
13
Cd
13.2
Cc
13.4
13.6
13.8
14
Ccol
C
FIG. 3. Entropy density (which is equal to the free energy at zero temperature) for the 5-coloring problem on an Erdős-Rényi
random graph (as defined in section III A) of average degree c. The planted entropy (in red) is compared to the quenched
one (in green) and the annealed one (in black). For c < cc the three curves agree. The quenched entropy goes to −∞ at the
colorability threshold ccol . The region between cd and cc is relevant only for cases with a first order phase transition as discussed
in Sec. III E. In mean field theory of spin glasses this region is called the d1RSB phase [41]. We also show the entropy of the
planted cluster below cc (black dashed), which is less than the total entropy, the difference is called the “complexity” Σ (in
blue) and is the logarithm of the number of pure states. Bottom: Sketch of the shape of the space of all valid colorings in the
planted case. At an average degree cd the space of solutions shatters into exponentially many clusters, the planted cluster being
one of them. Beyond cc the planted cluster contains more solutions than all the others together. At cs the last non-planted
cluster disappears (this is the coloring threshold in a non planted, purely random model). Figure taken from [182].
Let us start with the concept of replica symmetry. For the purpose of the present manuscript we can say that a
probability measure P (x|y) is replica symmetric if a MCMC that is initialized at equilibrium, i.e. in a configuration
sampled uniformly at random from P (x|y), is able to sample P (x|y) (close to) uniformly in a number of iterations
that is linear in the size of the system. A probability measure P (x|y) exhibits dynamical one-step of replica symmetry
breaking (d1RSB) if MCMC initialized at equilibrium explores in linear time only configurations that belong into an
exponentially small (in N ) fraction of the equilibrium configurations. In both RS and d1RSB the distance between
two configurations sampled from the Boltzmann probability measure is the same number with probability going to
one (as N → ∞).
We speak about static one-step replica symmetry breaking (1RSB) if MCMC initialized at equilibrium is able to
explore configurations belonging to a finite fraction of all equilibrium configurations. Moreover the distance between
two randomly sampled configurations can take (with probability going to one) one of two possible values.
A probability measure P (x|y) corresponds to full-step replica symmetry breaking (FRSB) if the distance between two randomly chosen configurations is distributed according to some probability distribution with continuous
non-trivial support. In terms of behavior of MCMC, in a FRSB regime the time needed to sample configurations
uniformly starting from equilibrium has a more complex behavior: relaxations are critical and a power-law behavior
is expected [289].
G.
No replica symmetry breaking in Bayes-optimal inference.
From the methodological point of view, the Nishimori line has one absolutely crucial property: the absence of glass
phase at equilibrium. In the words of the mean-field theory of spin glasses, there is no static replica symmetry breaking
on the Nishimori line. Note that there might be, and indeed is in all the cases where the underlying transition is of
the 1st order, the dynamical one-step RSB phase (d1RSB) [41]. The marginals in the d1RSB phase are, however, still
27
exactly described by the belief propagation algorithm (as described in section III), and this is what matters to us in
the Bayes-optimal inference.
A replica symmetry breaking phase (or equivalently the static spin glass phase) can be defined by the non-selfaveraging of the overlap q between two configurations randomly sampled from the Boltzmann distribution. Depending
on the realization of the sampling, the Hamming distance between the two configurations is with nonzero probability
different from its average.
In section I B 3 we derived that, on the Nishimori line, the overlap m between a typical configuration and the planted
one is equal to the overlap between two typical configurations q. With similar arguments one can derive equality of the
distribution (over realizations of the sampling) P (q) of the overlaps between two configurations randomly sampled from
the posterior, and the distribution of the magnetizations P (m), which is the overlap between the planted configuration
and a random sample from the posterior [238]. In physics, the magnetization is always argued to be self-averaging
(the same is true for the free energy density, and the magnetization is its derivative with respect to the magnetic
field), and hence on the Nishimori line the overlap is also self-averaging. From this, one argues that P (q) is trivial,
which in mean-field spin glasses indicates the absence of a spin glass phase. However, from a rigorous point of view,
self-averaging of the magnetization remains an open problem.
Another way to see the problem is to remark that, given that the spin glass susceptibility is equal to the ferromagnetic
one for planted systems (see for instance [184, 185]), then it is hard to reconcile the presence of an equilibrium
ferromagnet phase (where the susceptibility is finite) with an equilibrium spin glass phase (where the susceptibility
diverges).
A complementary (rigorous) argument was presented by Montanari [220] who showed that in sparse systems (i.e.
each observations depends on a bounded number of variables) in the Bayes-optimal setting, two point correlations
decay. In the static spin glass phase these correlations could not decay.
This means that we can exclude the presence of a static equilibrium transition to a glass phase on the Nishimori
line. Note, however, that the dynamics can still be complicated, and in fact as we mentioned, a dynamic spin glass
(d1RSB) phase can appear. As soon as we deviate from the Bayes optimal setting and mismatch the prior, the model
or their parameters, the stage for glassiness to arrive is open. And indeed as we see in Fig. 2 there are regions with
RSB when β 6= β ∗ , typically when we take the temperature too small. When studying inference with a mismatched
prior or model we thus always have to keep this in mind before making claims about exactness.
H.
Applications of planting for studies of glasses
Now that we have explored the concept of planting and quiet planting, and its connection to inference, we will show
that it also has a number of consequences for the study of structural glasses in physics. While these are not directly
relevant to inference, it illustrates in many ways the fascinating properties of the planted ensemble and the deep
connection between inference, optimization problems and the statistical physics of glasses, and we thus summarize
these properties in this section.
1.
Equilibration for free in glasses
Let us discuss the physical consequences of the fact that in quietly planted systems the planted configuration
behaves exactly the same way as any other equilibrium configuration. We can hence literarily equilibrate for free.
This is particularly interesting in the study of glassy models. As a matter of fact, the very concept of glassiness can
be defined as super-fast growth of the equilibration time as the glass transition temperature is approached [41].
The core of the interest in glasses is concerned by finite (usually three) dimensional models. In those cases the
annealed and quenched free energies are known not to coincide. A class of theories of glasses is, however, based on
mean-field (i.e. systems living on random geometries of fully connected graphs) disordered systems that present a
discontinuous phase transition. The analog of cd in these systems is called the dynamical glass transition, and the
analog of cc is called the Kauzmann glass temperature [32]. Properties of the equilibration time are well studied and
it turns out that the equilibration time (i.e. number of Monte Carlo steps per variable in the thermodynamic limit)
diverges as the dynamical temperature is approached [222] and is exponentially large in the system size between the
dynamical and Kauzmann temperature. This in practice means that numerical simulations of these models are very
cumbersome.
In most of the mean-field models used for the studies of glasses, the quiet planting works. Numerical simulations
studying the dynamics from equilibrium can hence be speeded up considerably by initializing in the planted configuration. A particular version from of this trick was first used in [221] where it enabled for the first time precise
28
simulation of the dynamics starting from equilibrium. It was later popularized by [182] and has now been used in a
considerable number of works. Without being exhaustive let us mention Refs. [12, 60, 61, 93, 202, 203].
Even if one does not care about the equivalence to the randomly-quenched ensemble one can use planting to obtain
equilibrium configurations even in general finite dimensional systems. Indeed the property of planted configuration
being the equilibrium one for a corresponding Boltzmann measure is general. This has been used in numerical studies
of so-called pinning in glasses, see e.g. [132].
2.
Following states and analyzing slow annealing
In a very influential work, Franz and Parisi [96] considered the so called “potential” method, that aimed to analyse
the properties of a single pure state in glasses. Perhaps more interestingly, this allows to describe the adiabatic
evolution of these glassy Gibbs states as an external parameter, such as the temperature, is slowly tuned. Again,
there is a deep connection with the planted ensemble which we shall now discuss. Ref. [182] studied the state that is
created in the planted coloring around the planted configuration. Adding to this the concept of quiet planting, this
framework allows us to study the shape of states in mean-field glassy systems and is equivalent to the Franz-Parisi
construction.
When a macroscopic system is in a given phase, and if one tunes a parameter, say the temperature, very slowly
then all observables, such as the energy or the magnetization in a magnet, will be given by the equilibrium equation
of state. In a glassy system, where there could be many such states, it is very interesting to see how they evolve
upon cooling, as they tend to fall off equilibrium very fast. This is for instance what is happening in the so-called
3-XOR-SAT problem, also called 3-spins spin-glass [114, 216, 265]. This is nothing but an Ising spin glass where the
two body interaction of Hamiltonian (30) is replaced by a three spin interaction (i.e. each term is a triplet Jijk Si Sj Sk ).
A plot that illustrates this phenomena is given in Fig. 4. The blue line is the equilibrium (in technical terms, the
one-step replica symmetry breaking solution) energy of the 3-regular hyper-graph (this means that every variables
has exactly 3 triplets of interactions, chosen randomly) 3-XOR-SAT problem as a function of temperature T . For
this system quiet planting works at all temperatures. The red lines are energies that we obtain when we plant
quietly (at temperature where the red lines crosses the blue one) and then lower or increase the temperature and
monitor the corresponding energy given by the dynamics. This procedure has been called state following in [183, 323],
because physically this is a study of the corresponding state at temperature at which the state is no longer among
the equilibrium ones.
Note that such state following was often described in physics of glasses on the level of a thought-experiment.
Methods for quantitative analysis existed only for the particular case of spherical p-spin model, via an exact solution
of the dynamics [41, 67, 68]. For other systems, in particular diluted ones, state following was considered to be an open
problem, see e.g. the concept suggested but not analyzed in [172]. The method for state following is mathematically
very much related to the concept of Franz-Parisi potential at two different temperatures [22, 96].
From an algorithmic point of view, the method of state following opens the exciting possibility of analyzing limiting
energy of a very slow annealing. It is well known that in a finite size system annealing always finds the ground
state, in a time that is in general exponential in the system size. A crucially interesting question is: What is the
lowest achievable energy when the annealing rate scales linearly with the size of the system? In mean-field models of
glasses equilibration is possible in a time that is linear in the system size down to the dynamical temperature Td (or
equivalently up to cd ), after that the dynamics gets blocked in one of the exponentially many states. If we take an
equilibrium state at Td − and follow it down to zero temperature the resulting energy is a very good candidate for
the behavior and for the limiting energy of very slow simulated annealing. Very slow annealing for the Ising p-spin
model was compared to the theoretical prediction from state following in [186] .
Paradox in state following. There is, however, a problem that appears when one tries to carry out the calculation
for following states that are at equilibrium close to Td . When following the state down in temperature, the replica
symmetry gets broken at some point (the boundary between full and dashed lines in Fig. 4). Moreover, at yet lower
temperature the solution correlated to the planted configuration disappears (in a mathematically analogous way as
when a spinodal line is encountered). Even using the 1RSB computation to describe the internal structure of the
state we were not able to follow it down to zero temperature. A possible explanation is that the correct description of
this region is the full-step replica symmetry breaking (FRSB). To test this hypothesis Ref. [294] analyzed the mixed
spherical p-spin model where the FRSB is tractable.
The findings of [294], however, raised even more questions. It is found by the Nishimori mapping to the model
with a ferromagnetic bias (as described in section II B) that there is no FRSB solution to the problem, yet there
is no magnetized 1RSB solution at very low temperature either. Yet in MCMC simulation one always confirms the
expectation that as temperature is lowered the bottom of the state is correlated to the reference configuration. In
other words, using the Nishimori mapping, the non-magnetized spin glass phase that is assumed to exist at low
29
FIG. 4. State following in regular 3-XOR-SAT. Figure taken from [323].
temperatures for ρ > ρc in Fig. 2 is not physical in the context of state following. These observations are hence an
inconsistency in the theory. Loose ends such as this one are always interesting and basically always eventually lead
to new theoretical development. This state following paradox should hence not be forgotten and should be revisited
as often as new possibilities open.
One such possibility might be related to the calculation involving a chain of reference (planted) configurations at
lower and lower temperatures as suggested in [97].
3.
Melting and glass transition
Using the planted models and extended form of the Nishimori conditions the analogy between glassy dynamics and
melting was studied in [184, 185]. Perhaps the theoretically most intriguing and very fundamental question in the
theory of glasses is: Is there a true glass transition in a finite-dimensional system (i.e. system where the graph of
interactions can be embedded into finite-dimensional Euclidean space)? Such an ideal glass transition has to come
with diverging time and length scales when approaching the transition. A definitive answer to this question is difficult
to obtain as both simulations and experiments are faced with the extremely slow dynamics. According to the random
first order theory of the glass transition [168, 169], there exist systems with time and length scales with a genuine
divergence at the ideal glass transition temperature. Moreover this theory is able to predict in which systems such
an ideal glass transition exists and in which it does not. However, the random first order theory is not free from
criticisms, see for instance [36] and references therein.
The question of existence of the ideal glass transition remains open, but in [185] it was derived that if there is a
finite-dimensional model with a first order phase transition on the Nishimori line (or equivalently in the Bayes optimal
inference) then there is an ideal glass transition as well. This result was reached by using the Nishimori mapping
from a planted model to a ferromagnet, sec. II B, and comparing their dynamical properties. In particular, if there is
a first order phase transition on the Nishimori line then it comes with two quite uncommon properties - absent latent
heat, and the melting dynamics being equivalent to the stationary equilibrium dynamics. Arguably, first order phase
transitions are easier to describe or exclude than the tricky glass transition. This is hence an interesting path towards
the fundamental question of existence of the ideal glass transition.
30
III.
PHASE DIAGRAM FROM THE CAVITY METHOD
Now we move to special cases of inference problems where the phase diagram and phase transitions can be computed
analytically exactly, and for which related algorithms more efficient than Monte-Carlo can be derived. We shall achieve
this goal using extended mean-field methods. Slightly abusively we will call mean-field systems those where these
methods lead to exact results.
Mean-field methods are often the first step in the understanding of systems in physics, and disordered systems
are no exception. In spin glasses, in particular, the mean-field model proposed in 1975 by D. Sherrington and S.
Kirkpatrick [285] has been particularly important. We refer to the classical literature [214] for the replica solution,
involving the so-called replica symmetry breaking, devised by Parisi. It is not that we are not interested in the replica
method, quite the contrary, the replica method is at the roots of many results in this review. We shall, however, focus
here in the alternative cavity method [211–213] since this leads more naturally to specific algorithms and algorithmic
ideas. The development of more efficient algorithms is the goal in inference problems.
There are two main types of lattices for which the mean-field theory is exact. The first type is the fully connected
lattice underlying the canonical Sherrington–Kirkpatrick model. Similar solutions have been derived for the p-spin
model [114] and for the Potts glass model [112], and these played a major role in the development of the mean-field
theory of the structural glass transition [165–167]. The linear estimation problem that we treat in Sec. VI is an
example of a problem on a fully connected lattice.
The second type of lattice for which the mean-field theory is exact is given by large random graphs with constant
average degree, a case commonly refereed to as “Bethe lattice” in the physics literature [304]. In 2001 Mézard and
Parisi [211, 212], using the cavity method, adapted the replica symmetry breaking scheme to solve models on such
sparse lattices. There is a number of reasons why Bethe lattices are interesting, the two major ones are: (a) because
of the finite degree, they provide a better approximation of the finite dimensional case and because the notions of
distance and neighboring can be naturally defined, and (b) the related random graphs are fundamental in many
interdisciplinary cases, for instance in computer science problems, which is our main motivation here.
The cavity method is a generalization of the Bethe and Onsager ideas [33, 242] for disordered systems. It was initially developed by Mézard, Parisi and Virasoro [213] as a tool to recover the solution of the Sherrington-Kirkpatrick
model without the use of replicas. It was subsequently developed by Mézard and Parisi to deal with the statistical
physics of disordered systems on random graphs [211]. There is a deep global connection between the Bethe approximation and the so-called belief propagation (BP) approach in computer science (in error correction [99] and Bayesian
networks [249]) as was realized in connection to error correcting codes by Kabashima and Saad [154], and was put
in a more general setting by [318]. For a more recent textbook that covers in detail the cavity method and belief
propagation see [217], for a very comprehensive explanation of belief propagation see [318].
A.
Random graphs
We will remind basic properties of sparse Erdős-Rényi (ER) [87, 88] random graphs, that we used briefly already
in sec. II A. At the core of the cavity method is the fact that such random graphs locally look like trees, i.e. there are
no short cycles going trough a typical node.
An ER random graph is taken uniformly at random from the ensemble, denoted G(N, M ), of graphs that have
N vertices and M edges. To create such a graph, one has simply to add random M edges to an empty graph.
Alternatively, one can also define the so called G(N, p) ensemble where an edge exists independently for each pair of
nodes with a given probability 0 < c/N < 1. The two ensembles are asymptotically equivalent in the large N limit,
when M = c(N − 1)/2. The constant c is called the average degree. We denote by ci the degree of a node i, i.e. the
number of nodes to which i is connected. The degrees are distributed according to Poisson distribution, with average
c.
Alternatively, one can also construct the so-called regular random graphs from the ensemble R(N, c) with N vertices
but where the degree of each vertex is fixed to be exactly c. This means that the number of edges is also fixed to
M = cN/2.
We will mainly be interested in the sparse graph limit, where N → ∞, c ∼ O(1), M ∼ N . The key point is that,
in this limit, such random graphs can be considered locally as trees [37, 139]. The intuitive argument for this result
is the following one: starting from a random site, and moving following the edges, in ` steps c` sites will be reached.
In order to have a loop, we thus need c` ∼ N to be able to come back on the initial site, and this gives ` ∼ log(N ).
31
B.
Belief propagation and the cavity method
Now that we have shown that random graphs are locally tree-like, we are ready to discuss the cavity method. We
will consider a system with Hamiltonian
X
X
H=−
Jij Si Sj −
hi Si
(56)
i
(ij)∈E
and derive the thermodynamics of this model on a tree. We have to deal with N variables Si with i = 1, . . . , N , and
M interactions, that we denote ψij (Si , Sj ), on each edge (ij) of the graph G(V, E), with ψij (Si , Sj ) = exp(βJij Si Sj ).
Further denoting ψi (Si ) = exp(βhi Si ) the “prior” information the Hamiltonian gives about Si , the total partition
sum reads
X
Y
Y
ψi (Si )
Z=
ψij (Si , Sj )
(57)
i
{Si=1,...,N }
(ij)∈E
To compute Z on a tree, the trick is to consider instead the variable Zi→j (Si ), for each two adjacent sites i and j,
defined as the partial partition function for the sub-tree rooted at i, when excluding the branch directed towards j,
with a fixed value Si of the spin variable on the site i. We also need to introduce Zi (Si ), the partition function of the
entire complete tree when, again, the variable i is fixed to a value Si . On a tree, these intermediate variables can be
exactly computed according to the following recursion
!
Y
X
Zi→j (Si ) = ψi (Si )
Zk→i (Sk )ψik (Si , Sk ) ,
(58)
Sk
k∈∂i\j
Y X
Zi (Si ) = ψi (Si )
Zj→i (Sj )ψij (Si , Sj ) ,
j∈∂i
(59)
Sj
where ∂i \ j denotes the set of all the neighbors of i, except spin j. In order to write these equations, the only
assumption that has been made was that, for all k 6= k 0 ∈ ∂i \ j, Zk→i (Sk ) and Zk0 →i (Sk0 ) are independent. On a
tree, this is obviously true: since there are no loops, the sites k and k 0 are connected only through i and we have
“cut” this interaction when considering the partial quantities. This recursion is very similar, in spirit, to the standard
transfer matrix method for a one-dimensional chain.
In practice, however, it turns out that working with partition functions (that is, numbers that can be exponentially
large in the system size) is somehow impractical, and we can thus normalize eq. (58) and rewrite these recursions in
terms of probabilities. Denoting ηi→j (Si ) as the marginal probability distribution of the variable Si when the edge
(ij) has been removed, we have
Zi (Si )
ηi (Si ) = P
.
0
S 0 Zi (S )
Zi→j (Si )
,
ηi→j (Si ) = P
0
S 0 Zi→j (S )
(60)
So that the recursions (58-59) now read
!
X
ψi (Si ) Y
ηi→j (Si ) =
ηk→i (Sk )ψik (Si , Sk ) ,
zi→j
Sk
k∈∂i\j
Y
X
ψi (Si )
ηi (Si ) =
ηj→i (Sj )ψij (Si , Sj ) ,
zi
j∈∂i
(61)
(62)
Sj
where the zi→j and zi are normalization constants defined by:
!
zi→j =
X
ψi (Si )
Si
zi =
X
Si
ψi (Si )
Y
X
k∈∂i\j
Sk
Y
j∈∂i
ηk→i (Sk )ψik (Si , Sk )
X
ηj→i (Sj )ψij (Si , Sj ) .
Sj
,
(63)
(64)
32
The iterative equations (61,62), and their normalization (63,64), are called the belief propagation equations. Indeed,
since ηi→j (Si ) is the distribution of the variable Si when the edge to variable j is absent, it is convenient to interpret
it as the “belief” of the probability of Si in absence of j. It is also called a “cavity” probability since it is derived
by removing one node from the graph. The belief propagation equations are used to define the belief propagation
algorithm
• Initialize the cavity messages (or “beliefs”) ηi→j (Si ) randomly or following a prior information ψi (Si ) if we have
one.
• Update the messages in a random order following the belief propagation recursion eq. (61,62) until their
convergence to their fixed point.
• After convergence, use the beliefs to compute the complete marginal probability distribution ηi (Si ) for each
variable. This is the belief propagation estimate on the marginal probability distribution for variable i.
Using the resulting
marginal distributions, one can compute, for instance, the equilibrium local magnetization via
P
mi = hSi i = Si ηi (Si )Si , or basically any other local quantity of interest.
At this point, since we have switched from partial partition sums to partial marginals, the astute reader could
complain that it seems that we have lost out prime objective: the computation of the partition function. Fortunately,
one can compute it from the knowledge of the marginal distributions. To do so, it is first useful to define the following
quantity for every edge (ij):
zij =
X
ηj→i (Sj )ηi→j (Si )ψij (Si , Sj ) =
Si ,Sj
zi
zj
=
,
zj→i
zi→j
(65)
where the last two equalities are obtained by plugging (61) into the first equality and realizing that it almost gives
eq. (64). Using again eqs. (61-64), we obtain
X
Y X
zi =
ψi (Si )
ηj→i (Sj )ψij (Si , Sj )
Si
Sj
j∈∂i
=
X
ψi (Si )
Si
Y
X
j∈∂i
Sj
P
Zj→i (Sj )
Si Zi (Si )
P
P
ψij (Si , Sj ) = Q
,
0
S 0 Zj→i (S )
j∈∂i
Sj Zj→i (Sj )
(66)
and along the same steps
P
Zj→i (Sj )
P
.
k∈∂j\i
Sk Zk→j (Sk )
zj→i = Q
Sj
(67)
P
For any spin Si , the total partition function can be obtained using Z = Si Zi (Si ). We can thus start from an
arbitrary spin i
X
Y X
Y
Y X
zj→i
Z=
Zi (Si ) = zi
Zj→i (Sj ) = zi
Zk→j (Sk ) ,
(68)
Si
j∈∂i
Sj
k∈∂j\i Sk
j∈∂i
and we continue to iterate this relation until we reach the leaves of the tree. Using Eq. (65), we obtain
Q
Y
Y
Y zj Y zk
zi
zj→i
··· = Q i
.
Z = zi
zk→j · · · = zi
zij
zjk
(ij) zij
j∈∂i
k∈∂j\i
j∈∂i
(69)
k∈∂j\i
We thus obtain the expression of the free energy in a convenient form, that can be computed directly from the
knowledge of the cavity messages, often called the Bethe free energy
X
X
f N = −T log Z =
fi −
fij ,
i
(ij)
(70)
fi = −T log zi , fij = −T log zij ,
33
where fi is a “site term” coming from the normalization of the marginal distribution of site i, and is related to the
change in Z when the site i (and the corresponding edges) is added to the system. fij is an “edge” term that can be
interpreted as the change in Z when the edge (ij) is added. This provides a convenient interpretation of the Bethe
free energy eq. (70): it is the sum of the free energy fi for all sites but, since we have counted each edge twice we
correct this by subtracting fij (see for instance [212]).
We have now entirely solved the problem on a tree. There is, however, nothing that prevents us from applying the
same strategy on any graph. Indeed the algorithm we have described is well defined on any graph, but we are not
assured that it gives exact results nor that it will converge. Using these equations on graphs with loops is sometimes
referred to as “loopy belief propagation” in Bayesian inference literature [318].
When do we expect belief propagation to be correct? As we have have discussed, random graphs are locally tree-like:
they are trees up to any finite distance. Further assuming that we are in a pure thermodynamic state, we expect that
we have short range correlations, so that the large O(log N ) loops should not matter in a large enough system, and
these equations should provide a correct description of the model.
Clearly, this must be a good approach for describing a system in a paramagnetic phase, or even a system with a
ferromagnetic transition (where we should expect to have two different fixed points of the iterations). It could be,
however, that there exists a huge number of fixed points for these equations: how to deal with this situation? Should
they all correspond to a given pure state? Fortunately, we do not have such worries, as the situation we just described
is the one arising when there is a glass transition. In this case, one needs to use the cavity method in conjunction with
the so-called “replica symmetry breaking” approach as was done by Mézard and Parisi [211, 212, 215]. In this review,
however, we discuss optimal Bayesian inference and we are guaranteed from the Nishimori conditions discussed in
sec. II G that no such replica symmetry breaking occurs. We can assume safely that in the Bayes-optimal setting
the simple belief propagation approach is always providing the correct description (unless there are other sources of
correlations in the system, e.g. the graph is not random but a lattice or the quenched disorder is correlated in a
non-local manner).
C.
Properties of the randomly-quenched ensemble
The first system on which we will illustrate how to use the cavity method and the related belief propagation to
compute the phase diagram is the Viana-Bray (V-B) model for the so-called diluted spin glass [304]. The reason we
discuss first the solution for this well known model is that, using the theory developed in sec. II, we will be able to
read off the solution for the planted spin glass model from the phase diagram of the randomly-quenched model, see
Fig. 2. Viana and Bray studied the Ising spin glass Hamiltonian (30) with the lattice being an Erdős-Rényi random
graph with average degree c. Interactions Jij are chosen from ±1 independently and uniformly at random and, for
simplicity, we consider the absence of a magnetic field. We call the ensemble of instances generated this way the
randomly-quenched ensemble.
The description of the V-B model within this approach was originally done by [44]. Instead of working with the
iteration (61), it is convenient to rewrite it in terms of an “effective” local cavity fields ui→j defined as
ηi→j (Si ) =
eui→j Si
.
2 cosh eui→j
(71)
With a bit of algebra, the belief propagation equation (61) can be written as
X
1
,
ui→j
= f (Jij , {uk→i
atanh tanh(βJij ) tanh β
uk→i
t
t−1
t−1 }k∈∂i\j ) =
β
(72)
k∈∂i\j
where we have added an index t that stands for a “time” iteration. As stated, iterating this procedure allows to actually
solve the problem on a tree, which is at the root of the cavity method. Here, however, we apply this procedure on the
graph, iteratively, hence the time indices in eq. (72). The local magnetization is then computed from the BP fixed
point using the hyperbolic tangent of the local field acting on the spin, and thus reads
X
mi = tanh (β
uj→i ) .
(73)
j∈∂i
Applying the same notation to the Bethe free energy per site (70) leads to
"
#
X
X
X
i
SβJij
j→i
i→j
βN f =
(ci − 1) log 2 cosh(βh ) −
log
2e
cosh(βh
+ Sβh
) ,
i
(ij)∈E
S=±1
(74)
34
P
P
where hi = k∈∂i uk→i and hi→j = k∈∂i\j uk→i , and ci is the degree of node i. Physical properties of the system
are computed from the fixed points of the BP equations (72). At this point, there are two different strategies to solve
the problem. The first one is to investigate numerically what is happening on a given realization of the problem, i.e.
to generate a large random graph and to actually run the BP iteration on it. This is actually very instructive (as we
shall see) but no matter how large our graph will be, it will not be at the thermodynamic limit.
The second strategy is to work on an infinite graph directly via a so called population dynamics [211]. We attempt
to describe the model by the knowledge of the probability distribution P(u) of the messages u on an edge with value
J over the (infinite) graph. Consider for instance the case of a random graph with fixed degree ci = c. Clearly, within
a pure Gibbs states, P(u) satisfies the BP equations (72) so that
P(u) =
1
Z
Z
P (J)dJδ[u − f (J, {ud }d=1,...,c−1 )]
c−1
Y
d=1
P(ud )dud
(75)
where f (.) is given by eq. (72). While this appears to be a difficult equation to solve explicitly, it can be conveniently
approximated to arbitrary precision with the so-called population dynamics method [211]. In this case, one models
the distribution P(u) by a population of M values
M
Ppop (u) =
1 X
δ(u − ul ) .
M
(76)
l=1
Plugging this expression in eq. (75) leads to the population dynamics algorithm:
• Initialize all values ul randomly.
• For a given element l of the population, draw c − 1 values of u from the population, and a value J of the coupling
from P (J), compute a new value unew = f (J, {u}), and replace the element ul by unew .
• Repeat the last step until Ppop (u) has converged to a good approximation of P(u).
This method is very simple and, given the population dynamics distribution, allows to compute the value of the free
energy per site via (70).
Either way, depending on the inverse temperature β in the Viana-Bray model, one observes one of the two following
possibilities: At low β there is a stable paramagnetic fixed point ui→j = 0 for all (ij) ∈ E. The corresponding
paramagnetic free energy reads
− βf (β, J) = log 2 +
c
[log (2 cosh β) − log 2] .
2
(77)
We do notice that the same expression is obtained for the annealed free energy for this model.
At high β (low temperature) the iterations of (72) do not converge on a single graph. The last statement, of course,
can not be checked simply when performing the population dynamics algorithm. A similar effect can, however, be
observed by performing two-populations dynamics with the same random numbers starting from infinitesimally close
initial populations. At high β the two populations will diverge to entirely different values [246]. This is a clear sign of
long range correlations and of the appearance of a spin glass phase. We are, of course, interested to know the precise
location of the phase transition.
Thouless [299] analyzed the threshold value of βc by linearizing the update (72) around the uniform fixed point,
thereby obtaining
X
ui→j
= tanh(βJij )
uk→i
(78)
t
t−1 .
k∈∂i\j
Thouless then argues that averaged over the disorder Jij the above equation gives zero, and when its square is averaged
we get
hu2t i = c tanh2 β hu2t−1 i ,
(79)
where we remind that c is the average degree of the ER graph, that is also equal to the average branching factor of
the tree that approximates locally the random graph. We thus expect that the paramagnetic phase becomes unstable
when c tanh2 β ≥ 1.
35
In fact, by analyzing such perturbations, one is investigating the behavior of the susceptibility. Indeed the spin
glass susceptibility can be written as
χSG ≡ β
∞
X
X
hS0 Si i2c ≈ β
cr hS0 Sr i2c .
(80)
r=0
i
In eq. (80), · · · represents a spatial average over the whole graph, h· · ·i represents a thermal average, Sr are all spins
at distance r from S0 , and Xc is the connected version of the correlation function X: hS0 Si ic = hS0 Si i − hS0 ihSi i.
The correlation can be computed using derivatives with respect to an external magnetic field on the spin S0
βhS0 Si i2c
=
∂hSi i
∂hh0 i
2
.
(81)
Since the physical field is a function of the cavity fields, we can monitor the propagation of the response using the
chain rule
∂hSr i
∂hSr i ∂hhr i ∂hhr−1 i
∂hh1 i
=
···
∂hhc0 i
∂hhr i ∂hhr−1 i ∂hhr−2 i
∂hh0 i
(82)
where we indicate with hi the cavity field going from spin i to spin i + 1 for simplicity and hr , hr−1 , ..., h0 is the
sequence of cavity fields on the shortest path from S0 to Sr (we are sure that there exists just one shortest path
because we are locally on a tree). To check whether or not the susceptibility defined in eq. (80) diverges, one should
thus simply analyze the square of the derivative of the belief propagation equations, as we have done in eq. (79). The
conclusion is that the critical temperature is given by the condition that a perturbation would grow, so that
1
ER
√
βc = atanh
.
(83)
c
Given the branching ratio is c − 1 for a random regular graph, we get in this case:
1
.
βcreg = atanh √
c−1
(84)
It follows that for β < βc the state of the Viana-Bray model is correctly described by the paramagnetic fixed point.
For β > βc the Bethe approximation fails and for the correct physical description the so-called replica symmetry
breaking framework needs to be applied [211]. Since, in this review, we are concerned by the Bayes-optimal inference
that effectively lives on the Nishimori line and does not need the concept of replica symmetric breaking, we refer the
reader interested in this to the textbook [217].
D.
Phase diagram of the planted spin glass
Now we compute the phase diagram of the planted spin glass, where the interactions J are chosen conditionally on
the planted configuration S∗ , following eq. (34). In this case the Bethe approximation and its linearization are still
described by (72) and (78). In the planted model the stability analysis of the paramagnetic fixed point is slightly
different, since the distribution from which the interactions Jij are taken is not random but depends on the planted
configuration. This is where our mapping to the Nishimori ensemble is making the problem rather easy to analyze
anyway. Indeed, according to our analysis in the last section, we simply need to ask whenever, on the Nishimori
∗
eβ
line with a bias ρ = 2 cosh
β ∗ , the system acquires a spontaneous magnetization. Or more simply, we just need to
compute the phase diagram when a ferromagnetic bias is present. Taking the average of (78) and counting properly
the probability of Jij given Si∗ and Sj∗ we immediately see that, while the spin glass transition is unchanged, the
ferromagnetic one is now given by
hut i = c(tanh β ∗ )(tanh β) hut−1 i .
(85)
According to this computation, the system thus starts to polarize towards the ferromagnetic configuration when
1
β > βF = atanh
.
(86)
c tanh β ∗
36
For the Bayes-optimal case (or equivalently, on the Nishimori line) when β ∗ = β, this immediately gives the transition
from the no-detectability phase to the recovery phase that we gave in eq. (43). We see now how easy it was to compute
it using the cavity method and the Nishimori mapping.
A couple of subtle but interesting points can be made about the phase diagram, always keeping in mind the fact
that the inference problem can be seen as looking for the ferromagnetic state on the Nishimori line.
• As β ∗ → ∞ (ρ = 1) we recover the ordinary ferromagnetic Ising model with its well known critical temperature.
√
• It is interesting to observe that there a multi-critical point at β = atanh(1/ c), where all the three conditions
(83), (86) and Bayes-optimality, β = β ∗ , meet. This point, of course, is on the Nishimori line, and is precisely
the transition point eq. (43).
• For the Bayes-optimal case, β ∗ = β, for all temperature higher than the transition point β ∗ < βc , the paramagnetic fixed point is the only fixed point of the recursion (72). Since the paramagnetic and annealed free
energies are equal, this is a case where the planting is quiet, as defined in section II E 3. This means that in
the whole region β ∗ < βc the planted ensemble behaves exactly the same as the randomly-quenched ensemble.
This is indeed reflected in the phase diagram in Fig. 2, where everything on the left of the multi-critical point
is independent of the value of ρ.
• Out of the Bayes-optimal case, i.e. when β ∗ 6= β, there exists at very low temperature (large β) a so-called
mixed phase where equilibrium configurations are correlated to the planted one (the ferromagnetic after our
mapping) but where the equilibrium phase is actually a spin glass one.
The phase diagram of the planted spin glass is summarized in Fig. 2 where the boundary between the paramagnet
(P) and the spin glass (SG) is plotted in green, between the paramagnet (P) and ferromagnet (F) in blue, and the
Nishimori line corresponding to the Bayes-optimal inference in red. We also sketch the position of the mixed phase
(M).
E.
Systems with first order phase transitions
In the above planted Ising spin glass model the underlying phase transition is of second order. In that case there
is no particular computational issue. Either inference is impossible, or it is possible and tractable. The hard phase
outlined in section (I B 5) is absent.
The picture is richer in problems when a discontinuous phase transition is found in the randomly-quenched ensemble.
Famously, this is the case in the low density parity check error correcting codes as described e.g. in [217, 238], and
has been analyzed also in details in the pioneering work of Nishimori and Wong [241]. The examples of clustering of
networks and compressed sensing (to which we dedicate sections V and VI) also fall in this category.
In order to observe such a phenomenon, we have to change the Hamiltonian eq. (30) and move from two-body
interactions to three-body interactions. This leads to the so-called 3-spin spin glass model, which we already mentioned
in sec. II H 2 under the name 3-XOR-SAT [114, 216, 265]. The Hamiltonian now reads
X
H(S) = −
Jijk Si Sj Sk ,
(87)
(ijk)∈E
Again, we can create this model on a random hyper-graph. For instance, we can consider a L regular hyper-graph
where each variable belongs to exactly L triplets, chosen randomly at random. Again we can study this model under
a planted setting, or equivalently, on the Nishimori line. What is changing with respect to the preceding section?
First of all, we need to rewrite the belief propagation equation. It is an exercise (see for instance [216, 265] or the
appendix A in [323]) to see that they now read
X
Y
mi→a = tanh β
Jb
mj→b
(88)
b∈∂i\a
j∈∂b\i
where a, b denote the triplets of interactions. Repeating the analysis of the previous section, one can obtain the phase
diagram which we show in Fig. 5.
At first sight, this phase diagram looks really similar to the one from Fig. 2. What has changed now is the fact
that the transition is a first order one. Let us first concentrate on the Nishimori line. As soon as T < Tspin there
are two fixed points of BP, a paramagnetic and a ferromagnetic one, and the one dominating the measure is the
37
2
Tspin.
P
T=1/β
1.5
TF
Tdyn.
1
TSG
0.5
SG
F
0
0.5
0.6
0.7
0.8
ρ=eβ*/2cosh β*
0.9
1
FIG. 5.
Phase diagram of the 3-XOR-SAT (or 3-spin glass) problem as a function of temperature T = 1/β and density
of ferromagnetic bonds ρ. This phase diagram is for random regular hyper-graphs of degree L = 5. The Nishimori line,
ρ = eβ /(2 cosh β), corresponds to the Bayes-optimal inference of a planted configuration. As in Fig. 2 there is the spin glass
(SG) phase where the system is in a glassy state, the ferromagnetic (F) phase, and the paramagnetic (P) phase (for simplicity,
the mixed phase is not shown). A major difference with respect to Fig. 2, however, is the fact that the ferromagnetic transition
is now a first order one, with a spinodal line. The spin glass transition is also a first-order one, with the so-called dynamic
transition happening before the static transition.
one with lower free energy. Only when T < TF does the free energy of the ferromagnetic fixed point become lower
than the paramagnetic one. This means that, for T < TF , a perfect sampling would give a configuration with a
positive magnetization. Using the inference-Nishimori connection, this means that inference is possible for T < TF ,
and impossible for T > TF : the equilibrium phase transition marks the information theoretic transition.
Now, let us try to actually perform inference in practice, for instance by running belief propagation or a MonteCarlo algorithm. Unfortunately, we have no idea about what is the planted configuration so at best we can initialize
randomly. The problem is that this will put us in the paramagnetic state, which exists and is locally stable at all
positive temperatures. Both BP or MCMC will simply explore this paramagnetic state, and, in any sub-exponential
time, will miss the existence of the ferromagnetic one. This is the classical picture of a first-order phase transition in
a mean-field model. Because of the first order transition, we are trapped into a high free energy minima that is stable
at all temperatures, and therefore, we cannot perform inference. While we should be able to find a configuration
correlated with the planted one (in the sense that it is information-theoretically possible), we cannot do it within the
class of existing polynomial algorithms. Related ideas, in the context of the glass transition, were originally present
in [95]. This illustrates the difference between a “possible” inference and an “easy” one in the sense of section I B 5.
XOR-SAT is actually a little bit special in the sense that the easy detectable phase (see below) is missing. There
is a spinodal transition (Tspin ) where the ferromagnetic phase becomes unstable, but not a second one where the
paramagnet become unstable. In most models, however, if the temperature is lowered a second spinodal is reached,
the paramagnetic phase becomes eventually unstable and inference is possible.
Let us thus discuss another example where such a transition arises. For this, we turn to the planted coloring
problem [182]. In planted coloring one first chooses a random assignment of q colors out of the q N possible ones.
Then, one puts M = cN/2 edges at random among all the edges that do not have the same color on their two
adjacent nodes. The Bayes optimal inference in this setting evaluates marginals over the set of all valid colorings of
the created graph. Computation of these marginals builds on previous works [234, 320], where the problem of coloring
random graphs was studied, and goes as follows: One considers belief propagation initialized in two different ways.
In the random initialization the messages are chosen at random but close to the paramagnetic ones. In the planted
initialization the initial values of the messages are set to point in the direction of the planted configurations. A fixed
point is reached from both these initializations and its free energy is evaluated. One obtains the following picture,
illustrated in Figs. 3 and 6 for planted 5-coloring
• Undetectable paramagnetic phase, for average degree c < cd . The planting is quiet and belief propagation
converges to the uniform fixed point both from random and planted initialization. We will call cd the dynamical
or the clustering phase transition [179, 182].
• Undetectable clustered phase, or dynamic one-step replica symmetry breaking (d1RSB) phase, for average degree
cd < c < cc . The planting is still quiet, but the set of solutions is split into exponentially many clusters that
are uncorrelated and all look the same as the one containing the planted solutions. We will refer to cc as the
detectability phase transition. In the randomly-quenched ensemble cc is the condensation transition [179, 182].
1
0.5
0.8
0.4
overlap, planted init
overlap, random init
fpara-fplanted
0.6
0.3
0.4
0.2
cd
cs
0.2
fpara-fplanted
normalized overlap
38
0.1
0
0
cc
12
13
14
15
c
16
17
18
FIG. 6. The phase transitions for planted 5-coloring. The overlap (normalized in such a way that for a random configuration
it is zero) achieved by BP initialized in the random and planted configuration. The difference between the corresponding free
energies is also plotted. Figure taken from [71].
• Hard detectable phase, for average degree cc < c < cs . This is arguably the most peculiar phase of the four. The
Bayes optimal marginals are correlated to the planted configuration, however, there is a metastable paramagnetic
phase to which belief propagation converges from a random initialization. A ghost of the quiet planting is visible
even for c > cc in the sense that the space of configurations looks exactly like the one of the randomly-quenched
ensemble except for the planted cluster.
• Easy detectable phase, for average degree c > cs . Belief propagation finds a configuration correlated to the
planted coloring. The phase transition cs can be located doing a local stability analysis of the uniform BP fixed
point, along the same lines as done for the planted spin glass in section III D. We will refer to cs as the spinodal
or the hard-easy phase transition. In the randomly-quenched ensemble, cs is the point starting from which belief
propagation stops converging, related to the local instability towards the spin glass phase [182, 320].
A careful reader will recognize that this corresponds to the situation we have described in the introduction when
we draw figure 1. The situation in this figure is that of a problem with a first order transition, where the possible
inference is defined by the equilibrium transition, and the easy/hard transition is defined by the presence of a spinodal
marking the instability of the paramagnet towards the planted state. This is arising in learning problems as diverses
as the planted perceptron, community detection with 4 or more communities, compressed sensing, error correction,
etc. If, however, the transition is a second order one, then these two transition coincides as in the case of the planted
spin glass (or, as we shall see, community detection with two equal sizes communities). We find it remarkable that
the century old classification of phase transition in physics, introduced by Paul Ehrenfest in 1933 [138], turns out to
be the right tool to discuss a deep phenomena of statistical inference and computational complexity.
F.
Comments on the phase transitions in various problems
One more question we want to clarify is whether the concept of quiet planting works for every inference problem.
Here the answer is both yes and no. First of all, for quiet planting to work, we need to be in a setting where belief
propagation has the uniform (sometimes confusingly called factorized) fixed point, where the messages in the fixed
point do not depend on the indices of the edge. If such a fixed point exists then the annealed free energy density equals
the quenched one [228], and consequently the argument about planting being quiet works. There are many problems
where BP on the randomly-quenched ensemble does not have a uniform fixed point, the most studied examples are
perhaps the problem of K-satisfiability (K-SAT), or the Ising model in a random external magnetic field hi .
Even in problems such as K-SAT, where canonical BP does not have a uniform fixed point [218], we can introduce
a reweighted Boltzmann measure, as done in [178], in such a way that the corresponding BP fixed point is uniform.
Planting according to this reweighted measure then leads to instances that are contiguous to the random ones. Note
39
that this reweighting and corresponding planted instances were used in the proof of [5, 6] and studied also in [23, 144].
The planted solution is not an equilibrium solution with respect to the canonical Boltzmann measure but with respect
to the reweighted measure.
Another comment is related to the following natural question: How to distinguish between problems where the
transition is continuous or discontinuous? And when it is discontinuous, how wide is the hard phase? Is there some
criteria simpler than performing the whole analysis? To distinguish between the first and second order transition, we
do not know of any other way. We only have a couple of rules of thumb - in Potts model with sufficient number of
states continuous transitions become discontinuous, and in models where more than two variables interact, e.g. the
p-spin model or K-SAT, the continuous transition becomes discontinuous for sufficiently large K and p [179]. But
there are binary pairwise problems such as the independent set problem where at sufficiently large average degree the
continuous transition also becomes discontinuous [17].
Interestingly, one can identify a class of constraint satisfaction problems (CSP) where the hard phase is very wide.
In [178] these problems were identified as those with cs → ∞. Previously, only XOR-SAT was known to have this
property [95], but in [89] a whole hierarchy of such problems was identified and is related to k-transitivity of the
constraints. A constraint is k-transitive if and only if the number of satisfying assignments compatible with any given
assignment of any k variables is the same. For all CSPs that are at least 2-transitive the hard phase extends over the
whole region of linearly many constraints, equivalently cs → ∞. Calling r the smallest number such that a constraint is
not r-transitive, the number of constraints predicted to render the problem tractable scales as N r/2 [89]. Interestingly,
this classification makes the K-XOR-SAT the hardest of all CSPs (despite the fact that it can be solved by Gaussian
elimination, which is not robust to any kind of noise). The notion of 2-transitivity, and consequently infinitely large
hard phase, extends also to continuous variables, and as it was recently pointed out, renders tensor factorization
and completion very difficult [266]. Physically, the corresponding systems are such when the paramagnetic phase is
locally stable down to zero temperature [95]. Examples of such systems indeed include the Ising p-spin (related to
K-XOR-SAT) or the spherical p-spin model related to tensor factorization.
Having examples where the hard phase is huge is very useful when thinking about formal reasons of why this phase
is hard. In [89] the authors proved hardness for a class of the so-called statistical algorithms using precisely this kind
of constraint satisfaction problems. Arguably, the class of statistical algorithms is too restrictive, but we believe that
the instances of CSP with such a large hard phase should be instrumental for other attempts in this direction. We
shall summarize in section IV B arguments for why we think there is something fundamentally hard about this phase.
In [324] we discussed the example of planted locked constraint satisfaction problems (CSP). These are particularly
interesting for a couple of reasons. First of all, the valid solutions are separated by extensive Hamming distance from
each other and therefore BP either converges to a paramagnet or points exactly towards the planted configuration.
Note that low-density parity check (LDPC) error correcting codes have precisely this property. Locked CSP can be
thought of as non-linear generalization of LDPC. A second reason why locked CSPs are interesting is that in some
cases the hard phase is very wide and moreover the planted assignment is with high probability the only one (perhaps
up to a global flip symmetry). The planted locked CSP are hence attractive candidates for a one-way function that
could find applications in cryptography. Having hard CSP instances that have exactly one valid assignment was used
in studies of quantum annealing where having non-degenerate ground state is greatly advantageous, see e.g. [127].
40
IV.
FROM PHYSICS INSIGHT TO NEW ALGORITHMIC IDEAS
In the introduction, sec. I B, we presented two main questions about inference that we aim to answer. One more
fundamental, about phase diagrams, phase transitions and related insights, the second more practical, about the
development of algorithms. This section is about algorithmic contributions obtained using the insight resulting from
the statistical physics analysis. For clarity, all is again illustrated on the planted spin glass model.
With the work of Mézard, Parisi and Zecchina on survey propagation [215], the statistical physics community
learned that the replica and cavity calculations are as powerful when used as algorithms as they are when used as
tools for analyzing the phase diagram. The connection between the Bethe approximation and belief propagation was
noticed earlier in connection to error correcting codes by Kabashima and Saad [154], and in a more general setting
by [318]. But looking at the volume of related literature, it is only after survey propagation that the whole statistical
physics community started to look for explicit algorithmic applications of the Bethe approximation.
Here we review some recent physics-related contributions to algorithmic development for Bayes-optimal inference.
A.
Non-backtracking spectral method
Notice, within the historical perspective, that the Bethe solution for the spin glass on a random graph was written
by Bowman and Levin [44], i.e. about 16 years before realizing that the same equation is called belief propagation
and that it can be used as an iterative algorithm in a number of applications.
The paper by Thouless [299], who analyzed the spin glass critical temperature, implicitly includes another very
interesting algorithm, that can be summarized by equation (78) where BP is linearized. The fact that this equation
has far-reaching implications when viewed as a basis for a spectral algorithm waited for its discovery even longer.
Notably until 2013 when the methodologically same algorithm was suggested for clustering of networks [180]. This
work will be discussed in sec. V D. We now turn back to our pedagogical example, the ±J planted spin glass, and
related spectral algorithm as described and analyzed in [274]. Related ideas also appeared in [325].
First note that a large class of inference algorithms is based on a spectral method of some kind. Probably the most
widely used is the so-called principal component analysis (PCA). Its operation can be thought of as revealing the
internal structure of the data in a way that best explains the variance in the data. In PCA the information is contained
in singular vectors (or eigenvectors for symmetric matrices) corresponding to the largest singular values. Resulting
spectral algorithms are used everyday to solve numerous real world problems. The same strategy of computing
leading singular (or eigen-) vectors of some matrix related to the data is applied in many settings including the
planted partitioning problem closely related to the planted spin glass model [64, 65], or in understanding financial
data [42, 190].
A natural question that arises is whether a spectral algorithm can find a configuration correlated to the planted
assignment in the planted spin glass in the whole region β ∗ > βc where belief propagation can asymptotically do so.
Facing this question, many experts would probably suggest to take the matrix Jij (with Jij = 0 when (ij) ∈
/ E) or
its associated Laplacian (perhaps normalized in some way) and compute the leading eigenvector. This is indeed a
good strategy, but on sparse Erdős-Rényi random graphs it does not work down to the threshold βc . The problem is
that for an Ising model living on a large sparse Erdős-Rényi graph the tail of the spectra of the commonly considered
matrices is dominated by eigenvectors localized on some small subgraphs (e.g. around high-degree nodes or on some
kind of hanging sub-trees).
The strategy in designing spectral method that does not have the problem with localized eigenvectors (at least
on a much larger class of graphs) is to try to mimic what BP is doing when departing from the paramagnetic fixed
point. Eq. (78) is precisely the linearization of BP around that fixed point and suggests a definition of a so-called
non-backtracking matrix B defined on directed edges as
Bi→j,k→l = Jij δil (1 − δjk ) .
(89)
Denoting M the number of edges in the non-directed graph, this is a 2M × 2M matrix. The term (1 − δjk ) means
that a walk that follows non-zero elements of the matrix is not going back on its steps, this motivates the name of
the matrix. Building on [40, 180], it was proven in [274] that the planted configuration is indeed correlated with the
signs of elements of the eigenvector of B corresponding to the largest (in module) eigenvalue as soon as β ∗ > βc .
Fig. 7 depicts the spectrum on the matrix B (89) for a planted spin glass model on an ER graph of average degree
c and having N = 2000 nodes. The Jij = ±1 were generated following the procedure described in sec. II A with
β ∗ = 0.55. We observe (the proof is found in [274]) that on sparse random graphs
√
• For β ∗ < βc = atanh(1/ c), corresponding to the paramagnetic phase of the planted
√ spin glass model: The
spectrum of B is the same as it would be for β ∗ = 0, bounded by a circle of radius c.
41
√
• For β ∗ > βc = atanh(1/ c), corresponding to the ferromagnetic phase of the planted spin glass model: The
spectrum of B has the same property as in the previous case, except for two eigenvalues on the real axes. One
of them is out of the circle and has value c tanh β ∗ . Its corresponding eigenvector is positively correlated with
the planted assignment.
The performance of the non-backtracking spectral algorithms will be further discussed and compared to other
algorithms for the example of clustering of networks in sec. V D. In sec. V G we will also illustrate how this method
performs on some non-random graphs. Just as all the other spectral algorithms, this method is not robust to adversarial
changes in a small part of the graph (such as planting a small dense sub-graph). This can hence be a disadvantage
with respect to methods that optimize some global cost function that is not affected by small local changes. We,
however, illustrate in Fig. 13 that spectrum of the non-backtracking matrix of many real graphs does not deviate
excessively from its behavior on random graphs.
4
4
3
c=3
3
2
2
1
1
0
-4
-3
-2
-1
c=8
0
0
1
2
3
4 -4
-3
-2
-1
0
-1
-1
-2
-2
-3
-3
-4
-4
1
2
3
4
FIG. 7. Spectrum of the non-backtracking operator
√ for the ±J planted spin glass model. Right: In the undetectable phase
where all eigenvalues are inside a circle of radius c. Left: In the detectable phase where one eigenvalue is outside the circle
on the real axe, its value is c tanh β ∗ . Figure taken from [274].
B.
On computational hardness of the hard phase
In section III E we described the four phases that we observe in the Bayes optimal setting of inference problems.
The most theoretically intriguing is the hard detectable phase that appears between the detectability transition cc ,
and the spinodal transition cs . In this phase both Monte Carlo and belief propagation, that aim to sample the Bayesoptimal posterior distribution, fail and get blocked in the metastable paramagnetic phase. The spinodal transition
cs is also a barrier for all known spectral algorithms. Some of the spectral methods including the one based on the
non-backtracking matrix saturate this threshold.
There are many other problems where a phase of the same physical origin was discovered. Among them the
best known are the LDPC error correcting codes [217] and the planted clique problem [9, 75, 142, 148]. Some
recent works prove hardness in other problems assuming the hardness of planted clique (in the corresponding region)
[31, 47, 118, 197]. There is a lot of current interest in this kind of results.
When a physicist sees a phenomenon as the one above she immediately thinks about universality and tends to
conjecture that the spinodal transition in inference problems will be a barrier for all polynomial algorithms. There is,
however, the case of planted XOR-SAT that presents the same phenomenology as described in section III E. Moreover,
the hard phase in planted XOR-SAT extends to infinity (the paramagnetic state is always locally stable). The work
of [89] actually suggests that the right scaling for the number of clauses at which planted K-XOR-SAT becomes
easy is when the number of clauses is at least O(N K/2 ). Yet, thanks to its linearity planted K-XOR-SAT can be
solved in polynomial time by Gaussian elimination. Hence the above conjecture about the hard phase is false. But
the physicist’s intuition is rarely entirely wrong, and in those rare cases when it is indeed wrong it is for absolutely
fundamental reasons. In our opinion, the question:
In which exact sense is the hard phase hard?
42
is the most fundamental question related to the interface between statistical physics, statistical inference and average
computational complexity.
A line of research that should be explored, and seems quite accesible, is proving or disproving that there is no way
to use belief propagation or Monte-Carlo sampling with a mismatched prior or model in order to improve the position
of the spinodal transition.
Let us also note here that the spinodals of first order phase transitions in Bayes-optimal inference problems create
a much cleaner algorithmic barrier than phase transitions described in optimization or in constraint satisfaction
problems such as K-satisfiability. In K-SAT the idea of hard instances being close to the satisfiability threshold goes
back to the works by [62, 219, 280]. However, even after more than 20 years of study it is still not clear what is the
limiting constraint density up to which polynomial algorithms are able to find satisfiable solutions in random K-SAT
(even at very large K). For a related discussion see [321].
C.
Spatial coupling
Spatial coupling is, from a physics point of view, a very natural strategy to beat the hard-easy spinodal transition
in systems where the structure of the graphical model can be designed. It is closely related to nucleation in physics.
An exponentially long living metastability is only possible in mean-field systems. Let us describe here why it is not
possible in systems embedded in finite Euclidean dimension d. Metastability means that there are two competing
phases. Let us call ∆f the free energy difference between them. Let us consider a large system in the metastable
state with a droplet of radius L (small compared to the size of the whole system) that by random fluctuation appears
to be in the stable state. Creating such a droplet will cost some energy due to its interface where the two phases do
not match. This is quantified by the surface tension Γ. The total energy balance can be written as
E(L) = ΓLd−1 − ∆f Ld .
(90)
Physical dynamics can be roughly thought of as gradient descent in this energy. Notice that E(L) is increasing for
L < L∗ = Γ(d − 1)/(d∆f ) and decreasing for L > L∗ . This means that if fluctuations created a droplet larger than
the critical size L∗ it will continue growing until it invades the whole system. The above argument works only in
geometries where the surface of a large droplet is much smaller than its volume. On random sparse graphs or fully
connected geometries it does not work. This is why nucleation does not exist in mean-field systems. Notice also that
nucleation is simpler for lower dimensions d, the lowest meaningful case being d = 1.
We can conclude that in order to avoid long living metastability we need to work with finite-dimensional systems.
But the situation is not so simple either. In systems where the mean-field geometries are used, they are used for
a good reason. In LDPC codes they assure large distance from spurious codewords, in compressed sensing random
projections are known to conserve well the available information, etc. To avoid metastability we need to mix the meanfield geometry on a local scale with the finite dimensional geometry on the global scale. This is exactly what spatial
coupling is doing. Originally developed for LDPC error correcting codes [188, 192] and closely related to convolutional
LDPC that have longer history [90], spatial coupling has been applied to numerous other settings. Again we will
discuss some of them in more details in section VI E. To introduce the concept here, we restrict ourselves to the
spatially coupled Curie-Weiss model as introduced in [121] and studied further in [49].
The Curie-Weiss (C-W) model in a magnetic field is arguably among the simplest models that present a first order
phase transition. The C-W model is a system of N Ising spins Si ∈ {−1, +1}, that are interacting according to the
fully connected Hamiltonian
HC−W (S) = −
N
X
J X
Si Sj − h
Si ,
N
i=1
(91)
hi,ji
where the notation h·, ·i is used to denote all unique pairs, J > 0 is the ferromagnetic interaction strength and h ∈ R
is the external magnetic field.
Let us briefly remind what is obvious to everybody trained in statistical physics. For positive magnetic field h > 0
the equilibrium magnetization m(J, h) > 0 is positive, and vice versa. There exists a critical value of the interaction
strength Jc = 1 such that: for J < Jc the limh→0+ m(J, h) = limh→0− m(J, h) = 0, and for J > Jc we have
limh→0+ m(J, h) > 0 > limh→0− m(J, h). The latter is a first order phase transition, in the “low temperature” regime
J > Jc the system keeps a non-zero magnetization even at zero magnetic field h. There exists a spinodal value of the
magnetic field
!
r
p
J −1
,
(92)
hs (J) = J(J − 1) − atanh
J
43
Chain of
C-W
systems
(w = 2)
1
=80
=160
=240
=320
=400
=480
=560
=640
0.5
z
m
+1
−1
+1
−1
+1
Single C-W
system
(n = 8)
-0.5
-1
−1
+1
0
+1
0
50
100
150
200
z
250
300
350
400
FIG. 8. Left: A schematic graphical representation of the spatially coupled C-W model. A chain of C-W models interacting
within a certain range w (w = 2 in the figure). In the zoomed part the fully-connected structure of the single C-W model is
shown. A connection between C-W models along the chain indicates connections between all spins contained in both models.
Right: The propagating wave obtained by Monte-Carlo heat-bath simulation for the following parameters: J = 1.4, L = 400,
n = 100, h = 0.05, w = 5, wseed = 10 and hseed = 0.3. Simulations were performed using random sequential update and
therefore time must be rescaled by τ = t/N . The figure is taken from [49].
with the following properties: If the magnetizations are initialized to negative values and the magnetic field is of
strength 0 < h < hs (J), then both local physical dynamics and local inference algorithms, such as the Monte-Carlo
sampling, will stay at negative magnetization m− (J, h) < 0 for times exponentially large in the size of the system.
The spinodal value of the magnetic field hs (J) acts as an algorithmic barrier to equilibration, when initialized at
Si = −1 for all i = 1, . . . , N , and hence to successful inference. For h > hs (J) it is, on the other hand, easy to reach
the equilibrium magnetization m+ (J, h).
D.
The spatially coupled Curie-Weiss model
We introduce here the construction from [121] that might not appear very transparent at first. But its idea is really
simple. We create a one-dimensional chain of C-W models, put a nucleation seed somewhere and let the sub-systems
interact in order to ensure that the nucleation proceeds to the whole chain.
As illustrated in Fig. 8, we consider a one-dimensional chain of (2L + 1) C-W systems, where each of the C-W
system has n spins (referred to as a “block”) and is labelled by the index z ∈ {−L, . . . , L}. The result is that a
configuration S of the full system is now given by the values of N = n(2L + 1) spins, each labelled by a compound
index:
S = {Siz ∈ {+1, −1} : i ∈ {1, . . . , n} , z ∈ {−L, . . . , L}} .
(93)
In a similar way, the uniform external magnetic field h for a single system is replaced by an external field profile hz .
As far as the coupling is concerned, every spin not only connects to all spins in the same location z but also all spins
within w blocks from z. The corresponding Hamiltonian is then
L
n
X
X
1 X
0
0
Hn,L (S) = −
Jzz Siz Sjz −
hz
Siz .
n
0
i=1
hiz,jz i
(94)
z=−L
The couplings between spins are Jzz0 = Jg (|z − z 0 |/w) /w, where the function
g satisfies the following condition
P+∞
1
−1
g (|x|) = 0, ∀ |x| > 1 and we choose its normalization to be w z=−∞ g w |z| = 1.
In order to ensure that the system nucleates, we must increase the field hz at some point on the chain. We choose
the magnetic field profile in such a way that hz = hseed > hsp in some small region given by z ∈ {0, . . . , wseed }.
Everywhere else, hz = h hsp , such that the average field strength is still small. To illustrate the nucleation in Fig. 8
we initialize to Siz = −1 everywhere and let the system evolve under Monte-Carlo dynamics. We indeed observe a
nucleation “wave” that invades the whole system in a time proportional to the number of blocks. The system can be
designed in a such a way that the spinodal regime entirely disappears [121].
44
E.
A brief history of time indices in the TAP equations
In physics of disordered system, the historically first form of iterative procedure that was used to compute marginals
(local magnetizations) in a densely connected spin glass model are the Thouless-Anderson-Palmer (TAP) equations
[300]. In the early statistical physics works on error correcting codes TAP equations were used for decoding. The
primary goal of the paper by Kabashima and Saad [154], who pointed out the connection between belief propagation
and the Bethe approximation, was to compare the performance of TAP and BP. In dense systems the fixed points of
both BP and TAP agree. However, it was reported e.g. in [150, 151] that TAP equations did not converge in some
cases even when BP does and there is no replica symmetry breaking. With the renewed interest in message passing
of dense systems that followed the work of Donoho, Maleki and Montanari on compressed sensing [81], it is definitely
worth making some up-to-date remarks on the TAP equations and their convergence.
The fixed point of the TAP equations
N
N
X
X
2
mi = tanh βh + β
Jij mj − β 2 mi
Jij
(1 − m2j )
(95)
j=1
j=1
describes the replica symmetric marginals of the following fully connected spin glass Hamiltonian
X
X
HSK (S) = −
Jij Si Sj − h
Si ,
(96)
i
hi,ji
˜
where Si are the Ising spins,
√ and Jij are independent random variables of zero mean and variance J/N . Hence a
typical value of Jij = O(1/ N ).
Fixed points of the TAP equations are stationary points of the corresponding TAP free energy [300]. Derived as
such, TAP equations do not include any time-indices and were mostly iterated in the simplest form of having time
t − 1 on all magnetizations on the right hand side, and t on the left hand side. At the same time, as we said above,
a number of authors reported that the TAP equations have problems to converge even when BP does converge and
replica symmetry is not broken [150, 151]. No satisfactory explanation for this existed in the literature until very
recently [39]. The reason behind this non-convergence was that the time indices were wrong and when done correctly,
convergence is restored, and on sufficiently dense systems TAP equations behave as nicely as BP.
For the Ising spin glass this was first remarked by Bolthausen [39], who proved the convergence of TAP with the
right time indices in a non-zero magnetic field (and out of the glassy spin glassy phase). His paper inspired the proof
technique of [26] but otherwise passed quite unnoticed. In the context of linear estimation, as discussed in Sec. VI,
the analog of the TAP equations is the approximate message passing algorithm (AMP) [82] (but see also [155]). AMP
is often thought of as the large degree limit of belief propagation and in that way of thinking the correct time indices
come up naturally. Here we re-derive TAP for the Ising spin glass as a large degree limit of BP to clarify this issue
of time indices.
P
Let us go back to the belief propagation for sparse graphs eq. (72). Define messages mti→j ≡ tanh (h + β k6=j uk→i
t−1 ).
√
Recall that Jij , and hence also u, is of order O(1/ N ) and expand to the leading order. BP then becomes
X
t−1
mti→j = tanh βh + β
Jki mk→i
.
(97)
k6=j
In this dense-graph message passing there is not yet any surprise concerning time indices. The TAP equations are
closed on full magnetizations (not cavity ones) defined as
!
X
t−1
t
mi = tanh βh + β
Jki mk→i .
(98)
k
In order to write the TAP equations we need to write messages in terms of the marginals up to all orders that
contribute to the final equations. We get by Taylor expansion with respect to the third term
!
mti→j = tanh βh + β
X
k
t−1
Jki mt−1
k→i − βJij mj→i
≈ mti − βJij mjt−1 [1 − (mti )2 ] ,
(99)
45
which leads to the iterative TAP equations
X
X
2
mti = tanh βh + β
Jij mt−1
− β 2 mt−2
Jij
[1 − (mjt−1 )2 ] .
j
i
j
(100)
j
Observe the time index (t − 2) before the last sum, which is not intuitive on the first sight. These equations are the
ones written and analyzed by Bolthausen [39].
In his proof, Bolthausen showed that the behavior of these iterative equations can be followed using what can
be called the state evolution of the TAP algorithm, at least outside of the spin glass phase (the algorithm does not
converge inside the spin glass phase). It works as follows: we consider the interactions Jij to have zero mean and
˜ . Going back to eq. (98), the argument of the tanh is a sum of independent variables (by the assumptions
variance J/N
of belief propagation) and thus follows a Gaussian distribution. Denoting m and q the mean and the variance of this
distribution of messages, two self-consistent equations can be derived
Z
p
2
1
e−z /2 tanh βz J˜ q t−1 + βh dz
(101)
mt = √
2π
Z
p
2
1
e−z /2 tanh2 βz J˜ q t−1 + βh dz
qt = √
(102)
2π
The fixed point of these equations provides the well known replica symmetric solution for the magnetization and the
spin overlap in the Sherrington-Kirkpatrick model, as was first derived in [285]. Since this is the correct solution
outside the spin glass phase (above the so-called dAT line [69], it is reassuring to see that we have a convergent
algorithm that will find the marginals fast in the densely connected case. We will return to this observation in section
VI C when we will discuss inference in densely connected models.
Note also that eqs. (101-102) correspond to a parallel update where all magnetizations are updated at the same
time. Very often, it has been found that such update might be problematic for convergence and that one should
update spins one at a time for better convergence properties. This, will also be discussed in VI C when discussing the
AMP algorithm [201]. In that case one chooses randomly at each time step which spin should be updated, then one
can still use eq. (100) with a careful book-keeping of the time-indices. The state evolution eqs. (102) become in that
case differential equations in time. Instead of q t+1 = f (q t ) as in eqs. (102) we have now dq/dt = f (q t ) − q t in the
continuous time limit. We show an exemple of the behavior of these two approaches in Fig. 9. While in this case the
parallel update was faster, we will see that sometimes the other one could be beneficial because it is smoother and
avoids e.g. the overshoot we see in the first iteration of the parallel update.
TAP with parallel update
TAP with random sequential update
0.8
0.7
0.7
0.6
0.6
0.5
0.5
q(t)
q(t)
0.8
0.4
0.3
0.4
0.3
Algorithm
0.2
Algorithm
0.2
State evolution
State evolution
0.1
0
0.1
0
5
10
t
15
0
0
5
10
15
t
FIG. 9. State evolution versus actual simulation for the TAP equations in the Sherrington-Kirkpatrick with parallel update
(left) and random sequential one (right), for a spin glass system with N = 5000, 1/β = T = 3, J˜ = 1 and h = 4. The variance
q of the distribution of messages is shown. In both cases, the state evolution equations (with the iterative form on the left and
the differential one on the right) perfectly describe the evolution of the algorithm.
These considerations stems from the derivation of TAP equations starting from the BP equations for random
interaction. To finish this section, we point out that there have been a large body of work in trying to write TAP
46
equations for interaction that are not random. We refer the interested reader to the book collection [244] for detailed
discussion of the TAP equations in this context. Along this line of thoughts, the adaptive TAP approach [245] allows
to learn the structure of the matrix J, while the structure of the Onsager term can be exactly computed depending
on the structure of the correlations (see e.g. [245, 247, 282]).
Since these approaches do not lead to a natural time ordering, it is worth mentioning that, very recently, these have
been revisited [243] in the light of the “time indices” subtlety, leading a much more satisfying theory of the dynamics
of the TAP equations on general matrices.
47
V.
CLUSTERING OF NETWORKS AND COMMUNITY DETECTION
Clustering is one of the most commonly studied problems in machine learning. The goal is to assign data points to
q clusters in a way that the cluster is somehow representative of the properties of the points inside it. About 15 years
ago, a large number of scientists got interested in the research on so-called complex networks, i.e. sets of nodes where
some pairs are connected by edges, arising in a wide range of settings. A natural question followed: How to cluster
networks? That is, how to assign nodes of a graph in groups such that the largest possible part of the edges can be
explain by a set of affinities between the clusters (instead of between the nodes themselves)? Typically, one considers
a number of clusters much smaller than the number of nodes. Part of the motivation for studying this question came
from studies of social networks which resulted in the name for this problem: community detection.
A.
Biased overview of community detection
The literature on community detection is vast and there are a number of reviews that make a great effort in putting
the existing works in chronological and merits perspective, e.g. [94]. Here we give a perspective biased towards the
approach of statistical physics of disordered systems.
In community detection we are given a graph G(V, E) of N nodes and M edges. We search to assign the nodes
into q groups, calling the assignment si ∈ {1, . . . , q}, i = 1, . . . , N . In the most commonly considered assortative
setting, the goal is that there are many edges between nodes of the same group and few edges among the groups.
Traditionally, in statistics or machine learning, this problem would be solved by a spectral relaxation. One takes
a matrix associated to the graph, its adjacency, its Laplacian or even something else, and compute its largest (or
smallest) eigenvalues and associated eigenvectors. The elements of these eigenvectors would then reveal in a visible
way the community structure [77, 91], for more details see the nice review [307].
An influential set of works on community detection is due to Mark Newman and collaborators who introduced the
so-called modularity function [63, 108, 237]
1 X
di dj
Q=
Aij −
δsi ,sj ,
(103)
2M
2M
hi,ji
where the sum is over all distinct pairs of nodes, di is the degree of node i. Community detection is performed by
finding an assignment of nodes that maximizes this modularity. Optimizing the modularity is a NP-hard problem. In
practice, however, many efficient heuristic algorithms were suggested for this task, see e.g. [94].
Modularity-based community detection has nice connections to statistical mechanics and spin glasses. Reichardt
and Bornholdt [263] pointed out that modularity maximization can be seen as a case of finding a ground state of a
particular Potts spin glass. In order to assess the statistical significance of community structure found by maximizing
the modularity, ref. [263] estimated the maximum value of modularity on a random graph, as the null model. Indeed,
on sparse random graphs the maximum value of modularity can be very large even when no community structure
whatsoever is present. To give an example, results on random graph partitioning [322] teach us that a random 3regular graph can be divided into two groups of roughly equal size in such a way that only about 11% of edges are
between the groups. This visually looks like a very good community structure with high modularity. Yet there were
no communities in the process that created the random graph. The need to compare the value of modularity to a null
model is not very satisfying. Ideally, one would like to have a method able to tell that the underlying graph does not
contain any communities if it did not.
PLet us assume that we believe there were q communities respectively taking a fraction na of nodes, a = 1, . . . , q,
a na = 1, and the probability that a given random node from community a is connected to a given random node
from community b is cab /N , where N is the notal number of nodes and cab is a system size independent q × q matrix.
There will be roughly cab na nb N edges between groups a and b, and caa n2a N/2 edges inside group a. The above very
natural assumption is called the sparse stochastic block model (SBM) [134, 314], and is commonly considered in this
setting. As Bayesian probability theory teaches us, the best possible way of inferring the hidden communities is to
enumerate all assignments of nodes into communities that have roughly the above sizes and respective number of
edges between groups. Subsequently one averages over all such assignments (factoring away the global permutational
symmetry between groups). This way one ends up with probabilities for a given node to belong to a given group. For
a truly random graph these probabilities will not be different from the fractions na .
From a computationally complexity point of view, exactly enumerating assignments as above is even harder than
finding optimal values of a cost function such as the modularity. However, from a statistical physics point of view
evaluating the above averages is equivalent to the computation of local magnetizations in the corresponding Potts
glass model. Notice, for instance, that planted coloring [182] is simply a special case of the stochastic block model
48
(caa = 0, and cab = ccd for all a 6= b, c 6= d). We shall summarize in the rest of this section the asymptotically exact
analysis of the stochastic block model [71, 72].
B.
Asymptotically exact analysis of the SBM
We explained all the methodology needed for the asymptotic analysis of the SBM on the example of the planted
spin glass in section II. In this section, we consider again the teacher-student scenario where the teacher generates
the observed graph from the stochastic block model. We denote θ = {na , cab } the set of parameters of the SBM. We
remind that na is the fraction of nodes taken by group a, a = 1, . . . , q, and cab /N is the probability that a given node
in group a connects to a given node in group b. Again we denote by θ∗ the actual values that the teacher used to
generate the graph G. The graph is represented by its adjacency matrix Aij , with N nodes and M edges. Further
si ∈ {1, . . . , q} are the Potts spins denoting the assignments to the respective groups.
The posterior likelihood of the SBM reads
P (s|G, θ) =
N
Y
Y
A
cs s 1−Aij
1
nsi
csi sj ij 1 − i j
Z(G, θ)
N
(104)
hi,ji
i=1
where the second product is over all distinct pairs, and we included a constant N −M into the partition function. We
need to evaluate the maximum-mean-overlap (MMO) estimator (22) in order to find a configuration that maximizes
the number of nodes that are assigned into the same group as the planted configuration. In a physics language, we
would rather think of the corresponding Potts model Hamiltonian
H(G, θ, s) =
N
X
i=1
log nsi +
Xh
hi,ji
cs s i
Aij log csi sj + (1 − Aij ) log 1 − i j
N
(105)
and the MMO estimator turns out to be computed from the local magnetizations of this Potts model.
When the parameters θ∗ are not known, we need to learn them according to the lines of section II D. We write
the posterior distribution on the parameters P (θ|G) and notice that it is concentrated around the maximum of the
partition function Z(G, θ). We then iteratively maximize this partition function by imposing the Nishimori conditions
N
1 X
E(δsi ,a ) ,
N i=1
X
1
=
E(δa,si δb,sj )
N na nb
na =
cab
(ij)∈E
caa =
X
2
E(δa,si δa,sj ) .
2
N na
(106)
when a 6= b ,
(107)
(108)
(ij)∈E
This learning algorithm was called expectation maximization learning in [71]. The expectations were computed using
belief propagation or Gibbs sampling as will be discussed subsequently. But even if the expectations were computed
exactly, this parameter learning algorithm depends strongly on the initial choice for the parameters θ. It is easily
blocked in a local maximum of the partition function. The size of the space of initial conditions that need to be
explored to find the global maximum is independent of N , but in practice this is a problem, and completely general
models cab are not so easy to learn. A considerable improvement can be obtained when the learning is initialized
based on the result of spectral clustering. This procedure was suggested in [326] and works nicely, in particular in
conjunction with the best known spectral clustering techniques [180] as will be discussed in sec. V D
Two algorithms that are (in the Bayes-optimal setting) asymptotically exact for computing the marginals and the
expectations in (106-108) were proposed in [71, 72]. One is the Gibbs sampling (or Markov Chain Monte Carlo),
another one is belief propagation (BP). We present here the belief propagation algorithm, as we have introduced it
in sec. III. It can also be used very directly to understand the behavior of the system and the phase transitions in
the thermodynamic limit. The BP algorithm for the SBM was first written by [124] but in that work the parameters
were fixed ad-hoc, not in the Bayes optimal way, nor by learning. Moreover, the algorithm was tested on a network
of only 128 nodes, which is too small to observe any phase transition.
Repeating the steps of sec. III, we now derive BP. For the posterior distribution (104) BP is written in terms
of the so-called messages ψsi→j
that are defined as probabilities that node i is assigned to group si conditioned on
i
49
the absence of edge (ij). Assuming asymptotic conditional independence of the probabilities ψsk→i
for k 6= j, i, the
k
iterative equations read
"
#
Y X
csi sk 1−Aik k→i
1
Aik
i→j
csi sk 1 −
ψsk
,
(109)
ψsi = i→j nsi
Z
N
s
k6=i,j
k
Pq
where Z i→j is a normalization constant ensuring that s=1 ψsi→j = 1. We further notice that because of the factor
1/N , messages on non-edges (ij) ∈
/ E depend only weakly on the target node and hence at the leading order the BP
equations simplify to
"
#
Y X
1
−hsi
i→j
k→i
ψsi = i→j nsi e
csk si ψsk
,
(110)
Z
s
k∈∂i\j
k
where we denote by ∂i the set of neighbors of i, and define an auxiliary external field as
hsi =
N
1 XX
csk si ψskk ,
N
s
k=1
(111)
k
where ψsi i is the BP estimate of the marginal probability of node i
Y X
1
,
ψsi i = i nsi e−hsi
csj si ψsj→i
j
Z
s
j∈∂i
(112)
j
with Z i being again a normalization. In order to find a fixed point of Eq. (110) we update the messages ψ i→j ,
recompute ψ j , update the field hsi by adding the new contribution, subtracting the old one, and repeat.
Once a fixed point of the BP equations is reached, its associated free energy (log-likelihood) fBP = − log Z(G, θ)/N
reads
1 X
c
1 X
log Z i +
log Z ij − ,
(113)
fBP (G, θ) = −
N i
N
2
(i,j)∈E
where
Z ij =
X
cab (ψai→j ψbj→i + ψbi→j ψaj→i ) +
a
a<b
Zi =
X
si
X
nsi e−hsi
YX
csj si ψsk→i
.
j
caa ψai→j ψaj→i
for
(i, j) ∈ E
(114)
(115)
j∈∂i sj
In [71, 72] the author argued that the above equations can be used to analyze the asymptotic performance of
Bayes optimal inference in the SBM, θ = θ∗ . The analysis goes as follow: Let us iterate BP (110) from two different
initializations:
• P
Random initialization, for each (ij) ∈ E we chose ψsi→j = na +si→j where i→j
is a small zero-sum perturbation,
s
i→j
= 0.
s s
• Planted initialization, each message is initialized as ψsi→j = δs,s∗i , where s∗i is the actual assignment of node to
groups (that is known only for the purpose of analysis).
The free energy (113) is computed for the two corresponding fixed points. The fixed point with smaller free energy
(larger likelihood) corresponds asymptotically to the performance of the Bayes-optimal inference. Here “asymptotically” is meant in the sense that the difference between the Bayes-optimal and the BP marginals goes to zero as
N → ∞. This is an important and nontrivial property, because in general the analysis of Bayes optimal inference in
such a high-dimensional setting is very difficult (sharp-P hard). Note that the asymptotic exactness of BP is so far
only a conjecture. Rigorous results showing the asymptotic exactness of BP are so far only partial, either at relatively
high average degrees [233], on in related problems for very low average degrees [224], or e.g. in ferromagnets [73].
Another property of belief propagation that makes it very convenient for the analysis of phase transitions and
asymptotic properties is the fact that BP, in a sense, ignores some part of the finite size effects. Let us compare BP
50
to Markov chain Monte Carlo methods. In the large size limit the performance and behaviour of BP and MCMC are
perfectly comparable, as demonstrated for the stochastic block model in [71]. But for finite sizes there are several
remarkable differences. Consider a set of parameters where the two initializations of BP do not give the same fixed
point. Even if we iterate BP for infinite time we will have two different fixed points. On the other hand, in Monte Carlo
Gibbs sampling for a system of size N , we will eventually find the lower free energy state (just as for every ergodic
Markov chain satisfying detailed balance). The time needed is in general exponentially large in N , but decreases
as the free energy barrier gets smaller and therefore locating the spinodal threshold precisely from finite system size
simulations is computationally more demanding compared to BP. Another manifestation of this property is that in the
paramagnetic phase on a graph with N nodes and M edges the free energy given by BP is always the same number.
Whereas the exact (finite N ) free energy typically has some fluctuations.
In the SBM, belief propagation provides an asymptotically exact analysis of Bayes-optimal inference. In the case
where the model or its parameters θ are mismatching the ones that generated the graph, this is not true in general.
The most striking example here is the MAP estimator, in other words the ground state of the corresponding Potts
model. Modularity maximization is a special case of this setting. The ground state could also be estimated by BP if we
introduced a temperature-like parameter 1/β in the posterior and tuned this temperature to zero. A related algorithm
was discussed recently in the context of modularity-based community detection in [327]. However, as we decrease this
temperature we will typically encounter a point βSG > 1 above which the replica symmetric approximation fails and
the corresponding belief propagation marginals are not any longer asymptotically exact. When this phase transition
at βSG is continuous this will manifest itself as non-convergence of the BP algorithm. Note that depending on the set
of parameters, this spin glass phase can be correlated to the orignal assignment (mixed phase) or not (pure spin glass
phase). This behavior is qualitatively the same as the one seen in the phase digram presented in Fig. 2. In the above
sense it is hence easier to compute the marginals at the Bayes-optimal value of the parameters (that may need to be
learned as described above) than to compute the ground state (or maximize the modularity) that typically is glassy.
C.
Examples of phase transitions in the SBM
The asymptotic analysis of the stochastic block model as described above leads to the discovery of phase transitions.
The most striking picture arises in the case when the parameters of the model are such that the average degree is the
same in every group. In a sense this is the hardest case for inference, because if groups differ in their average degree,
then a simple histogram of degrees gives some information about the correct assignment into groups.
Mathematically, the same average degree condition is expressed as
X
b
cab nb = c
∀b = 1, . . . , q ,
(116)
where c is the average degree. It is straightforward to see that when (116) holds, then
ψsi→j = ns
(117)
is always a (so-called uniform) fixed point (in general not the only one) of belief propagation (110).
Let us now return to the concept of quiet planting explained for the planted spin glass in sec. II E 3 and note
that all we needed for quiet planting to work was precisely the existence of such a uniform fixed point. It is hence
straightforward that the phase transitions described in sections II C and III E also exist in Bayes-optimal inference
in the stochastic block model. Let us give the example of assortative community structure with q = 4 equally sized
groups, na = 1/q for all a = 1, . . . , q, average degree c = 16, and affinity parameters such that caa = cin and
cab = cout for all a 6= b. We define a noise-like parameter = cout /cin . When = 0 the model generates four entirely
disconnected groups, and when = 1 the generation process corresponds to an Erdős-Rényi graph with no information
on the community structure. Without having any experience with phase transition and quiet planting, one might
anticipate that whenever < 1 the Bayes optimal inference will give a better overlap than a random assignment of
nodes into groups. This expectation is wrong and the phase transition that happens instead is illustrated in Fig. 10.
For other values of parameters we can observe the triplet of discontinuous phase transition discussed in sec. III E.
Actually, the planted coloring example, that was presented in Fig. 6, is a case of the stochastic block model with
q = 5 equally sized groups and extremely dis-assortative structure cin = 0.
For the general case of the sparse SBM, the order of the phase transition, and the locations of the dynamical cd
and detectability cc phase transitions, have to be analyzed case by case mostly numerically. However, the spinodal
transition has a neat analytical expression in the case of the same average degree in every group (116). This expression
is obtained as in sec. III C by analyzing the local stability of the uniform fixed point. Belief propagation (110) linearized
51
700
N=100k, BP
N=70k, MC
0.8
0.6
q=4, c=16
0.4
0.2
detectable
N=10k
N=100k
600
500
convergence time
normalized overlap
1
undetectable
400
q=4, c=16
300
200
100
0
0
0.2
0.4
0.6
ε= cout/cin
0.8
1
εc
0
0
0.2
0.4
0.6
ε= cout/cin
0.8
1
FIG. 10. Left: Normalized overlap between the MMO estimator and the original assignment to groups. We compare the
BP result to Monte Carlo Gibbs sampling. Both agree well, but we see the absence of finite size corrections in BP in the
undetectable phase. Right: The BP convergence time, illustrating the critical slowing down around the phase transition.
Figures are both taken from [71].
around this fixed point reads
i→j
=
a
q
X
X
Tab k→i
,
b
(118)
−1 .
(119)
b=1 k∈∂i\j
where the q × q matrix Tab is defined as
T ab = na
c
ab
c
The phase transition where the paramagnetic fixed point ceases to be locally stable happens when
cλ2T = 1 ,
(120)
where λT is the largest eigenvalue of the matrix T .
In case of equally sized groups in the assortative-diagonal case where caa = cin and cab = cout for a 6= b, the largest
eigenvalue of the T ab matrix (119) is λT = (cin − cout )/(cq) from which (120) implies for the spinodal phase transition
[72]
√
(121)
|cin − cout | = q c .
When the difference between cin and cout is larger, BP detects communities and gives asymptotically the Bayesoptimal overlap. When the difference is smaller BP fails. Depending on the parameters and the order of the phase
transition the detectability phase transition itself can either be given by eq. (120) or happen at strictly smaller
|cin − cout |, giving rise to the hard phase discussed in sec. IV B.
For the planted graph coloring (i.e. the extremely dis-assortative SBM, cin = 0), the detectability phase transition
is of second order for q ≤ 3 and of first order for q ≥ 4. The spinodal transition happens at cs = (q − 1)2 . In the
large q limit the dynamical and detectability phase transitions can be deduced from [320] while keeping in mind the
concept of quiet planting
cd = q(log q + log log q) + O(1) ,
cc = 2q log q − log q − 2 log 2 + o(1) .
(122)
(123)
For the assortative case cin > cout the phase transition is of second order for q ≤ 4 and of first order (although a
very weak one for small q) for q ≥ 5. The limit of dense graphs (average probability for an edge to exist is p, pin/out
are probabilities of connection within/between groups), and large number of groups q → ∞ was analyzed in [193] and
is related to critical temperatures in the dense Potts glass [50, 113]. The regime in which both the dynamical and
detectability phase transitions appear is
N (pin − pout )2 = γ(q)p(1 − p) ,
(124)
52
with the dynamical transitions at γd (q) = 2q log q[1 + o(1)], the detectability transition at γc (q) = 4q log q[1 + o(1)],
and the spinodal threshold at γs (q) = q 2 . In this limit we see that the hard phase between γs (q) and γc (q) is very
wide (compared to the overall scale).
The prediction and location of these phase transitions in the stochastic block model from [71, 72] was followed by a
remarkable mathematical development where the location of the threshold was made rigorous for q = 2 [205, 231, 232].
The detectable side was proven for generic q in [40].
To briefly comment on the case where the average degree of every group is not the same: BP is still conjectured to
be asymptotically exact, but the phase transitions get smeared. As recently studied in [328], the second order phase
transitions disappear, and the first order phase transitions shrink into a multi-critical point beyond which they also
disappear. The situation is similar to the semi-supervised clustering, where the assignment of a fraction of nodes is
known [329]. From a physics point of view it is relevant to note that this problem is an inference interpretation of the
particle pinning that was recently used quite extensively for studies of structural glasses [53].
D.
Spectral redemption for clustering sparse networks
Any reader that did no skip section IV A is now clearly anticipating the use of the non-backtracking matrix for
spectral clustering of networks. Indeed, in [180] the following spectral clustering method was proposed
• Construct the non-backtracking matrix
Bi→j,k→l = δil (1 − δjk ) .
(125)
and compute its largest (in module) eigenvalues until they start to have a non-zero imaginary part. Denote k
the number of those large real eigenvalues.
• Consider the k P
corresponding 2M -dimensional eigenvectors ui→j and create a set of N -dimensional vectors by
resuming ui = k∈∂i uk→i . Then view the k vectors ui as N points in k dimensional space and cluster those
points into k groups using some off-the-shell clustering algorithm such as k-means.
Note in particular that k here is the estimator for the number of groups. Up to this point we were assuming that
the number of groups was known. The performance of this non-backtracking based spectral algorithms compared
to BP, and five other spectral algorithms on clustering of sparse graphs, is illustrated in Fig. 11. All these spectral
algorithms look for the largest (or smallest) eigenvalue of the associated matrix and then assign nodes into groups
based on the sign of the corresponding eigenvector. The five matrices that are most commonly used for clustering are
the adjacency matrix A, the Laplacian L = D − A (where D is a diagonal matrix with the node degrees di on the
diagonal), normalized Laplacian L̃ = D−1/2 LD−1/2 , random walk matrix Wij = Aij /di , and the modularity matrix
Qij = Aij − di dj /(2M ). The spectra of all these matrices on sparse graphs is affected by the existence of large
eigenvalues associated to localized eigenvactors. This fact deteriorated their performance with respect to both BP
and the non-backtracking based algorithm.
Ref. [180] analyzed the performance of the above algorithm on networks generated from the stochastic block model
with equal average degree in every group. A large part of results that were not rigorous in [180] were very remarkably
made rigorous in [40].
On a random graph taken from the configurational model, i.e. at random with a given degree distribution P (d), in
the large
√ size limit, the spectrum of the non-backtracking matrix is (except one eigenvalue) included inside a circle of
radius c̃. Here c̃ is the average excess degree of the graph computed as
P 2
d P (d)
c̃ = Pd
− 1.
(126)
d dP (d)
The one eigenvalue that is not inside this circle is real and has value c̃. This random graph spectrum should be
contrasted with the spectrum of the adjacency matrix (and other related ones) which for graphs with constant
average and unbounded maximum degree is not even bounded in the thermodynamic limit.
Now, consider a graph generated by the SBM with
√ equal average degree for every group. Denote λT an eigenvalue
of the matrix T , eq. (119), then as long as λT > 1/ c̃ the non-backtracking matrix will have purely real eigenvalues
at positions cλT . This way we obtain up to q − 1 additional eigenvalues on the real axes out of the circle of radius
√
c̃. Note that in the assortative-diagonal case the matrix T has q − 1 degenerate eigenvalues λT = (cin − cout )/(qc),
and therefore the non-backtracking matrix has a (q − 1)-times degenerate eigenvalue at (cin − cout )/q which merges
into the bulk below the spinodal phase transition. The spectrum of the non-backtracking operator is illustrated in
Fig. 12.
53
1
1
Overlap
0.8
Non−backtracking
Normalized Laplacian
Modularity
Random Walk
Adjacency
Laplacian
BP
0.8
0.6
0.6
0.4
0.4
0.2
0.2
0
2.5
3
3.5
4
4.5
cin−cout
5
5.5
0
2
6
4
6
8
10
12
Average Degree
FIG. 11. The accuracy of spectral algorithms based on different linear operators, and of belief propagation, for two groups of
equal size. On the
√ left, we vary cin − cout while fixing the average degree c = 3; the detectability transition given by occurs
at cin − cout = 2 3 ≈ 3.46. On the right, we set cout /cin = 0.3 and vary c; the detectability transition is at c ≈ 3.45. Each
point is averaged over 20 instances with N = 105 . The spectral algorithm based on the non-backtracking matrix B achieves
an accuracy close to that of BP, and both remain large all the way down to the transition. Standard spectral algorithms based
on the adjacency matrix, modularity matrix, Laplacian, normalized Laplacian, and the random walk matrix all fail well above
the transition, giving a regime where they do no better than chance. Figure is taken from [180].
1.5
1.0
0.5
-1
µc
1
2
3
-0.5
-1.0
-1.5
FIG. 12. The spectrum of the non-backtracking matrix B for a network generated by the block model with N = 4000, cin = 5,
and cout = 1. The leading eigenvalue is at c√= 3, √
the second eigenvalue is close to µc = (cin − cout )/2 = 2, and the bulk of the
spectrum is confined to the disk of radius c = 3. The spectral algorithm that labels vertices according to the sign of B’s
second eigenvector (summed over the incoming edges at each vertex) labels the majority of vertices correctly. Figure is taken
from [180].
Consequently, when the average degree of every group is the same, condition (116), the phase transition of the
spectral algorithm coincides with the spinodal phase transition of the belief propagation. In other words as long the
average degree is larger than given by eq. 120 the eigenvectors corresponding to the eigenvalues on the real axes out
of the circle contain information about he ground-truth communities.
We need to stress that the agreement between the spectral and belief propagation thresholds arises only for the
case of equal average degree in every group. When the condition (116) is not satisfied there is information about the
groups on the degree and belief propagation explores this information in an optimal way. What concerns the phase
transitions, they tend to be smeared (second order phase transitions disappear, and first order ones become weaker)
[328]. The spectral methods, on the other hand, always have a threshold where the informative eigenvalues merge
back into the bulk. At this spectral phase transition, when (116) does not hold, BP does not witness any particular
change in performance.
54
Let us compare properties of belief propagation and the non-backtracking (NB) based spectral clustering. The
positive point for BP is that it gives a slightly better overlap. The positive points for NB-spectral clustering, that are
shared with other spectral methods, e.g. all those tested in Fig. 11, are that it does not need to get any information
about the model parameters nor the number of groups. It works in time linear in the number of groups, as opposed to
quadratic for BP. It is not expected to be affected by the presence of small loops which usually cause the failure of the
Bethe approximation. An inconvenience that is shared by all the considered spectral methods is that the performance
can deteriorate considerably upon small adversarial changes to the graph (such as planting a small dense subgraph).
Approaches based on some global optimization, such as BP, are expected to be robust towards small local changes.
Another inconvenient property of the non-backtracking matrix is that it can be rather large. However, this is no
obstacle at all because the spectrum (different from ±1) of B can be computed from the spectrum of the following
2N × 2N matrix B 0 [180]
0 D−1
0
B =
,
(127)
−1
A
where A is the adjacency matrix and D is the diagonal matrix with Dii = di being the degree of node i.
The matrix B 0 is smaller but still non-symmetric. The power method for computing the largest eigenvalues works
much better for symmetric matrices. In [271], it was suggested to use the so-called Bethe Hessian matrix defined as
BH(r) = (r2 − 1)1 − rA + D ,
(128)
√
with r = c̃ (or more general the squared root of the leading eigenvalue of (127) which is usually not problematic
to compute) as a basis for spectral clustering. In particular, to take eigenvectors of all of its negative eigenvalues.
Ref. [271] showed that this has all the advantages of the non-backtracking matrix while being N × N and symmetric.
The name Bethe Hessian is motivated by the relation between the Hessian of the Bethe free energy and this matrix
[226, 264], which is particularly apparent when generalized to the weighted case.
E.
More on the non-backtracking operator
The non-backtracking operator as defined above is relatively well known in mathematics, often called the Hashimoto
directed edges adjacency matrix [120]. It is used in a number of works as a tool in proofs, see e.g. [8, 24, 98, 159,
252, 287]. However, using it as a basis for a spectral algorithm was first suggested in [180]. The idea of using the
linearization of BP was inspired by the earlier work of [66]. Since then, a considerable number of papers on promising
algorithmic applications of the non-backtracking matrix on sparse networks started appearing, see [204, 229, 325].
Here, we briefly discuss selected contributions in this direction.
Looking at Fig. 12, we observe a very interesting structure in the bulk spectral density of the non-backtracking
operator. Motivated by this, ref. [272] applied the cavity method designed to compute spectral densities of nonHermitian random matrices [268] to study the non-backtracking spectral√density. A phase transition in the spectral
density was observed to happen on the boundary of the circle of radius c̃. Such a phase transition is absent in the
spectral density of traditional matrices such as the adjacency, Laplacian, random walks etc. This is no proof since we
are only concerned with spectral density not with sub-dominantly numerous eigenvalues, but from a physics point of
view a phase transition is a strong indication that the properties of the non-backtracking matrix on a random graph
are fundamentally different from those of traditional matrices. Ref. [272] also observed that the eigenvalues that
are close to the center of the circle are not visible in the asymptotic calculation of the spectral density. Numerical
investigations revealed that eigenvectors corresponding to these eigenvalues are localized. More detailed understanding
of this localization and its physical consequences is an interesting direction of future work.
In [10], the non-backtracking spectral clustering method was generalized to clustering of sparse hypergraphs. These
are graphs where instead of connected pairs of nodes, one connects k-uples of nodes. Interestingly, using the nonbacktracking based spectral method detection can be done also e.g. in planted constraint satisfaction problems without
even knowing the form of the corresponding constraints. Moreover, for the same reasons as in clustering, the method
works down to previously computed spinodal thresholds [324]. It was also observed that the overlap (performance)
of the spectral method is always a continuous function of the noise parameter, even if the Bayes optimal belief
propagation has a strongly discontinuous transition. This further supports our previous suggestion that the spectral
methods are extremely useful to point towards the right direction, but message-passing based methods should then
be used to improve the precision. Note that the hypergraph clustering can also be formulated as tensor factorization
that became recently popular in connection with interdependent networks or networks with different kinds of edges
that are referred to as multiplexes in the literature on complex networks.
55
Another exciting application of the non-backtracking matrix, or the Bethe Hessian, is the matrix completion problem
[273]. In matrix completion one aims to estimate a low-rank matrix from a small random sample of its entries. This
problem became popular in connection to collaborative filtering and related recommendation systems. Of course, the
fewer entries are observed, the harder the problem is. Ref. [273] shows that with the use of the non-backtracking
matrix the number of entries from which the low-rank structure is detectable is smaller than with other known
algorithms. The way the matrix completion problem is approached is by realizing that a Hopfield model with a Hebb
learning rule can be viewed as a rank r structure that is able to recover the r patterns as distinct thermodynamic
phases [136]. One views the observed entries of the low-rank matrix to complete as a sparse Hopfield model and
estimate the patterns from the leading eigenvectors of the corresponding Bethe Hessian. In a second stage, we use
this as a initialization for more standard matrix completion algorithms in the spirit of [162], and achieve overall better
performance.
Another problem where the leading eigenvalue of the non-backtracking matrix is relevant is percolation on networks
[161]. In percolation, we start with a graph G and remove each of its edges with probability 1 − p. Then we look
at the largest component of the remaining graph and ask whether it covers an extensive fraction of all nodes. The
value of pc below which the giant component does not exist anymore with high probability is called the percolation
threshold. Percolation thresholds were computed for many geometries including random graphs. A remarkable result
concerns any sequence of dense graphs for which it was proven that the threshold is given by the inverse of the leading
eigenvalue of the adjacency matrix [38]. This result does not extend to sparse graphs, where a counter-example is
given by random regular graphs in which the leading eigenvalue is the degree whereas the percolation threshold is the
inverse of the degree minus one.
Based on a message-passing algorithm for percolation, [161] argues that on a large class of sparse locally tree-like
graphs the percolation threshold is given by the inverse of the leading eigenvalue of the non-backtracking matrix.
Specifying for what exact class of graphs this result holds is still an open problem. Ref. [161] also proved that the
non-backtracking matrix provides a lower bound on the percolation threshold for a generic graph (with the exception
of trees). An analogous result was obtained independently by [119]. Due to these results, the non-backtracking matrix
was used to study percolation on real networks [229, 255, 267].
F.
The deceptiveness of the variational Bayes method
Both in physics and in computer science, a very common and popular method for approximate Bayesian inference
is the so-called naive mean field method, often called variational Bayesian inference [27, 244]. For clustering of
sparse networks generated by the stochastic block model this method was studied e.g. in [59, 133]. Away from the
detectability threshold the variational Bayes method gives very sensible results. Deep in the undetectable phase (very
small |cin − cout |) the predicted magnetizations are close to zero, reflecting the undetectability. Deep in the detectable
phase the variational Bayesian method is able to estimate the correct group assignment, and is statistically consistent
as proven in [59]. The performance of variational Bayes in the neighborhood of the detectability transition was studied
in [326]. The conclusions of the comparison to belief propagation are quite interesting. It is not surprising that BP
is more precise and converges faster on sparse graphs. What is surprising is that the naive mean-field inference fails
in a very deceptive way: the marginals it predicts are very biased towards a configuration that is not correlated
with the planted (ground truth) configuration. Ref. [326] quantifies this by introducing the so-called illusive overlap
and comparing it to the actual one. This means that it can be rather dangerous to use naive mean-field for testing
statistical significance of the results it provides and this could have profound consequences beyond the problem of
clustering networks. Especially when trying to use the variational Bayes method in the vicinity of the corresponding
detectable threshold, i.e. in regimes where the number of samples is barely sufficient for estimation. Monte Carlo
methods do not have this problem, but are in general slower.
G.
Real networks are not generated by the SBM
Our study of clustering of networks is asymptotically exact for the case when the networks were generated by the
stochastic block model. In view of actual data sets and realistic applications this is never the case.
One important discrepancy between the stochastic block model and real networks is that the latter often have
communities with a broad degree distribution, whereas the degree distribution of every group generated by the SBM
is Poissonian. A variant of the stochastic block model that takes the degree distribution into account is the so-called
degree corrected stochastic block model [160]. An asymptotically exact belief propagation can also easily be written
for this model, see [317]. In Ref. [317] it was also shown that the corresponding Bethe free energy is a great tool for
model selection between the canonical and degree corrected stochastic block models, and that classical model selection
56
30
3
20
Polblogs q=2
Overlap: 0.8640
10
-40
4
2
3
Dolphins q=2
Overlap: 0.7419
1
-20
20
40
60
80
-4
-2
1
2
-10
Karate q=2
Overlap: 1
2
4
6
8
-4
-2
-1
-1
2
4
6
8
10
12
-2
-20
-2
-30
-3
-3
-4
4
4
3
4
Football q=12
Overlap: 0.9163
2
Adjnoun q=2
Overlap: 0.6250
3
2
1
-4
-2
-1
-2
-3
-4
Polbooks q=3
Overlap: 0.7571
2
1
2
4
6
8
10
12
-5
5
-2
-4
10
-4
-2
-1
2
4
6
8
10
12
-2
-3
-4
FIG. 13. Spectrum of the non-backtracking matrix in the complex plane for some real networks commonly used as benchmarks
for community detection, taken from [7, 108, 171, 196, 236, 319]. The radius of the circle is the square root of the largest
eigenvalue, which is a heuristic estimate of the bulk of the spectrum. The overlap is computed using the signs of the second
eigenvector for the networks with two communities, and using k-means for those with three and more communities. The
non-backtracking operator detects communities in all these networks, with an overlap comparable to the performance of other
spectral methods. As in the case of synthetic networks generated by the stochastic block model, the number of real eigenvalues
outside the bulk appears to be a good indicator of the number q of communities. Figure taken from [180].
criteria do not apply to this problem because of the limited information per degree of freedom. Interestingly, BP for
the degree-corrected SBM also leads to the same non-backtracking matrix upon linearization.
There are of course many more discrepancies between real networks and those generated by the (degree corrected)
SBM. However, when observing the spectrum of the non-backtracking operator on some of the commonly used
benchmarks for community detection, we conclude that the properties derived for the SBM basically stay conserved
also in these examples, see Fig. 13. To conclude this section, let us quote the British mathematician and Bayesian
statistician George E. P. Box (1919-2013): “Essentially, all models are wrong, but some are useful”.
57
VI.
LINEAR ESTIMATION AND COMPRESSED SENSING
The inference problem addressed in this section is the following. We observe a M dimensional vector yµ , µ =
1, . . . , M that was created component-wise depending on a linear projection of an unknown N dimensional vector xi ,
i = 1, . . . , N by a known matrix Fµi
!
N
X
yµ = fout
Fµi xi ,
(129)
i=1
where fout is an output function that includes noise. Examples are the additive white Gaussian noise (AWGN) there
the output is given by fout (x) = x + w with w being random Gaussian variables, or the linear threshold output where
fout (x) = sign(x − κ), with κ being a threshold value. The goal in linear estimation is to infer x from the knowledge
PN
of y, F and fout . An alternative representation of the output function fout is to denote zµ = i=1 Fµi xi and talk
about the likelihood of observing a given vector y given z
P (y|z) =
M
Y
µ=1
A.
Pout (yµ |zµ ) .
(130)
Some examples of linear estimation
To put the immense number of important applications of the above setting into perspective, let us give an illustrative
example from medical diagnostics. Consider we have data about M cancer-patients that took a given treatment. Then
yµ is a variable rating the effectiveness of the treatment for a patient µ. Our goal is to predict the outcome of the
treatment for a new patient that did not undergo treatment yet. To do that, we collect medical and personal data
about each of the M patients, their medical history, family situation, but also for instance Facebook feeds, where
they worked, what sport they did in their life, what education they have, how they rate their relationship with their
wife, and all kinds of other information that is on first sight seems totally unrelated to the eventual outcome of
the treatment. This information is collected in the vector F~µ . We aim to find a set of coefficients xi such that the
PN
observations yµ can be well explained by zµ = i=1 Fµi xi for some Pout (yµ |zµ ).
In the way we set the problem, it is reasonable to think that most patient-attributes will have no effect on the
outcome of the treatment, i.e. for many i’s the xi will be close to zero. In a method, called LASSO (which stands
for least absolute shrinkage and selection operator) [123, 301], that is used everyday for similar types of problems (in
banking or insurance industry for instance) the coefficients xi are such that
!2
M
N
N
X
X
X
LASSO :
min
yµ −
Fµi xi
+λ
|xi | ,
(131)
x
µ=1
i=1
i=1
where λ is a regularization parameter that controls the trade-off between a good fit and the sparsity of the vector
x. The advantage of LASSO is that it is a convex problem that can be solved efficiently and with strong guarantees
using linear programming. LASSO and related linear regression problems are well studied in statistics and machine
learning, instead of reviewing the literature we refer the reader to the excellent textbook [122].
LASSO can be framed in the Bayesian probabilistic setting by considering
Pout (yµ |zµ ) = √
PN
2
i=1 Fµi xi )
2∆
(yµ −
1
e−
2π∆
(132)
and the prior of xi being the Laplace distribution
`1
PX
(xi ) = e−λ|xi | /2 .
(133)
Another sparsity inducing prior is a combination of strict zeros and fraction ρ of non-zeros
PX (xi ) = ρφ(xi ) + (1 − ρ)δ(xi ) ,
(134)
where φ(xi ) is the distribution of non-zero elements. Linear programming does not work for priors such as (134), but
the statistical-physics framework is more general in this aspect.
58
In many applications, the output variables yµ are binary (interpreted in the above example as the treatment
worked or did not), then the most commonly considered output distribution is the logistic function. The conditional
dependence then is
logistic
Pout
(yµ = 1|zµ ) =
1
.
1 + e−zµ
(135)
The example of this linear estimation setting that is the best known in statistical physics is the perceptron [269],
as introduced briefly in sec. I B 2. In the perceptron setting, components of the unknown vector x are interpreted as
synaptic weights (that are in general not sparse). The rows of the matrix F are patterns presented to the perceptron.
The measurements yµ are binary and correspond to classifications of the patterns (think of some patterns being
pictures of cats yµ = +1
dogs yµ = −1). The goal is to learn the synaptic weights
P and the others being pictures of P
xi in such a way that i Fµi xi > 0 when yµ = +1, and i Fµi xi < 0 when yµ = −1. The random version of the
problem has been considerably studied by physicists, see for instance [102, 103, 170, 286, 310].
Another example well known and studied in the statistical physics literature is the code-division multiple access
(CDMA) problem in which N users simultaneously aim to transmit their binary signal xi ∈ {±1} to one base
station. In order
√ to be able to deal with interference one designs a so-called spreading code associated to each user
Fµi ∈ {±1}/ N , µ = 1, . . . , M , that spreads the signal of each user among M samples. The base station then receives
a noisy linear combination of all the signals yµ , eq. (129). The goal is to reconstruct the signal of every user xi .
Interestingly, the two examples above are the first inference problems on which, to our knowledge, the strategy
discussed in the present chapter has been introduced in a statistical physics. The seminal work of Marc Mézard
derived the TAP equations for the perceptron problem as early as in 1989 [210]. It has then been generalized to many
other cases, mainly by Kabashima and co-workers (see [152, 155] again for the perceptron and [150] for randomly
spreading CDMA).
In this chapter, we shall follow these pioneering works and adopt the Bayesian probabilistic framework for the
above generalized linear estimation problem. In general this is algorithmically much more involved than the convex
relaxations of the problem, and consequently less widely studied. The cases where interesting theoretical analysis can
be done are limited to special classes of matrices Fµi (mostly with random iid elements, but see the discussion in
section VI F). Nevertheless the resulting algorithms can (as always) be used as heuristic tools for more general data
Fµi . One problem where the above linear estimation problem truly involves a matrix Fµi with random iid elements
is compressed sensing [78] to which we dedicate the next sections.
B.
Compressed sensing
Compressed sensing, a special case of the above linear estimation setting, triggered a major evolution in signal
acquisition. Most signals of interest are compressible and this fact is widely used to save storage place. But so far
compression is made only once the full signal is measured. In many applications (e.g. medical imaging using MRI or
tomography) it is desirable to reduce the number of measurements as much as possible (to reduce costs, time, radiation
exposition dose etc.). It is hence desirable to measure the signal directly in the compressed format. Compressed sensing
is a concept implementing exactly this idea by designing new measurement protocols. In compressed sensing one aims
to infer a sparse N -dimensional vector x from the smallest possible number M of linear projections y = F x. The
M × N measurement matrix F is to be designed.
Most commonly used techniques in compressed sensing are based on a development that took place in 2006 thanks
to the works of Candes, Tao, and Donoho [54, 55, 78]. They proposed to find the vector x satisfying the constraints
y = F x which has the smallest `1 norm (i.e., the sum of the absolute values of its components). This `1 minimization
problem can be solved efficiently using linear programming techniques. They also suggested using a random matrix
for F . This is a crucial point, as it makes the M measurements random and incoherent. Incoherence expresses the
idea that objects having a sparse representation must be spread out in the domain in which they are acquired, just as
a Dirac or a spike in the time domain is spread out in the frequency domain after a Fourier transform. Rather strong
and generic theoretical guarantees can be derived for the `1 -based sparse reconstruction algorithms [123]. These ideas
have led to a fast, efficient and robust algorithm, and this `1 -reconstruction is now widely used, and has been at the
origin of the burst of interest in compressed sensing over the last decade.
It is possible to compute exactly the performance of `1 -reconstruction in the limit N → ∞, and the analytic
study shows the appearance of a sharp phase transition [83]. For any signal with density ρ = K/N , where K is the
number of non-zero elements, the `1 -reconstruction gives the exact signal with probability one if the measurement rate,
α = M/N , is above the so-called Donoho-Tanner line, α > α`1 (ρ) > ρ, see Fig. 15 (black dashed line). Interestingly,
the very same transition can be found within the replica method [101, 157] but can also be computed, rigorously, with
the approximate message passing technics that we discuss in this chapter [81].
59
If one is willing to forget the (exponentially large) computational cost, sparse signal reconstruction from random
linear projections is possible whenever α > ρ (this can be done for instance by exhaustively checking all possible
positions of the non-zero elements and trying to invert the reduced system of linear equations). A challenging
question is whether there is a compressed sensing design and a computationally tractable algorithm that can achieve
reconstruction down to such low measurement rates. In a series of works [173, 175], complemented by an impressive
rigorous proof of [79], this question was answered positively under the assumption that the empirical distribution
of the non-zero elements of the signal is known. This contribution to the theory of compressed sensing is explained
below and uses many of the statistical physics elements from section II.
C.
Approximate message passing
Approximate message passing (AMP) is the algorithmic procedure behind the aforementioned work of [81] on linear
estimation and compressed sensing. Origins of AMP trace back to statistical physics works on the perceptron problem
[210] and on the code division multiple access (CDMA) [150].
What nowadays is known as the generalized approximate message passing (G-AMP) algorithm, i.e. the generalization to arbitrary element-wise output functions, was derived for the first time, to our knowledge, in connection
to perceptron neural network in [155, 210]. The algorithm presented in [155] had the correct time-indices and good
performance was illustrated on the CDMA problem. However, for an unknown reasons, that paper did not have much
impact and many of its results waited to be rediscovered in the connection to compressed sensing much later. The
opinion that this type of algorithm has invincible convergence problems prevailed in the literature until the authors of
[45] demonstrated impressive performance of a related message passing algorithm for learning in binary perceptron.
AMP is closely related to relaxed-belief propagation (r-BP) [256]. The algorithm of [45] can actually be seen as a
special case of r-BP.
Donoho, Maleki and Montanari introduced AMP for compressed sensing [81], they also established the name. Very
remarkably, the rigorous proof of its performance for random matrices F was done in [25, 26]. A comprehensive
generalization to arbitrary priors and element-wise output function is due to Rangan [257], who established the name
generalized approximate message passing for the version with a general output channel. Many variants of this approach
have then been considered (see for instance, without being exhaustive, [92, 158, 175, 288]).
As is clear from the earlier papers [150], AMP is actually very closely related to the TAP equations [300] presented
in section IV E. The main difference between TAP and AMP is that TAP was originally derived for a model with
Ising spins and pairwise interactions, whereas AMP is for continuous variables and N -body interactions.
FIG. 14. Factor graph of the linear estimation problem corresponding to the posterior probability (136). Circles represent the
unknown variables, whereas squares represent the interactions between variables.
60
1.
Relaxed Belief propagation
A detailed statistical physics derivation of the AMP algorithm was presented in [175] for the Gaussian output
channel, and for general output as a part of the yet more general problem of matrix factorization in [153]. Here we
shall describe the main elements necessary for the reader to understand where the resulting algorithm comes from.
We are given the posterior distribution
P (x|y, F ) =
M
N
Y
1 Y
Pout (yµ |zµ )
PX (xi ) ,
Z µ=1
i=1
where
zµ =
N
X
Fµi xi ,
(136)
i=1
where the matrix Fµi has independent random entries (not necessarily identically distributed) with mean and variance
O(1/N ). This posterior probability distribution corresponds to a graph of interactions yµ between variables (spins)
xi called the graphical model as depicted in Fig. 14.
A starting point in the derivation of AMP is to write the belief propagation algorithm corresponding to this graphical
model, as done originally in [21]. The matrix Fµi plays the role of randomly-quenched disorder, the measurements yµ
are planted disorder. As long as the elements of Fµi are independent and their mean and variance of of order O(1/N )
the corresponding system is a mean field spin glass. In the Bayes-optimal case (i.e. when the prior is matching the
true empirical distribution of the signal) the fixed point of belief propagation with lowest free energy then provides
the asymptotically exact marginals of the above posterior probability distribution.
For model such as (136), where the interaction are not pairwise, the BP equation we presented in sec. III B need
to be slightly modified, for examples see [217], but the logic is very similar. BP still implements a message-passing
scheme between nodes in the graphical model of Fig. 14, ultimately allowing one to compute approximations of the
posterior marginals. Messages mi→µ are sent from the variable-nodes (circles) to the factor-nodes (squared) and
subsequent messages mµ→i are sent from factor-nodes back to variable-nodes that corresponds to algorithm’s current
“beliefs” about the probabilistic distribution of the variables xi . It reads
Y
1
PX (xi )
mγ→i (xi )
(137)
mi→µ (xi ) =
zi→µ
γ6=µ
Z Y
X
1
mµ→i (xi ) =
[dxj mj→µ (xj )] Pout (yµ |
Fµl xl )
(138)
zµ→i
j6=i
l
While being easily written, this BP is not computationally tractable, because every interaction involves N variables,
and the resulting belief propagation equations involve (N − 1)-uple integrals.
Two facts enable the derivation of a tractable BP algorithm: the central limit theorem, on the one hand, and
a projection of the messages to only two of their moments (also used in algorithms such as Gaussian BP [284] or
non-parametric BP [293]). This results in the so-called relaxed-belief-propagation (r-BP) [256]: a form of equations
that is tractable and involves a pair of means and variances for every pair variable-interaction.
Let us go through the steps that allow
P to go from BP to r-BP and first concentrate our attention to eq. (138).
Consider the variable zµ = Fµi xi + j6=i Fµj xj . In the summation in eq. (138), all xj s are independent, and
P
2
thus we expect, according to the central limit theorem, j6=i Fµj
xj to be a Gaussian random variable with mean
P
P
2
ωµ→i = j6=i Fµj aj→µ and variance Vµ→i = j6=i Fµj vj→µ , where we have denoted
Z
Z
ai→µ ≡ dxi xi mi→µ (xi )
vi→µ ≡ dxi x2i mi→µ (xi ) − a2i→µ .
(139)
We thus can replace the multi-dimensional integral in eq. (138) by a scalar Gaussian one over the random variable z:
Z
(z−ωµ→i −Fµi xi )2
−
2Vµ→i
.
(140)
mµ→i (xi ) ∝ dzµ Pout (yµ |zµ )e
We have replaced a complicated multi-dimensional integral involving full probability distributions by a single scalar
one that involves only first two moments.
2 2
One can further simplify the equations: The next step is to rewrite (z − ωµ→i − Fµi xi )2 = (z − ωµ→i )2 + Fµi
xi −
√
2(z − ωµ→i )Fµi xi Fµi xi and to use the fact that Fµi is O(1/ N ) to expand the exponential
Z
(z−ω
)2
− 2V µ→i
µ→i
mµ→i (xi ) ∝ dzµ Pout (yµ |zµ )e
(141)
1
2 2
2 2
× 1 + Fµi
xi − 2(z − ωµ→i )Fµi xi + (z − ωµ→i )2 Fµi
xi + o(1/N ) .
2
61
At this point, it is convenient to introduce the output function gout , defined via the output probability Pout as
R
(z−ω)2
dzPout (y|z) (z − ω) e− 2V
.
(142)
gout (ω, y, V ) ≡
R
(z−ω)2
V dzPout (y|z)e− 2V
The following useful identity holds for the average of (z − ω)2 /V 2 in the above measure:
R
(z−ω)2
dzPout (y|z) (z − ω)2 e− 2V
1
2
+ ∂ω gout (ω, y, V ) + gout
(ω, y, V ) .
=
R
(z−ω)2
V
−
V 2 dzPout (y|z)e 2V
(143)
Using definition (142), and re-exponentiating the xi -dependent terms while keeping all the leading order terms, we
obtain (after normalization) that the iterative form of equation eq. (138) reads:
s
t
(Bµ→i
)2
x2
x
t
i At
√i
Atµ→i − 2N
µ→i +Bµ→i N − 2At
µ→i
mµ→i (t, xi ) =
e
(144)
2πN
with
t
t
t
Bµ→i
= gout (ωµ→i
, yµ , Vµ→i
) Fµi
t
t
2
and Atµ→i = −∂ω gout (ωµ→i
, yµ , Vµ→i
) Fµi
.
Given that mµ→i (t, xi ) can be written with a quadratic form in the exponential, we just write it as a Gaussian
distribution. We can now finally close the equations by writing (137) as a product of these Gaussians with the prior
mi→µ (xi ) ∝ PX (xi )e
−
(xi −Ri→µ )2
2Σi→µ
(145)
so that we can finally give the iterative form of the r-BP algorithm Algorithm 1 (below) where we denote by ∂ω (resp.
∂R ) the partial derivative with respect to variable ω (resp. R) and we define the “input” functions as
R
(x−R)2
dx x PX (x) e− 2Σ
fa (Σ, R) = R
,
fv (Σ, R) = Σ∂R fa (Σ, R) .
(146)
(x−R)2
dx PX (x) e− 2Σ
Algorithm 1 relaxed Belief-Propagation (r-BP)
Input: y
Initialize: ai→µ (t = 0),vi→µ (t = 0), t = 1
repeat
r-BP Update of {ωµ→i , Vµ→i }
Vµ→i (t) ←
X
2
Fµj
vj→µ (t − 1)
(147)
Fµj aj→µ (t − 1)
(148)
j6=i
ωµ→i (t) ←
X
j6=i
r-BP Update of {Aµ→i , Bµ→i }
Bµ→i (t) ← gout (ωµ→i (t), yµ , Vµ→i (t)) Fµi ,
2
Aµ→i (t) ← −∂ω gout (ωµ→i (t), yµ , Vµ→i (t))Fµi
(149)
(150)
r-BP Update of {Rµ→i , Σµ→i }
1
Aν→i (t)
X
Ri→µ (t) ← Σi→µ (t)
Bν→i (t)
Σi→µ (t) ← P
(151)
ν6=µ
(152)
ν6=µ
AMP Update of the estimated partial marginals ai→µ (t),vi→µ (t)
ai→µ (t) ← fa (Σi→µ (t), Ri→µ (t)) ,
(153)
vi→µ (t) ← fv (Σi→µ (t), Ri→µ (t)) .
(154)
t←t+1
until Convergence on ai→µ (t),vi→µ (t)
output: Estimated marginals mean and variances:
P
Bν→i
1
, Pν Aν→i
A
ν→i
ν
ν
P
Bν→i
fv P A1 ν→i , Pν Aν→i
ν
ν
ai ← fa
vi ←
P
,
(155)
.
(156)
62
2.
From r-BP to G-AMP
While one can implement and run the r-BP, it can be simplified further without changing the leading order behavior
of the marginals by realizing that the dependence of the messages on the target node is weak. This is exactly what
we have done to go from the standard belief propagation to the TAP equations in sec. IV E.
After the corresponding Taylor expansion the corrections add up into the so-called Onsager reaction terms [300].
The final G-AMP iterative equations are written in terms of means ai and their variances ci for each of the variables
xi . The whole derivation is done in a way that the leading order of the marginal ai are conserved. Given the BP was
asymptotically exact for the computations of the marginals so is the G-AMP in the computations of the means and
variances. Let us define
X
ωµt+1 =
Fµi ati→µ ,
(157)
i
Vµt+1
X
=
2 t
Fµi
vi→µ ,
(158)
i
1
t+1 ,
µ Aµ→i
P
t+1
µ Bµ→i
.
= P t
µ Aµ→i
Σt+1
=P
i
(159)
Rit+1
(160)
First we notice that the correction to V are small so that
X
X
2 t
2 t
Vµt+1 =
Fµi
vi→µ ≈
Fµi
vi .
i
(161)
i
From now on we will use the notation ≈ for equalities that are true only in the leading order.
From the definition of A and B we get
"
#−1 "
#−1
X
X
t+1
t+1
2
2
t+1
t+1
Σi = −
Fµi ∂ω gout (ωµ→i , yµ , Vµ→i )
≈ −
Fµi ∂ω gout (ωµ , yµ , Vµ )
µ
µ
#−1
"
Ri = −
X
2
Fµi
∂ω gout (ωµt+1 , yµ , Vµt+1 )
µ
×
"
X
#
t+1
t+1
Fµi gout (ωµ→i
, yµ , Vµ→i
)
µ
Then we write that
t+1
t+1
gout (ωµ→i
, yµ , Vµ→i
) ≈ gout (ωµt+1 , yµ , Vµt+1 ) − Fµi ati→µ ∂ω gout (ωµt+1 , yµ , Vµt+1 )
≈
gout (ωµt+1 , yµ , Vµt+1 )
−
Fµi ati ∂ω gout (ωµt+1 , yµ , Vµt+1 )
(162)
(163)
So that finally
#−1
"
(Σi )
t+1
= −
X
2
Fµi
∂ω gout (ωµt+1 , yµ , Vµt+1 )
(164)
µ
"
Rit+1
= (Σi )
t+1
×
#
X
µ
= ati + (Σi )t+1
Fµi gout (ωµ , yµ , Vµ ) −
X
2
Fµi
ai ∂ω gout (ωµ , yµ , Vµ )
Fµi gout (ωµt+1 , yµ , Vµt+1 )
(165)
(166)
µ
Now let us consider ai→µ :
t
t
ati→µ = fa (Ri→µ
, Σi→µ ) ≈ fa (Ri→µ
, Σi )
≈
≈
t
fa (Rit , Σi ) − Bµ→i
fv (Rit , Σi )
ati − gout (ωµt , yµ , Vµt )Fµi vit
(167)
(168)
(169)
63
So that
ωµt+1 =
X
i
Fµi ati −
X
2 t
gout (ωµt , yµ , Vµt )Fµi
vi =
X
i
i
Fµi ai − Vµt gout (ωµt , yµ , Vµt )
(170)
Algorithm 2 Generalized Approximate Message Passing (G-AMP)
Input: y
0
Initialize: a0 ,v0 , gout,µ
,t=1
repeat
AMP Update of ωµ , Vµ
Vµt ←
X
2 t−1
Fµi
vi
(171)
t−1
− Vµt gout,µ
Fµi at−1
i
(172)
i
ωµt ←
X
i
AMP Update of Σi , Ri , gout,µ
t
gout,µ
← gout (ωµt , yµ , Vµt )
"
#−1
X 2
t
t
t
Σi ← −
Fµi ∂ω gout (ωµ , yµ , Vµ )
(173)
(174)
µ
Rit ← at−1
+ Σti
i
X
t
Fµi gout,µ
(175)
µ
AMP Update of the estimated marginals ai , vi
ati ← fa (Σti , Rit )
(176)
vit ← fv (Σti , Rit )
(177)
t←t+1
until Convergence on a,v
output: a,v.
An important aspect of these equations to note is the index t − 1 in the Onsager reaction term in eq. (172) that is
crucial for convergence and appears for the same reason as in the TAP equations in sec. IV E. Note that the whole
algorithm is comfortably written in terms of matrix multiplications only, this is very useful for implementations where
fast linear algebra can be used.
3.
The potential
Using the usual Bethe approach, one can also compute the corresponding potential, and show that these message
passing equations actually aim to minimize the following Bethe free energy [174, 262, 306]:
X
X
F Bethe ({Σi }, {Ri }) =
DKL (Qi ||PX ) +
DKL (Mµ ||Pout )
i
µ
1 X
+
log 2πVµ∗ + 1 + Vµ∗ ∂ω gout (ωµ∗ , yµ , Vµ∗ ) ,
2 µ
(178)
R
with Vµ∗ and ωµ∗ satisfying their respective fixed-point conditions, and DKL (Q||P ) = [log Q(x) − log P (x)]dQ(x)
being the Kullback-Leibler divergences between two probability distributions. The two Kullback-Leibler divergences
above are taken with respect to probability distributions
Mµ (z; ωµ∗ , yµ , Vµ∗ ) =
Qi (x; Ti , Σi ) =
1
1
−
Pout (yµ |z) p
e
ZM
2πVµ∗
(x−Ti )2
1
−
PX (x)e 2Σi ,
ZQ
∗ )2
(z−ωµ
∗
2Vµ
(179)
(180)
where ZM and ZQ are the corresponding normalizations. Clearly these two probability distributions are related to
the denominators in (146) and (142).
64
4.
Note on the convergence and parameter learning
When G-AMP converges, its performance in terms of mean-squared error is usually considerably better than the
performance of other existing algorithms. The AMP equations are proven to converge in the idealized setting of large
systems, random iid matrices F with zero mean entries, and when the prior PX either corresponds to the empirical
distribution of the actual signal x∗ or to the convex relaxation of the problem [25, 26]. However, for mismatched
signal distributions, for non-random matrices, positive mean random matrices or small system-sizes they might have
unpleasant convergence issues.
Justified by the promise of a strikingly better performance, it is an important subject of current research how to
make the G-AMP algorithm the most robust as possible, while keeping its performance and speed. Several recent
works followed this line of research. In [52] it was identified why AMP has convergence problems for the simple case
of iid matrices with elements of non-zero mean. For means of the matrix elements larger than some critical value,
there is an anti-ferromagnetic-like (i.e. corresponding to a negative eigenvalue of the Hessian matrix) instability of
the Nishimori line. Finite size effects always cause a slight mismatch between the true and the empirical distribution
of the signal, and hence a slight deviation from the Nishimori line. The instability amplifies the deviation and causes
the iteration of the AMP equation to fail and go far away from the Nishimori line. Even though we know that in the
Bayes-optimal case the correct fixed point must satisfy the Nishimori conditions.
The simplest way to fix this problem for matrices of non-zero mean is the so-called mean removal as used in many
of the early implementations and discussed in [306]. However, the instability of the Nishimori line is a problem that
can cause failure of the algorithm in a wider range of settings than the non-zero mean matrices, e.g. for some of the
structured matrices. It is therefore useful to study other strategies to stabilize the Nishimori line in order to improve
the convergence. We observed that randomization of the update is doing just that. Combining the sequential update
while keeping track of the correct time-indices, Ref. [201] designed the so-called Swept-AMP (SwAMP). Ref. [175]
studied the expectation-maximization learning of parameters for the compressed sensing problem when the number
of parameters is small, typically finite while the system size grows. Since the parameter-learning is precisely imposing
the validity of the Nishimori conditions, this is another way that greatly stabilizes the Nishimori line and improves
the convergence.
An additional approach to stabilize the Nishimori line is to slow down the iterations via damping, i.e. taking a
linear combination of the new and old values when updating the messages. Such damping can, however, slow down
the iteration-process considerably. In order to avoid a drastic slow-down we take advantage of the following related
development.
The authors of [174] studied the Bethe free energy associated to G-AMP and wrote a variational form of the
Bethe free energy (178) for which the G-AMP fixed points are local minima (not only generic stationary points such
as saddles). This free energy can be optimized directly leading to a provably converging (but considerably slower)
method to find the fixed points. This variational Bethe free energy can be used to slow down the iterations in a way
to keep decreasing the free energy - this we call adaptative damping in [306]. Other groups suggested several closely
related convergence improvements [260, 262].
Combining the above implementation features - adaptive damping, sequential update, mean removal and parameter
learning - leads to the current state-of-the art implementation of the G-AMP algorithm for linear estimation problems
[1].
5.
Examples of priors and outputs
The G-AMP algorithm is written for a generic prior on the signal PX (as long as it is factorized over the elements)
and a generic element-wise output channel Pout . The algorithm depends on their specific form only trough the function
fa and gout defined by (146) and (142). It is useful to give a couple of explicit examples.
The sparse prior that is most commonly considered in probabilistic compressed sensing is the Gauss-Bernoulli prior,
that is when in (134) we have φ(x) = N (x, σ) Gaussian with mean x and variance σ. For this prior the input function
fa reads
(T −x)2
ρ e− 2(Σ+σ)
faGauss−Bernoulli (Σ, T ) =
(1 −
√
Σ
3
(xΣ + T σ)
(Σ+σ) 2
√
(T −x)2
T2
Σ
−
ρ)e 2Σ + ρ √Σ+σ
e− 2(Σ+σ)
,
(181)
The most commonly considered output channel is simply additive white Gaussian noise (AWGN) (132). The output
65
function then reads
AW GN
gout
(ω, y, V ) =
y−ω
.
∆+V
(182)
As we anticipated above, the example of linear estimation that was most broadly studied in statistical physics is
the case of the perceptron problem discussed in detail e.g. in [310]. In the perceptron problem each of the M N dimensional patterns Fµ is multiplied by a vector of synaptic weights xi in order to produce an output yµ according
to
yµ = 1
N
X
if
Fµi xi > κ ,
(183)
i=1
yµ = −1
otherwise ,
(184)
where κ is a threshold value independent of the pattern. The perceptron is designed to classify patterns, i.e. one
starts with a training set of patterns and their corresponding outputs yµ and aims to learn the weights xi in such a
way that the above relation between patterns and outputs is satisfied. To relate this to the linear estimation problem
above, let us consider the perceptron problem in the teacher-student scenario where the teacher perceptron generated
the output yµ using some ground-truth set of synaptic weights x∗i . The student perceptron knows only the patterns
and the outputs and aims to learn the weights. How many patterns are needed for the student to be able to learn the
synaptic weights reliably? What are efficient learning algorithms?
In the simplest case where the threshold is zero, κ = 0 one can redefine the patterns Fµi ← Fµi yµ in which case the
corresponding redefined output is yµ = 1. The output function in that case reads
ω2
perceptron
gout
(ω, V
e− 2V
1
,
)= √
2πV H − √ω
V
(185)
where
Z
∞
H(x) =
x
t2
dt
√ e− 2 .
2π
(186)
Note also that the perceptron problem formulated this way is closely related to what is called 1-bit compressed sensing
in the signal processing literature [43, 316].
In physics a case of a perceptron that was studied in detail is that of binary synaptic weights xi ∈ {±1} [102, 103,
117]. To take that into account in the G-AMP we consider the binary prior PX (x) = [δ(x − 1) + δ(x + 1)]/2 which
leads to the input function
T
fabinary (Σ, T ) = tanh
.
(187)
Σ
D.
Replica free energy, state evolution and phase transitions
1.
State Evolution
The performance of the G-AMP algorithm above can be analyzed in the large-size limit, for matrices with independent entries. The corresponding analysis was baptized state evolution in [82]. Rigorous results on the state evolution
are provided in [25, 26, 140] for both AMP and G-AMP.
A statistical physics derivation of the state evolution in the spirit of the cavity method is given in [175] and is
formally very related to the state evolution of the TAP equation in sec. IV E. The main point in the derivation of the
state evolution is not to start from the G-AMP algorithm but from its message passing form, eqs. (148-154). One
then recalls that in belief propagation we assumed conditional independence of incoming messages which translates
into independence of terms present in sums of an extensive number of terms in the message passing. One then utilizes
the central limit theorem, enabling a reduction to the mean and variance of these terms matters. One simply needs
to keep track of the evolution of this mean and variance.
In statistical physics terms, the state evolution exactly corresponds to the derivation of the replica symmetric
equations via the cavity method. Since the cavity and replica methods always yield the same result, the state
evolution equations are exactly equivalent to the equations stemming from the replica method for the corresponding
66
problem. For compressed sensing this is illustrated in [175]. For the perceptron problem all the replica symmetric
equations and related results, e.g. in [310], can be obtained as a special case of the G-AMP state evolution using the
perceptron output.
Let us sketch briefly the derivation of state evolution for G-AMP. We follow the derivation of [153] for the specific
case of linear estimation. In order to write the equations for the MMSE for a general output functions, we need to
rewrite the output function using y = h(z, w) where w is a random variable distributed according to PW (w). Then
Z
Pout (y|z) = dwPW (w)δ[y − h(z, w)] .
(188)
P 2
P
Now we consider the random variable Vµ = i Fµi
vi . On average its value is i vi /N whereas its is variance o(1).
It thus concentrates around its mean as N → ∞ and therefore
1 X
vi .
(189)
V t+1 =
N i
P
P
Let us now turn our attention to ωµ→i = j6=i Fµi ai→j and zµ→i = j6=i Fµi si . These quantities are, according the
the assumptions of the cavity method, sums of uncorrelated variables, so they must converge to (correlated) Gaussian
variables with co-variances reading
E[ω 2 ] = E[a2 ] = q ,
E[zω] = E[sa] = m ,
E[z 2 ] = E[s2 ] = q0 ,
(190)
(191)
(192)
where we defined the order parameters m and q as the average correlation between the estimator a and the groundtruth signal s, and as a self-correlation for the estimator a, respectively.
Let us now see how Ri behaves (note the difference between the letters ω and w):
X
X
Ri
=
Bµ→i =
Fµi gout (ωµ→i , yµ , Vµ→i )
Σi
µ
µ
X
X
=
Fµi gout (ωµ→i , h(
Fµj sj + Fµi si , wµ ), V )
µ
=
(194)
j6=i
X
µ
(193)
Fµi gout (ωµ→i , h(
X
Fµj sj , wµ ), V ) + si αm̂ ,
(195)
j6=i
where we define
m̂ = Eω,z,w [∂z gout (ω, h(z, w), V )] .
(196)
p
Ri
= N (0, 1) αq̂ + si αm̂
Σi
(197)
We further write
with N (0, 1) being a random Gaussian variables of zero mean and unit variance, and where
2
q̂ = Eω,z,w gout
(ω, h(z, w), V )
(198)
while the computation of Σ simply gives, thanks to concentration,
1
= αq̂ .
Σ
Now we can close all equations by computing q and m from their definitions (190-191) as
q = Es ER,Σ fa2 (Σ, R)
m = Es [ER,Σ sfa (Σ, R)]
(199)
(200)
(201)
In the Bayes-optimal case, the Nishimori consition imposes q = m and thus provides further simplifications. In fact,
one can show that q̂ is also equal to m̂ [153]. In this case, the state evolution leads to the asymptotic Bayes-optimal
minimum mean squared error (MMSE) for the corresponding inference problem. Just as it did for the planted spin
67
glass or the clustering of networks above. Since in linear estimation we are working with a fully connected graphical
model the asymptotic analysis can be written in terms of only 2 scalar parameters [175].
Finally, writing explicitly the integrals, the state evolution reads
ξ2
Z
Z
ξ
e− 2 2
1
t+1
,x + √
,
(202)
m
= dx PX (x) dξ √ fa
αm̂t
2π
αm̂t
(z−p)2
p2
Z
Z
−
e− 2mt e 2(ρ0 −mt )
t
m̂ = − dw PW (w) dp dz p
∂p gout (p, h(z, w), ρ0 − mt ) ,
(203)
2π mt (ρ0 − mt )
where we denote ρ0 = N EF (F 2 )EPX (x2 ). For the sparse prior (134) where the non-zero elements have the second
moment equal to x2 and the matrix F having iid elements of zero mean and variance 1/N we get ρ0 = ρx2 . For the
usual AWGN output (132) the second equation simplifies considerably into
m̂t =
1
∆ + ρx2 − mt
.
(204)
The mean-squared error of the AMP estimator is then MSE = ρx2 − mt . To evaluate the performance of AMP
initialized in a way unrelated to the unknown signal, the state evolution is iterated until convergence starting from
mt=0 = ρ2 x2 , where x is the mean of the distribution φ(x) in (134).
2.
Free energy
The same procedure we use to derive the state evolution can be used to write a scalar expression for the Bethe
free energy (178). Alternatively, the same expression can be obtained using the replica method, as e.g. in [153]. The
average Bethe free energy for the generalized linear estimation problem can be written as f Bethe = minm,m̂ Φ(m, m̂)
where
R
R
√
√
√
√
Φ(m, m̂) =α dy Dξ DsPout (y|s ρ0 − m + ξ m) log DxPout (y|x ρ0 − m + ξ m)
hR
√
√ i
R
αm̂s2
αm̂x2
αmm̂
−
+ DξdsPX (s)e− 2 +ξs αm̂ log dxPX (x)e− 2 +ξx m̂
(205)
2
The free energy can be also written as a function of the MSE E. In the case of the AWGN channel with variance
∆, for instance, it reads f Bethe = minE Φ(E) where [175]:
α(ρx2 − E)
α α
Φ(E) = − − log (∆ + E) −
2
2
2(∆ + E)
Z
Z
Z
√
α
α
x
√
∆+E x(s− 2 )+ξx ∆+E
+ dsPX (s) Dξ log
dx e
PX (x) .
(206)
Actually, this formula is, again, very close to the one discussed in the context of CDMA by Tanaka [298] and Guo and
Verdú [116]. An equivalent form has been derived for compressed sensing by Verdú [315]. Much more can be done
with the replica method in this context (see, e.g. [261, 295]).
3.
Phase transitions
What we want to discuss here is again the appearance of the very generic phenomena we have discussed already in
the first chapter, and illustrated on Fig. 1: the presence of two different transitions when trying to solve the Bayesian
optimal problem.
In Fig. 15 we plot down to what measurement rate α = M/N AMP recovers a signal of density ρ for several distributions of non-zero signal elements φ(x). We observe a first order phase transition that depends on the distribution
of non-zeros.
Interestingly, the corresponding line is always better than the `1 transition, as shown in [80], where the distribution
maximizing the spinodal line for Bayes-optimal inference was computed. This is to be expected, since we might hope
that the spinodal for the Bayesian optimal problem is the “best” spinodal, though there is no generic proof of this.
One might wonder how a measurement noise is affecting this diagram? Ref. [175] further studied how the optimal
inference performance is influenced by measurement noise, mismatch between the prior and the actual signal distribution, how the performance changes under expectation maximization learning of the parameters, such as ρ or the mean
68
1
α=M/N
0.8
0.6
0.4
ℓ1
N(0,1)
N(1,1)
N(-1,0.1)+N(1,0.1)
0.2
0
0
0.2
0.4
0.6
0.8
1
ρ=K/N
FIG. 15. Phase diagram for the AMP reconstruction in the optimal Bayesian case when the signal model is matching the
empirical distribution of signal elements. The elements of the M × N measurement matrix F are iid variables with zero mean
and variance 1/N . The spinodal transition αs (ρ) is computed with the state evolution and plotted for the following signal
distributions: φ(x) = N (0, 1) (green), φ(x) = N (1, 1) (blue) φ(x) = [N (−1, 0.1) + N (1, 0.1)]/2 (magenta). The data are
compared to the Donoho-Tanner phase transition α`1 (ρ) (dashed) for `1 reconstruction that does not depend on the signal
distribution, and to the theoretical limit for exact reconstruction α = ρ (red). Figure taken from [175].
x and the variance σ of the signal distribution. Ref. [16] explored the role of approximately sparse signals when the
δ(xi ) in (134) is replaced by a narrow Gaussian. The influence of an uncertainty in the elements of the matrix F was
studied in [176]. As expected from statistical physics, where in the presence of disorder first order phase transitions
become weaker, second order or disappear entirely, the same happens to the first order phase transition in compressed
sensing under presence of various kinds of noise. Qualitative results are illustrated in [16, 175, 176].
Existence of phase transition is not restricted to compressed sensing. In fact, the very same phenomenology has
been discussed and identified as early as in the 90’s in the context of learning a rule via a perceptron with binary
weights [103, 210, 281, 310]. In the present setting, this corresponds to a binary prior for the signal, and a perceptron
channel given by eq. (185). In this case, it becomes possible to identify perfectly the signal when the ratio between the
number of examples M and the number of variables N is larger than α ≈ 1.245, a result first derived by Derrida and
Gardner [103]. It is only beyond α ≈ 1.49, however, that the G-AMP algorithm actually succeed in finding the hidden
rule (i.e. the planted weights). All these classic results can thus be recovered by the present, unifying, approach.
Note that, for this problem, the G-AMP (or TAP) equations has been written and proposed as early as in 1989 in
[210]. The more recent application of either relaxed belief propagation [45] or G-AMP to actual data [330] has also
been considered.
Another classical example of this approach with a similar phenomenology is given by the classical work of Tanaka
and Kabashima in the context of CDMA [150, 298].
E.
Spatial coupling
As discussed in section IV B, if nothing is changed about the graphical model corresponding to compressed sensing,
the 1st order phase transition presented in Fig. 15 causes a barrier that is conjectured to be unbeatable by any
polynomial algorithm. In an idealized setting of compressed sensing, however, the design of the measurement matrix
Fµi is entirely up to us. Thanks to this, compressed sensing belongs to the class of problems where spatial coupling
can be successfully applied. This was the main point of Ref. [173], where spatially coupled measurement matrices
were designed that allowed AMP to reconstruct the signal in compressed sensing anytime the measurement rate α
is larger than the density of non-zero elements in the signal ρ, i.e. all the way above the red line in Fig. 15. This
threshold saturation was soon after proven rigorously in [79]. We should note that spatial coupling for compressed
sensing was considered even before in [187], but in that work the observed improvement was limited and it did not
69
seem possible to achieve the information theoretical limit α = ρ.
A spatially coupled matrix suggested in [173] is illustrated in Fig. 16. The physical mechanism of why the spinodal
transition disappears with such spatially-coupled measurement matrices is exactly the same as for the spatially coupled
Currie-Weiss model discussed in section IV C. The first block of the matrix has a larger effective measurement rate and
plays the role of a nucleation seed. The interactions in the blocks next to the diagonal then ensure that the seed grows
into the whole system. Ref. [175] further discusses the choice of parameters of the spatially coupled measurement
matrices, other related designs and the effect of noise. With the goal of optimizing the spatially coupled matrices the
properties of the nucleation and necessary properties of the seed in compressed sensing were studied in [51].
0y1
B
B
B
B
@
0
C B
C B
C=B
C B
A @
F
1
J1
J2
1
J1
0
J2
1
J1
J2
1
J1
1 0s1
0
J2
1
J1
J2
1
J1
J2
1
J1
J2
1
: unit coupling
: coupling J1
: coupling J2
: no coupling (null elements)
C B
C B
C⇥ B
C B
A B
B
B
B
B
B
B
B
@
C
C
C
C
C
C
C
C
C
C
C
C
A
FIG. 16. Example of a spatially coupled measurement matrix F for compressed sensing. The elements of the matrix are taken
independently at random, their variance has the block diagonal structure as illustrated in the figure. The figure it taken from
[175].
There are several problems in the literature closely formally related to compressed sensing to which spatial coupling
was applied. One of them is, again, the CDMA defined in sec. VI A. When the spreading code is not random iid but
correspond to a spatially coupled random matrix then the performance of the CDMA system can be further improved
[277, 297].
Another problem formally very closely related to compressed sensing is the real valued error correction where the
codeword is corrupted by gross errors on a fraction of entries and a small noise on all the entries. In [18] it was shown
that the error correction and its robustness towards noise can be enhanced considerably using AMP and spatial
coupling.
A similar application of these ideas has allowed for the development of the spatially coupled sparse superposition
codes [14, 15, 19, 147, 270]. These are efficient coding schemes using exactly the compressed sensing setting with a
twist: the variables are L-dimensional and are forced to point into one corner of the L-hypercube. These codes, with
the associated AMP decoder, have been shown to be fast and reliable, and actually reach the Shannon limit for the
AWGN channel in the large L limit. It is actually expected that they are Shannon achieving for any memoryless
channel.
F.
Non-random matrices and non-separable priors
1.
Structured priors
AMP as described above uses the empirical distribution of signal elements. A natural question is whether its
performance can be improved further if more information is known about the signal. There are works that have
investigated this question and combined AMP with more complex, structured priors. Utilizing such structured priors
is the key to leveraging many of the advancements recently seen in statistical signal representation. For instance
techniques such as hybrid AMP [259, 278] have shown promising results. In those works, an additional graphical
model is introduced in order to take into account the correlation between variables. The improvement, for instance
for image denoising, can be spectacular.
70
Another possible approach to structured priors is not to model the correlations directly, but instead to utilize a
bipartite construction via hidden variables, as in the restricted Boltzmann machine (RBM) [128, 129]. If a binary
RBM can be trained to model the support pattern of a given signal class, then the statistical description of the
RBM admits its use within AMP, as was shown recently in [302]. This is particularly interesting since RBMs are
the building blocks of “deep belief networks” [29] and have recently sparked a surge of interest, partly because of
the efficient algorithms developed to train them (e.g. contrastive divergence (CD) [129]). The possibility, shown in
[302], to incorporate deep learned priors into generalized linear problems such as compressed sensing appears very
promising.
2.
Orthogonal matrices
Using non-random matrices for the linear projection is crucial to explore the range of applicability. The `1 minimization based approaches provably work well for a much larger class of matrices than random ones [123]. It
is therefore important to explore performance of the approach stemming from statistical physics for more general
matrices. There has been a collection of works in this direction, for instance in the analysis of the so-called perceptron
learning with correlated pattern (which form the sensing matrix) [152, 286].
As we have mentioned in Sec. IV E, there are many works in the context of the Ising model that have extended the
TAP approach for more generic random matrices. Similar contributions have recently emerged in the context of AMP
as well: the recent work on the so-called S-AMP algorithm that extends the approximate message-passing algorithm
to general matrix ensembles when there is a certain well-defined large system size limit, based on the so-called Stransform of the spectrum of the measurement matrix [48]. Older works, in fact, already used a similar approach for
the CDMA problem [296]. Analysis of performances in this direction is also doable [156, 195]. This shows that there
are still interesting open problems in this direction and we feel that there is still much more to be done in the context
of AMP beyond random matrices.
3.
Reconstruction in discrete tomography
Signal reconstruction in x-ray computed tomography (CT) is another problem that belongs to the class of linear estimation. The matrix Fµi corresponding to computed tomography has a very particular structure. Classical
reconstruction algorithms such as filtered back-projection require the corresponding linear system to be sufficiently
determined. Therefore, the number of measurements needed is generally close to the number of pixels to be reconstructed. Nevertheless, many applications would benefit from being able to reconstruct from a smaller number of
angles. Ref. [110] designed and tested a message passing algorithm that is adapted to the measures performed in
CT and is able to provide a reliable reconstruction with a number of measurements that is smaller than what was
required by previous techniques. This algorithm is another example of how inspiration from physics can provide a
promising algorithmic approach.
Each measurement in x-ray CT is well approximated by a sum of absorption coefficient of the material under
investigation along a straight line of the corresponding x-ray. Let us discretize the object under study into small
elements, and assign a discrete value xi ∈ {1, . . . , q} of the absorption coefficient to every small element i. One
component of the measurement in CT is then
X
yµ =
xi ,
(207)
i∈µ
where the sum is over all elements that lie on the line µ. The structural information that in general allows us to
decrease the number of measurement is that two pixels that lie next to each other mostly have the same value. This
expectation is represented in the following posterior distribution
M
P
Y
X
1
δ yµ −
P (x|y) =
xi eJµ (ij)∈µ δxi ,xj ,
(208)
Z µ=1
i∈µ
where the notation (ij) ∈ µ means the elements i and j are next to each other on the line corresponding to measurement
µ, Jµ is the interaction constant.
Ref. [110] writes the belief propagation algorithm for estimating marginals of this posterior distribution. However,
one update of this belief propagation involves intractable sums over all configurations of variables on one line. We
realize that these sums correspond to the partition functions of one-dimensional Potts model with fixed magnetization.
71
A one-dimensional line is a tree and for trees partition functions can be computed exactly with belief propagation.
We use this to design an algorithm that uses BP within BP. The resulting implementation is comparably fast to the
competing convex relaxation based algorithms and provides better reconstruction and robustness to noise. We believe
that overall the algorithm suggested in [110] provides a promising perspective for reconstruction in x-ray computed
tomography.
G.
An application in physics: Optics in complex media
We would like to close the loop with physics and mention one explicit physics experiment in the field of compressed
optics where AMP and related tools have been especially useful.
Wave propagation in complex media is a fundamental problem in physics, be it in acoustics, optics, or electromagnetism. In optics, it is particularly relevant for imaging applications. Indeed, when light passes through a
multiple-scattering medium, such as a biological tissue or a layer of paint, ballistic light is rapidly attenuated, preventing conventional imaging techniques, and random scattering events generate a so-called speckle pattern that is
usually considered useless for imaging.
Recently, however, wavefront shaping using spatial light modulators has emerged as a unique tool to manipulate
multiply scattered coherent light, for focusing or imaging in scattering media [230]. This makes it possible to measure
the so-called transmission matrix of the medium [254], which fully describes light propagation through the linear
medium, from the modulator de vice to the detector. Interestingly, these transmission matrices, due to the fact that
they account for the multiple scattering and interference, are essentially iid random matrices [254], which is exactly
what AMP techniques have been designed for.
An important issue in most such recent experiments lies in accessing the amplitude and phase of the output field,
that in optics usually requires a holographic measurement, i.e., the use of a second reference beam. The phase and
amplitude of the measured field can then be extracted by simple linear combinations of interference patterns with a
phase shifted or off-axis reference. This however poses the unavoidable experimental problem of the interferometric
stability of the reference arm. In order to be able to avoid entirely the use of the reference beam, we need to solve a
problem that looks like compressed sensing: finding x (sparse or binary) such that y = |F x| where F is complex and
iid. Combining the G-AMP and the SwAMP approach for phase retrieval [201, 279], it has been possible to precisely
do this, to calibrate efficiently, and to perform a reference-less measurement of the transmission matrix of a highly
scattering material. Within this framework, the work of [84] showed that the transfer matrix can still be retrieved
for an opaque material (a thin layer of white paint), and can then be used for imaging or focusing, using the opaque
complex material as a lens. This was experimentally validated.
As a result the full complex-valued transmission matrix of a strongly scattering material can be estimated, up to a
global phase, with a simple experimental setup involving only real-valued inputs and outputs in combination with the
techniques presented in this review. We believe this illustrates the power of the techniques presented in this review,
and shows that statistical physics does not only provide interesting perspective to inference problems, but also that
inference techniques can be useful for physics experiments in return.
H.
Beyond linear estimation: matrix factorization
The technics that have been used in the context of (generalized) linear estimation are not limited to this setting.
In fact, AMP can also be applied to the more challenging setting of matrix factorization.
Decomposing a matrix into a product of two matrices of given properties is another problem with an huge range
of applications. Mathematically stated, we are given a noisy measurement of a matrix Y of dimension M × N and
we aim to decompose it into a product Y = DX, where D is a M × R matrix, and X is a R × N matrix. Depending
on the application the problem is subject to various constraints such as sparsity of D or X, low rank R, or specific
models for the noise. Such constraints turn the matrix factorization into an algorithmically extremely challenging
problem. Let us list several examples of potential applications
• Dictionary learning: Many modern signal processing devices take a great advantage of the fact that almost every
class of signals of interest has a sparse representation, i.e. there exists a basis such that the observed signal can
be written as a sparse linear combination of atoms (i.e. columns of the corresponding matrix) from this basis.
Widely known, studied and used examples include the wavelet basis for images, of Fourier basis for acoustic
signals. For other examples of interest, especially those which utilize uncommon or novel data types, a suitable
basis for sparse representation may not yet be known. The goal of dictionary learning [85, 199, 200] is to learn
such a basis purely from samples of the data. Naturally the more samples are available the easier is this task.
72
The challenge is hence to design an algorithm that would be able to learn the dictionary efficiently from the
smallest possible number of samples.
In the mathematical statement of the problem the N columns of matrix Y represent the N samples of the data
for which we aim to find a sparse representation. M is then the dimension of each of the data points. The
dictionary learning problem is a decomposition of the matrix Y into a product Y = DX + W , where D is the
dictionary of R so-called atoms, and X is the sparse representation of the data, having only a fraction ρ of
non-zero elements in each column. W is a possible noise that may be taking into account the approximative
nature of the sparse representation.
• Feature learning in neural network: Arguably one of the most impressive engineering success of the last decade
is the design of the so called “deep-learned” neural networks [30, 130, 131]. Using these, computers are now
able to recognize people on video, to tell a dog from a cat on a picture, to process speech efficiently and even
to answer complicated questions. At the roots of this efficiency is the ability to learn features. The simplest
building block of deep neural networks can be seen as a matrix factorization problem with an output Pout (y|z)
corresponding to the activation functions.
• Blind source separation: A typical example of blind source separation [28, 149] are M sensors recording noise
at a party during time N . We then aim to separate the speech of the R participants of the party based on these
recordings. Recording at one of the sensors is an unknown linear combination of the speech of all the present
people. A particularly challenging, but also very interesting application-wise, is when the number of sensors is
smaller than the number of people, M < R. The individual speeches may then be in principle recovered only if
they can be represented as sparse in some known basis.
• Sparse PCA: In PCA the aim is to approximate the matrix Y by a low-rank matrix, i.e. columns of Y are
written as a linear combination of only few elements. This reduces the information needed to describe the data
contained in Y . This is usually done via the singular value decomposition (SVD). In some applications it is
desirable that the linear combination to be sparse [145, 331], thus further reducing the necessary information.
The sparsity constraint cannot be straightforwardly included in the SVD and hence alternative algorithmic
approaches are needed.
1.
Learning random sparse dictionaries
The matrix factorization problem can again be studied within the teacher-student scenario: The teacher generates
a the matrices F and X with random iid entries, from some probability distribution (possibly sparse) and provides
to the student element-wise measurement of the product F X via an output function Pout (Y |F X). Let us consider
the dimensions M, N, R to be all of the same order. Both the AMP [153, 177, 248, 305] and the replica approach
[153, 177, 275, 276] can be applied to such problems. In this the problem of identifiability of F and X from Y , the
corresponding sample complexity and the associated Bayes-optimal mean squared error can be derived. We refer to
these works for the phase diagrams for special cases of the problem such as the sparse dictionary learning, blind source
separation, sparse PCA, robust PCA or matrix completion.
These results are very striking because in a number of relevant cases (including the dictionary learning and source
separation) it is observed that the ideal performance predicted by this approach is far better than the one achieved
by current algorithms. This suggests that there is a lot to gain in algorithmic efficiency in such problems. Making
the algorithm of [153, 248] more robust along the lines described above for AMP for linear estimation and applying
it to more realistic data is a very promising subject of future work.
2.
Low rank decomposition
In some part of the applications mentioned above (e.g. the PCA) the dimension R, corresponding to the rank
of matrix Y , is considered as very small. In those cases it is reasonable to consider the limit where M and N
are comparable and both N, M → ∞ whereas R = O(1). The graphical model corresponding to this case has Rdimensional variables and pairwise interactions. Problems that can be represented via this graphical model include
variations of PCA, submatrix localization problem, biclustering, clustering.
Again AMP type of algorithm can be written for a general element-wise output function Pout (Y |F X), and the
associated state evolution leads to an analysis of the phase diagram, phase transitions and optimal mean squared
errors. Interestingly in this approach the prior distribution can be any R-variate distribution, thus allowing for
instance for priors representing membership to one or more clusters.
73
A problem corresponding to the rank one matrix factorization, clustering in the Gaussian mixture model, was studied
long time ago using the replica method [20, 35, 311]. It was presented as a prototypical example of unsupervised
learning, see e.g. chapter 8 in the textbook [86]. Note, however, that many more settings closer to current unsupervised
learning architectures should be analyzed. The revival of interest in low-rank matrix factorization was linked with the
development of the corresponding approximate message passing algorithm by Rangan and Fletcher for rank R = 1
[258], and Matsushita and Tanaka for general rank [206] followed by rigorous work of [76]. Recently, a version for
generic rank and generic output (as in G-AMP) has been presented and analysed in [193, 194]. The resulting algorithm
when tested numerically has good convergence properties and these works hence open the stage to further studies and
applications along the line discussed in this review.
As a final comment, it is to be noted that the phase transitions observed in the matrix factorization setting are
again of the same kind that we have described all along this review, schematically in Fig. 1. In many problems, the
presence of distinct algorithmic and information theoretic thresholds is observed. In this setting, it is particularly
interesting to characterize when spectral methods [11], the most widely used approach for low rank factorization, are
optimal, and when they are not. As shown in, e.g. [75, 193, 194, 223], there are a number of situations where they are
sub-optimal. This is a valuable information one can get from the physics approach, that should motivate the creation
of new algorithms approaching the best possible performance.
74
VII.
PERSPECTIVES
This review discusses recent progress in the statistical physics approach to understanding of different problems of
statistical inference. Just as theoretical physics often focuses on understanding of idealized models that represent
in a simplified way the salient features of realistic settings, we focus on the teacher-student setting under which the
Bayes-optimal inference can be analyzed without ambiguity, and the phase transition can be identified and discussed.
The main concept that we presented in detail is the mathematical relation between the Bayes-optimal inference and
properties of disordered systems on the co-called Nishimori line. We discussed in detail various phase transitions that
arrise in the Bayes-optimal inference and related them to statistical and computational thresholds in the corresponding
inference problems. The resulting picture, that we explain in detail on the simple example of the planted spin glass, is
relevant for a large class of inference problems. In later sections we illustrate this picture on the problem of clustering
of sparse networks and on the problem of generalized linear estimation. In today’s scientific landscape the number of
very interesting data-driven problems is considerable and, in the authors’ opinion, in a considerable number of them
the strategy presented in this review will bring exciting new results.
Let us recall here the main theme of the present review. For all the presented problems, the central scientific questions that one can attempt to answer are: (1) Under what conditions is the information contained in the measurements
sufficient for a satisfactory inference to be possible? (2) What are the most efficient algorithms for this task? In the
probabilistic setting discussed here, what is striking is that all of these problems have a very similar phenomenology,
that also appeared earlier when physicists studied error correcting codes and perceptron learning, and that often these
two questions can be precisely answered. When the amount of information is too low a successful inference of the
signal is not possible for any algorithm: the corresponding information is simply insufficient. On the other hand, for
large enough amount of data, inference is possible, and the two regime are separated by a sharp phase transition.
Finally, and perhaps more importantly, there is often (as in first order transition in physics), an intermediate regime
where a successful inference is in principal possible but algorithmically hard. We have shown many example where
these transitions can be computed. A first order phase transition is always linked to appearance of computational
hardness, whereas a second order phase transition is not.
There is a deep value in this knowledge: If these transitions were not known, one could not know how good are the
state-or-the-art algorithms. To give an example, it is the very knowledge of the precise Shannon bound that is driving
research in information theory: if we know that we could be better in principle, when it is worth working on better
algorithms. As we have discussed in the very last chapter, for instance, present day matrix factorization algorithms
are often very far from optimality.
It is thus natural that, next to the theoretical understanding, an important outcome of the research reviewed in
this manuscript are new algorithmic ideas that are applicable beyond the idealized setting for which the presented
theoretical results hold. This translation of theoretical physics calculations into algorithmic strategies is very fruitful,
and promising direction of research. What is, perhaps, more challenging is to push such algorithmic ideas into actual
applications. From a point of view of a physicist, some of the problems presented here may seem very applied, but
in reality the research presented here is positioned on the theoretical side of the fields of statistics, machine learning
or signal processing. There is still a large step towards actual applications. Surely this is a very exciting direction to
explore and many of the researchers active in this field do explore it.
Let us now discuss some of the limitations and challenges related to the results and concepts presented in this
review.
On the algorithmic side a large part of the algorithms that stem from the research presented in this review involve
some kind of message passing derived under various assumptions of correlation decay. These assumptions rarely
hold of real datasets and hence sometimes this causes a failure of the corresponding algorithm. In our opinion the
theoretical potential of these message passing algorithms is very large, but much work needs to be done in order to
render these algorithms more robust and suitable for a generic problem.
Computer science and related fields mostly aim at understanding the worst possible setting or to prove something
under the weakest possible conditions. Analysis in a setting such as the teacher-student scenario usually does not
seem generic enough. In machine learning and more applied parts of computer science the focus is on a couple of
benchmark-datasets (such as e.g. the MNIST database). Simple models that aim to represent reasonably well the
generic situation are a domain of physical sciences. This is how progress in physics is made, and in our opinion this
is transferable to other fields as well.
Physicists can use with large effectiveness tools and methods from statistical physics of disordered systems such as
glasses and spin glasses. The advantage of these methods is that, unlike any other existing generic method, they are
able to describe the Bayes-optimal probabilistic inference. The disadvantage is that they are not fully rigorous and
are therefore sometimes viewed with dismissal or skepticism by mathematically oriented researchers. On the other
hand, rigorous proofs of results obtained with the cavity and replica method are a great mathematical challenge and
many partial results were obtained that have led to new mathematical tools making the whole probability theory more
75
powerful. One of the goals of the presented research is to transfer the physics reasoning and insight to mathematicians
in order to help them to craft plausible proof strategies.
abbreviation definition section
meaning
MMO
IB4
maximum mean overlap
MMSE
IB4
minimum mean squared error
MAP
IB4
maximum a posterior estimator
MCMC
IC3
Monte Carlo Markov chain
BP
IC3
belief propagation
RS
II F
replica symmetry
RSB
II F
replica symmetry breaking
1RSB
II F
one-step replica symmetry breaking
d1RSB
II F
dynamical one-step replica symmetry breaking
FRSB
II F
full-step replica symmetry breaking
XOR-SAT
II H 2
XOR-satisisfiability problem
ER
III A
Erdős-Rényi
V-B
III C
Viana-Bray model
LDPC
III F
low-density parity-check codes
CSP
III F
constraint satisfaction problem
K-SAT
III F
K-satisfiability problem
PCA
IV A
principal component analysis
C-W
IV C
Curie-Weiss model
TAP
IV E
Thouless-Anderson-Palmer
SBM
VA
stochastic block model
AWGN
VI
additive white Gaussian noise
CDMA
VI A
code-division multiple access
LASSO
VI A
least absolute shrinkage and selection operator
AMP
VI C
approximate message passing
G-AMP
VI C
generalized approximate message passing
r-BP
VI C 1
relaxed belief propagation
RBM
VI F
restricted Boltzmann machine
CT
VI F 3
computed tomography
TABLE I. Glossary of abbreviations with the number of the section where the term is defined.
76
ACKNOWLEDGEMENTS
We would like to express our gratitude to all our colleagues with whom many of the results presented here have been
obtained. In particular, we wish to thank Maria-Chiara Angelini, Jean Barbier, Francesco Caltagirone, T. Castellani,
Michael Chertkov, Andrea Crisanti, Leticia F. Cugliandolo, Laurent Daudet, Aurelien Decelle, Angelique Drémeau,
Laura Foini, Silvio Franz, Sylvain Gigan, Emmanuelle Gouillart, Jacob E. Jensen, Yoshyiuki Kabashima, Brian Karrer,
Lukas Kroc, Srinivas Kudekar, Jorge Kurchan, Marc Lelarge, Antoine Liutkus, Thibault Lesieur, Luca Leuzzi, André
Manoel, David Martina, Marc Mézard, Andrea Montanari, Cristopher Moore, Richard G. Morris, Elchanan Mossel,
Joe Neeman, Mark Newman, Hidetoshi Nishimori, Boshra Rajaei, Sundeep Rangan, Joerg Reichardt, Federico RicciTersenghi, Alaa Saade, David Saad, Ayaka Sakata, Fran cois Sausset, Christian Schmidt, Christophe Schülke, Guilhem
Semerjian, Cosma R. Shalizi, David Sherrington, Alan Sly, Phil Schniter, Eric W. Tramel, Rudiger Urbanke, Massimo
Vergassola, Jeremy Vila, Xiaoran Yan, Sun Yifan, Francesco Zamponi, Riccardo Zecchina and Pan Zhang.
A part of this manuscript has been written during a visit to UC Berkeley as part of the semester long program
on ”Counting Complexity and Phase Transitions”, at the Simons Institute for the Theory of Computing, which we
thanks for the kind hospitality.
Finally, we also wish to thank Alena and Julie for their continuous support and for letting us work, if only from
time to time, on this review.
FUNDINGS
Part of the research leading to the presented results has received funding from the European Research Council
under the European Union’s 7th Framework Programme (FP/2007-2013/ERC Grant Agreement 307087-SPARCS).
[1] Gamp in matlab. http://gampmatlab.wikia.com/wiki/Generalized Approximate Message Passing.
[2] Emmanuel Abbe, Afonso S Bandeira, Annina Bracher, and Amit Singer. Decoding binary node labels from censored
edge measurements: Phase transition and efficient recovery. Network Science and Engineering, IEEE Transactions on,
1(1):10–22, 2014.
[3] Emmanuel Abbe and Andrea Montanari. Conditional random fields, planted constraint satisfaction and entropy concentration. In Approximation, Randomization, and Combinatorial Optimization. Algorithms and Techniques, pages 332–346.
Springer, 2013.
[4] Dimitris Achlioptas and Amin Coja-Oghlan. Algorithmic barriers from phase transitions. In Foundations of Computer
Science, 2008. FOCS’08. IEEE 49th Annual IEEE Symposium on, pages 793–802. IEEE, 2008.
[5] Dimitris Achlioptas and Cristopher Moore. The asymptotic order of the random k-SAT threshold. In Foundations of
Computer Science, 2002. Proceedings. The 43rd Annual IEEE Symposium on, pages 779–788. IEEE, 2002.
[6] Dimitris Achlioptas and Yuval Peres. The threshold for random K-SAT is 2k log 2 − o(k). Journal of the American
Mathematical Society, 17(4):947–973, 2004.
[7] Lada A Adamic and Natalie Glance. The political blogosphere and the 2004 us election: divided they blog. In Proceedings
of the 3rd international workshop on Link discovery, pages 36–43. ACM, 2005.
[8] Noga Alon, Itai Benjamini, Eyal Lubetzky, and Sasha Sodin. Non-backtracking random walks mix faster. Communications
in Contemporary Mathematics, 9(04):585–603, 2007.
[9] Noga Alon, Michael Krivelevich, and Benny Sudakov. Finding a large hidden clique in a random graph. Random Structures
and Algorithms, 13(3-4):457–466, 1998.
[10] Maria Chiara Angelini, Francesco Caltagirone, Florent Krzakala, and Lenka Zdeborová. Spectral detection on sparse
hypergraphs. arXiv preprint arXiv:1507.04113, 2015.
[11] Jinho Baik, Gérard Ben Arous, and Sandrine Péché. Phase transition of the largest eigenvalue for nonnull complex sample
covariance matrices. Annals of Probability, pages 1643–1697, 2005.
[12] Victor Bapst, Guilhem Semerjian, and Francesco Zamponi. Effect of quantum fluctuations on the coloring of random
graphs. Physical Review A, 87(4):042322, 2013.
[13] David Barber. Bayesian reasoning and machine learning. Cambridge University Press, 2012.
[14] Jean Barbier and Florent Krzakala. Replica analysis and approximate message passing decoder for superposition codes.
In Information Theory (ISIT), 2014 IEEE International Symposium on, pages 1494–1498. IEEE, 2014.
[15] Jean Barbier and Florent Krzakala. Approximate message-passing decoder and capacity-achieving sparse superposition
codes. arXiv preprint arXiv:1503.08040, 2015.
[16] Jean Barbier, Florent Krzakala, Marc Mézard, and Lenka Zdeborová. Compressed sensing of approximately-sparse signals:
Phase transitions and optimal reconstruction. In Communication, Control, and Computing (Allerton), 2012 50th Annual
Allerton Conference on, pages 800–807. IEEE, 2012.
77
[17] Jean Barbier, Florent Krzakala, Lenka Zdeborová, and Pan Zhang. The hard-core model on random graphs revisited. In
Journal of Physics: Conference Series, volume 473, page 012021. IOP Publishing, 2013.
[18] Jean Barbier, Florent Krzakala, Lenka Zdeborová, and Pan Zhang. Robust error correction for real-valued signals via
message-passing decoding and spatial coupling. In Information Theory Workshop (ITW), 2013 IEEE, pages 1–5. IEEE,
2013.
[19] Jean Barbier, Christophe Schülke, and Florent Krzakala. Approximate message-passing with spatially coupled structured
operators, with applications to compressed sensing and sparse superposition codes. Journal of Statistical Mechanics:
Theory and Experiment, 2015(5):P05013, 2015.
[20] N Barkai and Haim Sompolinsky. Statistical mechanics of the maximum-likelihood density estimation. Physical Review
E, 50(3):1766, 1994.
[21] Dror Baron, Shriram Sarvotham, and Richard G Baraniuk. Bayesian compressive sensing via belief propagation. Signal
Processing, IEEE Transactions on, 58(1):269–280, 2010.
[22] A. Barrat, S. Franz, and G. Parisi. Temperature evolution and bifurcation of metastables states in mean-field spin glasses,
with connections with structural glasses. J. Phys. A, 30:5593–5612, 1997.
[23] Wolfgang Barthel, Alexander K Hartmann, Michele Leone, Federico Ricci-Tersenghi, Martin Weigt, and Riccardo
Zecchina. Hiding solutions in random satisfiability problems: A statistical mechanics approach. Physical Review Letters,
88(18):188701, 2002.
[24] Hyman Bass. The Ihara-Selberg zeta function of a tree lattice. International Journal of Mathematics, 3(06):717–797,
1992.
[25] Mohsen Bayati, Marc Lelarge, Andrea Montanari, et al. Universality in polytope phase transitions and message passing
algorithms. The Annals of Applied Probability, 25(2):753–822, 2015.
[26] Mohsen Bayati and Andrea Montanari. The dynamics of message passing on dense graphs, with applications to compressed
sensing. Information Theory, IEEE Transactions on, 57(2):764–785, 2011.
[27] Matthew James Beal. Variational algorithms for approximate Bayesian inference. University of London, 2003.
[28] Adel Belouchrani, Karim Abed-Meraim, Jean-François Cardoso, and Eric Moulines. A blind source separation technique
using second-order statistics. Signal Processing, IEEE Transactions on, 45(2):434–444, 1997.
[29] Yoshua Bengio. Learning deep architectures for AI. Foundations and Trends R in Machine Learning, 2(1):1–127, 2009.
[30] 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.
[31] Quentin Berthet and Philippe Rigollet. Computational lower bounds for sparse pca. arXiv preprint arXiv:1304.0828,
2013.
[32] Ludovic Berthier and Giulio Biroli. Theoretical perspective on the glass transition and amorphous materials. Reviews of
Modern Physics, 83(2):587, 2011.
[33] H. A. Bethe. Statistical physics of superlattices. Proc. Roy. Soc. London A, 150:552–575, 1935.
[34] Hans A Bethe. Statistical theory of superlattices. Proceedings of the Royal Society of London. Series A, Mathematical
and Physical Sciences, 150(871):552–575, 1935.
[35] Michael Biehl and Andreas Mietzner. Statistical mechanics of unsupervised structure recognition. Journal of Physics A:
Mathematical and General, 27(6):1885, 1994.
[36] Giulio Biroli and Jean-Philippe Bouchaud. The random first-order transition theory of glasses: a critical assessment.
Structural Glasses and Supercooled Liquids: Theory, Experiment, and Applications, pages 31–113, 2012.
[37] B. Bollobás. Random graphs. Cambridge University Press, second edition, 2001.
[38] Béla Bollobás, Christian Borgs, Jennifer Chayes, Oliver Riordan, et al. Percolation on dense graph sequences. The Annals
of Probability, 38(1):150–183, 2010.
[39] Erwin Bolthausen. An iterative construction of solutions of the TAP equations for the Sherrington–Kirkpatrick model.
Communications in Mathematical Physics, 325(1):333–366, 2014.
[40] Charles Bordenave, Marc Lelarge, and Laurent Massoulié. Non-backtracking spectrum of random graphs: community
detection and non-regular ramanujan graphs. In Foundations of Computer Science (FOCS), 2015 IEEE 56th Annual
Symposium on, pages 1347–1357. IEEE, 2015.
[41] J.-P. Bouchaud, L. Cugliandolo, J. Kurchan, and M Mézard. Out of equilibrium dynamics in spin glasses and other glassy
systems. In A. P. Young, editor, Spin Glasses and Random Fields, pages 161–223. World Scientific, Singapore, 1998.
[42] Jean-Philippe Bouchaud and Marc Potters. Theory of financial risk and derivative pricing: from statistical physics to
risk management. Cambridge university press, 2003.
[43] Petros T Boufounos and Richard G Baraniuk. 1-bit compressive sensing. In Information Sciences and Systems, 2008.
CISS 2008. 42nd Annual Conference on, pages 16–21. IEEE, 2008.
[44] DR Bowman and K Levin. Spin-glass theory in the bethe approximation: Insights and problems. Physical Review B,
25(5):3438, 1982.
[45] A. Braunstein and R. Zecchina. Learning by Message Passing in Networks of Discrete Synapses. Physical Review Letters,
96(3):030201, 2006.
[46] Joseph D Bryngelson and Peter G Wolynes. Spin glasses and the statistical mechanics of protein folding. Proceedings of
the National Academy of Sciences, 84(21):7524–7528, 1987.
[47] T Tony Cai, Tengyuan Liang, and Alexander Rakhlin. Computational and statistical boundaries for submatrix localization
in a large noisy matrix. arXiv preprint arXiv:1502.01988, 2015.
[48] Burak Cakmak, Ole Winther, and Bernard H Fleury. S-amp: Approximate message passing for general matrix ensembles.
In Information Theory Workshop (ITW), 2014 IEEE, pages 192–196. IEEE, 2014.
78
[49] Francesco Caltagirone, Silvio Franz, Richard G Morris, and Lenka Zdeborová. Dynamics and termination cost of spatially
coupled mean-field models. Physical Review E, 89(1):012102, 2014.
[50] Francesco Caltagirone, Giorgio Parisi, and Tommaso Rizzo. Dynamical critical exponents for the mean-field potts glass.
Physical Review E, 85(5):051504, 2012.
[51] Francesco Caltagirone and Lenka Zdeborová. Properties of spatial coupling in compressed sensing. arXiv preprint
arXiv:1401.6380, 2014.
[52] Francesco Caltagirone, Lenka Zdeborová, and Florent Krzakala. On convergence of approximate message passing. In
Information Theory (ISIT), 2014 IEEE International Symposium on, pages 1812–1816. IEEE, 2014.
[53] Chiara Cammarota and Giulio Biroli. Ideal glass transitions by random pinning. Proceedings of the National Academy of
Sciences, 109(23):8850–8855, 2012.
[54] Emmanuel J Candè and Michael B Wakin. An introduction to compressive sampling. Signal Processing Magazine, IEEE,
25(2):21–30, 2008.
[55] Emmanuel J Candes and Terence Tao. Near-optimal signal recovery from random projections: Universal encoding
strategies? Information Theory, IEEE Transactions on, 52(12):5406–5425, 2006.
[56] JM Carlson, JT Chayes, L Chayes, JP Sethna, and DJ Thouless. Bethe lattice spin glass: the effects of a ferromagnetic
bias and external fields. i. bifurcation analysis. Journal of statistical physics, 61(5-6):987–1067, 1990.
[57] JM Carlson, JT Chayes, James P Sethna, and DJ Thouless. Bethe lattice spin glass: the effects of a ferromagnetic
bias and external fields. ii. magnetized spin-glass phase and the de almeida-thouless line. Journal of statistical physics,
61(5-6):1069–1084, 1990.
[58] Tommaso Castellani, Florent Krzakala, and Federico Ricci-Tersenghi. Spin glass models with ferromagnetically biased
couplings on the Bethe lattice: analytic solutions and numerical simulations. The European Physical Journal B-Condensed
Matter and Complex Systems, 47(1):99–108, 2005.
[59] Alain Celisse, Jean-Jacques Daudin, Laurent Pierre, et al. Consistency of maximum-likelihood and variational estimators
in the stochastic block model. Electronic Journal of Statistics, 6:1847–1899, 2012.
[60] Patrick Charbonneau, Yuliang Jin, Giorgio Parisi, Corrado Rainone, Beatriz Seoane, and Francesco Zamponi. Numerical
detection of the gardner transition in a mean-field glass former. Physical Review E, 92(1):012316, 2015.
[61] Patrick Charbonneau, Yuliang Jin, Giorgio Parisi, and Francesco Zamponi. Hopping and the stokes–einstein relation
breakdown in simple glass formers. Proceedings of the National Academy of Sciences, 111(42):15025–15030, 2014.
[62] Peter Cheeseman, Bob Kanefsky, and William M. Taylor. Where the Really Hard Problems Are. In Proc. 12th IJCAI,
pages 331–337, San Mateo, CA, USA, 1991. Morgan Kaufmann.
[63] Aaron Clauset, Mark EJ Newman, and Cristopher Moore. Finding community structure in very large networks. Physical
Review E, 70(6):066111, 2004.
[64] Amin Coja-Oghlan. Graph partitioning via adaptive spectral techniques. Combinatorics, Probability and Computing,
19(02):227–284, 2010.
[65] Amin Coja-Oghlan and André Lanka. Finding planted partitions in random graphs with general degree distributions.
SIAM Journal on Discrete Mathematics, 23(4):1682–1714, 2009.
[66] Amin Coja-Oghlan, Elchanan Mossel, and Dan Vilenchik. A spectral approach to analysing belief propagation for 3colouring. Combinatorics, Probability and Computing, 18(06):881–912, 2009.
[67] A Crisanti, H Horner, and H-J Sommers. The sphericalp-spin interaction spin-glass model. Zeitschrift für Physik B
Condensed Matter, 92(2):257–271, 1993.
[68] L. F. Cugliandolo and J. Kurchan. Analytical solution of the off-equilibrium dynamics of a long-range spin glass model.
Phys. Rev. Lett., 71:173, 1993.
[69] J. R. L. de Almeida and D. J. Thouless. Stability of the Sherrington-Kirkpatrick solution of a spin-glass model. J. Phys.
A, 11:983–990, 1978.
[70] Bruno De Finetti. Theory of Probability. A critical introductory treatment. John Wiley & Sons, 1979.
[71] Aurelien Decelle, Florent Krzakala, Cristopher Moore, and Lenka Zdeborová. Asymptotic analysis of the stochastic block
model for modular networks and its algorithmic applications. Physical Review E, 84(6):066106, 2011.
[72] Aurelien Decelle, Florent Krzakala, Cristopher Moore, and Lenka Zdeborová. Inference and phase transitions in the
detection of modules in sparse networks. Physical Review Letters, 107(6):065701, 2011.
[73] Amir Dembo, Andrea Montanari, et al. Ising models on locally tree-like graphs. The Annals of Applied Probability,
20(2):565–592, 2010.
[74] Arthur P Dempster, Nan M Laird, and Donald B Rubin. Maximum likelihood from incomplete data via the em algorithm.
Journal of the royal statistical society. Series B (methodological), pages 1–38, 1977.
[75] Yash Deshpande and Andrea Montanari. Finding hidden cliques of size\ sqrt {N/e} in nearly linear time. Foundations
of Computational Mathematics, pages 1–60, 2015.
[76] Yateendra Deshpande and Andrea Montanari. Information-theoretically optimal sparse PCA. In Information Theory
(ISIT), 2014 IEEE International Symposium on, pages 2197–2201. IEEE, 2014.
[77] William E Donath and Alan J Hoffman. Lower bounds for the partitioning of graphs. IBM Journal of Research and
Development, 17(5):420–425, 1973.
[78] David L Donoho. Compressed sensing. Information Theory, IEEE Transactions on, 52(4):1289–1306, 2006.
[79] David L Donoho, Adel Javanmard, and Andrea Montanari. Information-theoretically optimal compressed sensing via
spatial coupling and approximate message passing. Information Theory, IEEE Transactions on, 59(11):7434–7464, 2013.
[80] David L Donoho, Iain Johnstone, and Andrea Montanari. Accurate prediction of phase transitions in compressed sensing
via a connection to minimax denoising. Information Theory, IEEE Transactions on, 59(6):3396–3433, 2013.
79
[81] David L Donoho, Arian Maleki, and Andrea Montanari. Message-passing algorithms for compressed sensing. Proceedings
of the National Academy of Sciences, 106(45):18914–18919, 2009.
[82] David L. Donoho, Arian Maleki, and Andrea Montanari. Message-passing algorithms for compressed sensing. Proc. Natl.
Acad. Sci., 106(45):18914–18919, 2009.
[83] David L Donoho and Jared Tanner. Sparse nonnegative solution of underdetermined linear equations by linear programming. Proceedings of the National Academy of Sciences of the United States of America, 102(27):9446–9451, 2005.
[84] Angélique Drémeau, Antoine Liutkus, David Martina, Ori Katz, Christophe Schülke, Florent Krzakala, Sylvain Gigan,
and Laurent Daudet. Reference-less measurement of the transmission matrix of a highly scattering material using a dmd
and phase retrieval techniques. Optics express, 23(9):11898–11911, 2015.
[85] Michael Elad and Michal Aharon. Image denoising via sparse and redundant representations over learned dictionaries.
Image Processing, IEEE Transactions on, 15(12):3736–3745, 2006.
[86] Andreas Engel and Christian P. L. Van den Broeck. Statistical Mechanics of Learning. Cambridge University Press, New
York, NY, USA, 2001.
[87] P. Erdős and A. Rényi. On random graphs. Publ. Math. Debrecen, 6:290–297, 1959.
[88] P. Erdős and A. Rényi. On the evolution of random graphs. Publ. Math. Inst. Hungar. Acad. Sci., 5:17–61, 1960.
[89] Vitaly Feldman, Will Perkins, and Santosh Vempala. On the complexity of random satisfiability problems with planted
solutions. In Proceedings of the Forty-Seventh Annual ACM on Symposium on Theory of Computing, pages 77–86. ACM,
2015.
[90] A. J. Felstrom and K.S. Zigangirov. Time-varying periodic convolutional codes with low-density parity-check matrix.
IEEE Trans. Inf. Theory, 45(6):2181 –2191, 1999.
[91] Miroslav Fiedler. Algebraic connectivity of graphs. Czechoslovak mathematical journal, 23(2):298–305, 1973.
[92] Alyson K Fletcher, Sundeep Rangan, Lav R Varshney, and Aniruddha Bhargava. Neural reconstruction with approximate
message passing (neuramp). In Advances in neural information processing systems, pages 2555–2563, 2011.
[93] Laura Foini, Florent Krzakala, and Francesco Zamponi. On the relation between kinetically constrained models of
glass dynamics and the random first-order transition theory. Journal of Statistical Mechanics: Theory and Experiment,
2012(06):P06013, 2012.
[94] Santo Fortunato. Community detection in graphs. Physics Reports, 486(3):75–174, 2010.
[95] S. Franz, M. Mézard, F. Ricci-Tersenghi, M. Weigt, and R. Zecchina. A ferromagnet with a glass transition. Europhys.
Lett., 55:465, 2001.
[96] Silvio Franz and Giorgio Parisi. Phase diagram of coupled glassy systems: A mean-field study. Phys. Rev. Lett.,
79(13):2486–2489, Sep 1997.
[97] Silvio Franz and Giorgio Parisi. Quasi-equilibrium in glassy dynamics: an algebraic view. Journal of Statistical Mechanics:
Theory and Experiment, 2013(02):P02003, 2013.
[98] Joel Friedman. A proof of Alon’s second eigenvalue conjecture and related problems. American Mathematical Soc., 2008.
[99] R. G. Gallager. Information theory and reliable communication. John Wiley and Sons, New York, 1968.
[100] Robert G. Gallager. Low-density parity check codes. IEEE Trans. Inform. Theory, 8:21–28, 1962.
[101] Surya Ganguli and Haim Sompolinsky. Statistical mechanics of compressed sensing. Physical Review Letters,
104(18):188701, 2010.
[102] E Gardner and B Derrida. Optimal storage properties of neural network models. Journal of Physics A: Mathematical
and General, 21(1):271, 1988.
[103] E Gardner and B Derrida. Three unfinished works on the optimal storage capacity of networks. Journal of Physics A:
Mathematical and General, 22(12):1983, 1989.
[104] Alan E Gelfand and Adrian FM Smith. Sampling-based approaches to calculating marginal densities. Journal of the
American statistical association, 85(410):398–409, 1990.
[105] Stuart Geman and Donald Geman. Stochastic relaxation, gibbs distributions, and the bayesian restoration of images.
IEEE Transactions on pattern analysis and machine intelligence, (6):721–741, 1984.
[106] A Georges, D Hansel, P Le Doussal, and J-P Bouchaud. Exact properties of spin glasses. ii. nishimori’s line: new results
and physical implications. Journal de Physique, 46(11):1827–1836, 1985.
[107] Antoine Georges and Jonathan S Yedidia. How to expand around mean-field theory using high-temperature expansions.
Journal of Physics A: Mathematical and General, 24(9):2173, 1991.
[108] Michelle Girvan and Mark EJ Newman. Community structure in social and biological networks. Proceedings of the
National Academy of Sciences, 99(12):7821–7826, 2002.
[109] Paul M Goldbart, Nigel Goldenfeld, and David Sherrington. Stealing the Gold-A Celebration of the Pioneering Physics
of Sam Edwards. Oxford University Press, 2005.
[110] Emmanuelle Gouillart, Florent Krzakala, Marc Mézard, and Lenka Zdeborová. Belief-propagation reconstruction for
discrete tomography. Inverse Problems, 29(3):035003, 2013.
[111] Peter Grassberger and Jean-Pierre Nadal. From statistical physics to statistical inference and back. Kluwer Academic,
1994.
[112] D. J. Gross, I. Kanter, and H. Sompolinsky. Mean-field theory of the potts glass. Phys. Rev. Lett., 55(3):304–307, Jul
1985.
[113] DJ Gross, I Kanter, and Haim Sompolinsky. Mean-field theory of the potts glass. Physical Review Letters, 55(3):304,
1985.
[114] D.J. Gross and M. Mézard. The simplest spin glass. Nucl. Phys. B, 240:431, 1984.
80
[115] Francesco Guerra and Fabio Lucio Toninelli. The thermodynamic limit in mean field spin glass models. Communications
in Mathematical Physics, 230(1):71–79, 2002.
[116] Dongning Guo and Sergio Verdú. Randomly spread cdma: Asymptotics via statistical physics. Information Theory, IEEE
Transactions on, 51(6):1983–2010, 2005.
[117] Géza Györgyi. First-order transition to perfect generalization in a neural network with binary synapses. Physical Review
A, 41(12):7097, 1990.
[118] Bruce Hajek, Yihong Wu, and Jiaming Xu. Computational lower bounds for community detection on random graphs.
arXiv preprint arXiv:1406.6625, 2014.
[119] Kathleen E Hamilton and Leonid P Pryadko. Tight lower bound for percolation threshold on an infinite graph. Physical
Review Letters, 113(20):208701, 2014.
[120] Ki-ichiro Hashimoto. Zeta functions of finite graphs and representations of p-adic groups. Automorphic forms and
geometry of arithmetic varieties., pages 211–280, 1989.
[121] S Hamed Hassani, Nicolas Macris, and Ruediger Urbanke. Chains of mean-field models. J. Stat. Mech.: Theor. and Exp.,
page P02011, 2012.
[122] Trevor Hastie, Robert Tibshirani, Jerome Friedman, and James Franklin. The elements of statistical learning: data
mining, inference and prediction. The Mathematical Intelligencer, 27(2):83–85, 2005.
[123] Trevor Hastie, Robert Tibshirani, and Martin Wainwright. Statistical learning with sparsity: the lasso and generalizations.
CRC Press, 2015.
[124] Matthew B Hastings. Community detection as an inference problem. Physical Review E, 74(3):035102, 2006.
[125] W Keith Hastings. Monte carlo sampling methods using markov chains and their applications. Biometrika, 57(1):97–109,
1970.
[126] Simon Heimlicher, Marc Lelarge, and Laurent Massoulié. Community detection in the labelled stochastic block model.
arXiv preprint arXiv:1209.2910, 2012.
[127] Itay Hen and AP Young. Exponential complexity of the quantum adiabatic algorithm for certain satisfiability problems.
Physical Review E, 84(6):061152, 2011.
[128] Geoffrey E. Hinton. Training products of experts by minimizing contrastive divergence. Neural Computation, 14(8):1771–
1800, August 2002.
[129] Geoffrey E. Hinton. A practical guide to training restricted boltzmann machines. Technical Report UTML TR 2010-003,
University of Toronto, Toronto, Canada, 2010.
[130] Geoffrey E Hinton, Simon Osindero, and Yee-Whye Teh. A fast learning algorithm for deep belief nets. Neural computation,
18(7):1527–1554, 2006.
[131] Geoffrey E Hinton and Ruslan R Salakhutdinov. Reducing the dimensionality of data with neural networks. Science,
313(5786):504–507, 2006.
[132] Glen M Hocky, Ludovic Berthier, and David R Reichman. Equilibrium ultrastable glasses produced by random pinning.
The Journal of chemical physics, 141(22):224503, 2014.
[133] Jake M Hofman and Chris H Wiggins. Bayesian approach to network modularity. Physical Review Letters, 100(25):258701,
2008.
[134] Paul W Holland, Kathryn Blackmond Laskey, and Samuel Leinhardt. Stochastic blockmodels: First steps. Social networks,
5(2):109–137, 1983.
[135] J. J. Hopfield. Neural networks and physical systems with emergent collective computational properties. Proc. Nat. Acad.
Sci. USA, 79:2554–2558, 1982.
[136] John J Hopfield. Neural networks and physical systems with emergent collective computational abilities. Proceedings of
the national academy of sciences, 79(8):2554–2558, 1982.
[137] Yukito Iba. The nishimori line and bayesian statistics. Journal of Physics A: Mathematical and General, 32(21):3875,
1999.
[138] Gregg Jaeger. The ehrenfest classification of phase transitions: Introduction and evolution. Archive for History of Exact
Sciences, 53(1):51–81, 1998.
[139] S. Janson, T. Luczak, and A. Rucinski. Random graphs. Wiley, New-York, 2000.
[140] Adel Javanmard and Andrea Montanari. State evolution for general approximate message passing algorithms, with
applications to spatial coupling. Information and Inference, page iat004, 2013.
[141] Edwin T Jaynes. Information theory and statistical mechanics. Physical Review, 106(4):620, 1957.
[142] Mark Jerrum. Large cliques elude the metropolis process. Random Structures & Algorithms, 3(4):347–359, 1992.
[143] Mark Jerrum and Gregory B Sorkin. Simulated annealing for graph bisection. In Foundations of Computer Science, 1993.
Proceedings., 34th Annual Symposium on, pages 94–103. IEEE, 1993.
[144] Haixia Jia, Cristopher Moore, and Doug Strain. Generating hard satisfiable formulas by hiding solutions deceptively. In
Proceedings of the National Conference on Artificial Intelligence, volume 20, page 384. Menlo Park, CA; Cambridge, MA;
London; AAAI Press; MIT Press; 1999, 2005.
[145] Iain M Johnstone and Arthur Yu Lu. Sparse principal components analysis. 2004.
[146] Michael I Jordan, Zoubin Ghahramani, Tommi S Jaakkola, and Lawrence K Saul. An introduction to variational methods
for graphical models. Machine learning, 37(2):183–233, 1999.
[147] Antony Joseph and Andrew R Barron. Least squares superposition codes of moderate dictionary size are reliable at rates
up to capacity. Information Theory, IEEE Transactions on, 58(5):2541–2557, 2012.
[148] Ari Juels and Marcus Peinado. Hiding cliques for cryptographic security. Designs, Codes and Cryptography, 20(3):269–280,
2000.
81
[149] Tzyy-Ping Jung, Scott Makeig, Colin Humphries, Te-Won Lee, Martin J Mckeown, Vicente Iragui, and Terrence J
Sejnowski. Removing electroencephalographic artifacts by blind source separation. Psychophysiology, 37(02):163–178,
2000.
[150] Yoshiyuki Kabashima. A CDMA multiuser detection algorithm on the basis of belief propagation. Journal of Physics A:
Mathematical and General, 36(43):11111, 2003.
[151] Yoshiyuki Kabashima. Propagating beliefs in spin-glass models. Journal of the Physical Society of Japan, 72(7):1645–1649,
2003.
[152] Yoshiyuki Kabashima. Inference from correlated patterns: a unified theory for perceptron learning and linear vector
channels. In Journal of Physics: Conference Series, volume 95, page 012001. IOP Publishing, 2008.
[153] Yoshiyuki Kabashima, Florent Krzakala, Marc Mézard, Ayaka Sakata, and Lenka Zdeborová. Phase transitions and sample
complexity in bayes-optimal matrix factorization. to appear in IEEE Trans. Inf. Theory, preprint arxiv:1402.1298, 2016.
[154] Yoshiyuki Kabashima and David Saad. Belief propagation vs. tap for decoding corrupted messages. EPL (Europhysics
Letters), 44(5):668, 1998.
[155] Yoshiyuki Kabashima and Shinsuke Uda. A BP-based algorithm for performing Bayesian inference in large perceptrontype networks. In Algorithmic Learning Theory, pages 479–493. Springer, 2004.
[156] Yoshiyuki Kabashima and Mikko Vehkapera. Signal recovery using expectation consistent approximation for linear
observations. In Information Theory (ISIT), 2014 IEEE International Symposium on, pages 226–230. IEEE, 2014.
[157] Yoshiyuki Kabashima, Tadashi Wadayama, and Toshiyuki Tanaka. A typical reconstruction limit for compressed sensing
based on lp-norm minimization. Journal of Statistical Mechanics: Theory and Experiment, 2009(09):L09003, 2009.
[158] Ulugbek Kamilov, Sundeep Rangan, Michael Unser, and Alyson K Fletcher. Approximate message passing with consistent
parameter estimation and applications to sparse learning. In Advances in Neural Information Processing Systems, pages
2438–2446, 2012.
[159] David R Karger, Sewoong Oh, and Devavrat Shah. Iterative learning for reliable crowdsourcing systems. In Advances in
neural information processing systems, pages 1953–1961, 2011.
[160] Brian Karrer and Mark EJ Newman. Stochastic blockmodels and community structure in networks. Physical Review E,
83(1):016107, 2011.
[161] Brian Karrer, MEJ Newman, and Lenka Zdeborová. Percolation on sparse networks. Physical Review Letters,
113(20):208702, 2014.
[162] Raghunandan H Keshavan, Andrea Montanari, and Sewoong Oh. Matrix completion from a few entries. Information
Theory, IEEE Transactions on, 56(6):2980–2998, 2010.
[163] R. Kikuchi. A theory of cooperative phenomena. Phys. Rev., 81:988–1003, 1951.
[164] S. Kirkpatrick, C. D. Gelatt Jr., and M. P. Vecchi. Optimization by simulated annealing. Science, 220:671–680, 1983.
[165] T.R. Kirkpatrick and D. Thirumalai. Dynamics of the structural glass transition and the p-spin-interaction spin-glass
model. Phys. Rev. Lett., 58:2091, 1987.
[166] T.R. Kirkpatrick and D. Thirumalai. p-spin-interaction spin-glass models: Connections with the structural glass problem.
Phys. Rev. B, 36:5388, 1987.
[167] T.R. Kirkpatrick, D. Thirumalai, and P.G. Wolynes. Scaling concepts for the dynamics of viscous liquids near an ideal
glassy state. Phys. Rev. A, 40:1045, 1989.
[168] TR Kirkpatrick and PG Wolynes. Connections between some kinetic and equilibrium theories of the glass transition.
Physical Review A, 35(7):3072, 1987.
[169] TR Kirkpatrick and PG Wolynes. Stable and metastable states in mean-field potts and structural glasses. Physical Review
B, 36(16):8552, 1987.
[170] W. Krauth and M. Mézard. Storage capacity of memory networks with binary couplings. J. Phys., 50:3057–3066, 1989.
[171] Valdis Krebs. Books about us politics. unpublished, http://www. orgnet. com, 2004.
[172] F. Krzakala and J. Kurchan. A landscape analysis of constraint satisfaction problems. Phys. Rev. E, 76:021122, 2007.
[173] F Krzakala, M Mézard, F Sausset, YF Sun, and L Zdeborová. Statistical-physics-based reconstruction in compressed
sensing. Physical Review X, 2(2):021005, 2012.
[174] Florent Krzakala, Andre Manoel, Eric W Tramel, and Lenka Zdeborová. Variational free energies for compressed sensing.
In Information Theory (ISIT), 2014 IEEE International Symposium on, pages 1499–1503. IEEE, 2014.
[175] Florent Krzakala, Marc Mézard, Francois Sausset, Yifan Sun, and Lenka Zdeborová. Probabilistic reconstruction in
compressed sensing: algorithms, phase diagrams, and threshold achieving matrices. Journal of Statistical Mechanics:
Theory and Experiment, 2012(08):P08009, 2012.
[176] Florent Krzakala, Marc Mézard, and Lenka Zdeborová. Compressed sensing under matrix uncertainty: Optimum thresholds and robust approximate message passing. In Acoustics, Speech and Signal Processing (ICASSP), 2013 IEEE International Conference on, pages 5519–5523. IEEE, 2013.
[177] Florent Krzakala, Marc Mézard, and Lenka Zdeborová. Phase diagram and approximate message passing for blind
calibration and dictionary learning. In Information Theory Proceedings (ISIT), 2013 IEEE International Symposium on,
pages 659–663. IEEE, 2013.
[178] Florent Krzakala, Marc Mézard, and Lenka Zdeborová. Reweighted belief propagation and quiet planting for random
k-sat. Journal on Satisfiability, Boolean Modeling and Computation, 8:149, 2014.
[179] Florent Krzakala, Andrea Montanari, Federico Ricci-Tersenghi, Guilhem Semerjian, and Lenka Zdeborová. Gibbs states
and the set of solutions of random constraint satisfaction problems. Proc. Natl. Acad. Sci. U.S.A, 104:10318, 2007.
[180] Florent Krzakala, Cristopher Moore, Elchanan Mossel, Joe Neeman, Allan Sly, Lenka Zdeborová, and Pan Zhang. Spectral
redemption in clustering sparse networks. Proceedings of the National Academy of Sciences, 110(52):20935–20940, 2013.
82
[181] Florent Krzakala, Federico Ricci-Tersenghi, Lenka Zdeborová, Riccardo Zecchina, Eric W. Tramel, and Leticia F. Cugliandolo. Statistical Physics, Optimization, Inference and Message-Passing Algorithms. Oxford University Press, 2015.
[182] Florent Krzakala and Lenka Zdeborová. Hiding quiet solutions in random constraint satisfaction problems. Physical
Review Letters, 102(23):238701, 2009.
[183] Florent Krzakala and Lenka Zdeborová. Following gibbs states adiabatically - the energy landscape of mean-field glassy
systems. EPL (Europhysics Letters), 90(6):66002, 2010.
[184] Florent Krzakala and Lenka Zdeborová. On melting dynamics and the glass transition. i. glassy aspects of melting
dynamics. The Journal of chemical physics, 134(3):034512, 2011.
[185] Florent Krzakala and Lenka Zdeborová. On melting dynamics and the glass transition. ii. glassy dynamics as a melting
process. The Journal of chemical physics, 134(3):034513, 2011.
[186] Florent Krzakala and Lenka Zdeborová. Performance of simulated annealing in p-spin glasses. In Journal of Physics:
Conference Series, volume 473, page 012022. IOP Publishing, 2013.
[187] S. Kudekar and H.D. Pfister. The effect of spatial coupling on compressive sensing. In Communication, Control, and
Computing (Allerton), pages 347–353, 2010.
[188] Shrinivas Kudekar, Thomas J Richardson, and Rüdiger L Urbanke. Threshold saturation via spatial coupling: Why
convolutional ldpc ensembles perform so well over the bec. IEEE Transactions on Information Theory, 57(2):803–834,
2011.
[189] C. Kwon and D. J. Thouless. Ising spin glass at zero temperature on the bethe lattice. Phys. Rev. B, 37(13):7649–7654,
1988.
[190] Laurent Laloux, Pierre Cizeau, Jean-Philippe Bouchaud, and Marc Potters. Noise dressing of financial correlation matrices. Physical Review Letters, 83(7):1467, 1999.
[191] M. Lelarge, L. Massoulié, and Jiaming Xu. Reconstruction in the labeled stochastic block model. In Information Theory
Workshop (ITW), 2013 IEEE, pages 1–5, Sept 2013.
[192] Michael Lentmaier, Arvind Sridharan, Daniel J Costello, and K Sh Zigangirov. Iterative decoding threshold analysis for
ldpc convolutional codes. IEEE Trans. Inf. Theory, 56(10):5274–5289, 2010.
[193] Thibault Lesieur, Florent Krzakala, and Lenka Zdeborová. Mmse of probabilistic low-rank matrix estimation: Universality
with respect to the output channel. In 2015 53rd Annual Allerton Conference on Communication, Control, and Computing
(Allerton), pages 680–687. IEEE, 2015.
[194] Thibault Lesieur, Florent Krzakala, and Lenka Zdeborová. Phase transitions in sparse pca. In Information Theory (ISIT),
2015 IEEE International Symposium on, pages 1635–1639. IEEE, 2015.
[195] Ting Liu, Chao-Kai Wen, Shi Jin, and Xiaohu You. Generalized turbo signal recovery for nonlinear measurements and
orthogonal sensing matrices. arXiv preprint arXiv:1512.04833, 2015.
[196] David Lusseau, Karsten Schneider, Oliver J Boisseau, Patti Haase, Elisabeth Slooten, and Steve M Dawson. The bottlenose dolphin community of doubtful sound features a large proportion of long-lasting associations. Behavioral Ecology
and Sociobiology, 54(4):396–405, 2003.
[197] Zongming Ma, Yihong Wu, et al. Computational barriers in minimax submatrix detection. The Annals of Statistics,
43(3):1089–1116, 2015.
[198] D. J. C. MacKay. Information theory, inference, and learning algorithms. Cambridge University Press, Cambridge, 2003.
[199] Julien Mairal, Francis Bach, Jean Ponce, and Guillermo Sapiro. Online dictionary learning for sparse coding. In Proceedings of the 26th annual international conference on machine learning, pages 689–696. ACM, 2009.
[200] Stéphane G Mallat and Zhifeng Zhang. Matching pursuits with time-frequency dictionaries. Signal Processing, IEEE
Transactions on, 41(12):3397–3415, 1993.
[201] Andre Manoel, Florent Krzakala, Eric Tramel, and Lenka Zdeborovà. Swept approximate message passing for sparse
estimation. In Proceedings of the 32nd International Conference on Machine Learning (ICML-15), pages 1123–1132,
2015.
[202] Romain Mari and Jorge Kurchan. Dynamical transition of glasses: From exact to approximate. The Journal of chemical
physics, 135(12):124504, 2011.
[203] Manuel Sebastian Mariani, Giorgio Parisi, and Corrado Rainone. Calorimetric glass transition in a mean-field theory
approach. Proceedings of the National Academy of Sciences, 112(8):2361–2366, 2015.
[204] Travis Martin, Xiao Zhang, and MEJ Newman. Localization and centrality in networks. Physical Review E, 90(5):052808,
2014.
[205] Laurent Massoulié. Community detection thresholds and the weak Ramanujan property. In Proceedings of the 46th
Annual ACM Symposium on Theory of Computing, pages 694–703. ACM, 2014.
[206] Ryosuke Matsushita and Toshiyuki Tanaka. Low-rank matrix reconstruction and clustering via approximate message
passing. In Advances in Neural Information Processing Systems, pages 917–925, 2013.
[207] DC Mattis. Solvable spin systems with random interactions. Physics Letters A, 56(5):421–422, 1976.
[208] Sharon Bertsch McGrayne. The Theory That Would Not Die: How Bayes Rule Cracked the Enigma Code, Hunted Down
Russian Submarines, and Emerged Triumphant from Two Centuries of Controversy. Yale University Press, 2011.
[209] N. Metropolis, A. W. Rosenbluth, M. N. Rosenbluth, A. H. Teller, and E. Teller. Equation of State Calculations by Fast
Computing Machines. The Journal of Chemical Physics, 21:1087–1092, June 1953.
[210] M Mézard. The space of interactions in neural networks: Gardner’s computation with the cavity method. Journal of
Physics A: Mathematical and General, 22(12):2181, 1989.
[211] M. Mézard and G. Parisi. The bethe lattice spin glass revisited. Eur. Phys. J. B, 20:217, 2001.
[212] M. Mézard and G. Parisi. The cavity method at zero temperature. J. Stat. Phys., 111:1–34, 2003.
83
[213] M. Mézard, G. Parisi, and M. A. Virasoro. SK model: The replica solution without replicas. Europhys. Lett., 1:77–82,
1986.
[214] M. Mézard, G. Parisi, and M. A. Virasoro. Spin-Glass Theory and Beyond, volume 9 of Lecture Notes in Physics. World
Scientific, Singapore, 1987.
[215] M. Mézard, G. Parisi, and R. Zecchina. Analytic and algorithmic solution of random satisfiability problems. Science,
297:812–815, 2002.
[216] M. Mézard, F. Ricci-Tersenghi, and R. Zecchina. Alternative solutions to diluted p-spin models and XORSAT problems.
J. Stat. Phys., 111:505, 2003.
[217] Marc Mézard and Andrea Montanari. Information, physics, and computation. Oxford University Press, 2009.
[218] R. Monasson and R. Zecchina. Entropy of the K-satisfiability problem. Phys. Rev. Lett., 76:3881–3885, 1996.
[219] Rémi Monasson, Riccardo Zecchina, Scott Kirkpatrick, Bart Selman, and Lidror Troyansky. Determining computational
complexity from characteristic “phase transitions”. Nature, 400(6740):133–137, 1999.
[220] A. Montanari. Estimating random variables from random sparse observations. European Transactions on Telecommunications, 19(4):385–403, 2008.
[221] A. Montanari and G. Semerjian. On the dynamics of the glass transition on bethe lattices. J. Stat. Phys., 124:103–189,
2006.
[222] A. Montanari and G. Semerjian. Rigorous inequalities between length and time scales in glassy systems. J. Stat. Phys.,
125:23, 2006.
[223] Andrea Montanari, Daniel Reichman, and Ofer Zeitouni. On the limitation of spectral methods: From the gaussian hidden
clique problem to rank-one perturbations of gaussian tensors. In Advances in Neural Information Processing Systems,
pages 217–225, 2015.
[224] Andrea Montanari and Devavrat Shah. Counting good truth assignments of random k-SAT formulae. In Proceedings of
the eighteenth annual ACM-SIAM symposium on Discrete algorithms, pages 1255–1264. Society for Industrial and Applied
Mathematics, 2007.
[225] Andrea Montanari and Rudiger Urbanke. Modern coding theory: The statistical mechanics and computer science point
of view. Complex Systems, Les Houches lecture notes, page 69, 2007.
[226] Joris M Mooij, Hilbert J Kappen, et al. Validity estimates for loopy belief propagation on binary real-world networks. In
NIPS, 2004.
[227] Cristopher Moore and Stephan Mertens. The nature of computation. Oxford University Press, 2011.
[228] T. Mora. Géométrie et inférence dans l’optimisation et en théorie de l’information. PhD thesis, Université Paris-Sud,
2007. http://tel.archives-ouvertes.fr/tel-00175221/en/.
[229] Flaviano Morone and Hernán A Makse. Influence maximization in complex networks through optimal percolation. Nature,
524:6568, 2015.
[230] Allard P Mosk, Ad Lagendijk, Geoffroy Lerosey, and Mathias Fink. Controlling waves in space and time for imaging and
focusing in complex media. Nature photonics, 6(5):283–292, 2012.
[231] Elchanan Mossel, Joe Neeman, and Allan Sly. Stochastic block models and reconstruction. arXiv preprint arXiv:1202.1499,
2012.
[232] Elchanan Mossel, Joe Neeman, and Allan Sly. A proof of the block model threshold conjecture. arXiv preprint
arXiv:1311.4115, 2013.
[233] Elchanan Mossel, Joe Neeman, and Allan Sly. Belief propagation, robust reconstruction and optimal recovery of block
models. In Proceedings of The 27th Conference on Learning Theory, pages 356–370, 2014.
[234] R. Mulet, A. Pagnani, M. Weigt, and R. Zecchina. Coloring random graphs. Phys. Rev. Lett., 89:268701, 2002.
[235] Radford M Neal. Probabilistic inference using markov chain monte carlo methods. Technical report, Department of
Computer Science, University of Toronto Toronto, Ontario, Canada, 1993.
[236] Mark EJ Newman. Finding community structure in networks using the eigenvectors of matrices. Physical Review E,
74(3):036104, 2006.
[237] Mark EJ Newman. Modularity and community structure in networks. Proceedings of the National Academy of Sciences,
103(23):8577–8582, 2006.
[238] H. Nishimori. Statistical Physics of Spin Glasses and Information Processing: An Introduction. Oxford University Press,
Oxford, UK, 2001.
[239] Hidetoshi Nishimori. Exact results and critical properties of the ising model with competing interactions. Journal of
Physics C: Solid State Physics, 13(21):4071, 1980.
[240] Hidetoshi Nishimori. Internal energy, specific heat and correlation function of the bond-random ising model. Progress of
Theoretical Physics, 66(4):1169–1181, 1981.
[241] Hidetoshi Nishimori and KY Michael Wong. Statistical mechanics of image restoration and error-correcting codes. Physical
Review E, 60(1):132, 1999.
[242] Lars Onsager. Crystal statistics. i. a two-dimensional model with an order-disorder transition. Physical Review, 65(34):117, 1944.
[243] Manfred Opper, Burak Cakmak, and Ole Winther. A theory of solving tap equations for ising models with general
invariant random matrices. Journal of Physics A: Mathematical and Theoretical, 49(11):114002, 2016.
[244] Manfred Opper and David Saad. Advanced mean field methods: Theory and practice. MIT press, 2001.
[245] Manfred Opper and Ole Winther. Tractable approximations for probabilistic models: The adaptive thouless-andersonpalmer mean field approach. Physical Review Letters, 86(17):3695, 2001.
84
[246] A. Pagnani, G. Parisi, and M. Ratiéville. Near optimal configurations in mean field disordered systems. Phys. Rev. E,
68:046706, 2003.
[247] Giorgio Parisi and Marc Potters. Mean-field equations for spin models with orthogonal interaction matrices. Journal of
Physics A: Mathematical and General, 28(18):5267, 1995.
[248] Jason T Parker, Philip Schniter, and Volkan Cevher. Bilinear generalized approximate message passingpart i: Derivation.
Signal Processing, IEEE Transactions on, 62(22):5839–5853, 2014.
[249] J. Pearl. Reverend bayes on inference engines: A distributed hierarchical approach. In Proceedings American Association
of Artificial Intelligence National Conference on AI, pages 133–136, Pittsburgh, PA, USA, 1982.
[250] R Peierls. Statistical theory of superlattices with unequal concentrations of the components. In Proceedings of the Royal
Society of London A: Mathematical, Physical and Engineering Sciences, volume 154, pages 207–222. The Royal Society,
1936.
[251] Carsten Peterson and James R. Anderson. A mean field theory learning algorithm for neural networks. Complex Systems,
1:995 – 1019, 1987.
[252] Henry D Pfister and Pascal O Vontobel. On the relevance of graph covers and zeta functions for the analysis of SPA
decoding of cycle codes. In Information Theory Proceedings (ISIT), 2013 IEEE International Symposium on, pages
3000–3004. IEEE, 2013.
[253] T Plefka. Convergence condition of the tap equation for the infinite-ranged ising spin glass model. Journal of Physics A:
Mathematical and general, 15(6):1971, 1982.
[254] SM Popoff, G Lerosey, R Carminati, M Fink, AC Boccara, and S Gigan. Measuring the transmission matrix in optics:
an approach to the study and control of light propagation in disordered media. Physical Review Letters, 104(10):100601,
2010.
[255] Filippo Radicchi. Predicting percolation thresholds in networks. Physical Review E, 91(1):010801, 2015.
[256] Sundeep Rangan. Estimation with random linear mixing, belief propagation and compressed sensing. In Information
Sciences and Systems (CISS), 2010 44th Annual Conference on, pages 1–6. IEEE, 2010.
[257] Sundeep Rangan. Generalized approximate message passing for estimation with random linear mixing. In Information
Theory Proceedings (ISIT), 2011 IEEE International Symposium on, pages 2168–2172. IEEE, 2011.
[258] Sundeep Rangan and Alyson K Fletcher. Iterative estimation of constrained rank-one matrices in noise. In Information
Theory Proceedings (ISIT), 2012 IEEE International Symposium on, pages 1246–1250. IEEE, 2012.
[259] Sundeep Rangan, Alyson K Fletcher, Vivek K Goyal, and Philip Schniter. Hybrid approximate message passing with
applications to structured sparsity. arXiv preprint arXiv:1111.2581, 2011.
[260] Sundeep Rangan, Alyson K Fletcher, Philip Schniter, and Ulugbek S Kamilov. Inference for generalized linear models
via alternating directions and bethe free energy minimization. In Information Theory (ISIT), 2015 IEEE International
Symposium on, pages 1640–1644. IEEE, 2015.
[261] Sundeep Rangan, Vivek Goyal, and Alyson K Fletcher. Asymptotic analysis of map estimation via the replica method
and compressed sensing. In Advances in Neural Information Processing Systems 22, pages 1545–1553. 2009.
[262] Sundeep Rangan, Philip Schniter, and Alyson Fletcher. On the convergence of approximate message passing with arbitrary
matrices. In Information Theory (ISIT), 2014 IEEE International Symposium on, pages 236–240. IEEE, 2014.
[263] Jörg Reichardt and Stefan Bornholdt. Statistical mechanics of community detection. Physical Review E, 74(1):016110,
2006.
[264] F. Ricci-Tersenghi. The bethe approximation for solving the inverse ising problem: a comparison with other inference
methods. J. Stat. Mech.: Th. and Exp., page P08015, 2012.
[265] F. Ricci-Tersenghi, M. Weigt, and R. Zecchina. Simplest random k-satifiablitity problem. Phys. Rev. E, 63:026702, 2001.
[266] Emile Richard and Andrea Montanari. A statistical model for tensor PCA. In Advances in Neural Information Processing
Systems, pages 2897–2905, 2014.
[267] Tim Rogers. Assessing node risk and vulnerability in epidemics on networks. EPL (Europhysics Letters), 109(2):28005,
2015.
[268] Tim Rogers and Isaac Pérez Castillo. Cavity approach to the spectral density of non-Hermitian sparse matrices. Physical
Review E, 79(1):012101, 2009.
[269] Frank Rosenblatt. The perceptron: a probabilistic model for information storage and organization in the brain. Psychological Review, 65(6):386, 1958.
[270] Cynthia Rush, Adam Greig, and Ramji Venkataramanan. Capacity-achieving sparse regression codes via approximate
message passing decoding. In Information Theory (ISIT), 2015 IEEE International Symposium on, pages 2016–2020.
IEEE, 2015.
[271] Alaa Saade, Florent Krzakala, and Lenka Zdeborová. Spectral clustering of graphs with the bethe hessian. In Advances
in Neural Information Processing Systems, pages 406–414, 2014.
[272] Alaa Saade, Florent Krzakala, and Lenka Zdeborová. Spectral density of the non-backtracking operator on random
graphs. EPL (Europhysics Letters), 107(5):50005, 2014.
[273] Alaa Saade, Florent Krzakala, and Lenka Zdeborová. Matrix completion from fewer entries: Spectral detectability and
rank estimation. In Advances in Neural Information Processing Systems, pages 1261–1269, 2015.
[274] Alaa Saade, Marc Lelarge, Florent Krzakala, and Lenka Zdeborová. Spectral detection in the censored block model. In
2015 IEEE International Symposium on Information Theory (ISIT), pages 1184–1188. IEEE, 2015.
[275] Ayaka Sakata and Yoshiyuki Kabashima. Sample complexity of bayesian optimal dictionary learning. In Information
Theory Proceedings (ISIT), 2013 IEEE International Symposium on, pages 669–673. IEEE, 2013.
85
[276] Ayaka Sakata and Yoshiyuki Kabashima. Statistical mechanics of dictionary learning. EPL (Europhysics Letters),
103(2):28008, 2013.
[277] Christian Schlegel and Dmitri Truhachev. Multiple access demodulation in the lifted signal graph with spatial coupling.
In IEEE International Symposium on Information Theory Proceedings (ISIT), pages 2989–2993. IEEE, 2011.
[278] Philip Schniter. Turbo reconstruction of structured sparse signals. In Information Sciences and Systems (CISS), 2010
44th Annual Conference on, pages 1–6. IEEE, 2010.
[279] Philip Schniter and Sundeep Rangan. Compressive phase retrieval via generalized approximate message passing. Signal
Processing, IEEE Transactions on, 63(4):1043–1055, 2015.
[280] Bart Selman, David G. Mitchell, and Hector J. Levesque. Generating hard satisfiability problems. Artif. Intell., 81(12):17–29, 1996.
[281] HS Seung, Haim Sompolinsky, and N Tishby. Statistical mechanics of learning from examples. Physical Review A,
45(8):6056, 1992.
[282] Maoz Shamir and Haim Sompolinsky. Thouless-anderson-palmer equations for neural networks. Physical Review E,
61(2):1839, 2000.
[283] C. E. Shannon. A mathematical theory of communication. Bell System Tech. Journal, 27:379–423, 623–655, 1948.
[284] Ori Shental, Paul H Siegel, Jack K Wolf, Danny Bickson, and Danny Dolev. Gaussian belief propagation solver for systems
of linear equations. In Information Theory, 2008. ISIT 2008. IEEE International Symposium on, pages 1863–1867. IEEE,
2008.
[285] D. Sherrington and S. Kirkpatrick. Solvable model of a spin-glass. Phys. Rev. Lett., 35:1792–1796, 1975.
[286] Takashi Shinzato and Yoshiyuki Kabashima. Perceptron capacity revisited: classification ability for correlated patterns.
Journal of Physics A: Mathematical and Theoretical, 41(32):324013, 2008.
[287] Sasha Sodin. Random matrices, nonbacktracking walks, and orthogonal polynomials. Journal of Mathematical Physics,
48(12):123503, 2007.
[288] Subhojit Som and Philip Schniter. Compressive imaging using approximate message passing and a markov-tree prior.
Signal Processing, IEEE Transactions on, 60(7):3439–3448, 2012.
[289] H. Sompolinsky and Annette Zippelius. Relaxational dynamics of the edwards-anderson model and the mean-field theory
of spin-glasses. Phys. Rev. B, 25:6860–6875, Jun 1982.
[290] N. Sourlas. Spin-glass models as error-correcting codes. Nature, 339:693–694, 1989.
[291] N. Sourlas. Spin-glasses, error-correcting codes and finite-temperature decoding. Europhys. Lett., 25:159–164, 1994.
[292] N. Sourlas. Statistical mechanics and error-correcting codes. In P. Grassberger and J.-P. Nadal, editors, From Statistical
Physics to Statistical Inference and Back, pages 195–204. Kluwer, 1994.
[293] Erik B Sudderth, Alexander T Ihler, Michael Isard, William T Freeman, and Alan S Willsky. Nonparametric belief
propagation. Communications of the ACM, 53(10):95–103, 2010.
[294] YiFan Sun, Andrea Crisanti, Florent Krzakala, Luca Leuzzi, and Lenka Zdeborová. Following states in temperature in
the spherical s+ p-spin glass model. Journal of Statistical Mechanics: Theory and Experiment, 2012(07):P07002, 2012.
[295] Koujin Takeda and Yoshiyuki Kabashima. Statistical mechanical analysis of compressed sensing utilizing correlated
compression matrix. In Information Theory Proceedings (ISIT), 2010 IEEE International Symposium on, pages 1538–
1542. IEEE, 2010.
[296] Koujin Takeda, Shinsuke Uda, and Yoshiyuki Kabashima. Analysis of cdma systems that are characterized by eigenvalue
spectrum. EPL (Europhysics Letters), 76(6):1193, 2006.
[297] Keigo Takeuchi, Toshiyuki Tanaka, and Tsutomu Kawabata. Improvement of BP-based CDMA multiuser detection by
spatial coupling. In IEEE International Symposium on Information Theory Proceedings (ISIT), pages 1489–1493. IEEE,
2011.
[298] Toshiyuki Tanaka. A statistical-mechanics approach to large-system analysis of cdma multiuser detectors. Information
Theory, IEEE Transactions on, 48(11):2888–2910, 2002.
[299] D. J. Thouless. Spin-glass on a Bethe lattice. Phys. Rev. Lett., 56:1082–1085, 1986.
[300] D. J. Thouless, P. W. Anderson, and R. G. Palmer. Solution of ‘solvable model of a spin-glass’. Phil. Mag., 35:593–601,
1977.
[301] Robert Tibshirani. Regression shrinkage and selection via the lasso. Journal of the Royal Statistical Society. Series B
(Methodological), pages 267–288, 1996.
[302] Eric W Tramel, Angélique Drémeau, and Florent Krzakala. Approximate message passing with restricted boltzmann
machine priors. Journal of Statistical Mechanics: Theory and Experiment, 2016(7):073401, 2016.
[303] Johannes D Van der Waals. The equation of state for gases and liquids. Nobel lectures in Physics, 1:254–265, 1910.
[304] L. Viana and A. J. Bray. Phase diagrams for dilute spin-glasses. J. Phys. C, 18:3037–3051, 1985.
[305] Jeremy Vila, Philip Schniter, and Joseph Meola. Hyperspectral image unmixing via bilinear generalized approximate
message passing. In SPIE Defense, Security, and Sensing, pages 87430Y–87430Y. International Society for Optics and
Photonics, 2013.
[306] Jeremy Vila, Philip Schniter, Sundeep Rangan, Florent Krzakala, and Lenka Zdeborová. Adaptive damping and mean
removal for the generalized approximate message passing algorithm. 2015 IEEE International Conference on Acoustics,
Speech and Signal Processing (ICASSP), pages 2021 – 2025, 2015.
[307] Ulrike Von Luxburg. A tutorial on spectral clustering. Statistics and computing, 17(4):395–416, 2007.
[308] Martin J Wainwright and Michael I Jordan. Graphical models, exponential families, and variational inference. Foundations
and Trends R in Machine Learning, 1(1-2):1–305, 2008.
[309] Larry Wasserman. All of statistics: a concise course in statistical inference. Springer Science & Business Media, 2013.
86
[310] Timothy LH Watkin, Albrecht Rau, and Michael Biehl. The statistical mechanics of learning a rule. Reviews of Modern
Physics, 65(2):499, 1993.
[311] TLH Watkin and J-P Nadal. Optimal unsupervised learning. Journal of Physics A: Mathematical and General, 27(6):1899,
1994.
[312] J. Wehr and M. Aizenman. Fluctuations of extensive functions of quenched random couplings. J. Stat. Phys., 60:287–306,
1990.
[313] P Weiss and G Foëx. Magnetization of ferromagnetic substances above the curie point. Archives des sciences physiques
et naturelles, 31:5–19, 1911.
[314] Harrison C White, Scott A Boorman, and Ronald L Breiger. Social structure from multiple networks. i. blockmodels of
roles and positions. American journal of sociology, pages 730–780, 1976.
[315] Yihong Wu and Sergio Verdú. Optimal phase transitions in compressed sensing. Information Theory, IEEE Transactions
on, 58(10):6241–6263, 2012.
[316] Yingying Xu, Yoshiyuki Kabashima, and Lenka Zdeborová. Bayesian signal reconstruction for 1-bit compressed sensing.
Journal of Statistical Mechanics: Theory and Experiment, 2014(11):P11015, 2014.
[317] Xiaoran Yan, Cosma Shalizi, Jacob E Jensen, Florent Krzakala, Cristopher Moore, Lenka Zdeborová, Pan Zhang, and
Yaojia Zhu. Model selection for degree-corrected block models. Journal of Statistical Mechanics: Theory and Experiment,
2014(5):P05007, 2014.
[318] J.S. Yedidia, W.T. Freeman, and Y. Weiss. Understanding belief propagation and its generalizations. In Exploring
Artificial Intelligence in the New Millennium, pages 239–236. Science & Technology Books, 2003.
[319] Wayne W Zachary. An information flow model for conflict and fission in small groups. Journal of anthropological research,
pages 452–473, 1977.
[320] L. Zdeborová and F. Krzakala. Phase transitions in the coloring of random graphs. Phys. Rev. E, 76:031131, 2007.
[321] Lenka Zdeborová. Statistical physics of hard optimization problems1. Acta Physica Slovaca, 59(3):169–303, 2009.
[322] Lenka Zdeborová and Stefan Boettcher. A conjecture on the maximum cut and bisection width in random regular graphs.
Journal of Statistical Mechanics: Theory and Experiment, 2010(02):P02020, 2010.
[323] Lenka Zdeborová and Florent Krzakala. Generalization of the cavity method for adiabatic evolution of gibbs states.
Physical Review B, 81(22):224205, 2010.
[324] Lenka Zdeborová and Florent Krzakala. Quiet planting in the locked constraint satisfaction problems. SIAM Journal on
Discrete Mathematics, 25(2):750–770, 2011.
[325] Pan Zhang. Nonbacktracking operator for the Ising model and its applications in systems with multiple states. Physical
Review E, 91(4):042120, 2015.
[326] Pan Zhang, Florent Krzakala, Jörg Reichardt, and Lenka Zdeborová. Comparative study for inference of hidden classes
in stochastic block models. Journal of Statistical Mechanics: Theory and Experiment, 2012(12):P12021, 2012.
[327] Pan Zhang and Cristopher Moore. Scalable detection of statistically significant communities and hierarchies, using message
passing for modularity. Proceedings of the National Academy of Sciences, 111(51):18144–18149, 2014.
[328] Pan Zhang, Cristopher Moore, and MEJ Newman. Community detection in networks with unequal groups. Physical
Review E, 93(1):012303, 2016.
[329] Pan Zhang, Cristopher Moore, and Lenka Zdeborová. Phase transitions in semisupervised clustering of sparse networks.
Physical Review E, 90(5):052802, 2014.
[330] Justin Ziniel, Philip Schniter, and Per Sederberg. Binary linear classification and feature selection via generalized approximate message passing. In Information Sciences and Systems (CISS), 2014 48th Annual Conference on, pages 1–6.
IEEE, 2014.
[331] Hui Zou, Trevor Hastie, and Robert Tibshirani. Sparse principal component analysis. Journal of computational and
graphical statistics, 15(2):265–286, 2006.
| 8cs.DS
|
Game Design and Analysis for Price based Demand Response: An
Aggregate Game Approach
arXiv:1508.02636v3 [q-fin.EC] 1 Feb 2016
Maojiao Ye, Student Member, IEEE, and Guoqiang Hu, Member, IEEE
Abstract— In this paper, an aggregate game is adopted for
the modeling and analysis of energy consumption control in
smart grid. Since the electricity users’ cost functions depend
on the aggregate energy consumption, which is unknown to
the end users, an average consensus protocol is employed to
estimate it. By neighboring communication among the users
about their estimations on the aggregate energy consumption,
Nash seeking strategies are developed. Convergence properties
are explored for the proposed Nash seeking strategies. For
energy consumption game that may have multiple isolated Nash
equilibria, a local convergence result is derived. The convergence is established by utilizing singular perturbation analysis
and Lyapunov stability analysis. Energy consumption control
for a network of heating, ventilation, and air conditioning
(HVAC) systems is investigated. Based on the uniqueness of
the Nash equilibrium, it is shown that the players’ actions
converge to a neighborhood of the unique Nash equilibrium
non-locally. More specially, if the unique Nash equilibrium is
an inner Nash equilibrium, an exponential convergence result
is obtained. Energy consumption game with stubborn players
is studied. In this case, the actions of the rational players can
be driven to a neighborhood of their best response strategies by
using the proposed method. Numerical examples are presented
to verify the effectiveness of the proposed methods.
Index Terms— Energy consumption control; aggregate game;
Nash equilibrium seeking
I. INTRODUCTION
Demand response schemes are designed to motivate the
electricity users to adjust their energy consumptions to the
desired profile based on supply conditions [1]. Due to the
benefits, such as improving system reliability, efficiency,
security and customer bill savings, great efforts have been
dedicated to control and optimization problems related to
demand response in smart grid (e.g., see [2]-[5] and the references therein). The electricity users are categorized as pricetakers and price-anticipating users in the existing literatures
[7]. The price-takers schedule their energy consumptions
regardless of their effect on the electricity price. In contrast,
the price-anticipating users consider the impact of their
energy consumptions on the electricity price. Industrial users
and commercial users with large energy consumption are
typical examples of price-anticipating users. Taking the price
as a function of the aggregate energy consumption brings
up coupled energy consumption control problems among the
electricity users.
M. Ye and G. Hu are with the School of Electrical and Electronic
Engineering, Nanyang Technological University, 639798, Singapore (Email:
[email protected], [email protected]).
This work was supported by Singapore Economic Development Board
under EIRP grant S14-1172-NRF EIRP-IHL.
Game theory is an effective modeling and analysis tool to
deal with the interactions among the electricity consumers
[2]. Game theoretical approaches such as stackelberg game,
evolutionary game, differential game, just to name a few,
have been extensively utilized to design and analyze demand
response schemes in smart grid (e.g., [8]-[21]). For example, in [8], Stackelberg game was leveraged to model the
interaction between the demand response aggregators and
generators. The Nash equilibrium of the game among the
generators was computed via solving a centralized quadratic
optimization problem. In [9], the authors considered peakto-average ratio (PAR) minimization and energy cost minimization by designing a non-cooperative energy consumption
game. By communicating among the players on their scheduled daily energy consumptions, a Nash seeking strategy
was proposed. A coupled-constraint game was used in [10]
for scheduling residential energy consumption. Noticing that
the best response algorithm suffers from privacy concern, a
gradient projection method based on estimations of the price
changing trend was proposed at the cost of computation time
and memory. Multiple utility companies were considered in
[11] where the authors used a two-level game approach to
manage the interactions. The competition among the utility
companies was described by a non-cooperative game. The
users were modeled as evolutionary game players. Considering the system dynamics, a differential game approach
was proposed in [12]. In this paper, the energy consumption
control is considered as an aggregate game [22]-[24] on
graph.
Related Work: The main objective of this paper is to
achieve Nash equilibrium seeking (e.g., see [26]-[30] and
the references therein for an incomplete reference list) in
energy consumption game. Different from many Nash seeking strategies where full communication among the players
is used, in [25] and [30], the Nash equilibrium is attained
by utilizing neighboring communication. An aggregate game
(e.g., see [22]-[25]) in which the players interact through
the sum of the players’ actions was investigated in [25].
By utilizing a gossip algorithm, the players can search for
the Nash equilibrium through neighboring communication.
This idea was further generalized in [30]. The players were
considered to be generally interacting with each other. The
game on graph was then solved via using a gossip algorithm.
Similar to [25] and [30], this paper leverages the idea of game
on graph to solve energy consumption game in smart grid.
This paper aims to solve energy consumption control for
a network of price-anticipating electricity users. Compared
with the existing works, the main contributions of the paper
A. Game Theory
are summarized as follows.
•
•
•
An aggregate game is adopted for the modeling and
analysis of energy consumption control in smart grid.
By using game on graph, the users update their actions based on the communication with their neighbors.
This scheme reduces the communication between the
electricity users with the centralized agent (e.g., energy
provider). Hence, the single-node congestion problem
is relieved.
An aggregate game that may admit multiple isolated
Nash equilibria is firstly considered. Based on an average consensus protocol, a Nash seeking strategy is
proposed for the players. An energy consumption game
for a network of heating, ventilation and air conditioning (HVAC) systems, in which the Nash equilibrium
is unique, is investigated. The Nash seeking strategy
is designed based on a primal-dual dynamics. More
specifically, if the unique Nash equilibrium is an inner
Nash equilibrium, an exponential convergence result
is derived. Energy consumption game with stubborn
players is studied. It is shown that with the presence of
stubborn players, the rational players’ actions converge
to a neighborhood of their best response strategies. The
proposed Nash seeking strategies serve as alternative
approaches for the gossip-algorithm in [25] to solve
aggregative game on graph.
The end-users only need to communicate with their
neighbors on their estimations of the aggregate energy
consumption. They don’t need to share their own energy
consumptions with their opponents. Hence, the privacy
of the electricity users is protected.
The rest of the paper is structured as follows. In Section
II, some related preliminaries are provided. System model
and the problem formulation are given in Section III. A
general energy consumption game that may have multiple
Nash equilibria is studied in Section IV without considering
the constraints. Energy consumption game among a network
of HVAC systems is investigated in Section V. In Section VI,
numerical examples are provided to verify the effectiveness
of the proposed methods. Conclusions and future directions
are stated in Section VII.
II. P RELIMINARIES
AND
N OTATIONS
In this paper, R represents for the set of real numbers, R+
stands for the set of non-negative real numbers and R++ is
the set of positive real numbers. Furthermore, diag{ki } for
i ∈ {1, 2, · · · , N } is defined as
k1 0 · · · 0
0 k2
,
diag{ki } = .
.
.
..
..
..
0 ···
kN
and [hi ]vec for i ∈ {1, 2, · · · , N } is defined as [hi ]vec =
[h1 , h2 , · · · , hN ]T .
Below are some definitions on game theory [24]-[34].
Definition 1: A game in normal form is defined as a triple
Γ , {N, X, C} where N = {1, 2, · · · , N } is the set of N
players, X = X1 × · · · × XN , Xi ⊂ R is the set of actions
for player i, and C = (C1 , · · · , CN ) where Ci is the cost
function of player i.
Definition 2: (Potential game) A game Γ is a potential
game if there exists a function P such that
∂P (li , l−i )
∂Ci (li , l−i )
=
,
(1)
∂li
∂li
∀i ∈ N. Furthermore, the function P is the potential
function. In (1), li denotes the action of player i and, l−i
denotes all the players’ actions other than the action of player
i, i.e., l−i = [l1 , · · · , li−1 , li+1 , · · · , lN ]T .
Definition 3: (Aggregate game) An aggregate game is a
normal form game with the players’ cost functions depending
only on its own action and a linear aggregate of the full action
profile.
Definition 4: (Nash equilibrium) Nash equilibrium is an
action profile on which no player can reduce its cost by
unilaterally changing its own action, i.e., an action profile
l∗ = (li∗ , l∗−i ) ∈ X is the Nash equilibrium if for all the
players
Ci (li∗ , l∗−i ) ≤ Ci (li , l∗−i ), ∀i ∈ N
(2)
for all li ∈ Xi .
For
Pan aggregate game Γ with the aggregate function being
l̄ = N
i=1 li , an equilibrium of the aggregate game Γ is an
action profile l∗ ∈ X on which ∀i ∈ N
Ci (li∗ , li∗ +
N
X
lj∗ ) ≤ Ci (li , li +
j=1,j6=i
N
X
lj∗ ), ∀li ∈ Xi
j=1,j6=i
∗
and
PN the∗ associated equilibrium aggregate is denoted as l̄ =
i=1 li .
B. Graph Theory
For a graph defined as G = (V, E) where E is the
edge set satisfying E ⊂ V × V with V = {1, 2, · · · , N }
being the set of nodes in the network, it is undirected if
for every (i, j) ∈ E, (j, i) ∈ E. An undirected graph
is connected if there exists a path between any pair of
distinct vertices. The element in the adjacency matrix A
is defined as aij = 1 if node j is connected with node i,
else, aij = 0. The neighboring set of agent i is defined as
Ni = {j ∈ V |(j, i) ∈ E}. The Laplacian matrix for the
graph L is defined as L = D̄ − A where D̄ is defined as a
diagonal matrix whose ith diagonal element
is equal to the
P
out degree of node i, represented by N
a
.
ij
j=1
C. Dynamic Average Consensus
Theorem 1: [39], [40] Let G be a connected, undirected
graph, L be the Laplacian of G. Then, for any constant ω ∈
RN , the state of the following system
ẋ
−I − L −L
x
ω
=
+
,
ẏ
L
0
y
0
and limin , limax are the minimal and maximal acceptable
energy consumption for user i, respectively. The problem
is defined as follows.
Problem 1: (Strategy Design for the Electricity Consumers in the Energy Consumption Game) In the energy
consumption game, the electricity users are the players. The
energy consumption li and Ci (li , l̄) are the action and cost
of player i, respectively. Player i’s objective is defined as
Generation network
Generator 2
Generator M
.......
Generator i
.......
Generator 1
wholesale price
demand
power transmission and
distribution
Energy providers
demand
retail price
.......
.......
EMC
AMI
EMC
AMI
EMC
AMI
.......
EMC
AMI
min Ci (li , l̄)
.......
user N
user j
user 1
.......
.......
EMC
AMI
li
subject to limin ≤ li ≤ limax , i ∈ N,
.......
.......
user i
user 2
User network
Fig. 1: A simplified illustration of the electricity buying and
selling model.
with arbitrary initial conditions x(0), y(0) ∈ RN remains
bounded and x(t) converges exponentially to N1 1T ω1 as
t → ∞ where 1 denotes an N dimensional column vector
composed of 1.
D. Saddle Point
The pair (x∗ , y∗ ) is the saddle point of F (x, y) if
F (x∗ , y) ≤ F (x∗ , y∗ ) ≤ F (x, y∗ ),
(3)
is satisfied [5], [6].
III. S YSTEM M ODEL AND P ROBLEM F ORMULATION
Consider an electricity market with a network of N
price-anticipating users. The simplified illustration of the
electricity buying and selling model is shown in Fig. 1.
The users are equipped with an energy-management
controller (EMC) and an advanced metering infrastructure
(AMI) [38]. The EMC is used to schedule the electricity
usage for the corresponding user. The AMI enables bidirectional communication among the electricity users and the
centralized agent (e.g., the energy provider). The communication between the electricity users and their neighbors can
be modeled by an undirected and connected graph. In this
paper, we suppose that the electricity users schedule their
energy consumption by minimizing their own costs.
Let li be the energy consumption of user i. Then, the cost
of user i can be quantified as [38],
Ci (li , l̄) = Vi (li ) + P (l̄)li ,
(4)
where ¯l is the aggregatePenergy consumption of all the
N
electricity users, i.e., l̄ = j=1 lj . Furthermore, Vi (li ) is the
load curtailment cost [38]. The term P (l̄)li represents the
billing payment for the consumption of energy li where the
price P (l̄) is a function of the aggregate energy consumption
l̄.
For each user, the energy consumption should be within
its acceptable range, i.e., li ∈ [limin , limax ] where limin < limax
where N = {1, 2, · · · , N } denotes the set of electricity
users. The aggregate energy consumption ¯l is unknown to
the players. Suppose that the pure-strategy Nash equilibrium
exists and is isolated. Furthermore, Ci , i ∈ N are smooth
functions over RN . Design a control strategy for the players
such that they can search for the Nash equilibrium.
Remark 1: In practice, providing the aggregate energy
consumption l̄ to all the users is challenging for the centralized agent when the users are dynamically updating their
actions. Hence, we consider the Nash equilibrium seeking
under the condition that the users have no access to the
aggregate energy consumption l̄. But the users are allowed
to communicate with their neighbors on the estimations of
the aggregate energy consumption. Furthermore, we suppose
that the centralized agent can broadcast the total number of
electricity users N to all the electricity users in the network.
In summary, the energy consumption control considered
in this paper is based on the following assumption.
Assumption 1: The electricity users can communicate
with their neighbors via an undirected and connected graph.
Furthermore, the total number of the electricity users, N , is
known to all the electricity users.
IV. E NERGY C ONSUMPTION G AME D ESIGN
A NALYSIS
AND
In this section, the energy consumption game is considered
in the general form (the pricing function is not specified). For
simplicity, the constraints limin ≤ li ≤ limax , i ∈ N, are not
considered in this section.
A. Game Analysis
Before we proceed to facilitate the subsequent analysis,
the following assumptions are made.
Assumption 2: [29] There exists isolated, stable Nash
equilibrium on which
∂Ci ∗
(l ) = 0,
∂li
∂ 2 Ci ∗
(l ) > 0, ∀i ∈ N
∂li2
where l∗ denotes the Nash equilibrium.
(5)
Assumption 3: [29] The matrix
2
∂ 2 C1
∂ C1 ∗
(l∗ ) · · ·
(l ) ∂l
∂l21
1 ∂l2
∂ 2 C2 ∗
2
∂l ∂l (l ) ∂∂lC22 (l∗ )
1
2
2
B=
..
..
.
.
2
∂ CN
∗
···
∂l1 ∂lN (l )
and
2
∂ C1
∂l1 ∂lN
(l∗ )
..
.
∂ 2 CN
∂l2N
(l∗ )
κ̇µ = 0,
(6)
is strictly diagonally dominant.
Inspired by [25], let Di denote player i’s estimation on the
aggregate energy consumption. By using the estimations, the
players’ objectives can be rewritten as:
Problem 2: Player i’s objective is
min Ci (li , Di ),
(7)
li
PN
where Di = j=1 lj , ∀i ∈ N.
In the following, a consensus based method will be
proposed to search for the Nash equilibrium of the energy
consumption game (without considering the constraints).
B. Nash Equilibrium Seeking for the Aggregate Energy Consumption Game
Based on the consensus protocol in [39]-[40], the Nash
seeking strategy for player i, i ∈ N is designed as
Ḋi = −Di −
κ̇i =
X
X
(Di − Dj ) −
j∈Ni
(Di − Dj )
X
(κi − κj ) + N li (8a)
j∈Ni
(8b)
j∈Ni
∂P (Di )
∂Vi
+ P (Di ) + li
),
l˙i = −k̄i (
∂li
∂Di
(8c)
where Ni denotes the neighboring set of player i, k̄i =
δki , i ∈ N, δ is a small positive parameter and ki is a fixed
positive parameter. Furthermore, κi , i ∈ N are intermediate
variables.
Writing (8) in the concatenated form gives
−I − L −L
D
Nl
Ḋ
=
+
(9a)
L
0
κ
0
κ̇
∂P (Di )
∂Vi
+ P (Di ) + li
]vec ), (9b)
l̇ = − δk([
∂li
∂Di
where k is defined as k = diag{ki }, i ∈ N, D, κ, l
are the concatenated vectors of Di , κi , li , respectively. Let
e , µ] be an N × N dimensional orthonormal matrix
U = [U
such that µT L = 0 where µ is an N dimensional column
e
κ
where e
κ is an N − 1
vector. Furthermore, let κ = U
κµ
dimensional column vector [5]. Then, the closed-loop system
can be rewritten as
#
"
e
Ḋ
−I − L −LU
D
Nl
=
+
(10a)
eT L
e
κ
0
ė
κ
U
0
l̇ = −δk([
∂P (Di )
∂Vi
+ P (Di ) + li
]vec ),
∂li
∂Di
(10b)
(11)
which indicates that κµ (t) = κ(0).
e e (l) are the quasi-steady states of
Suppose that De (l), κ
e , respectively, i.e.,
D and κ
"
#
e
−I − L −LU
De (l)
Nl
+
= 0, (12)
eT L
e e (l)
κ
0
U
0
e#e (l) is unique for fixed l as
for fixed l. "Note that De (l), κ
e
−I − L −LU
the matrix
is Hurwitz [40]. Then, by
T
e
U L
0
direct calculation, it can be derived that
−I − L −L
L
0
De (l)
e
U [(e
κ (l))T , κµ ]T
+
Nl
0
= 0,
(13)
which indicates that De (l) is the equilibrium of the original
system. Hence,
PN by using Theorem 1, it can be concluded that
De (l) = 1 i=1 li for fixed l.
Theorem 2: Suppose that Assumptions 1-3 hold. Then,
there exists a positive constant δ ∗ such that for every
e (t)) converges exponentially to
0 < δP< δ ∗ , (l(t), D(t), κ
N
e ∗
∗
∗
e (l )) under (10) given that ||l(0) − l∗ ||,
(l , 1 i=1 li , κ
PN ∗
||D(0) − 1 i=1 li ||, ||e
κ(0) − κ
ee (l∗ )|| are sufficiently small.
Proof: Let τ = δt. Then, in the τ -time scale, the reducedsystem is
#
dD "
e
δ dτ
−I − L −LU
D
Nl
=
+
(14a)
κ
eT L
e
κ
0
δ de
U
0
dτ
∂P (Di )
∂Vi
dl
+ P (Di ) + li
]vec ).
= −k([
dτ
∂li
∂Di
(14b)
Quasi-steady state analysis: letting δ = P
0 freezes D and
e at the quasi-steady state on which Di = N
κ
j=1 lj , ∀i ∈ N.
Hence, the reduced-system is
∂P (Di )
∂Vi
dl
+ P (Di ) + li
]vec )
= − k([
dτ
∂li
∂Di
∂Ci (li , l−i )
]vec ).
= − k([
∂li
(15)
Linearizing (15) at the Nash equilibrium point l∗ gives,
dl
= −kB(l − l∗ ),
(16)
dτ
where −kB is Hurwitz by Assumption 3 and the Gershgorin
Circle Theorem [48]. Hence, the equilibrium point l∗ is locally exponentially stable under (15), i.e., there exist positive
constants ̺1 and ̺2 such that (denote the trajectory of (15)
as lr (τ ))
||lr (τ ) − l∗ || ≤ ̺1 e−̺2 τ ||lr (0) − l∗ ||,
(17)
given that ||lr (0) − l∗ || is sufficiently small.
e e (l)) satisfies,
Boundary-layer analysis: Since (De (l), κ
"
#
e
−I − L −LU
De (l)
Nl
+
= 0, (18)
eT L
e e (l)
κ
0
U
0
e
e e (l)
it can be derived
" that D (l), κ
# are linear functions of l
e
−I − L −LU
as the matrix
is non-singular.
eT L
U
0
Let
′
D = D − De
(19)
′
e = κ − κe .
κ
Then,
"
δ
′
dD
dτ′
de
κ
dτ
#
"
#
e
D′ + De
−I − L −LU
=
eT L
e′ + κ
ee
κ
U
0
"
T #
∂DeT
Nl
∂l
+
−δ
T
0
e eT
∂κ
From the analysis, it can be seen that all the states in (8)
stay bounded and l(t) produced by (8) converges to the Nash
equilibrium under the given conditions.
In this section, a Nash seeking strategy is proposed without
requiring the uniqueness of the Nash equilibrium. In the following section, energy consumption game for HVAC systems
where the Nash equilibrium is unique will be considered.
V. E NERGY C ONSUMPTION G AME AMONG A N ETWORK
OF HVAC S YSTEMS
For HVAC systems, the load curtailment cost may be
modeled as [38]
Vi (li ) = ν i ξ 2i (li − l̂i )2 , i ∈ N,
∂l
N
X
∂Vi
+ P (Di′ +
lj )
× (−k([
∂li
j=1
P
∂P (Di′ + N
j=1 lj )
]vec ))
+ li
PN
′
∂(Di + j=1 lj )
"
#
e
D′
−I − L −LU
=
eT L
e′
κ
U
0
#
"
T
N
∂DeT
X
∂Vi
′
∂l
(−k([
−δ
lj )
+
P
(D
+
T
i
e eT
∂κ
∂li
j=1
∂l
P
∂P (Di′ + N
j=1 lj )
]vec )).
+ li
P
N
∂(Di′ + j=1 lj )
(20)
Hence, in t-time scale
"
"
#
′ #
dD
e
D′
−I − L −LU
dt′
=
′
eT L
de
κ
e
κ
U
0
dt
#
"
eT T
∂D
∂Vi
∂l
(−k([
−δ
T
e eT
∂κ
∂li
∂l
(21)
N
X
′
lj )
+ P (Di +
j=1
PN
∂P (Di′ + j=1 lj )
+ li
]vec )).
PN
∂(Di′ + j=1 lj )
Letting δ = 0 gives the boundary-layer model of (14) as
"
#
=
Since the matrix
"
′
dD
dt′
de
κ
dt
"
−I − L
eT L
U
−I − L
eT L
U
#" ′ #
e
D
−LU
′
.
(22)
0
e
κ
#
e
−LU
is Hurwitz, the equi0
′
′
e =0
librium point of the boundary-layer model D = 0, κ
is exponentially stable, uniformly in (t, l).
Therefore, by Theorem 11.4 in [42], it can be concluded
that there exists
a positive constant δ ∗ such that for all 0 <
PN
∗
∗
e e (l∗ )) is exponentially stable under
δ < δ , (l , 1 i=1 li∗ , κ
(10).
(23)
> 0 and ˆli is
where ν i and ξ i are thermal coefficients,
the energy needed to maintain the indoor temperature of the
HVAC system. When the pricing function is
ν i ξ 2i
P (l̄) = a
N
X
li + p0 ,
(24)
i=1
2ν ξ 2
i i
where a is a non-negative constant and a < mini∈N N −3
for
N > 3, the uniqueness of the Nash equilibrium is ensured
[38].
Based on this given model, a Nash seeking strategy
will be proposed in this section for the players to search
for the unique Nash equilibrium (by assuming that a <
2ν i ξ 2i
for N > 3 in the rest of the paper).
mini∈N N −3
A. Nash Equilibrium Seeking for Energy Consumption Game
of HVAC Systems
Lemma 1: The energy consumption game is a potential
game with a potential function being
Q(l) =
N
X
ν i ξ 2i (li −l̂i )2 +
N
X
i=1
i=1
a(
N
X
lj )li +
j=1,j6=i
N
X
(ali2 +p0 li ).
i=1
(25)
Proof: Noticing that
∂Q(l)
∂Ci (l)
=
, ∀i ∈ N,
∂li
∂li
(26)
the conclusion can be derived by using the definition of
potential game.
Based on the primal-dual dynamics in [43], the Nash
seeking strategy for player i is designed as
X
X
Ḋi = −Di −
(Di − Dj ) −
(κi − κj ) + N li
j∈Ni
κ̇i =
X
j∈Ni
(27a)
(Di − Dj )
(27b)
j∈Ni
∂Vi
+ P (Di ) + ali − η i1 + η i2 )
l˙i = −k̄i (
∂li
η̇ i1 = m̄i1 η i1 (limin − li )
η̇ i2 = m̄i2 η i2 (li − limax ), i ∈ N
(27c)
(27d)
(27e)
where m̄ij = δmij for all i ∈ N, j ∈ {1, 2}, mij are fixed
positive parameters and η ij (0) > 0, i ∈ N, j ∈ {1, 2}.
By introducing the orthonormal matrix U as in Section
IV-B, the system in (27) can be rewritten as
#
"
e
Ḋ
−I − L −LU
D
Nl
=
+
(28a)
eT L
e
κ
0
ė
κ
U
0
∂Vi
+ P (Di ) + ali − η i1 + η i2 )
∂li
= m̄i1 η i1 (limin − li )
l̇i = −k̄i (
η̇ i1
η̇ i2 = m̄i2 η i2 (li −
limax ), i
∈N
(28b)
(28c)
(28d)
and
κ̇µ = 0.
(29)
Theorem 3: Suppose that Assumption 1 is satisfied. Then,
there exists β ∈ KL such that for each pair of strictly positive
real number (∆, v), there exists δ(∆, v) > 0, such that
||χ(t)|| ≤ β(||χ(0)||, δt) + v,
(30)
for all t ≥ 0, ||χ(0)|| ≤ ∆ under (28) given that η ij (0) >
0,P
i ∈ N, j ∈ {1, 2}. In (30), χ(t) = [(l(t) − l∗ )T , (D(t) −
N
e e (l∗ ))T , (η(t) − η ∗ )T ]T , and η, η ∗
1 i=1 li∗ )T , (e
κ(t) − κ
are defined in the subsequent proof.
Proof: Following the proof of Theorem 2 by using singular
perturbation analysis, the reduced-system in τ -time scale is
given by
N
X
dli
lj + p0 + ali − η i1 + η i2 )
= −ki (2ν i ξ 2i (li − l̂i ) + a
dτ
j=1
dη i1
= mi1 η i1 (limin − li )
dτ
dη i2
= mi2 η i2 (li − limax ), i ∈ N.
dτ
(31)
According to Lemma 1, the energy consumption game is
a potential game. Hence, the Nash seeking can be achieved
by solving
min Q(l)
l
subject to
limin
≤ li ≤
limax , i
(32)
∈ N.
In the following, we show that (31) can be used to solve
the problem in (32).
the Lagrangian function as L(l, η) = Q(l) +
PDefine
N
min
− li ) + η i2 (li − limax )) where η =
i=1 (η i1 (li
2N
[η 11 , η 12 , η 21 , η 22 , · · · , η i1 , η i2 , · · · , η N 1 , η N 2 ]T ∈ R+
.
The dual problem for the minimization problem in (32) can
be formulated as
max min L(l, η).
η≥0
1
VL = (l − l∗ )T k−1 (l − l∗ )+
2
2
N X
X
1
(η ij − η ∗ij − η ∗ij log(ηij ) + η ∗ij log(η ∗ij )),
m
ij
i=1 j=1
(34)
where (l∗ , η ∗ ) is the saddle point of L(l, η), it can be shown
that the saddle point of L(l, η) is globally asymptotically
stable under (31) by Corollary 2 of [43] given that η ij (0) >
0, i ∈ N, j ∈ {1, 2}.
Hence, the strategy in (31) enables l to converge to the
Nash equilibrium of the potential game asymptotically.
Combining this result with the exponential stability of the
boundary-layer system (see Theorem 2 for boundary-layer
analysis), the result can be derived by using Lemma 1 in
[44] (see also [45] for more details).
B. Nash Equilibrium Seeking for Energy Consumption Game
of HVAC Systems with A Unique Inner Nash Equilibrium
In the following, energy consumption game with a unique
inner Nash equilibrium is considered 1 . If the constraints
do not affect the value of the Nash equilibrium, the Nash
seeking strategy can be designed as
X
X
Ḋi = −Di −
(Di − Dj ) −
(κi − κj ) + N li
j∈Ni
X
j∈Ni
(35a)
(Di − Dj )
(35b)
∂Vi
+ P (Di ) + ali ).
l˙i = −k̄i (
∂li
(35c)
κ̇i =
j∈Ni
By introducing the orthonormal matrix U as in Section
IV-B, the system in (35) can be rewritten as
Ḋ
ė
κ
=
"
−I − L
eT L
U
l̇i = −k̄i (
e
−LU
0
#
D
e
κ
+
Nl
0
∂Vi
+ P (Di ) + ali ), i ∈ N
∂li
a
···
···
..
.
a
a
..
.
· · · 2ν N ξ 2N + 2a
.
(36a)
(36b)
and
κ̇µ = 0.
The Hessian matrix of Q(l) is
2ν 1 ξ 21 + 2a
a
2
a
2ν
ξ
+ 2a
2
2
H =
..
.
a
(33)
l
2ν ξ 2
i i
for N > 3, H is positive definite by
Since a < mini∈N N −3
the Gershgorin Circle Theorem [48]. Hence, Q(l) is strictly
convex in l as its Hessian matrix is positive definite. Noting
that the inequality constraints are linear, the problem in (32)
has strong duality [47]. Hence, l∗ is the optimal solution to
2N
the problem in (32) if and only if there exists η ∗ ∈ R+
such
∗
∗
that (l , η ) is the saddle point of L(l, η) by the saddle point
theorem [47].
By defining the Lyapunov candidate function as [43]
(37)
Theorem 4: Suppose that Assumption 1 is satisfied. Then,
there exists a positive constant δ ∗ such that for every
1 To make it clear, in this paper, we say that the Nash equilibrium is
an inner Nash equilibrium if the Nash equilibrium satisfies limin < li∗ <
i
= 0, ∀i ∈ N.
limax , ∀i ∈ N, i.e., the Nash equilibrium is achieved at ∂C
∂l
i
e (t)) converges exponentially to
0 < δP< δ ∗ , (l(t), D(t), κ
N
e e (l∗ )) under (36).
(l∗ , 1 i=1 li∗ , κ
Proof: Following the proof of Theorem 2 by using singular
perturbation analysis, the reduced-system at τ -time scale is
given by
Furthermore,
∂Vj
+ P (Dj ) + alj − η j1 + η j2 )
l˙j = −k̄j (
∂lj
η̇ j1 = m̄j1 η j1 (ljmin − lj )
(44)
η̇ j2 = m̄j2 η j2 (lj − ljmax ), j ∈ N, j 6= i
N
X
dli
lj + p0 + ali ), i ∈ N.
= −ki (2ν i ξ 2i (li − ˆ
li ) + a
dτ
j=1
(38)
From (38), it can be derived that
ėl = −kHel,
(39)
where el(t) = l − l∗ .
Hence, it can be shown that in τ -time scale
s
maxi∈N {ki } − mini∈N {ki }λmin {H}τ e
||el(τ )|| ≤
e
||l(0)||,
mini∈N {ki }
(40)
by defining the Lyapunov candidate function as VL =
1eT −1e
l [29].
2l k
Combining this result with the exponential stability of the
boundary-layer system (see Theorem 2 for boundary-layer
analysis), the conclusion can be derived by Theorem 11.4 in
[42] (see the proof of Theorem 2 for more details) .
Remark 2: When the constraints do not affect the value
of the Nash equilibrium, the updating strategy in (35) is a
special case of the one in (8). It can be seen that the result
in Theorem 4 is stronger than the result in Theorem 2 under
the given HVAC model.
C. Energy Consumption Game of HVAC Systems with Stubborn Players
In this section, a special case where some players commit
to the coordination process while keeping a constant energy
consumption is considered. Without loss of generality, we
suppose that player i is a stubborn player and updates its
action according to
X
X
Ḋi = −Di −
(Di − Dj ) −
(κi − κj ) + N lis ,
j∈Ni
κ̇i =
X
j∈Ni
(41a)
(Di − Dj ),
(41b)
if all the rational players adopt (27), and
∂Vj
l˙j = −k̄j (
+ P (Dj ) + alj ), j ∈ N, j 6= i
∂lj
if all the rational players adopt (35).
Corollary 1: Suppose that Assumption 1 is satisfied.
Then, there exists β ∈ KL such that for each pair of strictly
positive real number (∆, v), there exists δ(∆, v) > 0, such
that
||χbr (t)|| ≤ β(||χbr (0)||, δt) + v,
(46)
for all t ≥ 0, ||χbr (0)|| ≤ ∆ under (42) and (44)
given that η kj (0) > 0, k ∈ N, k 6= i, j ∈ {1, 2}.
P
br
In (46), χbr (t) = [ľ−i (t)T , (D(t) − 1( N
j=1,j6=i lj +
e
s T
br T T
e (lbr
lis ))T , (e
κ(t) − κ
−i , li )) , (η −i (t) − η −i ) ] , ľ−i (t) =
T
[ľ1 (t), · · · , ľi−1 (t), ľi+1 (t), · · · , ľN (t)] , ľj (t) = lj (t) −
ljbr , j ∈ N, j 6= i and ljbr denotes the best response strategy of
br
player j, lbr
−i is the concatenated vector of lj , i ∈ N, j 6= i,
br
η −i , η −i are defined in the subsequent proof.
Furthermore, if the constraints ljmin ≤ lj ≤ ljmax , j ∈
N, j 6= i do not affect the values of the best response
∗
strategies, then, there exists a positive constant
PNδ suchbrthat
∗
for every 0 < δ < δ , (ľ−i (t), D(t) − 1( j=1,j6=i lj +
s
e (t)− κ
e e (lbr
lis ), κ
−i , li )) converges exponentially to zero under
(42) and (45).
Proof: Following the proof of Theorem 2 by using singular
perturbation analysis to get the reduced-system for both
cases. The first part of the Corollary can be derived by
noticing that the following is satisfied,
′
∂Cj
∂Q (l−i )
=
,
∂lj
∂lj
′
Q (l−i ) =
Ḋ
ė
κ
=
"
e
−I − L −LU
T
e
U L
0
#
+
N
X
ν j ξ 2j (lj − l̂j )2 +
N
X
j=1,j6=i
D
e
κ
+
Nl
0
, (42)
a(
N
X
lk )lj
k=1,k6=i,k6=j
(alj2 + (p0 + alis )lj ).
j=1,j6=i
(48)
Define
′
′
L (l−i , η −i ) = Q (l−i ) +
N
X
η j1 (ljmin − lj )
j=1,j6=i
+
N
X
η j2 (lj −
(49)
ljmax ),
j=1,j6=i
in which the ith component of l is fixed to be lis , and
κ̇µ = 0.
N
X
j=1,j6=i
j∈Ni
(47)
for all j 6= i, where
lis
where
is the constant energy consumption of player
i. Furthermore, all the rational players adopt (35) if the
constraints do not affect the value of all the players’ best
response strategies, else, all the rational players adopt (27).
By introducing the orthonormal matrix U as in Section
IV-B, then
(45)
2(N −1)
(43)
where η −i ∈ R+
denotes the concatenated vector of
η jk , j ∈ N and j 6= i, k ∈ {1, 2}.
Then, the dual problem is
5
1
′
max min L (l−i , η−i ).
η −i ≥0 l−i
Noticing that lbr
−i is the best response strategies of the rational
2(N −1)
players if and only if there exists η br
such that
−i ∈ R+
′
br
br
(l−i , η −i ) is the saddle point of L (l−i , η −i ), the rest of the
proof follows that of Theorem 3.
If the constraints ljmin ≤ lj ≤ ljmax , j ∈ N, j 6= i do not
affect the value of the best response strategies and all the
rational players adopt (35), then, the reduced-system at the
quasi-steady state is
∂Cj (l−i )
dlj
= −kj
dτ
∂lj
= −kj [2ν j ξ 2j (lj − ˆ
lj ) + a
N
X
lj + p0 + alis + alj ]
2
4
3
Fig. 2: Communication graph for the electricity users
TABLE I: Parameters in the simulation.
ˆ
li (kWh)
limax (kWh)
limin (kWh)
a
p0 (MU/kWh)
user 1
50
60
40
user 2
55
66
44
user 3
60
72
46
0.04
5
user 4
65
78
52
user 5
70
84
56
j=1,j6=i
= −kj [2ν j ξ 2j ľj + a
N
X
ˇ
lj + aľj ], i ∈ N.
j=1,j6=i
Writing (50) in the concatenated form gives
A. Simulation Setup
dľ−i
= −kH1 ľ−i ,
dτ
(51)
where H1 ∈ RN −1 × RN −1 is defined as
h1
H1 =
.
..
a
a
..
.
···
a
hi−1
hi+1
..
.
···
VI. N UMERICAL E XAMPLES
(50)
..
.
hN
2ν ξ 2
,
i i
with hj = 2ν j ξ 2j + 2a. Since a < mini∈N N −3
for N > 3,
it can be derived that the matrix H1 is symmetric and strictly
diagonally dominant with all the diagonal elements positive.
Hence, the matrix −kH1 is Hurwitz by the Gershgorin Circle
Theorem [48].
By similar analysis in Theorem 4, the conclusion can be
derived.
Remark 3: In this Corollary, only one player (i.e., player
i) is supposed to be stubborn. However, this is not restrictive
as similar results can be derived if multiple stubborn players
exist.
Remark 4: In the proposed Nash seeking strategy, the
players only communicate with its neighbors on Di and κi .
They do not communicate on their own energy consumption
li . Hence, the proposed Nash seeking strategy does not
lead to privacy concern for the users. The work in [29]
provided an extremum seeking method to seek for the Nash
equilibrium in non-cooperative games. However, the method
in [29] can’t be directly implemented for Nash seeking
in the energy consumption game if the aggregate energy
consumption is not directly available to the players.
In this section, we consider a network of 5 commercial/industrial users that are equipped with HVAC systems.
The electricity users communicate with each other via an
undirected and connected graph as shown in Fig. 2. The cost
function for electricity user i is
Ci (li , l̄) = ν i ξ 2i (li − ˆli )2 + P (l̄)li ,
(52)
PN
where the pricing function P (l̄) = a i=1 li + p0 [38].
Without loss of generality, suppose that ν i ξ 2i for all i ∈ N
are normalized to 1 in the simulation. For i ∈ {1, 2, 3, 4, 5},
limin = 0.8l̂i and limax = 1.2l̂i except that in Section VI-B,
l1min and l1max are set separately. The parameters are listed
in Table I2 .
B. Energy Consumption Control of HVAC Systems
In this section, we suppose that l1min = 45 and l1max =
55. By direct computation, it can be derived that the Nash
equilibrium is l∗ = (45, 46.4, 51.3, 56.2, 61.1)(kWh). The
equilibrium aggregate is l̄∗ = 259.9(kWh). Hence, the Nash
equilibrium is not an inner Nash equilibrium.
The simulation results by using the seeking strategy in
(27) are shown in Figs. 3-4.
Fig. 3 shows the users’ electricity energy consumptions
produced by the proposed seeking strategy in (27) and Fig.
4 indicates that the users’ estimations on the aggregate
energy consumptions converge to the actual aggregate energy
consumption.
From the simulation results, it can be seen that the energy
consumptions produced by the proposed method converge to
the Nash equilibrium of the energy consumption game.
2 MU
stands for Monetary Unit.
70
D. Energy Consumption Control with Stubborn Players
50
40
30
20
User 1’s electricity energy consumption
User 2’s electricity energy consumption
User 3’s electricity energy consumption
User 4’s electricity energy consumption
User 5’s electricity energy consumption
10
0
0
5
10
Time (Seconds)
15
20
Estimations of the aggregate energy consumption
Fig. 3: The users’ energy consumptions produced by the
proposed seeking strategy in (27).
300
250
VII. C ONCLUSIONS
200
150
100
The actual aggregate energy consumption
User 1’s estimation on the aggregate energy consumption
User 2’s estimation on the aggregate energy consumption
User 3’s estimation on the aggregate energy consumption
User 4’s estimation on the aggregate energy consumption
User 5’s estimation on the aggregate energy consumption
50
0
0
5
10
Time (Seconds)
15
20
Fig. 4: The actual aggregate energy consumption and the
users’ estimations on the aggregate energy consumption
produced by the proposed seeking strategy in (27).
C. Energy Consumption Control of HVAC Systems with A
Unique inner Nash equilibrium
In this section, we consider the energy consumption
game under the setting in Section VI-A. By direct calculation, it can be derived that the Nash equilibrium is
l∗ = (41.5, 46.4, 51.3, 56.2, 61.1)(kWh). The equilibrium
aggregate is l̄∗ = 256.7(kWh). The Nash equilibrium is an
inner Nash equilibrium and the seeking strategy in (35) is
used in the simulation. The simulation results produced by
the seeking strategy in (35) are shown in Figs. 5-6.
70
60
Users’ energy consumptions
In this section, we suppose that player 5 is a stubborn
player that commits to a constant energy consumption l5s =
100(kWh). Then, player 1-4’s best response strategies are
40.8(kWh), 45.7(kWh), 50.6(kWh), and 55.5(kWh), respectively. The aggregate energy consumption is 292.7(kWh). In
the simulation, the rational players adopt (35) to update their
actions. The stubborn player uses the seeking strategy in (41).
The simulation results produced by the proposed method are
shown in Figs. 7-8.
It can be seen that with the presence of the stubborn player,
all the other players’ actions converge to the best response
strategies with respect to the stubborn action.
AND
F UTURE W ORK
This paper considers energy consumption control among
a network of electricity users. The problem is solved by
using an aggregative game on an undirected and connected
graph. To estimate the aggregate energy consumption, which
is supposed to be unknown to the players during the Nash
seeking process, an average consensus protocol is employed.
The convergence property is analytically studied via using
singular perturbation and Lyapunov stability analysis. A
general energy consumption game where multiple Nash equilibria may exist is firstly considered. A Nash seeking strategy
based on consensus is proposed to enable the users to search
for the Nash equilibrium. Energy consumption control of
HVAC systems with linear pricing functions is then studied.
Convergence results are provided. Furthermore, stubborn
players are investigated and it is shown that the rational
players’ actions converge to the best response strategies.
For future directions, the following aspects would be
considered:
1) The design of incentive provoking mechanisms. As
Nash solution is usually not efficient from the systemlevel perspective, socially optimal solution might be
preferred if coordination is allowed. Incentive provoking mechanisms can be designed to motivate the electricity users to coordinate such that system efficiency
can be improved [5].
2) Analysis of the energy consumption game with the
existence of cheaters. This includes the detection of
50
40
30
User 1’s electricity energy consumption
User 2’s electricity energy consumption
User 3’s electricity energy consumption
User 4’s electricity energy consumption
User 5’s electricity energy consumption
20
10
0
0
5
Time (Seconds)
10
Fig. 5: The users’ energy consumptions produced by the
proposed seeking strategy in (35).
From the simulation results, it can be seen that the
users’ energy consumptions converge to the unique Nash
equilibrium.
Estimations on the aggregate load energy consumption
Users’ energy consumptions
60
300
250
200
150
100
50
0
0
The actual aggregate energy consumption
User 1’s estimation on the aggregate energy consumption
User 2’s estimation on the aggregate energy consumption
User 3’s estimation on the aggregate energy consumption
User 4’s estimation on the aggregate energy consumption
User 5’s estimation on the aggregate energy consumption
5
Time (Seconds)
10
Fig. 6: The actual aggregate energy consumption and the
users’ estimations on the aggregate energy consumptions
produced by the seeking strategy in (35).
120
Users’ energy consumptions
100
80
60
40
User 1’s electricity energy consumption
User 2’s electricity energy consumption
User 3’s electricity energy consumption
User 4’s electricity energy consumption
User 5’s electricity energy consumption
20
0
0
5
Time (Seconds)
10
Fig. 7: The users’ energy consumptions produced by the
proposed method (i.e., the rational players adopt (35) and
the stubborn player adopts (41)).
cheaters, the design of penalty (e.g., [49]) or reward
algorithms to prevent cheating behaviors, etc.
3) Nash seeking for energy consumption game under
various communication conditions.
R EFERENCES
[1] L. Gelazanska and K. Gamage, “Demand side management in smart
grid: a review and proposals for future direction,” Sustainable Cites
and Society, Vol. 11, pp. 22-30, 2014.
[2] J. Vardakas, N. Zorba and C. Verikoukis, “A survey on demand
response programs in smart grid: pricing methods and optimization
algorithms,” IEEE Communications Surveys and Tutorials, Vol. 17,
pp. 152-178, 2014.
[3] G. Wen, G. Hu, J. Hu, X. Shi and G. Chen, “Frequency regulation
of source-grid-load systems: a compound control strategy,” IEEE
Transactions on Industrial Informatics, accepted, to appear, DOI:
10.1109/TII.2015.2496309.
[4] W. Yu, G. Wen, X. Yu, Z. Wu and J. Lv, “Bridging the gap between
complex networks and smart grids,” Journal of Control and Decision,
Vol. 1, pp. 102-114, 2014.
[5] M. Ye and G. Hu, “Distributed extremum seeking for constrained networked optimization and its application to energy consumption control
in Smart Grid,” IEEE Transactions on Control Systems Technology,
accepted, to appear, DOI. 10.1109/TCST.2016.2517574.
[6] M. Ye and G. Hu, “A distributed extremum seeking scheme for
networked optimization,” IEEE Conference on Decision and Control,
pp. 4928-4933, 2015.
[7] P. Samadi, H. Mohsensian-Rad, R. Schober and V. Wong, “Advanced
demand side management for future smart grid using mechanism
design,” IEEE Transactions on Smart Grid, Vol. 3, pp. 1170-1180,
2012.
[8] E. Nekouei, T. Alpcan and D. Chattopadhyay, “Game-theoretic frameworks for demand response in electricity markets,” IEEE Transactions
on Smart Grid, Vol. 6, pp. 748-758, 2014.
The aggregate energy consumption
300
250
200
150
100
50
0
0
The actual aggregate energy consumption
User 1’s estimation on the aggregate energy consumption
User 2’s estimation on the aggregate energy consumption
User 3’s estimation on the aggregate energy consumption
User 4’s estimation on the aggregate energy consumption
User 5’s estimation on the aggregate energy consumption
5
Time (Seconds)
10
Fig. 8: The actual aggregate energy consumption and the
users’ estimations on the aggregate energy consumption
produced by the proposed method (i.e., the rational players
adopt (35) and the stubborn player adopts (41)).
[9] H. Mohsenian-Rad, V. Wong, J. Jatskevich, R. Schober and A. Garcia, “Autonomous demand-side management based on game-theoretic
energy consumption scheduling for the future smart grid,” IEEE
Transactions on Smart Grid, Vol. 1, pp. 320-331, 2010.
[10] R. Deng, Z. Yang, J. Chen, N. Asr and M. Chow, “Residential energy
consumption scheduling: a coupled contraint game approach,” IEEE
Transactions on Smart Grid, Vol. 5, pp. 1340-1350, 2014.
[11] B. Chai, J. Chen, Z. Yang and Y. Zhang, “Demand response management with multiple utility companies: a two level game approach, ”
IEEE Transactions on Smart Grid, Vol. 5, pp. 722-731, 2014.
[12] N. Forouzandehmehr, M. Esmalifalak. H. Mohsenian-Rad and Z. Han,
“Autonomous demand response using stochastic differential games,”
IEEE Transactions on Smart Grid, Vol. 6, pp. 291-300, 2015
[13] C. Wu, H. Mohsenian-Rad and J. Huang, “Vehicle-to-aggregator
interaction game,” IEEE Transactions on Smart Grid, Vol. 3, pp. 434442, 2012.
[14] W. Tysgar, B. Chai, C. Yuen, D. Smith, K. Wood, Z. Yang and H. Poor,
“Three-party energy management with distributed energy resources in
smart grid,” IEEE Transactions on Industrial Electronics, Vol. 62, pp.
2487-2497, 2015.
[15] F. Meng and X. Zeng, “A stackelberg game-theoretic approach to
optimal real-time pricing for smart grid,” Springer Soft Computing,
Vol. 17, pp. 2365-2830, 2013.
[16] W. Tushar, J. Zhang, D. Smith, H. Poor and S. Thiebaux, “Prioritizing
consumers in smart grid: a game theoretic approach,” IEEE Transactions on Smart Grid, Vol. 5, pp. 1429-1438, 2014.
[17] S. Bu, F. Yu, Y. Chai, and X. Liu, “When the smart grid meets
the energy efficient communications: green wireless cellular networks
powered by the smart grid,” IEEE Transactions on Wireless Communication, Vol. 11, pp. 3014-3024, 2012.
[18] G. Asimakopoulou, A. Dimeas and N. Hatziargyriou, “Leader-follower
strategies for energy management of multi-microgrids,” IEEE Transactions on Smart Grid, Vol. 4, pp. 1909-1916, 2013.
[19] P. Yang, G. Tang and A. Nehorai, “A game-theoretic approach for
optimal time-of-use electricity pricing,” IEEE Transactions on Power
System, Vol. 28, pp. 884-892, 2013.
[20] L. Chen, N. Li, S. Low and J. Doyle, “Two market models for demand
response in power networks,” IEEE SmartGridComm, pp. 397-402,
2010.
[21] S. Maharjan, Q. Zhu, Y. Zhang, S. Gjessing and T. Barsar, “Dependable demand response management in the smart grid: a stackelberg
game approach,” IEEE Transactions on Smart Grid, Vol. 4, pp. 120132, 2013.
[22] M. Jesen, “Aggregate games and best-reply potentials,” Econom.
Theory, Vol. 43, pp. 45-66, 2010.
[23] D. Acemoglu and M. Jensen, “Aggregate comparative statics,” Games
and Economic Behavior, vol. 81, pp. 27-49, 2013.
[24] D. Martimort and L. Stole, “Representing equilibrium aggregates in
aggregate games with applications to common agency,” Games and
Economic Behavior, Vol. 76, pp. 753-772, 2012.
[25] J. Koshal, A. Nedic and U. Shanbhag, “A gossip algorithm for aggregate games on graphs,” IEEE Conference on Decision and Control,
pp. 4840-4845, 2012.
[26] M. Zhu and E. Frazzoli, “On distributed equilibrium seeking for generalized convex games,” IEEE Conference on Decision and Control,
pp. 4858-4863, 2012.
[27] B. Gharesifard and J. Cortes, “Distributed convergence to Nash
equilibria in two-network zero-sum games,” Automatica, Vol. 49, pp.
1683-1692, 2013.
[28] S. Stankovic, K. H. Johansson, and D. M. Stipanovic, “Distributed
seeking of Nash equilibria with applications to mobile sensor networks,” IEEE Transactions on Automatic Control, Vol. 57, pp.904-919,
2012
[29] P. Frihauf, M. Krstic and T. Basar, “Nash equilibrium seeking in noncoopearive games,” IEEE Transactions on Automatic Control, Vol. 57,
pp. 1192-1207, 2012.
[30] F. Salehisadaghiani and L. Pavel, “Nash equilibrium seeking by a
gossip-based algorithm,” IEEE Conference on Decision and Control,
pp. 1155-1160, 2014.
[31] M. Ye and G. Hu, “Game design for price based demand response,”
IEEE Power and Energy Society General Meeting, pp. 1-5, 2015.
[32] M. Ye and G. Hu, “Distributed seeking of time-varying Nash equilibrium for non-cooperative games,” IEEE Transactions on Automatic
Control, Vol. 60, pp. 3000-3005, 2015.
[33] J. Nash, “Non-cooperative games,” Ann. Math., Vol. 54, pp. 286-295,
1951.
[34] D. Monderer and L. Shapley, “Potential games,” Games and Economic
Behavior, Vol. 14, pp. 124-143, 1996.
[35] P. Dubey, O. Haimanko and A. Zapechelnyuk, “Strategic complements
and substitutes and potential games,” Games and Economic Behavior,
Vol. 54, pp. 77-94, 2006.
[36] A. Ozguler and A. Yildiz, “Foraging swarms as Nash equilibria of
dynamic games,” IEEE Transactions on Cybernetics, Vol. 44, pp. 979987, 2014.
[37] Y. Yang and X. Lim “Towards a snowdrift game optimization to vertex
cover of networks,” IEEE Transactions on Cybernetics, Vol. 43m pp.
948-956, 2013.
[38] K. Ma, G. Hu and C. Spanos, “Distributed energy consumption control
via real-time pricing feedback in smart grid,” IEEE Transactions on
Control Systems Technology, Vol. 22, pp. 1907-1914, 2014.
[39] R. Freeman, P. Yang and K. Lynch, “Stability and convergence properties of dynamic average consensus estimators,” IEEE Conference on
Decision and Control, pp. 398-403, 2006.
[40] A. Menon and J. Baras, “Collaborative extremum seeking for welfare
optimization,” IEEE Conference on Decision and Control, pp. 346351, 2014.
[41] A. Wood and B. Wollenberg, Power generation, Operation and Control,
Wiley, New York, 1996.
[42] H. Khailil, Nonlinear System, Prentice Hall, 3rd edtion, 2002.
[43] H. Durr, E. Saka and C. Ebenbauer, “A smooth vector field for
quadratic programming,” IEEE Conference on Decision and Control,
pp. 2515-2520, 2012.
[44] Y. Tan, D. Nesic and I. Mareels, “On non-local stability properties of
extremum seeking control,” Automatica, Vol. 42, pp. 889-903, 2006.
[45] A. Teel, L. Moreau, and D. Nesic, “A unified framework for input-tostate stability in systems with two time scales,” System and Control
Letters, Vol. 48, pp. 1526-1544 , 2003.
[46] Z. Feng, G. Hu and G. Wen, “Distributed consensus tracking for multiagent systems under two types of attacks,” International Journal of
Robust and Nonlinear Control, Vol. 26, pp. 896-918, 2016.
[47] S. Boyd and L. Vandenberghe, Convex optimization, Cambradge
University Press, 2004.
[48] R. Horn and C. Johnson, Matrix Analysis, Cambridge,
U.K.:Cambridge Univ. Press, 1985.
[49] K. Ma, G. Hu and C. Spanos, “A cooperative demand response scheme
using punishment mechanism and application to industrial refrigerated
warehouses,” IEEE Transactions on Industrial Informatics, Vol. 11, pp.
1520-1531, 2015.
| 3cs.SY
|
Towards Effective Low-bitwidth Convolutional Neural Networks∗†
Bohan Zhuang, Chunhua Shen, Mingkui Tan, Lingqiao Liu, Ian Reid
arXiv:1711.00205v2 [cs.CV] 17 Nov 2017
Abstract
This paper tackles the problem of training a deep
convolutional neural network with both low-precision
weights and low-bitwidth activations. Optimizing a
low-precision network is very challenging since the
training process can easily get trapped in a poor local minima, which results in substantial accuracy loss.
To mitigate this problem, we propose three simple-yeteffective approaches to improve the network training.
First, we propose to use a two-stage optimization strategy to progressively find good local minima. Specifically, we propose to first optimize a net with quantized
weights and then quantized activations. This is in contrast to the traditional methods which optimize them
simultaneously. Second, following a similar spirit of
the first method, we propose another progressive optimization approach which progressively decreases the
bit-width from high-precision to low-precision during
the course of training. Third, we adopt a novel learning
scheme to jointly train a full-precision model alongside
the low-precision one. By doing so, the full-precision
model provides hints to guide the low-precision model
training. Extensive experiments on various datasets
( i.e., CIFAR-100 and ImageNet) show the effectiveness
of the proposed methods. To highlight, using our methods to train a 4-bit precision network leads to no performance decrease in comparison with its full-precision
counterpart with standard network architectures ( i.e.,
AlexNet and ResNet-50).
Contents
1. Introduction
2
2. Related work
2
3. Methods
3.1. Quantization function revisited . . . . . .
3
3
∗ B. Zhuang, C. Shen, L. Liu and I. Reid are with The University of Adelaide, Australia. M. Tan is with South China University of Technology, China.
† Correspondence to C. Shen (e-mail: [email protected]).
1
3.2. Two-stage optimization . . . . . . . . . .
3.3. Progressive quantization . . . . . . . . .
3.4. Guided training with a full-precision network . . . . . . . . . . . . . . . . . . . .
3.5. Remark on the proposed methods . . . .
3.6. Implementation details . . . . . . . . . .
3
4
4
5
6
4. Experiment
4.1. Evaluation on ImageNet . . . . . . . . .
4.2. Evaluation on Cifar100 . . . . . . . . . .
4.3. Ablation study . . . . . . . . . . . . . . .
6
6
7
7
5. Conclusion
9
1. Introduction
target problem. Motivated by these observations, we
propose to train a full-precision network alongside the
target low-precision network. Also, in contrast to standard knowledge distillation methods, we do not require
to pre-train the guidance model. Rather, we allow the
two models to be trained jointly from scratch since we
discover that this treatment enables the two nets adjust
better to each other.
Compared to several existing works that achieve
good performance when quantizing both weights and
activations [13, 22, 29, 36], our methods is more considerably scalable to the deeper neural networks [9,10].
For example, some methods adopt a layer-wise training
procedure [29], thus their training cost will be significantly increased if the number of layers becomes larger.
In contrast, the proposed method does not have this issue and we have experimentally demonstrated that our
method is effective with various depth of networks (i.e.,
AlexNet, ResNet-50).
The state-of-the-art deep neural networks [9,17, 25]
usually involve millions of parameters and need billions of FLOPs during computation. Those memory
and computational cost can be unaffordable for mobile
hardware device or especially implementing deep neural networks on chips. To improve the computational
and memory efficiency, various solutions have been proposed, including pruning network weights [7, 8], low
rank approximation of weights [16, 33], and training a
low-bit-precision network [4, 35–37]. In this work, we
follow the idea of training a low-precision network and
our focus is to improve the training process of such a
network. Note that in the literature, many works adopt
this idea but only attempt to quantize the weights of
a network while keeping the activations to 32-bit floating point [4, 35, 37]. Although this treatment leads
to lower performance decrease comparing to its fullprecision counterpart, it still needs substantial amount
of computational resource requirement to handle the
full-precision activations. Thus, our work targets the
problem of training network with both low-bit quantized
weights and activations.
The solutions proposed in this paper contain three
components. They can be applied independently or
jointly. The first method is to adopt a two-stage training process. At the first stage, only the weights of a network is quantized. After obtaining a sufficiently good
solution of the first stage, the activation of the network
is further required to be in low-precision and the network will be trained again. Essentially, this progressive
approach first solves a related sub-problem, i.e., training a network with only low-bit weights and the solution of the sub-problem provides a good initial point
for training our target problem. Following the similar idea, we propose our second method by performing
progressive training on the bit-width aspect of the network. Specifically, we incrementally train a serial of
networks with the quantization bit-width (precision)
gradually decreased from full-precision to the target
precision. The third method is inspired by the recent
progress of mutual learning [34] and information distillation [1, 11, 21, 23, 31]. The basic idea of those works
is to train a target network alongside another guidance
network. For example, The works in [1, 11, 21, 23, 31]
propose to train a small student network to mimic the
deeper or wider teacher network. They add an additional regularizer by minimizing the difference between
student’s and teacher’s posterior probabilities [11] or
intermediate feature representations [1, 23]. It is observed that by using the guidance of the teacher model,
better performance can be obtained with the student
model than directly training the student model on the
2. Related work
Several methods have been proposed to compress
deep models and accelerate inference during testing.
We can roughly summarize them into four main categories: limited numerial percision, low-rank approximation, efficient architecture design and network pruning.
Limited numerical precision When deploying
DNNs into hardware chips like FPGA, network quantization is a must process for efficient computing and
storage. Several works have been proposed to quantize
only parameters with high accuracy [4, 35, 37]. Courbariaux et al. [4] propose to constrain the weights
to binary values (i.e., -1 or 1) to replace multiplyaccumulate operations by simple accumulations. To
keep a balance between the efficiency and the accuracy,
ternary networks [37] are proposed to keep the weights
to 2-bit while maintaining high accuracy. Zhou et al.
[35] present incremental network quantization (INQ) to
efficiently convert any pre-trained full-precision CNN
model into low-precision whose weights are constrained
to be either powers of two or zero. Different from these
methods, a mutual knowledge transfer strategy is proposed to jointly optimize the full-precision model and
its low-precision counterpart for high accuracy. What’s
more, we propose to use a progressive optimization approach to quantize both weights and activations for
better performance.
Low-rank approximation Among existing works,
some methods attempt to approximate low-rank filters
in pre-trained networks [16, 33]. In [33], reconstruction
error of the nonlinear responses are minimized layerwisely, with subject to the low-rank constraint to re2
duce the computational cost. Other seminal works attempt to restrict filters with low-rank constraints during training phrase [20,27]. To better exploit the structure in kernels, it is also proposed to use low-rank tensor decomposition approaches [5, 20] to remove the redundancy in convolutional kernels in pretrained networks.
Efficient architecture design The increasing
demand for running highly energy efficient neural
networks for hardware devices have motivated the
network architecture design. GoogLeNet [26] and
SqueezeNet [14] propose to replace 3x3 convolutional
filters with 1x1 size, which tremendously increase the
depth of the network while decreasing the complexity a
lot. ResNet [9] and its variants [10, 30] utilize residual
connections to relieve the gradient vanishing problem
when training very deep networks. Recently, depthwise separable convolution employed in Xception [3]
and MobileNet [12] have been proved to be quite effective. Based on it, ShuffleNet [32] generalizes the group
convolution and the depthwise separable convolution
to get the state-of-the-art results.
Pruning and sparsity Substantial effort have been
made to reduce the storage of deep neural networks in
order to save the bandwidth for dedicated hardware
design. Han et al. [7, 8] introduce “deep compression”, a three stage pipeline: pruning, trained quantization and Huffman coding to effectively reduce the
memory requirement of CNNs with no loss of accuracy.
Guo et al. [6] further incorporate connection slicing to
avoid incorrect pruning. More works [18,19,28] propose
to employ structural sparsity for more energy-efficient
compression.
process and the activation quantization process as follows:
Quantization on weights:
wq = Q(
xq = Q(f (x)).
(3)
Back-propagation with quantization function:
In general, the quantization function is nondifferentiable and thus it is impossible to directly apply
the back-propagation to train the network. To overcome this issue, we adopt the straight-through estimator [2, 13, 36] to approximate the gradients calculation.
∂z
Formally, we approximate the partial gradient ∂zrq with
an identity mapping, namely
can be approximated by
∂zq
∂zr
≈ 1. Accordingly,
∂l
∂l ∂zq
∂l
=
≈
.
∂zr
∂zq ∂zr
∂zq
∂l
∂zr
(4)
3.2. Two-stage optimization
With the straight-through estimator, it is possible
to directly optimize the low-precision network. However, the gradient approximation of the quantization
function inevitably introduces noisy signal for updating network parameters. Strictly speaking, the approximated gradient may not be the right updating direction. Thus, the training process will be more likely
to get trapped at a poor local minima than training a
full precision model. Applying the quantization function to both weights and activations further worsens
the situation.
To reduce the difficulty of training, we devise a twostage optimization procedure: at the first stage, we
only quanitze the weights of the network while setting
the activations to be full precision. After the converge
(or after certain number of iterations) of this model,
we further apply the quantization function on the activations as well and retrain the network. Essentially,
the first stage of this method is a related subproblem of
the target one. Compared to the target problem, it is
easier to optimize since it only introduces quantization
In this section, we will first revisit the quantization
function in the neural network and the way to train
it. Then we will elaborate our three methods in the
subsequent sections.
3.1. Quantization function revisited
A common practise in training a neural network
with low-precision weights and activations is to introduce a quantization function. Considering the general
case of k-bit quantization as in [36], we define the quantization function Q(·) to be
1
round((2k − 1)zr )
2k − 1
(2)
tanh(w)
+ 12 to obIn other words, we first use 2 max(|tanh(w)|)
tain a normalized version of w and then perform the
quantization, where tanh(·) is adopted to reduce the
impact of large values.
Quantization on activations:
Same as [36], we first use a clip function f (x) =
clip(x, 0, 1) to bound the activations to [0, 1]. After
that, we conduct quantize the activation by applying
the quantization function Q(·) on f (x).
3. Methods
zq = Q(zr ) =
1
tanh(w)
+ ).
2 max(|tanh(w)|) 2
(1)
where zr ∈ [0, 1] denotes the full-precision value and
zq ∈ [0, 1] denotes the quantized value. With this quantization function, we can define the weight quantization
3
guidance loss
full-precision
⨁
⨁
⨁
⨁
L1
low-precision
⨁
⨁
⨁
⨁
L2
Figure 1: Demonstration of the guided training strategy. We use the residual network structure for illustration.
3.3. Progressive quantization
function on weights. Thus, we are more likely to arrive
at a good solution for this sub-problem. Then, using it
to initialize the target problem may help the network
avoid poor local minima which will be encountered if
K
we train the network from scratch. Let Mlow
be the
high-precision model with K-bit. We propose to learn
k
a low-precision model Mlow
in a two-stage manner with
K
Mlow serving as the initial point, where k < K. The
detailed algorithm is shown in Algorithm 1.
The aforementioned two-stage optimization approach suggests the benefits of using a related easy optimized problem to find a good initialization. However,
separating the quantization of weights and activations
is not the only solution to implement the above idea. In
this paper, we also propose another solution which progressively lower the bitwidth of the quantization during
the course of network training. Specifically, we progressively conduct the quantization from higher precisions
to lower precisions (e.g., 32-bit → 16-bit → 4-bit →
2-bit). The model of higher precision will be used the
the starting point of the relatively lower precision, in
analogy with annealing.
Let {b1 , ..., bn } be a sequence precisions, where bn <
bn−1 , ..., b2 < b1 , bn is the target precision and b1 is set
to 32 by default. The whole progressive optimization
k
procedure is summarized in as Algorithm 2. Let Mlow
be the low-precision model with k-bit and Mf ull be the
full precision model. In each step, we propose to learn
k
Mlow
, with the solution in the (i − 1)-th step, denoted
K
by Mlow
, serving as the initial point, where k < K.
Algorithm 1: Two-stage optimization for k-bit
quantization
1
2
3
4
5
6
7
8
9
10
11
Input: Training data {(xi , yi )}N
i=1 ; A K-bit
K
precision model Mlow
.
k
Output: A low-precision deep model Mlow
with
weights Wlow and activations being
quantized into k-bit.
Stage 1: Quantize Wlow :
for epoch = 1, ..., L do
for t = 1, ...T do
Randomly sample a mini-batch data;
Quantize the weights Wlow into k-bit by
calling some quantization methods with
K-bit activations;
3.4. Guided training with a full-precision network
The third method proposed in this paper is inspired by the success of using information distillation [1, 11, 21, 23, 31] to train a relatively shallow network. Specifically, these methods usually use a teacher
model (usually a pretrained deeper network) to provide guided signal for the shallower network. Following
this spirit, we propose to train the low-precision network alongside another guidance network. Unlike the
work in [1, 11, 21, 23, 31], the guidance network shares
the same architecture as the target network but is pretrained with full-precision weights and activations.
However, a pre-trained model may not be necessarily
optimal or may not be suitable for quantization. As a
result, directly using a fixed pretrained model to guide
Stage 2: Quantize activations:
Initialize Wlow using the converged k-bit weights
from Stage 1 as the starting point;
for epoch = 1, ..., L do
for t = 1, ...T do
Randomly sample a mini-batch data;
Quantize the activations into k-bit by
calling some quantization methods while
keeping the weights to k-bit;
4
where k · k denotes some proper norms.
Let Lθ1 and Lθ2 be the cross-entropy classification
losses for the full-precision and low-precision model,
respectively. The guidance loss will be added to Lθ1
and Lθ2 , respectively, resulting in two new objectives
for the two networks, namely
Algorithm 2: Progressive quantization for accurate CNNs with low-precision weights and activations
Input: Training data {(xj , yj )}N
j=1 ; A pre-trained
32-bit full-precision model Mf ull as
baseline; the precision sequence
{b1 , ..., bn } where
bn < bn−1 , ..., b2 < b1 = 32.
bn
Output: A low-precision deep model Mlow
.
b1
1 Let Mlow = Mf ull , where b1 = 32;
2 for i = 2, ...n do
3
Let k = bi and K = bi−1 ;
k
4
Obtain Mlow
by calling some quantization
K
methods with Mlow
being the input;
(6)
L2 (Wlow ) = Lθ2 + λR(Wf ull , Wlow ).
(7)
and
where λ is a balancing parameter. Here, the guidance
loss R can be considered as some regularization on Lθ1
and Lθ2 .
In the learning procedure, both Wf ull and Wlow
will be updated by minimizing L1 (Wf ull ) and
L2 (Wlow ) separately, using a mini-batch stochastic
gradient descent method. The detailed algorithm is
shown in Algorithm 3. A high-bit precision model
K
k
is used as an initialization of Mlow
, where K >
Mlow
k. Specifically, for the full-precision model, we have
K = 32. Relying on Mf ull , the weights and activations
k
of Mlow
can be initialized by equations (2) and (3),
respectively.
Note that the training process of the two networks
are different. When updating Wlow by minimizing
L2 (Wlow ), we use full-precision model as the initialization and apply the forward-backward propagation rule
in Section 3.1 to fine-tune the model. When updating
Wf ull by minimizing L1 (Wf ull ), we use conventional
forward-backward propagation to fine-tune the model.
the target network may not produce the best guidance
signals. To mitigate this problem, we do not fix the
parameters of a pretrained full precision network as in
the previous work [34].
By using the guidance training strategy, we assume that there exist some full-precision models with
good generalization performance, and an accurate lowprecision model can be obtained by directly performing
the quantization on those full-precision models. In this
sense, the feature maps of the learned low-precision
model should be close to that obtained by directly
doing quantization on the full-precision model. To
achieve this, essentially, in our learning scheme, we can
jointly train the full-precision and low-precision models. This allows these two models adapt to each other.
We even find by doing so the performance of the fullprecision model can be slightly improved in some cases.
Formally, let Wf ull and Wlow be the full-precision
model and low-precision model, respectively. Let
µ(x; Wf ull ) and ν(x; Wlow ) be the nested feature maps
(e.g., activations) of the full-precision model and lowprecision model, respectively. To create the guidance signal, we may require that the nested feature
maps from the two models should be similar. However, µ(x; Wf ull ) and ν(x; Wlow ) is usually not directly comparable since one is full precision and the
other is low-precision.
To link these two models, we can directly quantize
the weights and activations of the full-precision model
by equations (2) and (3). For simplicity, we denote
the quantized feature maps by Q(µ(x; Wf ull )). Thus,
Q(µ(x; Wf ull )) and ν(x; Wlow ) will become comparable. Then we can define the guidance loss as:
R(Wf ull , Wlow ) =
L1 (Wf ull ) = Lθ1 + λR(Wf ull , Wlow ).
Algorithm 3: Guided training with a fullprecision network for k-bit quantization
1
2
3
4
5
6
Input: Training data {(xi , yi )}N
i=1 ; A pre-trained
32-bit full-precision model Mf ull ; A k-bit
k
precision model Mlow
.
k
Output: A low-precision deep model Mlow
with
weights and activations being quantized
into k bits.
k
Initialize Mlow
based on Mf ull ;
for epoch = 1, ..., L do
for t = 1, ...T do
Randomly sample a mini-batch data;
Quantize the weights Wlow and
activations into k-bit by minimizing
L2 (Wlow );
Update Mf ull by minimizing L1 (Wf ull );
3.5. Remark on the proposed methods
1
k Q(µ(x; Wf ull ))−ν(x; Wlow )k2 ,
2
(5)
The proposed three approaches tackle the difficulty
in training a low-precision model with different strate5
gies. They can be applied independently. However, it
is also possible to combine them together. For example,
we can apply the progressive quantization to any of the
steps in the two-stage approach; we can also apply the
guided training to any sub-step in the progressive training. Detailed analysis on possible combinations will be
experimentally evaluated in the experiment section.
2. “TS”: We apply the two-stage optimization strategy described in Sec. 3.2 and Algorithm 1 to quantize the weights and activations. We denote the
first stage as Stage1 and the second stage as
Stage2.
3. “PQ”: We apply the progressive quantization
strategy described in Sec. 3.3 and Algorithm 2 to
continuously quantize weights and activations simultaneously from high-precision (i.e., 32-bit) to
low-precision.
3.6. Implementation details
In all the three methods, we quantize the weights
and activations of all layers except that the input data
are kept to 8-bit. Furthermore, to promote convergence, we propose to add a scalar layer after the last
fully-connected layer before feeding the low-bit activations into the softmax function for classification. The
scalar layer has only one trainable small scalar parameter and is initialized to 0.01 in our approach.
During training, we randomly crop 224x224 patches
from an image or its horizontal flip, with the per-pixel
mean subtracted. We don’t use any further data augmentation in our implementation. We adopt batch
normalization (BN) [15] after each convolution before
activation. For pretraining the full-precision baseline
model, we use Nesterov SGD and batch size is set
to 256. The learning rate starts from 0.01 and is divided by 10 every 30 epochs. We use a weight decay
0.0001 and a momentum 0.9. For weights and activations quantization, the initial learning rate is set to
0.001 and is divided by 10 every 10 epochs. We use a
simple single-crop testing for standard evaluation. Following [31], for ResNet-50, we add only two guidance
losses in the 2 last groups of residual blocks. And for
AlexNet, we add two guidance losses in the last two
fully-connected layers.
4. “Guided”: We implement the guided training approach as described in Sec. 3.4 and Algorithm 3
to independently investigate its effect on the final
performance.
5. “PQ+TS”: We further combine PQ and TS together to see whether their combination can improve the performance.
6. “PQ+TS+Guided”: This implements the full
model by combining PQ, TS and Guided modules together.
7. “PQ+TS+Guided**”: Based on PQ+TS+Guided, we use full-precision weights for the first
convolutional layer and the last fully-connected
layer following the setting of [36, 37] to investigate
its sensitivity to the proposed method.
4.1. Evaluation on ImageNet
We further train and evaluate our model on
ILSVRC2012 [24], which includes over 1.2 million images and 50 thousand validation images. We report 4bit and 2-bit precision accuracy for both AlexNet and
ResNet-50. The sequence of bit-width precisions are
set as {32, 8, 4, 2}. The results of INQ [35] are directly
cited from the original paper. We did not use the sophisticated image augmentation and more details can
be found in Sec. 3.6. We compare our model to the
32-bit full-precision model, INQ, DoReFa-Net and the
baseline approach described in Sec. 3.1. For INQ, only
the weights are quantized. For DoReFa-Net, the first
convolutional layer uses the full-precision weights and
the last fully-connected layer use both full-precision
weights and activations.
Results on AlexNet: The results for AlexNet are
listed in Table 1. Compared to competing approaches,
we achieve steadily improvement for 4-bit and 2-bit
settings. This can be attributed to the effective progressive optimization and the knowledge from the fullprecision model for assisting the optimization process.
Furthermore, our 4-bit full model even outperforms
the full-precision reference by 0.7% on top-1 accuracy.
4. Experiment
To investigate the performance of the proposed
methods, we conduct experiments on Cifar100 and ImageNet datasets. Two representative networks, different precisions AlexNet and ResNet-50 are evaluated
with top-1 and top-5 accuracy reported. We use a
variant of AlexNet structure [17] by removing dropout
layers and add batch normalization after each convolutional layer and fully-connected layer. This structure is
widely used in previous works [36, 37]. We analyze the
effect of the guided training approach, two-stage optimization and the progressive quantization in details
in the ablation study. Seven methods are implemented
and compared:
1. “Baseline”: We implement the baseline model
based on DoReFa-Net as described in Section 3.1.
6
66
This may be due to the fact that on this data, we may
not need a model as complex as the full-precision one.
However, when the expected bit-width decrease to 2bit, we observe obvious performance drop compared
to the 32-bit model while our low-bit model still brings
2.8% top-1 accuracy increase compared to the Baseline
method.
Results on ResNet-50: The results for ResNet50 are listed in Table 2.
For the full-precision
model, we implement it using Pytorch following the
re-implementation provided by Facebook1 . Comparatively, we find that the performance are approximately
consistent with the results of AlexNet. Similarly, we
observe that our 4-bit full model is comparable with the
full-precision reference with no loss of accuracy. When
decreasing the precision to 2-bit, we achieve promising
improvement over the competing Baseline even though
there’s still an accuracy gap between the full-precision
model. Similar to the AlexNet on ImageNet dataset,
we find our 2-bit full model improves more comparing
with the 4-bit case. This phenomenon shows that when
the model becomes more difficult to optimize, the proposed approach turns out to be more effective in dealing with the optimization difficulty. To better understand our model, we also draw the process of training
for 2-bit ResNet-50 in Figure 3 and more analysis can
be referred in Sec. 4.3.
accuracy
65.5
65
Baseline
Stage1
Stage2
Stage2+Guided
64.5
64
5
10
15
20
25
30
epochs
70
60
accuracy
50
40
30
Baseline
Stage1
Stage2
Stage2+Guided
20
10
0
0
20
40
60
80
100
epochs
Figure 2: Validation accuracy of 4-bit AlexNet on Cifar100 using (a): the fine-tuning strategy; (b): learning
from scratch strategy. Stage2+Guided means we combine the methods Stage2 and Guided together during
optimization to investigate the effect of the guided training on the final performance.
4.2. Evaluation on Cifar100
Cifar100 is an image classification benchmark containing images of size 32x32 in a training set of 50,000
and a test set of 10,000. We use the AlexNet for our
experiment. The quantitative results are reported in
Table 3. From the table, we can observe that the
proposed approach steadily outperforms the competing method DoReFa-Net. Interestingly, the accuracy
of our 4-bit full model also surpasses its full precision
model. We speculate that this is due to 4-bit weights
and activations providing the right model capacity and
preventing overfitting for the networks.
learning from scratch. This indicates that the initial
point for training low-bitwidth model is crutial for obtaining good accuracy. In addition, the gap between
the Baseline and TS is obvious (i.e., 2.7 % in our experiment) with learning from scratch. This justifies
that the two-stage optimization strategy can effectively
help the model converge to a better local minimum.
4.3. Ablation study
The effect of quantizing all layers: This set of
experiments is performed to analyze the influence for
quantizing the first convolutional layer and the last
fully-connected layer. Several previous works [37] argue to keep these two layers precision as 32-bit floating
points to decrease accuracy loss. By comparing the
results of PQ+TS+Guided** and PQ+TS+Guided in
Table 4 and Table 5, we notice that the accuracy gap
between the two settings is not large, which indicates
that our model is not sensitive to the precision of these
two layers. It can be attributed to two facts. On one
hand, fine-tuning from 32-bit precision can drastically
decrease the difficulty for optimization. On the other
In this section, we analyze the effects of different
components of the proposed model.
Learning from scratch vs. Fine-tuning: To analyze the effect, we perform comparative experiments
on Cifar100 with AlexNet using learning from scratch
and fine-tuning strategies. The results are shown in
Figure 2, respectively. For convenience of exposition,
this comparison study is performed based on method
TS. First, we observe that the overall accuracy of finetuning from full-precision model is higher than that of
1 https://github.com/facebook/fb.resnet.torch
7
Accuracy
Top1
Top5
Full precision
57.2%
80.3%
5-bit (INQ)
57.4%
80.6%
4-bit (DoReFa-Net)
56.2%
79.4%
4-bit (Baseline)
56.8%
80.0%
4-bit (PQ+TS+Guided)
58.0%
81.1%
2-bit (DoReFa-Net)
48.3%
71.6%
2-bit (Baseline)
48.8%
72.2%
2-bit (PQ+TS+Guided)
51.6%
76.2%
2-bit (Baseline)
67.7%
84.7%
2-bit (PQ+TS+Guided)
70.0%
87.5%
Table 1: Top1 and Top5 validation accuracy of AlexNet on ImageNet.
Accuracy
Top1
Top5
Full precision
75.6%
92.2%
5-bit (INQ)
74.8%
91.7%
4-bit (DoReFa-Net)
74.5%
91.5%
4-bit (Baseline)
75.1%
91.9%
4-bit (PQ+TS+Guided)
75.7%
92.0%
2-bit (DoReFa-Net)
67.3%
84.3%
Table 2: Top1 and Top5 validation accuracy of ResNet-50 on ImageNet.
Accuracy
Top1
Top5
Full precision
65.4%
88.3%
4-bit (DoReFa-Net)
64.9%
88.5%
4-bit (Baseline)
65.0%
88.5%
4-bit (PQ+TS+Guided)
65.8%
88.6%
2-bit (DoReFa-Net)
63.4%
87.5%
2-bit (Baseline)
63.9%
87.6%
2-bit (PQ+TS+Guided)
64.6%
87.8%
Table 3: Top1 and Top5 validation accuracy of AlexNet on Cifar100.
Method
4-bit (TS)
4-bit (PQ)
4-bit (PQ+TS)
4-bit (Guided)
4-bit (PQ+TS+Guided)
4-bit (PQ+TS+Guided**)
2-bit (TS)
2-bit (PQ)
2-bit (PQ+TS)
2-bit (Guided)
2-bit (PQ+TS+Guided)
2-bit (PQ+TS+Guided**)
top-1
57.7%
57.5%
57.8%
57.3%
58.0%
58.1%
50.7%
50.3%
50.9%
50.0%
51.6%
52.5%
top-5
81.0%
80.8%
80.8%
80.4%
81.1%
81.2%
74.9%
74.8%
74.9%
74.1%
76.2%
77.3%
the guided training strategy further ease the instability
during training.
The effect of the two-stage optimization strategy: We further analyze the effect of each stage in
the TS approach in Figure 2 and Figure 3. We take
the 2-bitwidth ResNet-50 on ImageNet as an example.
In Figure 3, Stage1 has the minimal loss of accuracy.
As for the Stage2, although it incurs apparent accuracy decrease in comparison with that of the Stage1,
its accuracy is consistently better than the results of
Baseline in every epoch. This illustrates that progressively seeking for the local minimum point is crutial for
final better convergence. We also conduct additional
experiments on Cifar100 with 4-bit AlexNet. Interestingly, taking the model of Stage1 as the initial point,
the results of Stage2 even have relative increase using
two different training strategies as mentioned above.
This can be interpreted by that further quantizing the
activations impose more regularization on the model
to overcome overfitting. Overall, the two-step optimization strategy still performs steadily better than
the Baseline method which proves the effectiveness of
this simple mechanism.
Table 4: Evaluation of different components of the proposed method on the validation accuracy with AlexNet
on ImageNet.
Method
4-bit (TS)
4-bit (PQ)
4-bit (PQ+TS)
4-bit (Guided)
4-bit (PQ+TS+Guided)
4-bit (PQ+TS+Guided**)
2-bit (TS)
2-bit (PQ)
2-bit (PQ+TS)
2-bit (Guided)
2-bit (PQ+TS+Guided)
2-bit (PQ+TS+Guided**)
top-1
75.3%
75.4%
75.5%
75.3 %
75.7%
75.9%
69.2%
68.8%
69.4%
69.0%
70.0%
70.8%
top-5
91.9%
91.8%
92.0%
91.7%
92.0%
92.4%
87.0%
86.9%
87.0%
86.8%
87.5%
88.3%
The effect of the progressive quantization strategy: What’s more, we also separately explore the progressive quantization (i.e., PQ) effect on the final performance. In this experiment, we apply AlexNet on
the ImageNet dataset. We continuously quantize both
weights and activations simultaneously from 32-bit→8bit→4-bit→2-bit and explictly illustrate the accuracy
change process for each precision in Figure 4. The
quantitative results are also reported in Table 4 and
Table 5. From the figure we can find that for the 8-bit
and 4-bit, the low-bit model has no accuracy loss with
respect to the full precision model. However, when
quantizing from 4-bit to 2-bit, we can observe significant accuracy drop. Despite this, we still observe 1.5%
Table 5: Evaluation of different components of the proposed method on the validation accuracy with ResNet-50
on ImageNet.
hand, the progressive optimization approach as well as
8
75
guided training strategy and the progressive learning
mechanism can benefit from each other for further improvement.
70
Joint vs. without joint: We further illustrate the
joint optimization effect on guided training in Figure 5. For explaning convenience, we implement it
based on the method Stage2+Guided and report the
2-bit AlexNet top-1 validation accuracy on ImageNet.
From the figure, we can observe that both the fullprecision model and its low-precision counterpart can
benefit from learning from each other. In contrast, if
we keep the full-precision model unchanged, apparent
performance drop is observed. This result strongly supports our assumption that the high-precision and the
low-precision models should be jointly optimized in order to obtain the optimal gradient during training. The
improvement on the full-precision model may due to
the ensemble learning with the low-precision model and
similar observation is found in [34] but with different
task.
accuracy
65
60
55
Baseline
Stage1
Stage2
Stage2+Guided
50
45
0
10
20
30
epochs
Figure 3: Validation accuracy of 2-bit ResNet-50 on ImageNet. Stage2+Guided means we combine the methods
Stage2 and Guided together during training.
relative improvement by comparing the top-1 accuracy
over the 2-bit baseline, which proves the effectiveness
of the proposed strategy. It is worth noticing that the
accuracy curves become more unstable when quantizing to lower bit. This phenomenon is reasonable since
the precision becomes lower, the value will change more
frequently during training.
60
accuracy
55
60
55
50
45
32-bit model
32-bit model with joint training
2-bit(without joint training)
2-bit(joint training)
accuracy
50
45
40
0
40
30
20
30
epochs
32-bit(PQ)
8-bit(PQ)
4-bit(PQ)
2-bit(PQ)
2-bit(Baseline)
35
10
Figure 5: The effect of the joint training strategy using
AlexNet on ImageNet.
25
5
10
15
20
25
30
epochs
5. Conclusion
Figure 4: Validation accuracy of the progressive quantization approach using AlexNet on ImageNet.
In this paper, we have proposed three novel approaches to solve the optimization problem for quantizing the network with both low-precision weights and
activations. We first propose a two-stage approach to
quantize the weights and activations in a two-step manner. We also observe that continuously quantize from
high-precision to low-precision is also beneficial to the
final performance. To better utilize the knowledge from
the full-precision model, we propose to jointly learn the
low-precision model and its full-precision counterpart
to optimize the gradient problem during training. Using 4-bit weights and activations for all layers, we even
The effect of the jointly guided training: We also
investigate the effect of the guided joint training approach explained in Sec. 3.4. By comparing the results
in Table 4 and Table 5, we can find that Guided method
steadily improves the baseline method by a promising
margin. This justifies the low-precision model can always benefit by learning from the full-precision model.
What’s more, we can find PQ+TS+Guided outperforms PQ+TS in all settings. This shows that the
9
outperform the performance of the 32-bit model on ImageNet and Cifar100 with general frameworks.
[16]
References
[1] J. Ba and R. Caruana. Do deep nets really need to be
deep? In Proc. Adv. Neural Inf. Process. Syst., pages
2654–2662, 2014.
[2] Y. Bengio, N. Léonard, and A. Courville. Estimating or propagating gradients through stochastic
neurons for conditional computation. arXiv preprint
arXiv:1308.3432, 2013.
[3] F. Chollet.
Xception:
Deep learning with
depthwise separable convolutions.
arXiv preprint
arXiv:1610.02357, 2016.
[4] M. Courbariaux, Y. Bengio, and J.-P. David. Binaryconnect: Training deep neural networks with binary
weights during propagations. In Proc. Adv. Neural Inf.
Process. Syst., pages 3123–3131, 2015.
[5] E. L. Denton, W. Zaremba, J. Bruna, Y. LeCun, and
R. Fergus. Exploiting linear structure within convolutional networks for efficient evaluation. In Proc. Adv.
Neural Inf. Process. Syst., pages 1269–1277, 2014.
[6] Y. Guo, A. Yao, and Y. Chen. Dynamic network
surgery for efficient dnns. In Proc. Adv. Neural Inf.
Process. Syst., pages 1379–1387, 2016.
[7] S. Han, H. Mao, and W. J. Dally. Deep compression: Compressing deep neural networks with pruning, trained quantization and huffman coding. Proc.
Int. Conf. Learn. Repren., 2016.
[8] S. Han, J. Pool, J. Tran, and W. Dally. Learning both
weights and connections for efficient neural network.
In Proc. Adv. Neural Inf. Process. Syst., pages 1135–
1143, 2015.
[9] K. He, X. Zhang, S. Ren, and J. Sun. Deep residual
learning for image recognition. In Proc. IEEE Conf.
Comp. Vis. Patt. Recogn., pages 770–778, 2016.
[10] K. He, X. Zhang, S. Ren, and J. Sun. Identity mappings in deep residual networks. In Proc. Eur. Conf.
Comp. Vis., pages 630–645. Springer, 2016.
[11] G. Hinton, O. Vinyals, and J. Dean.
Distilling
the knowledge in a neural network. arXiv preprint
arXiv:1503.02531, 2015.
[12] A. G. Howard, M. Zhu, B. Chen, D. Kalenichenko,
W. Wang, T. Weyand, M. Andreetto, and H. Adam.
Mobilenets: Efficient convolutional neural networks
for mobile vision applications.
arXiv preprint
arXiv:1704.04861, 2017.
[13] I. Hubara, M. Courbariaux, D. Soudry, R. El-Yaniv,
and Y. Bengio. Binarized neural networks. In Proc.
Adv. Neural Inf. Process. Syst., pages 4107–4115, 2016.
[14] F. N. Iandola, S. Han, M. W. Moskewicz, K. Ashraf,
W. J. Dally, and K. Keutzer. Squeezenet: Alexnetlevel accuracy with 50x fewer parameters and¡ 0.5 mb
model size. arXiv preprint arXiv:1602.07360, 2016.
[15] S. Ioffe and C. Szegedy. Batch normalization: Accelerating deep network training by reducing internal co-
[17]
[18]
[19]
[20]
[21]
[22]
[23]
[24]
[25]
[26]
[27]
[28]
[29]
[30]
[31]
10
variate shift. In Proc. Int. Conf. Mach. Learn., pages
448–456, 2015.
Y.-D. Kim, E. Park, S. Yoo, T. Choi, L. Yang, and
D. Shin. Compression of deep convolutional neural
networks for fast and low power mobile applications.
arXiv preprint arXiv:1511.06530, 2015.
A. Krizhevsky, I. Sutskever, and G. E. Hinton. Imagenet classification with deep convolutional neural networks. In Proc. Adv. Neural Inf. Process. Syst., pages
1097–1105, 2012.
V. Lebedev and V. Lempitsky. Fast convnets using
group-wise brain damage. In Proc. IEEE Conf. Comp.
Vis. Patt. Recogn., pages 2554–2564, 2016.
B. Liu, M. Wang, H. Foroosh, M. Tappen, and M. Pensky. Sparse convolutional neural networks. In Proc.
IEEE Conf. Comp. Vis. Patt. Recogn., pages 806–814,
2015.
A. Novikov, D. Podoprikhin, A. Osokin, and D. P.
Vetrov. Tensorizing neural networks. In Proc. Adv.
Neural Inf. Process. Syst., pages 442–450, 2015.
E. Parisotto, J. L. Ba, and R. Salakhutdinov. Actormimic: Deep multitask and transfer reinforcement
learning. Proc. Int. Conf. Learn. Repren., 2016.
M. Rastegari, V. Ordonez, J. Redmon, and A. Farhadi.
Xnor-net: Imagenet classification using binary convolutional neural networks. In Proc. Eur. Conf. Comp.
Vis., pages 525–542, 2016.
A. Romero, N. Ballas, S. E. Kahou, A. Chassang,
C. Gatta, and Y. Bengio. Fitnets: Hints for thin deep
nets. Proc. Int. Conf. Learn. Repren., 2015.
O. Russakovsky, J. Deng, H. Su, J. Krause,
S. Satheesh, S. Ma, Z. Huang, A. Karpathy, A. Khosla,
M. Bernstein, et al. Imagenet large scale visual recognition challenge. Int. J. Comp. Vis., 115(3):211–252,
2015.
K. Simonyan and A. Zisserman. Very deep convolutional networks for large-scale image recognition. In
Proc. Int. Conf. Learn. Repren., 2015.
C. Szegedy, W. Liu, Y. Jia, P. Sermanet, S. Reed,
D. Anguelov, D. Erhan, V. Vanhoucke, and A. Rabinovich. Going deeper with convolutions. In Proc. IEEE
Conf. Comp. Vis. Patt. Recogn., pages 1–9, 2015.
C. Tai, T. Xiao, Y. Zhang, X. Wang, et al. Convolutional neural networks with low-rank regularization.
Proc. Int. Conf. Learn. Repren., 2016.
W. Wen, C. Wu, Y. Wang, Y. Chen, and H. Li. Learning structured sparsity in deep neural networks. In
Proc. Adv. Neural Inf. Process. Syst., pages 2074–2082,
2016.
J. Wu, C. Leng, Y. Wang, Q. Hu, and J. Cheng. Quantized convolutional neural networks for mobile devices.
In Proc. IEEE Conf. Comp. Vis. Patt. Recogn., pages
4820–4828, 2016.
S. Zagoruyko and N. Komodakis. Wide residual networks. arXiv preprint arXiv:1605.07146, 2016.
S. Zagoruyko and N. Komodakis. Paying more attention to attention: Improving the performance of convo-
[32]
[33]
[34]
[35]
[36]
[37]
lutional neural networks via attention transfer. Proc.
Int. Conf. Learn. Repren., 2017.
X. Zhang, X. Zhou, M. Lin, and J. Sun. Shufflenet:
An extremely efficient convolutional neural network
for mobile devices. arXiv preprint arXiv:1707.01083,
2017.
X. Zhang, J. Zou, K. He, and J. Sun. Accelerating
very deep convolutional networks for classification and
detection. IEEE Trans. Pattern Anal. Mach. Intell.,
38(10):1943–1955, 2016.
Y. Zhang, T. Xiang, T. M. Hospedales, and
H. Lu.
Deep mutual learning.
arXiv preprint
arXiv:1706.00384, 2017.
A. Zhou, A. Yao, Y. Guo, L. Xu, and Y. Chen. Incremental network quantization: Towards lossless cnns
with low-precision weights. Proc. Int. Conf. Learn.
Repren., 2017.
S. Zhou, Y. Wu, Z. Ni, X. Zhou, H. Wen, and
Y. Zou. Dorefa-net: Training low bitwidth convolutional neural networks with low bitwidth gradients.
arXiv preprint arXiv:1606.06160, 2016.
C. Zhu, S. Han, H. Mao, and W. J. Dally. Trained
ternary quantization. Proc. Int. Conf. Learn. Repren.,
2017.
11
| 1cs.CV
|
1
Asymptotic Signal Detection Rates
with 1-bit Array Measurements
arXiv:1711.00739v2 [cs.IT] 20 Feb 2018
Manuel S. Stein
Abstract—This work considers detecting the presence of a
band-limited random radio source using an antenna array
featuring a low-complexity digitization process with single-bit
output resolution. In contrast to high-resolution analog-to-digital
conversion, such a direct transformation of the analog radio
measurements to a binary representation can be implemented
hardware and energy-efficient. However, the probabilistic model
of the binary receive data becomes challenging. Therefore, we
first consider the Neyman-Pearson test within generic exponential
families and derive the associated analytic detection rate expressions. Then we use a specific replacement model for the binary
likelihood and study the achievable detection performance with 1bit radio array measurements. As an application, we explore the
capability of a low-complexity GPS spectrum monitoring system
with different numbers of antennas and different observation
intervals. Results show that with a moderate amount of binary
sensors it is possible to reliably perform the monitoring task.
Index Terms—1-bit ADC, analog-to-digital conversion, array
processing, exponential family, GPS, Neyman-Pearson test, quantization, detection, spectrum monitoring
I. I NTRODUCTION
Since, in 1965, Gordon E. Moore predicted a doubling
in computational capability every two years, chip companies
have kept pace with this prognosis and set the foundation
for digital systems which today allow processing high-rate
radio measurements by sophisticated algorithms. In conjunction with wireless sensor arrays, this results in advanced
signal processing capabilities, see, e.g. [1]. Unfortunately, in
the last decades, the advances regarding the analog circuits
forming the radio front-end were much slower. Therefore, in
the advent of the Internet of things (IoT), where small and
cheap objects are supposed to feature radio interfaces, cost
and power consumption of wireless sensors are becoming an
issue. In particular, analog-to-digital conversion [2], [3] shows
to set constraints on the digitization rate and the number of
antenna elements under strict hardware and power budgets.
In this context, we consider detection of a band-limited
source with unknown random structure by a sensor array
providing single-bit radio measurements. We formulate the
processing task as a binary hypothesis test regarding exponential family models and derive expressions for the asymptotic
detection rates. The results are used to determine the design
of a low-complexity GPS spectrum monitoring system.
This work was supported by the German Academic Exchange Service
(DAAD) with funds from the German Federal Ministry of Education and
Research (BMBF) and the People Program (Marie Skłodowska-Curie Actions)
of the European Union’s Seventh Framework Program (FP7) under REA grant
agreement no. 605728 (P.R.I.M.E. - Postdoctoral Researchers International
Mobility Experience).
M. S. Stein is with the Chair for Stochastics, Universität Bayreuth, Germany
(e-mail: [email protected]).
Note that detection with quantized signals has found attention in distributed decision making [4] where data is collected
through sensors at different locations and quantization is used
to diminish the communication overhead between sensors and
terminal node [5]–[7]. Optimum quantization for detection
is the focus of [8]–[11], while [12] considers the detection
performance degradation due to hard-limiting. For discussions
on symbol detection for quantized communication see, e.g.,
[13]–[18]. Detection for cognitive radio with single-antenna
1-bit receivers is considered in [19] while array processing
with 1-bit measurements is analyzed in [20]–[22].
II. P ROBLEM F ORMULATION
We consider a receive situation where a narrow-band random wireless signal is impinging on an array of S sensors,
y = γAx + η.
(1)
Each sensor features two outputs (in-phase and quadrature),
M
such that the receive
TvectorT yT ∈ Y = R , M =S 2S, can be
with y I , y Q ∈ R . Likewise,
decomposed y = y I y Q
T
∈ R2 consists of two
the random source x = xI xQ
zero-mean signal components with covariance matrix
(2)
Rx = Ex xxT = I,
where Eu [·] denotes the expectation concerning the probability distribution p(u) and I the identity matrix. The steering
T
matrix A = AT
∈ RM×2 , with AI , AQ ∈ RS×2
AT
Q
I
models a uniform linear sensor array response (half carrierwavelength inter-element distance) for a narrow-band signal
arriving from direction ζ ∈ R, such that
cos 0
sin 0
cos π sin (ζ)
sin π sin (ζ)
AI =
(3)
..
..
.
.
cos (S − 1)π sin (ζ) sin (S − 1)π sin (ζ)
and
AQ =
− sin 0
− sin π sin (ζ)
..
.
− sin (S − 1)π sin (ζ)
cos 0
cos π sin (ζ)
..
.
cos (S − 1)π sin (ζ)
.
(4)
The parameter γ ∈ R characterizes the source strength in
relation to the additive zero-mean sensor noise η ∈ RM with
(5)
Rη = Eη ηη T = I.
2
Due to the properties of the source and noise signals, the
receive data (1) can be modeled by a Gaussian distribution
exp − 12 y T R−1
y (γ)y
y ∼ py (y; γ) = p
,
(6)
(2π)M det (Ry (γ))
with covariance matrix
Ry (γ) = Ey;γ yy T = γ 2 AAT + I.
(7)
Based on the likelihood (6), the problem of signal detection
can be stated as the decision about which of the two models
H0 : y ∼ py (y; γ0 ),
H1 : y ∼ py (y; γ1 )
has generated the data (K independent array snapshots)
Y = y 1 y2 . . . yK ∈ Y K .
(8)
H0 : z ∼ pz (z; θ0 ),
(9)
(10)
where sign (u) is the element-wise signum function, i.e.,
(
+1 if [y]i ≥ 0
[z]i =
(11)
−1 if [y]i < 0.
Note, that (10) characterizes a low-complexity digitization
process without feedback and is, therefore, distinct from
sigma-delta conversion [23], where a single fast comparator
with feedback is used to mimic a high-resolution ADC.
Modeling the output of (10) by its exact parametric probability distribution function, requires computing the integral
Z
pz (z; γ) =
py (y; γ)dy
(12)
C
while the probability of correctly deciding for H1 is given by
Z
PD =
pZ (Z; θ1 )dZ.
(16)
Approaching the decision problem (14) under the desired test
size PFA and maximum PD , the Neyman-Pearson theorem
shows that it is optimum to use the likelihood ratio test [24]
L(Z) =
pZ (Z; θ1 )
> ξ′
pZ (Z; θ0 )
IN THE
E XPONENTIAL FAMILY
Consider the multivariate parametric exponential family
pz (z; θ) = exp β T (θ)φ(z) − λ(θ) + κ(z) ,
(13)
where θ ∈ RD constitute its parameters, β(θ) : RD → RL the
natural parameters, φ(z) : RM → RL the sufficient statistics,
λ(θ) : RD → R the log-normalizer and κ(z) : RM → R the
(17)
for the assignment of the critical region
C = {Z : L(Z) > ξ ′ },
(18)
′
while the decision threshold ξ is determined through
Z
PFA =
pZ (Z; θ 0 )dZ.
(19)
{Z:L(Z)>ξ ′ }
Based on the ratio (17), a test statistic T (Z) : Z K → R can
be formulated such that the binary decision is performed by
(
H0 if T (Z) ≤ ξ
decide
.
(20)
H1 if T (Z) > ξ
To analyze the performance of (20), it is required to characterize the distribution of T (Z) and evaluate (15) and (16). As the
data Z consists of K independent samples, the test statistic
can be factorized into a sum of independent components
T (Z) =
M
for all 2 points in Z = B , where Y(z) characterizes the
subset in Y which by (10) is transformed to z. Additionally,
(12) requires the orthant probability, which for M > 4 is an
open problem. As the multivariate Bernoulli model resulting
from (10) is part of the exponential family like (6), in the
following we resort to discussing the considered processing
task for generic data models within this broad class. Without
exactly specifying the quantized likelihood (12), this will allow
us to analyze the asymptotically achievable detection rates
with low-complexity 1-bit array measurements.
III. D ECISIONS
(14)
is to be performed. To this end, we assign a critical region
C ⊂ Z K and decide in favor of H1 if the observed data
satisfies Z ∈ C. The probability of erroneously deciding for
H1 while H0 is the true data-generating model is calculated
Z
PFA =
pZ (Z; θ0 )dZ,
(15)
Y(z)
M
H1 : z ∼ pz (z; θ1 )
C
Using model (1) in practical applications implies that a highresolution analog-to-digital converter (ADC) is available for
each output channel. Significant savings are possible when
using a converter with 1-bit output resolution. Such a receiver
can be modeled by an ideal sampling device with infinite
resolution followed by a hard-limiter
z = sign (y),
carrier measure. Given a data set Z ∈ Z K of the form (9),
the simple binary hypothesis test between
K
X
t(z k )
(21)
k=1
such that, by the central limit theorem, the test statistic in the
large sample regime follows the normal distribution
a
(T (Z) − µi )2
1
exp −
,
(22)
p T (Z)|Hi = √
2σi2
2πσi
a
where by = we denote asymptotic equality. Through the mean
and standard deviation of the test statistic
µi = EZ;θi [T (Z)] ,
r
h
σi =
EZ;θi
2 i
T (Z) − µi ,
the asymptotic performance is then given by
ξ − µ1
a
,
PD = Pr {T (Z) > ξ|H1 } = Q
σ1
ξ − µ0
a
PFA = Pr {T (Z) > ξ|H0 } = Q
,
σ0
(23)
(24)
(25)
(26)
3
where Q(u) denotes the Q-function. Consequently, for a
desired PFA , the decision threshold is
a
ξ = Q−1 (PFA ) σ0 + µ0 ,
resulting in the asymptotic probability of detection
µ1 − µ0
σ0
a
−1
.
−
PD (PFA ) = Q Q (PFA )
σ1
σ1
(27)
(28)
2
arcsin (Σy (θ)) ,
π
−1
−1
Σy (θ) = diag (Ry (θ)) 2 Ry (θ) diag (Ry (θ)) 2 .
Rz (θ) =
Further, the evaluation of (35) requires determining the matrix
h
T i
(43)
C(θ) = Ez;θ vec zz T vec zz T
(29)
IV. R ESULTS
the log-likelihood ratio for exponential family models is
ln L(Z) =
K
X
k=1
T
b φ(z k ) − K(λ(θ 1 ) − λ(θ 0 )),
(30)
such that with the empirical mean of the sufficient statistics
φ̄ =
K
1 X
φ(z k ),
K
(42)
which is possible [21] by the arcsine law and the orthant
probability of the quadrivariate Gaussian distribution [26].
Writing
b = β(θ1 ) − β(θ0 ),
is obtained by the arcsine law [25, pp. 284],
(31)
We apply the results to discuss the design of a 1-bit array
GPS spectrum monitoring system. The task is to check a spectral band with two-sided bandwidth B = 2.046 Mhz, centered
at 1.57 GHz, for a source signal from a specific direction.
The receiver samples the low-pass filtered receive signal at
Nyquist rate, i.e., fs = B, such that K = 2046 array snapshots
are available within a millisecond. The upper triangular area
k=1
a likelihood-based test statistic is
0
T (Z) = b φ̄.
The mean and standard deviation of the test are
r
1 T
T
b Rφ (θi )b,
µi = b µφ (θ i ), σi =
K
(32)
−2
(33)
χ [dB]
T
−4
where
µφ (θ) = Ez;θ [φ(z)] ,
h
i
Rφ (θ) = Ez;θ φ(z)φT (z) − µφ (θ)µT
φ (θ).
(34)
(35)
For Gaussian models (6),
1
β(θ) = − vec R−1
y (θ) ,
2
φ(y) = vec yy T ,
µφ (θ) = vec (Ry (θ)) ,
(36)
(37)
(38)
where the matrix Φ eliminates duplicate and diagonal statistics. This is potentially suboptimal as (39) does, in general,
not contain all sufficient statistics of a multivariate Bernoulli
distribution. The missing statistics are absorbed in the carrier
of (13) and, therefore, do not contribute to the decision
process. For the natural parameter difference (29), we use
(40)
as it maximizes the distance of the asymptotic outcome in (32)
under both hypotheses. The mean of the statistics (39),
µφ (θ) = Ez;θ [φ(z)] = Φ vec (Rz (θ)) ,
−8
5
and the matrix (35) can be determined through Isserlis’ theorem. If the model (13) is unspecified, a favorable choice of
φ(z) and b has to be found. For the 1-bit analysis, we use
(39)
φ(z) = Φ vec zz T ,
−1
b = R−1
φ (θ 1 )µφ (θ 1 ) − Rφ (θ 0 )µφ (θ 0 )
SNR = −15
SNR = −18
SNR = −21
SNR = −24
−6
(41)
10
15
Number of Sensors S
dB
dB
dB
dB
20
Fig. 1. Quality vs. Array Size (K = 2046, ζ = 45◦ )
under the receiver operating characteristic (ROC) curve,
Z 1
χ=2
PD (u)du − 1,
(44)
0
determines the system quality regarding the detection task.
√ For
different signal-to-noise ratios (SNR), γ0 = 0 vs. γ1 = SNR,
Fig. 1 shows the 1-bit system quality for an exemplary setting
with K = 2046, ζ = 45◦ , versus the number of array elements
S. While for very weak sources (SNR = −24 dB) more than
S = 20 sensors are required to provide high performance, at a
power level of SNR = −15 dB already S = 5 antennas suffice
to operate close to a perfect monitoring system with χ = 1. To
determine a favorable observation length, for S = 8, ζ = 30◦ ,
Fig. 2 shows χ for different numbers of samples. While
reliable detection with SNR = −24 dB requires sampling
more than 10 ms, the decision at SNR = −15 dB can be
made trustworthy within less than 1 ms. Fig. 3 depicts the
analytic and simulated performance (using 105 realizations)
with S = 8, K = 100, ζ = 15◦ . At moderate sample size, the
4
R EFERENCES
χ [dB]
0
−2
−4
SNR = −15
SNR = −18
SNR = −21
SNR = −24
−6
2
dB
dB
dB
dB
4
6
8
Observation Time [ms]
10
Fig. 2. Quality vs. Observation Time (S = 8, ζ = 30◦ )
asymptotic results already show a good correspondence with
the Monte-Carlo simulations.
1
0.8
PD
0.6
0.4
0.2
0
PFA = 10−3
PFA = 10−4
−18 −16 −14 −12 −10 −8
Signal-to-Noise Ratio [dB]
−6
Fig. 3. Analysis vs. Simulation (S = 8, K = 100, ζ = 15◦ )
V. C ONCLUSION
We have derived achievable detection rates with a large
number of radio measurements obtained with a lowcomplexity sensor array performing 1-bit analog-to-digital
conversion. Discussing the simple binary hypothesis test in the
framework of the exponential family enables circumventing
the intractability of the 1-bit array likelihood. Its difficult
characterization forms a fundamental obstacle to the application of analytic tools in statistical signal and information
processing to problems with coarsely quantized multivariate
data. Using the analytic results to determine the spectrum
monitoring capability with wireless multi-antenna receivers
shows that, under the right system design, radio frequency
measurements from binary arrays are sufficient to perform the
processing task of signal detection in a fast and reliable way.
[1] H. Krim, M. Viberg, ”Two decades of array signal processing research:
The parametric approach,” IEEE Signal Process. Mag., vol. 13, no. 4,
pp. 67–94, Jul. 1996.
[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] B. Murmann, “ADC performance survey 1997-2017,” [Online]. Available: http://web.stanford.edu/˜murmann/adcsurvey.html
[4] R. R. Tenney, N. R. Sandell, “Detection with distributed sensors,” IEEE
Trans. Aerosp. Electron. Syst., vol. 17, no. 4, pp. 501–510, Jul. 1981.
[5] J. Fang, Y. Liu, H. Li, S. Li, “One-bit quantizer design for multisensor
GLRT fusion,” IEEE Signal Process. Lett., vol. 20, no. 3, pp. 257–260,
Mar. 2013.
[6] D. Ciuonzo, G. Papa, G. Romano, P. Salvo Rossi, P. Willett, “One-bit
decentralized detection with a Rao test for multisensor fusion,” IEEE
Signal Process. Lett., vol. 20, no. 9, pp. 861–864, Sept. 2013.
[7] H. Zayyani, F. Haddadi, M. Korki, “Double detector for sparse signal
detection from one-bit compressed sensing measurements,” IEEE Signal
Process. Lett., vol. 23, no. 11, pp. 1637-1641, Nov. 2016.
[8] S. Kassam, “Optimum quantization for signal detection,” IEEE Trans.
Commun., vol. 25, no. 5, pp. 479–484, May 1977.
[9] H. V. Poor, J. B. Thomas, “Optimum data quantization for a general
signal detection problem,” in Asilomar Conference on Circuits, Systems
and Computers, Pacific Grove, 1977, pp. 299-303.
[10] B. Aazhang, H. Poor, “On optimum and nearly optimum data quantization for signal detection,” IEEE Trans. Commun., vol. 32, no. 7, pp.
745-751, Jul 1984.
[11] W. A. Hashlamoun, P. K. Varshney, “Near-optimum quantization for
signal detection,” IEEE Trans. Commun., vol. 44, no. 3, pp. 294–297,
Mar. 1996.
[12] P. Willett, P. F. Swaszek, “On the performance degradation from onebit quantized detection,” IEEE Trans. Inf. Theory, vol. 41, no. 6, pp.
1997–2003, Nov. 1995.
[13] G. Foschini, R. Gitlin, S. Weinstein, “Optimum detection of quantized
PAM data signals,” IEEE Trans. Commun., vol. 24, no. 12, pp. 1301–
1309, Dec. 1976.
[14] O. Dabeer, U. Madhow, “Detection and interference suppression for
ultra-wideband signaling with analog processing and one bit A/D,” in
Asilomar Conference on Signals, Systems and Computers, Pacific Grove,
2003, pp. 1766–1770.
[15] A. Mezghani, M. S. Khoufi, J. A. Nossek, “Maximum likelihood
detection for quantized MIMO systems,” in International ITG Workshop
on Smart Antennas, Vienna, 2008, pp. 278–284.
[16] S. Wang, Y. Li, 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.
[17] J. Choi, J. Mo, R. W. Heath, “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] S. Jacobsson, G. Durisi, M. Coldrey, U. Gustavsson, C. Studer,
”Throughput analysis of massive MIMO uplink with low-resolution
ADCs,” IEEE Trans. Wireless Commun., vol. 16, no. 6, pp. 4038–4051,
Jun. 2017.
[19] A. Ali, W. Hamouda, “Low power wideband sensing for one-bit quantized cognitive radio systems,” IEEE Wireless Commun. Lett., vol. 5, no.
1, pp. 16-19, Feb. 2016.
[20] O. Bar-Shalom, A. J. Weiss, “DOA estimation using one-bit quantized
measurements,” IEEE Trans. Aerosp. Electron. Syst., vol. 38, no. 3, pp.
868–884, Jul. 2002.
[21] M. Stein, K. Barbé, J. A. Nossek, “DOA parameter estimation with 1bit quantization - Bounds, methods and the exponential replacement”,
in International ITG Workshop on Smart Antennas, Munich, 2016, pp.
1-6.
[22] C. L. Liu, P. P. Vaidyanathan, “One-bit sparse array DOA estimation,”
in IEEE International Conference on Acoustics, Speech and Signal
Processing (ICASSP), New Orleans, 2017, pp. 3126–3130.
[23] P. M. Aziz, H. V. Sorensen, J. van der Spiegel, “An overview of sigmadelta converters,” IEEE Signal Process. Mag., vol. 13, no. 1, pp. 61–84,
Jan. 1996.
[24] S. M. Kay, Fundamentals of Statistical Signal Processing, Volume 2:
Detection Theory. Upper Saddle River, NJ: Prentice-Hall, 1998.
[25] J. B. Thomas, Introduction to Statistical Communication Theory. Hoboken, NJ: John Wiley & Sons, 1969.
[26] M. Sinn, K. Keller, “Covariances of zero crossings in Gaussian processes,” Theory Probab. Appl., vol. 55, no. 3, pp. 485–504, 2011.
| 7cs.IT
|
Please cite this article as:
Cuevas, E., Sención-Echauri, F., Zaldivar, D., Pérez-Cisneros, M. Multi-circle detection on images using artificial bee
colony (ABC) optimization, Soft Computing, 16 (2), (2012), pp. 281-296.
Multi-Circle Detection on Images Using Artificial
Bee Colony (ABC) Optimization
Erik Cuevas, Felipe Sención-Echauri, Daniel Zaldivar and Marco Pérez-Cisneros
Departamento de Ciencias Computacionales
Universidad de Guadalajara, CUCEI
Av. Revolución 1500, Guadalajara, Jal, México
{erik.cuevas, felipe.sencion, 1daniel.zaldivar, marco.perez}@cucei.udg.mx
Abstract
Hough transform (HT) has been the most common method for circle detection, exhibiting robustness, but adversely
demanding considerable computational effort and large memory requirements. Alternative approaches include
heuristic methods that employ iterative optimization procedures for detecting multiple circles. Since only one circle
can be marked at each optimization cycle, multiple executions must be enforced in order to achieve multi-detection.
This paper presents an algorithm for automatic detection of multiple circular shapes that considers the overall process
as a multi-modal optimization problem. The approach is based on the artificial bee colony (ABC) algorithm, a swarm
optimization algorithm inspired by the intelligent foraging behavior of honey bees. Unlike the original ABC
algorithm, the proposed approach presents the addition of a memory for discarded solutions. Such memory allows
holding important information regarding other local optima which might have emerged during the optimization
process. The detector uses a combination of three non-collinear edge points as parameters to determine circle
candidates. A matching function (nectar- amount) determines if such circle candidates (bee-food-sources) are actually
present in the image. Guided by the values of such matching functions, the set of encoded candidate circles are
evolved through the ABC algorithm so that the best candidate (global optimum) can be fitted into an actual circle
within the edge-only image. Then, an analysis of the incorporated memory is executed in order to identify potential
local optima, i.e., other circles. The proposed method is able to detect single or multiple circles from a digital image
through only one optimization pass. Simulation results over several synthetic and natural images, with a varying
range of complexity, validate the efficiency of the proposed technique regarding its accuracy, speed, and robustness.
Keywords: Circle detection; artificial bee colony; nature-inspired algorithms; intelligent image processing.
1. Introduction
The problem of detecting circular features holds paramount importance for image analysis in industrial
applications such as automatic inspection of manufactured products and components, aided vectorization
of drawings, target detection, etc. [1]. Solving common challenges for object localization is normally
approached by two techniques: deterministic and stochastic. The former includes the application of
Hough transform-based methods [2], the use of geometric hashing and other template or model-based
matching techniques [3, 4]. On the other hand, stochastic techniques include random sample consensus
techniques [5], simulated annealing [6] and genetic algorithms (GA) [7].
Template and model matching techniques were the first approaches to be applied to shape detection.
Although several methods have now been developed for solving such a problem [8], shape coding
techniques and a combination of shape properties have been successfully tested on representing different
objects. Their main drawbacks are related to the contour extraction step from real images and to their
deficiencies in dealing with pose invariance except for very simple objects.
The circle detection in digital images is commonly solved through the Circular Hough Transform [9]. A
typical Hough-based approach employs an edge detector and some edge information to infer locations
and radii values. Peak detection is then performed by averaging, filtering and histogramming within the
transform space. Unfortunately, such an approach requires a large storage space as the 3-D cells include
parameters (x, y, r) that augment the computational complexity and yield a low processing speed. The
accuracy of parameters for the extracted circles is poor, particularly under noisy conditions [10].
In the particular case of a digital image holding a significant width and height, and some densely
populated edge pixels, the required processing time for Circular Hough Transform makes it prohibitive to
1
Please cite this article as:
Cuevas, E., Sención-Echauri, F., Zaldivar, D., Pérez-Cisneros, M. Multi-circle detection on images using artificial bee
colony (ABC) optimization, Soft Computing, 16 (2), (2012), pp. 281-296.
be deployed in real time applications. In order to overcome such a problem, some other researchers have
proposed new approaches following Hough transform principles, yielding the probabilistic Hough
transform [11], the randomized Hough transform (RHT) [12], the fuzzy Hough transform [13] and some
other topics as is widely discussed by Becker in [14].
As an alternative to Hough Transform-based techniques, the problem of shape recognition in computer
vision has also been handled through optimization methods. Ayala–Ramirez et al. presented a GA based
circle detector [15] that is capable of detecting multiple circles over real images, but fails frequently while
detecting imperfect circles. On the other hand, Dasgupta et al. [16] have recently proposed an automatic
circle detector using the bacterial foraging optimization algorithm (BFOA) as optimization procedure.
However, both methods employ an iterative scheme to achieve multiple-circle detection, which executes
the algorithm as many times as the number of circles to be found demands. Only one circle can be found
at each run yielding quite a long execution time.
An impressive growth in the field of biologically inspired meta-heuristics for search and optimization has
emerged during the last decade. Some bio-inspired examples like genetic algorithm (GA) [17] and
differential evolution (DE) [18] have been applied to solve complex optimization problems, while swarm
intelligence (SI) has recently attracted interest from several fields. The SI core lies in the analysis of the
collective behavior of relatively simple agents working on decentralized systems. Such systems typically
gather an agent’s population that can communicate with each other while sharing a common environment.
Despite a non-centralized control algorithm regulating its behavior, the agent can solve complex tasks by
analyzing a given global model and harvesting cooperation to other agents. Therefore, a novel global
behavior evolves from interaction among agents as can be seen in case of ant colonies, animal herding,
bird flocking, fish schooling, honey bees, bacteria, and many more. Swarm-based algorithms, such as
particle swarm optimization [19], ant colony optimization [20] and bacterial foraging optimization
algorithm (BFOA) [21] have already been successfully applied to several engineering applications.
Karaboga has recently presented one bee-swarm algorithm for solving numerical optimization problems,
which is known as the artificial bee colony (ABC) method [22]. Inspired by the intelligent foraging
behavior of a honeybee swarm, the ABC algorithm consists of three essential components: food source
positions, nectar- amount and different honey bee classes. Each food source position represents a feasible
solution for the problem under consideration, and the nectar-amount of a food source represents the
quality of such solution corresponding to its fitness value. Each class of bees symbolizes one particular
operation for generating new candidate food source positions (candidate solutions).
The ABC algorithm starts by producing a randomly distributed initial population (food source locations).
After initialization, an objective function evaluates whether such candidates represent an acceptable
solution (nectar-amount) or not. Guided by the values of such an objective function, the candidate
solutions are evolved through different ABC operations (honey bee types). When the fitness function
(nectar-amount) cannot be further improved after a maximum number of cycles, its related food source is
assumed to be abandoned and replaced by a new randomly chosen food source location. However, in
order to contribute towards the solution of multi-modal optimization problems, our proposal suggests that
such abandoned solutions are not to be discarded; instead, they are to be arranged into a so-called
“exhausted-source memory” that contains valuable information regarding global and local optima that
have been emerging with the evolution of optimization.
Although ABC draws several similarities with other bio-inspired algorithms, there are some significant
issues to be discussed: ABC does not depend upon the best member within the population in order to
update the particle’s motion as is done by PSO [23]; it does not require all particles for computing
parameters such as the pheromone concentration, which determines the overall performance, as is
demanded by ACO [24]. In contrast, ABC uses randomly chosen particles to calculate new motion
vectors, contributing towards augmenting the population diversity. Similar to DE, ABC does require a
selection operation that allows individuals to access a fair chance of being elected for recombination
(diversity). However, ABC holds a second modification operation that follows a random “roulette
selection”, allowing some privileges for best located solutions and augmenting the convergence speed
[25]. In contrast to the local particle modifications executed by BFOA, ABC employs operators that
tolerate modifications over the full search space for each parameter, avoiding typical oscillations around
the optimum produced by BFOA [26]. The performance of ABC algorithm has been compared with other
optimization methods such as GA, DE and PSO [27,28]. The results showed that ABC can produce
optimal solutions and thus is more effective than other methods in several optimization problems. Such
2
Please cite this article as:
Cuevas, E., Sención-Echauri, F., Zaldivar, D., Pérez-Cisneros, M. Multi-circle detection on images using artificial bee
colony (ABC) optimization, Soft Computing, 16 (2), (2012), pp. 281-296.
characteristics have motivated the use of ABC to solve different sorts of engineering problems such as
signal processing [29,30], flow shop scheduling [31], structural inverse analysis [32], clustering [33,34],
vehicle path planning [35] and electromagnetism [36].
This paper presents an algorithm for the automatic detection of multiple circular shapes from complicated
and noisy images, which does not take into consideration the conventional-Hough transform principles.
The detection process is approached as a multi-modal optimization problem. The ABC algorithm searches
the entire edge-map looking for circular shapes by considering a combination of three non-collinear edge
points that represent candidate circles (food source locations) in the edge-only image of the scene. An
objective function is used to measure the existence of a candidate circle over the edge-map. Guided by the
values of such an objective function, the set of encoded candidate circles are evolved through the ABC
algorithm so that the best candidate can be fitted into the most circular shape within the edge-only image.
A subsequent analysis of the incorporated exhausted-source memory is then executed in order to identify
potential useful local optima (other circles). The approach generates a fast sub-pixel detector that can
effectively identify multiple circles in real images despite circular objects exhibiting significant occluded
sections. Experimental evidence shows the effectiveness of the method for detecting circles under various
conditions. A comparison with one state-of-the-art GA-based method [15], the BFOA [16] and the RHT
algorithm [12] on different images has been included to demonstrate the performance of the proposed
approach. Conclusions of the experimental comparison are validated through statistical tests that support
the discussion suitably.
The paper is organized as follows: Section 2 provides information regarding the ABC algorithm. Section
3 depicts the implementation of the proposed circle detector. The complete multiple-circle detection
procedure is presented in Section 4. Experimental outcomes after applying the proposed approach are
stated in Section 5 and some relevant conclusions are discussed in Section 6.
2. Artificial Bee Colony (ABC) algorithm
The ABC algorithm assumes the existence of a set of operations that may resemble some features of the
honey bee behavior. For instance, each solution within the search space includes a parameter set
representing food source locations. The “fitness value” refers to the food source quality that is strongly
linked to the food’s location. The process mimics the bee’s search for valuable food sources yielding an
analogous process for finding the optimal solution.
2.1 Biological bee profile
The minimal model for a honey bee colony consists of three classes: employed bees, onlooker bees and
scout bees. The employed bees will be responsible for investigating the food sources and sharing the
information with recruit onlooker bees. They, in turn, will make a decision on choosing food sources by
considering such information. The food source having a higher quality will have a larger chance to be
selected by onlooker bees than those showing a lower quality. An employed bee, whose food source is
rejected as low quality by employed and onlooker bees, will change to a scout bee to randomly search for
new food sources. Therefore, the exploitation is driven by employed and onlooker bees while the
exploration is maintained by scout bees. The implementation details of such bee-like operations in the
ABC algorithm are described in the next sub-section.
2.2 Description of the ABC algorithm
Resembling other swarm based approaches, the ABC algorithm is an iterative process. It starts with a
population of randomly generated solutions or food sources. The following three operations are applied
until a termination criterion is met [28]:
1. Send the employed bees.
2. Select the food sources by the onlooker bees.
3. Determine the scout bees.
2.2.1 Initializing the population
3
Please cite this article as:
Cuevas, E., Sención-Echauri, F., Zaldivar, D., Pérez-Cisneros, M. Multi-circle detection on images using artificial bee
colony (ABC) optimization, Soft Computing, 16 (2), (2012), pp. 281-296.
The algorithm begins by initializing N p food sources. Each food source is a D-dimensional vector
containing the parameter values to be optimized, which are randomly and uniformly distributed between
and the upper initial parameter bound x high
.
the pre-specified lower initial parameter bound x low
j
j
x j ,i = xlow
+ rand(0,1) ⋅ ( x high
− x low
j
j
j );
j = 1, 2,K , D; i = 1, 2,K , N p .
(1)
with j and i being the parameter and individual indexes respectively. Hence, x j ,i is the jth parameter of the
ith individual.
2.2.2 Send employed bees
The number of employed bees is equal to the number of food sources. At this stage, each employed bee
generates a new food source in the neighborhood of its present position as follows:
v j ,i = x j ,i + φ j ,i ( x j ,i − x j , k );
k ∈ {1, 2,K , N p } ; j ∈ {1, 2,K , D}
(2)
x j ,i is a randomly chosen j parameter of the ith individual and k is one of the N p food sources, satisfying
the condition i ≠ k . If a given parameter of the candidate solution vi exceeds its predetermined
boundaries, that parameter should be adjusted in order to fit the appropriate range. The scale factor φ j ,i is
a random number between [−1,1] . Once a new solution is generated, a fitness value representing the
profitability associated with a particular solution is calculated. The fitness value for a minimization
problem can be assigned to each solution vi by the following expression:
1
fiti = 1 + J i
1 + abs ( J )
i
if J i ≥ 0
(3)
if J i < 0
where J i is the objective function to be minimized. A greedy selection process is thus applied between
vi and xi . If the nectar- amount (fitness) of vi is better, then the solution xi is replaced by vi ; otherwise,
xi remains.
2.2.3 Select the food sources by the onlooker bees
Each onlooker bee (the number of onlooker bees corresponds to the food source number) selects one of
the proposed food sources, depending on their fitness value, which has been recently defined by the
employed bees. The probability that a food source will be selected can be obtained from the following
equation:
Probi =
fiti
(4)
Np
∑ fit
i
i =1
where fiti is the fitness value of the food source i, which is related to the objective function value ( J i )
corresponding to the food source i. The probability of a food source being selected by onlooker bees
increases with an increase in the fitness value of the food source. After the food source is selected,
onlooker bees will go to the selected food source and select a new candidate food source position inside
the neighborhood of the selected food source. The new candidate food source can be expressed and
calculated by (2). In case the nectar-amount, i.e., fitness of the new solution, is better than before, such
position is held; otherwise, the last solution remains.
4
Please cite this article as:
Cuevas, E., Sención-Echauri, F., Zaldivar, D., Pérez-Cisneros, M. Multi-circle detection on images using artificial bee
colony (ABC) optimization, Soft Computing, 16 (2), (2012), pp. 281-296.
2.2.4 Determine the scout bees
If a food source i (candidate solution) cannot be further improved through a predetermined trial number
known as “limit”, the food source is assumed to be abandoned and the corresponding employed or
onlooker bee becomes a scout. A scout bee explores the searching space with no previous information,
i.e., the new solution is generated randomly as indicated by (1). In order to verify if a candidate solution
has reached the predetermined “limit”, a counter Ai is assigned to each food source i. Such a counter is
incremented consequent to a bee-operation failing to improve the food source’s fitness.
2.3 Exhausted-source memory
Though the classic ABC algorithm eliminates the abandoned food sources, our approach considers the
exhausted food sources (solutions) through avoiding their discharge by saving them into the exhaustedsource memory. Such recorded solutions contain valuable information regarding global and local optima
that emerged during the optimization process.
3. Circle detection using ABC
3.1 Data preprocessing
The ABC circle detector involves a pre-processing stage that requires marking the object’s contour by
applying a single-pixel edge detection method. For our purpose, such a task is accomplished by the
classical Canny algorithm. Then, the locations of the found edge pixels are stored within the
{
}
vector P = p1 , p2 ,K , pE p , E p being the total number of edge pixels in the image.
3.2 Individual representation
In order to construct each candidate circle C (or food-source within the ABC framework), indexes i1 , i2
and i3 representing three edge points previously stored in vector P must be combined. Therefore, each
food-source is encoded as one circle
C = { pi1 , pi2 , pi3 } , which is characterized by three
points pi1 , pi2 and pi3 that lie on its own circumference. Such candidate circle is labeled as a potential
solution for the detection problem. Considering the configuration of the edge points in Fig. 1, the circle
centre ( x0 , y0 ) and the radius r of C can be calculated using simple geometric equations [37].
pi1
pi2
r
( x0 , y0 )
C
pi3
Fig. 1. Circle candidate (individual) built from the combination of points pi1 , pi2 and pi3 .
3.3 Objective function
5
Please cite this article as:
Cuevas, E., Sención-Echauri, F., Zaldivar, D., Pérez-Cisneros, M. Multi-circle detection on images using artificial bee
colony (ABC) optimization, Soft Computing, 16 (2), (2012), pp. 281-296.
In order to calculate the matching error produced by a candidate circle C, its circumference coordinates
are calculated as a virtual shape that must be validated, i.e., confirm whether C really exists in the edgemap. Such circumference coordinates are grouped within the test set S = {s1 , s2 ,K , sN s } , with
N s representing the number of points over which the existence of an edge point, corresponding to C ,
should be verified.
In this approach, the set S is generated by the midpoint circle algorithm (MCA) [38]. The MCA, which is
considered to be the quickest method providing a sub-pixel precision [39,40], calculates the required
points for digitally drawing a circle. Considering that the function f Circle ( x, y ) = x 2 + y 2 − r 2 defines a
circular primitive, MCA introduces an error e as a measurement for the deviation of the halfway pixel
position (sub-pixel distance) characterized by e = fCircle ( x, y ) , with e being zero for locations lying on the
circumference, positive for those outside and negative for those occupying the interior. The minimum
error (i.e., the error shown by the pixel lying closer to the ideal circumference) is used to decide which
pixel should be set next as the best circle boundary approximation. On the other hand, the computation
time of MCA is reduced by considering the symmetry among circles. Circle sections lying at adjacent
octants within one quadrant, are symmetric with respect to the 45° line dividing two octants. Taking
advantage of such symmetry property, MCA generates all pixel positions around a circle by calculating
only the first octant. Therefore, other octants are inferred from the first one by using simple symmetry
relationships. For more details, see [41].
The objective function J(C) represents the matching error produced between the pixels S (calculated by
MCA) of the circle candidate C and the pixels that actually exist in the edge image, yielding:
Ns
∑ E (s )
(5)
v
J (C ) = 1 −
v =1
Ns
where E ( sv ) is a function that verifies the pixel existence in the sv location ( sv = ( xv , yv )) , with sv ∈ S
and N s being the number of pixels lying on the perimeter corresponding to C currently under testing.
Hence, the function E ( sv ) is defined as:
1 if the pixel ( xv , yv ) is an edge point
E ( sv ) =
otherwise
0
Pi 2
(6)
Pi 2
C
Pi1
Pi1
Pi 3
(a)
(b)
Pi 3
(c)
Fig. 2. Procedure to evaluate the objective function J(C): (a) the original edge-map, (b) the virtual shape generated by
MCA considering C = { pi1 , pi2 , pi3 } . (c) shows the comparison operation between the original edge-map shown in (a)
and the virtual shape presented by (b).
A value near to zero of J(C) implies a better response from the “circularity” operator and represents a
better nectar- amount within the ABC framework. Fig. 2 shows the procedure to evaluate the objective
function J(C). First, three edge points (as they are exhibited by Fig. 2a) encode a candidate circle C.
Then, by using the MCA, a circular virtual shape is built (see Fig. 2b). Such virtual shape, which is
6
Please cite this article as:
Cuevas, E., Sención-Echauri, F., Zaldivar, D., Pérez-Cisneros, M. Multi-circle detection on images using artificial bee
colony (ABC) optimization, Soft Computing, 16 (2), (2012), pp. 281-296.
characterized by the vector S, groups a determined number N s of pixel coordinates that are shown by Fig.
2b as 56. Finally, the virtual shape is compared with the original edge-map, point by point, in order to
find matches between virtual and edge pixels. Fig. 2c shows the comparison operation between the
original edge-map (Fig. 2a) and the virtual shape (Fig. 2b). As a result, only eighteen edge pixels are
Ns
common to both images yielding:
∑ E ( x , y ) = 18 and therefore, J (C ) ≈ 0.67 .
v
v
v =1
3.3. ABC Implementation
The implementation of the proposed algorithm can be summarized by the following steps:
Step 1:
Apply the Canny filter to the original image and store edge pixels within vector P.
Step 2:
Initialize required parameters of the ABC algorithm. Set the colony’s size, the abandonment
limit and the maximum number of cycles.
Step 3:
Initialize N C circle candidates Cb (original food sources) with b ∈ (1,K , N C ) , and clear all
counters Ab .
Step 4:
Step 5:
Step 6:
Step 7
Step 8:
Obtain the matching fitness (food source quality) for each circle candidate Cb using (5).
Repeat steps 6 to 10 until a termination criterion is met.
Modify the circle candidates as stated by (2) and evaluate its matching fitness (send employed
bees onto food sources). Likewise, update all counters Ab .
Calculate the probability value Probb for each circle candidate Cb . Such probability value
will be used as a preference index by onlooker bees (4).
Generate new circles candidates (using the (2)) from current candidates according to their
probability Probb (Send onlooker bees to their selected food source). Likewise, update
counters Ab .
Step 9:
Step 10:
Step 11:
Obtain the matching fitness for each circle candidate Cb and calculate the best circle
candidate (solution).
Stop modifying the circle candidate Cb (food source) whose counter Ab has reached its
counter “limit” and save it as a possible solution (global or local optimum) in the exhaustedsource memory. Clear Ab and generate a new circle candidate according to (1).
Analyze solutions previously stored in the exhausted-source memory (see Section 5). The
memory holds solutions (any other potential circular shape in the image) generated through
the evolution of the optimization algorithm.
In ABC algorithm, the steps 6 to 10 are repeated until a termination criterion is met. Typically, two stop
criteria have been employed for meta-heuristic algorithms: either an upper limit of the fitness function is
reached or an upper limit of the number of generations is attained [42]. The first criterion requires an
extensive knowledge of the problem and its solutions [43]. On the contrary, by considering the stop
criterion based on the number of generations, feasible solutions may be found by exploring the search
space through several iterations. For our purpose, the number of iterations as stop criterion is employed in
order to allow the multi-circle detection. Hence, if a solution representing a valid circle appears at early
stages, it would be stored in the exhausted-source memory and the algorithm continues detecting other
feasible solutions until depleting the iteration number. Therefore, the main issue is to define a fair
iteration number, which should be big enough to allow finding circles at the image and small enough to
avoid an exaggerated computational cost. For this study, such a number was experimentally defined as
300.
4. The multiple-circle detection procedure
The original ABC algorithm considers the so-called abandonment limit, which aims to stop the local
exploration for a candidate solution after a trial number is reached. All “stuck solutions”, i.e., those that
do not improve further during the optimization cycle are supposed to be discarded and replaced by other
randomly generated solutions. However, this paper proposes the use of an “exhausted-source memory” to
7
Please cite this article as:
Cuevas, E., Sención-Echauri, F., Zaldivar, D., Pérez-Cisneros, M. Multi-circle detection on images using artificial bee
colony (ABC) optimization, Soft Computing, 16 (2), (2012), pp. 281-296.
store information regarding local optima that represent possible solutions for the multi-circle detection
problem.
Several heuristic methods have been employed for detecting multiple circles as an alternative to classical
Hough transform-based techniques [15, 16]. Such strategies imply that only one circle can be marked per
optimization cycle, forcing a multiple execution of the algorithm in order to achieve multiple-circle
detection. The surface representing J(C) holds a multimodal nature, which contains several global and
local optima that are related to potential circular shapes in the edge-map. This paper aims to solve the
objective function J(C) using only one optimization procedure by assuming the multi-detection problem
as a multimodal optimization issue.
The multi-detection problem can be summarized as follows: guided by the values of a matching function,
the set of encoded circle candidates are evolved through the ABC algorithm and the best circle candidate
(global optimum) is considered to be the first detected circle over the edge-only image. Then, an analysis
of the incorporated exhausted-source memory is executed in order to identify other local optima (other
circles). The analysis includes two operations: arraignment and extraction. In the arraignment, food
sources that are held by the memory are organized in descending order according to their J(C). Once the
exhausted-source memory has been arranged, the goal is to extract circles considered to be different (local
optima) from it. Such discrimination is accomplished by comparing all elements in the arranged memory.
Several local optima (i.e., circles slightly shifted or holding small deviations) can represent the same
circle. Therefore, a distinctiveness factor Esdi is required to measure the mismatch between two given
circles (food-sources) as follows:
Esdi =
( xA − xB ) + ( y A − yB ) + ( rA − rB )
2
2
2
(7)
where ( x A , y A ) and rA are the coordinates of the centre and radius of the circle C A respectively, while
( xB , yB ) and
rB are the corresponding parameters of the circle CB . In order to decide whether two circles
must be considered different, a threshold value EsTH is defined as follows:
ESTH = α
( cols − 1)
2
+ ( rows − 1) + ( rmax − rmin )
2
2
(8)
where rows and cols refer to the number of rows and columns in the image respectively. rmax and rmin are
the maximum and minimum radii for representing feasible candidate circles, while α is a sensitivity factor
affecting the discrimination between circles. A high value of α allows circles to be significantly different
and still be considered as the same shape while a low value would imply that two circles with slight
differences in radii or positions could be considered as different instances. The EsTH value calculated by
(8) allows discriminating circles with no consideration about the image size.
In order to find “sufficiently different” circles, the elements of the arranged exhausted-source memory
must be contrasted. Each element is compared with the others by using (7). A circle is considered
different enough if its distinctiveness factor Esdi (found in the comparison) surpasses the threshold EsTH .
The multiple-circle detection procedure can thus be described as follows:
Step 1:
The best solution found by the ABC algorithm and all candidate circles held by the
exhausted-source memory are organized in a decreasing order of their matching fitness,
yielding a new vector M C = {C1 ,K , CNe } , with Ne being the size of the exhausted-source
Step 2:
memory plus one.
The candidate circle C1 showing the highest matching fitness is identified as the first
circle shape CS1 as it is stored within a vector Ac of actual circles.
Step 3:
The distinctiveness factor Esdi for the candidate circle Cm (element m in M C ) is
8
Please cite this article as:
Cuevas, E., Sención-Echauri, F., Zaldivar, D., Pérez-Cisneros, M. Multi-circle detection on images using artificial bee
colony (ABC) optimization, Soft Computing, 16 (2), (2012), pp. 281-296.
compared with every element in Ac . If Esdi > ESTH is true for each pair of solutions
(those present in Ac and in the candidate circle Cm ), then Cm is considered as a new
circle shape CS and is added to the vector Ac . Otherwise, the next circle candidate Cm +1
is evaluated and Cm is discarded.
Step 4:
Step 3 is repeated until all Ne candidate circles in M C have been analyzed.
Summarizing the overall procedure, Fig. 3 shows the outcome of the ABC-based circular detector. The
input image (Fig. 3a) has a resolution of 256 x 256 pixels and shows two circles and two ellipses with a
different circularity factor. Fig. 3b presents the detected circles with a red overlay. Fig. 3c shows the
candidate circles held by the exhausted-source memory after the optimization process. Fig. 3d presents
the resulting image after the previously described discrimination procedure has been applied.
(b)
50
50
45
45
Radius
Radius
(a)
40
40
35
35
30
200
30
200
300
150
y Coordinate
0
200
100
100
50
300
150
200
100
y Coordinate
x Coordinate
100
50
0
x Coordinate
(c)
(d)
Fig. 3. ABC-based circular detector performance: (a) the original image and (b) all detected circles as an overlay (c)
Candidate circles held by the exhausted-source memory after optimization process and (d) remaining circles after the
discrimination process.
5. Experimental Results
Experimental tests have been developed in order to evaluate the performance of the circle detector. The
experiments address the following tasks:
(1) Circle localization,
(2) Shape discrimination,
(3) Circular approximation: occluded circles and arc detection.
Table 1 presents the parameters for the ABC algorithm in this study. They have been retained for all test
images after being experimentally defined.
Colony size
Abandonment limit
Number of cycles
20
100
300
Table 1. ABC detector parameters
9
α
0.05
limit
30
Please cite this article as:
Cuevas, E., Sención-Echauri, F., Zaldivar, D., Pérez-Cisneros, M. Multi-circle detection on images using artificial bee
colony (ABC) optimization, Soft Computing, 16 (2), (2012), pp. 281-296.
All the experiments have been executed over a Pentium IV 2.5 GHz computer under C language
programming. All the images are pre-processed by the standard Canny edge detector using the imageprocessing toolbox for MATLAB R2008a.
(a)
(b)
(c)
(d)
Fig. 4. Circle localization over synthetic images. The image (a) shows the original image while (b) presents the
detected circles as an overlay. The image in (c) shows a second image with salt & pepper noise and (d) shows
detected circles as a red overlay.
5.1 Circle localization
5.1.1. Synthetic images
The experimental setup includes the use of several synthetic images of 320x240 pixels. All images
contain varying amounts of circular shapes and some have also been contaminated by added noise so as to
increase the complexity of the localization task. The algorithm is executed over 100 times for each test
image, successfully identifying and marking all circles in the image. The detection has proved to be
robust to translation and scaling, requiring less than 1s. Fig. 4 shows the outcome after applying the
algorithm over two images taken from the experimental set.
5.1.2. Natural images
This experiment tests the circle detection on real-life images. Twenty five test images of 640x480 pixels
have been captured using a digital camera under an 8-bit color format. Each natural scene includes
circular shapes that have been pre-processed through the Canny edge detection algorithm before being fed
to the ABC procedure. Fig. 5 shows a multiple-circle detection over a natural image.
10
Please cite this article as:
Cuevas, E., Sención-Echauri, F., Zaldivar, D., Pérez-Cisneros, M. Multi-circle detection on images using artificial bee
colony (ABC) optimization, Soft Computing, 16 (2), (2012), pp. 281-296.
(a)
(b)
65
Radius
Radius
65
60
55
250
60
55
250
200
250
200
150
y Coordinate
100
200
250
200
150
150
y Coordinate
x Coordinate
(c)
150
100
x Coordinate
(d)
Fig. 5. Circle detection algorithm over natural images: (a) the original image (b) the detected circles as a red overlay
(c) candidate circles lying at the exhausted-source memory after the optimization and (d) detected circles after
finishing the discrimination process.
(a)
(b)
Fig. 6. Shape discrimination over synthetic images: (a) the original image contaminated by salt & pepper noise (b)
detected circles as an overlay.
5.2. Shape discrimination tests
This section discusses the detector’s ability to differentiate circular patterns over any other shape, which
might be present in the image. Fig. 6 shows five different synthetic shapes within an image of 540x300
pixels that has been contaminated by salt & pepper noise. Fig. 7 repeats the experiment over real-life
images.
11
Please cite this article as:
Cuevas, E., Sención-Echauri, F., Zaldivar, D., Pérez-Cisneros, M. Multi-circle detection on images using artificial bee
colony (ABC) optimization, Soft Computing, 16 (2), (2012), pp. 281-296.
(a)
(b)
Fig. 7. Shape discrimination in real-life images: (a) original image and (b) the detected circle as an overlay.
5.3 Circular approximation: occluded circles and arc detection.
The ABC detector algorithm is able to detect occluded or imperfect circles as well as partially defined
shapes such as arc segments. The relevance of such functionality comes from the fact that imperfect
circles are commonly found in typical computer vision applications. Since circle detection has been
considered as an optimization problem, the ABC algorithm allows finding circles that may approach a
given shape according to fitness values for each candidate. Fig. 8a shows some examples of circular
approximation. Likewise, the proposed algorithm is able to find circle parameters that better approach an
arc or an occluded circle. Fig. 8b and 8c show some examples of this functionality. A small value for
J(C), i.e., near zero, refers to a circle while a slightly bigger value accounts for an arc or an occluded
circular shape. Such a fact does not represent any trouble as circles can be shown following the obtained
J(C) values.
5.4 Performance evaluation
In order to enhance the algorithm analysis, the ABC proposed algorithm is compared with the BFAOA
and the GA circle detectors over a set of common images.
The GA algorithm follows the proposal of Ayala-Ramirez et al., which considers the population size as
70, the crossover probability as 0.55, the mutation probability as 0.10 and the number of elite individuals
as 2. The roulette wheel selection and the 1-point crossover operator are both applied. The parameter
setup and the fitness function follow the configuration suggested in [15]. The BFAOA algorithm follows
the implementation from [16] considering the experimental parameters as: S=50, N c = 100 ,
N s = 4, N ed = 1, Ped = 0.25 , d attract = 0.1 , wattract = 0.2 , wrepellant = 10 hrepellant = 0.1 , λ = 400 and ψ = 6 .
Such values are found to represent the best configuration set according to [16].
(a)
12
Please cite this article as:
Cuevas, E., Sención-Echauri, F., Zaldivar, D., Pérez-Cisneros, M. Multi-circle detection on images using artificial bee
colony (ABC) optimization, Soft Computing, 16 (2), (2012), pp. 281-296.
(b)
(c)
Fig. 8. ABC approximating circular shapes and arc sections.
Images rarely contain perfectly-shaped circles. Therefore, aiming for a test on accuracy for a single-circle,
the detection is challenged by a ground-truth circle, which is determined manually from the original edgemap. The parameters ( xtrue , ytrue , rtrue ) of the reference circle are computed considering the best matching
circle that has been previously defined over the edge-map. If the centre and the radius of the detected
circle are defined as ( xD , yD ) and rD , then an error score (Es) can be chosen as follows:
Es=η ⋅ ( xtrue − xD + ytrue − yD ) + µ ⋅ rtrue − rD
The central point difference
(x
true
− xD + ytrue − yD
(9)
) represents the centre shift for the detected circle as it
is compared with a benchmark circle. The radio mismatch ( rtrue − rD ) accounts for the difference between
their radii. η and µ represent two weighting parameters, which are to be applied separately to the central
point difference and to the radius mismatch for the final error Es. In this study, they are chosen as
η = 0.05 and µ = 0.1 . This particular choice ensures that the radii difference would be strongly weighted
in comparison to the difference in the central circular positions of manually detected and machinedetected circles. In order to use an error metric for multiple-circle detection, the averaged Es produced
from each circle in the image is considered. Such criterion, defined as the multiple error (ME), is
calculated as follows:
1 NC
ME=
⋅ ∑ Es R
NC R =1
(10)
where NC represents the number of circles actually present the image. In case the ME is less than 1, the
algorithm is considered successful; otherwise it is said to have failed in the detection of the circle set.
Notice that for η = 0.05 and µ = 0.1 , it yields ME<1, which accounts for a maximal tolerated average
difference on radius length of 10 pixels, whereas the maximum average mismatch for the centre location
can be up to 20 pixels. In general, the success rate (SR) can thus be defined as the percentage of achieving
success after a certain number of trials.
13
Please cite this article as:
Cuevas, E., Sención-Echauri, F., Zaldivar, D., Pérez-Cisneros, M. Multi-circle detection on images using artificial bee
colony (ABC) optimization, Soft Computing, 16 (2), (2012), pp. 281-296.
Original image
Best case
(a)
(b)
(c)
(d)
(e)
(f)
14
Worst case
Please cite this article as:
Cuevas, E., Sención-Echauri, F., Zaldivar, D., Pérez-Cisneros, M. Multi-circle detection on images using artificial bee
colony (ABC) optimization, Soft Computing, 16 (2), (2012), pp. 281-296.
Fig. 9. Test images and their detected circles using the ABC detector. The results show the best and the worst case
obtained throughout 35 runs.
Fig. 9 shows six images that have been used to compare the performance of the GA-based algorithm [15],
the BFOA method [16] and the proposed approach. The performance is analyzed by considering 35
different executions for each algorithm over six images. Table 2 presents the averaged execution time, the
success rate (SR) in percentage and the averaged multiple error (ME). The best entries are bold-cased in
Table 2. Closer inspection reveals that the proposed method is able to achieve the highest success rate
with the smallest error, and still requires less computational time for most cases. Fig. 9 also exhibits the
resulting images after applying the ABC detector. Such results present the best and the worst cases
obtained throughout 35 runs.
A non-parametric statistical significance proof called Wilcoxon’s rank sum test for independent samples
[44-46] has been conducted at 5% significance level on the multiple error (ME) data of Table 2. Table 3
reports the p-values produced by Wilcoxon’s test for the pair-wise comparison of multiple error (ME)
between two groups. One group corresponds to ABC versus GA and the other corresponds to ABC versus
BFOA, one at a time. As a null hypothesis, it is assumed that there is no significant difference between
the mean values of the two groups. The alternative hypothesis considers a significant difference between
the mean values of both groups. All p-values reported in the table are less than 0.05 (5% significance
level), which is a strong evidence against the null hypothesis, indicating that the best ABC mean values
for the performance are statistically significant and have not occurred by chance.
Image
(a)
(b)
(c)
(d)
(e)
(f)
Averaged execution time ± Standard
deviation (s)
Success rate (SR) (%)
Averaged ME ± Standard deviation
GA
BFOA
ABC
GA
BFOA
ABC
GA
BFOA
ABC
2.23±(0.41)
3.15±(0.39)
4.21±(0.11)
5.11±(0.43)
6.33±(0.34)
7.62±(0.97)
1.71±(0.51)
2.80±(0.65)
3.18±(0.36)
3.45±(0.52)
4.11±(0.14)
5.36±(0.17)
0.21±(0.22)
0.36±(0.24)
0.20±(0.19)
1.10±(0.24)
1.61±(0.17)
1.95±(0.41)
94
81
79
93
87
88
100
95
91
100
94
90
100
98
100
100
100
98
0.41±(0.044)
0.51±(0.038)
0.48±(0.029)
0.45±(0.051)
0.81±(0.042)
0.92±(0.075)
0.33±(0.052)
0.37±(0.032)
0.41±(0.051)
0.41±(0.029)
0.77±(0.051)
0.88±(0.081)
0.22±(0.033)
0.26±(0.041)
0.15±(0.036)
0.25±(0.037)
0.37±(0.055)
0.41±(0.066)
Table 2. The averaged execution-time, success rate and the averaged multiple error for the GA-based algorithm, the
BFOA method and the proposed ABC algorithm, considering the six test images shown in Fig. 9
Image
(a)
(b)
(c)
(d)
(e)
(f)
p-Value
ABC vs. GA
ABC vs. BFOA
1.8061e-004
1.7454e-004
1.7981e-004
1.7788e-004
1.6989e-004
1.7012e-004
1.8288e-004
1.9011e-004
1.8922e-004
1.8698e-004
1.9124e-004
1.9081e-004
Table 3. p-values from Wilcoxon’s test, comparing ABC with GA and BFOA over the ME from Table 2.
Fig. 10 demonstrates the relative performance of ABC in comparison with the RHT algorithm following
the proposal in [12]. All images belonging to the test are complicated and contain different noise
conditions. The performance analysis is achieved by considering 35 different executions for each
algorithm over the three images. The results, exhibited in Fig. 10, present the median-run solution (when
the runs were ranked according to their final ME value) obtained throughout the 35 runs. On the other
hand, Table 4 reports the corresponding averaged execution time, success rate (in %), and average
multiple error (using (10)) for ABC and RHT algorithms over the set of images. Table 4 shows a decrease
in performance of the RHT algorithm as noise conditions change. Yet the ABC algorithm holds its
performance under the same circumstances.
15
Please cite this article as:
Cuevas, E., Sención-Echauri, F., Zaldivar, D., Pérez-Cisneros, M. Multi-circle detection on images using artificial bee
colony (ABC) optimization, Soft Computing, 16 (2), (2012), pp. 281-296.
Image
(I)
(II)
(III)
Average time ± Standard deviation (s)
Success rate
(SR) (%)
RHT
7.82±(0.34)
8.65±(0.48)
10.65±(0.48)
RHT
100
70
18
ABC
0.20±(0.31)
0.23±(0.28)
0.22±(0.21)
ABC
100
100
100
Average ME ± Standard deviation
RHT
0.19±(0.041)
0.47±(0.037)
1.21±(0.033)
ABC
0.20±(0.021)
0.18±(0.035)
0.23±(0.028)
Table 4. Average time, success rate and averaged error for ABC and HT, considering three test images
I
II
III
Original image
RHT
ABC
Fig. 10. Relative performance of RHT and ABC
6. Conclusions
This paper has presented an algorithm for the automatic detection of multiple circular shapes from
complicated and noisy images without considering the conventional Hough Transform principles. The
detection process is considered to be similar to a multi-modal optimization problem. In contrast to other
heuristic methods that employ an iterative procedure, the proposed ABC method is able to detect single or
multiple circles over a digital image by running only one optimization cycle. The ABC algorithm
searches the entire edge-map for circular shapes by using a combination of three non-collinear edge
16
Please cite this article as:
Cuevas, E., Sención-Echauri, F., Zaldivar, D., Pérez-Cisneros, M. Multi-circle detection on images using artificial bee
colony (ABC) optimization, Soft Computing, 16 (2), (2012), pp. 281-296.
points as candidate circles (food positions) in the edge-only image. A matching function (objective
function) is used to measure the existence of a candidate circle over the edge-map. Guided by the values
of this matching function, the set of encoded candidate circles is evolved using the ABC algorithm so that
the best candidate can fit into an actual circle. A novel contribution is related to the exhausted-source
memory that has been designed to hold “stuck” solutions which, in turn, represent feasible solutions for
the multi-circle detection problem. A post-analysis on the exhausted-source memory should indeed detect
other local minima, i.e., other potential circular shapes. The overall approach generates a fast sub-pixel
detector, which can effectively identify multiple circles in real images despite circular objects exhibiting a
significant occluded portion.
Classical Hough Transform methods for circle detection use three edge points to cast a vote for the
potential circular shape in the parameter space. However, they would require a huge amount of memory
and longer computational times to obtain a sub-pixel resolution. Moreover, HT-based methods rarely find
a precise parameter set for a circle in the image [47]. In our approach, the detected circles hold a sub-pixel
accuracy inherited directly from the circle equation and the MCA method.
In order to test the circle detection performance, both speed and accuracy have been compared. Score
functions are defined by (9)-(10) in order to measure accuracy and effectively evaluate the mismatch
between manually-detected and machine-detected circles. We have demonstrated that the ABC method
outperforms both the GA (as described in [15]) and the BFOA (as described in [16]) within a statistically
significant framework (Wilcoxon test). In contrast to the ABC method, the RHT algorithm [12] shows a
decrease in performance under noisy conditions. Yet the ABC algorithm holds its performance under the
same circumstances.
Finally, Table 2 indicates that the ABC method can yield better results on complicated and noisy images
compared with the GA and the BFOA methods. However, the aim of this study is not to beat all the circle
detector methods proposed earlier, but to show that the ABC algorithm can effectively serve as an
attractive method to successfully extract multiple circular shapes.
References
[1] da Fontoura Costa, L., Marcondes Cesar Jr., R., 2001. Shape Análisis and Classification. CRC Press,
Boca Raton FL.
[2] Yuen, H., Princen, J., Illingworth, J., Kittler, J., 1990. Comparative study of Hough transform
methods for circle finding. Image Vision Comput. 8 (1), 71–77.
[3] Iivarinen, J., Peura, M., Sarela, J., Visa, A., 1997. Comparison of combined shape descriptors for
irregular objects. In: Proc. 8th British Machine Vision Conf., Cochester, UK, pp. 430–439.
[4] Jones, G., Princen, J., Illingworth, J., Kittler, J., 1990. Robust estimation of shape parameters. In:
Proc. British Machine Vision Conf., pp. 43– 48.
[5] Fischer, M., Bolles, R., 1981. Random sample consensus: A paradigm to model fitting with
applications to image analysis and automated cartography. CACM 24 (6), 381–395.
[6] Bongiovanni, G., and Crescenzi, P.: Parallel Simulated Annealing for Shape Detection, Computer
Vision and Image Understanding, vol. 61, no. 1, pp. 60-69, 1995.
[7] Roth, G., Levine, M.D., 1994. Geometric primitive extraction using a genetic algorithm. IEEE Trans.
Pattern Anal. Machine Intell. 16 (9), 901–905.
[8] Peura, M., Iivarinen, J., 1997. Efficiency of simple shape descriptors. In: Arcelli, C., Cordella, L.P., di
Baja, G.S. (Eds.), Advances in Visual Form Analysis. World Scientific, Singapore, pp. 443–451.
[9] Muammar, H., Nixon, M., 1989. Approaches to extending the Hough transform. In: Proc. Int. Conf.
on Acoustics, Speech and Signal Processing ICASSP_89, vol. 3, pp. 1556–1559.
[10] T.J. Atherton, D.J. Kerbyson, “Using phase to represent radius in the coherent circle Hough
transform”, Proc, IEE Colloquium on the Hough Transform, IEE, London, 1993.
17
Please cite this article as:
Cuevas, E., Sención-Echauri, F., Zaldivar, D., Pérez-Cisneros, M. Multi-circle detection on images using artificial bee
colony (ABC) optimization, Soft Computing, 16 (2), (2012), pp. 281-296.
[11] Shaked, D., Yaron, O., Kiryati, N., 1996. Deriving stopping rules for the probabilistic Hough
transform by sequential analysis. Comput. Vision Image Understanding 63, 512–526.
[12] Xu, L., Oja, E., Kultanen, P., 1990. A new curve detection method: Randomized Hough transform
(RHT). Pattern Recognition Lett. 11 (5), 331–338.
[13] Han, J.H., Koczy, L.T., Poston, T., 1993. Fuzzy Hough transform. In: Proc. 2nd Int. Conf. on Fuzzy
Systems, vol. 2, pp. 803–808.
[14] Becker J., Grousson S., Coltuc D., 2002. From Hough transforms to integral transforms. In: Proc. Int.
Geoscience and Remote Sensing Symp., 2002 IGARSS_02, vol. 3, pp. 1444–144.
[15] Ayala-Ramirez, V., Garcia-Capulin, C. H., Perez-Garcia, A. and Sanchez-Yanez, R. E. Circle
detection on images using genetic algorithms. Pattern Recognition Letters, 2006, 27, pp. 652–657.
[16] Dasgupta, S., Das, S., Biswas A. and Abraham, A. Automatic circle detection on digital images whit
an adaptive bacterial foraging algorithm. Soft Computing, 2009, DOI 10.1007/s00500-009-0508-z.
[17] Holland, J.H., Adaptation in Natural and Artificial Systems, University of Michigan Press, Ann
Arbor, MI, 1975.
[18] K. Price, R. Storn, A. Lampinen, Differential Evolution a Practical Approach to Global
Optimization, Springer Natural Computing Series, 2005.
[19] J. Kennedy, R. Eberhart, Particle swarm optimization, in: IEEE International Conference on Neural
Networks (Piscataway, NJ), 1995, pp. 1942–1948.
[20] M. Dorigo, V. Maniezzo, A. Colorni, Positive feedback as a search strategy, Technical Report 91016, Politecnico di Milano, Italy, 1991.
[21] Liu Y, Passino K. Biomimicry of social foraging bacteria for distributed optimization: models,
principles, and emergent behaviors. J Optim Theory Appl 115(3):603–628, 2002.
[22] D. Karaboga. An idea based on honey bee swarm for numerical optimization, technical reportTR06,Erciyes University, Engineering Faculty, Computer Engineering Department 2005.
[23] N. Karaboga. A new design method based on artificial bee colony algorithm for digital IIR filters.
Journal of the Franklin Institute 346 (2009) 328–348.
[24] S. L. Ho, S. Yang. An artificial bee colony algorithm for inverse problems. International Journal of
Applied Electromagnetics and Mechanics, 31 (2009) 181–192.
[25] Ying-ping Chen, Pei Jiang. Analysis of particle interaction in particle swarm optimization.
Theoretical Computer Science 411(21), 2010, 2101-2115.
[26] Hongnian Zang, Shujun Zhang, Kevin Hapeshia. A Review of Nature-Inspired Algorithms. Journal
of Bionic Engineering 7(1), 2010, S232-S237.
[27] Josef Tvrdík. Adaptation in differential evolution: A numerical comparison. Applied Soft Computing
9(3), 2009, 1149-1155.
[28] Arijit Biswas, Swagatam Das, Ajith Abraham, Sambarta Dasgupta. Stability analysis of the
reproduction operator in bacterial foraging optimization. Theoretical Computer Science 411, 2010, 21272139.
[29] D. Karaboga, B. Basturk. On the performance of artificial bee colony (ABC) algorithm. Applied soft
computing, Volume 8, Issue 1, January 2008, Pages 687-697.
18
Please cite this article as:
Cuevas, E., Sención-Echauri, F., Zaldivar, D., Pérez-Cisneros, M. Multi-circle detection on images using artificial bee
colony (ABC) optimization, Soft Computing, 16 (2), (2012), pp. 281-296.
[30] D. Karaboga, B. Akay. A comparative study of Artificial Bee Colony algorithm. Applied
Mathematics and Computation 214 (2009) 108-132.
[31] Bresenham, J.E.: A Linear Algorithm for Incremental Digital Display of Circular Arcs.
Communications of the ACM 20, 100-106. (1987).
[32] Van Aken, J R. Efficient ellipse-drawing algorithm, IEEE Comp, Graphics applic., 2005. 4, (9), pp.
24-35.
[33] Aytug, H., Koehler, G.J.: New stopping criterion for genetic algorithms. European Journal of
Operational Research 126 (2000) 662–674.
[34] Greenhalgh, D., Marshall, S.: Convergence criteria for genetic algorithms. SIAM Journal on
Computing 20 (2000) 269–282.
[35] Wilcoxon F (1945) Individual comparisons by ranking methods. Biometrics 1:80–83.
[36] Garcia S, Molina D, Lozano M, Herrera F (2008) A study on the use of non-parametric tests for
analyzing the evolutionary algorithms’ behaviour: a case study on the CEC’2005 Special session on real
parameter optimization. J Heurist. doi:10.1007/s10732-008-9080-4.
[37] J. Santamaría, O. Cordón, S. Damas, J.M. García-Torres, A. Quirin, Performance Evaluation of
Memetic Approaches in 3D Reconstruction of Forensic Objects. Soft Computing, DOI: 10.1007/s00500008-0351-7, in press (2008).
19
| 9cs.NE
|
Synthesising Choreographies from Local Session Types
(extended version)
Julien Lange and Emilio Tuosto
arXiv:1204.2566v1 [cs.PL] 11 Apr 2012
Department of Computer Science, University of Leicester, UK
Abstract. Designing and analysing multiparty distributed interactions can be
achieved either by means of a global view (e.g. in choreography-based approaches)
or by composing available computational entities (e.g. in service orchestration).
This paper proposes a typing systems which allows, under some conditions, to
synthesise a choreography (i.e. a multiparty global type) from a set of local session types which describe end-point behaviours (i.e. local types).
1
Introduction
Communication-centred applications are paramount in the design and implementation
of modern distributed systems such as those in service-oriented or cloud computing.
Session types [8] and their multiparty variants [7, 9] offer an effective formal framework
for designing, analysing, and implementing this class of applications. Those theories
feature rather appealing methodologies that consists of (i) designing a global view of
the interactions – aka global type –, (ii) effective analysis of such a global view, (iii)
automatic projection of the global view to local end-points – aka local types –, and (iv)
type checking end-point code against local types. Such theories guarantee that, when
the global view enjoys suitable properties (phase (ii)), the end-points typable with local
types enjoy e.g., liveness properties like progress.
A drawback of such approaches is that they cannot be applied when the local types
describing the communication patterns of end-points are not obtained by an a priori
designed global view. For instance, in service-oriented computing, one typically has
independently developed end-points that have to be combined to form larger services.
Hence, deciding if the combined service respects its specification becomes non trivial.
To illustrate this, we introduce a simple example used throughout the paper.
Consider a system SBS = b1 [P1 ] | s1 [S1 ] | b2 [P2 ] | s2 [S2 ] consisting of two buyers
(b1 and b2 ) and two servers (s1 and s2 ) running in parallel, so that
P1 = t1 !order.p1 ?price.r?price.(c1 !.t1 !addr ⊕ c2 !.no1 !)
P2 = t2 !order.p2 ?price.r!price.(c2 ?.t2 !addr + c1 ?.no2 !)
Si = ti ?order.pi !price.(ti ?addr + noi ?), i ∈ {1, 2}
is the behaviour of b1
is the behaviour of b2
is the behaviour of si
with a!e (resp. a?e) representing the action of sending (resp. receiving) a message of
type e on a channel a (we omit e when the message is immaterial), ⊕ representing an
internal choice, and + a choice made by the environment.
Intuitively, the overall behaviour of SBS should be that either b1 or b2 purchase from
their corresponding sellers. A natural question arises: is SBS correct? Arguably, it is not
immediate to decide this by considering the end-point behaviours.
We propose to construct a global view of distributed end-points like SBS by a straightforward extension of the multiparty session types introduced in [9]. Such types formalise a global view of the behaviour which, for SBS , resembles the informal diagram
below, where the choreography of the overall protocol becomes much clearer.
An advantage of our approach is that we can reuse the results of the theory of multiparty session types to prove properties of end-points e.g. safety and progress. In fact,
we show that when the choreography can be constructed, its projections correspond to
the initial end-points. Therefore, the well-formedness of the synthesised global choreography guarantees progress and safety properties of end-points.
The extraction of session types from programs has been studied extensively [6, 8, 9].
We assume in this work that such session types are readily available before addressing
the construction of a global type.
Contributions. We introduce a theory whereby, under some conditions, it is possible to
assign a global type to a set of local types. If a global type can be constructed from a
set of local types, we show that it is unique (Theorem 2) and well-formed (Theorem 3).
In addition, we show that typability is preserved by reduction (Theorem 4). Our theory
also guarantees progress and safety properties (Theorems 5 and 6). We also show that
the projections of a constructed global type are equivalent to the original system (Theorem 7). Finally, we show that for every well-formed global types, an equivalent global
type can be assigned to its projections (Theorem 8).
Synopsis. In § 2, we give the syntax and semantics of the local types from which it is
possible to construct a global type. In § 3, we present an extension of the global types
in [9]. In § 4, we introduce a typing systems for local types, and we give our main
results. Finally, in § 5 we conclude, and discuss related and future work.
2
Local Types
We use CCS-like processes (with guarded external and internal choices) to infer a global
type from local types that correspond to the participants in the inferred choreography.
2
Hereafter, P is a denumerable set of participant names (ranged over by s, r, n, . . . ) and
C is a denumerable set of channel names (ranged over by a, b, . . . ).
Syntax. The syntax of local types below is parametrised wrt basic data types such as
bool, int, . . . (ranged over by e):
S, T
::= S | S0 | n[P] | a : ρ | 0
P, Q
::=
L
i∈I ai !ei .Pi
| ∑i∈I ai ?ei .Pi | µ x.P | x
A system S consists of the parallel composition of processes and queues. A process
n[P] is a behaviour P identified by n ∈ P; we assume that the participant names are all
different. A behaviour is either a guarded external choice, a guarded internal choice,
L
or a recursive process. An internal choice i∈I ai !ei .Pi is guarded by output prefixes
ai !ei representing the sending of a value of sort ei on channel ai . An external choice
∑i∈I ai ?ei .Pi is guarded by input prefixes ai ?ei representing the reception of a value of
type ei on channel ai . We adopt asynchronous (order-preserving) communications and
assume that the channels in the guards of choices are pairwise distinct; moreover
def
=
0
L
i∈∅ ai !ei .Pi
=
∑i∈∅ ai ?ei .Pi
Finally, in a recursive behaviour µ x.P, all occurrences of x in P are bound and prefixguarded; also, we consider closed behaviours only that is, behaviours with no free occurrences of recursion variables. We assume that bound variables are pairwise distinct.
A program is a system with no queues, while a runtime system is a system having
exactly one queue a : ρ per channel name a ∈ C in S. In the following, S, T, . . . denote
either a program or runtime system.
Semantics. The semantics of local types is given by the labelled transition system (LTS)
whose labels are
λ ::= α | X | a · e | e · a | n[α] | n : α
where
α ::= a!e | a?e
Label α indicates either sending or reception by a process. Label X indicates termination, a · e and e · a respectively indicate push and pop operations on queues. Label
n[α] indicates a communication action done by participant n while n : a!e and n : a?e
indicate a synchronisation between n and a queue.
Assume the usual laws for commutative monoids for | and 0 on systems and µ x.P ≡
λ
P[µ x.P/x]. The LTS −→ is the smallest relation closed under the following rules:
[ INT ]
[ PUSH ]
L
i∈I ai !ei .Pi
a·e
n[a?e]
a j ?ej
−→ Pj
a : ρ −→ a : ρ · e
[ EXT ]
[ POP ]
[ OUT ]
n:a?e
e·a
a : e · ρ −→ a : ρ
X
0 −→ 0
[ END ]
P −→ P0
α
T −→ T 0
[ BOX ]
n:a!e
S | T −→ S0 | T 0
λ
α
[ EQ - S ]
P −→ P0
α
S −→ S0
λ
3
n[α]
n[P] −→ n[P0 ]
S ≡ T −→ T 0 ≡ S0
P ≡ Q −→ Q0 ≡ P0
[ EQ - P ]
j∈I
a·e
S −→ S0
T −→ T 0
S | T −→ S0 | T 0
∑i∈I ai ?ei .Pi −→ Pj
n[a!e]
e·a
S −→ S0
[ IN ]
a j !ej
Rules [ INT ] and [ EXT ] are trivial. By [ PUSH ] (resp. [ POP ]), a queue receives a (resp. sends the
first) datum (resp. if any). Processes can synchronise with queues according to rules [ IN ]
and [ OUT ]. The remaining rules are rather standard. Let S −→ iff there are S0 and λ s.t.
λ
λ ...λn
λ
1
S −→ S0 and =⇒
(resp. =⇒) be the reflexive transitive closure of −→ (resp. −→).
3
Global Types
A global type G specifies an ordering of the interactions in a choreography. The syntax
for global types in [9] is extended with a generalised sequencing in the following syntax:
G ::= s → r : ahei.G | G ; G 0 | G + G 0 | G | G 0 | µ χ.G | χ | 0
The prefix s → r : ahei represents an interaction where s ∈ P sends a value of sort e to
r ∈ P on a ∈ C (we let ι range over interactions s → r : ahei and assume that s 6= r). The
production G ; G 0 indicates a generalised form of sequencing, where the interactions in
G 0 are enabled only after the ones in G . The production G + G 0 indicates a (exclusive)
choice of interactions. Concurrent interactions are written G | G 0 . A global type µ χ.G ,
indicates a recursive type, where χ is bound in G . We assume that global types are
closed and that recursion is guarded. We often omit trailing occurrences of 0.
Example 1. The first two interactions between bi and si in the example of § 1 are
Gi = bi → si : ti horderi.si → bi : pi hpricei i ∈ {1, 2}
(3.1)
The type Gi says that a participant bi sends a message of type order to participant si
on channel ti , then si replies with a message of type price on channel pi .
The smallest equivalence relation satisfying the laws for commutative monoids for
| , +, and 0 and the axioms below is the structural congruence for global types:
G ;0 ≡ G
0;G ≡ G
(G ; G 0 ) ; G 00 ≡ G ; (G 0 ; G 00 )
ι.(G ; G 0 ) ≡ (ι.G ) ; G 0
µ χ.G ≡ G [µ χ.G /χ]
The syntax of global types may specify behaviours that are not implementable. The
rest of this section borrows from [5] and [9] and adapts the requirements a global type
must fulfil to ensure that the ordering relation it prescribes is indeed feasible.
3.1
Channel Usage and Linearity
It is paramount that no race occurs on the channels of a global type (i.e. a datum sent
on a channel is received by its intended recipient). As in [9], we require that a global
type is linear, that is actions on channels shared by different participants are temporally
ordered. For this, we use generic environments (ranged over by C) which keep track of
channel usage. Such environments are trees defined as follows:
c
c
•
C
root only
C1
C is a child of c
C2
C1 and C2 are children of c
4
Each node c has a label c of the form ◦, s → r : a, or µχ respectively representing
the root of choice or concurrent branches, an interaction between s and r on a, and a
recursive behaviour. We write c ∈ C if c is a node in C. We use as a wild-card when
some of the components of a label are immaterial, e.g. → : a matches any label
representing an interaction on a. Given a tree C, we write c1 ≺ c2 , if c1 , c2 ∈ C and c2
is a node in the sub-tree rooted at c1 . We adapt the definitions in [9] to our framework.
Definition 1 (Dependency relations [9]) Fix C, we define the following relations:
c1 ≺II c2 if c1 ≺ c2 and ci = si → r : ai i ∈ {1, 2}
c1 ≺IO c2 if c1 ≺ c2 and c1 = s1 → r : a1 and c2 = r → s2 : a2
c1 ≺OO c2 if c1 ≺ c2 and ci = s → ri : a i ∈ {1, 2}
An input dependency from c1 to c2 is a chain of the form c1 ≺φ1 . . . ≺φk c2 (k ≥ 0) such
that φi ∈ {II, IO} for 1 ≤ i ≤ k − 1 and φk = II. An output dependency from c1 to c2 is
a chain c1 ≺φ1 . . . ≺φk c2 (k ≥ 1) such that φi ∈ {OO, IO}.
Definition 2 (Linearity [9]) C is linear if and only if whenever c1 ≺ c2 with c1 = →
: a and c2 = → : a then there is both input and output dependencies from c1 to c2 .
We also define a function ? to append trees as follows
c
c
?C0 =
c
,
C0 ?C0
C0
c
?C0 =
• ?C = C,
C1
C2
C1 ?C0 C2 ?C0
and a partial function to append a tree C0 to a tree C while preserving linearity: CC0 =
C ?C0 if C ?C0 is linear, otherwise CC0 = ⊥. Also, let T (G ) be the total function (cf.
Appendix A.1) which returns a tree C corresponding to the use of channels in G .
3.2
Well-formed Global Types
We define the conditions for a global type to be well-formed. We write P (G ) (resp.
C (G )) for the set of participant (resp. channel) names in G , and fv(G ) for the set of
free variables in G , similarly for a system S. We give a few accessory functions. Let
def
R(G ) = {s → r : a | G ≡ (s → r : ahei.G1 + G2 | G3 ) ; G4 }
(
def FP (G1 ) ∪ FP (G2 ), G = G1 | G2
FP (G ) =
{P (G )},
otherwise
FO ({s, r} ∪ P , G1 ),
G = s → r : ahei.G1
F
(∅,
G
)
∪
F
(∅,
G
),
G = G1 | G2
O
1
O
2
G = G1 + G2 and FO (P , G1 ) = FO (P , G2 )
FO (P , G1 ),
def
FO (P , G ) = FO (P , G1 ),
G = µ χ.G1
FO (∅, G2 ),
G = G1 ; G2
{P },
G = 0 or G = χ
⊥,
otherwise
5
R(G ) is the ready set of G . FP (G ) is the family of sets of its participants running in
different concurrent branches. That is, N ∈ FP (G ) iff all n ∈ N are in a same top-level
thread of G . FO (G , P ) is the family of sets of participants of G , so that for all N, M ∈
FO (P , G ), the participants in N and those in M are in different concurrent branches in
def
the last part of G ; define FO (G ) = FO (∅, G ). Note that FO ( , ) is a partial function.
Example 2. Let Gi, j = b1 → b2 : ci hi.(bi → si : ti haddri | bj → sj : no j hi) describe
each of the branches of the or box in the example of § 1, where i 6= j ∈ {1, 2}, then
R(G1,2 ) = {b1 → b2 : c1 } , FP (G1,2 ) = {b1 , s1 , b2 , s2 } , FO (G1,2 ) = {{b1 , s1 }, {b2 , s2 }}
The global type below corresponds to the whole protocol of § 1
G = (G1 | G2 ) ; b2 → b1 : rhpricei.(G1,2 + G2,1 )
hence R(G ) = {bi → si : ti }i=1,2 , FP (G ) = FP (G1,2 ), and FO (G ) = FO (G1,2 ).
Well-formedness. The well-formedness of a global type G depends on how it uses
channels; a judgement of the form C `G states that G is well-formed according to the
channel environment C (cf. § 3.1); G is well-formed if • ` G can be derived from the
rules given in Fig. 1. We assume that each premise of the rules in Fig. 1 does not hold if
any of the functions used are not defined (e.g., in [ WF - ;], if FO (G ) = ⊥ then C `G ; G 0 is
not derivable). Hereafter, we assume that a node c is fresh (i.e. c 6∈ C). The environment
C permits to tackle one of the main requirements for a global type to be well-formed:
there should not be any race on channels. In the following, we discuss the rules of Fig. 1,
which are grouped according to three other requirements: sequentiality, single threaded,
and knowledge of choice.
Sequentiality [5]. Rules [ WF -.], [ WF - ;] and [ WF - ;-0] ensure that sequentiality is preserved.
In [ WF -.], there must be an ordering dependency between a prefix and its continuation so
that it is possible to implement each participant so that at least one action of the first
prefix always happens before an action of the second prefix. More concretely, we want
to avoid global types of the form, e.g.
s1 → r1 : ahei.s2 → r2 : bhe0 i 8
where, evidently, it is not possible to guarantee that s2 sends after r1 receives on a.
Since we are working in an asynchronous setting, we do not want to force both send
and receive actions of the first prefix to happen before both actions of the second one.
Rule [ WF - ;] requires the following for generalised sequencing. (i) For each pair of “first”
participants in G 0 , there exist two concurrent branches of G such that these two participants appear in different branches. This is to avoid global types of the form, e.g.
(s1 → r1 : ahei | s2 → r2 : bhei) ; s1 → r1 : chei 8
since there is no possible sequencing between the prefix on b and the one on c. (ii) For
all top-level concurrent branches in G , there is a participant in that branch which is also
in one of the branches of G 0 . This requirement discards global types of the form, e.g.
(s1 → r1 : ahei | s2 → r2 : bhei | s3 → r3 : chei) ; s1 → r2 : dhei 8
6
[ WF -.]
∀ s0 → r0 : ∈ R(G ) : {s0 , r0 } ∩ {s, r} 6= ∅
[ WF - ;]
C c `G
c=s→r:a
C `s → r : ahei.G
∀ s → r : ∈ R(G 0 ) . ∃N1 6= N2 ∈ FO (G ) . s ∈ N1 ∧ r ∈ N2
∀N ∈ FP (G ) . ∃N 0 ∈ FP (G 0 ) . N ∩ N 0 6= ∅
C `G
C T (G ) `G 0
C `G ; G 0
0
C (G ) ∩ C (G 0 ) = ∅
C `G
C `G 0
C `G | G 0
χ ∈ fv(G ) ⇒ #FO (G ) = 1
C ? c `G
c=χ
[ WF -µ χ]
C `µ χ.G
C `G
CC(χ)
[ WF - ;-0]
[ WF -χ]
C `G ; 0
C `χ
[ WF -0] C `0
C `G
C `G 0
∀ s → r : a ∈ R(G ).∀ s0 → r0 : b ∈ R(G 0 ) . s = s0 ∧ a 6= b
C `G + G 0
[ WF - | ]
[ WF -+]
P (G ) ∩ P (G ) = ∅
Fig. 1. Rules for Well-formedness
since it is not possible to enforce an order between s3 and r3 and the others. (iii) G and
G 0 are also well-formed. Observe that (i) implies that for G ; G 0 to be well-formed, G is
of the form G1 | G2 , with G1 6= 0 and G2 6= 0. Both [ WF -.] and [ WF - ;] are only applicable
when linearity is preserved. Finally, rule [ WF - ;-0] is a special case of G ; G 0 .
Single threaded [9]. A participant should not appear in different concurrent branches
of a global type, so that each participant is single threaded. This is also reflected in
the calculus of § 2, where parallel composition is only allowed at the system level.
Therefore, in [ WF - | ], the participant (resp. channel) names in concurrent branches must
be disjoints. Rule [ WF -µ χ] adds a new node in C to keep track of recursive usage of the
channels, and requires that G is single threaded, i.e. concurrent branches cannot appear
under recursion. If that was the case, a participant would appear in different concurrent
branches of the unfolding of a recursive global type. Rule [ WF -χ] unfolds C at χ to ensure
that the one-time unfolding of C preserves linearity (see [9] for details).
Knowledge of choice [5, 9]. Whenever a global type specifies a choice of two sets of
interactions, the decision should be made by exactly one participant. For instance,
s1 → r1 : a1 hei.G1
s2 → r2 : a2 he0 i.G2
+
8
specifies a choice made by two participants. Indeed, s1 is the one making a decision
in the first branch, while s2 makes a decision in the second one; this kind of choreographies cannot be implemented (without using hidden interactions). Also, we want
to avoid global types where a participant n behaves differently in two choice branches
without being aware of the choice made by others. For instance, in
s → r : ahei.n → r : chei.G1
+
s → r : bhei.n → r : dhei.G2
7
8
where n ignores the choice of s and behaves differently in each branch. On the other
hand, we want global types of following form to be accepted.
s → r : ahei.n → s : bhei.s → n : chei.n → r : dhei
+
s → r : a0 hei.n → s : bhei.s → n : c0 hei.n → r : d 0 hei
4
Indeed, in this case n behaves differently in each branch, but only after “being informed” by s about the chosen branch.
Together with the projection map defined below, rule [ WF -+] guarantees that “knowledge of choice” is respected. In particular, the rule requires that the participant who
makes the decision is the same in every branch of a choice, while the channels guarding
the choice must be distinct.
Definition 1 ( ). The projection of a global type G wrt. n ∈ P (G ) is defined as
0
a?e.G n ,
0
a!e.G n ,
G 0 n ,
G1 n ]G2 n ,
def
G n = Gi n ,
G1 n [G2 n /0],
µ χ.G 0 n ,
G,
⊥,
if G = s → n : ahei.G 0
if G = n → r : ahei.G 0
if G = s → r : ahei.G 0 and s 6= n 6= r
if G = G1 + G2
if G = G1 | G2 and n 6∈ P (G j ), i 6= j ∈ {1, 2}
if G = G1 ; G2
if G = µ χ.G 0
if G = χ or G = 0
otherwise
We say that a global type is projectable if G n is defined for all n ∈ P (G ).
The projection map is similar to the one given in [9], but for the generalised sequencing
case and the use of ] to project choice branches. Observe that if G = G1 ; G2 , we
replace 0 by the projection of G2 in the projection of G1 . Function ] basically merges
(if possible) the behaviour of a participant in different choice branches; ] is defined
only when the behaviour is the same in all branches, or if it differs after having received
enough information about the branch which was chosen. The definition of ] is given
in Appendix A.2. A global type may be projected even if is not well-formed, but in that
case none of the properties given below are guaranteed to hold.
4
Synthesising Global Types
We now introduce a typing systems to synthesise a global type G from a system S so that
S satisfies safety and progress properties (e.g. no race on channels and no participant
gets stuck). Also, the set of typable systems corresponds exactly to the set of systems
obtained by projecting well-formed global types. To synthesise G from a system S, a
careful analysis of what actions can occur at each possible state of S is necessary.
8
If S ≡ n[P] | S0 then S(n) denotes P (if S 6≡ n[P] | S0 then S(n) = ⊥). We define the
ready set of a system as follows:
{ai |i ∈ I} ∪ R(S0 ) if S ≡ r[∑i∈I ai ?ei .Pi ] | S0
{ai |i ∈ I} ∪ R(S0 ) if S ≡ s[Li∈I ai !ei .Pi ] | S0
R(S) =
if S ≡ a : e · ρ | S0
{a} ∪ R(S0 )
∅
if S ≡ 0
def
We overload R( ) on behaviours in the obvious way and define R(S) = {a ∈ C | a ∈
R(S) or a ∈ R(S)} and S l ⇐⇒ ∃ a ∈ C : a ∈ R(S) ∧ a ∈ R(S); we write S 6l if S l does
not hold.
4.1
Validation Rules
A judgement of the form A ; Γ ; C ` S I G says that the system S forms a choreography
defined by a global type G , under the environments A, Γ, and C. The environment A is a
superset of the channel names used in S, and corresponds to the channels S is entitled to
use. The environment Γ maps participant names and local recursion variables to global
recursion variables (◦ is the empty context Γ). The channel environment C records the
use of channels. Hereafter, we use · for the disjoint union of environments.
Programs. A global type G can be synthesised from the program S if the judgement
C (S) ; ◦ ; • ` S I G
(stating that S is entitled to use all its channels in empty environments) is derivable from
the rules in Fig. 2 (driven by the ready set of S and the structure of its processes).
Rule [.] validates prefixes provided that the system is entitled to linearly use the
channel a, that the continuation is typable, and that no other interactions are possible in
S. For instance, [.] does not apply to
s1 [a!e.P1 ] | r1 [a?e.Q1 ] | s2 [b!e.P2 ] | r2 [b?e.Q2 ] 8
because there is no ordering relation between the actions on a and b; in this case either
[ | ] or [ ;] should be used.
Rule [ | ] validates concurrent branches when they can be validated using a partition A1
and A2 of the channels (recall that P (S) ∩ P (S0 ) = ∅).
Rule [ ;] splits the system into two sequential parts and it relies on the function split( )
defined in § 4.2; for now it suffices to notice that linearity is checked for in the second
part of the split by adding the channel environment corresponding to G1 to C (recall that
CC0 is undefined if C ?C0 is not linear).
Rule [⊕] introduces the global type choice operator, it requires that both branches are
typable and that no other interactions are possible in S.
Rule [+] allows to discharge a branch of an external choice; together with the premises
of [ | ], rule [+] discards systems such as the one on left below (due to a race on b) but
9
[.]
{a} ∪ A ; Γ ; C c ` s[P] | r[Q] | S I G
[|]
[ ;]
c=s→r:a
S 6l
{a} ∪ A ; Γ ; C ` s[a!e.P] | r[a?e.Q] | S I s → r : ahei.G
A2 ; ◦ ; C ` S 0 I G 0
A1 ; ◦ ; C ` S I G
A1 ∩ A2 = ∅
A1 ∪ A2 ; Γ ; C ` S | S0 I G | G 0
A ; ◦ ; C ` S1 I G1
split(S) = (S1 , S2 )
A ; ◦ ; C T (G1 ) ` S2 I G2
A ; Γ ; C ` S I G1 ; G2
[⊕]
A ; Γ ; C ` s[P] | S I G
R(Q) ⊆ A
A ; Γ ; C ` r[P] | S I G
S 6l
A ; Γ ; C ` r[P + Q] | S I G
∃ 1 ≤ i, j ≤ k . (ni [Pi ] | nj [Pj ]) l
A ; Γ · (n1 , x1 ) : χ, . . . , (nk , xk ) : χ ; C ? µχ ` n1 [P1 ] | . . . | nk [Pk ] I G
A ; Γ ; C ` n1 [µ x1 .P1 ] | . . . | nk [µ xk .Pk ] I µ χ.G
[X]
S≡S
[eq]
S 6l
A ; Γ ; C ` s[P ⊕ Q] | S I G + G 0
[+]
[µ]
A ; Γ ; C ` s[Q] | S I G 0
0
∀1 ≤ i ≤ k . Γ(ni , xi ) = χ CC(µχ)
A ; Γ ; C ` n1 [x1 ] | . . . | nk [xk ] I χ
A ; Γ ; C ` S0 I G
A; Γ; C ` S I G
[0]
∀n ∈ P (S) . S(n) = 0
C (S) = ∅
A; Γ; C ` S I 0
Fig. 2. Validation Rules for Programs
permits those like the one on the right (as only the channels guarding the choice must
be in A).
r1 [a?e + b?e] | s2 [b!e] | r2 [b?e] 8 s1 [a!e] | r1 [a?e + c?e.b?e] | s2 [b!e] | r2 [b?e] 4
Rules [µ] and [ X ] handle recursive systems. The former rule “guesses” the participants
involved in a recursive behaviour. If two of them interact, [µ] validates the recursion
provided that the system can be typed when such participants are associated to the
global recursion variable χ (assuming that χ is not in Γ). Rule [ X ] checks that all the
participants in the recursion have reached a local recursion variable corresponding to
the global recursion, and that the unfolding of C on µχ preserves linearity; for this we
define C(µχ) to be the subtree of C rooted at the deepest node of C labelled by µχ (note
that this node is unique since bound variables are distinct).
Rule [0] only applies when all the participants in S end while [eq] validates a system up
to structural congruence.
Theorem 1 (Decidability). Typability is decidable.
The proofs follows from the fact that the typing is done wrt to the (finite) partitions of
channels in a system, and that the number of required behaviour unfoldings is finite.
Theorem 2 (Unique typing). If A ; Γ ; C ` S I G and A ; Γ ; C ` S I G 0 then G ≡ G 0 .
10
Theorem 3 (Well-formedness). If A ; Γ ; C ` S I G then • `G and G is projectable.
The proofs for these two theorems are by induction on the structure of the derivation.
Runtime system. In order to have subject reduction for our typing systems, queues
have to be handled effectively; we use a distinguished participant name ∗ to denote an
anonymous participant. Assume ∗ 6∈ P and write ∗ → r : ahei.G to specify that there is
a message of sort e on channel a for participant r.
Example 3. Let S = n[a!e] | s[b!e.a?e] | r[b?e] | a : [] | b : []. Consider the judgement
A ; Γ ; C ` S I s → r : bhei.n → s : ahei
If S evolves to S0 = n[0] | s[b!e.a?e] | r[b?e] | a : e | b : [], the identity of the sender n
is lost. However, the judgement
A ; Γ ; C ` S0 I s → r : bhei.∗ → s : ahei
types S0 using ∗.
Runtime systems can be handled by slightly extending Def. 1 so that we have1
c1 ≺OO c2 if c1 ≺ c2 and c1 = ∗ → r : a and c2 = s → r : a
and by adding two rules to the validation rules for handling queues:
[ρ]
{a} ∪ A ; ◦ ; C c ` a : ρ | r[P] | S I G
c=∗→r:a
S 6l
{a} ∪ A ; Γ ; C ` a : e · ρ | r[a?e.P] | S I ∗ → r : ahei.G
[[]]
A; Γ; C ` S I G
A ; Γ ; C ` a : [] | S I G
Rule [ρ] is similar to rule [.], except that a non-empty queue replaces the sender, and Γ
is emptied. Rule [[]] simply allows to remove empty queues from the system.
λ
Theorem 4. If A ; ◦ ; C ` S I G , S −→ S0 , and C (λ) 6∈ C then A ; ◦ ; C ` S0 I G 0
The proof is by case analysis on the different types of transitions a system can make. The
recursive case follows from the fact that reduction preserves closeness of behaviours.
4.2
Splitting Systems
The purpose of systems’ splitting is to group participants according to their interactions.
For this we use judgements of the form
Ψ; Θ ` S m Ω
(4.1)
which reads as “S splits as Ω under Ψ and Θ”. The environment Ψ is a set of (pairwise
disjoint) ensembles that is disjoint sets N ⊆ P (S) containing participants that interact
with each other for a while; and then some of them may interact with participants in
other ensembles in Ψ. The environment Θ is a set of (pairwise disjoint) duos, that is
1
This extension makes sense since the order of messages is preserved in the calculus.
11
two-element sets of participants {s, r ∈ P (S) : r 6= s} representing the first participants
able to interact once the first part of the split is finished. Under suitable conditions,
one could identify when n ∈ N has to interact with a participant of another ensemble.
In other words, one can divide S(n) as n[P1 · ε · P2 ]: the interactions in P1 happen with
participants in the ensemble of n, while P2 starts interacting with a participant in another
ensemble. Finally, the environment Ω assigns behaviours augmented with a separator ε
to participant names, and lists of sorts to queues a.
Θ
Given a judgement as (4.1), we say that N, M ∈ Ψ are Θ-linked (N _ M in symbols)
Ω
/ also, we say that n, m ∈ P (S) are Ω-linked (n _ m in symiff ∃D ∈ Θ : N ∩ D ∩ M 6= 0;
def
/ We define S[N] = ∏n∈N n[S(n)] | ∏a∈C (S) a : S(a).
bols) iff C (Ω(n)) ∩ C (Ω(m)) 6= 0.
Definition 2. The judgement Ψ; Θ ` S m Ω is coherent if it can be derived from the
rules in Fig. 3, Θ 6= ∅, and for all N ∈ Ψ, S[N] l and the following conditions hold
∃!n ∈ N : (∃!m ∈ N \ {n} : S[N \ {n}] 6l ∧ S[N \ {m}] 6l) or (S[N \ {n}] 6l)
(4.2)
~ is total on N and ↔Θ is total on Ψ
def Θ ∗
Θ
(4.3)
def Ω +
where ↔Θ = _ is the reflexive and transitive closure of _ and ~ = _ is the transiΩ
tive closure of _.
Essentially, Def. 2 ensures that rule [ ;] is the only rule of Fig. 2 applicable when the
system can be split. Condition (4.2) ensures that, in each ensemble N, there is a unique
pair of synchronising participants or there is a unique participant that can synchronise
with a queue a. Condition (4.3) is the local counterpart of the well-formedness rule for
global types of the form G ; G 0 . The totality of ~ on N guarantees that the participants
in an ensemble share channels. The totality of ↔Θ on Ψ guarantees that each ensemble
in Ψ has one “representative” which is one of the first participants to interact in the
second part of the split. Together with condition Θ 6= ∅, the condition on ↔Θ ensures
that there are (at least) two ensembles of participants in Ψ. Note that (4.3) also ensures
that all the set of participants in Ψ are interdependent (i.e. one cannot divide them into
independent systems, in which case rule [ | ] should be used).
A judgement (4.1) is to be derived with the rules of Fig.3 (we omit rules for commutativity and associativity of systems). The derivation is driven by the structure of up
to two processes in S, and whether they are in the same ensemble and/or form a duo.
Rule [ε] marks two processes m and n as “to be split” when m and n form a duo in
Θ and are in different ensembles of Ψ. The continuation of the system is to be split as
well, with m and n removed from the system and from the environments.
Rule [sync] records in Ω the interactions of participants in a same ensemble of Ψ.
Rule [+] discharges the branch of an external choice for participants in a same ensemble
while [⊕] deals with internal choice. The premise Ω Ω0 holds only when Ω and Ω0
have the same domain and differ only up to external choice, i.e. for each n either its
split is the same in both branches, or its split is an external choice (guarded by different
channels); Ω t Ω0 merges Ω and Ω0 accordingly (cf. Appendix A.3). The additional
premise s[P ⊕ P0 ] | r[Q] l is required so that the split is done before a branching if a
participant cannot interact with one of its peer in N after the branching.
12
n ∈ N, m ∈ M
[ε]
(n[P] | m[Q]) l
s, r ∈ N
[sync]
Ψ · N; Θ ` s[P] | r[Q] | S m Ω · s : π · r : ϕ
Ψ · N; Θ ` s[a!e.P] | r[a?e.Q] | S m Ω · s : a!e.π · r : a?e.ϕ
m, n ∈ N
(m[P] | n[Q]) l
[+]
[⊕]
Ψ · N\{n} · M\{m}; Θ ` S m Ω
Ψ · N · M; Θ · {n, m} ` n[P] | m[Q] | S m Ω · n : ε · m : ε
Ψ · N; Θ ` m[P] | n[Q] | S m Ω · m : π
Ψ · N; Θ ` m[P + P0 ] | n[Q] | S m Ω · m : π
n, m ∈ N
(n[P ⊕ P0 ] | m[Q]) l
Ω Ω0
Ψ · N; Θ ` n[P] | m[Q] | S m Ω · n : π
Ψ · N; Θ ` n[P0 ] | m[Q] | S m Ω0 · n : ϕ
Ψ · N; Θ ` n[P ⊕ P0 ] | m[Q] | S m Ω t Ω0 · n : π ⊕ ϕ
[ax]
{∅}; ∅ ` 0 m ∅
(n[P] | S) 6l
[rem]
r∈N
[q]
[0]
P 6≡ 0
Ψ \ n; Θ ` S m Ω
Ψ; Θ ` n[0] | S m Ω · n : 0
Ψ \ n; Θ ` S m Ω
Ψ; Θ ` n[P] | S m Ω · n : ε
Ψ · N; Θ ` r[P] | S m Ω · r : π · a : ρ
Ψ · N; Θ ` r[a?e.P] | a : e · ρ | S m Ω · r : a?e.π · a : e · ρ
Fig. 3. Splitting Systems.
Rule [ax] terminates a derivation (all environments emptied) while [0] completes the split
of a process (abusing notation, Ψ \ n denotes the removal of n from any N ∈ Ψ).
Rule [rem] marks a process to be split when it cannot interact with anyone in S. The
premise P 6≡ 0 allows to differentiates a process which terminates after the split, from
others which terminate before. In the latter case, rule [0] is to be used.
Rule [q] records in Ω interactions with non-empty queues.
We now define a (partial) function split which splits a system into two parts.
Definition 3 (split( )). Let Ψ; Θ ` S m Ω be a coherent judgement. Define split(S) =
(S1 , S2 ) where
– ∀n ∈ P (S). S1 (n) = S(n) − Ω(n) and S2 (n) = S(n) % Ω(n)
– ∀a ∈ C (S). S1 (a) = Ω(a) and S2 (a) = S(a) \ Ω(a)
if S(n) % Ω(n) 6= ⊥ for all n ∈ P (S), and split(S) = ⊥ otherwise.
The auxiliay funtions − and % used in Def. 3 are defined in Appendix A.3; we
give here their intuitive description. Let n ∈ P (S), and Ψ; Θ ` S m Ω be a coherent
judgement. Function S(n) − Ω(n) returns the “first part” of the split of n, that is the
longest common prefix of S(n) and Ω(n), while S(n) % Ω(n) is partial and returns the
the remaining part of the behaviour of S(n) after Ω(n).
Example 4. Taking SBS as in § 1, we have split(SBS ) = (S1 , S2 ) such that
S1 (b1 ) = t1 !order.p1 ?price
S1 (si ) = ti ?order.pi !price
S2 (b1 ) = r?price.(c1 !.t1 !addr ⊕ c2 !.no1 !)
S2 (si ) = ti ?addr + noi ?
Note that {{b1 , s1 }, {b2 , s2 }} ; {{b1 , b2 }} ` SBS m Ω is coherent.
13
4.3
Properties of Synthesised Global Type
Progress and safety. If a system is typable, then it will either terminate or be able to
make further transitions (e.g. if there are recursive processes).
X
Theorem 5. If A ; ◦ ; C ` S I G then S −→ S0 , or ∀ n ∈ P (S) . S(n) = 0, or S −→.
Let us add the rule [ ERROR ] below to the semantics given in § 2.
a?e0
[ ERROR ]
e·a
T −→ T 0
S −→ S0
S | T −→ error
e 6= e0
Theorem 6. If A ; ◦ ; C ` S I G , then S is race free and S =⇒−→ error is not possible.
The proofs of Theorems 5 and 6 are by contradiction, using Theorem 4.
Behavioural equivalences. We show that there is a correspondence between the original
system and the projections of its global type. First, let us introduce two relations.
α
α
Definition 3 ( . and ≈) P . Q if and only if Q −→ Q0 then there is P0 such that P −→
α
α
α
P0 . Also, S ≈ T iff whenever S −→ S0 then T −→ T 0 and S0 ≈ T 0 ; and whenever T −→ T 0
α
then S −→ S0 and S0 ≈ T 0 where α ∈ {n : a!e, n : a?e, X}.
The behaviour of a participant in S is a simulation of the projection of a synthesised
global type from S onto this participant. Intuitively, the other direction is lost due to
rule [+], indeed external choice branches which are never chosen are not “recorded” in
the synthesised global type.
Lemma 1. If A ; ◦ ; C ` S I G then ∀n ∈ S . G n . S(n).
The proof is by case analysis on the transitions of S, using Theorem 4.
Since the branches that are not recorded in a synthesised global type are only those
which are never chosen, we have the following result.
Theorem 7. If A ; ◦ ; C ` S I G then ∏n∈P (S) n[G n ] ≈ S.
The proof is by case analysis on the transitions of S, using Theorem 4 and Lemma 1.
Our completeness result shows that every well-formed and projectable global type
is inhabited by the system consisting of the parallel composition of all its projections.
Theorem 8. If • ` G and G is projectable, then there is G 0 ≡ G such that A ; Γ ; C `
∏n∈P (G ) n[G n ] I G 0 .
The proof is by induction on the structure of (well-formed) G .
5
Concluding Remarks and Related Work
We presented a typing system that, under some conditions, permits to synthesise a
choreography (represented as global type) from a set of end-point types (represented
as local types). The synthesised global type is unique and well-formed; moreover, its
projections are equivalent to the original local session types. We have shown safety and
progress properties for the local session types. Finally, the derivatives of local types
which form a choreography can also be assigned a global type (subject reduction).
14
Related work. A bottom-up approach to build choreographies is studied in [11]; this
work relies on global and local types, but uses local and global graphs. A local graph
is similar to a local type while a global graph is a disjoint union of family of local
graphs. We contend that global types are more suitable than global graphs to represent
choregraphies; in fact, differently from the approach in [11], our work allows us to reuse
most of the theories and techniques based on multiparty global types.
Our work lies on the boundary between theories based on global types (e.g. [1, 5, 7,
9]) and the ones based on the conversation types [3]. Our work relies on the formalism
of global types, but uses it the other way around. We start from local types and construct
a global type. We have discussed the key elements of the global types in § 3.
Conversation types [3] abandon global views of distributed interactions in favour of
a more flexible type structure allowing participants to dynamically join and leave sessions. The approach in [6] fills the gap between the theories based on session types and
those based on behavioural contracts [4] (where the behaviour of a program is approximated by some term in a process algebra). We are also inspired from [12], where session
types are viewed as CCS-like “projections” of process behaviours. The approach of considering local types as processes is similar to ours. However, the theory of [12] is based
on a testing approach. The connectedness conditions for a choreography given in [2] is
similar to our notion of well-formed global type.
Future work. We aim to extend the framework so that global types can be constructed
from session types which features name passing and restriction. We also plan to refine
the theory and use it in a methodology so that if a choreography cannot be synthesised,
the designers are given indications on why this has failed. Finally, we are considering
implementing an algorithm from the rules of Fig. 2 and Fig. 3, and integrate it in an
existing tool [10] implementing the framework from [1].
References
1. L. Bocchi, K. Honda, E. Tuosto, and N. Yoshida. A theory of design-by-contract for distributed multiparty interactions. In CONCUR, 2010.
2. M. Bravetti, I. Lanese, and G. Zavattaro. Contract-driven implementation of choreographies.
In TGC, 2008.
3. L. Caires and H. T. Vieira. Conversation types. In ESOP, 2009.
4. S. Carpineti, G. Castagna, C. Laneve, and L. Padovani. A formal account of contracts for
web services. In WS-FM, 2006.
5. G. Castagna, M. Dezani-Ciancaglini, and L. Padovani. On global types and multi-party
sessions. In FMOODS/FORTE, 2011.
6. G. Castagna and L. Padovani. Contracts for mobile processes. In CONCUR, 2009.
7. P.-M. Deniélou and N. Yoshida. Dynamic multirole session types. In POPL, 2011.
8. K. Honda, V. T. Vasconcelos, and M. Kubo. Language primitives and type discipline for
structured communication-based programming. In ESOP, 1998.
9. K. Honda, N. Yoshida, and M. Carbone. Multiparty asynchronous session types. In POPL,
2008.
10. J. Lange and E. Tuosto. A modular toolkit for distributed interactions. In PLACES, 2010.
11. D. Mostrous, N. Yoshida, and K. Honda. Global principal typing in partially commutative
asynchronous sessions. In ESOP, 2009.
12. L. Padovani. On projecting processes into session types. MSCS, 22:237–289, 2012.
15
A
Additional Definitions
In this section, we give the definitions of the accessory functions used in the main
sections of the paper.
A.1
Linearity
Definition 4 (T ( )).
s→r:a
µχ
T (s → r : ahei.G ) =
T (µ χ.G ) =
T (G )
T (G )
◦
◦
T (G + G 0 ) =
T (G | G 0 ) =
T (G ) T (G 0 )
T (G ) T (G 0 )
T (G ; G 0 ) = T (G ) ? T (G 0 )
T (0) = T (χ) = •
The function T ( ) returns a channel environment corresponding to a global type.
A.2
Projections
Definition 5 ( ] ).
if P = ∑i∈I ai ?ei .Pi0 and Q = ∑ j∈J a j ?ei .Q0j
and ∀i ∈ I. ∀ j ∈ J. ai 6= a j and I, J 6= ∅
L
L
if P = i∈I ai !ei .Pi0 and Q = j∈J a j !ei .Q0j
P]Q =
and ∀i ∈ I. ∀ j ∈ J. ai 6= a j and I, J 6= ∅
0
0
ahei.(P ] Q ), if P = ahei.P0 and Q = ahei.Q0 ∈ {!, ?}
P,
if P ≡ Q
⊥,
otherwise
P + Q,
P ⊕ Q,
The function merges the behaviour of a participant in different choice branches. In the
first two cases, it merges two guarded internal (resp. external) choices, if their sets of
guard channels are disjoint. In the third case, the function merges the continuation of
both processes, if both are guarded by the same prefix. Note that it is a partial function,
e.g. it might be the case that a participant behaves differently in two branches without
being aware of which branch was chosen, in which case the projection of that participant
is undefined.
16
Definition 6 (Substitution). The substitution P[Q/R], where R is 0.
a!e.(P0 [Q/R])
if P = a!e.P0
a?e.(P0 [Q/R])
if P = a?e.P0
P1 [Q/R] ⊕ P2 [Q/R] if P = P1 ⊕ P2
P[Q/R] = P1 [Q/R] + P2 [Q/R] if P = P1 + P2
µ x.(P0 [Q/R])
if P = µ x.P0
Q
if P = R = 0
P
otherwise
Substitution is used in the projection map.
A.3
Splitting Systems
Omitted rules in Fig. 3:
[com]
Ψ; Θ ` S0 | S m Ω
Ψ; Θ ` S | S0 m Ω
[com]
Ψ; Θ ` (S | T ) | T 0 m Ω
Ψ; Θ ` S | (T | T 0 ) m Ω
Definition 7 ( ). Ω Ω0 holds if and only if ∀n ∈ P (Ω) ∪ P (Ω0 ) either
Ω(n) ≡ Ω0 (n) or
(Ω(n) ≡ ∑ ai ?ei .Pj and Ω0 (n) ≡ ∑ a j ?ej .Pj )
i∈I
j∈ j
where ∀i ∈ I. ∀ j ∈ J. ai 6= a j .
The boolean function holds only if two maps Ω differ wrt external choice.
Definition 8 ( t ).
Ω = Ω0 t Ω1 is defined only if Ω0 Ω1 holds, in which case
∀n ∈ P (Ω0 ) ∪ P (Ω1 ) . Ω(n) = merge(Ω0 (n), Ω1 (n))
where
(
P
if P ≡ Q
merge(P, Q)
P + Q if P ≡ ∑i∈I ai ?ei .Pj and Q ≡ ∑ j∈ j a j ?ej .Pj
The function t merges two Ω maps, if holds.
Definition 9 ( − ).
L
(k,k0 )∈K ak !ek .(Pk − Qk0 )
P−Q = ∑(k,k0 )∈K ak ?ek .(Pk − Qk0 ) + ∑n∈N an ?en .Pn
0,
where K = {(i, j) ∈ I × J | ai = a j }
and
if P ≡
L
i∈I ai !ei .Pi
L
j∈J a j !ej .Q j
if P ≡ ∑i∈I ai ?ei .Pi and Q ≡ ∑ j∈J a j ?ej .Q j
if P ≡ 0 and Q ≡ 0, or Q = ε,
N = {n ∈ I | ∀ j ∈ J. an 6= a j }.
17
and Q ≡
The function − computes the first part of a split behaviour, given the original
behaviour, e.g. S(n) and its prefix in Ω(n). The case for external choices keeps the
branches from the original behaviour which do not appear in Ω(n). The rationale is that
even if some branches are never “chosen” in the system, they might still induce e.g.
races and therefore they need to be taken into account in the main system.
Definition 10 ( % ).
P0
P0
P % Q = 0,
P,
⊥,
if P ≡ i∈I ai !ei .Pi and Q ≡ j∈J a j !ej .Q j
if P ≡ ∑i∈I ai ?ei .Pi and Q ≡ ∑ j∈J a j ?ej .Q j
if P ≡ 0 and Q ≡ 0,
if Q = ε,
otherwise
L
with P0 defined as follows
(
Pi % Q j with (i, j) ∈ K
P0 =
⊥
L
if ∀(i, j)(k, l) ∈ K . Pi % Q j ≡ Pk % Ql
otherwise
where K = {(i, j) ∈ I × J | ai = a j }
The function % computes the second part of a split behaviour. Essentially, it returns the “rest” of a behaviour after Ω(n). Note that if Ω(n) is a branching behaviour,
then the rest must be the same in all branches (since only one behaviour can be returned), e.g. S(n) % Ω(n) = ⊥, if
S(n) = a!e.b!e ⊕ c!e.d!e and
A.4
Ω(n) = a!e.ε ⊕ c!e.ε
Results
Definition 11. There is a race in S if and only if there is S =⇒ S0 such that ∃ a ∈ R(S0 )
such that
∃ {n, m} ∈ P (S) :
a ∈ R(S(n)) and a ∈ R(S(m)) or
18
a ∈ R(S(n)) and a ∈ R(S(m))
B
Definitions Used in the Proofs
Definition 12 (Connected - R).
– Two participants are connected in a system S if (n, m) ∈ RS
(n, m) ∈ RS ⇐⇒ C (S(n)) ∩ C (S(m)) 6= ∅ or ∃n0 . (n, n0 ) ∈ RS ∧ (n0 , m) ∈ RS
– Two participants are connected in a global type G if (n, m) ∈ RG
(n, m) ∈ RG ⇐⇒ C (G n ) ∩ C (G m ) 6= ∅ or ∃n0 . (n, n0 ) ∈ RG ∧ (n0 , m) ∈ RG
Definition 13 (Projections with queues).
e · G 0 a ,
if G = ∗ → a : rhei.G 0
G 0 a ,
if G = s → r : ahei.G 0
G1 a ]G2 a , if G = G1 + G2
G a = Gi a ,
if G = G1 | G2 and a 6∈ C (G j ), i 6= j ∈ {1, 2}
G
·
G
,
if
G = G1 ; G2
1 a
2 a
[],
if ∗ → r : ahei 6∈ G
⊥,
otherwise
Definition 14 ( ] ).
(
ρ1 , if ρ1 = ρ2
ρ1 ] ρ2 =
⊥, otherwise
C
Proofs for Theorem 1 (Decidability)
Typability is decidable.
Proof. The typing systems is decidable from the fact that the ready set of a system, the
number of participants, and their behaviours are finite. Here, we show that the number
of behaviour unfoldings needed to type a system is also finite.
Let (non-recursive) behaviour context C [ ] defined as follows
C [ ] ::=
M
ai !ei .Ci [ ] |
i∈I
∑ ai ?ei .Ci [ ]
| [ ] | µ x.P | x
i∈I
and (possibly recursive) behaviour context C0 [ ] defined as follows
C0 [ ] ::=
M
i∈I
ai !ei .Ci0 [ ] |
∑ ai ?ei .Ci0 [ ]
| [ ] | µ x.C0 [ ] | x
i∈I
Let unfoldi (P) be the ith unfolding of P, defined as follows
unfoldi (P) =
unfold1 (unfoldi−1 (P))
L
i∈I ai !ei .unfold1 (Pi )
unfold1 (P) = ∑i∈I ai ?ei .unfold1 (Pi )
P[µ x.P/x]
19
i>1
L
if P ≡ i∈I ai !ei .Pi
if P ≡ ∑i∈I ai ?ei .Pi
if P ≡ µ x.P
The need for unfolding occurs whenever a recursive participants interact with another participant, while not all the participants feature directly a recursive behaviour.
In this case, we need to unfold some participants (rule [eq]), then use rules [⊕], [+], [.],
and/or [0] until rule [µ] is applicable. Note that rules [ ;], [ | ] and [ρ] cannot be used under
recursion.
Consider the following system
S = S0 | S1
where
S0 = n1 C1 µ x.C10 [x] | . . . | nj C j µ x.C0j [x]
S1 =
]nj+1 [µ x.C0j+1 [x]]
| ... |
nk [µ x.C0j+k [x]]
(C.1)
(C.2)
S l, S0 6l, and there is exactly one n ∈ P (S) such that
S ≡ n[S(n)] | T
and
T 6l
Let Ci for j ≤ i ≤ j + k be the empty context, we can rewrite S such that
S ≡ ∏ ni Ci µ x.Ci0 [x]
I = {i | 1 ≤ i ≤ j + k}
i∈I
Given S as above, we define |i|x to be the smallest k such that Ci [x] is a sub-tree
of unfoldk (Ci0 [x]), |i|x = ⊥ if there is no such k. Note that |i|x must be smaller than
the length of Ci [0] (since recursion is guarded). If one |i|x is not defined, then S is not
typable.
We also defined M = max{|i|x | i ∈ I}, and K(i) = M − |i|x . We can unfold each
behaviour so that all of them are unfolded to the same extent, let
S∗ ≡ ∏ ni unfoldK(i) (Ci µ x.Ci0 [x] )
i∈I
We show that
A ; Γ ; C ` S∗ I G ⇒ A ; Γ ; C ` S I G
By definition of unfold( ), and since C [ ] does not contain recursive definition, we
have
S∗ ≡ ∏ ni Ci unfoldK(i) (µ x.Ci0 [x])
(C.3)
i∈I
≡ ∏ ni Ci Ci0 Ci0 . . .Ci0 µ x.Ci0 [x] . . .
(C.4)
i∈I
Where in (C.4), Ci0 [ ] has been unfolded K(i) times. It is easy to see that S∗ is typable
if
∏ ni
0 0
Ci Ci Ci . . .Ci0 [0] . . .
i∈I
and
∏ ni
0
Ci [0]
i∈I
are typable themselves, note that rule [eq] does not need to be used to unfold the lefthand side system, since it is recursion free; and there is exactly one recursion less in the
right hand side.
In fact, if we would unfold (C.3) once more, we would not get more chances to type
S∗ since it would amount to add the sub-derivation of the right-hand side to one of the
left-hand side.
20
D
Proofs for Lemma 2 (Uniqueness)
If A ; Γ ; C ` S I G and A ; Γ ; C ` S I G 0 then G ≡ G 0 .
Proof. The proof is by case analysis. We show that every time one rule from Fig.2 is
applicable, either no other rule is applicable, or the derivation produces an equivalent
global type.
Due to their syntactic restriction and the condition S6l, the cases for rules [.], [⊕], and
[µ] are straightforward. In addition, the cases for rule [+] is easy since it does not affect
G . The cases for rules [ X ] and [0] are trivial.
The case for rule [eq] follows from the fact that associativity and commutativity in
S do not affect G . In addition, if one unfold behaviour once more, we have the result
since µ χ.G ≡ G [µ χ.G /χ].
The interesting part of the proof is to show that [ | ] and [ ;] are mutually exclusive.
In fact, if [ | ] is applicable, [ ;] cannot be used because ~ and ↔Θ must be total on
Ψ and Θ 6= ∅. If a system S could be separated in two sub-system by [ | ], these two
conditions could not hold. If [ ;] is applicable, it means that it is not possible to split the
participant in two totally independent sub-systems, and therefore [ | ] is not applicable.
Finally, observe that by Lemma 16 the split is unique.
E
Proofs for Theorem 3 (Well-formedness)
If A ; Γ ; C ` S I G then • `G and G is projectable.
Proof. The proof is by induction on the derivation A ; Γ ; C ` S I G . We make a case
analysis on the last rule used.
Case [.]. We have
G = s → r : ahei.G 0
S = s[a!e.P] | r[a?e.Q] | S0
and
– WF. We show that we have
∀ n1 → n2 : ∈ R(G 0 ) . {s, r} ∩ {n1 , n2 } 6= ∅
by contradiction. By IH, we know that
{a} ∪ A ; Γ ; C s → r : a ` s[P] | r[Q] | S0 I G 0
If we had G 0 ≡ n1 → n2 : bhe0 i.G0 + G1 , with ni 6= s and ni 6= r with i ∈ {1, 2},
then we would have
S0 ≡ n1 [b!e0 .P00 ⊕ P10 ]n2 [b?e0 .Q00 + Q01 ] | S00
which is in contradiction with the premise S0 6l.
By Lemma 25, the result above and since C s → r : a is defined, we have • `G .
– Projection. By Def. 1, we have that G s = a!e.G 0 s , G r = a?e.G 0 r , and G n =
G 0 n , for s 6= n 6= r.
21
Case [⊕]. We have
G = G0 + G1 and S = s[P ⊕ Q] | S0
– WF. Observe that we have that all the guard channels are disjoint by definition of
processes. We have to show that
∀(n1 , n2 ) ∈ R(G0 ).∀(n01 , n02 ) ∈ R(G1 ) . n1 = n01
i.e. that for all prefixes in G0 and G1 , s is the sender. In other words, the participant
who makes the internal choice must be the same in all the branches. If that was not
the case, we would have
S0 ≡ n1 [b!e.Q0 ⊕ Q1 ] | n2 [b?e.Q2 + Q3 ] | S00
which is in contradiction with the premise S0 6l.
By Lemma 25, the result above and C `G0 and C `G1 by IH, we have • `G .
– Projection. We have to prove that
∀n ∈ P (G0 + G1 ). G0 n ]G1 n 6= ⊥
By IH, we have that both G0 n and G1 n exist. There are two cases where the
projection does not exist (i 6= j ∈ {0, 1} in the following):
• Gin = a!e.P and G jn = b?e0 .Q. This cannot happen by definition of behaviours,
i.e. such projections could only come from behaviours of the form (a!e.P) ⊕(b?e0 .Q),
which is not syntactically correct.
• Gi n = ∑k∈K ak ?ek .Pk and G j n = ∑k0 ∈K 0 ak0 ?ek0 .Pk0 , with ak = ak0 for some k
and k0 . This cannot happen by definition of behaviours, i.e. such projections
could only come from behaviours of the form a?e.P0 + a?e.P1 which is not a
syntactically correct behaviour. The same reasoning applies for internal choice.
Note that we have G s = G0 s ⊕ G1 s .
Case [+]
– WF. By induction hypothesis, we have C `G and by Lemma 25, we have • `G .
– Projection. By induction hypothesis.
Case [ | ] We have
G = G0 | G1 and S = S0 | S1
– WF. We have to show that
P (G0 ) ∩ P (G1 ) = ∅ and C (G0 ) ∩ C (G1 ) = ∅
By definition of systems, we know that there cannot be two participants of the same
name in a same system, since S | S0 is a system we have that
P (S0 ) ∩ P (S1 ) = ∅
and by Lemma 27, we have P (G0 ) ∩ P (G1 ) = ∅. By Lemma 29, the premise A0 ∩
A1 , we have C (G0 ) ∩ C (G1 ) = ∅. We have C `G0 and C `G1 by assumption, with
the result above and Lemma 25, we have • `G .
22
– Projection. For all n ∈ P (G ), G n is defined by IH and since P (G0 ) ∩ P (G1 ) = ∅.
Case [ ;]. We have
G = G0 ; G1 and C T (G0 ) and split(S) = (S0 , S1 )
– WF. We have to show that
∀ n1 → n2 : ∈ R(G1 ) . ∃N1 6= N2 ∈ FO (G0 ) . n1 ∈ N1 ∧ n2 ∈ N2
(E.1)
∀N ∈ FP (G0 ) . ∃N 0 ∈ FP (G1 ) . N ∩ N 0 6= ∅
(E.2)
and
We know that split(S) = (S0 , S1 ) therefore there is Ψ; Θ ` S m Ω coherent. We
show (E.1) first. We first show that
s → r : a ∈ R(G1 ) ⇐⇒ {s, r} ∈ Θ
We start with
ni → nj : ∈ R(G1 ) ⇐⇒ S1 (ni ) | S1 (nj ) l
(⇒) If ni → nj : ∈ R(G1 ) then we must have G1 = ((ni → nj : ahei.G2 + G3 ) | G4 ) ; G5
(by definition of R). And by Lemma 26, this implies that S1 (ni ) = a!e.P ⊕ P0 and
S1 (nj ) = a?e.Q + Q0 . Thus we have S1 (ni ) | S1 (nj ) l.
(⇐) If S1 (ni ) | S1 (nj )l, we must have S1 (ni ) = a!e.P ⊕ P0 and S1 (nj ) = a?e.Q + Q0
and since G1 is well-formed by IH, we have the required result by Lemma 26 and
the definition of R.
Now, let us show that
S1 (ni ) | S1 (nj ) l ⇐⇒ {ni , nj } ∈ Θ
(E.3)
(⇒) Since the processes interact, we know that they are 6= 0. Moreover, since they
appear in the second part of the split, the following rule must be applied so that
S1 (ni ) = S(ni ) % Ω(ni ) and S1 (nj ) = S(nj ) % Ω(nj ) ( 6= ⊥ by assumption) with
Ω(ni ) and Ω(nj ) ending with ε.
[ε]
ni ∈ N, nj ∈ M
ni [S1 (ni )] | nj [S1 (nj )] l
Ψ · N\{n} · M\{m}; Θ ` S m Ω
Ψ · N · M; Θ · {ni , nj } ` ni [S1 (ni )] | nj [S1 (nj )] | S m Ω · ni : ε · nj : ε
which gives us the expected result. Note that the rule [rem] cannot be applied because
the processes interact with each other we do not have n[P] | S 6l.
(⇐) If {ni , nj } ∈ Θ then, the rule [ε] must also be applied (the axiom cannot be
reach with Θ 6= ∅), since the rule requires that ni [S1 (ni )] | nj [S1 (nj )] l, we have
the expected result.
Let us now show that
∀{ni , n j } ∈ Θ . ∃Ni 6= N j ∈ FO (G0 ) . ni ∈ N j ∧ n j ∈ N j
23
(E.4)
First we show that
N ∈ FO (G0 ) ⇒ ∃!N 0 ∈ Ψ : N ∩ N 0 6= ∅
(E.5)
By Lemma 20, we have that ∀N ∈ FO (G0 ) . RG0 is total on N. By Lemma 21, we
have that (n, m) ∈ RG0 ⇒ (n, m) ∈ ~. By Lemma 12, we have that (n, m) ∈ ~ ⇒
{n, m} ∈ N ∈ Ψ. Therefore we have
N ∈ FO (G0 ) ⇒ ∃N 0 ∈ Ψ : N ∩ N 0 6= ∅
and N 0 is unique since Ψ is a set of pairwise disjoint sets.
We finalise the proof of (E.1), by showing (E.4) by contradiction. If we had
∃{n1 , n2 } ∈ Θ .∃N ∈ FO (G0 ) . {n1 , n2 } ∈ N
by Lemma 11, we have that ∃N1 6= N2 ∈ Ψ : n1 ∈ N1 ∈ Ψ and n2 ∈ N2 ∈ Ψ which
is contradiction with (E.5).
Let us now show (E.2), i.e.
∀N ∈ FP (G0 ) . ∃N 0 ∈ FP (G1 ) : N ∩ N 0 6= ∅
We first show that
FP (G ) = Ψ
By definition of FP , we have that
G=
∏
GN with GN 6≡ GN0 | GN00 .
N∈FP (G )
Therefore, by Lemma 20, RGN is total on P (GN ) for each N. By Lemma 21, we
have that ~ is total on each P (GN ) as well, and by Lemma 12, we have
N ∈ Ψ ⇐⇒ P (GN ) = N
Now, we show that
∀N ∈ Ψ. ∃N 0 ∈ FP (G1 ). N ∩ N 0 6= ∅
by contradiction, if we had
∃N ∈ Ψ.N ∩ P (G1 ) = ∅
by Lemma 11, we must have N 6= M ∈ Ψ such that {n, m} ∈ Θ and n ∈ N and m ∈ M.
Since n ∈ N is also in Θ, we must have S1 (n) | S1 (m)l which implies that S1 (n) ≡ 0
and therefore n ∈ P (G1 ), by Lemma 27. Thus N ∩ P (G1 ) 6= ∅
From (E.1), (E.2), the fact that by assumption C T (G0 ) is defined, and Lemma 25
we have • `G .
– Projection. By IH, we have that for all n G0 n and G1 n exist, thus we have that
G0 n [G1 n /0] is defined for all n. Note that since • `G0 , we have #(FO (G0 )) > 1
(cf. E.1). Therefore, by Lemma 23, fv(G1 ) = ∅. In addition, by Lemma 22, we
have bv(S) = ∅ (since the split(S) by assumption), and by Lemma 24, bv(G0 ) =
∅. Thus, every branch in G0 ends with 0, and all the projections of G0 end with 0.
24
Case [µ] We have
G = µ χ.G0 and S = n1 [µ x1 .P1 ] | . . . | nk [µ xk .Pk ] and
– WF. We have to show that
χ ∈ fv(G 0 ) ⇒ #FO (G 0 ) = 1
we can apply Lemma 23 and we have the result directly. Observe that the recursion
is prefix guarded since ∃ i, j : ni [Pi ] | mj [Pj ] l holds.
– Projection. By induction hypothesis.
Case [ X ] We have
G = χ and ∀n ∈ P (S).S(n) = xn
We have to show that CC(χ), which follows directly from the premises of rule [ X ].
Case [0] Trivial.
Case [ρ] We have
G = ∗ → r : ahei.G 0
S = r[a?e.Q] | a : e · ρ | S0
and
– WF. We show that we have
∀ n1 → n2 : ∈ R(G 0 ) . ni = r
with i ∈ {1, 2}
by contradiction. By IH, we know that
{a} ∪ A ; Γ ; C s → r : a ` r[Q] | a : ρ | S0 I G 0
If we had G 0 ≡ n1 → n2 : bhe0 i.G0 + G1 , with ni 6= r with i ∈ {1, 2}, then we would
have
S0 ≡ n1 [b!e0 .P00 ⊕ P10 ]n2 [b?e0 .Q00 + Q01 ] | S00
which is in contradiction with the premise S0 6l.
By Lemma 25, the result above, and since C ∗ → r : a is defined, we have • `G .
– Projection. By Def. 1, we have that G r = a?e.G 0 r , and G n = G 0 n , for n 6= r.
F
Proofs for Theorem 4 (Subject Reduction)
λ
If A ; ◦ ; C ` S I G , S −→ S0 , and C (λ) 6∈ C then A ; ◦ ; C ` S0 I G 0
Proof. There are three cases to consider.
n[a!e]
a·e
Case 1. S ≡ S1 | S2 with S1 −→ S10 and S2 −→ S20
In this case, S must have the following form
S ≡ n[a!e.P ⊕ P0 ] | a : ρ | T
and S0 ≡ n[P] | a : ρ · e | T
and by Lemma 2 we have that A ; ◦ ; C ` S0 I G 0 .
25
n[a?e]
e·a
Case 2. S ≡ S1 | S2 with S1 −→ S10 and S2 −→ S20
In this case, S must have the following form
S = n[a?e.P + P0 ] | a : e · ρ | T
and
S0 = n[P] | a : ρ | T
and by Lemma 2 we have that A ; ◦ ; C ` S0 I G 0
X
Case 3. S −→ S in this case, we have A ; ◦ ; C ` S0 I G 0 trivially since S0 ≡ S.
Case 4 (unfolding). If S ≡ n[µ x.P] | S0 and S0 ≡ n[P[µ x.P/x]] | S0 then G = µ χ.G0 ,
and ∀m ∈ P (S0 ) . S0 (m) ≡ µ x.Q. We have the result by unfolding all the participants in
S0 wiht rule [eq], so to have G 0 = G0 [µ χ.G0 /χ].
In the other direction, we have the result by folding all the participants in P (S0 ).
Case 5 (commutativity and associativity). If S ≡ S0 , then we have the result with rule
[eq].
Lemma 2. The following holds:
1. A ; ◦ ; C ` s[a!e.P ⊕ P0 ] | a : ρ | T I G ⇒ A ; ◦ ; C ` s[P] | a : ρ · e | T I G 0
2. A ; ◦ ; C ` r[a?e.P + P0 ] | a : e · ρ | T I G ⇒ A ; ◦ ; C ` r[P] | a : ρ | T I G 0
with → : a 6∈ C in 1 and 2.
Corollary 1.
G n = (a!e.G0 n ) ⊕ G1 n ⇒ G 0 n = G0 n
and
G n = a?e.G0 n ⇒ G 0 n = G0 n
Proof. We show that 1 and 2 hold by contradiction.
Case 1. Assume
A ; ◦ ; C ` s[a!e.P ⊕ P0 ] | a : ρ | S I G is derivable.
A ; ◦ ; C ` s[P] | a : ρ · e | S I G 0 is not.
Take S ≡ r[a?e.Q] | S0 , ρ = [], and S0 6l. We must have the following sub-derivation for
G
[.]
..
.
A ; ◦ ; C s → r : a ` s[P] | a : ρ | r[a?e.Q] | S0 I G0
A ; ◦ ; C ` s[a!e.P] | a : ρ | r[a?e.Q] | S0 I s → r : ahei.G0
Consider the non-derivable judgement for
[ρ]
(F.1)
G0
⊥
A ; ◦ ; C ∗ → r : a ` s[P] | a : ρ | r[a?e.Q] | S0 I G00
A ; ◦ ; C ` s[P] | a : e | r[a?e.Q] | S0 I ∗ → r : ahei.G00
26
(F.2)
Where the rule [ρ] must be applicable here, since the only difference with the above
system is C ∗ → r : a which is defined since → : a 6∈ C.
We have a contradiction here since
s[P] | a : ρ | r[a?e.Q] | S0 is derivable in (F.1).
while
s[P] | a : ρ | r[a?e.Q] | S0 is not derivable in (F.2).
Case 2. Assume
A ; ◦ ; C ` r[a?e.P + P0 ] | a : e · ρ | S I G is derivable.
A ; ◦ ; C ` r[P] | a : ρ | S I G 0 is not.
Take S such that S 6l. We must have the following sub-derivation for G
[ρ]
..
.
A ; ◦ ; C ∗ → r : a ` r[P] | a : ρ | S I G0
A ; ◦ ; C ` r[a?e.P] | a : e · ρ | S I ∗ → r : ahei.G0
(F.3)
This induces a contradiction since we would have the following for G 0
[ρ]
⊥
A ; ◦ ; C ` r[P] | a : ρ | S I G00
(F.4)
Note that the lack of ∗ → r : a in C does not affect since → : a 6∈ C.
G
Proofs for Theorem 5 (Progress)
X
If A ; ◦ ; C ` S I G then S −→ S0 , or ∀ n ∈ P (S) . S(n) = 0, or S −→.
Proof. The proof is by contradiction. If we had A ; Γ ; C ` S I G and S =⇒ S0 , with S0 6l,
S0 6= 0 and ∃n ∈ P (S0 ) : S0 (n) 6= 0. By Theorem 4, we should have A ; Γ ; C ` S0 I G 0 .
Let us take S0 ≡ n[a!e.P] | S00 (with S0 6l). No rule from Fig. 2 is applicable for this
process, and therefore S0 is not typable.
H
Proofs for Theorem 6 (Safety)
If A ; ◦ ; C ` S I G , then S is race free and S =⇒−→ error is not possible.
Proof. No error. The proof is by contradiction. Assume we have A ; Γ ; C ` S I G and
S =⇒ S0 with
S0 ≡ a : e · ρ | r[a?e0 .Q] | S00 and e 6= e0
so that S −→ error. By Theorem 4, we should have A ; Γ ; C ` S0 I G 0 , however, no
rule is applicable for S0 . Indeed we have S l, but e 6= e0 .
27
No Race. Straightforward by contradiction with Theorem 4, the following is not typable
S0 ≡ n[a?e.P] | m[a?e.Q] | S00
due to the condition on C, the premise S 6l in [.], and the fact that the set of channels
must be disjoint in concurrent branches. The other case (i.e. two sends) is similar.
Lemma 3. If A ; Γ ; C ` S I G , then, ∀ s → r : a ∈ FP (G ) : , either
∃!{s, r} ∈ P (S) : a ∈ R(S(s)) and a ∈ R(S(r)) or ∃!(r, a) ∈ P (S)× C (S) : a ∈ R(S(s))
Proof. By straightforward induction on the derivation. Each case follows by definition
of FP (G ), and the premise S 6l.
I
Proofs for Lemma 1 and Theorem 7 (Equivalences)
If A ; ◦ ; C ` S I G then ∀n ∈ S . G n . S(n).
Proof. Let B be a binary relation on processes defined as follows
(P, Q) ∈ B ⇐⇒ A ; Γ ; C ` S I G and ∃n ∈ P (S). G n = P and S(n) = Q
Let us show that B is a simulation.
a!e
– If G n −→ P1 then G n ≡ a!e.P1 ⊕ P2 and by Lemmas 30 and 32 we have that S(n) =
a!e
a!e.Q1 ⊕ Q2 and thus S(n) −→ Q1
Now we have to show that (P1 , Q1 ) ∈ B, i.e.
A ; Γ ; C ` S0 I G 0 with G 0 n = P1 and S0 (n) = Q1
Pose
S0 ≡ n[Q1 ] | a : e |
S(m)
∏
m6=n∈P (S)
by Lemma 2, A ; Γ ; C ` S0 I G 0 , and by Corollary 1, we have G 0n = P1 , as required.
a?e
– If G n −→ P1 then G n ≡ a?e.P1 + P2 and by Lemmas 30 and 32 we have that
a?e
S(n) = a?e.Q1 + Q2 and thus S(n) −→ Q1
Now we have to show that (P1 , Q1 ) ∈ B, i.e.
A ; Γ ; C ` S0 I G 0 with G 0 n = P1 and S0 (n) = Q1
Pose
S ≡ n[a?e.Q1 + Q2 ] | a : e · ρ |
∏
S(m)
m6=n∈P (S)
and
S0 ≡ n[Q1 ] | a : ρ |
∏
S(m)
m6=n∈P (S)
by Lemma 2, A ; Γ ; C ` S0 I G 0 , and by Corollary 1, we have G 0n = P1 , as required.
28
X
– If G n −→ P1 then G n ≡ 0 and P1 = 0, thus n 6∈ P (G ) and by Lemma 27 this means
X
that S(n) = 0, we then have S(n) −→ 0.
Lemma 4. The following holds:
1. If A ; Γ ; C ` n[a!e.P ⊕ P0 ] | S I G then G n ≡ a!e.Q ⊕ Q0
2. If A ; Γ ; C ` n[a?e.P + P0 ] | S I G then either G n ≡ a?e.Q + Q0 , or A ; Γ ; C `
n[P0 ] | S I G
Proof. The proof of 1 is by Lemma 26, and the proof of 2 is by Lemma 26 and rule [+].
If A ; ◦ ; C ` S I G then ∏n∈P (S) n[G n ] ≈ S.
Proof. For this proof we pose T = ∏m∈P (G ) m[G m ] | ∏b∈C (G ) b : G b .
n[a!e]
a·e
S sends. Assume S ≡ S0 | S1 , S0 −→ and S1 −→. We have
S ≡ n[a!e.P ⊕ P0 ] | a : ρ | S00
(I.1)
S0 ≡ n[P] | a : ρ · e | S00
(I.2)
and
n[a!e]
By Lemma 4, G n ≡ a!e.Q ⊕ Q0 , thus G n −→, and by Lemma 34 G a = ρ.
We then have (note that a ∈ C (G ))
T ≡ n[a!e.Q ⊕ Q0 ] | a : ρ |
m[G m ] |
∏
m6=n∈P (G )
b : G b
∏
(I.3)
b6=a∈C (G )
and by definition of −→,
T 0 ≡ n[Q] | a : ρ · e |
m[G m ] |
∏
m6=n∈P (G )
b : G b
∏
(I.4)
b6=a∈C (G )
Let us now show that
A ; Γ ; C ` S0 I G 0 with T 0 ≡
m[G 0 m ] |
∏
m∈P (G 0 )
∏
b : G 0 b
(I.5)
b∈C (G 0 )
By Lemma 2 we know that A ; Γ ; C ` S0 I G 0 , we have that G 0 n ≡ Q, by Corollary 1,
and G 0 a ≡ ρ · e by Lemma 34.
n[a!e]
a·e
T sends. Assume T ≡ T0 | T1 , T0 −→ and T1 −→. We have
T ≡ n[a!e.Q ⊕ Q0 ] | a : ρ |
∏
m[G m ] |
m6=n∈P (G )
∏
b : G b
(I.6)
b6=a∈C (G )
and
T 0 ≡ n[Q] | a : ρ · e |
∏
m6=n∈P (G )
29
m[G m ] |
∏
b6=a∈C (G )
b : G b
(I.7)
n[a!e]
By Lemma 1, we have S(n) −→, and since, by Lemma 28, a ∈ C (G ) ⇒ a ∈ C (S),
a·e
there is a queue a in S. Note that a queue a can always make a transition a : ρ0 −→
0
(regardless of ρ ). By Lemma 34, S(a) = ρ.
Therefore, we must have
S ≡ n[a!e.P ⊕ P0 ] | a : ρ | S00
(I.8)
And by definition of −→, we have
S0 ≡ n[P] | a : ρ · e | S00
(I.9)
Finally, we have
A ; Γ ; C ` S0 I G 0 with T 0 ≡
m[G 0 m ] |
∏0
m∈P (G )
b : G 0 b
∏0
(I.10)
b∈C (G )
by Lemma 2.
n[a?e]
e·a
S receives. Assume S ≡ S0 | S1 , S0 −→ and S1 −→. We have
S ≡ n[a?e.P + P0 ] | a : e · ρ | S00
(I.11)
S0 ≡ n[P] | a : ρ | S00
(I.12)
and
By Lemma 4 and (I.11), we have either
G n ≡ a?e.Q + Q0
(I.13)
A ; Γ ; C ` n[P0 ] | a : e · ρ | S00 I G
(I.14)
or
However, by assumption we have A ; Γ ; C ` S I G , with S as in (I.11), therefore (I.14)
cannot hold by Lemma 5.
e·a
a?e
By Lemma 34, we have that G a −→ and by (I.13), G n −→. By definition of −→,
we have
T ≡ n[a?e.Q ⊕ Q0 ] | a : e · ρ |
m[G m ] |
∏
m6=n∈P (G )
∏
b : G b
(I.15)
b : G 0 b
(I.16)
b6=a∈C (G )
Let us now show that
A ; Γ ; C ` S0 I G 0 with T 0 ≡
∏0
m[G 0 m ] |
m∈P (G )
∏0
b∈C (G )
By Lemma 2 we know that A ; Γ ; C ` S0 I G 0 , we have that G 0 n ≡ Q, by Corollary 1,
and G 0 a ≡ e · ρ by Lemma 34.
n[a?e]
e·a
T receives. Assume T ≡ T0 | T1 , T0 −→ and T1 −→. We have
T ≡ n[a?e.Q ⊕ Q0 ] | a : e · ρ |
∏
m[G m ] |
m6=n∈P (G )
∏
b : G b
(I.17)
b6=a∈C (G )
and
T 0 ≡ n[Q] | a : ρ |
m[G m ] |
∏
m6=n∈P (G )
30
∏
b6=a∈C (G )
b : G b
(I.18)
a?e
By Lemma 1, we have S(n) −→, and by Lemma 34, we have S(a) = ρ, therefore
S ≡ n[a?e.P + P0 ] | a : e · ρ | S00
(I.19)
S ≡ n[P] | a : ρ | S00
(I.20)
and
We now have to show that
A ; Γ ; C ` S0 I G 0 with T 0 ≡
∏
m[G 0 m ] |
m∈P (G 0 )
∏
b : G 0 b
(I.21)
b∈C (G 0 )
as before, we have G 0 n = P by Corollary 1 and G 0 a = ρ by Lemma 34.
X
X
End. If S −→ S, then S ≡ 0 and G ≡ 0, and vice versa if T −→ T .
Lemma 5. If
A ; Γ ; C ` r[a?e.P + P0 ] | a : e · ρ | S I G
is derivable then
A ; Γ ; C ` r[P0 ] | a : e · ρ | S I G 0
is not derivable.
Proof. Assume P0 ≡ b?e.P00 . We show that we must have
G ≡ (∗ → r : ahei.G0 | G1 ) ; G2
and the derivation of G must have the following form
..
.
A0 ; ◦ ; C ` r[P0 ] | a : ρ0 | S00 I G0
[.]
A0 ; ◦ ; C ` r[a?e.P0 ] | a : e · ρ0 | S00 I ∗ → r : ahei.G0
[+]
A0 ; ◦ ; C ` r[a?e.P0 + P00 ] | a : e · ρ0 | S00 I ∗ → r : ahei.G0
..
.
[|]
[ ;]
A1 ; ◦ ; C ` S01 I G1
A ; ◦ ; C ` r[a?e.P0 + P00 ] | a : e · ρ0 | S0 I ∗ → r : ahei.G0 | G1
A ; Γ ; C ` r[a?e.P + P0 ] | a : e · ρ | S I G
..
.
G2
where we must have
– S l or G1 = G2 = 0 and S0 l or S01 ≡ 0.
– split(S) = (r[a?e.P0 + P00 ] | a : e · ρ0 | S0 , ) (the second part of the split does not
matter)
– b 6∈ R(S) (by Lemma 7)
– A0 ∩ A1 = ∅ and b 6∈ A2 because of rules [ | ] and [+]
31
Now let us discuss a derivation for G 0 . Since we have b 6∈ R(S), we must have
split(r[P0 ] | a : e · ρ | S) = (S1 , S2 ) such that S1 (r) = P0 , and S1 (a) = e · ρ, if the split
does exist. If it does, we have Ψ; Θ ` r[P0 ] | a : e · ρ | S m Ω such that ∀N ∈ Ψ . r 6∈ N.
Therefore, the split for the rest of the system is the same as in the other derivation.
Again, we can divide the system using [ | ] if need be such that we get
A0 ; ◦ ; C ` r[P00 ] | a : e · ρ0 | S00 I ∗ → r : ahei.G0
with S00 6l therefore no rule is applicable for this judgement, and the derivation does not
exist.
Lemma 6. Let S a system such S 6l, and a, b 6∈ R(S) the following is not derivable
T ≡ s1 [a!e.P1 ] | s2 [b!e0 .P2 ] | r[a?e.Q + b?e0 .Q0 ] | S
Proof. We show this by contradiction. Given T as above, the only rule applicable is
[+] on r either selecting the branch on a or on b. Therefore, the following should be
derivable
[.]
[+]
..
.
A ; Γ ; C ` s1 [P1 ] | s2 [b!e0 .P2 ] | r[Q] | S I G
A ; Γ ; C ` s1 [a!e.P1 ] | s2 [b!e0 .P2 ] | r[a?e.Q] | S I s1 → r : ahei.G
A ; Γ ; C ` T I s1 → r : ahei.G
and, we must have a, b ∈ A and s1 [P1 ] | r[Q] | Sl. And the other derivation as the form:
[.]
[+]
..
.
A ; Γ ; C ` s1 [a!e.P1 ] | s2 [P2 ] | r[Q] | S I G
A ; Γ ; C ` s1 [a!e.P1 ] | s2 [b!e0 .P2 ] | r[b?e0 .Q0 ] | S I s2 → r : bhe0 i.G 0
A ; Γ ; C ` T I s2 → r : bhe0 i.G 0
where we have a, b ∈ A and s2 [P2 ] | r[Q] | S l. However, this is clearly in contradiction
with Theorem 2, i.e.
s1 → r : ahei.G 6≡ s2 → r : bhe0 i.G 0
Lemma 7. Let S a system such S 6l, and a, b 6∈ R(S) the following is not derivable
T ≡ s1 [a!e.P1 ] | b : e0 · ρ | r[a?e.Q + b?e0 .Q0 ] | S
Proof. The proof is similar to the one of Lemma 7 where b : e0 · ρ replaces s2 [b!e0 .P2 ].
J
Proofs for Theorem 8 (Completeness wrt G )
If • `G and G is projectable, then there is G 0 ≡ G such that A ; Γ ; C ` ∏n∈P (G ) n[G n
] I G 0.
32
Proof. By Lemma 8, with P = P (G ) and Γ = ◦ since G is closed by assumption.
Lemma 8. Let Proj(G , P) = ∏n∈P n[G n ] with P (G ) ⊆ P, and if G n = 0, then n is not
in P.
If C `G , G is projectable, and ∀ χ ∈ fv(G ) . ∀ n ∈ P . ∃ (n, χ) : χ ∈ Γ then
C (G ) ; Γ ; C ` Proj(G , P) I G 0 with G ≡ G 0
Proof. We show this by induction on the structure of G . Let
S ≡ ∏ G n
SQ ≡
n∈P
∏
G n
A = C (G )
n∈P\Q
G ≡ s → r : ahei.G0 + G1 .
By definition of projection, we have
S ≡ s[a!e.G0 s ⊕ G1 s ] | r[a?e.G0 s + G1 s ] | S{s,r}
We can apply rules [⊕], [+] (twice) and [.] in order to have the result, i.e.
by IH
[.]
A ; Γ ; Ca ` s[G0 s ] | r[G0 s ] | S{s,r} I G0
by IH
A ; Γ ; C ` s[a!e.G0 s ] | r[a?e.G0 s ] | S{s,r} I G 0
A ; Γ ; C ` Proj(G1 , P) I G1
A; Γ; C ` S I G
with G 0 = s → r : ahei.G0 , and Ca = C s → r : a note that the later is defined by
C `G . Observe that S{s,r} 6l otherwise that would mean that ∃ s0 → r0 : b ∈ R(G0 ) such
that {s, r} ∩ {s0 , r0 } = ∅ which is in contradiction with C `G . Finally, it is obvious that
a∈A
and
R(Gi r ) ⊆ A
i ∈ {0, 1}
since A = C (G ).
G ≡ G0 | G1 .
We have S of the form, by definition of projections (and well-formedness)
S ≡ Proj(G0 , P0 ) | Proj(G1 , P1 ) with P0 ∩ P1 = ∅
Note that since #(FO (G )) > 1, we have fv(G ) = ∅, therefore, by IH, we have
Ai ; ◦ ; C ` Proj(Gi , P) I Gi0
and
Gi ≡ Gi0 and Ai = C (Gi )
We have the result by applying rule [ | ]:
by IH
by IH
A0 ; ◦ ; C ` Proj(G0 , P0 ) I G1
A1 ; ◦ ; C ` Proj(G1 , P1 ) I G1
A1 ∪ A2 ; Γ ; C ` S I G
By well-formedness we have, A0 ∩ A1 = ∅, and Lemmas 28 and 28 guarantee that each
Ai is large enough.
33
G ≡ µ χ.G 0 .
By definition of projections, we have
S ≡ ∏ n[µ χ.G 0 n ]
n∈P
Since G 0 is prefix-guarded, there must be s, r ∈ P (G 0 ) such that
s[µ χ.G 0 s ] | r[µ χ.G 0 r ] l
Therefore, rule [µ] is applicable here
A ; Γ0 ; C ` s[G 0 s ] | r[G 0 r ] |
∏
n[G 0 n ] I G 0
n∈P\{s,r}
0
A ; Γ ; C ` s[µ χ.G s ] | r[µ χ.G 0 r ] | S{s,r} I G
where
Γ0 = Γ · (s, χ) : χ), (r, χ) : χ · ΓS
and
ΓS such that ∀ n ∈ P (S{s,r} ).Γ(n, χ) = χ
The rest follows by induction hypothesis.
G = χ. Then, we have
S ≡ ∏ n[χ]
n∈P
By assumption, we have ∀ n ∈ P . ∃ (n, χ) : χ ∈ Γ, hence we can apply rule
are done.
[X]
and we
G ≡ G0 ; G1 . Then we have
S ≡ ∏ G0 n [G1 n /0]
n∈P
Let us show that split(S) = (S0 , S1 ) and
S0 ≡ ∏ G0 n = Proj(G0 , P) and
n∈P
S1 ≡ ∏ G1 n = Proj(G1 , P)
n∈P
Since G is well-formed, we have Ψ; Θ ` S m Ω coherent if we pose
Ψ = {N|N ∈ FP (G0 )}
and
Θ = {{s, r}|s → r : a ∈ R(G1 )}
For each Gi a top level concurrent branch of G0 , we have that
{P (Gi )}; ∅ ` Proj(Gi , P (Gi )) m Ωi
is derivable by Lemma 15, since Proj(Gi , P (Gi )) is typable by IH. In addition, we have
Gi n = Ωi (n)[0/ε].
By construction, we have
Ψ0 ; Θ ` Proj(G0 , P) m Ω0
with Ψ0 = {N 0 |∃N ∈ Ψ : N 0 ⊆ N}.
34
with ∀n ∈ P . Ω0 (n) = ε
Finally, we have ∀n ∈ P . S(n) % Ω(n) 6= ⊥ since the same suffix G1 n is added to
each branch of a behaviour, and S(n) % Ω(n) = G1 n . We have the required result by IH
and rule [ ;].
G ≡ 0.
We have
S ≡ ∏ n[0]
n∈P
and the results holds by rule [0].
K
K.1
Accessory Results
Linearity
Lemma 9. If C s → r : a is defined, then either
1.
→ : a 6∈ C,
s→r:a
2. C( → : a) =
,
C0
∗→r:a
3. C( → : a) =
, or
C0
4. ∃ → r : and → s : ∈ C
Proof. The proofs of 1 and 2 follow directly from Def. 2 and the definition of C( ). The
proof of 3 follows by definition of C( ) and the fact that, by Def. 1, there cannot be an
output dependency from s0 → r0 : b and ∗ → r : a since
s → r : b 6≺OO ∗ → r : a
and
s → r0 : b 6≺IO ∗ → r : a
since s 6= ∗ 6= r
Therefore, ∗ → r : a must be the first prefix with label → : a in C. The proof of 4
follows from the Def. 2. In fact, since whenever the sender/receiver are different on
two nodes with common channel there must be two dependency relation we have the
following cases. In the following, we assume that there is no prefix on a in the ellipsis.
If only the senders are different, i.e. the following appears on a path in C
s0 → r : a . . . s → r : a
then we have s0 → r : a ≺II s → r : a and we must have at least a node between the two
such that, e.g. s0 → r : a ≺OO s0 → s : b ≺IO s → r : a and we have the result. If only
the receivers are different, we have
s → r0 : a . . . s → r : a
35
and we have s → r0 : a ≺OO s → r : a and we must have at least one node between the
two such that s → r0 : a ≺IO r0 → r : b ≺II s → r : a, and we have the result. If both
sender and receiver are different, i.e.
s0 → r0 : a . . . s → r : a
then we need two nodes c1 and c2 such that, (i) c1 = → r : b, otherwise there would
be a ≺II relation in the input dependency, (ii) c2 = → s : c otherwise there would be a
≺IO relation in the output dependency (note that ≺OO is only defined if the channels are
the same). In fact, there must an input dependency, e.g.
s0 → r0 : a ≺IO r0 → s1 : b1 ≺IO s1 → s2 : b2 ≺IO s2 → r : b3 ≺II s → r : a
and an output dependency, e.g.
s0 → r0 : a ≺IO r0 → s : c ≺IO s → r : a
Observe that, in the first chain, we have s2 → r : b3 , and r0 → s : c in the second.
Actually, the shortest (input) chain when both pair of participants are different is
s0 → r0 : a ≺IO r0 → s : c1 ≺IO s → r : c2 ≺II s → r : a
where we also have r0 → s : c1 ≺IO s → r : a, for the output chain. Notice that in this
case we have r0 → s : c1 and s → r : c2 in C.
Corollary 2. If linearity holds on C and ∗ → n : a ∈ C, then
∗→n:a
C( → : a) =
C0
Lemma 10. If C ∗ → r : a ` G then C ` G and if Γ ; C ∗ → r : a ; S ` G I then
Γ ; C ; S ` G I.
Proof. We have to show that if ∗ → r : a is involved in a input or output dependency,
then there is another dependency between the same two nodes without ∗ → r : a. Note
first that by Lemma 9, we know that ∗ → r : a 6∈ C, therefore, if there is a need to have
a chain of the form
∗ → r : a ≺ . . . ≺ s0 → r0 : a
this need disappears with ∗ → r : a (dependencies are needed only between nodes with
a common channel). Thus, if ∗ → r : a is the first node in a dependency chain, then the
result holds trivially. Observe that ∗ → r : a 6≺OO s → r0 : b and s → r0 : b 6≺OO ∗ → r : a
for any a, b since ∗ 6= s, for the same reason, we have s → r : b 6≺IO ∗ → r : a.
Finally, we have the following cases, where the left hand side describes the dependency involving a and the right hand side shows that the dependency between the two
external nodes still exists without the node on a.
s → r : b ≺II ∗ → r : a ≺II s0 → r : c
⇒
s → r : b ≺II s0 → r : c
s → r : b ≺II ∗ → r : a ≺IO r → s0 : c
⇒
s → r : b ≺IO r → s0 : c
36
K.2
Splitting systems
Lemma 11.
{n, m} ⊆ Θ ⇒ ∃N 6= M ∈ Ψ : n ∈ N ∧ m ∈ M
Proof. Direct from rules [ε] and [ax].
Lemma 12. If A ; Γ ; C ` S I G and Ψ; Θ ` S m Ω is derivable and coherent, then
(n, m) ∈ ~ ⇐⇒ {n, m} ⊆ N with N ∈ Ψ
Proof. (⇐) follows directly from the fact that the judgement is coherent. (⇒) The proof
is by contradiction. Assume that there is (n, m) ∈ ~ with n ∈ N and m ∈ M, where
N 6= M ∈ Ψ (assuming C (Ω(n)) ∩ C (Ω(m)) 6= ∅, without loss of generality).
Let us consider the following judgement:
Ψ0 · N · M; Θ ` n[a!e.P1 ] | m[a?e.P2 ] | S0 m Ω0 · n : a!e.π · m : a?e.ϕ
Note that rule [ε] is not applicable here since we do not have n : ε and m : ε. Looking
at the rules we see that only two applications of [sync] would introduce n : a!e.π and
m : a?e.ϕ.
For this rule to be applicable we must have S0 ≡ n0 [a?e.P3 ] | m0 [a!e.P4 ] | S00 , and
0
n ∈ N, m0 ∈ M.
By definition of split and since rule ; must be applied for A ; Γ ; C ` S I G to
hold, we should be able to derive
A ; Γ ; C ` n[a!e.Q1 ] | m[a?e.Q2 ] | n0 [a?e.Q3 ] | m0 [a!e.Q4 ] | S00 I G
However, this is not derivable due to the obvious race on channel a!
Lemma 13. If
split(r[P] | a : ρ | T ) = (S00 , S10 ) such that
S00 (r) = P0
and
S00 (a) = ρ0
then
split(r[a?e.P] | a : e·ρ | T ) = (S0 , S1 ) such that
S0 (r) = a?e.P0
and
S0 (a) = e·ρo
In addition, ∀n ∈ P (T ) : Si (n) = Si0 (n), for i ∈ {0, 1}.
Proof. Straightforward (runtime rule)
Lemma 14. If A ; Γ ; C ` s[a!e.P ⊕ P0 ] | a : ρ | T I G and
split(s[a!e.P ⊕ P0 ] | a : ρ | T ) = (S1 , S2 )
such that
S1 (s) = a!e.P0 ⊕ P00
and
S1 (a) = ρ
then we have
split(s[P] | a : ρ · e | T ) = (S10 , S20 ) such that
S10 (s) = P0
and
and ∀n 6= s ∈ P (S).S1 (n) = S10 (n) and S2 (n) = S20 (n), ditto ∀a ∈ C (S).
37
S10 (a) = ρ · e
Proof. We must have a judgement of the form
Ψ; Θ ` s[a!e.P ⊕ P0 ] | a : ρ | T m Ω · s : a!e.P0 ⊕ P00 · r : a?e.Q0 ⊕ Q00 · a : ρ
Since the system is derivable, we have to use [q] until the queue is empty (otherwise
linearity would not be preserved). Then only use [sync] to remove the send on a, therefore
both ρ and the send should be on the same part of the split. After reduction, we have
the following judgement
Ψ; Θ ` s[P] | a : ρ · e | T m Ω · s : P0 · r : a?e.Q0 ⊕ Q00 · a : ρ · e
Since r was able to receive e from s before, it must be able to receive it from the queue
as well (note that there is not restriction on wrt Ψ with action on queues).
Lemma 15. If A ; Γ ; C ` S I G then {P (S)}; ∅ ` S m Ω and Ω(n)[0/ε] ≡ S(n).
Proof. Straightforward induction on the derivation of {P (S)}; Θ ` S m Ω. Note that
for each rule of the split, there is a rule in the inference system, whose premises are
always weaker. Also, [ε] is not applicable (since Θ is empty) and Ω keeps track of
everything that happens (module the branches which are not taken).
Lemma 16. If A ; Γ ; C ` S I G and Ψ; Θ ` S m Ω is coherent, then ∀ Ψ0 , Θ0 , Ω0 such
that Ψ0 ; Θ0 ` S m Ω0 is coherent: Ψ = Ψ0 , Θ = Θ0 and Ω = Ω0 .
Proof. First, recall that by Lemma 12:
(n, m) ∈ ~ ⇐⇒ {n, m} ⊆ N with N ∈ Ψ
We first show that Ψ (such that the split is coherent) is unique. Note that because of
condition (4.2) and the fact that A ; Γ ; C ` S I G , it is not possible to have a coherent
judgement with Ψ0 like Ψ except for two sets in Ψ being merged (subdivided) in Ψ0 .
Indeed, the number of interacting pairs of participants is fixed in S. The only changes
one can do in Ψ are as follows
1. Add n in N ∈ Ψ (with n not in Ψ).
2. Remove n in N ∈ Ψ (with n in Ψ).
3. Permute n ∈ N ∈ Ψ with m ∈ M ∈ Ψ.
We now show that any of these changes makes the judgement not coherent.
Case 1. Assume Ψ = N · Ψ0 and Ψ0 = {n} ∪ N · Ψ0 . This means that Ω0 (n) 6= 0 and
0
Ω (n) 6= ε (otherwise ~ would not be total on {n} ∪ N), and Ω(n) = 0, or Ω(n) = ε, or
n only received from queues. This means that n interact with participants in Ω0 but not
in Ω, which means that at some point in the derivation of Ω we have n[ ] | S0 6l while
we have n[ ] | S0 l in the derivation for Ω0 . However, since G is typable, there cannot
be races in the systems and therefore, the interaction between pair of participants must
be the same in both derivations of Ω and Ω0 .
Case 2. Assume that Ψ = {n} ∪ N · Ψ0 and Ψ0 = N · Ψ0 , the case is similar to the
previous one, we have that at some point in the derivation of Ω0 we have n[ ] | S0 6l
while we have n[ ] | S0 l in the derivation for Ω.
38
Case 3. Assume that Ψ = {m} ∪ N · {n} ∪ M · Ψ0 and Ψ0 = {n} ∪ N · {m} ∪ M · Ψ0 .
We know that
∀n0 ∈ N . C (Ω(n0 )) ∩ C (Ω(m)) = ∅
and ∀m0 ∈ M . C (Ω(m0 )) ∩ C (Ω(n)) = ∅
since the original judgement is coherent and the system typable. This implies that we
cannot have ~ total neither on {n} ∪ N nor on {m} ∪ M. Therefore the judgement is not
coherent.
We now consider the changes that one can make on Ω. Note that changes on Ω
must be done on pairs of participants since they always interact by pair (except for
those which interact only with queues but that has no effect on Ψ or Θ).
1. There is n1 , n2 such that Ω(ni ) is a prefix of Ω0 (ni ). A pair of elements in Ω0 can
only be longer if one merges two in Ψ, which is possible (see above).
2. There is n1 , n2 such that Ω0 (ni ) is a prefix of Ω(ni ). A pair of elements in Ω0 can
only be shorter if two a set in Ψ is subdivided into two sets which is not allowed
either.
We consider changes that can be made on Θ:
1. Add {n, m} to Θ, then, that pair will not allow the derivation to reach the axiom,
unless sets in Ψ are sub-divided, which is not possible.
2. Remove {n, m} from Θ, then, there will be a pair missing to reach the axiom (plus
↔Θ might not be total any more); unless two sets in Ψ are merged, but this is not
possible.
3. Permutation in Θ would only possible if one could permute participants in sets of
Ψ which, by above, is not possible.
K.3
Others
Lemma 17. If G 6≡ G0 | G1 and • `G then RG is total on P (G ).
Proof. We show this by induction on the structure of G .
G ≡ s → r : ahei.G 0 . By definition (s, r) ∈ RG and by IH RG 0 is total on P (G 0 ). Since
G is well-formed we have
∀ n1 → n2 : ∈ R(G 0 ) . {n1 , n2 } ∩ {s, r}
Thus there is (s, ni ) ∈ RG or (r, ni ) ∈ RG and we have the required result by definition
of RG and Def. 1.
G ≡ G0 + G1 . By IH, RGi is total on P (Gi ) for i ∈ {0, 1}. Since G is well-formed we
have
∀ s → r : a ∈ R(G ).∀ s0 → r0 : b ∈ R(G 0 ) . s = s0 ∧ a 6= b
i.e. s = s0 ∈ Gi , and we have the required result by definition of RG and Def. 1.
G ≡ s → r : ahei.(G0 | G1 ). By IH RGi is total on P (Gi ). Since G is well-formed we
must have s ∈ P (Gi ) and r ∈ P (G j ) with i 6= j ∈ {0, 1}. We have the required result
since we have (s, r) ∈ RG by definition of RG and Def. 1.
39
G ≡ G0 ; G1 . Observe that by IH and by definition of FP , we have that ∀N ∈ FP (Gi ) . RGi
is total on N, with i ∈ {0, 1}.
Since the projection is defined as G n = G0 n [G1 n /0], we have that
∀ N0 × N1 ⊆ FP (G0 ) × FP (G1 ) : RG is total on N0 ∪ N1 if there is n ∈ N0 ∩ N1
Since RG is a transitive relation, let us define a transitive relation on the intersection of
sets of participants from G0 and G1 :
N0 ∩ N1 6= ∅ or
(N0 , N1 ) ∈ W
⇐⇒
∃ (M0 , M1 ) : (N0 , M1 ) ∈ W and (M0 , N1 ) ∈ W
It is easy to see that RG is total on any N0 ∪ N1 whenever (N0 , N1 ) ∈ W, thus
[
RG is total on
N0 ∪ N1
(N0 ,N1 )∈W
Let us show that
(N0 , N1 ) ∈ FP (G0 ) × FP (G1 ) ⇒ (N0 , N1 ) ∈ W
Since G is well-formed we have
∀ s → r : ∈ R(G1 ) . ∃N1 6= N2 ∈ FO (G0 ) . s ∈ N1 ∧ r ∈ N2
(K.1)
∀N ∈ FP (G0 ) . ∃N 0 ∈ FP (G1 ) . N ∩ N 0 6= ∅
(K.2)
and
Therefore, by (K.1) and the fact that s → r : ∈ R(G1 ) ⇒ {s, r} ⊆ N ∈ FP (G1 ), we have
∀N1 ∈ FP (G1 ) . ∃N0 6= N00 ∈ FP (G0 ) : N0 ∩ N1 6= ∅ and N00 ∩ N1 6= ∅
(K.3)
By (K.2), we have
∀N0 ∈ FP (G0 ) . ∃N1 ∈ FP (G1 ) : N0 ∩ N1 6= ∅
(K.4)
Now assume that there is (N0 , N1 ) ∈ FP (G0 ) × FP (G1 ) such that (N0 , N1 ) 6∈ W, this is
contradiction with (K.3) and (K.4).
G ≡ µ χ.G . By IH
Other cases. The cases where G = 0 or G = χ are trivial.
Lemma 18. If A ; Γ ; C ` S I G and • `G then the following holds
∀n ∈ P (S) : C (S(n)) ⊆ C (G n )
Proof. Straightforward induction on the validation rules.
Lemma 19. If A ; Γ ; C ` S I G and RG is total on P (G ) then RS is total on P (S).
Proof. By Lemmas 17 and 18 and the definition of RS .
40
Lemma 20. If • `G then ∀N ∈ FO (G ) . RG is total on N.
Proof. By Lemma 17 and the definition of FO .
Lemma 21. If A ; Γ ; C ` S I G0 ; G1 , C `G0 , Ψ; Θ ` S m Ω is coherent, and split(S) 6=
⊥ then
∀n, m ∈ P (S) : (n, m) ∈ RG0 ⇒ (n, m) ∈ ~
Proof. Straightforward by definitions of RG0 and ~. Note that external choice branches
which do not appear in Ω do not appear in G0 either.
Lemma 22. If bv(S) 6= ∅ then split(S) = ⊥.
Proof. If bv(S) 6= ∅, we must have
S ≡ n[P] | S0 where µ x.P0 is a suffix of P
The result follows from the fact that there is no rule in Fig. 3 which “removes” recursive definition. Therefore, it not possible to derive a split whenever there is a recursion
definition in the system to be split.
Lemma 23. If A ; Γ ; C ` S I µ χ.G and χ ∈ fv(G ) then #(FO (G )) = 1.
Proof. The proof follows from the fact that the context Γ is emptied each time the rule
[ | ] is used in the derivation (this rule is the only one introducing concurrent branches).
In addition, for the axiom [ X ] to be used in the derivation one must have ( , ) : χ ∈ Γ.
Therefore, the only way one could have #(FO (G )) > 1 (i.e. at least two concurrent
branches in G ) is if χ does not appear in G .
Lemma 24. If A ; Γ ; C ` S I G and bv(S) = ∅ then bv(G ) = ∅
Proof. By straightforward induction on the rules of Fig. 2.
Lemma 25. If C `G then • `G
Proof. This follows from the fact that •C is always defined.
Lemma 26. If A ; Γ ; C ` S I G and G is well-formed
G ≡ ((n → m : ahei.G1 + G2 ) | G3 ) ; G4 ⇐⇒ S ≡ n[a!e.P ⊕ P0 ] | m[a?e.Q + Q0 ] | S0
Proof. (⇒) Assume that
A ; Γ ; C ` S I ((n → m : ahei.G1 + G2 ) | G3 ) ; G4
is derivable. We show that either a rule introducing the corresponding operator is applicable or that an equivalent G can be inferred.
41
– If A ; Γ ; C ` S I G 0 ; G4 is derivable then we have either split(S) = ⊥ then G4 = 0,
thus
A; Γ; C ` S I G0
or split(S) 6= ⊥, in this case, we must have
split(S) = (S1 , S10 )
with
A ; ◦ ; C ` S1 I G 0 and A ; ◦ ; C T (G 0 ) ` S10 I G4
– Assume G 0 = G 00 | G3 , if A ; Γ ; C1 ` S1 I G 0 is derivable then we must have either
S1 ≡ S2 | S3 , A = A1 ∪ A2 , and A1 ∩ A2 = ∅ such that
A1 ; ◦ ; C ` S2 I G 00
and A2 ; ◦ ; C ` S3 I G3
are derivable, or S3 = 0 and G3 ≡ 0.
– Assume G 00 = G 000 + G2 , we must have either
S2 = n[P0 ⊕ P00 ] | S4
(K.5)
and
A1 ; ◦ ; C ` n[P0 ] | S4 I G 000
and
A1 ; ◦ ; C ` n[P00 ] | S4 I G2
are derivable, or P00 ≡ 0 and G2 ≡ 0.
– Assume G 000 = n → m : ahei.G1 , we must have
P0 ≡ a!e.P
and
S4 ≡ m[a?e.Q + Q0 ] | S0
(K.6)
and
A1 ; ◦ ; C ` n[P0 ] | m[a?e.Q + Q0 ] | S0 I G1
derivable.
Putting (K.5) and (K.6) together, we have that
S ≡ n[a!e.P ⊕ P0 ] | m[a?e.Q + Q0 ] | S0
(⇐) Assume
A ; Γ ; C ` n[a!e.P ⊕ P0 ] | m[a?e.Q + Q0 ] | S0 I G
(K.7)
We show that either a rule introducing the corresponding operator is applicable or that
an equivalent G can be inferred.
– Either split(S) = ⊥ in which case G ≡ G 0 ; 0 or
split(S) = (S1 , S10 )
and we must have G4 6≡ 0 and
A ; Γ ; C ` S1 I G
with
S1 (n) = a!e.P0 ⊕ P00
and
S1 (m) = a?e.Q0 + Q00
and S10 (n) = P1
and
S10 (m) = Q1
such that
a!e.P ⊕ P0 ≡ (a!e.P0 ⊕ P00 )[P1 /0] and
42
a?e.Q + Q0 ≡ (a?e.Q0 + Q00 )[Q1 /0]
– Either there is
S = n[a!e.P0 ⊕ P00 ] | m[a?e.Q0 + Q00 ] | S1 | S2
G 0 = G 00 | G3 and A = A1 ∪ A2 , and A1 ∩ A2 = ∅
A1 ; ◦ ; C ` n[a!e.P0 ⊕ P00 ] | m[a?e.Q0 + Q00 ] | S1 I G 00
and
A2 ; ◦ ; C ` S2 I G3
where S1 6l (this is a sound assumption, since one could apply
times as necessary to obtain this), or G3 ≡ 0.
– Either there is G 00 ≡ G 000 + G2 such that
[ ;]
and
[|]
as many
A ; ◦ ; C1 ` n[a!e.P0 ] | m[a?e.Q0 + Q00 ] | S1 I G 000
and
A ; ◦ ; C1 ` n[P0 ] | m[a?e.Q0 + Q00 ] | S1 I G2
or G2 = 0.
– For
A ; ◦ ; C1 ` n[a!e.P] | m[a?e.Q + Q0 ] | S1 I G 000
to be derivable, we must have G 000 ≡ n → m : aheiG1 .
Putting all the pieces together, we have the required result.
Lemma 27. If A ; Γ ; C ` S I G then
∀n ∈ P (S). S(n) 6= 0 : n ∈ P (S) ⇐⇒ n ∈ P (G )
and
∀n ∈ P (S). S(n) = 0 ⇐⇒ n 6∈ P (G )
Proof. Straightforward.
Lemma 28. If A ; Γ ; C ` S I G then C (G ) ⊆ C (S).
Proof. Straightforward induction on the derivation.
Lemma 29. If A ; Γ ; C ` S I G then C (S) ⊆ A.
Proof. Straightforward induction on the derivation.
Lemma 30. If G is well-formed and projectable then
– if G n ≡ a!e.P ⊕ Q then there is a branch of G such that the first prefix on n is
n → m : ahei.
– if G n ≡ a?e.P + Q then there is a branch of G such that the first prefix on n is
m → n : ahei.
Proof. By definition of Projection.
43
Lemma 31. If A ; Γ ; C ` S | a : e · ρ I G is derivable then for each branch in G the
first prefix on a is ∗ → n : ahei.
Proof. Follows from the fact that for a common channel, [ρ] must be used before [.] (see
Lemma 2).
Lemma 32. If A ; Γ ; C ` S I G and there is a branch in G such that n → m : ahei is
the first prefix on n (resp. m), then S(n) = a!e.P ⊕ P0 (resp. S(m) = a?e.Q + Q).
Proof. Straightforward.
Lemma 33. If fv(S) = ∅, and S −→ S0 , then fv(S0 ) = ∅, i.e. reduction of systems
preserves closeness.
Proof. Follows directly from the semantics of the calculus.
Lemma 34. If a ∈ C (G ) then
A ; Γ ; C ` a : ρ | S I G ⇐⇒ then G a = ρ
In addition, a 6∈ C (G ) ⇒ ρ = [].
Proof. Follows from rules [ρ] and [0].
44
| 6cs.PL
|
Graph Based Recommendations: From Data
Representation to Feature Extraction and Application
arXiv:1707.01250v1 [cs.IR] 5 Jul 2017
Amit Tiroshi, Tsvi Kuflik, Shlomo Berkovsky, Mohamed Ali (Dali) Kaafar
Abstract
Modeling users for the purpose of identifying their preferences and then personalizing services on the basis of these models is a complex task, primarily
due to the need to take into consideration various explicit and implicit signals,
missing or uncertain information, contextual aspects, and more. In this study, a
novel generic approach for uncovering latent preference patterns from user data
is proposed and evaluated. The approach relies on representing the data using
graphs, and then systematically extracting graph-based features and using them
to enrich the original user models. The extracted features encapsulate complex
relationships between users, items, and metadata. The enhanced user models
can then serve as an input to any recommendation algorithm. The proposed approach is domain-independent (demonstrated on data from movies, music, and
business recommender systems), and is evaluated using several state-of-the-art
machine learning methods, on different recommendation tasks, and using different evaluation metrics. The results show a unanimous improvement in the
recommendation accuracy across tasks and domains. In addition, the evaluation
provides a deeper analysis regarding the performance of the approach in special
scenarios, including high sparsity and variability of ratings.
Keywords:
Recommender systems, graph-based recommendations, feature extraction,
graph metrics.
1. Introduction
Recommender systems aim at helping users find relevant items among a large
variety of possibilities, based on their preferences (Adomavicius and Tuzhilin,
2005). In many cases, these personal preferences are inferred from patterns
that emerge from data about the users’ past interactions with the system and
with other users, as well as additional personal characteristics available from
different sources. These patterns are typically user-specific and are based on
the metadata of both the users and items, as well as on the interpretation of
the observed user interactions (Kobsa, 2001; Zukerman and Albrecht, 2001).
Eliciting user preferences is a challenging task because of issues such as changes
in user preferences, contextual dependencies, privacy constraints, and practical
Preprint submitted to Elsevier
July 6, 2017
data collection difficulties (Ricci et al., 2011). Moreover, the collected data
may be incomplete, outdated, imprecise, or even completely inapplicable to
the recommendation task at hand. In order to address these issues, modern
recommender systems attempt to capture as much data as possible, and then,
apply data mining and other inference techniques to elicit the desired preferences
Cantador et al. (2015). Several techniques can be applied for the pattern-mining
task, among which are techniques originated in machine learning and statistics,
e.g., clustering and regression, or those that evolved in information retrieval and
user modeling (Mobasher, 2007).
Regardless of the technique exploited by a recommender system, it is inherently bound by the available user data and the features extracted/elicited from
it. One major question that arises in this context is how to engineer1 meaningful
features from often noisy user data? Features may be manually engineered by
domain experts. This approach is considered expensive and non-scalable because of the deep domain knowledge that is necessary, the creativity required
to conceive new features, and the time needed to populate and evaluate the
contribution of the features. A notable example of this challenge is provided by
the Netflix Prize winning team, in their recap: “while major breakthroughs in
the competition were achieved by uncovering new features underlying the data,
those became rare and very hard to get” (Koren, 2009).
An alternative to manual feature engineering is automatic feature engineering, which is a major area of research in machine learning (Guyon et al., 2006),
particularly in the domains of image recognition (Nixon, 2008; Due Trier et al.,
1996) and text classification (Scott and Matwin, 1999). So far, automatic feature engineering has mainly focused on either algebraic combinations of existing features, e.g., summation or averaging of existing features (Markovitch and
Rosenstein, 2002), finding domain specific feature generators, e.g., for character
recognition in image processing (Nixon, 2008; Due Trier et al., 1996), or eliciting
latent features as in the SVD (Klema and Laub, 1980) and PCA (Wold et al.,
1987) methods. The algebraic approaches for automatic feature engineering
manage to produce large quantities of features; however, the relationships between the engineered features and the underlying patterns in the data are often
not interpretable (Kotsiantis et al., 2006). For example, if averaging the ratings
for items with the sum of some other arbitrary feature improves predictions,
the reasons for this improvement will not necessarily be clear. Similarly, the
latent feature discovery techniques do not provide sufficient insight regarding
the representation or meaning of those features (Koren et al., 2009).
In this work, a novel framework is proposed that uses graph-based representation properties to generate additional features from user modeling data
of recommender systems, with the objective of improving the accuracy of the
1 Feature engineering is sometimes also referred to in the literature as feature extraction,
generation, and discovery, depending on the field of research. In this paper, it broadly refers
to the task of adding new features to a dataset, regardless of the manner in which it is done
(e.g., manual vs. automatic).
2
generated recommendations. The proposed framework is underpinned by the
idea of examining a tabular recommender system’s data from the graph theorybased perspective, which represents entities and their relationships as a graph
and allows the extraction of a suite of new features computed using established
graph-based metrics. The extracted features encapsulate information about the
relationships between entities in the graph and lead to new patterns uncovered
in the data. In most cases, they are also interpretable; for example, a node’s
degree (number of edges to other nodes) represents the importance of the node
in the graph, while the path length between two nodes communicates their relatedness (the shorter the path - the more related are the nodes). The approach
is domain-independent and can be applied automatically.
The proposed framework offers several benefits for automatic feature extraction. Given a new dataset, it is usually impossible to determine a-priori which
graph representations will yield the most informative set of features for the recommendation generation. Thus, the proposed framework provides a systematic
method for generating and assessing various graph representations, their contribution to the newly extracted features, and, in turn, to the accuracy of the
generated recommendations. Additionally, since the number of nodes and relationship types in each graph representation is different, an exhaustive method
of distilling the possible graph metrics from each representation is proposed.
Two case studies are conducted to gather extensive empirical evidence and
demonstrate how graph features supplement existing feature sets, improve the
accuracy of the recommendations, and perform adequately as stand-alone outof-the-box features. The case studies answer the following questions:
• How does the use of graph features affect the performance of rating predictions and recommendation generation in different domains and tasks?
• How are the recommendations affected by the sub-graph and its representation used to generate the graph features?
Multiple datasets, multiple machine learning mechanisms, and multiple evaluation metrics are used across the case studies, in order to demonstrate the effectiveness of the approach. Overall, the results show that graph-based representation and automatic feature extraction allow for the generation of more precise
recommendations. A comparison across various graph schemes is conducted
and the justification for systematic feature extraction is established. Hence,
this work concludes the line of research presented earlier in (Tiroshi et al., 2013,
2014a,b) and provides a complete picture that validates the applicability of the
proposed graph-based feature generation approach to recommender systems.
The rest of the paper is structured as follows. Next, the necessary background is provided, and related work is described. Then, the graph representation and graph-based feature extraction process is formalized, and its advantages
and disadvantages are discussed. Two case studies demonstrating the contribution of the graph-based features to the recommendation process are then
presented. Through these, the overall performance of the framework, as well as
3
the performance of certain graph representations and feature subsets, is evaluated. Finally, the implications of the findings are discussed, together with the
suggested future work.
2. Background and Related Work
Graphs have been exploited in recommender system for many tasks, mainly
due to their ability to represent many entities of different types and their relationships in a simple data structure that offers a broad variety of metrics and
reasoning techniques. In this section we provide a general background on the
use of graphs in recommender systems, followed by specific aspects of graph
representation in recommender systems and feature engineering.
2.1. Graph-Based Recommender Systems
In recent years, especially since social networks were identified as a major
source for freely available personal information, graphs and networks data structures have been used as tools for user modeling, especially since they combine
different entities and links into one simple structure capturing the links between
the entities. This section aims at giving the readers an idea about how graph
techniques are used in graph-based user modeling and recommender systems.
Given the vast amount of studies (a search for “graph-based” and “recommender
systems” in Google Scholar yielded 225 results for 2016 alone), this is only a
brief presentation of recent studies and not an in-depth survey.
What was clearly noticeable was that most of the graph-based representations were defined for a specific problem, in specific domains, and in many cases
they applied variants of random walk as the only graph feature used for recommendations. (Pham et al., 2015) suggested to use a simple graph representation
for recommending groups to users, tags to groups, and events to users, using
a general graph-based model called HeteRS, while considering the recommendation problem as a query dependent node proximity problem. (Portilla et al.,
2015) applied random walk for predicting YouTube videos watching, on a graph
composed of videos as nodes and the link representing the appearance of videos
in the recommendation lists. (Wu et al., 2015) suggested the use of a heterogeneous graph for representing contextual aspects in addition to items and users,
and used random walk for context-aware recommendation. (Lee et al., 2015) applied random walk for finding top-K paths from an origin user node to an item
node in a heterogeneous graph, as a way for identifying the best items for recommendation. Still, these works used the PageRank algorithm for the purpose
of generating the recommendations. (Lee et al., 2013) used an enhanced version
of personalised PageRank algorithm to recommend items to target users and
proposed to reduce the size of the graph by clustering nodes and edges. (Shams
and Haratizadeh, 2016) also applied personalised PageRank over the user/item
graph augmented with pairwise ranking for items recommendation.
In addition to the wide use of random walk based algorithms, there is a variety of task-specific representations and metrics. It is interesting to note that
4
even for a specific task, a variety of approaches was suggested. For instance,
for song/playlist recommendations, (Benzi et al., 2016) combined graph-based
similarity representation of playlists and songs with classical matrix factorization to improve the recommendations. (Ostuni et al., 2015) took a different
approach and suggested to use tags and sound description represented as a
knowledge graph, from which similarity of nodes was extracted using a specific
metric they defined. (Mao et al., 2016) suggested using graph representation for
music tracks recommendations, where they represented by graphs the relative
preferences of users, e.g., pair-wise preference of tracks. They used the graph as
a representation for user preferences for tracks and calculated the probability
of a user liking a track based on the probability that s/he likes the in-linked
tracks.
Some researchers suggested to use graph representations as an alternative to
the classical collaborative and hybrid recommenders. (Moradi et al., 2015) used
clustering of graph representation of users and items for generating a model
for item- and user-based collaborative filtering. (Bae et al., 2015) used graphs
for representing co-occurrence of mobile apps, as logged from users mobile devices, and the similarity of user graphs was used for finding a neighborhood
and generation recommendations. (Cordobés et al., 2015) also addressed the
app recommendation problem and explored the potential of graph representation for several variants of recommendation strategies for recommending apps
to users through banners on webpages. (Park et al., 2015) proposed a graph
representation for linking item based on their similarity; hence, having a graph
that links items while the weight on the edges represents their similarity. Users
were linked to items they rated, such that items most similar to the items rated
by the users could be recommended. (Lee and Lee, 2015) suggested an approach
for graph-based representation of the user-item matrix, where links among items
represent the positive user ratings, and use entropy to find the items to recommend to users, thus, introducing serendipity into the recommendation process.
(Hong and Jung, 2016) used affinity between users for creating a user graph,
where users are nodes and edges represent affinity for the purpose of group
recommendations of movies Said et al. (2011).
A highly relevant line of work focuses on enriching recommender systems
dataset with information extracted from graph representation of the data, which
is called MetaPaths. A good recent example study is the work of (Vahedian
et al., 2016). The author suggested to enrich a classical recommender systems
dataset (in their case DBLP authors/papers dataset) with what the so-called
metapath data links extracted from citations network. They added this information to the existing set of features, then applied classical matrix factorization,
and showed an improvement to the results using only the original data. Our
framework can be considered as a generalized variant of (Vahedian et al., 2016),
where a specific set of metrics was extracted from the graph representation of
the data and matrix factorization was applied for recommendation generation
purposes. The studies presented in this work used a variety of metrics, datasets,
and recommendation methods.
Additional applications of graphs for recommendations include domains of
5
cultural heritage, tourism, social networks, and more. (Chianese and Piccialli,
2016) used graphs for representing context evolution in cultural heritage: nodes
modeled states and transitions between the nodes were based on observation
of user behavior Bohnert et al. (2008). (Shen et al., 2016) used graphs for
representing tourist attractions and their similarity, where different graphs could
represent content-based, collaborative, and social relationships. (Jiang et al.,
2016) used graph techniques for trust prediction in social networks. (Godoy
and Corbellini, 2016) reviewed the use of folksonomies, which can be naturally
seen as user-item-tag graphs, in recommender systems. As we see, graphs-based
approaches in user modeling and recommender systems have become highly
popular and there is a growing numbers of tools that enable analysis of large
graphs. We refer an interested reader to (Batarfi et al., 2015) and (Zoidi et al.,
2015) for recent and encompassing reviews of the area.
2.1.1. Similarity Measurement Using Graphs and their Application
Previous research on recommender systems that use graph representations
focused on measuring the similarity of two entities in the data (user-to-item,
user-to-user, or item-to-item), and tried to associate this with a score or rating (Amatriain et al., 2011). Graph-based similarity measurement is based on
metrics extracted from a graph-based representation (Desrosiers and Karypis,
2011). Two key approaches for measuring similarity using graphs are path-based
and random walk-based.
In the path-based similarity, the distance between two graph nodes can be
measured using the shortest path and/or the total number of paths between
the two. The definition of the shortest path may include a combination of
the number of edges transitively connecting the two nodes in question and the
weights of these edges if exist, e.g., if a user is connected to an item and the
user’s rating for the item as the edge label. Shortest paths can then be computed
for a user node and an item node in question, in order to quantify the extent
to which the user prefers the item. The “number of paths” approach works
similarly, by calculating the number of paths between the two nodes as a proxy
for their relatedness (the more paths, the more related they are). However, this
approach is more computationally intensive.
Random walks can be used to compute similarity by estimating the probability of one node being reached from another node, given the available graph
paths. The more probable it is that the target node can be reached from the
source node, the higher is the relatedness of the two nodes. Random walks can
be either unweighted (equal probability of edges) or weighted (edges having different probabilities based on their label, e.g., rating) (Desrosiers and Karypis,
2011).
Examples of recommendation studies in which the approaches detailed above
were applied can be found in (Li and Chen, 2009; Lösch et al., 2012; Konstas
et al., 2009), as well as in Section 2.1. (Li and Chen, 2009) reducted the recommendation problem was to a link prediction problem. That is, the problem
of finding whether a user would like an item was cast as a problem of finding
6
whether a link exists between the user and item in the graph. A similarity measure between user and item nodes was computed using random walks. Items
were then ranked based on their similarity scores, such that top scoring items
were recommended to users. Using classification accuracy metrics, this approach
was shown to be superior to other non-graph based similarity ranking methods.
A similar walking distance metric was used in (Lösch et al., 2012), complemented by graph structure metrics such as the number of sub trees. These
metrics were used for the purpose of link prediction and property value prediction in RDF semantic graphs, using a learning technique based on an SVM.
Experimental results showed that the graph features varied in their performance
based on the graph structure on which they operated, for example, full versus
partial subtrees. It was also noted that the newly defined features were not
dataset-specific, but could be applied to any RDF graph The graph structures
in the context of RDF are less applicable to those used in the approach proposed in this work, because the recommendation dataset graphs do not follow
a hierarchical model of RDFs. In the presented approach, any feature value
is connected to other features values based on co-occurrence in the dataset,
without the need for matching a predefined structure or scheme.
Finally, (Konstas et al., 2009) developed a graph-based approach for generating recommendations in social datasets like Last.fm. The work focused on
optimizing a single graph algorithm (random walk with restarts) and its parameters, such as the walk restart. The reported results show an improvement in
recommendations using the random walk approach, compared to the baseline
collaborative filtering. In the presented work, random walks on a graph, although with static parameters, are represented by the PageRank score feature.
The above studies are also extended in this work by generalizing the adoption of
graph metrics beyond random walks and their use for similarity measurements,
and they are not bound to specific graph structures, such as RDF trees.
2.1.2. Representing social data and trust using graphs
Other studies involving graph approaches in recommender systems primarily
addressed the context of representing social, semantic, and trust data. In some
studies, only the graph representation was used as the means to query the
data, e.g., neighboring nodes and the weights of edges connecting to them (Ma
et al., 2009), while others utilize both the graph representation and graph-based
reasoning methods (Massa and Avesani, 2007; Quercia et al., 2014).
A survey of connection-centric approaches in recommender systems (Perugini et al., 2004) exemplifies how the data of an email network (Schwartz and
Wood, 1993) and of a co-occurrence in Web documents (Kautz et al., 1997) can
be represented in graphs. The graph representation of the email interactions
between users defines each user as a node and edges connect users, who corresponded via email. In the case of Web documents, people are again represented
as nodes and edges connect people, who are mentioned in the same document.
When these graphs are established, they can be used to answer recommendationrelated queries. In the email graph, a query regarding the closeness of users can
be answered using a similarity or distance metric, such as those mentioned in
7
the previous section. In the Web co-occurrence graph, a query regarding people
sharing interests can be answered by counting their common neighbors (assuming the co-occurrence in the type of Web documents collected is an indicator of
shared interests).
Other graph representation variants are hypergraphs (Berge and Minieka,
1973). They differ from graphs by allowing an edge, denoted by a hyperedge, to
connect with multiple nodes. Hypergraphs have been proposed in the context
of recommendation generation, for the purpose of representing complex associations, such as social tagging (Jäschke et al., 2007; Berkovsky et al., 2007; Bu
et al., 2010; Tan et al., 2011), where a tag is attached to an item by a user. If
the tag, user, and item are represented by nodes, at least two edges are required
to represent the association between the three entities2 . This association can
be represented by a hyperedge connecting the three nodes. In these studies,
similarity metrics, e.g., a modified hypergraph PageRank, are then composed
based on this structure and used for the recommendation generation. Results
presented in (Jäschke et al., 2007) show that the similarity metrics from hypergraphs led to better recommendations than variations that did not utilize the
properties of the hypergraph representation.
Prior works focusing on the means of incorporating trust between users for
the sake of improving the recommendations were surveyed in (O’Donovan and
Smyth, 2005). For example, (Ma et al., 2009) proposed a graph representation
encapsulating trust between users. The representation modeled users as the
graph nodes and the trust relationships between them were reflected by the
weights on the edges. Data extracted from the graph, e.g., who trusts whom
and to what extent, was used in the recommendation process, and it was shown
to improve the generated recommendations. However, the graph was used only
to represent the data and propagate the trust scores.
Another usage of graphs for recommendation purposes is in the case of
geospatial recommendations. Quercia et al. used graphs to find the shortest
path between geographical locations, while also maximizing the enjoyment of
the path for the user (Quercia et al., 2014). Locations were represented as nodes
and connected to each other based on geographical proximity. Nodes were also
ranked based on how pleasant (beautiful, quiet, happy) the locations were. Finally, a route that optimizes the shortness and pleasantness was computed based
on a graph method and recommended to the user. In this work, both graphbased representation and graph theory methods are used for recommendation
generation.
2.2. Feature Engineering for Recommendations
As mentioned at the beginning of the section, another group of related works
that covers automatic feature engineering. According to Guyon et al., “feature
extraction addresses the problem of finding the most compact and informative
2 A single edge between the user and item can be labeled with the chosen tag, but then the
reuse of tags by other users or for other items becomes less comprehensible.
8
set of features, to improve the efficiency or data storage and processing” (Guyon
et al., 2006). Basic features are a result of quantitative and qualitative measurements, while new features can be engineered by combining these or finding
new means to generate additional measurements. In the big data era, the possibilities of engineering additional features, as well as their potential importance,
have risen dramatically.
Feature engineering (also referred to in the literature as feature extraction,
composition, or discovery) can be performed either manually or automatically.
In the manual method, domain experts analyze the task for which the features are required, e.g., online movie recommendation versus customer churn
prediction, and conceive features that may potentially inform the task. The engineering process involves aggregating and combining features already present
in the data, in order to form new, more informative features. This approach,
however, does not scale well because of the need for a human expert, the time it
takes to compose features, and the sheer number of possibilities for the new features (Domingos, 2012). Conversely, automatic feature extraction, the process
of algorithmically extracting new features from a dataset, does scale up well.
Many features can be engineered in a short time using a variety of engineering
methods. Coupling automatic feature engineering with automatic feature selection (Kohavi and John, 1997) (the process of separating between useful and not
useful features) can lead to faster and more accurate recommendation models.
A basic approach for engineering new features from the existing ones is to
combine them using arithmetic functions. In one study that evaluated this
approach, arithmetic functions, such as min, max, average, and others, were
used (Markovitch and Rosenstein, 2002). The study also presented a specific
language for defining features, where the features were described by a set of inputs, their types, construction blocks, and the produced output. A framework
for generating a feature space using the feature language as input was evaluated. The evaluation showed that the framework outperformed legacy feature
generation algorithms in terms of accuracy. The main difference between the
framework presented at (Markovitch and Rosenstein, 2002) and its predecessors was that the framework was generic and applicable to multiple tasks and
machine learning approaches.
Additional automatic feature engineering methods that are domain-specific
were surveyed in (Nixon, 2008; Due Trier et al., 1996) for image recognition and
in (Scott and Matwin, 1999) for text classification purposes. An example of a
feature engineering method for image recognition is quantifying the amount of
skin color pixels in an image in order to classify whether it contains a human
face or not (Garcia and Tziritas, 1999), whereas for text classification a bag-ofwords (frequency of occurrence of each word in a document) can be generated
for every document and used to describe it.
A different suite of methods for eliciting new features, which is also applicable
to recommender systems, is latent features computation. Methods such as SVD
(Klema and Laub, 1980) and PCA (Wold et al., 1987) can be used to compute
new features and support the generation of recommendations by decomposing
the available data into components and matching composing factors, i.e., the
9
latent features. When the data is decomposed and there exists a set of latent
features that can recompose it with a certain error rate, missing features and
ratings can be estimated (Amatriain et al., 2011). Although it has been shown
that this approach successfully improves the accuracy of the recommendations
(Bennett and Lanning, 2007), it is limited in the interpretability of the latent
features found (Koren et al., 2009).
The current work defines an automatic and recommendation task agnostic
feature engineering process, which is based on graph-based representation of
a recommender system data. The details of this process are provided in the
following section.
3. Graph Based Data Modeling for Recommendation Systems
In this section, an approach for enhancing recommendations based on representing the data as a graph is presented. This representation allows a set of
graph algorithms to be applied and a set of graph-related metrics, which offer
a new perspective on the data and allow the extraction of new features, to be
deduced. Following a brief overview of the approach, the structure of recommender system datasets is formalized (Section 3.1). Then a detailed description
of porting data from a classical tabular representation to a graph-based representation is given (Section 3.2). An elaboration of methods for generating
multiple graph representations follows (Section 3.2.2) and finally the process
of exhaustively distilling graph features from these representations is outlined
(Section 3.3).3
The input to the process (illustrated in Figure 1) is a tabular recommender
system dataset and the output is a set of graph-based features capturing the
relationships between the dataset entities from the graph perspective. The first
step deals with the generation of a complete graph representation of the data:
the tabular data is converted into a representation where the dataset entities
are nodes, connected based on their co-occurrence in the data. Next, a set of
partial representations is derived from the complete graph: first the basic representation containing only user and item nodes, and then additional alternative
representations, each with a unique combination of relationships filtered from
the complete graph. The partial representations are passed to the next step,
where the extraction of the graph features is performed. Finally, the newly
generated graph-based features are used to supplement the original features
available in the dataset and this extended data is fed into the recommender
system for the generation of predictions or recommendations.4 In the following
sub-sections the above steps of the feature extraction process are elaborated.
3 An
open
source
package
implementing
the
approach
is
released
at
http://amitti.github.io/GraphRecSys/.
4 Note that although the process of selecting features that are more predictive for the task
at hand (i.e., feature selection) is outside the scope of the propose approach, it is addressed
indirectly by the features extracted from the partial graph representations.
10
Figure 1: Graph modeling and feature extraction flow chart
3.1. The Structure of a Recommender System Dataset
In (Burke, 2007; Ricci et al., 2011), classical recommendation approaches are
categorized into several key groups: collaborative filtering, content based filtering, demographic, knowledge-based, community-based, and hybrid approaches.
We first consider the representation of the input data used by these approaches,
which can be converted into a tabular form as follows:
• In collaborative filtering, the data is represented as a matrix of user feedback on items (matrix dimensions are users×items), where both the users
and the items are denoted by their unique identifiers and the content of
the matrix reflects the feedback of the users for the items, e.g., numeric
ratings or binary consumption logs.
• In content based filtering, the items are modeled using a set of features,
e.g., terms or domain features. Here, the matrix dimensions include the
identifiers of the users, as well as the identifiers of the content features,
and the values represent the preferences of the users for the features. The
model also contains a second matrix with item identifiers and the same
content features. The values in this matrix represent the weights of the
features in each item.
• In demographic recommenders, the demographic features of the users are
exploited in order to assign them to a group with a known set of preferences. Hence, in essence, it is analogous to the representation of content11
based recommender systems, where a user’s demographic features are used
instead of individual user’s characteristics and preferences.
• Two variants of knowledge-based recommenders – case-based and constraintbased – break the items into weighted features, e.g., the price of a product
and the importance of the price for the user. This model can be represented by two matrices, one contains the items’ weighted features and
the second contains the users’ ranking of the features importance. In the
items matrix, each column represents a feature, each row represents an
item, and the values are the strength, or how representative the feature is
of the item. Similarly, in the users matrix, each column represents a feature, each row represents a user, and the values represent the importance
of the feature for the given user.
• Community-based recommenders combine information regarding users’ social/trust relations with their ratings. Therefore, ratings of a trusted or
socially close user are weighted heavier than those of a less trusted one.
The items rating information can be represented in a matrix identically
to the one described in the collaborative filtering approach. The trust
or social relations weights between users can be represented by a second
matrix, where the rows and columns are represent users and the values
quantify the degree of the relationship between them. The values of the
matrix diagonal are 1, since users fully trust themselves, while the rest of
the matrix can be either symmetric or directional.
• Finally, hybrid approaches combine some of the above stand-alone recommendation models and, therefore, can be represented using the matrix
representation.
The datasets used by the above approaches, which we denote by D, contain
two key types of entities. The first refers to the entity for which the recommendations are generated, i.e., the user; it is referred to as the source entity and
denoted by DS . The second refers to the entity that is being recommended, e.g.,
item, content, product, service, or even another user. This entity is referred to
as the target entity and denoted by DT . This notation follows the primary goal
of a recommender system: to recommend a target item to the source of the
recommendation request. 5 Additional data available in the datasets typically
represent the features of the source and/or the target entity, or the relationships
between the two. The feature set is denoted by DF .
For example, in a movie recommenderation dataset, DS refers to the system
users and DT to the recommendable movies. Any available features describing
either the users or the movies are denoted by DF . User features can be the user’s
5 This definition will also be useful when moving to a graph representation, where metrics
are defined relative to source and target vertices. An alternative definition of “target users”
would have led to confusion and would have broken traditional definitions of metrics, e.g.,
shortest path measured from source to target and not vice versa.
12
age, gender, and location, while movie features can be genre, director, language,
and length. A practical assumption is made that in a tabular recommender
dataset, all the features associated with an entity are stored in the same table
as the entity itself. That is, the gender of a user is stored in the user table rather
than in the movie table. A formal representation of the entities and their features
in the above example is D = {DS , DT , DF }, where DS = userid , DT = movieid ,
and the features DF are split into DF = {DF S , DF T } as follows: DF S =
{fs1 = age, fs2 = gender, fs3 = location} and DF T = {ft1 = genre, ft2 =
director, ft3 = language, ft3 = length}.
It should be noted that the source and target entities can have common
features (Berkovsky, 2006; Berkovsky et al., 2008). For example, in the case of
a restaurant recommendation task, the source entity (user) and target entity
(business) can both have the “location” feature. The role of the source/target
entities and features can also change according to the recommendation task at
hand. In the restaurant recommendation example, when the task is to recommend restaurants to users, the users are the source entity, the restaurants are
the target entity, and location is a feature of both. However, if the task was to
recommend a location, e.g., tourist destinations, for a user to visit based on the
restaurants in that location, then the source entity would still be the users, the
target entity would be the locations, and the restaurants would be the features
of the locations.
An important aspect that needs to be considered is the relationship observed
between the entities, e.g., the fact that a user watched, rated, tagged, or favored
a movie. Relationships can be established not only between a source and a target
entity, but also between two source/target entities. Examples of relationships
between two user entities are the directional followee-follower relationship or the
non-directional friendship. Relationships between two movies can be established
because they are directed by the same director, are in the same language, and
so forth. Relationships between entities are defined using the tuple (source ∈
DS , {f eatures} ∈ DF , target ∈ DT ). For example, the availability of user
ratings for a movie is defined by relrating = (user, value, movie) and friendship
between two users is defined by relf riend = (user, {∅}, user).6 The set of all
possible relationships in a dataset is denoted by DR = {reli }, such as in the
movies example DR = {rel1 = rating, rel2 = f riendship}.
Given the above formalization of entities, features, and relationships, a recommendation task implies the prediction of a relationship between entities. For
example, the task of a movie recommender can be considered as the prediction
of the relrating relationship. This relationship can be numeric (star rating) or
binary (interested or not interested), but the recommendations delivered to the
users are guided by the predicted values of relrating . If, on the contrary, the
system is a social recommender that recommends online friends, then the relationship in question is relf riend and its task is to recommend a set of candidate
friends.
6 Additional
friendship features, such as duration or strength, can also be included.
13
In addition to the original data that is available to the recommender, more
features can be generated and distilled, thus, enriching the dataset. For example, two popular features frequently computed in rating-based recommendation
datasets are the average rating of a user and the average rating for an item.
These features are associated with the users and items, stored in the relevant
tables, and they are used to refine, e.g., normalize, the predicted ratings and
improve the quality of the recommendations (Schafer et al., 1999). The question addressed in this work is whether the availability of additional, supposedly
more complex, features that encompass more information and stem from graph
representation of the data can contribute to the accuracy of the predictions and
the quality of the recommendations. In the following sub-sections, the details
of extracting and populating features are provided.
3.2. Transforming a Tabular Representation into a Graph-based Representation
3.2.1. Basic graph representation for recommender systems data
When moving from the tabular to the graph-based representation of a recommender system dataset, there are multiple graph design considerations. Three
key design questions are:
1. Should the graph encompass all the available data? What parts of the
dataset are important and need to be represented by the graph?
2. Which entities from the selected data should be represented by graph
vertices and which entities by graph edges?
3. How should the edges be defined? Should they be directed or undirected?
Should they be labeled? What should the labels be?
Regarding the first question, it is probable that the decision regarding the
data to be represented in the graph is data-dependent. For some domains,
datasets, and recommendation tasks, certain parts of the data may be more
informative than others. Since the space of possible graph-based data representations is too large for determining a-priori the most suitable scheme, a possible
alternative is to start with a graph model based on the entire data, and then,
to systematically extract all sub-graph representations and their features. This
leads to automatic coverage of the entire search space, inherently uncovering
the representations that produce the most effective features. Then, the most
informative feature set can be selected.
To answer the second and third questions, an intuitive modeling approach is
used. Namely, the graph model considers all the source, target, and feature entities as vertices, while their links and relationships between features (including
user feedback on items) are the edges. If the information about the relationship
is binary, e.g., the item is viewed or not, the edges are not labeled. Otherwise, the edges labels communicate the information about the relationship, e.g.,
rating or type of association. In most cases, the edges are not directed, as information about a feature connected to an entity or about an entity connected to a
feature is equivalent. Although this work does not consider directed edges, the
proposed approach can be extended to support this (outlined in Section 6.2).
14
Source
Target
Source
Target
Source
Target
Features
a) Bipartite
Source
Target
Features
b) Bipartite [with features]
c) Non-Bipartite
d) Non-Bipartite due to
shared features
Figure 2: Examples of two types of graph schemes for representing a recommender system
dataset: bipartite (a,b) and non-bipartite (c,d). In (b) the red block confines the multi-part
bipartite graph component. In (c) and (d) the red edges break the bipartite structure.
Based on the above abstraction of recommender systems datasets, the following basic graph representation emerges. User and item entities are represented
by the graph vertices, and edges connect a user and an item vertex when an
association between the two is available. This association can be explicit (ratings or likes) or implicit (content or user view). This graph is called a bipartite
graph (West et al., 2001), because it can be split into two partitions consisting
of the source and target entity vertices, i.e., the users and items, respectively
(Figure 2-A).
The basic representation can be extended by adding additional features as
new graph vertices and linking them to the existing vertices. For example, if
user locations are provided, each location can be represented by a vertex and
the users associated with the locations are linked to their vertices. A similar
situation may occur in the target partition of the graph, e.g., the target entity of
movies and a variety of their content features: genre, actors, keywords, and more
(Figure 2-B). Adding the feature vertices still preserves the bipartite nature of
the graph, but the partition with the added features gets virtually split into two
groups of vertices: the entities themselves and their features.
The situation changes, however, when adding information within the source
or target partitions, e.g., user-to-user social links or item-to-item links of the
domain taxonomy. This information introduces new links within the partitions,
which break the bipartite structure (Figure 2-C). Additional information that
may break the bipartite structure is the common features shared between the
source and target partitions. For example, in the movie domain, the items may
be linked to their genres, while the users may also express their preferences
towards the genres. Thus, links to the genre vertices are established from both
the user and item partitions (Figure 2-D) and the graph is no longer bipartite.
Note that each of the four schemes shown in Figure 2 potentially generates
different sets of features and the values of the features also vary.
15
Following is an outline7 of a high-level approach for generating the complete
graph, which includes all the data and relationships of a recommender dataset.
The algorithm scans all the tables in the dataset, and for each column that
is not a source entity column, target entity column, or feedback column (e.g.,
ratings) it generates a graph node for every unique value appearing in the column. Thus, every unique userid and movieid is assigned to a graph vertex,
as well as every actor, director, movie genre, keyword, and so forth. Features
that are non-categorical, e.g., movie budget, can be discretized using a simple
binning, e.g., under $10M, $10M-$20M, $20M-$30M, etc. When the range of
values is unknown, the discretization can split the values based on their observed distribution, e.g., four equal-sized quarters, each containing 25% of the
data. Upon discretizing the values in the columns and creating the nodes, all
the nodes matching the values that appear in the same row are connected by
edges to the source and target nodes of the same row, if they are available in
the table. The result is a graph that contains all the values of the features as
the graph nodes, which are connected to the source and target entities based on
their co-occurrence in the data.
3.2.2. Multiple sub-graph representations
Despite being included in a dataset, not all the features are necessarily informative and contribute to the accuracy of the recommendations. Certain
features may be noisy or bear little information, thus, hindering the recommendation process. For example, if a feature is sparsely populated, its values
are identical across users, or it is populated only across a certain subset of
users, then this feature is unlikely to help the recommender and may not be
included in the graph representation. However, it is hard to assess the contribution of the features in advance with a high degree of certainty. This leads
to the idea of automatically deriving multiple sub-graph representations from
the complete graph and extracting the graph features for each sub-graph first,
and selecting the most informative ones in a later stage. Specifically, all the
possible sub-graphs are exhaustively generated and their features are extracted.
Each sub-graph represents a combination of features influenced by the entities
and relationships included in the graph. The process is presented in detail in
Algorithm 1.
The input to the algorithm is the complete graph representation CompleteGraph,
which was discussed at the end of Section 3.2.1, and the edge P redEdge representing the relationship reli being predicted. The function GenerateEdgeCombinations
invoked in line 1 returns all the possible combinations of different types of graph
edges. Note that this function receives also the type of the predicted edges
P redEdge. This is done in order to preserve the P redEdge edges in all the
sub-graphs. Namely, this type of edges will not be included in the combinations
7 For readability purposes, the pseudo codes in this paper omit several pre-processing steps
and technical optimizations. The exact implementation details can be found in the accompanying library.
16
Algorithm 1: Generate sub-graphs and extract features
input : CompleteGraph - complete graph representation of the dataset
P redEdge - edge type of the relationship being predicted
output: ExtractedGraphF eatures - set of features extracted from various
sub-graph representations
1
2
GraphEdgeT ypeCombinations ← GenerateEdgeCombinations({EdgeT ypes},
P redEdge)
ExtractedGraphF eatures ← ∅
7
foreach EdgeCombination ∈ GraphEdgeT ypeCombinations do
SubGraph ← RemoveEdgesFromGraph(CompleteGraph, EdgeCombination)
SubGraphF eatures ← ExtractGraphFeatures(SubGraph, P redEdge)
ExtractedGraphF eatures ← (ExtractedGraphF eatures ∪
SubGraphF eatures)
end
8
return ExtractedGraphF eatures
3
4
5
6
that are removed from the complete graph and, therefore, will be present in all
the sub-graphs.
Upon generating all the possible edge type combinations, the set is iterated over and the function RemoveEdgesFromGraph is invoked to create a
sub-graph SubGraph by removing the combination EdgeCombination from
CompleteGraph (line 4). Then, the function ExtractGraphFeatures is invoked to extract from SubGraph the set of possible graph features referred to
as SubGraphF eatures (line 5, to be elaborated in Section 3.3) and append
SubGraphF eatures to the set of features ExtractedGraphF eatures (line 6).
Finally, in line 8 the algorithm returns ExtractedGraphF eatures – the set of
all the possible graph features from all the possible sub-graphs.
Source
Target
Feature
Source
Target
Features
G1
Source
Target
Features
G2=G1−[r2]
G3=G1−[r3]
Source
Target
Features
G4=G1−[r2,r3]
Figure 3: Four sub-graph schemes that are generated from the complete schema based on the
relationship permutations. Dashed lines represent links removed from the graph.
The execution of Algorithm 1 is illustrated by an example in Figure 3. Consider a graph G = (V, E), where V = {VS ∪ VT ∪ VF } is the set of vertices of
the source entities VS = {VS1 , ..., VSm }, target entities VT = {VT 1 , ..., VT n }, and
domain feature values VF = {VF 1 , ..., VSk }. In addition, E = {rel1 , rel2 , rel3 }
17
is the set of graph edges, reflecting three relationship types: rel1 is the sourcetarget relationship being predicted; rel2 is the relationship between the target
entities and domain features; and rel3 is the relationship between the source
vertices. In graph terminology, the recommendation task is to predict the label
(or the existence) of an edge rel1 (i, j) between a source vertex VSi and a target
vertex VT j .
For this graph, the set GraphEdgeCombinations created by GenerateEdgeCombinations
includes GraphEdgeCombinations = {{∅}, {rel2 }, {rel3 }, {rel2 , rel3 }}. These
are the combinations of edges that are removed from the graph while creating
sub-graphs, whereas the predicted relationship rel1 is preserved in all the subgraphs. Removing these combinations of edges, function RemoveEdgesFromGraph
generates four variants of SubGraph shown in Figure 3: G1 ← CompleteGraph−
{∅}, G2 ← CompleteGraph − {rel2 }, G3 ← CompleteGraph − {rel3 }, and
G4 ← CompleteGraph − {rel2 , rel3 }. Note that G1 is the complete graph,
whereas other sub-graphs have either rel2 or rel3 , or both removed. For each
SubGraph, function ExtractGraphFeatures is invoked to extract the respective
feature set SubGraphF eatures and all the extracted feature sets are appended
to ExtractedGraphF eatures.
3.3. Distilling Graph Features
The function ExtractGraphFeatures in line 5 of Algorithm 1 received a subgraph derived from the complete representation and was invoked to extract a set
of graph-based features. Moreover, this function was invoked for all the possible
sub-graphs, to ensure that all the possible graph features are extracted. The
graph-based features are extracted using a number of functions, each calculating
a different graph metric. These functions, referred as generators, are divided
into several, families according to the number of graph vertices they process.
Source
Target
Features
Source
Features
f (source)
f (target)
PageRank
PageRank
Centrality
Centrality
Average Neighborhood DegreeAverage Neighborhood Degree
a) Single node generators
Target
f (source, target)
Shortest-Path
Fraction of shared Neighbors
b) Dual node generators
Source
Target
Features
f (source, target, node type)
Fraction of shared Neighborsof a certain type
c) Multiple node generators
Figure 4: Key graph-based feature generator families and their instances
The main steps of ExtractGraphFeatures are detailed in Algorithm 2,
which uses three types of generators:
18
Algorithm 2: Extract graph features from a sub-graph
input : SubGraph - sub-graph derived from the complete graph representation
P redEdge - edge type of the relationship being predicted
output: ExtractedSubGraphF eatures - set of features extracted from
SubGraph
1
2
3
4
5
6
7
8
9
10
11
12
ExtractedSubGraphF eatures ← ∅
SubGraphP redictedEdges ← ExtractPredictedEdges(SubGraph,P redEdge)
foreach (SourceEntity,T argetEntity) of Edge ∈ SubGraphP redictedEdges
do
foreach 1-Function in 1-VertexGenerators do
SourceF eatures ← 1-Function(SourceEntity)
T argetF eatures ← 1-Function(T argetEntity)
ExtractedSubGraphF eatures ← (ExtractedSubGraphF eatures ∪
SourceF eatures ∪ T argetF eatures)
end
foreach 2-Function in 2-VertexGenerators do
SourceT argetF eatures ← 2-Function(SourceEntity,T argetEntity)
ExtractedSubGraphF eatures ← (ExtractedSubGraphF eatures ∪
SourceT argetF eatures)
end
13
M ultipleEntityCombinations ←
ExtractEntityCombinations({V ertexT ypes})
14
foreach EntityCombination ∈ M ultipleEntityCombinations do
N ← |EntityCombination|
foreach N-Function in N-VertexGenerators do
M ultipleEntityF eatures ← N-Function(SourceEntity,
T argetEntity, EntityCombination)
ExtractedSubGraphF eatures ← (ExtractedSubGraphF eatures ∪
M ultipleEntityF eatures)
end
end
15
16
17
18
19
20
21
22
return ExtractedSubGraphF eatures
end
• 1-VertexGenerators are applied to a single graph vertex, either the
source or the target entity, and compute features of this vertex only, e.g.,
the PageRank score (Figure 4-A).
• 2-VertexGenerators are applied to a pair of vertices, the source and the
target entities, and compute graph-based relationships between the two,
e.g., the shortest path (Figure 4-B).
• N-VertexGenerators are applied to N >2 vertices, two of which are the
source and target entities and the rest are not. An example function from
this family is “number of vertices of type X, which are common neighbors
19
of the source and target vertices” (Figure 4-C).
Section 3.3.1 lists the functions from each generator family that were used. Note
that these are executed iteratively, in order to generate all the possible graph
features. By no means this list of functions is exhaustive; it only exemplifies a
number of popular functions that were used, but many more functions can be
conceived and added.
At the initial stage of Algorithm 2, edges belonging to the predicted relationship are copied to the SubGraphP redictedEdges set (line 2). For each Edge
in this set, the generators are invoked as follows. The 1-VertexGenerators
functions are invoked in lines 5 and 6, respectively, on the SourceEntity and
T argetEntity vertices of Edge. Applying these functions to other vertices is
unlikely to produce features that can contribute to the prediction of the desired relationship, while leading to significant computational overheads. Hence,
1-VertexGenerators are restricted to these two vertices only. The 2-VertexGenerators
are applied in line 10 to the pairs of vertices SourceEntity and T argetEntity.
Then, the ExtractEntityCombinations function is invoked in line 13, in order
to create a set of all the possible entity combinations of vertices, M ultipleEntityCombinations.
These combinations necessarily involve SourceEntity and T argetEntity, and in
addition any other type of graph vertices. For each combination EntityCombination
of size N (line 15), the relevant N-VertexGenerators generators are invoked
in line 17. Features extracted by 1-VertexGenerators, 2-VertexGenerators,
and N-VertexGenerators are all appended to ExtractedSubGraphF eatures.
Note that the value of N determines the N-VertexGenerators functions that
are invoked and the relationships they uncover. Again, two of the N vertices are
necessarily SourceEntity and T argetEntity, whereas the third vertex can be of
any other entity linked to either of them. For instance, for N = 3 in the movie
recommendation task and entities of user, item, and location, the relationship
can be “the number of cinema locations that the user has visited and where the
movie is screened”. The generator considers the user and movie vertices, and
then, scans all the location vertices and identifies those, with edges connected
to both. It should be noted that more complex relationships with a higher value
of N can be considered. Since a broad range of combinations is possible, the
N-VertexGenerators extract a large number of features that surpasses by far
the set of features that can be engineered manually.
3.3.1. Distilled graph features
The set of metrics selected for implementation in this work and used for the
evaluation of the approach is now given in detail. The metrics are those that
are commonly implemented in widely used graph analysis libraries – NetworkX
(Hagberg et al., 2008), igraph (Csardi and Nepusz, 2006), and Gephi (Bastian
et al., 2009)) – and used in social network analysis and measurement works
(Wilson et al., 2009; Lewis et al., 2008). It is important to stress that this
set of metrics is only a portion of those that could be used and serves only as
an example. The space of all graph metrics is large, as can be seen in (Costa
et al., 2007; Wasserman, 1994; Coffman et al., 2004), and, thus, could not be
20
exhaustively evaluated within the scope of this work.
The set of 1-VertexGenerators functions were implemented and used for
evaluation are degree centrality (Borgatti and Halgin, 2011), average neighbor
degree (Barrat et al., 2004), PageRank score (Page et al., 1999), clustering
coefficient (Latapy et al., 2008), and node redundancy (Latapy et al., 2008).
These metrics are referred to as the basic graph features. Following is a brief of
the 1-VertexGenerators functions.
• Degree Centrality (Borgatti and Halgin, 2011) (or, simply, node degree)
quantifies the importance of a vertex through the number of other vertices to which it is connected. Hence, in the bipartite graph, the degree
centrality of a user vertex Si is the activity of i, i.e., the number of items
with which Si is associated, and, vice versa, for an item vertex Tj it is the
popularity of j, i.e., the number of users who are associated with Tj . In a
graph that includes metadata, the number of metadata vertices associated
with either the user or the item vertex are added to the degree centrality
score. The degree of centrality of a vertex v is denoted by Deg(v).
• Average Neighbor Degree (Barrat et al., 2004) measures the average degree
of vertices to which a vertex is connected. In the bipartite graph, this
metric conveys for Si – the average popularity of items with which Si is
associated, and for Tj – the average activity of users who are associated
with Tj . Formally, if N (v) denotes the set of neighbors of a vertex v, then
the average neighbor degree is
AvgN ghDeg(v) =
X
1
Deg(v)
|N (v)|
(1)
u∈N (v)
In a graph with metadata, the average neighbor degree of a user/item vertex also incorporates the popularity of the metadata features with which
it is associated.
• PageRank (Page et al., 1999) is a widely-used recursive metric that quantifies the importance of graph vertices. For a user vertex Si , the PageRank
score is computed through PageRank scores of a set of item vertices {Tj }
with which Si is associated and vice versa Thus, the PageRank score of a
user vertex Si can be expressed as
X
P ageRank(Si ) =
Tj ∈N (Si )
P ageRank(Tj )
,
Deg(Tj )
(2)
i.e., the PageRank score of Si depends on the PageRanks of each item
vertex Tj connected to Si , divided by the degree of Tj . In a graph with
metadata, the PageRank scores of user/item vertices are also affected by
the PageRank of the metadata vertices to which they are connected.
• Clustering Coefficient (Latapy et al., 2008) measures the density of the
immediate subgraph of a vertex as the ratio between the observed and
21
possible number of cliques of which the vertex may be a part. Since
cliques of a size greater than two are impossible in the bipartite graph,
ClustCoef measures the density of shared neighbors with respect to the
total number of neighbors of the vertex. The formal definition for the
bipartite graphs is
P
|N (v)∩N (u)|
u∈N (N (v)) |N (v)∪N (u)|
ClustCoef (v) =
|N (N (v))|
(3)
• Node Redundancy (Latapy et al., 2008) is applicable only to bipartite
graphs and shows the fraction of pairs of neighbors of a vertex that is
linked to the same other vertices. This metric quantifies for user vertex
Sa - the portion of pairs of items with which a is associated that are
also both associated with another user b. Likewise, for item vertex Tx ,
it quantifies the portion of pairs of users associated with x and also both
associated with another item y. If a vertex, node redundancy of which
is computed, is removed from the graph, the metric reflects the fraction
of its neighbors that will still be connected to each other through other
vertices. Intuitively, in a bipartite graph N odeRed(v) can be seen as the
portion of connected ‘squares’ of which v is a part, among all the potential
‘squares’.
Next, multiple-vertex generator functions are detailed. Specifically, the following functions from the 2-VertexGenerators and N-VertexGenerators families were implemented:
• Shortest Path (Floyd, 1962). Unlike the above feature generators that
operate on a single vertex, shortest path receives a pair of graph vertices:
a source entity and a target entity. It evaluates the distance, i.e., the lowest
number of edges, between the two vertices. The distance communicates
the proximity of the vertices in the graph, as is a proxy for their similarity
or relatedness. A short distance indicates high relatedness, e.g., more
items shared between users or more features for items, while a longer
distance indicates low relatedness.
• Shared Neighbors of Type X. This is one of the N-VertexGenerators
functions, which receives three parameters: source entity vertex, target
entity vertex, and entity type X. It returns the fraction of neighbors
shared between the source and target vertices that areof the desired type
X. The fraction is computed relatively to the union of the source vertex
neighbors with the target vertex neighbors. Note that this feature cannot
be populated for graphs that do not have a sufficient variety of entities
connected to the source and target vertices. For example, the generator
is inapplicable for a graph having only the source and target entities.
• Complex relationships across entities. Apart from the above mentioned
generators, system designers may define other N-VertexGenerators functions, which could extract valuable features. For example, it may be
22
beneficial for a movie recommender to extract the portion of users, who
watched movies from genres g1 , g2 directed by person p, and released between years t1 and t2 . It is clear that it is impossible to exhaustively list
all the combinations of such features: this is domain- and applicationdependent. Hence, the task of defining these complex generators is left
open-ended and invites system designers to use the provided library and
develop their own feature generators.
To recap, each of the above 1-VertexGenerators and 2-VertexGenerators
is applied to every source and target vertex and generates features associated
with the vertex or a pair of vertices. In addition, N-VertexGenerators is applied
to the source and target vertices and all the possible combinations of other entity
types. Recall that this is done for every sub-graph extracted from the complete
graph8 and the complexity of the feature generation task becomes clear.
3.3.2. Quantifying the number of graph features
Here, the number of graph features that can be extracted from a recommender system dataset using the proposed approach is quantified. The quantification illustrates the coverage and computational complexity of the extraction
process. Considering Algorithms 1 and 2, it becomes evident that the number
of extracted features primarily depends on two key elements: the number of
sub-graphs that are extracted from the complete graph and the number of entities in each sub-graph. Both of these are derived from the number of entities
and relationships in the dataset.
Building on the recommender dataset analysis given in Section 3.1, each
dataset D contains two entities, DS and DT , and |DR | relationships. The
features are extracted from the complete graph and also from the sub-graphs.
The latter are generated from any non-ordered combination of relationships
from DR , but necessarily contain the predicted relationship reli , such that the
overall number of relationships is M = |DR | − 1. Hence, the number of possible
sub-graph combinations is
M
X
M
x=0
x
=
M
X
(M )!
,
x!(M − x)!
x=0
(4)
the sum of the numbers of combinations of size x that can be produced, where
x ∈ [0, M ].
For each of these sub-graphs, let us assume that F1 features can be generated
by 1-VertexGenerators function for DS and DT individually, and F2 features
can be generated by 2-VertexGenerators for the pair (DS , DT ). On top of
these, at least M = |DR | − 1 more complex by N-VertexGenerators functions
can be applied to every sub-graph, as per the number of relationships available
8 Some generators should be applied in a different manner to certain sub-graphs, e.g., Deg
and ClustCoef generators in bipartite and non-bipartite graphs.
23
5632
Number of Features
6000
5000
4000
2816
3000
2000
1000
00
352
11 22 44 88 176
2
1408
704
4
6
8
Number of Relationships
10
Figure 5: Number of extracted graph features versus the number of relationships in the dataset
in the data. This brings the overall number of generated graph-based features
to the order of
M
X
(M )!
(2F1 + F2 + M ) ×
(5)
x!(M − x)!
x=0
The library that accompanies this work defines F1 = 5 single node generators and F2 = 1 dual node generator. For illustrative purposes only, Figure
5 plots the number of extracted features, which is exponential with the number of relationships |DR |. For example, the number of features extracted from
a dataset having |DR | ≤ 4 relationships is smaller than 100. However, for a
dataset with |DR | = 10 relationships, the number of extracted features exceeds
5,000. Clearly, engineering all these features manually would require considerable resources, whereas the proposed approach is fully automated.
4. Experimental Setting and Datasets
It is important to highlight that the product of the presented approach
is graph-based features that help to generate recommendations using existing
recommendation methods. These features can either be used as stand-alone
features, i.e., the only source of information for the recommendation generation,
or be combined with other features. Hence, the baseline for comparison in the
evaluation part is the performance of common recommendation methods when
applied without the newly generated features.
To present solid empirical evidence, the contribution of the graph feature
extraction to the accuracy of the recommendations was evaluated using three
machine learning methods: Random Forest (Breiman, 2001), Gradient Boosting
(Friedman, 2000), and Support Vector Machine (SVM) (Gunn et al., 1998).
Both Random Forest and Gradient Boosting are popular ensemble methods
that have been shown to be accurate and won recommendation (Koren, 2009)
and general prediction (Yu et al., 2010) competitions. The methods are also
implemented in widely used machine-learning libraries (Pedregosa et al., 2011;
24
Case Study I –
Overall contribution of
graph-based features
Case Study II –
Performance of different
graph schemes
Dataset I - Last.fm
Dataset II - Yelp
Dataset III - Yelp II
Dataset IV - OSN
Dataset V - Movielens
Table 1: Mapping of datasets to case studies
Hall et al., 2009), and were shown to perform well in prior recommender systems
works (Jahrer et al., 2010; Bellogı́n et al., 2013; Töscher et al., 2009).
In the next section, two case studies showing the contribution of graph-based
features are presented. These case studies demonstrate the value of the proposed
graph-based approach when applied to a range of recommendation tasks and
application domains. Case study I evaluates the performance of the graph-based
approach, evaluating its contribution in different domains and tasks. Case study
II focuses on the impact of representing data using different graph schemes on
the recommendations. Altogether, five datasets were used across the case studies
and the mapping between the datasets and case studies is laid out in Table 1.
In the following sub-sections, a brief characterization of the datasets, as well as
the overview of the recommendation tasks and evaluation metrics, is provided.
4.1. Dataset I – Last.fm
The first dataset is of users’ relevance feedback provided for music performers via the Last.fm online service. The dataset is publicly available9 and was
obtained by (Cantador et al., 2011). The dataset consists of 1,892 users and
17,632 artists whom the users tagged and/or listened to. More than 95% of
users in the dataset have 50 artists listed in their profiles as a result of the
method used to collect the data. There are 11,946 unique tags in the dataset,
which were assigned by users to artists 186,479 times. Each user assigned on
average 98.56 tags, 18.93 of which are distinct. Each artist was assigned 14.89
tags on average, of which 8.76 are distinct. The dataset also contains social
information regarding 12,717 bidirectional friendship linkss established between
Last.fm users, based on common music interests or real life friendship.
A brief characterisation of the dataset is shown in Figure 6. Figure 6a illustrates the distribution of the number of friends per user. The average number
of user-to-user edges is low, which is illustrated by the vast majority of users
having less than 10 friends and about half of users having less than four friends.
9 http://grouplens.org/datasets/hetrec-2011/
25
(a) Distribution of the number of
friends per user
(b) Distribution of the average number
of listens per user/artist and overall
Figure 6: Last.fm data characteristics
Intuitively, a friendship edge between two users can be an indicator of similar
tastes, and as such, friendship-based features are expected to affect the recommendations. Figure 6b, shows the distributions of the number of listens per
artist, user, and in total. It can be observed that the overall and per artist
distribution are highly similar. The user-based distribution resembles the same
behaviour, but drops faster. This aligns with the intuition that the number of
users who listen to several hundreds of artists is smaller than the number of
artists who are listened by several hundreds of users (Haupt, 2009).
There are four relationships in the Last.fm dataset: [user, listens, artist],
[user, uses, tag], [tag, used, artist], and [user, friend, user]. The task defined
for this dataset was to predict the artists to whom a users will listen the most,
i.e., the predicted relationship was [user, listens, artist]. This task requires first
predicting the number of times each user will listen to each artist, then ranking
the artists, and choosing the top K artists. Based on the sub-graph generation
process detailed in Algorithm 1 and the relationship being predicted, the data
can be represented via eight graph schemes in general. Four graph schemas that
incorporate the source and target entities were evaluated:
• A bipartite graph that includes users and artists only, denoted as the
baseline (BL)
• A non-bipartite graph that includes users, social links, and artists (BL+F)
• A non-bipartite graph that includes users, artists, and tags assigned by
users to artists (BL+T)
• A graph that includes all the entities and relationships: users, tags, artists
and social links (BL+T+F).
The four graphs are illustrated in Figure 7. For each of the graphs, two sets
of features were generated: basic features, as well as a set of extended features
associated with the auxiliary data being included. The generated features are
used as the input for a Gradient Boosting Decision Tree regressor (Friedman,
2000), trained to predict the number of listens for a given user-artist pair.
26
User
Listens
Artist
User
Listens
Artist
User
Listens
Artist
Friend
User
Baseline + Tags (BL+T)
Artist
Friend
Tag
Tag
Baseline (BL)
Listens
Baseline + Friends (BL+F)
Baseline + Tags + Friends (BL+T+F)
Figure 7: Graph representations for dataset I (Last.fm)
A 5-fold cross validation was performed. Users with fewer than five ratings
were pruned, to ensure that every user has at least one rating in the test set
and four in the training set. For each training fold, a graph was created for each
graph model shown in Figure 7. For each user, in the test set, a candidate set of
artists was created by selecting artists out of the set of the artists listened to by
the user and complementing these by randomly selected artists. For example,
a candidate set of 100 artists included 10 artists listened to by the user and 90
random artists. Three different candidate sizes were evaluated: 50, 100, and
150.
Then, a regressor was used to predict the number of listens for each artist
in the candidate set, rank the set, and compute precision at 10 (P@10) as
the performance metric (Shani and Gunawardana, 2011). If candidate set CS
consists of the artists selected from a user’s artist set denoted by U A and the
randomly selected artists set RA, then P@10 is computed by P @10 = (U A ∩
top 10 artists(U A ∪ RA))/10, where top K artists is the list of top-K artists in
CS ranked according to the predicted number of listens. Finally, an average
of the P@10 scores across all the users in the test set is computed. In order
to evaluate the significance in the performance of the various graph schemes
feature sets, a two-sided t-test was applied on the results.
4.2. Dataset II – Yelp (from RecSys-2013)
The second dataset is of users relevance feedback given for businesses, such
as restaurants, shops, and services. The dataset was released by Yelp for the
RecSys-2013 Challenge (Blomo et al., 2013), and is publicly available.10 For the
analysis, users with less than five reviews were filtered out, which resulted in
9,464 users providing 171,003 reviews and the corresponding ratings for 11,197
businesses. The average number of reviews per user is 18.07 and the average
number of reviews per business is 15.27. A key observation regarding this dataset
is the distribution of ratings, which were almost all positive (more than 60% of
ratings were at least 4 stars on a 5-star scale), and the low variance of ratings
across businesses and users. This phenomenon is common in star rating datasets,
where users tend to review fewer items that they did not like.
Figure 8 summarizes the basic statistics of users and businesses in the Yelp
dataset. Figure 8a illustrates the distribution of the number of reviews and
10 https://www.kaggle.com/c/yelp-recsys-2013/data
27
(a) Distribution of the number of reviews per user
(b) Distribution of the number of reviews per business
Figure 8: Yelp dataset characteristics
Users
Rated
Businesses
Users
Rated
Businesses
Bipartite
Component
Business
Metadata
(a)
(b)
Figure 9: Graph representations for dataset II (Yelp - RecSys-2013 Challenge)
ratings per user. A long tail distribution of the number of businesses a user
reviewed can be observed, with more than 75% of the users providing less than
10 reviews. Likewise, we observe in Figure 8b the distribution of the number
of reviews a business received. Only 24% of businesses attract more than 10
reviews, while only a few businesses (less than 2%) have a relatively high number
of reviews (more than 100). Despite the high number of categories in the data,
the average number of categories with which a business is associated is only
2.68. Every business is also associated with a single location.
The task defined for this dataset is the one originally defined for the RecSys2013 challenge, i.e., to predict the ratings a user will assign to businesses. Two
graph models were implemented and evaluated based on this dataset: a bipartite
model with sets of vertices U and B representing users and businesses and a tripartite 11 model with sets of vertices U, B, and M representing users, businesses,
and metadata items, respectively. The high-level graph representation models
are illustrated in Figure 9, while the detailed presentation of the sub-graphs will
be given in Section 4.3, in which the follow-up dataset is presented.
The features generated for this dataset were aggregated into three groups:
11 The use of ‘tripartite’ is slightly inconsistent with the canonic definition, such that the
“bipartite graph with metadata nodes” notation would be more appropriate. For the sake of
brevity, the bipartite and tripartite terminology is used.
28
• Basic features that include only the unique identifiers of users {ui } ∈ U
and businesses {bj } ∈ B.
• Manual features that include the number of reviews by ui , average rating
of ui , number of reviews for bj , number of categories |{m}| with which
bj is associated, average number of businesses in {m}, average rating of
businesses in {m}, the main category12 of bj , average degree of businesses
associated with the main category of bj , average degree of businesses in
{m}, and the location of bj .
• Graph features that include the degree centrality, average neighbor degree,
PageRank score, clustering coefficient, and node redundancy. These features were generated for both user nodes ui and business nodes bj , whereas
an additional shortest path feature was computed for the pairs of (ui ,bj ).
In this case, a Random Forest regression model (Breiman, 2001) was applied
for the generation of the predictions of users ratings for businesses. At the
classification stage, the test data items were run through all the trees in the
trained forest. The value of the predicted rating was computed as a linear
combination of the scores of the terminal nodes reached when traversing the
trees. It should be noted that the ensemble of trees in Random Forest and the
selection of the best performing feature in each node inherently eliminate the
need for feature selection. Since every node uses a single top performing feature
for decision making, the most predictive features are naturally selected in many
nodes and the ensemble of multiple trees virtually replaces the feature selection
process.
A 5-fold cross validation was performed. For each fold, the predictive model
was trained using both the original features encapsulated in the dataset and the
new graph features. The basic and manual groups of features were populated
directly from the reviews, whereas the graph features were populated from the
bipartite and tripartite graph representations and augmentrf the former groups
of features. Predictive accuracy of various combinations of features was measured using the widely-used
q RMSE metric (Shani and Gunawardana, 2011),
P
(ŷ −y )2
t
t
n
, where n is the number of predictions, ŷt
computed as RM SE =
n
are the predicted values, and yt are the actual user ratings. A two-sided t-test
was applied to validate the statistical significance of the results.
4.3. Dataset III – Yelp II (with social links)
The third dataset is an extension that was released by Yelp to the previous
dataset. The new version contains more users, businesses and reviews (although
their distribution still resembles the one shown in Figure 8), and, more importantly, new information regarding users’ social links. The distribution of the
12 Each business in the Yelp dataset is associated with multiple categories, some having
an internal hierarchy. The main category is the most frequent root category a business was
associated with.
29
Distribution of social links among users
0.30
0.25
0.20
0.15
0.10
0.05
0.000
20
40
60
80
100
Figure 10: Yelp II Dataset characteristics – distribution of social links
Friends
User
Friends
Rated
Name
Business
User
Category
Name
[1] Full Graph (Non-Bipartite)
Rated
User
Business
Friends
Rated
Category
[2] Without Social Links
(Bipartite)
Business
User
Category
Name
[3] Without Name Links (NonBipartite)
Rated
Business
[4] Without Category Links
(Non-Bipartite)
Friends
User
Rated
Business
Name
User
Rated
Business
User
Rated
Business
User
Rated
Business
Category
[5] Without Social Links and
Category Links (Bipartite)
[6] Without Social Links and
Name Links (Bipartite)
[7] Without Metadata (NonBipartite)
[8] Without Metadata and
Social Links (Bipartite)
Figure 11: Graph representations for dataset III (Yelp II)
social links among users is illustrated in Figure 10. It can be seen that the social links follow a long tail distribution, where most users have a small number
of links: 29% with no links, 57% with less than 20 links, and only a few users
with more than 20 links. The social links also break the bipartite structure of
the first Yelp dataset, which influences the generated graph features.
The task for this dataset is identical to that of the first Yelp dataset, i.e.,
predicting users ratings for businesses. Eight graph models were generated and
evaluated based on this dataset. The models are illustrated in Figure 11 and,
depending on the availability of the user-to-user friendship edges, categorized as
bipartite or non-bipartite. The complete graph is shown in the top-left schema.
In the following three schemes one type of edges is missing: either social links,
user names, or categories. In the next three, two types of edges are misisng:
social and categories, social and names, and names and categories. Finally, in
the bottom-right graph all three are missing.
The generated features presented in Section 3.3.1 are referred to in the evaluation of this dataset as the basic features. These features are aggregated into
groups, based on the graph scheme from which they were extracted. For example, all the features extracted from the graph named “without category links”
in Figure 11 were grouped into a combination having the same name. Another
30
evaluated combination includes the union of all the features generated from all
the graph schemes, and this is named “all graph features”. Finally, the union
of “all graph features” with the “basic features” is referred to as “all features”.
A 5-fold cross validation was performed. For each fold, the predictive models
were trained using graph features extracted from each of the above feature sets.
The evaluation was conducted three times, each time training the models using
a different method (Random Forests, Gradient Boosting, and SVM), in order to
evaluate how the choice of method impacts the results. Predictive accuracy of
various feature combinations was measured using the RMSE metric (Shani and
Gunawardana, 2011), and a two-sided t-test was applied to validate statistical
significance.
4.4. Dataset IV – OSN
The fourth dataset is an Online Social Network (OSN) profile dataset that
was collected from six large networks: Facebook, LinkedIn, Last.fm, Blogger,
YouTube, and LiveJournal. The profiles were manually linked and matched
to each other by the users themselves, as they mentioned their user names on
other OSNs. The lists of user interests were then extracted from the OSNs
and categorized into five domains: movies, music, books, TV, and general.
The categorization was explicitly made by the users on Blogger, Facebook, and
YouTube; all Last.fm interests were categorized as music; no categorization was
available on LinkedIn and LiveJournal, so that there interests were treated as
general. Users having one interest only and interests mentioned by one user
only were filtered, such that the resultant dataset contained 21,880 users with
an average of 1.49 OSNs and 19.46 interests per user.
It can be observed in Table 2 that the most common interests in user profiles
are from the music and general domains. Table 3 shows that Facebook is, by
far, the OSN with the most listed interests. Table 4 shows the number of
users who have at least one interest for a domain and OSN combination, with
the right-most column indicating the total number of users. The OSN with
the largest number of profiles is Facebook, followed by LinkedIn and Last.fm.
Music interests are the most common across the Facebook and YouTube profiles,
while general interests are the most common in Blogger profiles. Finally, Table
5 shows the average number of interests a user has in each domain and OSN.
The task defined for this dataset was to predict the interests of the users,
based on their partial profiles. The data were represented using a single graph
model because of the availability of only two entities: users and interests. The
model is bipartite graph G = {U, I, E}, where users U = {ui } and interests
I = {ii } are the vertices. User vertices are connected to interest vertices with
an edge if the interest is mentioned in one of the available OSN profiles, i.e., E
= {eij | if ui listed ij }). The edges are labeled by the OSN(s), in which the
interest was listed.
From the graph, a set of graph and manually engineered features was extracted. They can be categorized into two groups: user features and interest
features. Each of these groups can be split into two sub-groups: basic manual
features (IB and U B) and graph-based features (IG and U G). The U B features
31
Domain
General interests
Movies
Musics
Total
154,245
54,382
139,307
Unique
13,053
5,190
21,255
Domain
Books
TV
All
Total
24,404
53,508
425,846
Unique
3,789
4,027
47,314
Table 2: Total number of interests and unique interests in each domain
Network
Blogger
Facebook
Last.fm
LinkedIn
Total
27,045
253,217
63,952
47,955
Unique
6,587
31,511
16,483
6,325
Network
Livejournal
YouTube
All
Total
30,924
2,753
425,846
Unique
5,198
1,561
47,314
Table 3: Total number of interests and unique interests in each OSN
Blogger
Facebook
Last.fm
LinkedIn
LiveJournal
YouTube
General interests
2,716
8,391
Movies
1,090
8,922
Music
1,370
10,453
7,042
Books
518
6,565
7,755
1,494
448
484
650
552
TV
9,619
Total
3,136
11,619
7,042
7,755
1,494
1,548
Table 4: Number of users who have at least one interest in each domain and OSN
Blogger
Facebook
Last.fm
LinkedIn
LiveJournal
YouTube
General interests
5.913
6.999
–
6.183
20.69
1.288
Movies
2.976
5.662
–
–
–
1.278
Music
4.676
6.509
9.082
–
–
1.395
Books
2.577
3.414
–
–
–
1.177
TV
–
5.562
–
–
–
–
Table 5: Average number of interests available per user in each domain and OSN
include the number of OSNs of which the user is a member (U B 1 ), the number
of user interests in each domain – books, TV, movies, music, general (U B 2 ,
U B 3 , U B 4 , U B 5 , U B 6 , respectively), and the total number of interests (U B 7 ).
The IB features include the number of users who liked the interest (IB 1 ), the
number of OSNs where the interest appears (IB 2 ), Boolean features signifying
whether the interest is listed on each OSN – Blogger, LinkedIn, Last.fm, LiveJournal, Facebook, YouTube (IB 3 , IB 5 , IB 6 , IB 7 , IB 8 , IB 9 , respectively),
and the domain to which the interest belongs (IB 4 ).
The graph-based features are identical for users and interests and contain:
Degree centrality (IG1 , U G2 ), Node redundancy (IG2 , U G1 ), Clustering coefficient (IG3 , U G4 ), Average neighborhood degree (IG4 , U G3 ), PageRank (IG5 ,
U G1 ), and the Shortest path feature computing the distance between a user-
32
interest pair. Additional features defined for this dataset are IGall = {IGi },
U Gall = {∪U Gi }, IBall = {IBi }, and U Ball = {BGi }. Finally, Iall =
{IBall ∪ IGall } and Uall = {U Ball ∪ U Gall }.
The experiments using the OSN dataset evaluated the effect of the features
on the predictions of the likelihood of a user to list an interest. A Random
Forest classifier was used and trained on the user-interest pairs augmented with
their features. Each pair was classified into the ‘like’ or ‘dislike’ classes. 10fold cross validation was applied for evaluation. For each fold, a graph was
built, and then, the above features were extracted and fed into the classifier.
Since no real disliked interests were in the data, random interests were selected
from the interests not listed by the user. The number of disliked interests was
equal to the number of liked interests for each user. The synthetic disliked
interests were used only to train the classifier and not used in the evaluation.
Precision was the metric chosen to evaluate the quality of predictions for a user:
P
P = T PT+F
P , where T P is the number of correctly and F P is the number of
incorrectly predicted interests (Shani and Gunawardana, 2011).
4.5. Dataset V – Movielens
Movielens (Lam and Herlocker, 2012) is a classical recommender systems
dataset studied in numerous prior works. In this work it is used to show that
the graph-based approach is as effective on legacy datasets as on more recent
datasets including social data. The 1M Ratings Movielens dataset consists of
1,000,209 ratings assigned by 6,040 users for 3,883 movies, on a discrete scale of 1
to 5 stars. Each user in the dataset rated at least 20 movies. The distribution of
ratings across users and movies is illustrated in Figures 12a and 12b, respectively.
The dataset contains metadata of both users and movies. The user metadata
includes the gender, occupation, zip code area, and age group, while the movie
metadata contains the genre(s) of the movies.
The task defined for this dataset was to predict what ratings would users
assign to movies. Based on the above description of the dataset, 32 graph
schemes were generated and evaluated (see Figure 13). The schemes are categorized based on the number of relationships that were removed from the complete
graph that contains all the entities and relationships. As can be seen, there are
four categories: schemes with a single node type removed, containing 5 subgraphs, schemes with 2 node types removed containing 10 sub-graphs, schemes
with 3 node types removed containing 10 more sub-graphs, and finally, schemes
with 4 node types removed containing 5 graphs. The minimal graph scheme is
the one from which all the entities and relationships were removed, except for
the source and target entities and the predicted ‘rating’ relationships.
A 5-fold cross validation was performed. For each fold, the predictive models were trained using graph features extracted from each of the above graph
schemes. The evaluations were conducted twice, training the models using the
Random Forest and Gradient Boosting approaches, in order to evaluate how
the choice of the learning method impacts the results. The predictive accuracy
of various combinations of the above feature sets was measured again using the
33
Distribution of ratings among movies
Distribution of ratings among users
0.08
Percentage of movies
Percentage of users
0.10
0.06
0.04
0.02
0.00
0
500
1000
1500 2000
0
2500
# of ratings
0.07
0.06
0.05
0.04
0.03
0.02
0.01
0.00
500 1000 1500 2000 2500 3000 3500
# of ratings
(a) Distribution of ratings across users (b) Distribution
movies
of
ratings
across
Figure 12: Movielens dataset characteristics
Gender
Gender
Occupation
User
Rated
Gender
User
Movie
Rated
Movie
Age
Age
Genre
ZIP
[1] Full Graph
Occupation
User
Rated
Rated
Movie
Age
Genre
ZIP
Graphs without a single relationship
(5 graphs)
Movie
User
Genre
ZIP
Graphs without three relationships
(10 graphs)
User
Rated
Movie
Graphs without four relationships
(5 graphs)
ZIP
Graphs without two relationships
(10 graphs)
User
Rated
Movie
[1] Minimal Graph
Figure 13: Graph representations for dataset V (Movielens). Each graph is an example of the
sub-graphs in the group.
RMSE and MAE predictive accuracy metrics (Shani and Gunawardana, 2011),
and a two-sided t-test was applied to validate statistical significance.
4.6. Summary of the datasets, features, and metrics
Table 6 summarizes this section and presents the experimental datasets,
number of source and target entities, various sub-graph schemes investigated,
number of extracted feature sets, groups of features, and evaluation metrics
exploited. The five datasets contain large numbers of users and items and cover
a broad range of data types, application domains, and recommendation tasks.
The datasets also contain both legacy and recently collected datasets, such that
the evaluation presented in the following section offers solid empirical validity.
34
Source and
Target Entities
Graph Schemes
Graphs
Feature Sets
Last.fm
1,892 (users)
17,632 (artists)
Bipartite + Non-bipartite
(w/ social links, w/ tags,
w/ social links+tags)
4
7
Yelp
9,464 (users)
11,197 (businesses)
Bipartite +
Bipartite with metadata
2
13
Yelp II
13,366 (users)
14,853 (businesses)
Bipartite +
Non-bipartite (w/ social links),
with and without metadeta
8
OSN
21,880 (users)
47,314 (interests)
Bipartite
1
Movielens
6,040 (users)
3,883 (movies)
Bipartite +
Bipartite with metadata
Dataset
32
Learning Method
Evaluation
Metric
Basic graph features,
extended graph features
Gradient Boosting
P@K
Basic graph features,
manually engineered
Random Forest
RMSE
13
Basic graph features
Random Forest,
Gradient Boosting,
Support Vector Machine
RMSE
8
Basic graph features,
manually engineered
Random Forest
Precision
Basic graph features
Random Forest,
Gradient Boosting
RMSE,
MAE
36
Extracted
Features
Table 6: Summary of datasets characteristics
5. Results and Analysis
5.1. Case Study I: Overall Contribution of the Graph-based Approach
This case study answers the broad question: How does the use of graph features affect the performance of rating predictions and recommendation generation in different domains and tasks? Each of the above datasets was represented
by graphs and graph-based features were extracted from the graphs using the
approach detailed in Section 3. For each dataset, a matching recommendation
task was defined as follows: for the Last.fm dataset the task was to predict
the artists to which users will listen; for the two Yelp datasets the task was to
predict user ratings for business; for the OSN dataset, to predict the interests in
user profiles; and, for the Movielens dataset, to predict user ratings for movies.
The tasks were performed and evaluated under three conditions:
• Prediction with versus without the newly extracted graph features
• Prediction with user-related versus item-related graph features
• Prediction using features of a bipartite graph versus extended graph schemes,
e.g, containing metadata.
All the evaluations were conducted using the N-fold cross validation methodology (Kohavi et al., 1995), with N = 5 folds in the Last.fm, both Yelps, and
Movielens datasets, and N = 10 in the OSN dataset. For each fold, the complete graph representation was generated based on the entities from both the
training and test sets, except for the relationships being predicted in the test set.
A two-sided t-test was conducted with the null hypothesis of having identical
expected values across the compared prediction sets. The tests assumed that
the predicted ratings using feature set A and predicted ratings using feature set
B were taken from the same population. The threshold used for a statistically
significant difference was p=0.05.
5.1.1. Dataset I – Last.fm Results
Four graph schemes were generated for the Last.fm dataset, as per the structure in Figure 7. For each graph scheme the set of basic graph features listed in
35
Figure 14: Precision of feature combinations using the four graphs - Last.fm dataset.
Section 3.3.1 was extracted and populated. The basic features encapsulate only
the user-artist listening data and denoted by F̂ . In addition, when the social
and tagging data is available, namely, in the BL+T, BL+F, BL+T+F schemes,
the set of extended features can be extracted. These features are denoted by
F , e.g., FBL+T denotes the set of extended features extracted from the graph
with the tagging data. Note that for the BL schema in Figure 7, having neither
social nor tagging data, the basic and extended feature sets are identical, i.e.,
FBL = F̂BL .
The P@10 results obtained for the extended feature sets extracted from
the four schemes are summarized in Figure 14. The boundaries of the boxes
represent the 25th and 75th percentile of the obtained P@10, and the average
P@10 is marked by the dot inside the boxes. The values of the average P@10
are also given. The baseline for comparison in this case is the performance
of the graph features extracted from the bipartite scheme FBL , which scored
P@10=0.336. A notable improvement, between 63% and 70%, was observed
when the extended feature sets were extracted. For instance, FBL+F , scored
P@10=0.555, which is an improvement of more than 65%. A combination of the
extended features using the graph that includes both social tags and friendships,
FBL+T +F is the best performing feature. This scored the highest P@10=0.571
and improved the baseline P@10 by as much as almost 70%.
In order to evaluate the significance of the results, a paired t-test was performed with each group of features, using the P@10 values obtained for each
of the four graphs. The results show that among the extended feature sets, all
the differences were significant, p<0.05. Thus, the inclusion of auxiliary tagging
and friendship data improved the accuracy of the prediction, while their combination including both components led to the most accurate predictions. More
importantly, the extraction of graph-based features was shown to consistently
and significantly boost the performance of the recommender, in comparison to
the variant not using the extracted features.
36
1
2
3
4
5
6
7
8
9
10
11
12
13
Features combination
Features
RMSE
Improvement
All Features
AllExcept Tripartite
AllExcept Basic
Manual and Bipartite
AllExcept Bipartite
Manual and tripartite
AllExcept Manual
All Graph
AllExcept Graph
Bipartite
Tripartite
Basic
Manual
Basic∪Manual∪Graph
Basic∪Manual∪Bipartite
Manual∪Graph
Manual∪Bipartite
Basic∪Manual∪Tripartite
Manual∪Tripartite
Basic∪Graph
Bipartite∪Tripartite
Basic∪Manual
1.0766
1.0775
1.0822
1.0850
1.0896
1.1073
1.1095
1.1148
1.1175
1.1188
1.1326
1.1809
1.1853
8.82%
8.75%
8.35%
8.11%
7.72%
6.22%
6.04%
5.59%
5.36%
5.25%
4.09%
N/A
-0.37%
Table 7: RMSE of selected feature combinations - Yelp dataset (baseline combination in light
gray).
5.1.2. Dataset II – Yelp Results
Improvements due to the use of the graph-based approach were also evident in experiments using the second dataset (Yelp). As per the description in
Section 4.2, basic (user and business identifiers), manual (number of reviews,
average rating, business categoriy and location), and graph-based features were
extracted and populated. The latter were further broken down into the bipartite
and tripartite features. In this dataset, the performance of the basic features related to the user-to-business associations serves as the baseline. Table 7 presents
the full results for all the feature combinations.
The largest improvement in the RMSE of business ratings prediction was an
8.82% decrease obtained for the combination of graph features with basic and
manually engineered ones (row 1). The similarity of the RMSE scores obtained
by the various combinations is explained primarily by the low variance of user
ratings in the dataset. Since most ratings are similar, they are highly predictable
using simple methods and there is only a limited space for improvement. A combination containing only the graph features (row 5) outperformed the baseline
performance by 5.59%. On the contrary, the use of manual features (row 13)
slightly deteriorated the accuracy of the predictions. This demonstrates the full
benefit of the graph-based approach: extracting the graph features took less
time than crafting the manual ones, and the graph features also outperformed
the manual ones.
An examination of the differences in the accuracy of the results obtained
when combining various groups of features revealed a number of findings. An
analysis of the performance of each group of features shows that the bipartite and tripartite feature sets performed noticeably better than the basic and
manual feature sets (rows 10 and 11 versus rows 12 and 13). A combination
of graph features (row 8) still outperforms slightly, although significantly, the
combination of the basic and manually engineered features (row 9). To analyze
37
Basic
Manual_Eng
Bipartite
Tripartite
0.07
1.00
0.07
0.10
1.00
Manual_Eng_and_Bipartite 0.12
All_Graph
All_Graph
0.12
1.00
AllExcept_Manual_Eng
Manual_Eng_and_Tripartite
Manual_Eng_and_Bipartite
Manual_Eng_and_Tripartite
AllExcept_Manual_Eng
AllExcept_All_Graph
All_Features
All_Features 1.00
AllExcept_All_Graph
1.00
0.10
1.00
Tripartite
0.19
1.00
Bipartite
0.19
Manual_Eng
1.00
1.00
Basic
1.00
Figure 15: Significance of the differences between feature combinations in the Yelp dataset.
White cells - significant, dark cells - not significant, p-value given.
the impact of the feature groups, each group was excluded from the overall set
of features and the change with respect to the All Features combination (row
1) was measured. When the graph features were excluded (row 9), the predictions were less accurate than when the basic (row 3) or manual features (row 7)
were excluded. This indicates that the graph features provide the most valuable
information, which is not covered by the basic and manual features.
The paired t-test performed using the RMSE values revealed that the majority of differences were significant, p<0.001. The insignificant differences are
highlighted in Figure 15. Three conclusions can be drawn from the insignificant
pairs: (1) the bipartite features are comparable to all the graph features, which
indicates the low contribution of the tripartite features; (2) all graph features
are comparable to all features except for manual (i.e., graph and basic features),
which indicates the low contribution of the basic features; and, (3) the combination of manual and bipartite features is comparable to the combination of
all the features, which indicates that the former two are the most informative
features in this case.
5.1.3. Dataset III – Yelp II (with social links) Results
The results of the evaluation using the extended Yelp II dataset that includes social links between users are in line with the results of the original Yelp
dataset. Table 8 lists the results of this evaluation for a selected set of feature
combinations: basic user and business features, feature of the complete graph,
features of all the sub-graphs, and the union of all the available features.
The results show that the combinations including graph features generally
outperform the basic feature sets. The best performing combination of graphbased features only, using all the features from all the sub-graph schemes (row 2,
RMSE=1.1417), achieves a 1.73% improvement over the baselines. When adding
the basic features to all the graph-based features, a slightly lower RMSE=1.1416
38
1
2
3
4
5
6
Features Subset
RMSE
Improvement
All Features
All Graph Features
Complete Graph
Business Features
User Features
Basic Features
1.1416
1.1417
1.1450
1.1465
1.1580
1.1619
1.73%
1.73%
1.45%
1.32%
0.33%
N/A
Table 8: RMSE of selected feature combinations - Yelp II dataset (baseline combination in
light gray).
(row 1) is obtained. Another noticeable difference is between the businessrelated features, which achieve RMSE=1.1465 and the user-related features,
which achieve RMSE=1.158 (rows 4 and 5, respectively). This intuitively indicates that the predicted ratings assigned to the businesses being predicted
are more informative than the ratings of the target user. Again, the achieved
improvements are generally modest, primarily due to the low variance of ratings
in the Yelp II dataset.
The performance differences between the evaluated combinations are mostly
significant, p<0.01, except for two pairs of feature sets. The difference between
business-related features and complete graph features is borderline, with p=0.07.
Also the difference between ‘All Features’ and ‘All Graph Features’ is expectedly
insignificant. This shows that the most important contribution to the predictive
accuracy comes from the graph features, while the addition of the basic features
improves the prediction only a little.
5.1.4. Dataset IV – OSN Results
In the fourth dataset of online social networks, user mentions of interests
in their profiles were predicted. Table 9 shows the precision scores achieved
by individual features listed in Section 4.4, as well as by a number of their
combinations. These features include the individual interest- and user-focused
graph features IG and UG; basic interest and graph features IB and UB; their
unions IG All, UG All, IB All, and UB All; as well as I All = IG All ∪ IB All
and U All = UG All ∪ UB All. The baseline here is the prediction using the
available user-interest features only.
Overall, item-related features were again seen to improve the precision of
the predictions more than user-related features. In fact, the accuracy of all
the item features is above the baseline precision of 0.56, while the accuracy
of all the user features is below the baseline. This can also be seen through
the comparison of the union of all item features I All with the union of user
features U All, which shows the superiority of the former. Zooming on the item
features, it can be observed that the graph-based item features IG outperform
most of the basic item features IB, except for IB1 (the number of users who
mentioned the interest), which turns out to be a reliable predictor. As a result,
the union of item-related graph features IG All obtains a higher precision than
39
Feature/Group
All
IG All
IG1
IG2
IG3
IB1
I All
IB All
IG4
IG5
IB2
IB3
Precision
0.6455
0.5821
0.5745
0.5734
0.5691
0.5687
0.5642
0.5599
0.5589
0.5560
0.5482
0.5307
Feature/Group
IB4
IB5
IB6
IB7
IB8
IB9
Baseline
SP
UB1
UB2
UB3
UB4
Precision
0.5287
0.5230
0.5221
0.5215
0.5206
0.5205
0.5128
0.5107
0.4771
0.4736
0.4646
0.4632
Feature/Group
UB5
UG1
UB6
UG All
UG3
UG2
UB7
UG4
UG5
UB All
U All
Precision
0.4529
0.4514
0.4507
0.4465
0.4440
0.4430
0.4405
0.4401
0.4400
0.4391
0.4376
Table 9: Average precision for individual features and feature combinations in the OSN dataset
IG - Interests Graph features, UG - Users Graph features, IB - Interests Basic (manual)
features, UB - Users Basic (manual) features.
100%
All
IB_All
IG_All
I_All
UB_All
UG_All
U_All
None
90%
CDF (%)
80%
70%
60%
50%
40%
30%
20%
0
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9
1
precision
Figure 16: Precision CDF of the various feature combinations - OSN dataset.
its basic feature counterpart IB All, 0.58 vs 0.56. Combining the graph features
with the basic ones produced the highest precision, 0.65. Hence, graph-based
features resulted in an improvement of the the interest predictions.
The cumulative distribution functions of the results obtained by selected
feature combinations is illustrated in Figure 16. The combined feature combination ‘All’, performs best by having the highest precision over large portions
of the data. The item-related graph-based feature combination is third best,
and it is very close to the combined graph-based feature set that comes second.
The performance of all user-related feature sets is visibly lower, which shows
another argument in favor of the extraction of the graph-based features.
5.1.5. Dataset V – Movielens Results
Finally, the experimentation with the Movielens dataset re-affirms the contribution of graph-based feature extraction to the recommendation generation.
The task in this dataset was to predict movie ratings, whereas the predictions
were evaluated using the MAE and RMSE predictive accuracy metrics. Table
10 summarizes the perofrmance of a selected group of features. The basic user40
1
2
3
4
5
Features Subset
RMSE
Improvement
MAE
Improvement
All Features
All Graph Features
Movie Features
Basic Features
User Features
1.0272
1.0362
1.0400
1.0722
1.0895
4.20%
3.36%
3.01%
N/A
-1.61%
0.8303
0.8349
0.8380
0.8838
0.8967
6.06%
5.53%
5.18%
N/A
-1.46%
Table 10: Performance of selected features combinations - Movielens dataset (baseline combination in light gray, rows are sorted by RMSE).
item pairs are compared here with the user and item features used individually,
all the extracted graph-based features, and the union of all of them, denoted by
‘All Features’ (row 1).
The already discussed superiority of item features over user features (row 3
versus row 5) can be clearly seen again. In this case, the former improve the
accuracy of the predictions by 3-5%, while the latter only deteriorate it. The
extraction of the graph-based features (row 2) also leads to an improvement of
3.36% and 5.53% relative to the baseline, using the RMSE and MAE metrics,
respectively. When combined with other features, the graph features achieve the
best result, which is RMSE=1.0272, or a 4.20% improvement over the baseline.
Those performance differences were statistically evaluated and found significant.
5.1.6. Performance across Learning Methods, Datasets, and Metrics
This case investigated the the impact of the graph-based features affect on
the accuracy of the recommendations. Although all the evaluations reported
so far show that using the graph-based features improves the accuracy of the
recommendations, the results cannot be fully corroborated yet, as the conducted
experiments use different learning methods, datasets, and evaluation metrics
(see Table 6). To confidently address the resarch question, the design of the
evaluation has overlaps in these factors, so that the contribution of the graph
features can be singled out.
The analysis below aims to establish whether the observed improvements
should be attributed to the information contributed by the graph features or
to the differences in the experimental settings, i.e., learning method, dataset,
and metric.The results of the experiments used in the analysis are summarized
in Table 11. In all the cases, the performance of the baseline approaches not
using the graph features, which were highlighted in light gray in all the tables,
is compared to the performance of all the graph-based based features, i.e, row
8 in Table 7, row 2 in Table 8, and row 2 in Table 10.
Included are the results of experiments using the Yelp, Yelp II, and Movielens
datasets, which were discussed in sections 5.1.2, 5.1.3, and 5.1.5, respectively.
That said, results in rows 3, 4, 5, and 7 of Table 11 are presented here for the first
time. This is due to the fact that previously reported Yelp experiments (both
datasets) used Random Forest as their learning method, while the Movielens
experiments used Graduate Boosting. Here, new Yelp II results with Gradient
41
Results
1
2
3
4
5
6
7
8
Dataset
Metric
Method
Yelp I
Yelp II
Yelp II
Yelp II
Movielens
Movielens
Movielens
Movielens
RMSE
RMSE
RMSE
RMSE
RMSE
RMSE
MAE
MAE
Random Forest
Random Forest
Gradient Boosting
SVM
Random Forest
Gradient Boosting
Random Forest
Gradient Boosting
Baseline
Graph Features
Improvement
1.1809
1.1619
1.2480
1.1818
1.1667
1.0722
0.9144
0.8838
1.1148
1.1417
1.1715
1.1783
1.0268
1.0362
0.8157
0.8349
5.59%
1.74%
6.13%
0.30%
11.90%
3.36%
10.79%
5.53%
Table 11: Summary of experiments and results for Case Study I.
Boosting and SVM, and new Movielens results with Random Forest are also
presented. It should also be highlighted that experiments using the Last.fm
and OSN datasets are excluded from the analysis, since, unlike the other three
experiments, they use classification accuracy metrics. As such, they differ in
two factors, dataset and evaluation metric, and are not comparable with the
other experiments.
In order to demonstrate that the improvement is not due to the selected
dataset, the metric and learning method were fixed, while the approaches using
different datasets were compared. Two evaluations sets are applicable to this
scenario: (1) Random Forest predictions evaluated with the RMSE metric, using
the Yelp, Yelp II, and Movielens datasets (rows 1, 2, and 5), and (2) Gradient
Boosting predictions also evaluated with RMSE, but using the Yelp II and
Movielens datasets (rows 3 and 6). The results of these experiments show an
improvement of 1.74% to 11.90%, which allows to eliminate the selected dataset
as a possible reason for improvement.
To demonstrate that the improvement is also not due to the selected machine
learning method, the dataset and metric were fixed, while the approaches using
different learning methods were compared. Three evaluation sets are applicable
to this scenario: (1) RMSE of business predictions using the Yelp II dataset,
where the learning methods are Random Forest, Gradient Boosting, and SVM
(rows 2, 3, and 4), (2) RMSE of movie rating predictions using the Movielens
dataset, where the methods are Random Forest and Gradient Boosting (rows
5 and 6), and (3) MAE of movie rating predictions using the the Movielens
dataset, where the methods are Random Forest and Gradient Boosting (rows
7 and 8). The results of these experiments show an improvement across all
experiments, ranging from 0.30% to 6.13% for the Yelp II dataset, and from
from 3.36% to 11.90% for the Movielens dataset. The low variance of ratings in
the Yelp datasets, which was discussed earlier, is the main reason for the low
improvement observed. This is particularly noticeable with the SVM method,
which struggles to linearly separate businesses with moderate ratings. Thus,
the learning method cannot be the reason for the accuracy improvement.
42
Finally, to demonstrate that the improvement is not due to the selected
evaluation metric, the dataset and method were fixed, while the performance
of approaches using different metrics was compated. Two evaluation sets are
applicable to this scenario: (1) Random Forest movie rating predictions using the Movielens dataset, evaluated using RMSE and MAE (rows 5 and 7),
and (2) Gradient Boosting movie rating predictions also using the Movielens
dataset, and also evaluated using the RMSE and MAE metrics (rows 6 and
8). The results of these experiments show a clear improvement across, ranging
from 3.36% to 11.90%, allowing to eliminate the selected evaluation metric as a
possible reason for improvement.
Summing up this causal analysis, all three hypotheses that the improved
performance is driven by the differences in the experimental settings (dataset,
learning method, and evaluation metric) were rejected. Thus, it can be concluded that the reason for the observed improvement lies in the inclusion of
graph-based features, contributing new information to the recommendation process.
5.2. Case Study II: Different Graph Schemes and their Impact on Recommendations
As mentioned in the Section 3, various sub-graphs and graph schemes can
be generated for each dataset. The feature extraction process will, thus, yield
a number of graph schemes, corresponding feature sets, and even the values of
the same graph features. This leads to to the second research question: How
are the recommendations affected by the sub-graph and its representation used
to generate the graph features? In order to answer this question, another set of
experiments was conducted.
In these experiments, the accuracy of recommendations when using various
graph schemes was evaluated using four datasets: Last.fm, both Yelp datasets,
and Movielens. The OSN dataset was not used here because it was represented
as a simple bipartite graph, lacking the desired number of entities and relationships. The recommendation tasks were identical to the previous experiments,
i.e., to predict listened artists in the Last.fm dataset, user ratings for businesses
in the two Yelp datasets and user ratings for movies in the Movielens dataset.
An N-fold cross validation methodology similar to the one reported in Section
5.1 was followed. Also, the same two-sided t-test statistical significance testing
was carried out.
5.2.1. Dataset I – Last.fm Results
The evaluations using the Last.fm dataset focused on the influence of the
social elements, i.e., friendship links and tags, on the obtained recommendation accuracy. In this dataset, the results of recommendations based on the
bipartite user-artist graph representation (BL in Figure 7) were compared with
those of three non-bipartite schemes, BL+T, BL+F, and BL+T+F, including,
respectively, the tags assigned by the users to the artists, social friendship links
between the users, and tags and friendship links alike. As mentioned in Section
43
Figure 17: Last.fm results: Precision of the extended (solid) versus the basic (dashed) feature
sets.
5.1.1, two sets of graph features were extracted for each schema: a set of basic
features F̂ and a set of extended features F . Although the basic feature set F̂
is shared across all the schemes, their values may change due to the presence of
additional graph nodes. The extended feature set F is composed of the basic
features along with new features that were extracted from the social links and
tags available in each schema. A detailed discussion of the extended feature set
can be found in (Tiroshi et al., 2014b).
Figure 17 shows the obtained P@10 scores averaged over all the users in
the test set, when using both the basic and extended feature sets. For each
representation, the solid boxes on the left denote the results obtained with the
extended features F , whereas the dotted boxes on the right present the results
obtained with the basic features F̂ . First, it can clearly be observed that the
inclusion of the social auxiliary data of either the assigned tags or friendships
links substantially improves P@10. When both the tags and friendship links are
included in the BL+T+F model, the highest average P@10 is observed. Both
in the basic and the extended feature sets, the BL+T and BL+F models obtain
comparable P@10 scores, showing the effect of the inclusion of auxiliary data in
the graph schemes. However, as noted in Section 4.1, the tag data includes more
than 186K tag assignments, whereas the friendship data consists of only 12K
user-to-user links. Since the obtained precision scores are comparable, a single
friendship link is more influential than a single artist tag and yields a greater
improvement in the recommendation accuracy. Looking at the significance tests
conducted within the basic and extended feature sets, significant differences,
p<0.05, were observed between all the pairs of extended features and all the
pairs of basic features except for the F̂BL+T and F̂BL+F pair.
When comparing the performance of the extended graph features to the performance of the corresponding basic features (solid boxes versus dashed boxes
in Figure 17, it can be seen that the extended sets consistently outperformed
the basic sets across all the four graph schemes, and the difference within the
pairs was statistically significant, p<0.05. In the BL+T scheme, the extended
graph features from improved on the basic features extracted from it by 10%,
44
8
10
11
12
Features combination
Features
RMSE
Improvement
All Graph
Bipartite
Tripartite
Basic
Bipartite∪Tripartite
1.1148
1.1188
1.1326
1.1809
5.59%
5.25%
4.09%
N/A
Table 12: Yelp results: RMSE of the bipartite versus the tripartite feature sets. Full results
are given in Table 7.
P@10=0.548 versus P@10=0.498, while in the BL+F scheme the improvement
was by 11.6%, P@10=0.555 versus P@10=0.497. The largest improvement was
noted in the BL+T+F scheme, where the extended graph features outperformed
the basic features by as much as 28.6%, P@10=0.571 versus P@10=0.444. Surprisingly, when the basic feature set F̂BL+T +F set was found achieve a lower
P@10 than F̂BL+T and F̂BL+F . A possible explanation for this can be that
including both types of social data but not extracting and populating the extended features leads to redundancy in the graph and degrades the performance
of the recommender.
5.2.2. Dataset II – Yelp Results
For the Yelp dataset and the task of business rating prediction, two graph
schemes were compared: a pure bipartite graph that contained only the users
and businesses, and a tripartite graph that, on top of user and business nodes,
also contained metadata nodes describing the businesses. The two graph schemes
are illustrated in Figure 9. The reason these were the only graph schemes created is that sparse features having a small number of unique features, were
filtered from the dataset. These features would have resulted in most of the
nodes of a group, e.g., users, being connected to a single node, which would
render it meaningless. For example, adding three “gender” nodes, male, female,
and unspecified, would have resulted in all users being connected to either one
of the three, essentially creating three large clusters in the graph.
The complete set of graph features was generated for both the bipartite and
tripartite representations. The results in Table 12 show the RMSE scores obtained for these feature sets. Note that these results are essentially extracted
from the results presented in Table 7 and their original row numbers are preserved. The experiments showed that the bipartite schema, not including the
metadata nodes, performed slightly but significantly better than the tripartite
schema with metadata, RMSE=1.1188 versus RMSE=1.1326. The relative improvement with respect to the baseline recommendations was 1.16% higher.
This difference in the performance of the schemas led to their unified feature
set, which is the All Graph, to outperform the two feature sets individually.
However, the superiority of All Graph was statistically significant only when
compared to the tripartite schema, as can be seen in Figure 15.
45
Features Subset
RMSE
Improvement
2
All Graph Features
1.1417
1.73%
7
3
8
9
Without Name Links
Complete Graph
Without Social Links
Without Social and Name Links
1.1450
1.1450
1.1463
1.1465
1.45%
1.45%
1.33%
1.32%
10
11
12
13
Without
Without
Without
Without
1.1508
1.1508
1.1519
1.1523
0.95%
0.94%
0.85%
0.82%
6
Basic Features
1.1619
N/A
Category Links
Metadata
Social and Category Links
Metadata and Social Links
Table 13: Yelp II results: RMSE of various sub-graph feature sets.
5.2.3. Dataset III – Yelp II (with social links) Results
The richer information provided by the Yelp II datasets allowed for the
creation of a larger set of sub-graphs. These are illustrated in Figure 11, where
various combinations of entities are removed from the complete graph. Thus, in
addition to the complete graph, seven sub-graph representations can be created
and the performance of the feature sets extracted from these can be compared.
The results of this experiment are presented in Table 13. The complete graph
and the seven sub-graphs are compared to the basic feature set and the union of
all the graph features, which were, respectively, the baseline and best performing
combination in Table 8. The numbering of rows already presented in Table 8 is
preserved (rows 2, 3, and 6), while the rows of all the sub-graphs from Figure 11
are numbered 7 to 13. The significance of the differences between the sub-graphs
is shown in Figure 18.
As can be clearly seen, the results of the various sub-graphs fell into two
groups, based on the significance tests. The groups were: sub-graphs containing
the ‘category’ relationship (“Without Name Links”, “Complete Graph”, “Without Social Links”, and “Without Social and Name Links”) and sub-graphs not
containing the ‘category’ relationship (“Without Category Links”, “Without
Metadata”, “Without Social and Category Links”, and “Without Metadata
and Social Links”). The former group of sub-graphs (rows 7, 3, 8, and 9 in
Table 13) performed significantly better than the latter (rows 10, 11, 12, and
13), which highlights the importance of business categories in predicting the
business ratings. This is also in line with the dominance of business features
over the user features that was already observed in Table 8. The union of all the
graph-based features extracted from all the sub-graphs (“All Graph Features”
in row 2) expectedly outperformed all other sub-graphs and feature sets. This
highlights the strength of the proposed approach in producing all the possible
features from all the possible sub-graph representations of the data rather than
identifying the optimal sub-graph and dealing with feature selection.
46
All Features
All Graph Features
Without Name Links
Full Graph
Without Social Links
Business Metrics
Without Social and Name Links
Without Category Links
Without Metadata
Without Social and Category Links
Without Metadata and Social Links
User Metrics
Basic Features
All Features 1.00 0.97
All Graph Features 0.97 1.00
1.00 0.94 0.14 0.06 0.11
Without Name Links
0.94 1.00 0.17 0.07 0.12
Full Graph
0.14 0.17 1.00 0.67 0.87
Without Social Links
0.06 0.07 0.67 1.00 0.79
Business Metrics
0.11 0.12 0.87 0.79 1.00
Without Social and Name Links
1.00 0.97 0.25 0.09
Without Category Links
0.97 1.00 0.27 0.10
Without Metadata
0.25 0.27 1.00 0.59
Without Social and Category Links
0.09 0.10 0.59 1.00
Without Metadata and Social Links
1.00
User Metrics
1.00
Basic Features
Figure 18: Significance of the differences between feature combinations in the Yelp II dataset.
White cells - significant, dark cells - not significant, p-value given.
5.2.4. Dataset V – Movielens Results
The Movielens dataset offered an even richer information about users and
items and allowed for the extraction of 32 sub-graph schemes. Only a small
sample of these is illustrated in Figure 13. The MAE and RMSE scores obtained
for the 32 sub-graphs are listed in Table 14 and the significance test results are
given in Figure 19. The sub-graphs are compared to the basic feature set and the
union of all the graph features, which were presented in Table 10 (rows numbered
2 and 4). The rows corresponding to the various sub-graph representations
are numbered 6 to 36. For the sake of clarity, the sub-graphs are denoted by
the entity types included rather than excluded. For example, “graph w/[Age,
Genre, Zip]” denotes the sub-graph with the ‘Age’, ‘Genre’, and ‘Zip’ entities,
which is identical to the complete graph with the ‘Occupation’ and ‘Gender’
entities excluded. In Figure 19, the names of the included entities are further
abbreviated, as detailed in the caption.
The significance test shows that the ‘genre’ relationship in Movielens subgraphs plays a similar role to the “category” relationship in Yelp. Sub-graphs
containing this relationship (rows 6 to 22) outperformed those, where it was excluded (rows 23 to 36), and the differences between the groups are significant. A
common link between the ‘category’ relationship in Yelp II and the ‘genre’ relationship in MovieLens is that they both divide the item space – be it businesses
or movies – into connected groups, which affects values of the item features.
In agreement with previous results, the feature set that unifies all the graph
features from all the sub-graph schemes (“All Graph Features”, row 2) achieves
the highest accuracy and outperforms any other feature set. Again, this is at-
47
2
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
4
Features Set
RMSE
Improvement
MAE
Improvement
All Graph Features
graph w/[Age, Gender, Genre, Zip]
graph w/[Age, Gender, Occupation, Zip]
graph w/[Gender, Genre, Occupation, Zip]
graph w/[Genre, Occupation]
graph w/[Age, Genre, Zip]
graph w/[Genre, Occupation, Zip]
graph w/[Age, Genre, Occupation]
graph w/[Age, Gender, Genre]
graph w/[Age, Gender, Genre, Occupation]
graph w/[Age, Genre, Occupation, Zip]
graph w/[Genre]
graph w/[Gender, Genre, Occupation]
graph w/[Age, Genre]
graph w/[Age, Gender, Genre, Occupation, Zip]
graph w/[Genre, Zip]
graph w/[Gender, Genre]
graph w/[Gender, Genre, Zip]
graph w/[Age]
graph w/[Zip]
graph w/[Age, Zip]
graph w/[Age, Occupation]
graph w/[Age, Gender]
graph w/[Age, Gender, Zip]
graph w/[Age, Occupation, Zip]
graph w/[Occupation, Zip]
graph w/[Occupation]
graph w/[Gender, Occupation, Zip]
graph w/[Gender, Zip]
graph w/[Gender]
graph w/[Gender, Occupation]
graph w/[Age, Gender, Occupation]
Basic Features
1.0362
1.0369
1.0373
1.0384
1.0410
1.0411
1.0411
1.0412
1.0412
1.0413
1.0413
1.0413
1.0413
1.0414
1.0414
1.0415
1.0416
1.0416
1.0425
1.0426
1.0426
1.0426
1.0427
1.0427
1.0427
1.0427
1.0428
1.0428
1.0431
1.0431
1.0432
1.0433
1.0722
3.36%
3.29%
3.25%
3.16%
2.91%
2.90%
2.90%
2.90%
2.89%
2.89%
2.89%
2.89%
2.88%
2.88%
2.88%
2.87%
2.86%
2.85%
2.77%
2.77%
2.76%
2.76%
2.76%
2.75%
2.75%
2.75%
2.75%
2.74%
2.72%
2.71%
2.70%
2.70%
N/A
0.8349
0.8353
0.8357
0.8365
0.8391
0.8386
0.8385
0.8390
0.8392
0.8390
0.8388
0.8393
0.8392
0.8395
0.8390
0.8388
0.8396
0.8391
0.8413
0.8407
0.8409
0.8412
0.8413
0.8408
0.8408
0.8410
0.8414
0.8409
0.8411
0.8418
0.8417
0.8417
0.8838
5.53%
5.48%
5.44%
5.35%
5.06%
5.12%
5.12%
5.07%
5.05%
5.07%
5.10%
5.04%
5.04%
5.02%
5.08%
5.09%
5.00%
5.06%
4.81%
4.88%
4.86%
4.82%
4.81%
4.87%
4.86%
4.84%
4.80%
4.85%
4.83%
4.75%
4.77%
4.77%
N/A
Table 14: Performance of selected features combinations - Movielens dataset (baseline combination in light gray, rows are sorted by RMSE).
tributed to the broad coverage of the proposed feature extraction mechanism,
which produces and aggregates promising feature combinations.
5.2.5. Summary
The purpose of this analysis was to analyze the differences driven by the subgraphs that are used for the feature extraction. To recap the results obtained
using the four datasets, the following was established.
• Features extracted from different graph schemes performed differently, not
following a certain pattern tied to the entities or relationships included in
48
All Features
All Graph Features
G w/[A,Gndr,Gnr,Z]
G w/[A,Gndr,O,Z]
G w/[Gndr,Gnr,O,Z]
Movie Features
G w/[Gnr,O]
G w/[A,Gnr,Z]
G w/[Gnr,O,Z]
G w/[A,Gnr,O]
G w/[A,Gndr,Gnr]
G w/[A,Gndr,Gnr,O]
G w/[A,Gnr,O,Z]
G w/[Gnr]
G w/[Gndr,Gnr,O]
G w/[A,Gnr]
G w/[A,Gndr,Gnr,O,Z]
G w/[Gnr,Z]
G w/[Gndr,Gnr]
G w/[Gndr,Gnr,Z]
G w/[A]
G w/[Z]
G w/[A,Z]
G w/[A,O]
G w/[A,Gndr]
G w/[A,Gndr,Z]
G w/[A,O,Z]
G w/[O,Z]
G w/[O]
G w/[Gndr,O,Z]
G w/[Gndr,Z]
G w/[Gndr]
G w/[Gndr,O]
G w/[A,Gndr,O]
Basic Features
User Features
All Features 1.00
All Graph Features 1.00
1.00
G w/[A,Gndr,Gnr,Z]
1.00
G w/[A,Gndr,O,Z]
1.00
G w/[Gndr,Gnr,O,Z]
1.00
Movie Features
1.000.850.900.680.580.490.360.440.350.180.200.120.110.06
G w/[Gnr,O]
0.851.000.950.820.720.640.490.560.450.250.290.170.160.09
G w/[A,Gnr,Z]
0.900.951.000.770.670.590.440.520.420.220.260.150.140.08
G w/[Gnr,O,Z]
0.680.820.771.000.890.840.660.730.600.350.420.250.230.14
G w/[A,Gnr,O]
0.580.720.670.891.000.960.780.830.700.430.520.310.290.18
G w/[A,Gndr,Gnr]
0.490.640.590.840.961.000.780.840.690.390.470.260.240.14
G w/[A,Gndr,Gnr,O]
0.360.490.440.660.780.781.000.980.870.520.650.370.340.21
G w/[A,Gnr,O,Z]
0.440.560.520.730.830.840.981.000.860.560.690.430.400.26
G w/[Gnr]
0.350.450.420.600.700.690.870.861.000.680.840.530.500.34
G w/[Gndr,Gnr,O]
0.180.250.220.350.430.390.520.560.681.000.790.830.790.59
G w/[A,Gnr]
0.200.290.260.420.520.470.650.690.840.791.000.600.560.37
G w/[A,Gndr,Gnr,O,Z]
0.120.170.150.250.310.260.370.430.530.830.601.000.960.75
G w/[Gnr,Z]
0.110.160.140.230.290.240.340.400.500.790.560.961.000.78
G w/[Gndr,Gnr]
0.060.090.080.140.180.140.210.260.340.590.370.750.781.00
G w/[Gndr,Gnr,Z]
1.000.760.630.660.760.490.440.460.360.250.08
G w/[A]
0.761.000.850.891.000.730.630.660.540.420.150.09
G w/[Z]
0.630.851.000.960.850.900.770.800.670.560.210.140.07
G w/[A,Z]
0.660.890.961.000.890.850.730.760.630.520.190.120.06
G w/[A,O]
0.761.000.850.891.000.730.630.660.540.420.150.09
G w/[A,Gndr]
0.490.730.900.850.731.000.840.870.710.570.180.110.05
G w/[A,Gndr,Z]
0.440.630.770.730.630.841.000.970.890.800.330.230.130.08
G w/[A,O,Z]
0.460.660.800.760.660.870.971.000.860.770.310.220.120.07
G w/[O,Z]
0.360.540.670.630.540.710.890.861.000.930.400.290.170.11
G w/[O]
0.250.420.560.520.420.570.800.770.931.000.380.260.140.08
G w/[Gndr,O,Z]
0.080.150.210.190.150.180.330.310.400.381.000.830.600.44
G w/[Gndr,Z]
0.090.140.120.090.110.230.220.290.260.831.000.760.58
G w/[Gndr]
0.070.06 0.050.130.120.170.140.600.761.000.81
G w/[Gndr,O]
0.080.070.110.080.440.580.811.00
G w/[A,Gndr,O]
1.00
Basic Features
1.00
User Features
Figure 19: Significance of the differences between feature combinations in the Movielens
dataset. White cells - significant, dark cells - not significant, p-value given. “G w/[]” denotes
sub-graphs that contain the listed entity types, where: A=Age, Gndr=Gender, Gnr=Genre,
O=Occupation, and Z=Zip.
the sub-graph. This means that it was not possible to conclude which
relationships lead to better results if included in the graph. We posit that
this is dataset-specific and may be affected by additional factors, such
as density of a specific feature, distribution of its values, domain-specific
considerations, and so forth. This finding comes through in the ‘category’
and ‘genre’ relationships in the Yelp II and Movielens datasets, but not
in the Yelp I dataset. Notably, the social links had a major contribution
in the Last.fm dataset, but not in the Yelp II dataset, possibly due to the
sparsity of the latter.
• Features extracted from the complete graph representations, i.e., those
containing all the relationships and entities in the dataset, were not necessarily the best performing feature sets. A negative example can be seen
in the basic features of the BL+F+T schema in Figure 17 that are dominated by the basic feature of BL+F and BL+T alike. Having said that, the
49
feature set that aggregated (that is, unified) all the graph features from
all the sub-graph schemes performed the best in the other three scenarios
in which it was evaluated: Yelp, Yelp II, and Movielens. We consider
this to be a strong argument in favor of using the proposed approach, as
its exhaustive nature allows to cover a range of features and necessarily
uncover the most informative ones, as well as their best combination.
The differences across the obtained results do not allow to generalize and
determine a priori the best performing sub-graph and feature set. Due to this,
the suggested approach of generating sub-graphs, populating features from each
of them, and then aggregating the features in the feature sets is more likely to
uncover the best performing feature combination. Note that this trades off with
computational overheads and potential scalability issues in large-scale datasets
(discussed in detail in Section 6.1.2). We believe that future research may unveil
rating patterns or characteristics of datasets, which may predict the contribution
of certain sub-graph, data entities, or even types of features.
6. Discussion and Conclusions
6.1. Discussion
The effectiveness of the graph-based approach for improving recommendations was demonstrated in the previous sections. It has been shown that precision and accuracy gains can be achieved by representing tabular data by graphs
and extracting new features from them. This contrasts and complements prior
approaches that improved recommendations by enhancing the recommendation
techniques themselves. Also established are the benefits of the graph-based
approach across recommendation domains, tasks, and metrics. These findings
show that the graph representation exploits indirect latent links in the data,
which lead to an improved recommendation accuracy. Finally, the approach is
generic and it can be applied to many recommender system datasets.
The suggested process is automatic and can be run end-to-end, from data
representation to feature extraction, without human intervention, unlike manual
feature extraction methods, which are often time consuming and requires domain expertise. Using the proposed graph-based approach, rich features, based
on intricate relationships between various data entities and sub-graph scheme
variations, can be systematically extracted from a dataset. This allows for a
better coverage of the features space with a considerable lower effort, as discussed in detail in Section 3. In the following sub-sections, the key limitations
and challenges encountered in the experiments and case studies are discussed.
6.1.1. Overfitting
Regarding concerns referring to possible overfitting due to the newly generated features, as long as the volume of available data greatly exceeds the number
of extracted features, there is little risk that the features will be the cause of
overfitting. The high diversity of unique data characteristics can hardly be captured in full by a smaller subset of features. Recommender system datasets tend
50
to be in the medium to large scale (tens of thousands to millions of data points),
while the number of features generated by the proposed approach is still in the
scale of tens to hundreds.
Additionally, machine learning methods such as Random Forests have internal mechanisms for feature selection and can filter out features that overfit.
They do so by training on a sample of the dataset and evaluating the performance of the features on the rest of the data. A feature that performs well
on the sample but underperforms on the test data is ranked low. In the evaluations, cross validation was used with at least N=5 folds, showing that the
models and features on which they are built are in fact generalizable. Moreover, it was shown that in cases of sparse data, which require a higher degree of
generalization, the graph features still outperformed other features.
6.1.2. Scalability
A possible disadvantage of the proposed approach is that some graph-based
computations, e.g., PageRank, are iterative and may take a long time to converge. In the age of Big Data, recommender system datasets are getting large
and this limitation may become a hurdle. The representation of the datasets
results in large graphs and the computation issue becomes a bigger problem.
A general approach for handling this issue in a deployed system would be to
extract the graph-based features offline, say, on a nightly basis, and use the precomputed values for real-time predictions. This may resolve the problem under
the reasonable assumption that the values do not change substantially too frequently. Another means to overcome the computational latency is through using
a distributed graph feature computation library. Such a library, e.g., Okapi13 ,
can use distributed tools in order extract the graph features.
Another factor that adds to the computational complexity of the approach
is the exhaustive search for new features. It should be noted that the complexity of the process of generating every possible sub-graph and populating the
matching feature combinations is exponential. The number of relationships in
current recommendation datasets (as surveyed in Section 3.1) is still manageable, and can be accommodated by the proposed approach. However in the
future, with additional data sources being integrated for recommendation purposes, this might become unsustainable and will require a long-term solution.
Two possible approaches for handling this issue are parallelization, e.g., each
sub-graph being processed by a different machine, and heuristics for pruning
less relevant sub-graph representations.
6.1.3. Initial Transition to the Graph Model
Another possible disadvantage of graph-based features is the possible need
for human intervention when generating the initial complete graph. Non-categorical
feature values, e.g., income or price, may generate a large number of vertices,
which would lead to a low connectivity of the graph, since not many users or
13 http://grafos.ml/okapi.html
51
items would share the exact value of the feature. This would lead to a very
sparse graph and will need to be addressed by a manual intervention by a domain expert, who can determine how the non-categorical values can be grouped
and categorized, e.g., by creating appropriate income or price buckets. A naive
solution for this might be to attempt to auto-categorize such features based on
the observed distribution of their values, e.g., first quarter, second quarter, and
so on. This may, however, mask the differences between fine-grained groups and
cause information loss.
Also to be acknowledged in this context is the historic human contribution
that was required in order to conceive the graph methods exploited in this work
for the generation of the various basic graph features: shortest path, degree,
PageRank, etc. Indeed, these methods took a considerable amount of time and
effort to evolve; however, they are reusable for generations and the overheads
related to their development have been shared across many subsequent applications, while manually engineered features would usually not be highly reusable.
Overall, when weighting the ease, quantity, and the possible contribution of the
graph-based features to the accuracy of the generated recommendations against
the above mentioned disadvantages, it can be concluded that it is worth to
generate and populate such features, when designing a recommendation engine.
6.2. Conclusions and Future Work
In this work, a new approach for improving recommendations was presented
and evaluated. Unlike many previous works, which focused on addressing the
recommendation problem by making improvements to the recommendation algorithms, the presented approach does so by suggesting a different way of looking
at the dataset used for recommendation. It proposed representing the datasets
using graphs and then to extract and populate new features from those graphs,
all in a systematic fashion, and feed the new features into existing recommendation algorithms. New features and relationships that were not visible in the
original tabular form can be thus uncovered. In this manner, applying this
approach may compliment classical recommendation approaches and further
enhance them.
The methodology, implementation, and analysis of the approach were described in detail and the approach was evaluated from two main perspectives:
overall contribution to recommendations and impact of various graph representations. The evaluation encompassed a number of datasets, recommendation
tasks, and evaluation metrics. Furthermore, the datasets belonged to four application domains (movies, music, businesses, and personal interests) that in
part included metadata and in part included social links. The recommendation
tasks varied from binary link predictions to star rating predictions. A number
of state-of-the-art classifiers and regressors were used for the generation of the
predictions. All in all, the presented evaluations examined the impact of the
graph representations and showed that the approach had a profound effect on
the accuracy of the recommendations.
The graph-based representation and features were shown to lead to the generation of more accurate recommendations. The variations in performance across
52
various graph schemes and the justification for systematically extracting them,
due to that, was established. The approach presented was implemented in a
library and is being provided as open source software for the community to use
and build on-top. Given such a library, the cost of generating additional features that can improve recommendations becomes substantially lower, in terms
of computation time and effort. It can be adopted as a natural first resort,
when given a dataset and recommendation task, or as a complementary aid to
enhance the standard manual feature engineering.
The conducted evaluations suggested and demonstrated the potential of the
proposed approach in improving the recommendations by exploiting the benefits
of links between entities and characteristics of entities extracted from the graph
representations. Therefore, this work lays the foundations for further exploring how graph-based features can enhance recommender systems and automatic
feature engineering in the more general context. Several variables were investigated in this work but many more require additional attention. The following
paragraphs identify several directions of exploration, which were identified as
possible research directions in future works.
• Temporal Aspects. Given a dataset that includs dated actions that are
not sparse, the time aspect can be used to build a different type of graphs.
Each graph will represent a snapshot in time and will either contain or
exclude a link between vertices based on whether it was available in the
dataset at that time. A combination of two temporally adjacent graphs
will reflect the evolution of the data over that period of time. The main
question in this setting is how such temporal graphs will affect the values
of features extracted from them and how a recommender systems that use
these features will perform in their respective recommendation tasks.
• Weighted and Labeled Graphs. Several features in a dataset can be used
to populate the edge labels when constructing the graph based representation. The labels, once set, can be taken into consideration in some graph
features being extracted. One example would be to calculate a weighted
PageRank score that will have jumps from a vertex to its neighbors based
on a skewed probability correlated with the weight on the edge linking to
the neighbor. This could lead to further improvement in the recommendations; however, this requires fine-tuning of initial weights on edges that
do not naturally have them, e.g., social relationship edges in the Last.fm
dataset.
• Directed Graphs. Similarly, in cases where the direction of the edges can
be important, the process can be extended to include this aspect by generating additional graph representations, with various combinations of the
edge directions. For example, in one variant, edges will be directed from
the source vertex to the target vertex, in another, in the opposite direction,
and in a third one there will be no direction. This will guarantee coverage
in terms of expressing the direction of the edges, and the performance of
the features in the various scenarios can be evaluated.
53
The effects of these modifications on the scalability of the approach can be
handled using the previously suggested methods, either by scaling the computations (e.g., computing the features of each subgraph in a separate process),
or using distributed graph computation libraries, or identifying heuristics for
pruning the feature and subgraph space.
References
Adomavicius, G. and Tuzhilin, A. (2005). Toward the next generation of recommender
systems: A survey of the state-of-the-art and possible extensions. Knowledge and
Data Engineering, IEEE Transactions on, 17(6):734–749.
Amatriain, X., Jaimes, A., Oliver, N., and Pujol, J. M. (2011). Data mining methods for recommender systems. In Recommender Systems Handbook, pages 39–71.
Springer.
Bae, D., Han, K., Park, J., and Yi, M. Y. (2015). Apptrends: A graph-based mobile
app recommendation system using usage history. In 2015 International Conference
on Big Data and Smart Computing, BIGCOMP 2015, Jeju, South Korea, February
9-11, 2015, pages 210–216.
Barrat, A., Barthelemy, M., Pastor-Satorras, R., and Vespignani, A. (2004). The
Architecture of Complex Weighted Networks. PNAS.
Bastian, M., Heymann, S., Jacomy, M., et al. (2009). Gephi: an open source software
for exploring and manipulating networks. ICWSM, 8:361–362.
Batarfi, O., Shawi, R. E., Fayoumi, A. G., Nouri, R., Beheshti, S., Barnawi, A., and
Sakr, S. (2015). Large scale graph processing systems: survey and an experimental
evaluation. Cluster Computing, 18(3):1189–1213.
Bellogı́n, A., Cantador, I., Dı́ez, F., Castells, P., and Chavarriaga, E. (2013). An
empirical comparison of social, collaborative filtering, and hybrid recommenders.
ACM Transactions on Intelligent Systems and Technology (TIST), 4(1):14.
Bennett, J. and Lanning, S. (2007). The netflix prize. In Proceedings of KDD cup and
workshop, volume 2007, page 35.
Benzi, K., Kalofolias, V., Bresson, X., and Vandergheynst, P. (2016). Song recommendation with non-negative matrix factorization and graph total variation. In
2016 IEEE International Conference on Acoustics, Speech and Signal Processing,
ICASSP 2016, Shanghai, China, March 20-25, 2016, pages 2439–2443.
Berge, C. and Minieka, E. (1973). Graphs and hypergraphs, volume 7. North-Holland
publishing company Amsterdam.
Berkovsky, S. (2006). Decentralized mediation of user models for a better personalization. In Adaptive Hypermedia and Adaptive Web-Based Systems, pages 404–408.
Springer Berlin/Heidelberg.
Berkovsky, S., Aroyo, L., Heckmann, D., Houben, G.-J., Kröner, A., Kuflik, T., and
Ricci, F. (2007). Providing context-aware personalization through cross-context
reasoning of user modeling data. UBIDEUM 2007.
54
Berkovsky, S., Kuflik, T., and Ricci, F. (2008). Mediation of user models for enhanced personalization in recommender systems. User Modeling and User-Adapted
Interaction, 18(3):245–286.
Blomo, J., Ester, M., and Field, M. (2013). Recsys challenge 2013. In Proceedings of
the 7th ACM conference on Recommender systems, pages 489–490. ACM.
Bohnert, F., Zukerman, I., Berkovsky, S., Baldwin, T., and Sonenberg, L. (2008).
Using interest and transition models to predict visitor locations in museums. AI
Communications, 21(2-3):195–202.
Borgatti, S. P. and Halgin, D. S. (2011). Analyzing Affiliation Networks. The Sage
handbook of social network analysis.
Breiman, L. (2001). Random forests. Machine learning, 45(1):5–32.
Bu, J., Tan, S., Chen, C., Wang, C., Wu, H., Zhang, L., and He, X. (2010). Music
recommendation by unified hypergraph: combining social media information and
music content. In Proceedings of the international conference on Multimedia, pages
391–400. ACM.
Burke, R. (2007). Hybrid web recommender systems. In The adaptive web, pages
377–408. Springer.
Cantador, I., Brusilovsky, P., and Kuflik, T. (2011). Second workshop on information
heterogeneity and fusion in recommender systems (hetrec 2011). In RecSys.
Cantador, I., Fernández-Tobı́as, I., Berkovsky, S., and Cremonesi, P. (2015). Crossdomain recommender systems. In Recommender Systems Handbook.
Chianese, A. and Piccialli, F. (2016). A smart system to manage the context evolution
in the cultural heritage domain. Computers & Electrical Engineering, pages –.
Coffman, T., Greenblatt, S., and Marcus, S. (2004). Graph-based technologies for
intelligence analysis. Communications of the ACM, 47(3):45–47.
Cordobés, H., Chiroque, L. F., Anta, A. F., Leiva, R. A. G., Morere, P., Ornella, L.,
Pérez, F., and Santos, A. (2015). Empirical comparison of graph-based recommendation engines for an apps ecosystem. IJIMAI, 3(2):33–39.
Costa, L. d. F., Rodrigues, F. A., Travieso, G., and Villas Boas, P. R. (2007). Characterization of complex networks: A survey of measurements. Advances in Physics,
56(1):167–242.
Csardi, G. and Nepusz, T. (2006). The igraph software package for complex network
research. InterJournal, Complex Systems, 1695(5).
Desrosiers, C. and Karypis, G. (2011). A comprehensive survey of neighborhoodbased recommendation methods. In Recommender systems handbook, pages 107–
144. Springer.
Domingos, P. (2012). A few useful things to know about machine learning. Communications of the ACM, 55(10):78–87.
55
Due Trier, Ø., Jain, A. K., and Taxt, T. (1996). Feature extraction methods for
character recognition-a survey. Pattern recognition, 29(4):641–662.
Floyd, R. W. (1962). Algorithm 97: shortest path. Communications of the ACM,
5(6):345.
Friedman, J. H. (2000). Greedy function approximation: A gradient boosting machine.
Annals of Statistics, 29:1189–1232.
Garcia, C. and Tziritas, G. (1999). Face detection using quantized skin color regions
merging and wavelet packet analysis. Multimedia, IEEE Transactions on, 1(3):264–
277.
Godoy, D. and Corbellini, A. (2016). Folksonomy-based recommender systems: A
state-of-the-art review. Int. J. Intell. Syst., 31(4):314–346.
Gunn, S. R. et al. (1998). Support vector machines for classification and regression.
ISIS technical report, 14.
Guyon, I., Gunn, S., Nikravesh, M., and Zadeh, L. (2006). Feature extraction. Foundations and applications.
Hagberg, A., Swart, P., and S Chult, D. (2008). Exploring network structure, dynamics, and function using networkx. Technical report, LANL.
Hall, M., Frank, E., Holmes, G., Pfahringer, B., Reutemann, P., and Witten, I. H.
(2009). The weka data mining software: an update. ACM SIGKDD explorations
newsletter, 11(1):10–18.
Haupt, J. (2009). Last.fm: People-powered online radio. Music Reference Services
Quarterly, 12(1-2):23–24.
Hong, M. and Jung, J. J. (2016). Mymoviehistory: Social recommender system by
discovering social affinities among users. Cybernetics and Systems, 47(1-2):88–110.
Jahrer, M., Töscher, A., and Legenstein, R. (2010). Combining predictions for accurate
recommender systems. In Proceedings of the 16th ACM SIGKDD international
conference on Knowledge discovery and data mining, pages 693–702. ACM.
Jäschke, R., Marinho, L., Hotho, A., Schmidt-Thieme, L., and Stumme, G. (2007). Tag
recommendations in folksonomies. In Knowledge Discovery in Databases: PKDD
2007, pages 506–514. Springer.
Jiang, W., Wang, G., Bhuiyan, M. Z. A., and Wu, J. (2016). Understanding graphbased trust evaluation in online social networks: Methodologies and challenges.
ACM Comput. Surv., 49(1):1–35.
Kautz, H., Selman, B., and Shah, M. (1997). Referral web: combining social networks
and collaborative filtering. Communications of the ACM, 40(3):63–65.
Klema, V. and Laub, A. J. (1980). The singular value decomposition: Its computation
and some applications. Automatic Control, IEEE Transactions on, 25(2):164–176.
Kobsa, A. (2001). Generic user modeling systems. User modeling and user-adapted
interaction, 11(1-2):49–63.
56
Kohavi, R. et al. (1995). A study of cross-validation and bootstrap for accuracy
estimation and model selection. In IJCAI, volume 14, pages 1137–1145.
Kohavi, R. and John, G. H. (1997). Wrappers for feature subset selection. Artificial
intelligence, 97(1):273–324.
Konstas, I., Stathopoulos, V., and Jose, J. M. (2009). On social networks and collaborative recommendation. In SIGIR, pages 195–202.
Koren, Y. (2009). The bellkor solution to the netflix grand prize. Netflix prize documentation, 81.
Koren, Y., Bell, R., and Volinsky, C. (2009). Matrix factorization techniques for
recommender systems. Computer, 42(8):30–37.
Kotsiantis, S., Kanellopoulos, D., and Pintelas, P. (2006). Data preprocessing for
supervised leaning. International Journal of Computer Science, 1(2):111–117.
Lam, S. and Herlocker, J. (2012). Movielens 1m dataset.
Latapy, M., Magnien, C., and Vecchio, N. D. (2008). Basic notions for the analysis of
large two-mode networks. Social Networks.
Lee, K. and Lee, K. (2015). Escaping your comfort zone: A graph-based recommender
system for finding novel recommendations among relevant items. Expert Syst. Appl.,
42(10):4851–4858.
Lee, S., Kim, S., and Park, S. (2015). A graph-based recommendation framework
for price-comparison services. In Proceedings of the 24th International Conference
on World Wide Web Companion, WWW 2015, Florence, Italy, May 18-22, 2015 Companion Volume, pages 59–60.
Lee, S., Park, S., Kahng, M., and Lee, S. (2013). Pathrank: Ranking nodes on a
heterogeneous graph for flexible hybrid recommender systems. Expert Syst. Appl.,
40(2):684–697.
Lewis, K., Kaufman, J., Gonzalez, M., Wimmer, A., and Christakis, N. (2008). Tastes,
ties, and time: A new social network dataset using facebook. com. Social networks,
30(4):330–342.
Li, X. and Chen, H. (2009). Recommendation as link prediction: a graph kernelbased machine learning approach. In Proceedings of the 9th ACM/IEEE-CS joint
conference on Digital libraries, pages 213–216. ACM.
Lösch, U., Bloehdorn, S., and Rettinger, A. (2012). Graph kernels for rdf data. In
The Semantic Web: Research and Applications, pages 134–148. Springer.
Ma, H., King, I., and Lyu, M. R. (2009). Learning to recommend with social trust
ensemble. In Proceedings of the 32nd international ACM SIGIR conference on Research and development in information retrieval, pages 203–210. ACM.
Mao, K., Chen, G., Hu, Y., and Zhang, L. (2016). Music recommendation using graph
based quality model. Signal Processing, 120:806–813.
57
Markovitch, S. and Rosenstein, D. (2002). Feature generation using general constructor
functions. Machine Learning, 49(1):59–98.
Massa, P. and Avesani, P. (2007). Trust-aware recommender systems. In Proceedings
of the 2007 ACM conference on Recommender systems, pages 17–24. ACM.
Mobasher, B. (2007). Data mining for web personalization. In The Adaptive Web,
Methods and Strategies of Web Personalization, pages 90–135.
Moradi, P., Ahmadian, S., and Akhlaghian, F. (2015). An effective trust-based recommendation method using a novel graph clustering algorithm. Physica A: Statistical
Mechanics and its Applications, 436:462 – 481.
Nixon, M. (2008). Feature extraction & image processing. Academic Press.
O’Donovan, J. and Smyth, B. (2005). Trust in recommender systems. In Proceedings
of the 10th international conference on Intelligent user interfaces, pages 167–174.
ACM.
Ostuni, V. C., Noia, T. D., Sciascio, E. D., Oramas, S., and Serra, X. (2015). A
semantic hybrid approach for sound recommendation. In Proceedings of the 24th
International Conference on World Wide Web Companion, WWW 2015, Florence,
Italy, May 18-22, 2015 - Companion Volume, pages 85–86.
Page, L., Brin, S., Motwani, R., and Winograd, T. (1999). The pagerank citation
ranking: Bringing order to the web. Technical Report 1999-66, Stanford InfoLab.
Park, Y., Park, S., Jung, W., and Lee, S. (2015). Reversed CF: A fast collaborative
filtering algorithm using a k-nearest neighbor graph. Expert Syst. Appl., 42(8):4022–
4028.
Pedregosa, F., Varoquaux, G., Gramfort, A., Michel, V., Thirion, B., Grisel, O.,
Blondel, M., Prettenhofer, P., Weiss, R., Dubourg, V., et al. (2011). Scikit-learn:
Machine learning in python. The Journal of Machine Learning Research, 12:2825–
2830.
Perugini, S., Gonçalves, M. A., and Fox, E. A. (2004). Recommender systems research:
A connection-centric survey. Journal of Intelligent Information Systems, 23(2):107–
143.
Pham, T. N., Li, X., Cong, G., and Zhang, Z. (2015). A general graph-based model
for recommendation in event-based social networks. In 31st IEEE International
Conference on Data Engineering, ICDE 2015, Seoul, South Korea, April 13-17,
2015, pages 567–578.
Portilla, Y., Reiffers, A., Altman, E., and Azouzi, R. E. (2015). A study of youtube recommendation graph based on measurements and stochastic tools. In 8th IEEE/ACM
International Conference on Utility and Cloud Computing, UCC 2015, Limassol,
Cyprus, December 7-10, 2015, pages 430–435.
Quercia, D., Schifanella, R., and Aiello, L. M. (2014). The shortest path to happiness:
Recommending beautiful, quiet, and happy routes in the city. In Proceedings of
the 25th ACM Conference on Hypertext and Social Media, HT ’14, pages 116–125.
ACM.
58
Ricci, F., Rokach, L., and Shapira, B. (2011). Introduction to recommender systems
handbook. Springer.
Said, A., Berkovsky, S., and De Luca, E. W. (2011). Group recommendation in context.
In Proceedings of the 2nd Challenge on Context-Aware Movie Recommendation,
pages 2–4. ACM.
Schafer, J. B., Konstan, J., and Riedl, J. (1999). Recommender systems in e-commerce.
In Proceedings of the 1st ACM conference on Electronic commerce, pages 158–166.
ACM.
Schwartz, M. F. and Wood, D. (1993). Discovering shared interests using graph analysis. Communications of the ACM, 36(8):78–89.
Scott, S. and Matwin, S. (1999). Feature engineering for text classification. In ICML,
volume 99, pages 379–388.
Shams, B. and Haratizadeh, S. (2016). Graph-based collaborative ranking. CoRR,
abs/1604.03147.
Shani, G. and Gunawardana, A. (2011). Evaluating recommendation systems. In
Recommender Systems Handbook, pages 257–297. Springer.
Shen, J., Deng, C., and Gao, X. (2016). Attraction recommendation: Towards personalized tourism via collective intelligence. Neurocomputing, 173:789–798.
Tan, S., Bu, J., Chen, C., and He, X. (2011). Using rich social media information
for music recommendation via hypergraph model. In Social media modeling and
computing, pages 213–237. Springer.
Tiroshi, A., Berkovsky, S., Kaafar, M. A., Chen, T., and Kuflik, T. (2013). Cross
social networks interests predictions based on graph features. In Proceedings of the
7th ACM Conference on Recommender Systems, RecSys ’13, pages 319–322. ACM.
Tiroshi, A., Berkovsky, S., Kaafar, M. A., Vallet, D., Chen, T., and Kuflik, T. (2014a).
Improving business rating predictions using graph based features. In Proceedings of
the 19th international conference on Intelligent User Interfaces, pages 17–26. ACM.
Tiroshi, A., Berkovsky, S., Kaafar, M. A., Vallet, D., and Kuflik, T. (2014b). Graphbased recommendations: Make the most out of social data. In User Modeling,
Adaptation, and Personalization, pages 447–458. Springer.
Töscher, A., Jahrer, M., and Bell, R. M. (2009). The bigchaos solution to the netflix
grand prize. Netflix prize documentation.
Vahedian, F., Burke, R. D., and
tended multi-relational matrix
International Florida Artificial
2016, Key Largo, Florida, May
Mobasher, B. (2016). Meta-path selection for exfactorization. In Proceedings of the Twenty-Ninth
Intelligence Research Society Conference, FLAIRS
16-18, 2016., pages 566–571.
Wasserman, S. (1994). Social network analysis: Methods and applications, volume 8.
Cambridge university press.
West, D. B. et al. (2001). Introduction to graph theory, volume 2. Prentice hall Upper
Saddle River.
59
Wilson, C., Boe, B., Sala, A., Puttaswamy, K. P., and Zhao, B. Y. (2009). User
interactions in social networks and their implications. In Proceedings of the 4th
ACM European conference on Computer systems, pages 205–218. Acm.
Wold, S., Esbensen, K., and Geladi, P. (1987). Principal component analysis. Chemometrics and intelligent laboratory systems, 2(1):37–52.
Wu, H., Yue, K., Liu, X., Pei, Y., and Li, B. (2015). Context-aware recommendation
via graph-based contextual modeling and postfiltering. Int. J. Distrib. Sen. Netw.,
2015:16:16–16:16.
Yu, H.-F., Lo, H.-Y., Hsieh, H.-P., Lou, J.-K., McKenzie, T. G., Chou, J.-W., Chung,
P.-H., Ho, C.-H., Chang, C.-F., Wei, Y.-H., et al. (2010). Feature engineering and
classifier ensemble for kdd cup 2010.
Zoidi, O., Fotiadou, E., Nikolaidis, N., and Pitas, I. (2015). Graph-based label propagation in digital media: A review. ACM Comput. Surv., 47(3):48.
Zukerman, I. and Albrecht, D. W. (2001). Predictive statistical models for user modeling. User Modeling and User-Adapted Interaction, 11(1-2):5–18.
60
| 2cs.AI
|
A non-interleaving process calculus for multi-party
synchronisation
Paweł Sobociński
ECS, University of Southampton
We introduce the wire calculus. Its dynamic features are inspired by Milner’s CCS: a unary prefix operation, binary choice and a standard recursion construct. Instead of an interleaving parallel
composition operator there are operators for synchronisation along a common boundary (;) and noncommunicating parallel composition (⊗) . The (operational) semantics is a labelled transition system
obtained with SOS rules. Bisimilarity is a congruence with respect to the operators of the language.
Quotienting terms by bisimilarity results in a compact closed category.
Introduction
Process calculi such as CSP, SCCS, CCS and their descendants attract ongoing theoretical attention because of their claimed status as mathematical theories for the study of concurrent systems. Several concepts have arisen out of their study, to list just a few: various kinds of bisimulation (weak, asynchronous,
etc.), various reasoning techniques and even competing schools of thought (‘interleaving’ concurrency
vs ‘true’ concurrency).
The wire calculus does not have a CCS-like interleaving ‘k’ operator; it has operators ‘⊗’ and ‘;’
instead. It retains two of Milner’s other influential contributions: (i) bisimilarity as extensional equational
theory for observable behaviour, and (ii) syntax-directed SOS. The result is a process calculus which is
‘truly’ concurrent, has a formal graphical representation and on which ‘weak’ bisimilarity agrees with
‘strong’ (ordinary) bisimilarity.
The semantics of a (well-formed) syntactic expression in the calculus is a labelled transition system
(LTS) with a chosen state and with an associated sort that describes the structure of the labels in the LTS.
In the graphical representation this is a ‘black box’ with a left and a right boundary, where each boundary
is simply a number of ports to which wires can be connected. Components are connected to each other
by ‘wires’—whence the name of the calculus. The aforementioned sort of an LTS prescribes the number
of wires that can connect to each of the boundaries.
Notationally, transitions have upper and lower labels that correspond, respectively, to the left and the
right boundaries. The labels themselves are words over some alphabet Σ of signals and an additional
symbol ‘ι’ that denotes the absence of signal.
The operators of the calculus fall naturally into ‘coordination’ operators for connecting components
and ‘dynamic’ operators for specifying the behaviour of components. There are two coordination operators, ‘⊗’ and ‘;’. The dynamic operators consist of a pattern-matching unary prefix, a CSP-like choice
operator and a recursion construct. Using the latter, one can construct various wire constants, such as
different kinds of forks. These constants are related to Bruni, Lanese and Montanari’s stateless connectors [5] although there they are primitive entities; here they are expressible using the primitive operations
of the calculus.
The wire calculus draws its inspiration form many sources, the two most relevant ones come from
the model theory of processes: tile logic [6] and the Span(Graph) [7] algebra. Both frameworks are
F. Bonchi, D. Grohmann, P. Spoletini, and E. Tuosto:
ICE’09 Structured Interactions
EPTCS 12, 2009, pp. 87–98, doi:10.4204/EPTCS.12.6
c Paweł Sobociński
This work is licensed under the
Creative Commons Attribution License.
88
A non-interleaving process calculus for multi-party synchronisation
aimed at modeling the operational semantics of systems. Tile logic inspired the use of doubly-labeled
transitions, as well as the use of specific rule formats in establishing the congruence property for bisimilarity. Work on Span(Graph) informs the graphical presentation adopted for the basic operators of the
wire calculus and inspired the running example. Similarly, Abramsky’s interaction categories [1] are a
related paradigm: in the wire calculus, processes (modulo bisimilarity) are the arrows of a category and
sequential composition is interaction. There is also a tensor product operation, but the types of the wire
calculus have a much simpler structure.
One can argue, however, that the wire calculus is closer in spirit to Milner’s work on process algebra than the aforementioned frameworks. The calculus offers a syntax for system specification. The
semantics is given by a small number of straightforward SOS rules. No structural equivalence (or other
theory) is required on terms: the equational laws among them are actually derived via bisimilarity. The
fact that operations built up from the primitive operators of the wire calculus preserve bisimilarity means
that there is also a shared perspective with work on SOS formats [2] as well as recent work on SOS-based
coordination languages [3].
An important technical feature is the reflexivity of labelled transition systems—any process can always ‘do nothing’ (and by doing so synchronise with ι’s on its boundaries). This idea is also important
in the Span(Graph) algebra; here it ensures that two unconnected components are not globally synchronised. The wire calculus also features ι-transitivity which ensures that unconnected components
are not assumed to run at the same speeds and that internal reduction is unobservable. Reflexivity and ιtransitivity together ensure that weak and strong bisimilarities coincide—they yield the same equivalence
relation on processes.
Structure of the paper. The first two sections are an exposition of the syntax and semantics of the
wire calculus. In §3 we model a simple synchronous circuit. The next two sections are devoted to the
development of the calculus’ theory and §6 introduces a directed variant.
1
Preliminaries
A wire sort Σ is a (possibly infinite) set of signals. Let ι be an element regarded as “absence of signal”,
with ι ∈
/ Σ. In order to reduce the burden of bureaucratic overhead, in this paper we will consider calculi
with one sort. The generalisation to several sorts is a straightforward exercise.
Before introducing the syntax of the calculus, let us describe those labelled transition systems (LTSs)
that will serve as the semantic domain of expressions: (k, l)-components. In the following #a means the
length of a word a. We shall use ι for words that consist solely of ι’s.
def
Definition 1 (Components). Let L = Σ + {ι} and k, l ≥ 0. A (k, l)-transition is a labelled transition
a
of the form −
→
where a, b ∈ L∗ , #a = k and #b = l. A (k, l)-component C is a pointed, reflexive and
b
ι-transitive LTS (v0 ,V, T ) of (k, l)-transitions. The meanings of the adjectives are:
• Pointed: there is a chosen state v0 ∈ V ;
ι
• Reflexive: for all v ∈ V there exists a transition v −
→
v;
ι
• ι-Transitive: if v −
→
v1 , v1 −
→
v2 and v2 −
→
v0 then also v −
→
v0 .
ι
ι
b
b
To the right is a graphical rendering of C ; a box with k wires on the left and l k { v0
}l
wires on the right:
When the underlying LTS is clear from the context we often identify a component with its chosen
state, writing v0 : (k, l) to indicate that v0 is (the chosen state of) a (k, l)-component.
ι
a
ι
a
..
.
..
.
Paweł Sobociński
89
Definition 2 (Bisimilarity). For two (k, l)-components C , C 0 , a simulation from C to C 0 is a relation
a
a
S that contains (v0 , v00 ) and satisfies the following: if (v, w) ∈ S and v −
→
v0 then ∃w0 s.t. w −
→
w0 and
b
b
(v0 , w0 ) ∈ S. A bisimulation is a relation S such that itself and its inverse S−1 are both simulations. We
say that C , C 0 are bisimilar and write C ∼ C 0 when there is some bisimulation relating C and C 0 .
Remark 3. One pleasant feature of reflexive and ι-transitive transition systems is that weak bisimilarity
agrees with ordinary (“strong”) bisimilarity, where actions labelled solely with ι’s are taken as silent.
Indeed, ι-transitivity implies that any weak (bi)simulation in the sense of Milner [9] is also an ordinary
(bi)simulation.
2
Wire calculus: syntax and semantics
The syntax of the wire calculus is given in (1) and (2) below:
P ::= Y | P ; P | P ⊗ P |
M
MP
| P + P | µY : τ. P
M ::= ε | x | λ x | ι | σ ∈ Σ | MM
(1)
(2)
All well-formed wire calculus terms have an associated sort (k, l) where k, l ≥ 0. We let τ range over
sorts. The semantics of τ-sorted expression is a τ-component. There is a sort inference system for the
syntax presented in Fig. 1. In each rule, Γ denotes the sorting context: a finite set of (i) process variables
with an assigned sort, and (ii) signal variables. We will consider only those terms t that have a sort
derived from the empty sorting context: i.e. only closed terms.
The syntactic categories in (1) are, in order: process variables, composition, tensor, prefix, choice
and recursion. The recursion operator binds a process variable that is syntactically labelled with a sort.
When we speak of terms we mean abstract syntax where identities of bound variables are ignored. There
are no additional structural congruence rules. In order to reduce the number of parentheses when writing
wire calculus expressions, we shall assume that ‘⊗’ binds tighter than ‘;’.
The prefix operation is specific to the wire calculus and merits an extended explanation: roughly, it
is similar to input prefix in value-passing CCS, but has additional pattern-matching features. A general
u
prefix is of the form v P where u and v are strings generated by (2): ε is the empty string, x is a free
occurrence of a signal variable from a denumerable set x = {x, y, z, . . . } of signal variables (that are
disjoint from process variables), λ x is the binding of a signal variable, ι is ‘no signal’ and σ ∈ Σ is a
signal constant. Note that for a single variable x, ‘λ x’ can appear several times within a single prefix.
Prefix is a binding operation and binds possibly several signal variables that appear freely in P; a variable
u
x is bound in P precisely when λ x occurs in u or v. Let bd( v ) denote the set of variables that are bound
u
u
by the prefix, i.e. x ∈ bd( v ) ⇔ λ x appears in uv. Conversely, let f r( v ) denote the set of variables that
appear freely in the prefix.
When writing wire calculus expressions we shall often omit the sort of the process variable when
def
writing recursive definitions. For each sort τ there is a term 0τ = µY : τ.Y . We shall usually omit the
subscript; as will become clear below, 0 has no non-trivial behaviour.
The operations ‘;’ and ‘⊗’ have graphical representations that are convenient for modelling physical
systems; this is illustrated below.
..
.
P
..
.
R
..
.
P
..
.
..
.
Q
..
.
..
.
90
A non-interleaving process calculus for multi-party synchronisation
Γ ` P : (k, n)
Γ ` R : (n, l)
Γ ` P : (k, l)
Γ ` P;R : (k, l)
Γ, X:τ ` X : τ
Γ,Y :τ 0 ` P : τ
Γ ` P⊗Q : (k+m, l+n)
#u=k, #v=l, f r( uv )∩bd( uv )=∅, f r( uv )⊆Γ
Γ`
Γ ` Q : (m, n)
Γ, bd( uv ) ` P : (k, l)
Γ ` µY :τ 0 . P : τ
Γ ` P:τ
u
v P : (k, l)
Γ ` Q:τ
Γ ` P+Q : τ
Figure 1: Sorting rules.
P−
→P
P−
→
R
ι
(R EFL )
a
R−
→
Q
b
P−
→
Q
b
a
P−
→
Q
c
c
R−
→
S
b
a
P;R −
→
Q;S
b
u
vP
−−
→ P|σ
v|σ
(P REF )
(C UT )
P[µY. P/Y ] −
→
Q
b
a
µY. P −
→
Q
b
P−
→
R
b
(ι L)
P−
→
Q
b
a
ac
(T EN )
a
ι
P−
→
Q R−
→
S
ι
ι
ι
ι
(ι R)
c
R−
→
S
d
P⊗R −
→ Q⊗S
bd
ι
(R EC )
ι
R−
→
Q
ι
P−
→
Q
b
a
a
u|σ
a
a
ι
ι
ι
P+R −
→ Q+S
P−
→
Q
b
(+ι )
(ab6=ι)
a
P+R −
→
Q
b
(+L)
Figure 2: Semantics of the wire calculus. Symmetric rule (+R) omitted.
The semantics of a term t : (k, l) is the (k, l)-component with LTS defined using SOS rules: (R EFL ),
and (R EC ) of Fig. 2. The rules (R EFL ), (ι L) and (ι R) guarantee
that the LTS satisfies reflexivity and ι-transitivity of Definition 1. Note that while any calculus with
global rules akin to (R EFL ), (ι L), (ι R) for ‘silent’ actions is automatically ι-transitive and reflexive with
the aforementioned result on bisimilarity (Remark 3), there are deeper reasons for the presence of these
rules: see Remark 5.
The rules (C UT ) and (T EN ) justify the intuition that ‘;’ is synchronisation along a common boundary
u
u
and ‘⊗’ is parallel composition sans synchronisation. For a prefix v P a substitution is a map σ : bd( v ) →
u
Σ + {ι}. For a term t with possible free occurrences of signal variables in bd( v ) we write t|σ for the term
u
resulting from the substitution of σ (x) for x for each x ∈ bd( v ). Assuming that u, v have no free signal
variables, we write u|σ , v|σ for the strings over Σ + {ι} that result from replacing each occurrence of
‘λ x’ with σ (x). The SOS rule that governs the behaviour of prefix is (P REF ), where σ is any substitution.
The choice operator + has a similar semantics to that of the CSP choice operator ‘@’, see for instance [4]:
a choice is made only on the occurrence of an observable action. Rules (+ι ), (+L) and (+R) make this
statement precise. The recursion construct and its SOS rule (R EC ) are standard. We assume that the
substitution within the (R EC ) rule is non-capturing.
As first examples of terms we introduce two wire constants1 , their graphical representations as well
as an SOS characterisation of their semantics.
(ι L), (ι R), (C UT ), (T EN ), (P REF ), (+ι ), (+L), (+R)
def
λx
I = µY. λ x Y : (1, 1)
def
λ xλ y
X = µY. λ yλ x Y : (2, 2)
a
I−
→
I
a
ab
(I D )
X −ba
→X
(T W )
Remark 4. Wire constants are an important concept of the wire calculus. Because they have a single
state, any expression built up from constants using ‘;’ and ‘⊗’ has a single state. Bisimilar wirings can
1 Those
terms whose LTS has a single state.
Paweł Sobociński
91
thus be substituted in terms not only without altering externally observable behaviour, but also without
combinatorially affecting the (intensional) internal state.
3
Global synchrony: flip-flops
As a first application of the wire calculus, we shall model the following circuit2 :
F0
F1
(3)
F0
where the signals that can be sent along the wire are {0, 1} (n.b. 0 is not the absence of signal, that is
represented by ι). The boxes labelled with F0 and F1 are toggle switches with one bit of memory, the
state of which is manifested by the subscript. They are simple abstraction of a flip-flop. The switches
synchronise to the right only on a signal that corresponds to their current state and change state according
to the signal on the left. The expected behaviour is a synchronisation of the whole system — here
a tertiary synchronisation. In a single “clock tick” the middle component will change state to 0, the
rightmost component to 1 and the leftmost component will remain at 0. F0 and F1 are clearly symmetric
and their behaviour is characterised by the SOS rules below (where i ∈ {0, 1}) .
i
Fi →
−i Fi
(iSETi)
1−i
Fi −−
→ F1−i
i
(iSET1 − i)
ι
Fi −
→
Fi
ι
(iR EFL )
In the wire calculus F0 and F1 can be defined by the terms:
def
0
1
1
def
0
1
0
0
1
F0 = µY. 0 Y + 0 µZ. ( 1 Z + 1 Y ) and F1 = µZ. 1 Z + 1 µY. ( 0 Y + 0 Z).
In order to give an expression for the whole circuit, we need two additional wire constants d and e.
They are defined below, together with a graphical representation and an SOS characterisation. Their
mathematical significance will be explained in Section 5.
def
d : (0, 2) = µY. λ xλ x Y
def
e : (2, 0) = µY.
λ xλ x
Y
d→
−
d
aa
aa
e −→ e
(d)
(e)
Returning to the example, the wire calculus expression corresponding to (3) can be written down in the
wire calculus by scanning the picture from left to right.
def
A = d ; (I ⊗ (F0 ; F1 ; F0 )) ; e.
What are its dynamics? Clearly A has the sort (0, 0). It is immediate that any two terms of this sort are
extensionally equal (bisimilar) as there is no externally observable behaviour. This should be intuitively
obvious because A and other terms of sort (0, 0) are closed systems; they have no boundary on which
an observer can interact. We can, however, examine the intensional internal state of the system and the
possible internal state transitions. The semantics is given structurally in a compositional way – so any
2 This
example was proposed to the author by John Colley.
92
A non-interleaving process calculus for multi-party synchronisation
non-trivial behaviour of A is the result of non-trivial behaviour of its components. Using
possible behaviour of F0 ; F1 is of the form (i) below.
x
F0 −
→
X
z
z
x
F1 −
→
Y
y
F0 ;F1 −
→
X;Y
y
0
x
F0 −
→
X
0
(C UT )
x
F1 −
→
F0
1
→
X;F0
F0 ;F1 −
1
(i)
x
→
X;F0
F0 ;F1 −
1
(C UT )
x
(C UT )
the only
1
→
F1
F0 −
0
→
X;F0 ;F1
F0 ;F1 ;F0 −
0
(ii)
(C UT )
(iii)
The dynamics of F0 and F1 imply that z = 0 and y = 1, and the latter also implies that Y = F0 , hence the
occurrence of (C UT ) must be of the form (ii) above. Composing with F0 on the right yields (iii).
wx
I⊗(F0 ;F1 ;F0 ) −w0
→ I⊗(X;F0 ;F1 )
(⊗)
(iv)
Next, tensoring with I yields (iv) above. In order for (C UT ) to apply after post-composing with e, w must
be equal to 0, yielding (iv) below.
0x
(I⊗(F0 ;F1 ;F0 ));e −→ (I⊗(X;F0 ;F1 ));e
(C UT )
d;(I⊗(F0 ;F1 ;F0 ));e →
− d;(I⊗(F0 ;F0 ;F1 ));e
(iv)
(C UT )
(v)
The final application of (C UT ) forces x = 0 and hence X = F0 , resulting in (v). It is easy to repeat this
argument to show that the entire system has three states — intuitively ‘1 goes around the ring’. It is
equally easy to expand this example to synchronisations that involve arbitrarily many components. Such
behaviour is not easily handled using existing process calculi.
Remark 5. Consider the wire calculus expression represented by (†).
F0
F1
F0
F0
(†)
F1
The two components are unconnected. As a consequence of (R EFL ) and (T EN ), a “step” of the entire system
is either a (i) step of the upper component, (ii) step of the lower component or (iii) a ‘truly-concurrent’
step of the two components. The presence of the rules (ι L) and (ι R) then ensures that behaviour is not
scheduled by a single global clock — that would be unreasonable for many important scenarios. Indeed,
in the wire calculus synchronisations occur only between explicitly connected components.
4
Properties
In this section we state and prove the main properties of the wire calculus: bisimilarity is a congruence
and terms up-to-bisimilarity are the arrows of a symmetric monoidal category W . In fact, this category
has further structure that will be explained in section 5.
ak
ak
a1
Let P ( −b→
) Q denote trace P −b→
P1 · · · Pk−1 −b→
Q, for some P1 , . . . , Pk−1 .
1
k
k
Lemma 6. Let P : (k, l) and Q : (l, m), then:
(i) If P ; Q −
→
R then R = P0 ; Q0 and there exist traces
b
a
P(−
→
) Pk −
→
Pl0 ( −
→
) P0 ,
c
el
d
ι
k
a
ι
d
e
k
l
Q ( −→
) Qk −
→
Q0l ( −→
) Q0 ;
ι
ι
b
c
Paweł Sobociński
93
(ii) If P ⊗ Q −
→
R then R = P0 ⊗ Q0 and there exist transitions:
b
a
a
1
P −b→
P0 ,
1
a
2
Q −b→
Q0 with a = a1 a2 and b = b1 b2 .
2
a
Proof. Induction on the derivation of P ; Q −
→
R. The base cases are: a single application of (C UT )
b
a
c
0
0
0
0
whence P −
→
P and Q −
→
Q and R = P ; Q , and a single application of (R EFL ) whence R = P ; Q but by
c
b
ι
ι
reflexivity we have also that P −
→
P and Q −
→
Q. The first inductive step is an application of (ι L) where
ι
ι
ι
a
P;Q −
→
R
and
R
−
→
S.
Applying
the
inductive
hypothesis to the first transition we get R = P00 ; Q00
ι
b
a
and suitable traces. Now we apply the inductive hypothesis again to R00 ; Q00 −
→
S and obtain S = P0 ; Q0
b
together with suitable traces. The required traces are obtained by composition in the obvious way. The
second inductive step is an application of (ι R) and is symmetric. Part (ii) involves a similar, if easier,
induction.
Theorem 7 (Bisimilarity is a congruence). Let P, Q : (k, l), R : (m, n), S : (l, l 0 ) and T : (k0 , k) and suppose
that P ∼ Q. Then:
(i) P ; S ∼ Q ; S and T ; P ∼ T ; Q;
(ii) P ⊗ R ∼ Q ⊗ R and R ⊗ P ∼ R ⊗ Q;
(iii)
u
vP
u
∼ v Q;
(iv) P + R ∼ Q + R and R + P ∼ R + Q.
Proof. Part (i) follows easily from the conclusion of Lemma 6, (i); in the first case one shows that
def
a
A = { (P ; R, Q ; R) | P, Q : (k, l), R : (l, m), P ∼ Q } is a bisimulation. Indeed, if P ; R −
→
P0 ; R0 then
b
0
0
0
there exists an appropriate trace from P to P and from R to R . As P ∼ Q, there exists Q with P0 ∼ Q0
a
and an equal trace from Q to Q0 . We use the traces from Q and R to obtain a transition Q ; R −
→
Q0 ; R0
b
0
0
0
0
via repeated applications of (C UT ) and (ι L,R). Clearly (P ; R , Q ; R ) ∈ A. Similarly, (ii) follows from
Lemma 6, (ii); (iii) and (iv) are straightforward.
λ x ...λ x
def
def
λ x ...λ x λ y ...λ y
Let Ik = µY. λ x11 ...λ xkk Y and Xk, l = µY. λ y11 ...λ ykl λ x11...λ xkl Y . Xk, l is drawn as
k{
..
.
..
.
}l
l{
..
.
..
.
}k
.
Lemma 8 (Categorical axioms). In the statements below we implicitly universally quantify over all
k, l, m, n, u, v ∈ N, P : (k, l), Q : (l, m), R : (m, n), S : (n, u) and T : (u, v).
(i) (P ; Q) ; R ∼ P ; (Q ; R);
(ii) P ; Il ∼ P ∼ Ik ; P;
(iii) (P ⊗ R) ⊗ T ∼ P ⊗ (R ⊗ T );
(iv) (P ⊗ S) ; (Q ⊗ T ) ∼ (P ; Q) ⊗ (S ; T );
(v) (P ⊗ R) ; Xl,n ∼ Xk,m ; (R ⊗ P);
(vi) Xk,l ; Xl,k ∼ Ik+l .
..
.
P
..
.
..
.
..
.
R
..
.
..
.
∼
..
.
..
.
R
..
.
..
.
..
.
P
..
.
Proof. (i) Here we use Lemma 6, (ii) to decompose a transition from (P ; Q) ; R into traces of P, Q and
a
R and then, using reflexivity, compose into a transition of P ; (Q ; R). Indeed, suppose that (P ; Q) ; R −
→
b
(P0 ; Q0 ) ; R0 . Then:
P ; Q (−
→
) Pk ; Qk
d
ι
i
a
−
→
c
d
i
Pl0 ; Q0l ( −
→
) P0 ; Q0 and R ( −→
) Rk
ei
ι
ι
c
−
→
b
e
i
R0l ( −
→
) R0 .
ι
94
A non-interleaving process calculus for multi-party synchronisation
Decomposing the first trace into traces of P and Q yields
P(−
→
) P1k1 −
→
) P1 · · · Pk ( −
→
)−
→
→
P1l0 1 ( −
(−
→
) Pl0 · · · P0
e1i
c1
d
c0
d0
e0
ι
ι
ι
a
ι
1i
i
ι
i
c
d1i
ι
d0
e
c0
e0
1
1i
i
i
Q01l1 ( −
→
) Q1 · · · Qk ( −
and Q ( −→ ) Q1k1 −d→
→
)−
→
(−
→
) Q0l · · · Q0 .
ι
ι
c
ι
1
Using reflexivity and (C UT ) we obtain
d
c
d0
e
e0
c0
1i
1
1i
i
i
Q ; R ( −−→
) Q1k1 ; R −→
Q01l1 ; R1 ( −→
) Q1 ; R1 · · · Qk ; Rk ( −→
)−
→
(−
→
) Q0l ; R0l · · · Q0 ; R0 .
ι
ι
ι
ι
ι
b
Now by repeated applications of (C UT ) followed by (ι L,R) we obtain the required transition P ; (Q ; R) −
→
b
0
0
0
P ; (Q ; R ). Similarly, starting with a transition from P ; (Q ; R) one reconstructs a matching transition
from (P ; Q) ; R.
aa0
Parts (ii) and (iii) are straightforward. For (iv), a transition (P ⊗ R) ; (Q ⊗ S) −−
→ (P0 ⊗ R0 ) ; (Q0 ⊗
bb0
0
S ) decomposes first into traces:
a
aa0
P ⊗ R ( d−
→
) Pk ⊗ Rk −−
→ Pl0 ⊗ R0l ( −
→
) P0 ⊗ R0
cc0
d0
e e0
ι
ι
i i
i i
di d0
ei e0
cc0
i
i
→ Q0l ⊗ Sl0 ( −−ι→
Q ⊗ S ( −−ι→
) Qk ⊗ Sk −−
) Q0 ⊗ S0
bb0
and then into individual traces of P, R, Q and S:
P(−
→
) Pk −
→
Pl0 ( −
→
) P0
c
ei
d
a
ι
ι
i
d
ι
i
i
d0
e
c0
e0
i
i
→
Sl0 ( −
→
) S0
S ( −→
) Sk −
ι
ι
b0
i
i
→
Q0l ( −
→
) Q0
Q ( −→
) Qk −
ι
ι
b
c
a0
R(−
→
) Rk −→
R0l ( −
→
) R0
c0
d0
e0
ι
a0
and hence transitions P ; Q −
→
P0 ; Q0 and R ; S −→
R0 ; S0 , which combine via a single application of
b
b0
a
aa0
→ (P0 ; Q0 ) ⊗ (R0 ; S0 ). The converse is similar. Parts (v) and (vi) are
to give (P ; Q) ⊗ (R ; S) −−
bb0
trivial.
(⊗)
As a consequence of (i) and (ii) there is a category W that has the natural numbers as objects and
terms of sort (k, l) quotiented by bisimilarity as arrows from k to l; for P : (k, l) let [P] to denote P’s
equivalence class wrt bisimilarity – then [P] : m → n is the arrow of W . The identity morphism on
m is [Im ]. Composition [P] : k → l with [Q] : l → m is [P ; Q] : k → m.3 Parts (iii) and (iv) imply
def
that W is monoidal with a strictly associative tensor. Indeed, on objects let m ⊗ n = m + n and on
def
arrows [P] ⊗ [Q] = [P ⊗ Q].4 The identity for ⊗ is clearly 0. Subsequently (v) and (vi) imply that
W is symmetric monoidal. Equations (i)–(iv) also justify the use of the graphical syntax: roughly,
rearrangements of the representation in space result in syntactically different but bisimilar systems.
5
Closed structure
Here we elucidate the closed structure of W : the wire constants d and e play an important role.
Lemma 9. d ⊗ I ; I ⊗ e ∼ I ∼ I ⊗ d ; e ⊗ I
∼
∼
.
3 Well-defined
4 Well-defined
due to Theorem 7.
due to Theorem 7.
Paweł Sobociński
95
Given k ∈ N, define recursively
d1
def
= d and dn+1
def
= d ; I ⊗ dn ⊗ I and dually e1
def
= e and en+1
def
= In ⊗ e ⊗ In ; en .
def
Let d0 , e0 = I0 = 0.
Lemma 10. For all n ∈ N, dn ⊗ In ; In ⊗ en ∼ In ∼ In ⊗ dn ; en ⊗ In .
Proof. For n = 0 the result is trivial. For n = 1, it is the conclusion of Lemma 9. For n > 1 the result
follows by straightforward inductions:
dn+1 ⊗ In+1 ; In+1 ⊗ en+1
= (d ; I ⊗ dn ⊗ I) ⊗ In+1 ; In+1 ⊗ (In ⊗ e ⊗ In ; en )
∼ d ⊗ In+1 ; I ⊗ dn ⊗ In+2 ; I2n+1 ⊗ e ⊗ In ; In+1 ⊗ en
∼ d ⊗ In+1 ; I ⊗ e ⊗ In ; I ⊗ dn ⊗ In ; In+1 ⊗ en
∼ (d ⊗ I ; I ⊗ e) ⊗ In ; I ⊗ (dn ⊗ In ; In ⊗ en )
∼ In+1
In+1 ⊗ dn+1 ; en+1 ⊗ In+1
= In+1 ⊗ (d ; I ⊗ dn ⊗ I) ; (In ⊗ e ⊗ In ; en ) ⊗ In+1
∼ In+1 ⊗ d ; In+2 ⊗ dn ⊗ I ; In ⊗ e ⊗ I2n+1 ; en ⊗ In+1
∼ In+1 ⊗ d ; In ⊗ e ⊗ I ; In ⊗ dn ⊗ I ; en ⊗ In+1
∼ In ⊗ (I ⊗ d ; e ⊗ I) ; (In ⊗ dn ; en ⊗ In ) ⊗ I
∼ In+1
With Lemma 10 we have shown that W is a compact-closed category [8]. Indeed, for all n ≥ 0 let
n; we have shown the existence of arrows
def
n? =
[dn ] : 0 → n ⊗ n?
[en ] : n? ⊗ n → 0
(4)
that satisfy [dn ] ⊗ idn ; In ⊗ [en ] = In = idn ⊗ [dn ] ; [en ] ⊗ In .5
While the following is standard category theory, it may be useful for the casual reader to see how this
data implies a closed structure. For W to be closed wrt ⊗ we need, for any l, m ∈ W an object l ( m
and a term
evl,m : ((l ( m) ⊗ l, m)
that together satisfy the following universal property: for any term P : (k ⊗ l, m) there exists a unique
(up to bisimilarity) term Cur(P) : (k, l ( m) such that the diagram below commutes:
O
[Cur(P)⊗Il ]
k⊗l
[evl,m ]
/
t9 m
tt
t
t
tt
tt [P]
t
t
(l ( m) ⊗ l
5 The (−)? operation on objects of W may seem redundant at this point but it plays a role when considering the more
elaborate sorts of directed wires in §6.
96
A non-interleaving process calculus for multi-party synchronisation
def
def
def
Owing to the existence of dn and en we have l ( m = m+l, evl,m = Im ⊗ el and Cur(P) = Ik ⊗ dl ; P ⊗ Il .
Indeed, it is easy to verify universality: suppose that there exists Q : (k, m+l) such that Q ⊗ Il ; evl,m ∼ P.
Then
Cur(P) = Ik ⊗ dl ; P ⊗ Il
∼ Ik ⊗ dl ; (Q ⊗ Il ; evl,m ) ⊗ Il
∼ Ik ; Q ; Im ⊗ (Il ⊗ dl ; el ⊗ Il )
∼Q
Starting from the situation described in (4) and using completely general reasoning, one can define a
contravariant functor (−)? : W op → W by A 7→ A? on objects and mapping f : A → B to the composite
B? ⊗ f ⊗A?
B? ⊗d
e ⊗A?
A
B
B? −−−−→
B? ⊗ A ⊗ A? −−−−−→ B? ⊗ B ⊗ A? −−
−→ A? .
Indeed, the fact that (−)? preserves identities is due to one of the triangle equations and the fact that
it preserves composition is implied by the other. We can describe (−)? on W directly as the up-tobisimulation correspondent of a structurally recursively defined syntactic transformation that, intuitively,
rotates a wire calculus term by 180 degrees. First, define the endofunction (−)? on strings generated
by (2) to be simple string reversal (letters are either free variables, bound variables, signals or ‘ι’).
Definition 11. Define an endofunction (−)? on wire calculus terms by structural recursion on syntax as
follows:
def
def
(R ; S)? = S? ; R?
u
def v? ?
u? R
( v R)? =
(R ⊗ S)? = S? ⊗ R?
def
(R + S)? = R? + S?
def
(µY. R)? = µY. R?
Notice that d? = e and, by a simple structural induction, P?? = P. This operation is compatible with
d and e in the following sense.
Lemma 12. Suppose that P : (k, l). Then dk ; P ⊗ Ik ∼ dl ; Il ⊗ P? . Dually, P ⊗ Il ; el ∼ Ik ⊗ P? ; ek .
The conclusion of the following lemma implies that for any k, l ≥ 0 one can define a function (−)? :
def
W (k, l) → W (l, k) by setting [P]? = [P? ].
b?
Lemma 13. If P : (k, l) then P? : (l, k). Moreover P −
→
R iff P? −a→
R? . Consequently P ∼ Q iff
?
b
P? ∼ Q? .
a
Proof. Structural induction on P.
Moreover, by definition we have:
([P]; [Q])? = [P ; Q]? = [(P ; Q)? ] = [Q? ; P? ] = [Q? ] ; [P? ] = [Q]? ; [P]?
and [Ik ]? = [Ik ], thus confirming the functoriality of (−)? : W op → W .
Paweł Sobociński
6
97
Directed wires
In the examples of §3 there is a suggestive yet informal direction of signals: the components set their state
according to the signal coming in on the left and output a signal that corresponds to their current state
on the right. It is useful to make this formal so that systems of components are not wired in unintended
ways. Here we take this leap, which turns out not to be very difficult. Intuitively, decorating wires with
directions can be thought as a type system in the sense that the underlying semantics and properties are
unchanged, the main difference being that certain syntactic phrases (ways of wiring components) are no
longer allowed.
def
def
Definition 14 (Directed components). Let D = {L, R} and subsequently Ld = (Σ + {ι}) × D. Abbrevi−
−
ate (a, L) ∈ Ld by ←
a and (a, R) by →
a . Let π : Ld → D be the obvious projection.
a
∗
Let k, l ∈ D . A (k, l)-transition is a labelled transition of the form −
→
where a, b ∈ Ld∗ , π ∗ (a) = k
b
∗
and π (b) = l. A (k, l)-component C is a pointed, reflexive and ι-transitive LTS (v0 ,V, T ) of (k, l)transitions.
The syntax of the directed version of the wire calculus is obtained by replacing (2) with (5) below:
signals are now either inputs (?) or outputs (!).
M ::= ε | D | MM
(5)
D ::= A? | A!
A ::= x | λ x | ι | σ ∈ Σ
def
def
def
def
Define a map (−) : D ∗ → D ∗ in the obvious way by letting L = R and R = L. Now define a function
d from terms generated by (5) to D ∗ recursively by letting d(ε) = ε, d(uv) = d(u)d(v), d(x?) = L and
d(x!) = R. The terms are sorted with the rules in Fig. 1, after replacing the rule for prefix with the
following:
du=k, dv=l, f r( uv )∩bd( uv )=∅, f r( uv )⊆Γ
Γ, bd( uv ) ` P : (k, l)
Γ ` uv P : (k, l)
The semantics is defined with the rules of Fig. 2 where labels are strings over Ld and (P REF ) is replaced
with ( D P REF ) below. In the latter rule, given a substitution σ , let eσ be the function that takes a prefix
component generated by (5) with no occurrences of free signal variables to Ld∗ , defined by recursion in
the obvious way from its definition on atoms: eσ (λ x?) = (σ x, L), eσ (λ x!) = (σ x, R), eσ (a?) = (a, L),
eσ (a!) = (a, R). Abusing notation, let (−) be the endofunction of Ld∗ that switches directions of components. Then:
u
vP
e u
σ
−−
→ P|σ
eσ v
( D P REF )
Below are some examples of wire constants in the directed variant of the calculus:
def
λ x!
IL = µY. λ x? Y : (L, L)
←
−
a
IL −←
→
− IL
(IL )
a
def
λ x?
IR = µY. λ x! Y : (R, R)
→
−
a
IR −→
→ IR
−
(IR )
a
def
dL = µY. λ x?λ x! Y : (ε, LR)
dL −←
−
→ dL
−→
−
a a
(dL )
98
A non-interleaving process calculus for multi-party synchronisation
def
eL = µY.
λ x?λ x!
Y : (RL, ε)
→
−
−
a←
a
eL −
−−→ eL
(eL )
At this point the Fx components of §3 can modelled in the directed wire calculus. The results of §4 and
§5 hold mutatis mutandis with the same proofs – this is not surprising as only the semantics of prefix and
the structure of sorts is affected.
7
Conclusion and future work
As future work, one should take advantage of the insights of Selinger [10]: it appears that one can
express Selinger’s queues and buffers within the wire calculus and thus model systems with various
kinds of asynchronous communication. Selinger’s sequential composition, however, has an interleaving
semantics.
From a theoretical point of view it should be interesting to test the expressivity of the wire calculus
with respect to well-known interleaving calculi such as CCS and non-syntactic formalisms such as Petri
nets.
Acknowledgement. The author thanks the anonymous referees and the participants in the ICE forum
for their suggestions, many of which have improved the paper. Early discussions with R.F.C. Walters
and Julian Rathke were fundamental in stabilising the basic technical contribution. Special thanks go to
Fabio Gadducci and Andrea Corradini.
References
[1] Samson Abramsky, Simon Gay & Rajogopal Nagarajan (1995): Interaction Categories and the Foundations
of Typed Concurrent Programming. In: Proceedings of the 1994 Marktoberdorf Summer School, NATO ASI
Series F. Springer.
[2] Luca Aceto, Willem Jan Fokkink & Chris Verhoef (1999): Structural Operational Semantics. In: Bergstra,
Ponse & Smolka, editors: Handbook of Process Algebra. Elsevier.
[3] Simon Bliudze & Joseph Sifakis (2008): A Notion of Glue Expressiveness for Component-Based Systems.
In: Proceedings of the 19th International Conference on Concurrency Theory (CONCUR’08), LNCS 5201.
Springer, pp. 508–522.
[4] Stephen D. Brookes, A. William Roscoe & David J. Walker (1988): An Operational Semantics for CSP.
Manuscript.
[5] Roberto Bruni, Ivan Lanese & Ugo Montanari (2005): Complete Axioms for Stateless Connectors. In: Algebra and Coalgebra in Computer Science, (CALCO ’05), LNCS 3629. pp. 98–113.
[6] Fabio Gadducci & Ugo Montanari (2000): The tile model. In: Proof, Language and Interaction: Essays in
Honour of Robin Milner. MIT Press, pp. 133–166.
[7] Piergiulio Katis, Nicoletta Sabadini & Robert Frank Carslaw Walters (1997): Span(Graph): an algebra of
transition systems. In: Proceedings of 6th International Conference on Algebraic Methodology and Software
Technology (AMAST ’97), LNCS 1349. Springer, pp. 322–336.
[8] G. M. Kelly & M. L. Laplaza (1980): Coherence for compact closed categories. Journal of Pure and Applied
Algebra 19, pp. 193–213.
[9] Robin Milner (1989): A calculus of communicating systems. Prentice Hall.
[10] Peter Selinger (1997): First-order axioms for asynchrony. In: Proceedings of 8th International Conference
on Concurrency Theory, CONCUR ’97, LNCS 1243. Springer, pp. 376–390.
| 6cs.PL
|
An Early Warning System for Bankruptcy Prediction:
lessons from the Venezuelan Bank Crisis
Financial Markets. Investments and Risk Management
Loren Trigo
Professor
Center for Production and Technological Innovation
Instituto de Estudios Superiores de Administración
Calles IESA, San Bernardino, Caracas, Venezuela
58 212 555 4498
58 212 555 4494 fax
[email protected]
Sabatino Costanzo
Professor
Center of Finance
Instituto de Estudios Superiores de Administración
Calles IESA, San Bernardino, Caracas, Venezuela
58 212 555 4498
58 212 555 4494 fax
[email protected]
An Early Warning System for Bankruptcy Prediction: lessons from the
Venezuelan Bank Crisis
Abstract
During the years 1993-94 Venezuela experienced a severe banking crisis which ended up with a total of 18 commercial banks
intervened by the government. Here we develop an early warning model for detecting credit related bankruptcy through
discriminant functions developed on financial and macroeconomic data predating the 1993-94 crisis. A robustness test with data
from 1996-2004 performed on these functions, shows high precision in Type 1 & 2 errors. The model calibrated on pre-crisis data
is even able to detect abnormal financial tension in the late Banco Capital by June 2000, many months before it was intervened and
liquidated.
Introduction
About banking crises
There are many definitions of what constitutes a banking crisis. Sundararajan and Baliño (1991) mention some of the definitions in
the relevant literature. Some examples of these definitions are the following:
•
“A money demand that cannot be satisfied by all participants in the short run” Schwartz (1985); Mirón (1986); Wolfon
(1986)
•
“Liquidation of loans that were granted during an economic boom” Veblen (1904); Mitchel (1941)
•
“Circumstance in which the suppliers of lending funds that in usual conditions are able to grant loans without any
difficulties find themselves unable to grant these loans in any form or term” Guttentag and Herring (1983),
Manikow(1986)
In addition, Morón and Aguero (2003) refer to the following concept in their work:
•
“A banking crisis is a circumstance where a bank or a group of banks becomes insolvent, with the consequence that the
value of their assets falls below the value of their obligations” Bell and pain (2000)
Main microeconomic causes of banking crises
Disproportionate increase of the size of the credit portfolio
Most banking crises follow an economic boom (García 1998), often accompanied by an increase in the size of the credit portfolios
of banks. In most of these cases, according to Barton, Newell and Wilson (2002), these new credits are often bad credits that can
ignite a crisis.
High number of expired loans
Given a booming credit portfolio plagued with problematic loans, the portfolio of expired loans increases in two ways: there is an
increase in the number of overdue loans and of loans under litigation. This of course affects the balance sheet; nevertheless, very
often this situation goes unnoticed because the banks dress up their financial statements to hide their expired loan portfolio
problems.
High number of highly correlated loans
Two loans are correlated if their collaterals are correlated. With respect to stock market bubbles, according to Barton, Newell and
Wilson (2002), several years of continued growth in the price of stocks can indicate the existence of a stock market bubble that
could burst. This is dangerous because 1. Many loans are backed up by shares of stock and a stock market crash will cause a
reduction in the collateral of these loans and 2. Many banks invest in stocks as well and stand to lose that investment if markets
become deeply bearish.
Main Macroeconomic Causes of Banking Crises
Economic instability
According to García (2000), one of the main causes of banking crises is an economic growth that is relatively unstable or with very
pronounced cycles covering short time periods. These cycles are not infrequent in oil exporting countries with oil price fluctuations
that make their economies highly volatile. One of the problems derived from these cycles is the abnormal growth of lending
portfolios during intervals of economic expansion, with good loans being granted at the beginning of the good times and bad loans
thereafter, resulting an a deteriorated lending portfolio whenever the good times end up being shorter than expected. The crucial
point of this argument is that this deterioration is not visible except after the economic slowdown, when the losses within lending
portfolio begin to materialize.
Previous economic deregulation
When interest rates are regulated it is not possible for banks to increase their client base by attractive lending rates that differ
significantly from the regulated rate. Demirguc-Kunt and Detragiache(1998) report that banking crises are more probable in
deregulated financial systems. They document a direct relation between the number of banking bankruptcies and the weakness of
regulatory control and the inefficacy or negligence of supervisory entities. Kaminski and Reinhart (1995) also point out the
coincidence of banking and stock market crises in deregulated economies.
Occurrence of a boom in the years prior to the crisis
According to Kindelberger (1978), “the Financial Crises are associated with the peaks of certain economic cycles… A financial crisis
is the culmination of a period of economic expansion that is about to end”. During the boom, the crisis is prepared by the
uncontrolled growth of the lending portfolios. Then the very growth of the lending portfolios cause a hike in lending interest rates
to the point where the demand for credit fades away. The higher interest rates can certainly cause a contraction of the economy. A
plunge in the price of bank assets and credit collaterals that safeguard the lending porfolios can in turn expose lending institutions
to bankruptcy risk.
Fall in the demand for money
Negative shocks and expectations of future devaluations produce reductions in the demand for money because economic agents try
to exchange their positions in local currency with equivalent positions in foreign currency or assets. These events occurred in
Venezuela in the years preceding the banking crisis, due to the political instability caused by two attempted coups d’etat in 1992 and
to the deep depression caused by the government’s budgetary deficit, the fall of the local stock and real estate markets. This fall in
the demand for money is very prejudicial for banks because it results in a steep reduction of the banking sector’s primary resource:
the size of the accumulated money mass deposited by bank customers. At the same time, because of the decrease in the demand
for money banks cannot replace quickly enough old expired loans with fresh loans that would warrant the continued profitability of
the loan portfolio.
Chronic government budget deficits
Chronic governement budget deficits are accompanied by interest rate hikes and inflation. “Due to the use by certain countries of
the exchenge rate as nominal anchor or inflation fighting tool, the presence of chronic budget deficits contributed to the real
appreciation that led to problems in the balance of payments and later to banking crises.”
Other Causes of Banking Crises
Contagion Effects and Bank Runs
According to Sundarajan and Baliño (1991), the contagion effect in the banking sector occurs when following the bankruptcy of a
given bank, the growing anxiety of the savings account owners affects other banks of reputation similar or worse than that of the
bank that went bankrupt. This phenomenon manifests itself in the form of bank runs by savings account owners on banks
suspected of being in financial trouble, bank runs that could result in the latter’s bankruptcy.
Faults in Bank Regulation and Supervision
To avoid the effects of a systemic or isolated banking crisis it is important that the entity in charge of supervising the banking
industry carry out its function in an effective manner. The relevant literature indicates that in many banking crises were made worse
by a complacent, irresponsible supervising entities. The manner or way in which the supervising entity reacts to a banking crisis is
also important: indecisiveness may highten public distrust or fear and an ever worsening contagion effect.. During the 3-month
long “closed-doors” intervention of Banco Latino in 1993-94, the supervising entities (SIB, Fogade, Ministerio de Hacienda and
BCV) gave the impression of being unprepared to deal with a banking crisis of that magnitude. This impression grew into a panic
that infected most of the banking system and resulted in the bankruptcy of 18 banks.
Consequences of the Banking Crisis
•
The consequences of the banking crises are numerous. Morón and Aguero (2003) cite the following:
•
They cause a great cost to the state in the recovery of deposits of the public (in countries with insured deposits) because of
the huge tranfer of funds to the bankrupt banks that must meet their obligations with their savings account owners.
•
They affect the system of payments, which is one of the main functions of banking institutions. As a result, transactions
become costly for all economic agents.
•
They cease to function as financial intermediaries, therefore many productive proyects fail to be realized for lack of
resources.
•
In case there is no deposit insurance, a great social cost results from the loss of savings. These social costs trigger political
tensions which if mishandled can degenerate into grave chaos.
•
The economy stops growing. Relevant literature shows that economic growth after a banking crisis is often negative.
•
Banking crises reduce the effect of monetary policies.
Can a banking crisis be predicted?
According to Berg and Pattiño (2000), the integration of the world markets, the velocity with which money moves from one
country to another, the globalization and the opening of capital make more complex the task of predicting banking crises. In their
study entitled “Dificultades para la predicción de crisis económicas”, Berg and Patiño argue that the banking crises in Mexico, Korea,
Thailand, Indonesia and Malasia share common causes, despite their exhibiting great differences that make difficult the construction
of a unique model of early crisis alert.
For example, in Thailand the main variables were
•
•
the ratio of deposit account deficits to the GNP
the variation in foreign reserves
whereas in Malasia the main variables were
•
•
•
the ratio of the deposit account deficits to the GNP
the variation in international reserves
the over-valuation of the exchange rate.
Finally Berg and Patiño study the crises in Korea and Indonesia and find a key role for
The ratio of the short term foreign debt and foreign reserves
Likewise, Kaminsky (1998) considers that in Latin America and in Thailand a key role is played by the great expansion of credit in
the banking system during the years prior to the crisis. He also argues that in Chile the banking crisis was to a great extent oriented
to the foreign sector. In Colombia a key role was played by
•
•
•
the m2 multiplier,
the domestic credit index with the GNP and
the m2 index on foreign reserves.
Berg and Pattiño (2000) conclude their work arguing that the models of prediction for banking crises must never be used in
isolation, but must be used with other suitable tools.
Brief summary of the 1993-1994 banking crisis in Venezuela
From 1992 on the fragility of the banking system in Venezuela begins to show. At that time Venezuela had just come out of an
economic Boom produced by the increase in the price of oil following the onset of the Gulf War (1990-1991). Then in February
and november of 1992 two attempts of coup d’etat occurred in Venezuela, both of which failed, creating a climate of extreme
economic, political and social instability. According to García (1998), prior to 1992 the Venezuelan Central Bank (BCV) carried out
a number of studies of the banking industry and had alerted the public regarding the fragility of that sector. In 1993 the BCV
informed the executive branch of the insolvency of a number of banking institutions. Despite these warnings, the oversight
agencies did not act quickly and thus allowed some banks to continue running deficits.
On December 22, 1993, the president of Banco Latino, the second largest bank in the country, resigns following rumors about the
precarious situation of the bank’s position. On January 12 of 1994 Banco Latino leaves the Bureau of Compensation and the next
day is intervened by the government behind closed doors. As was pointed out above, the magnitude of a banking crisis depends in
large measure on the way in which supervisory and regulatory agencies react. Given this, the intervention of Banco Latino behind
closed doors created a climate of uncertainty and panic among thousand of account owners of this financial institution. When the
public began to suspect that there were other banks in similar or worse conditions, a series of bank runs occurred, and the
contagion effect became generalized.
The regulatory and supervisory agencies’ lack of capacity to deal with the impending crisis became evident. During the first months
of the crisis a series of contraditcions plagued the agencies that were responsible for attending to the crisis; namely, the Venezuelan
Central Bank, Superintendencia de Bancos, Consejo Bancario Nacional, Fogade and the Ministry of Finance.
In june of 1994, the supervisory agency could no longer able to keep a secret that it was intervening Banco Maracaibo,
Construccion, Metropolitano, La Guaira, Amazonas and Bancor.
In these interventions they committed more mistakes, among which Garcia(1998) point out:
•
The granting of financial help under the assumption that the basic problem was one of lack of liquidity (and not one of
insolvency).
•
The granting of help without removal of management. In many cases these cash injections wound up in related accounts,
offshore banking accounts and foreign currency.
Some time after, these institutions were closed and liquidated anyway.
In August of 1994 new interventions occurred, this time affecting Banco Venezuela, Consolidado, Progreso, Italo-Venezuelan,
Profesional and Principal. But when Banco de Venezuela and Consolidado were intervened, a new open door model was put into
practice: the institutions were disinfected, bank management was removed from office and the banks were sold to interested
investors. A similar model was used with Banco Progreso, but when the confidence of account holders was not recovered –among
other things, the president of Banco Progresso was out of the country—the agency decided to tranfer the accounts to state owned
institutions. The rest of the banks intervened that month had the same fate as Banco Progreso: they were liquidated and closed.
At the end of 1994 and beginning 1995, Banco Andino, Empresarial, Union and Principal were intervened, from which the first and
third were liquidated and closed, while the second was forced to look for a partner that would give it capital or else sell off part of
itself. Despite this, the bank continued having solvency problems and was hard pressed by account owners and by competent
agencies to the point that the president had to travel to Colombia to a meeting with the directives of Banco Ganadero in order to
obtain written statement of intent to buy. The merger never took place but the news caused the situation to stabilize and the bank
runs ceased.
Behavior of some macroeconomic variables during the period of the Crisis.
The local interest rate spread
Although there are various ways of calculating that value, we decided to do so as the difference between the active average rate and
the passive average weighted annual rate.
The behavior of this variable should correspond (in theory at least) to the fact that as the crisis approached, risk loving banks would
want to rise their passive rate (diminishing the spread) in order to capture funds with which to pay the high levels of debt that they
had contracted. Let us see how events occurred in Venezuela during the years 89-93:
Spread Interest Rates
10.00
Spread
8.00
6.00
4.00
2.00
1990
1991
1992
1993
Año
As may be observed, the spread diminishes until 1992 and then bounces back up 50% by the end of 1993. One possible
explanation is that the troubled banks –in their need to capture more deposits—rised their passive (outbound) rates and then in
order to honor those passive rates had to rise their active (inbound) rates to a level where they sacrificed the quality of their credit
portfolio; that is, they allowed their credit portfolio to deteriorate in order to obtain more profits that would allow them to honor
their debts.
The real effective exchange rate
The money in circulation (m1) is the quantity composed by “coin and bills” and the “deposits in view,” whereas monetary liquidity
(m2) is equal to the sum of “m1,” the “savings deposits” and the “saving certificates.” This indicator shows us how in the years
previous to the crisis, because of the high passive (outbound) rates, the number of deposits placed in banks increased (specially
savings deposits and savings certificates), as one can see in the following graph:
m2 / m1
Monetary Liquidity Index (m2) / Circulant (m1)
4.00
3.50
3.00
2.50
2.00
1.50
1.00
0.50
1990
1991
1992
1993
AÑO
The graph shows that after 1992 this indicator moves up, with a clear preference on the part of the public to maintain savings in
banks instead of holding cash.
Percent change of IGAEM
As this is an indicator of economic activity, it behaves as the Internal Gross Product in times of banking crises, with a tendency to
decrease in the years leading to the crisis, which damages in a significant way the balances of the banks at the time when the
economic bubble bursts. Let us see how the IGAEM behaved during the period 1990-1993:
Growth IGAEM
106.00
104.00
102.00
100.00
98.00
96.00
94.00
92.00
90.00
1990
1991
1992
1993
Año
This indicator move just as expected. After a substantial increase in the period 1990-1991 (probably caused by the Persian Gulf
War), it began to come down at the end of 1992 (possibly because of the political instability following two attempted coup d’etats).
During the year prevous to the crisis, the percent change in this variable was negative.
Percent change in foreign reserves
In the years predating the crisis one can see how the level of foreign reserves (minus gold) is diminishing considerably. This drop is
due –as was explained in the case of the effective real interest rate—to the flight of capital caused by the social and political
upheaval during 1992 and to the perception of a deteriorating balance sheet in the banking sector.
Percentual Change International Reserves
100.00
80.00
60.00
40.00
20.00
(20.00)
1990
1991
1992
1993
Methodology
1.
2.
Our model is based on the one Developer by Sanz (1997).
Sanz (1997) builds 5 models with discriminant análisis using the financial data from the Venezuelan Banking Crisis previous to
1994. She uses the following variables:
Financial info. on 17 broken banks and 17 not broken 1 year before the crisis.
• Variables: A, B, C, D, E
Financial info. on 17 broken banks and 17 not broken 2 years before the crisis.
• Variables Financieras: F, G, H, I
Financial info. on 18 broken banks and 18 not broken 3 years before the crisis.
• Variables: J
Financial info. on 17 broken banks and 17 not broken 4 years before the crisis.
• Variables: K, L, M
Financial info. on 16 broken banks and 16 not broken, 5 years before the crisis.
• Variables: N, C
Advantages and limitations of Sanz work (1997)
Advantages
•
If we apply the five Sanz models (comparable to 5 different “syndromes”) to a given bank, the model according to which
this bank falls into bankruptcy (the “syndrome” best fitting the Bank) will also indicate, by construction, the life left to this
given bank.
Limitations
•
These five models are difficult to generalize and even to compare because they are based on different indicators (i.e.: each
“syndrome” is a set of “simptoms” that is disjoint with respect to the other “syndromes”). Only the indicator C appears in
more than one.
Indicators used
•
•
•
Our model, based in Discriminant Análisis, uses 5 of the financial variables used by Sanz, plus 5 more relevant variables of
macro-economic type.
The financial variables were selected considering the statistical significance found by Sanz(1997) for each one of them.
All the variables are conceptually justifiable and compatible with the relevant literatura on banks crises.
Advantages of the new model
The incorporation of macro-economic variables allows:
•
•
•
The possibility of using a single model along all the different years of the experiment, because these variables capture the
time-changes of the context common to all banks. (notice that for a given bank is much less indicative of structural problems to be in
critical condition during a global crisis than during bountiful times).
The possibility of evaluating the precision of the model year after year in comparable terms.
The possibility of extending this model to an international one where we could study banking and money fenomena
common to several interconnected but distinctly different economies, as it is the case of a banking crisis contagion among
countries whose economies are interconnected.
Classic financial indicators used in the discriminants model
•
•
•
•
•
Other assets/ total assets: measures the assets quality.
Financial outflows/average total inflows: effective passive interest rate, measures the average cost incurred by financial
institutions to pay for the resources provided by the public.
Operative margin/average assets: measures the assets profitability. Reflects the capacity of the institution to carry out
normal activities.
Operative margin/average equity: indicates the equity profitability and the ability of the institution to face its normal
operative activities.
Financial inflows/financial outflows: measures the spread generated by the difference between the rates paid to the public
and the ones collected from the credit portfolio and other investments.
Macroeconomic indicators used in the discriminants
•
•
•
•
•
Local interest rates spread: the differential between the weighted annual active interest rates and the passive ones..
Effective real Exchange rate: derived from the inflation adjusted nominal exchange rate.
Monetary liquidity index (m2) /circulating money (m1): indicates the public preference for savings over cash.
IGAEM percentual change: percentual change of the Venezuelan Leading Economic Indicator
Internacional reserves percentual changes (minus gold stock).
The procedure
•
•
We calibrate the model on “pre-crisis” data (c. 70 banks at the beginning, reduced to 36 because of data availability and
quality reasons)
We test the model on “post-crisis” data (68 Banks)
About the “pre-crisis”, 1990-1993 data used to calibrate the discriminants model
•
•
Data from 36 banks:
Each intervened bank was compared with a non-intervened with similar average capitalization, to obtain 18 intervened
banks and 18 non-interveened banks.
What does it really mean to calibrate the model? A word about the model resulting from the “pre-crisis” calibration:
After changing the axes of the vector space where each point represents a bank, we obtain the equation of a hiper-plane separating
two “Types of banks”:
-0.32F1 + 0.20F2 – 0.18F3 – 0.32F4 – 0.99F5 + 0.38M1 – 0.95M2 – 0.5M3 – 0.26M4 + 0.01M5 ≥ -190.395
Where F is a Financial indicador, M a Macro-economic Indicator and -190.395 the bankruptcy “treshold”
In practice, given a bank characterized wby the indicators F1, … , M5, whenever the weighted sum of its indicators is greater than
-190.395, the diagnosis is one of imminent bakruptcy, and vice-versa.
Linear Discriminant Análisis (Plate)
How the set of “post-crisis”, 1996-2004 data is used to test the discriminant model:
–
The formula Developed above is used to “probe” each of the 68 banks every 6 months.
At last, the indicator gives us a warning about the Banco Capital :
Year Capital Median Mean
StDev
2000 -150
-294
-335
144
1999 -272
-404
-464
620
1998 -270
-347
-367
111
1997 -289
-445
-465
-445
1996 -199
-327
-346
137
Banco Capital will be the only one among 68 others breaking the -190.395 threshold in 1996 and then again in June 2000.
For the other years, it falls quite close to the threshold. The Govt. closes Banco Capital on December 2000.
The change in Model Precision along the “pre-crisis” years:Adicional Conditions: error of Type I v.s. error of Type II
Model obtained from data collected starting 4, 3, 2 and 1 years before the crisis (140 observations).4 years before c.
3 years before c.
2 years before c.
1 year before c.
Yes
No
%
Yes
No
%
Yes
No
%
Yes
No
%
Yes
58
12
83
47
6
89
34
1
97
16
1
94
No
22
48
69
18
35
66
12
23
66
7
10
59
Accuracy
75.71%
77.36%
81.43%
76.47%
Note:
•
•
If we require that Error of Type I (bankruptcy omission) to be less than the Error of Type II (false bankruptcy alarm), the
model generates false alarms in a 10% of the “post-crisis” data.
This is compatible with the relevant literature: it is well known that most of the Early Warning Systems often generate
“false alarms”.
Bibliography
•
Berg, Andrew y Pattiño, Catherine, Dificultades para la predicción de crisis económicas, Fondo Monetario Internacional
Washington, 2000.
•
Dominic Barton, Roberto Newell y Gregory Wilson, Dangerous Markets: Managing in Financial Crises. 2003.
•
García Gustavo, Lecciones de la Crisis Bancaria de Venezuela, Ediciones IESA, 1998.
•
Kaminsky, Graciela y Reinhart, Carmen, The Twin Crises: The Causes of Banking and Balance-of-Payments Problems, American
Economic Review, 1995.
•
Kaminsky, Graciela, Currency and Banking Crises: The Early Warnings of Distress, Board of Governors of the Federal Reserve
System, 1998.
•
Kindelberger Charles, Manias, panics and crashes, New York, NY: Basic Books Inc, 1978.
•
Minsky, Hyman P., The Financial-Instability Hypothesis: Capitalist Process and the Behavior of the Economy, Cambridge University
Press, 1982.
•
Miron J.A., Financial Panics, the Seasonality of the Nominal Interest Rate, and the Founding of the FED, American Economic
Review, 1986.
•
Morón Eduardo y Aguero, Loo-kung, Sistema de Alerta Temprana de Fragilidad Financiera, Universidad del Pacifico,
2003.
•
Paracare Elsy y Zany Victor Fermín, Un Indicador Mensual de Actividad Económica (IGAEM), Serie Enero 2000 del
Banco Central de Venezuela.
•
Schwartz, Anna, Real and Pseudo financial crises, New York: San Martin’s Press, 1985.
| 5cs.CE
|
Static Analysis of Lockless Microcontroller C Programs
Eva Beckschulze
Sebastian Biallas
Stefan Kowalewski
Embedded Software Laboratory
RWTH Aachen University, Germany
{lastname}@embedded.rwth-aachen.de
Concurrently accessing shared data without locking is usually a subject to race conditions resulting
in inconsistent or corrupted data. However, there are programs operating correctly without locking
by exploiting the atomicity of certain operations on a specific hardware. In this paper, we describe
how to precisely analyze lockless microcontroller C programs with interrupts by taking the hardware
architecture into account. We evaluate this technique in an octagon-based value range analysis using
access-based localization to increase efficiency.
1
Introduction
Static analysis based on abstract interpretation [7] is a formal method that found its way into practice
by several commercial code analysis tools. Proving the absence of run-time errors in microcontroller
programs is of particular importance as microcontrollers are often deployed in safety-critical systems.
However, C code analyzers usually do not cope with C extensions and hardware-specific control prevalent
in microcontroller programs. This control is not only necessary for data input/output but also needed to
implement interrupt service routines (ISRs), which allows some form of concurrency and can be used for
asynchronous hardware communication and periodic tasks. Since the control functions of the hardware
are often exposed through normal memory accesses, a sound analysis of microcontroller programs has
to reflect these registers in its memory model. On the Atmega16 [2], for example, it is possible to
enable/disable interrupts by a write to the SREG register which is located at the memory address 0x5F.
Beside these peculiarities in programming microcontrollers, software engineers often rely on additional
assumptions outside the scope of standard C semantics on how the compiler will translate a program and
on how the microcontroller behaves w. r. t. the atomicity of some operations. For example, they might
omit the locking of shared data because an 8 bit read/write is always executed atomically (such algorithms
are typically called lockless [10] or lockfree). This saves program space on the controller as well as
execution time but makes a precise analysis on C code level particularly challenging. In this paper, we
deal with atomicity assumptions when analyzing interrupts. We exploit the characteristics of interrupts
w. r. t. concurrency to design an efficient fixed-point computation.
1.1
Concurrency Induced by Interrupts
Compared to concurrency implemented by threads concurrency induced by interrupts exhibits some
essential differences [17]. While threads can preempt each other, interrupts can preempt the main program
but the main program cannot interrupt an ISR. An interrupt can only trigger if both the specific interrupt is
enabled and interrupts are globally enabled. Locks to guarantee atomic sections that are not interrupted
are, therefore, implemented by disabling interrupts globally. By default interrupts are disabled in ISRs
such that an ISR runs to completion. Explicitly enabling interrupts in ISRs is allowed but due to the
F. Cassez, R. Huuck, G. Klein and B. Schlich (eds.):
Systems Software Verification Conference 2012 (SSV 2012)
EPTCS 102, 2012, pp. 103–114, doi:10.4204/EPTCS.102.10
© Beckschulze et. al.
This work is licensed under the
Creative Commons Attribution License.
104
Static Analysis of Lockless Microcontroller C Programs
limited stack size an error-prone approach. In this paper, we concentrate on programs without such
nested interrupts. Considering all these specifics of interrupts we can design a more precise analysis of
microcontroller C code.
1.2
Analysis Framework
We consider interrupts in the context of a data flow analysis evaluating pointers and value ranges based
on the octagon abstract domain [13], in which the relations between variables (memory locations) x, y
are expressed as inequalities ±x ± y ≤ c, where c is a constant. To consider hardware dependencies, our
memory model is augmented with hardware-specific knowledge [4] so as to capture, e. g., the setting
or resetting of interrupt enable bits. In this paper, we show how to extend this analysis to interrupts by
including hardware specifics and taking the C semantics into account.
1.3
Contribution and Outline
The contribution of this paper is twofold: (a) We develop a set of rules which lockless C programs
must follow to behave predictable under different compilers. (b) We present a combined analysis of
value ranges, pointers and interrupts for lockless microcontroller C programs. This analysis combines
ramifications of the C memory model with understanding of the underlying hardware to allow a sound
representation of lockless code.
Our paper is laid out as follows. First, Sect. 2 introduces our technique exemplified on a lockless
UART driver. Then, Sect. 3 details the notion of atomicity we implemented to analyze such programs
on a C and hardware-specific level. Our analysis is described in Sect. 4 and is evaluated in Sect. 5. The
papers ends with a survey of related work in Sect. 6 and a concluding discussion in Sect. 7.
2
Motivating Example
In this section, we introduce a UART driver that operates without locking shared data by exploiting the
atomicity of certain operations on the specific hardware architecture. We discuss different approaches to
analyze such a program.
2.1
Lockless UART driver
Consider the source code excerpt in Fig. 1 implementing the receiver of a UART (Universal Asynchronous
Receiver Transmitter) driver that is supposed to run on an AVR microcontroller1 . The driver uses a FIFO
rx_buff to buffer incoming data software-based in addition to the hardware-implemented buffer. An
integer variable rx_in (rx_out) is used as an index to denote the position where the next byte is to be
stored (where the next byte is to be read). The function getNextPos is called to increment the index by
one or to reset it to 0 when the index is out of bounds. Reading a byte out of the hardware register called
UDR and storing it in the FIFO buffer is performed by ISR, an interrupt service routine that is triggered by
hardware. The function getByte returns the data located at position rx_out. We assume that the global
interrupt enable bit is set initially and remains set all the time, while the specific interrupt used by the
UART is disabled when the buffer is full (line 36) and enabled when there is at least one free position
(line 27). Hence, the functions getByte and isEmpty might be interrupted anywhere in between two
1 This
is a slightly modified excerpt of the code found here http://www.mikrocontroller.net/topic/101472#882716
Beckschulze et. al.
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
# define vu8 ( x )
105
(*( volatile uint8 *) &( x ) )
uint8 rx_buff [ RX0_SIZE ];
uint8 rx_in ;
uint8 rx_out ;
uint8 getNextPos ( uint8 pos , size ) {
pos ++;
if ( pos >= size ) {
return 0;
}
return pos ;
}
uint8 isEmpty () {
return rx_out == vu8 ( rx_in ) ;
}
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
uint8 getByte () {
uint8 data ;
while ( isEmpty () ) ;
data = rx_buff [ rx_out ];
// get byte
rx_out = getNextPos ( rx_out , RX0_SIZE ) ;
URX0_IEN = 1;
// enable RX interrupt
return data ;
}
ISR ( USART0_RX_vect ) {
uint8 i = rx_in ;
i = getNextPos (i , RX0_SIZE ) ;
if ( i == rx_out ) {
// buffer overflow
URX0_IEN = 0;
// disable RX interrupt
return ;
}
rx_buff [ rx_in ] = UDR ;
rx_in = i ;
}
Figure 1: UART driver
operations. By way of contrast, the ISR always runs to completion due to the automatic global interrupt
disable implemented by hardware. As both the main program and the ISR access rx_out and rx_in and
one of the accesses is a write access, these variables could be subject to a data race. However, as reading
and writing an 8-bit variable on an 8-bit processor architecture is atomic, locking is unnecessary in this
case.
2.2
Analyzing Interrupts
A typical question verified by static analyses is that of all array accesses being within the bounds of the
array. This requires a value range analysis of variables to determine possible values for variables used as
indices to access an array. Interrupts have to be considered during the analysis in two ways: First, as calls
to ISRs are not visible in the code, ISR code has to be added to the control flow and taken into account
appropriately. For example, this can be done by nondeterministic calls to the ISR between two control
flow nodes. Second, we need to deal with shared variable accesses, i. e., an access to a variable in the
main program that might be performed incompletely before an interrupt is triggered. As this may result in
corrupted data, care has to be taken for such race conditions to design a sound value range analysis. Next,
we discuss the assumptions made during analysis, first in case the analysis is designed for an arbitrary
hardware platform and second in case of hardware specifics that can be used to refine the analysis.
2.2.1
Hardware Agnostic Approach
Static analyzers for C code usually do not consider any hardware specifics such as interrupts. To analyze
microcontroller C programs using generic static analyzers, the user is advised to annotate the program
to provide the analyzer with further constraints. Without an appropriate annotation, for example, the
analyzer will take the ISR function as dead code since it cannot see the implicit calls by the hardware.
Further annotations are required to deal correctly with atomic sections that can be defined by toggling the
interrupt bit in some microcontroller dependent absolute address. The analysis of ISRs is integrated into
another analysis by interleaving all expressions outside of atomic sections with non-deterministic calls to
ISRs. If the analysis encounters an expression in the main program that accesses a variable that is also
accessed by an ISR and one of the accesses is a write access then there is a race condition. Therefore, to
be sound, a value range analysis unaware of the hardware architecture has to assume that this variable
106
Static Analysis of Lockless Microcontroller C Programs
may take any value within its type bounds. We find such race conditions in the UART example in Fig. 1:
• Read access to rx_in in line 18, concurrent write access to rx_in in line 40
• Write access to rx_out in line 26, concurrent read access to rx_out in line 35
Unfortunately, assuming type bounds for these variables that are used as an index to the buffer rx_buffer
results in a presumed array out of bounds access which is spurious. In the next section we discuss how
considering a hardware-specific memory model can refine the analysis and avoid this false alarm.
2.2.2
Considering Hardware Specifics
Programming microcontrollers is intrinsically tied to dealing with the specific hardware. Taking hardware
specifics into account as well when designing static analyses avoids tedious user annotations while
increasing precision of analyses. In a hardware-specific memory model, an access to an absolute address
(register) is linked to the semantics of this register [4]. This way, interrupts can be identified and added
automatically to the control flow wherever they may occur. We delay the discussion of reducing the
number of ISR analyses further to Sect. 4. With respect to the shared read/write access mentioned above,
hardware considerations enhance the precision of the analysis significantly. Knowing that the target
platform is an 8-bit architecture, we conclude that reading or writing to 8-bit variables will always be
performed atomically. Thus, rx_in and rx_out always have consistent values. To make sure that all
possible values are considered when reading rx_in, it is sufficient to compute a fixed-point over the ISR
in advance and propagate it non-deterministically. Similarly, before writing to rx_out, the analysis of the
ISR takes account of the old abstract value of rx_out while the analysis of the ISR after the write access
will consider its new abstract value. Note that it is not the C expression that we assume to be atomic, but
the eventual load or store instruction executed by the processor. In the following section, we will detail
this notion of atomicity.
3
Requirements for Lockless Code
Precisely analyzing shared data in concurrent programs on a high abstraction level such as C code is
usually not possible as we are unaware how a compiler translates a C code expression into processor
instructions. On the other hand, stable concurrent programs should not depend on a certain compiler
version or compilation options. In this section, we try to infer basic rules under which lockless C programs
can operate in a well-defined manner. We then enrich these rules by certain hardware specifics: In the
last section, e. g., we argued that writing or reading to an 8-bit variable cannot be interrupted on an 8-bit
hardware architecture, which gives rise to a basic form of atomic access.
With the rules derived in this section, we achieve two goals: First, we can detect program errors
(which might manifest themselves depending on compiler specifics) by checking whether a program
follows these rules. Second, we can formulate the foundations of a sound and precise analysis for each
program which follows these rules.
3.1
Atomicity at the Level of C
In the absence of locks, we assume that all data shared between the main function and the ISRs is
performed using volatile accesses only (i. e., either by casting the access to a volatile data type or declaring
the variable volatile in the first place). Accessing non-volatile shared data is prone to failure and thus
reported, since an optimizing compiler might eliminate seemingly unnecessary reloads and dead stores.
Beckschulze et. al.
107
...
Translation of one C expression
Figure 2: Possible Control Flow on Instruction Level
In the following, we call a C expression competing if it contains an access to volatile data. The term
competing stresses that competing expressions should be evaluated in a well-defined order.
C compilers are allowed to schedule loads, stores and calculations of expressions in arbitrary order
as long as this does not alter the visible behavior of the program (as-if rule). In the presence of volatile
accesses, however, at least all volatile objects must be stable at sequence points [11, Sect. 5.1.2.3]. To
illustrate this, consider the shared variables a, b and the statement a = ++b; which writes twice to shared
data between sequence points. Let us further assume that we have the precondition a ≤ b. Although
we know that a and b must be stable after this statement, the assignment operator forms no sequence
point, and a compiler might thus translate this expression into one of the following two pseudo-assembly
snippets:
1
2
3
4
LOAD
INC
STORE
STORE
temp , b
temp
b , temp
a , temp
1
2
3
4
LOAD
INC
STORE
STORE
temp , b
temp
a , temp
b , temp
In the right snippet, a is stored before b. An ISR invoked between line 3 and line 4 might thus observe
that a > b, while a ≤ b is an invariant in the left snippet. Hence, we want to detect and warn about such
statements.
Now, consider an arbitrary C expression between two sequence points. Fig. 2 shows the control
flow graph for such an expression assuming that the compiler translated it into a certain sequence of
instructions (white nodes), where each instruction might be followed by one or more calls to an ISR (gray
nodes). Between the two sequence points, we do not know how the compiler translated the expression.
This gives rise to two requirements for writing stable lockless code:
1. The effect of the execution of an ISR between two sequence points must be covered by executing
the ISR at the sequence point before or after the expression.
2. The effect of the execution of an ISR must not depend on the exact instructions generated by the
compiler.
The first requirement stems from the fact that the analysis – as well as the programmer – can predict the
program state only at sequence points, and thus, the ISR behavior should only depend on this predictable
state. The second requirement can be regarded as a corollary of the first: If a compiler creates a different
set of instructions (because of, e. g., different options), we still want requirement 1 to hold. We call
expressions that fulfill these requirements well-formed.
Observe that the well-formedness of expressions depends on two distinct properties: How much
freedom the compiler has to translate an expression (especially scheduling loads and stores) and which
atomic primitives are provided by the underlying hardware. We will defer the discussion of the latter to
Sect. 3.2 and for now assume that all loads and stores of shared variables cannot be interrupted by the
ISRs. We will first formally define well-formed expressions inductively:
108
Static Analysis of Lockless Microcontroller C Programs
1. A constant expression expr ::= const and a variable expression expr ::= v is well-formed.
2. A unary expression expr ::=
expr1 with
∈ {−, !,˜} is well-formed iff expr1 is well-formed.
3. Let expr ::= expr0 expr1 be a binary expression, ∈ {+, −, /, ∗, %, , , |, &, ˆ, <, <=, >, >=,
==, ! =}. Then, expr is well-formed iff (a) expr0 and expr1 are well-formed and (b) at most one of
expr0 and expr1 is competing.
4. Let expr ::= expr0 ./ expr1 be a comma or logic expression, ./ ∈ {||, &&,“,”}. Since ./ forms a
sequence point, expr is well-formed iff expr0 and expr1 are well-formed.
5. Let expr ::= fun(expr0 , . . . , exprn ) be a function call. Function calls are sequence points, yet the order in which the arguments are evaluated is not defined. Thus, expr is well-formed if expr0 , . . . , exprn
are well-formed and at most one expri is competing. Additionally, expr is competing iff the body of
fun accesses shared data.
6. An assignment expr ::= lvalue = expr1 is well-formed iff (a) lvalue is not competing and expr1 is
well-formed or (b) expr1 is well-formed and does not write shared data (this does not span into
functions which are being called).
Intuitively, a well-formed expression is an expression in which the order of all accesses of shared data
is determined by the C standard, i. e., the evaluation of competing expressions is ordered by sequence
points. Additionally, we require that we have at most one write to shared data in a well-formed expression
(follows from 6.). Furthermore, note that the second part of bullet point 5. forbids delicate cases such as a
= f()+g() where f and g access shared data (the order in which f and g are invoked is not defined). Yet,
expressions such as a = f()+1 are well-formed even if a is shared and f is competing.
The latter point deserves a more detailed study: The assignment operator = does not form a sequence
point. Hence, expressions such as a = b = 0 are not well-formed, since the store to a and b can be
performed in arbitrary order. However, the compiler must only generate one store to the left-hand-side
of an assignment, which depends on the evaluation of the right-hand-side. Thus, if the right-hand-side
contains no writes to shared data, such as in a = b, an assignment is well-formed. We allow this construct
because it is typically used in lockless code. Alternatively, this construct could be avoided by introducing
temporary (unshared) variables.
The design rules that we derive to achieve robustness of lockless programs against compiler reordering
are simple. Function call expressions should be either full expressions or should adhere to bullet point 5.
Further, the programmer is advised to avoid subexpression with side effects (accessing a volatile object
is a side effect, too). By splitting up complex expression in several simple ones without side effects in
subexpression or by at least encapsulating them in simple assignments the desired order of evaluation is
made explicit. These are well-known design rules that usually aim at well-defined (compiler-independent)
behavior of single-threaded code. However, it also contributes to well-formed expressions which matter in
lockless concurrent programs only.
Finally, we require the compiler to behave sensible. That is, in essence, that loads and stores to volatile
data are translated to elementary load and store operations (one instruction where possible). Note that this
requirement is, to the best of our knowledge, fulfilled by all compilers used in industry and is exactly what
a programmer expects – and exploits – when writing lockless code. In the next section, we will connect
the general requirements of C to the specific offerings of the hardware to create a concise memory model.
3.2
Atomicity on the Hardware Level
Crucial for our approach is a hardware model that reflects atomicity at the assembly level. In the last
section, we assumed that each elementary load and store of shared data can be performed atomically,
Beckschulze et. al.
109
i. e., it cannot be interrupted. Depending on the type and size of the data object, this is in general not the
case. However, accessing data of the size of general purpose registers can usually be performed by one
instruction, which then is not interruptible. Loading and storing data of different sizes, on the other hand,
is usually performed by a sequence of instructions. Thus, an ISR might interrupt this sequence and read
or modify corrupted data. Our analysis allows to configure the size of atomic data types beforehand and
aims to detect cases where atomic accesses are required but not possible.
Additionally, some processors provide atomic primitives such as compare-and-swap as one instruction.
These instructions can only be accessed using compiler specific functions, which are usually built-in
or provided as inline assembly. Our analysis can simulate these instructions atomically if the functions
provided by the compiler are annotated appropriately. With this structure in place, we can now describe
our analysis in the context of interrupts and lockless code.
4
Designing Analyses Considering Interrupts
Analyzing ISRs and the sound handling of shared data requires adaptations of the existing data flow
analysis. In the following, we first describe the original analysis and subsequently discuss how to integrate
ISR analysis and shared data handling.
4.1
Original Analysis
The original analysis is a combined analysis of pointers and value range analysis based on octagons [13].
It is a fixed-point computation iterating over the nodes in the control flow graph (CFG) and propagating
the computed results along control flow edges until old and new results coincide. The analysis is flow and
context sensitive, i. e., it determines an abstract value for each node in the CFG and distinguishes different
calling contexts for function calls. The octagon analysis evaluates nodes concerning value ranges while an
address taking node is handled by the pointer analysis.
In order to increase efficiency of the octagon based analysis, we apply access-based localization, a
technique that reduces the abstract state that is propagated when analyzing a function to the subset of
the abstract state that is actually accessed in this function or in subroutines [3, 15]. The set of accessed
variables in each function is over-approximated by a pre-analysis based on a flow-insensitive pointer
analysis.
4.2
Invoking Interrupts in Fixed-Point Computation
For efficiency reasons, ISRs shall only be analyzed if necessary. Therefore, we make use of our memory
model [4] in order to find accesses to absolute addresses that correspond to setting or clearing the global
or individual interrupt bits. The current status of the interrupt bits is added to the abstract state propagated
along the CFG. The naïve approach would be to trigger ISR analysis in between two arbitrary nodes
where the interrupt is enabled. However, the execution of an ISR does not necessarily affect all subsequent
operations of the main program. Precisely stated, an ISR affects an operation in the main program if
and only if the operation accesses data that might be written by the ISR. In order to incorporate such
dependencies ISRs have to be analyzed before analyzing a shared data access. In our analysis, we trigger
ISR analysis after each node that enables interrupts (globally or individually). This way we also ensure
that ISRs are analyzed at least once between two atomic sections.
With respect to the control flow, the ISR analysis corresponds to an extra node calling the ISRs added
between an interrupt-enabling node and its successor node while the direct connection between these
110
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
Static Analysis of Lockless Microcontroller C Programs
main () {
...
sei () ;
...
a = b;
...
cli () ;
...
}
ISR () {
...
b = c;
...
}
sei();
// interrupt enable
// shared data access
ISR();
...
// interrupt disable
a=b
ISR();
...
// shared data access
cli();
Figure 3: Code example (left) and its control flow graph (right) enhanced with necessary non-deterministic
calls to the ISR that need to be considered in the analysis
nodes is kept to consider the case that no interrupt is triggered (cf. Fig 3). The analysis applies the join
operation to the abstract states of both branches. We compute a fixed-point over all ISRs that might
be triggered considering every possible order and frequency of interrupts in order to over-approximate
the number of possibly occurring interrupts. Indeed, on an AVR microcontroller each ISR execution is
followed by at least one main program instruction. However, the exact number of possible ISR executions
cannot be determined on C level, as we are unaware how the compiler translates a given code sequence
into instructions.
Still, analyzing interrupts at one location between two atomic sections is not sufficient if shared data
is written in the main program and read by the ISR. Therefore, we trigger another analysis of ISRs after
analyzing a shared access node (cf. Fig 3). As previously noted, interrupts in the analysis are considered
non deterministic but possibly occurring infinitely often. Note that analyzing ISRs both before and after a
shared access node is sufficient if the requirements in Sect. 3.1 are fulfilled.
Finally, to speed up the analysis, we apply access-based localization as explained above also to
ISRs. For this purpose, we arrange the pre-analysis collecting accessed variables to analyze ISRs as
well. Additionally, we subdivide accessed variables into read and written variables in order to be able to
distinguish the different cases of shared access.
4.3
Sound Analysis of Shared Data
Shared data handling depends on whether a full expression (cf. [11, Sect. 6.8]) is well-formed. As our
control flow nodes often represent subexpressions of a full expression, we add to each node whether the
corresponding full expression is well-formed. Table 1 reviews the cases of shared accesses and indicates
the behavior of the analysis. We omitted the case where both the main program and an ISR only read the
same variable as it does not raise any problem.
The first row of the table shows the case that the shared accesses are atomic, the main program
writes(reads) and an ISR read (writes) the shared variables and the corresponding full expression is
well-formed. In this case the octagon analysis is performed as usual. All issues of concurrency are in this
case considered by triggering the ISR analysis somewhere before and immediately after this node. In the
second row we consider a well-formed full expression that includes an atomic write of a variable that is
also written in an ISR. Though it does not cause corrupted data, one write may immediately overwrite the
other one. As this might be unintended by the programmer, we issue a warning that data loss might occur.
The third row considers the case that the analysis encounters a full expression where all shared accesses
are atomic but the expression is not well-formed. In this case the analysis sets all shared variables to type
Beckschulze et. al.
111
atomic
access
access
type
well-formed
expression
analysis behavior
yes
yes
yes
no
r/w
w/w
*
*
yes
yes
no
*
no special behavior
issue a warning
set shared variables to type bounds, issue a warning
set shared variables to type bounds, issue a warning
Table 1: Cases of shared access
Program
Loc
# global vars
Time
# of ISR analyses
# Warnings
(legitimate)
# Warnings
(spurious)
UART buffer
RGB-LED
Traffic Light
175
95
68
32
22
5
7.3 s
0.8 s
0.1 s
136
27
30
0
1
0
1
0
0
Table 2: Results of the case study
bounds thereby overapproximating any possible order of execution. Note that these overapproximations
are also propagated to the subsequent ISR analysis. We issue a warning that this expression is unspecified
behavior. The case that the access is non-atomic (row 4) is handled by the analysis the same way. Here,
we issue the warning that a non-atomic access might result in inconsistent or corrupted data.
5
Case Studies
Our implementation is written in JAVA and builds on the Eclipse Framework. We used it to analyze several
microcontroller programs written for the AVR ATmega 16 microcontroller unit. For this processor our
analysis assumes that only 8 bit memory accesses are performed atomically.
The results of the case study are presented in Table 2. For each checked program, we provide the lines
of code, number of global variables, time for analysis, number of times we have to check the ISRs and
the number of (spurious) warnings. The analyzer was able to prove the absence of array out of bounds
accesses in the UART buffer. Yet, we still had a spurious warning, since this program writes to the data
buffer without locking from both the ISR and the main function. The correctness of such an operation
has to be checked manually. In the RGB-LED program, we found an unlocked shared access to a 16-bit
variable, which was a legitimate warning. Finally, the Traffic Light program, controlling an intersection
with two traffic lights, could be checked without triggering a warning.
6
Related Work
Traditionally, model checking has been used to verify concurrent programs such as in [9] where partial
order reduction is used to increase efficiency. Schlich et al. [18] implement this technique for model
checking embedded software on the assembly level. Atig et al. [1] describe how to model check in the
presents of a weak memory model, which corresponds to the lockless programs described in this paper.
Regehr and Cooprider [17] describe how to map microcontroller programs (interrupt-driven code) to
thread-driven code. In particular, they point out the differences between threads and interrupts and show
112
Static Analysis of Lockless Microcontroller C Programs
how to exploit existing techniques for verification tools for multithreaded code to verify interrupt-driven
embedded software. Cooprider [6] describes how to increase efficiency by only analyzing ISRs at certain
locations. His approach, however, is restricted to properly locked programs.
Pratikakis et al. describe in [16] the LOCKSMITH tool which can check automatically for proper locking of shared data in an abstract interpretation framework. We focus on the verification of microcontroller
programs even in the absence of locks.
Recently, Miné presented an extensive work [14] on analyzing embedded real-time parallel C programs
based on abstract interpretation. He defines the semantics of expressions based on interference, i. e.,
whenever a thread reads a variable x, all abstract values another thread might set x to are also considered.
These interference sets are non-relational and flow-insensitive information while we interchange relational
and flow-sensitive information between the main program and ISRs. Due to the flow-insensitivity dealing
with the order of execution of C expressions is superfluous in Miné’s approach, while it is essential in
our approach. Additionally, his approach differs from ours in considering all primitive statements to be
atomic independent of types and the underlying hardware. This way shared data access is not handled
correctly in case of incomplete assignments.
In [8] Fehnker at al. extend the generic (unsound) static analyzer Goanna to detect microcontroller
specific bugs. In their approach, the CFG is labeled with occurrences of syntactic constructs of interest,
while the desired behavior is put into a CTL formula that can be checked by model checking techniques.
Their work focuses on integrating hardware information to specify and check simple rules that should
be followed when programming the specific microcontroller. Instead, this paper advocates the use of a
hardware-specific memory model to enhance precision of data flow analyses and to avoid false alarms.
Finally, there is an ongoing effort to formalize memory models for existing languages, which recently
cumulated in a memory model for the new C++ standard [5]. Yet, programmers of microcontroller software
still rely on non-strictly defined semantics and “sensible” compiler behavior mixed with knowledge about
the underlying hardware. Our approach aims to implement these assumptions, which are sometimes quite
subtle, in a verification framework.
7
Concluding Discussion
This paper advocates a static analysis for lockless microcontoller C programs combining different
techniques to make the approach precise as well as tractable. To achieve precision for such programs, it is
necessary to deal both with C semantics and hardware specifics. Our memory model reflects what the
user can (and will) expect from the compiler on the one side and what atomic primitives the hardware can
provide on the other side. Using the derived rules for well-formed expressions, we can detect latent bugs
that would manifest themselves only in certain circumstances (such as different compiler options) and we
can detect bugs that result from improper communication between the main program and the ISR using
our fine-grained value analysis.
Still – as Meyers and Alexandrescu [12] point out – thread-unaware languages such as C pose inherent
difficulties to write thread-aware code. It makes proper (manual) synchronization exceptionally hard
since all corner cases of the languages have to be considered. Additionally, as we have shown in this
paper, it poses obstacles for the program analysis, since the order of certain operation is often unclear.
In such cases, either imprecise non-relational analyses have to be deployed or a careful analysis of all C
expressions is necessary.
Beckschulze et. al.
113
Acknowledgements
This work was supported, in part, by the DFG research training group 1298 Algorithmic Synthesis of
Reactive and Discrete-Continuous Systems and by the DFG Cluster of Excellence on Ultra-high Speed
Information and Communication, German Research Foundation grant DFG EXC 89. Further, the work
of Sebastian Biallas was supported by the DFG. We thank Volker Kamin for sharing his thoughts on the
ideas described in this paper, and the anonymous referees for their helpful comments.
References
[1] M. F. Atig, A. Bouajjani, S. Burckhardt & M. Musuvathi (2010): On the verification problem for weak
memory models. In: Proceedings of the 37th annual ACM SIGPLAN-SIGACT symposium on Principles of
programming languages, POPL ’10, ACM, New York, NY, USA, pp. 7–18, doi:10.1145/1706299.1706303.
[2] Atmel Corporation (2009): The Atmel 8-bit AVR Microcontroller with 16K Bytes of In-system Programmable
Flash. www.atmel.com/atmel/acrobat/doc2466.pdf.
[3] E. Beckschulze, J. Brauer & S. Kowalewski (2012): Access-Based Localization for Octagons. Electronic
Notes in Theoretical Computer Science 287(0), pp. 29 – 40, doi:10.1016/j.entcs.2012.09.004. Proceedings of
the Fourth International Workshop on Numerical and Symbolic Abstract Domains, (NSAD 2012).
[4] E. Beckschulze, J. Brauer, A. Stollenwerk & S. Kowalewski (2011): Analyzing Embedded Systems
Code for Mixed-Critical Systems Using Hybrid Memory Representations. In: 14th IEEE Int. Symp.
on Object/Component/Service-Oriented Real-Time Distributed Computing Workshops, IEEE, pp. 33–40,
doi:10.1109/ISORCW.2011.40.
[5] H.-J. Boehm & S. V. Adve (2008): Foundations of the C++ concurrency memory model. In: Proceedings of
the 2008 ACM SIGPLAN conference on Programming language design and implementation, PLDI ’08, ACM,
New York, NY, USA, pp. 68–78, doi:10.1145/1375581.1375591.
[6] N. Cooprider (2008): Data-flow analysis for interrupt-driven microcontroller software. Ph.D. thesis, The
University of Utah, Salt Lake City, UT, USA.
[7] P. Cousot & R. Cousot (1977): Abstract Interpretation: A Unified Lattice Model for Static Analysis
of Programs by Construction or Approximation of Fixpoints. In: POPL, ACM Press, pp. 238–252,
doi:10.1145/512950.512973.
[8] A. Fehnker, R. Huuck, B. Schlich & M. Tapp (2009): Automatic Bug Detection in Microcontroller Software
by Static Program Analysis. In: SOFSEM 2009: Theory and Practise of Computer Science,Spindleruv Mlýn,
Czech Republic, Lecture Notes in Computer Science 5404, Springer, pp. 267–278, doi:10.1007/978-3-54095891-8_26.
[9] P. Godefroid (1996): Partial-Order Methods for the Verification of Concurrent Systems - An Approach to the
State-Explosion Problem. LNCS 1032, Springer, doi:10.1007/3-540-60761-7.
[10] T. E. Hart, P. E. McKenney & A. D. Brown (2006): Making Lockless Synchronization Fast: Performance
Implications of Memory Reclamation. In: 20th IEEE International Parallel and Distributed Processing
Symposium, Rhodes, Greece, doi:10.1109/IPDPS.2006.1639261.
[11] ISO (1999): ANSI/ISO/IEC 9899-1999: Programming Languages — C. International Organization for
Standardization, Geneva, Switzerland.
[12] S. Meyers & A. Alexandrescu (2004):
C++ and the perils of double-checked locking.
http://www.aristeia.com/Papers/DDJ_Jul_Aug_2004_revised.pdf.
[13] A. Miné (2006): The Octagon Abstract Domain. Higher-Order and Symbolic Computation 19(1), pp. 31–100,
doi:10.1007/s10990-006-8609-1.
[14] A. Miné (2012): Static Analysis of Run-Time Errors in Embedded Real-Time Parallel C Programs. Logical
Methods in Computer Science 8(1–26), pp. 1–63.
114
Static Analysis of Lockless Microcontroller C Programs
[15] H. Oh, L. Brutschy & K. Yi (2011): Access analysis-based tight localization of abstract memories. In:
Proceedings of the 12th International Conference on Verification, Model Checking, and Abstract Interpretation,
VMCAI’11, Springer-Verlag, Berlin, Heidelberg, pp. 356–370, doi:10.1007/978-3-642-18275-4_25.
[16] P. Pratikakis, J. S. Foster & M. Hicks (2011): LOCKSMITH: Practical static race detection for C. ACM Trans.
Program. Lang. Syst. 33(1), pp. 3:1–3:55, doi:10.1145/1889997.1890000.
[17] J. Regehr & N. Cooprider (2007): Interrupt Verification via Thread Verification. Electron. Notes Theor.
Comput. Sci. 174(9), pp. 139–150, doi:10.1016/j.entcs.2007.04.002.
[18] B. Schlich, T. Noll, J. Brauer & L. Brutschy (2009): Reduction of Interrupt Handler Executions for Model
Checking Embedded Software. In: Haifa Verification Conference (HVC 2009), Haifa, Israel, LNCS 6405,
Springer, pp. 5–20, doi:10.1007/978-3-642-19237-1_5.
| 6cs.PL
|
From Regexes to Parsing Expression Grammars
Sérgio Medeiros
arXiv:1210.4992v1 [cs.FL] 17 Oct 2012
Department of Computer Science – UFS – Aracaju – Brazil
Fabio Mascarenhas
Department of Computer Science – UFRJ – Rio de Janeiro – Brazil
Roberto Ierusalimschy
Department of Computer Science – PUC-Rio – Rio de Janeiro – Brazil
Abstract
Most scripting languages nowadays use regex pattern-matching libraries.
These regex libraries borrow the syntax of regular expressions, but have an informal semantics that is different from the semantics of regular expressions,
removing the commutativity of alternation and adding ad-hoc extensions
that cannot be expressed by formalisms for efficient recognition of regular
languages, such as deterministic finite automata.
Parsing Expression Grammars are a formalism that can describe all deterministic context-free languages and has a simple computational model.
In this paper, we present a formalization of regexes via transformation to
Parsing Expression Grammars. The proposed transformation easily accommodates several of the common regex extensions, giving a formal meaning to
them. It also provides a clear computational model that helps to estimate
the efficiency of regex-based matchers, and a basis for specifying provably
correct optimizations for them.
Keywords: regular expressions, parsing expression grammars, natural
semantics, pattern matching, regexes
Email addresses: [email protected] (Sérgio Medeiros), [email protected] (Fabio
Mascarenhas), [email protected] (Roberto Ierusalimschy)
Preprint submitted to Science of Computer Programming
October 19, 2012
1. Introduction
Regular expressions are a concise way for describing regular languages
with an algebraic notation. Their syntax has seen wide use in pattern matching libraries for programming languages, where they are used to specify the
pattern against which a user is trying to match a string, or, more commonly,
the pattern that a user is searching for in a string.
Regular expressions used for pattern matching are known as regexes [1,
2], and while they look like regular expressions they often have different
semantics, based on how the pattern matching libraries that use them are
actually implemented.
A simple example that shows this semantic difference are the regular
expressions a|aa and aa|a, which both describe the language {a, aa}. It is
trivial to prove that the | operator of regular expressions is commutative given
its common semantics as the union of sets. But the regexes a|aa and aa|a
behave differently for some implementations of pattern matching libraries
and some subjects.
The standard regex libraries of the Perl and Ruby languages, as well as
PCRE [3], a regex library with bindings for many programming languages,
give different results when matching these two regexes against the subject
aa. In all three libraries the first regex matches just the first a of the subject,
while the second regex matches the whole subject. With the subject ab both
regexes give the same answer in all three libraries, matching the first a, but
we can see that the | operator for regexes is not commutative.
This behavior of regexes is directly linked to the way they are usually implemented, by trying the alternatives in a | expression in the order they appear and backtracking when a particular path through the expression makes
the match fail.
A naive implementation of regex matching via backtracking can have
exponential worst-case running time, which implementations try to avoid
through ad-hoc optimizations to cut the amount of backtracking that needs
to be done for common patterns. These ad-hoc optimizations lead to implementations not having a cost model of their operation, which makes it
difficult for users to determine the performance of regex patterns. Simple
modifications can make the time complexity of a pattern go from linear to
exponential in unpredictable ways [4, 5].
Regexes can also have syntactical and semantical extensions that are difficult, or even impossible, to express through pure regular expressions. These
2
extensions do not have a formal model, but are informally specified through
how they modify the behavior of an implementation based on backtracking. The meaning of regex patterns that use the extensions may vary among
different regex libraries [6], or even among different implementations of the
same regex library [7].
Practical regex libraries try to solve performance problems with ad-hoc
optimizations for common patterns, but this makes the implementation of a
regex library a complex task, and is another source of unpredictable performance, as different implementations can have different performance characteristics.
A heavily optimized regex engine, RE2 [8], uses an implementation based
on finite automata and guarantees linear time performance, but it relies on
ad-hoc optimizations to handle more complex patterns, as a naive automatabased implementation can have quadratic behavior [9]. More importantly, it
cannot implement some common regex extensions [8].
Parsing Expression Grammars (PEGs) [10] are a formalism that can express all deterministic context-free languages, which means that PEGs can
also express all regular languages. The syntax of PEGs is based on the syntax of regular expressions and regexes, and PEGs have a formal semantics
based on ordered choice, a controlled form of backtracking that, like the |
operation of regexes, is sensitive to the ordering of the alternatives.
We believe that ordered choice makes PEGs a suitable base for a formal
treatment of regexes, and show, in this paper, that we can describe the
meaning of regex patterns by conversion to PEGs. Moreover, PEGs can
be efficiently executed by a parsing machine that has a clear cost model
that we can use to reason about the time complexity of matching a given
pattern [11, 12]. We can then use the semantics of PEGs to reason about the
behavior of regexes, for example, to optimize pattern matching and searching
by avoiding excessive backtracking. We believe that the combination of the
regex to PEG conversion and the PEG parsing machine can be used to build
implementations of regex libraries that are simpler and easier to extend than
current ones.
The main contribution of this paper is our formalization of a simple,
structure-preserving translation from plain regular expressions to PEGs that
can be used to translate regexes to PEGs that match the same subjects. We
present a formalization of regular expressions as patterns that match prefixes
of strings instead of sets of strings, using the framework of natural semantics [13]. In this semantics, regular expressions are just a non-deterministic
3
form of the regexes used for pattern matching. We show that our semantics is equivalent to the standard set-based semantics when we consider the
language of a pattern as the set of prefixes that it matches.
We then present a formalization of PEGs in the same style, and use it
to show the similarities and differences between regular expressions, regexes,
and PEGs. We then define a transformation that converts a regular expression to a PEG, and prove its correctness. We also show how we can improve
the transformation for some classes of regexes by exploiting their properties
and the greater predictability and control of performance that PEGs have,
improving the performance of the resulting PEGs. Finally, we show how
our transformation can be adapted to accommodate four regex extensions
that cannot be expressed by regular expressions: independent expressions,
possessive and lazy repetition, and lookahead.
There are procedures for transforming deterministic finite automata and
right-linear grammars to PEGs [11, 14] and, as there are transformations from
regular expressions to these formalisms, we could have used these existing
procedures as the basis of an implementation of regular expressions in PEG
engines. But the transformations of regular expressions to these formalisms
cover just a subset of regexes, not including common extensions, including
those covered in Section 6 of this paper. The direct transformation we present
here is straightforward and can cover regex extensions.
In the next section, we present our formalizations of regular expressions
and PEGs, and discuss when a regular expression has the same meaning when
interpreted as a regular expression and as a PEG, along with the intuition
behind our transformation. In Section 3, we formalize our transformation
from regular expressions to PEGs and prove its correctness. In Section 4
we show how we can reason about the performance of PEGs to improve the
PEGs generated by our transformation in some cases. In Section 5 we show
how our approach compares to existing regex implementations with some
benchmarks. In Section 6 we show how our transformation can accommodate
some regex extensions. Finally, in Section 7 we discuss some related work
and present our conclusions.
2. Regular Expressions and PEGs
Given a finite alphabet T , we can define a regular expression e inductively
as follows, where a ∈ T , and both e1 and e2 are also regular expressions:
e = ε
a
e1 e2
4
e1 | e2
e∗1
Traditionally, a regular expression can also be ∅, but we will not consider
it; ∅ is not used in regexes, and any expression with ∅ as a subexpression can
be either rewritten without ∅ or is equal to ∅.
Note that this definition gives an abstract syntax for expressions, and
this abstract syntax is what we use in the formal semantics and proofs.
In our examples we use a concrete syntax that assumes that juxtaposition
(concatenation) is left-associative and has higher precedence than |, which
is also left-associative, while ∗ has the highest precedence, and we will use
parentheses for grouping when these precedence and associativity rules get
in the way.
The language of a regular expression e, L(e), is traditionally defined
through operations on sets. Intuitively, the languages of ε and a are singleton sets with the corresponding symbols, the language of e1 e2 is given by
concatenating all strings of L(e1 ) with all strings of L(e2 ), the language of
e1 | e2 is the union of the languages of e1 and e2 , and S
the language of e∗1 is the
i
0
Kleene closure of the language of e1 , that is, L∗ = ∞
i=0 L where L = {ε}
and Li = LLi−1 for i > 0 [15, p. 28].
We are interested in a semantics for regexes, the kind of regular expressions used for pattern matching and searching, so we will define a matching
RE
RE
relation for regular expressions, ❀. Informally, we will have e xy ❀ y if and
only if the expression e matches the prefix x of input string xy.
RE
Formally, we define ❀ via natural semantics, using the set of inference
RE
rules in Figure 1. We have e xy ❀ y if and only if we can build a proof tree
for this statement using the inference rules. The rules follow naturally from
the expected behavior of each expression: rule empty.1 says that ε matches
itself and does not consume the input; rule char.1 says that a symbol matches
and consumes itself if it is the beginning of the input; rule con.1 says that
a concatenation uses the suffix of the first match as the input for the next;
rules choice.1 and choice.2 say that a choice can match the input using
either option; finally, rules rep.1 and rep.2 say that a repetition can either
match ε and not consume the input or match its subexpression and match
the repetition again on the suffix that the subexpression left.
The following lemma proves that the set of strings that expression e
RE
matches is the language of e, that is, L(e) = {x ∈ T ∗ | ∃y e xy ❀ y, y ∈ T ∗ }.
Lemma 1. Given a regular expression e and a string x, for any string y we
RE
have that x ∈ L(e) if and only if e xy ❀ y.
Proof. (⇒): By induction on the complexity of the pair (e, x). Given the
5
Empty String
RE
ε x ❀x
(empty.1)
RE
Concatenation
e1 xyz ❀ yz
Character
RE
RE
e1 | e2 xy ❀ y
e2 xy ❀ y
(choice.1)
RE
e1 | e2 xy ❀ y
RE
Repetition
RE
e∗ x ❀ x
(con.1)
RE
RE
Choice
(char.1)
RE
e2 yz ❀ z
e1 e2 xyz ❀ z
e1 xy ❀ y
RE
a ax ❀ x
(rep.1)
e xyz ❀ yz
(choice.2)
RE
e∗ yz ❀ z
RE
e∗ xyz ❀ z
, x 6= ε (rep.2)
RE
Figure 1: Natural semantics of relation ❀
pairs (e1 , x1 ) and (e2 , x2 ), the first pair is more complex than the second
one if and only if either e2 is a proper subexpression of e1 or e1 = e2 and
|x1 | > |x2 |. The base cases are (ε, ε) and (a, a), and their proofs follow
by application of rules empty.1 and char.1, respectively. Cases (e1 e2 , x)
and (e1 | e2 , x) use a straightforward application of the induction hypothesis
on the subexpressions, followed by application of rule con.1 or one of the
choice rules. Case (e∗ , ε) follows directly from rule rep.1, while for case
(e∗ , x), where x 6= ε, we know by the definition of the Kleene closure that
x ∈ Li (e1 ) with i > 0, where Li (e1 ) is L(e1 ) concatenated with itself i
times. This means that we can decompose x into x1 x2 , with a non-empty x1 ,
where x1 ∈ L(e1 ) and x2 ∈ Li−1 (e1 ). Again by the definition of the Kleene
closure this means that x2 ∈ L(e∗1 ). The proof now follows by the induction
hypothesis on (e1 , x1 ) and (e∗1 , x2 ) and an application of rule rep.2.
RE
(⇐): By induction on the height of the proof tree for e xy ❀ y. Most
cases are straightforward; the interesting case is when the proof tree concludes
with rule rep.2. By the induction hypothesis we have that x ∈ L(e1 ) and
y ∈ L(e∗1 ). By the definition of the Kleene closure we have that y ∈ Li (e1 ), so
xy ∈ Li+1 (e1 ) and, again by the Kleene closure, xy ∈ L(e∗1 ), which concludes
the proof.
Salomaa [16] developed a complete axiom system for regular expressions,
6
where any valid equation involving regular expressions can be derived from
the axioms. The axioms of system F1 are:
e1 | (e2 | e3)
e1 (e2 e3 )
e1 | e2
e1 (e2 | e3)
(e1 | e2 )e3
e|e
εe
∅e
e|∅
e∗
e∗
=
=
=
=
=
=
=
=
=
=
=
(e1 | e2 ) | e3
(e1 e2 )e3
e2 | e1
e1 e2 | e1 e3
e1 e3 | e2 e3
e
e
∅
e
ε | e∗e
(ε | e)∗
(1)
(2)
(3)
(4)
(5)
(6)
(7)
(8)
(9)
(10)
(11)
Salomaa’s regular expressions do not have the ε case; the original axioms
use ∅∗ , which has the same meaning, as the only possible proof trees for ∅∗
use rep.1. The following lemma shows that these axioms are valid under our
semantics for regular expressions, if we take e1 = e2 to mean that e1 and e2
match the same sets of strings.
Lemma 2. For each of the axioms of F1 , if l is the expression on the left
RE
side and r is the expression on the right side, we have that l xy ❀ y if and
RE
only if r xy ❀ y.
Proof. Trivially true for axiom 8, as there are no proof trees for either the
left or right sides of this axiom. For axioms 1 to 7 and for axiom 9 it is
straightforward to use the subtrees of the proof tree of one side to build a
proof tree for the other side. We can prove the validity of axiom 11 by an
straightforward induction on the height of the proof trees for each side.
For axiom 10, we need to prove the identity a∗ a = aa∗ , by induction on
the heights of the proof trees. From this identity the left side of axiom 10
follows by taking the subtrees of rep.2 and combining them with con.1 into
a tree for aa∗ , which means we have a tree for a∗ a that we can use to build
a tree for the right side using choice.2. The right side follows from getting
a tree for aa∗ from a tree for a∗ a using the identity, then taking its subtrees
and using rep.2 to get a tree for a∗ .
7
Parsing expression grammars borrow the syntax of regular expressions.
A parsing expression is also defined inductively, extending the inductive definition of regular expressions with two new cases, A for a non-terminal, and
!e for a not-predicate of expression e. A PEG G is a tuple (V, T, P, pS ) where
V is the set of non-terminals, T is the alphabet (set of terminals), P is a
function from V to parsing expressions, and pS is the parsing expression that
the PEG matches (its starting parsing expression). We will use the notation
G[p] for a grammar derived from G where pS is replaced by p while keeping
V , T , and P the same. We will refer to both regular expressions and parsing
expressions as just expressions, letting context disambiguate between both
kinds.
While the syntax of parsing expressions is similar to the syntax of regular expressions, the behavior of the choice and repetition operators is very
different. Choice in PEGs is ordered; a PEG will only try to match the right
side of a choice if the left side cannot match any prefix of the input. Repetition in PEGs is possessive; a repetition will always consume as much of
the input as it can match, regardless of whether this leads to failure or a
shorter match for the whole pattern1 . To formally define ordered choice and
possessive repetition we also need a way to express that an expression does
not match a prefix of the input, so we need to introduce fail as a possible
outcome of a match.
PEG
Figure 2 gives the definition of ❀, the matching relation for PEGs. As
PEG
with regular expressions, we say that G xy ❀ y to express that the grammar
G matches the prefix x of input string xy, and the set of strings that a PEG
PEG
matches is its language: L(G) = {x ∈ T | ∃y G xy ❀ y, y ∈ T ∗ }.
We mark with a ∗ the rules that have been changed from Figure 1, and
mark with a + the rules that were added. Unmarked rules are unchanged
from Figure 1, except for the trivial change of adding the parameter G to
the relation. We have six new rules, and two changed rules. New rules
char.2 and char.3 generate fail in the case that the expression cannot
match the symbol in the beginning of the input. New rule con.2 says that a
concatenation fails if its left side fails. New rule var.1 says that to match a
non-terminal we have to match the parsing expression associated with it in
this grammar’s function from non-terminals to parsing expressions (the non1
Possessive repetition is a consequence of ordered choice, as e∗ is the same as expression
A where A is a fresh non-terminal and P (A) = eA | ε.
8
PEG
Empty String
PEG
(empty.1)
G[P (A)] x ❀ X
Non-terminal
PEG
G[ε] x ❀ x
Terminal
(char.1)
PEG
, b 6= a (char.2)+
PEG
G[a] ax ❀ x
PEG
G[a] ε ❀ fail
PEG
PEG
G[p2 ] y ❀ X
PEG
G[p1 ] x ❀ fail
(con.1)
PEG
G[p1 p2 ] xy ❀ X
G[p1 ] xy ❀ y
PEG
PEG
G[p1 ] x ❀ fail
(choice.1)
PEG
(rep.1)∗
G[p] xyz ❀ yz
PEG
G[p∗ ] yz ❀ z
PEG
G[p∗ ] x ❀ x
(rep.2)
G[p∗ ] xyz ❀ z
PEG
PEG
Not Predicate
(choice.2)∗
G[p1 | p2 ] x ❀ X
PEG
PEG
G[p] x ❀ fail
PEG
G[p2 ] x ❀ X
PEG
G[p1 | p2 ] xy ❀ y
Repetition
(con.2)+
G[p1 p2 ] x ❀ fail
PEG
Ordered Choice
(char.3)+
PEG
G[b] ax ❀ fail
G[p1 ] xy ❀ y
Concatenation
(var.1)+
G[A] x ❀ X
G[p] x ❀ fail
PEG
(not.1)+
G[!p] x ❀ x
G[p] xy ❀ y
PEG
(not.2)+
G[!p] xy ❀ fail
PEG
Figure 2: Definition of Relation ❀ through Natural Semantics
terminal’s “right-hand side” in the grammar). New rules not.1 and not.2
say that a not predicate never consumes input, but fails if its subexpression
matches a prefix of the input.
The change in rule con.1 is trivial and only serves to propagate fail, so
we do not consider it an actual change. The changes to rules choice.2 and
rep.1 are what actually implements ordered choice and possessive repetition,
respectively. Rule choice.2 says that we can only match the right side of
the choice if the left side fails, while rule rep.1 says that a repetition only
stops if we try to match its subexpression and fail.
It is easy to see that PEGs are deterministic; that is, a given PEG G can
only have a single result (either fail or a suffix of x) for some input x, and
only a single proof tree for this result. If the PEG G always yields a result
for any input in T ∗ then we say that G is complete [10]. PEGs that are not
complete include any PEG that has left recursion and PEGs with repetitions
e∗ where e matches the empty string. From now on we will assume that any
PEG we consider is complete unless stated otherwise. The completeness of
9
a PEG can be proved syntactically [10].
The syntax of the expressions that form a PEG are a superset of the
syntax of regular expressions, so syntactically any regular expression e has a
corresponding PEG Ge = (V, T, P, e), where V and P can be anything. We
can prove that L(Ge ) ⊆ L(e) by a simple induction on the height of proof
PEG
trees for Ge xy ❀ y, but it is easy to show examples where L(Ge ) is a
proper subset of L(e), so the regular expression and its corresponding PEG
have different languages.
For example, expression a | ab has the language {a, ab} as a regular expression but {a} as a PEG, because on an input with prefix ab the left side
of the choice always matches and keeps the right side from being tried. The
same happens with expression a (b | bb), which has language {ab, abb} as a
regular expression and {ab} as a PEG, and on inputs with prefix abb the left
side of the choice keeps the right side from matching.
The behavior of the PEGs actually match the behavior of the regexes a | ab
and a (b | bb) on Perl-compatible regex engines. These engines will always
match a | ab with just the first a on subjects starting with ab, and always
match a (b | bb) with ab on subjects starting with abb, unless the order of the
alternatives is reversed.
A different situation happens with expression (a | aa) b. As a regular
expression, its language is {ab, aab} while as a PEG it is {ab}, but the
PEG fails on an input with prefix aab. Regex engines will backtrack and
try the second alternative when b fails to match the second a, and will also
match aab, highlighting the difference between the unrestricted backtracking
of common regex implementations and PEGs’ restricted backtracking.
If we take the previous regular expression, (a | aa) b, and distribute b over
the two alternatives, we have ab | aab. This expression now has the same
language when we interpret it either as a regular expression, as a PEG, or
as a regex.
If we change a | ab and a (b | bb) so they have the prefix property2 , by adding
an end-of-input marker $, we have the expressions (a | ab) $ and a (b | bb) $.
Now their languages as regular expressions are {a$, ab$} and {ab$, abb$},
respectively, but the first expression fails as a PEG on an input with prefix
ab and the second expression fails as a PEG on an input with prefix abb.
Both (a | ab) $ and a (b | bb) $ match, as regexes, the same set of strings that
2
There are no distinct strings x and y in the language such that x is a prefix of y.
10
form their languages as regular expressions. They are in the form (e1 | e2 ) e3 ,
like (a | aa) b, so we can distribute e3 over the choice to obtain e1 e3 | e2 e3 . If
we do that the two expressions become a$ | ab$ and a (b$ | bb$), respectively,
and they now have the same language either as a regular expression, as a
PEG, or as a regex.
We will say that a PEG G and a regular expression e over the same
alphabet T are equivalent if the following conditions hold for every input
string xy:
PEG
RE
G xy ❀ y ⇒ e xy ❀ y
RE
PEG
e xy ❀ y ⇒ G xy 6 ❀ fail
(12)
(13)
That is, a PEG G and a regular expression e are equivalent if L(G) ⊆ L(e)
and G does not fail for any string with a prefix in L(e). In the examples
above, regular expressions a | ab, a (b | bb), a$ | ab$, a (b$ | bb$), and ab | aab
are all equivalent with their corresponding PEGs, while (a | ab) $, a (b | bb) $,
and (a | aa) b are not.
Informally, a PEG and a regular expression will be equivalent if the PEG
matches the same strings as the regular expression when the regular expression is viewed as a regex under the common “leftmost-first” semantics of
Perl-compatible regex implementations. If a regular expression can match
different prefixes of the same subject, due to the non-determinism of the
choice and repetition operations, the two conditions of equivalence guarantee that an equivalent PEG will match one of those prefixes.
Regexes are deterministic, and will also match just a single prefix of
the possible prefixes a regular expression can match. Our transformation
will preserve the ordering among choices, so the prefix an equivalent PEG
obtained with our transformation matches will be the same prefix a regex
matches.
While equivalence is enough to guarantee that a PEG will give the same
results as a regex, equivalence together with the prefix property yields the
following lemma:
Lemma 3. If a regular expression e with the prefix property and a PEG G
are equivalent then L(G) = L(e).
Proof. As our first condition of equivalence says that L(G) ⊆ L(e), we just
need to prove that L(e) ⊆ L(G). Suppose there is a string x ∈ L(e); this
RE
means that e xy ❀ y for any y. But from equivalence this means that
11
PEG
PEG
G xy 6 ❀ fail. As G is complete, we have G xy ❀ y ′. By equivalence, the
prefix of xy that G matches is in L(e). Now y cannot be a proper suffix of
y ′ nor y ′ a proper suffix of y, or the prefix property would be violated. This
means that y ′ = y, and x ∈ L(G), completing the proof.
We can now present an overview on how we will transform a regular expression e into an equivalent PEG. We first need to transform subexpressions
of the form (e1 | e2 ) e3 to e1 e3 | e2 e3 . We do not need to actually duplicate
e3 in both sides of the choice, potentially causing an explosion in the size
of the resulting expression, but can introduce a fresh non-terminal X with
P (X) = e3 , and distribute X to transform (e1 | e2 ) e3 into e1 X | e2 X.
Transforming repetition is trickier, but we just have to remember that
∗
e1 e2 ≡ (e1 e∗1 | ε) e2 ≡ (e1 e∗1 e2 ) | e2 . Naively transforming the first expression
to the third does not work, as we end up with e∗1 e2 in the expression again,
but we can add a fresh non-terminal A to the PEG with P (A) = e1 A | e2
and then replace e∗1 e2 with A in the original expression. The end result of
repeatedly applying these two transformation steps until we reach a fixed
point will be a parsing expression that is equivalent to the original regular
expression.
As an example, let us consider the regular expression b∗ b$. Its language
is {b$, bb$, . . .}, but when interpreted as a PEG the language is ∅, due to
possessive repetition. If we transform the original expression into a PEG
with starting parsing expression A and P (A) = bA | b$, it will have the same
language as the original regular expression; for example, given the input
bb$, this PEG matches the first b through subexpression b of b A, and then
A tries to match the rest of the input, b$. So, once more subexpression b of
b A matches b and then A tries to match the rest of the input, $. Since both
b A and b$ fail to match $, A fails, and thus b A fails for input b$. Now we
try b$, which successfully matches b$, and the complete match succeeds.
If we now consider the regular expression b∗ b, which has {b, bb, . . .} as its
language, we have that it also has the empty set as its language if we interpret
it as a PEG. A PEG with starting parsing expression A and P (A) = bA | b
also has {b, bb, . . .} as its language, but with an important difference: the
regular expression can match just the first b of a subject starting with bb, but
the PEG will match both, and any other ones that follow, so we do not have
RE
PEG
e xy ❀ y implying G xy ❀ y. But the behavior of the PEG corresponds the
behavior of regex engines, which use greedy repetition, where a repetition will
always match as much as it can while not making the rest of the expression
12
Π(ε, Gk ) = Gk
Π(a, Gk ) = Gk [a pk ]
Π(e1 e2 , Gk ) = Π(e1 , Π(e2 , Gk ))
Π(e1 | e2 , Gk ) = G2 [p1 | p2 ], where G2 = (V2 , T, P2 , p2 ) = Π(e2 , (V1 , T, P1 , pk ))
and
Π(e∗1 ,
(V1 , T, P1 , p1 ) = Π(e1 , Gk )
Gk ) = G , where G = (V1 , T, P1 ∪ {A → p1 | pk }, A) with A ∈
/ Vk and
(V1 , T, P1 , p1 ) = Π(e1 , (Vk ∪ {A}, T, Pk , A))
Figure 3: Definition of Function Π, where Gk = (Vk , T, Pk , pk )
fail.
The next section formalizes our transformation, and proves that for any
regular expression e it will give a PEG that is equivalent to e, that is, it
will yield a PEG that recognizes the same language as e if it has the prefix
property.
3. Transforming Regular Expressions to PEGs
This section presents function Π, a formalization of the transformation
we outlined in the previous section. The function Π transforms a regular
expression e using a PEG Gk that is equivalent to a regular expression ek to
yield a PEG that is equivalent to the regular expression e ek .
The intuition behind Π is that Gk is a continuation for the regular expression e, being what should be matched after matching e. We use this
continuation when transforming choices and repetitions to do the transformations of the previous section; for a choice, the continuation is distributed
to both sides of the choice. For a repetition, it is used as the right side for the
new non-terminal, and the left side of this non-terminal is the transformation
of the repetition’s subexpression with the non-terminal as continuation.
For a concatenation, the transformation is the result of transforming the
right side with Gk as continuation, then using this as continuation for transforming the left side. This lets the transformation of (e1 | e2 ) e3 work as
expected: we transform e3 and then use the PEG as the continuation that
we distribute over the choice.
We can transform a standalone regular expression e by passing a PEG
with ε as starting expression as the continuation; this gives us a PEG that
is equivalent to the regular expression e ε, or e.
13
Figure 3 has the definition of function Π. Notice how repetition introduces
a new non-terminal, and the transformation of choice has to take this into
account by using the set of non-terminals and the productions of the result of
transforming one side to transform the other side, so there will be no overlap.
Also notice how we transform a repetition by transforming its body using
the repetition itself as a continuation (through the introduction of a fresh
non-terminal), then building a choice between the transformation and the
body and the continuation of the repetition. The transformation process is
bottom-up and right-to-left.
We will show the subtler points of transformation Π with some examples.
In the following discussion, we use the alphabet T = {a, b, c}, and the continuation grammar Gk = (∅, T, ∅, ε) that is equivalent to the regular expression ε. In our first example, we use the regular expression (a | b | c)∗ a (a | b | c)∗,
which matches an input that has at least one a.
We first transform the second repetition by evaluating Π((a | b | c)∗, Gk );
we first transform a | b | c with a new non-terminal A as continuation, yielding
the PEG aA | bA | cA, then combine it with ε to yield the PEG A where A
has the production below:
A → aA | bA | cA | ε
Next is the concatenation with a, yielding the PEG aA. We then use this
PEG as continuation for transforming the first repetition. This transformation uses a new non-terminal B as a continuation for transforming a | b | c,
yielding aB | bB | cB, then combines it with aA to yield the PEG B with the
productions below:
B → aB | bB | cB | aA
A → aA | bA | cA | ε
When the original regular expression matches a given input, we do not
know how many a’s the first repetition matches, because the semantics of
regular expressions is non-deterministic. Regex implementations commonly
resolve ambiguities in repetitions by the longest match rule, where the first
repetition will match all but the last a of the input. PEGs are deterministic
by construction, and the PEG generated by Π obeys the longest match rule.
The alternative a A of non-terminal B will only be tried if all the alternatives
fail, which happens in the end of the input. The PEG then backtracks until
14
the last a is found, where it matches the last a and proceeds with non-terminal
A.
The regular expression (b | c)∗ a (a | b | c)∗ defines the same language as the
regular expression of the first example, but without the ambiguity. Now Π
with continuation Gk yields the following PEG B:
B → bB | cB | aA
A → aA | bA | cA | ε
Although the productions of this PEG and the previous one match the
same strings, the second PEG is more efficient, as it will not have to reach
the end of the input and then backtrack until finding the last a. This is
an example on how we can use our semantics and the transformation Π to
reason about the behavior of a regex. The relative efficiency of the two PEGs
is an artifact of the semantics, while the relative efficiency of the two regexes
depends on how a particular engine is implemented. In a backtracking implementation it will depend on what ad-hoc optimizations the implementation
makes, in an automata implementation they both will have the same relative
efficiency, at the expense of the implementation lacking the expressive power
of some regex extensions.
The expressions in the two previous examples are well-formed. A regular
expression e is well-formed if it does not have a subexpression e∗i where ε ∈
L(ei ). If e is a well-formed regular expression and Gk is a complete PEG then
Π(e, Gk ) is also complete. In Section 3.1 we will show how to mechanically
obtain a well-formed regular expression that recognizes the same language as
a non-well-formed regular expression while preserving its overall structure.
We will now prove that our transformation Π is correct, that is, if e is
a well-formed regular expression and Gk is a PEG equivalent to a regular
expression ek then Π(e, Gk ) is equivalent to e ek . The proofs use a small
technical lemma: each production of PEG Gk is also in PEG Π(e, Gk ), for any
regular expression e. This lemma is straightforward to prove by structural
induction on e.
We will prove each property necessary for equivalence separately; equivalence will then be a direct corollary of those two proofs. To prove the first
property we need an auxiliary lemma that states that the continuation grammar is indeed a continuation, that is if the PEG Π(e, Gk ) matches a prefix
x of a given input xy then we can split x into v and w with x = vw and Gk
matching w.
15
Lemma 4. Given a well-formed regular expression e, a PEG Gk , and an
PEG
input string xy, if Π(e, Gk ) xy ❀ y then there is a suffix w of x such that
PEG
Gk wy ❀ y.
Proof. By induction on the complexity of the pair (e, xy). The interesting
case is e∗ . In this case Π(e∗ , Gk ) gives us a grammar G = (V1 , T, P, A),
PEG
where A → p1 | pk . By var.1 we know that G[p1 | pk ] xy ❀ y. There are
now two subcases to consider, choice.1 and choice.2.
PEG
For subcase choice.2, we have G[pk ] xy ❀ y. But then we have that
PEG
Gk [pk ] xy ❀ y because any non-terminal that pk uses to match xy is in both
G and Gk and has the same production in both. The string xy is a suffix of
itself, and pk is the starting expression of Gk , closing this part of the proof.
PEG
For subcase choice.1 we have Π(e, Π(e∗ , Gk )) xy ❀ y, and by the
PEG
induction hypothesis Π(e∗ , Gk ) wy ❀ y. We can now use the induction
hypothesis again, on the length of the input, as w must be a proper suffix of
PEG
x. We conclude that Gk w ′ y ❀ y for a suffix w ′ of w, and so a suffix of x,
ending the proof.
The following lemma proves that if the first property of equivalence holds
between a regular expression ek and a PEG Gk then it will hold for e ek and
Π(e, Gk ) given a regular expression e.
Lemma 5. Given two well-formed regular expressions e and ek and a PEG
PEG
RE
PEG
Gk , where Gk wy ❀ y ⇒ ek wy ❀ y, if Π(e, Gk ) vwy ❀ y then
RE
e ek vwy ❀ y.
Proof. By induction on the complexity of the pair (e, vwy). The interesting
case is e∗ . In this case, Π(e∗ , Gk ) gives us a PEG G = (V1 , T, P, A), where
PEG
A → p1 | pk . By var.1 we know that G[p1 | pk ] vwy ❀ y. There are now two
PEG
subcases, choice.1 and choice.2 of ❀.
PEG
For subcase choice.2, we can conclude that Gk vwy ❀ y because pk is
the starting expression of Gk and any non-terminals it uses have the same
RE
production both in G and Gk . We now have ek vwy ❀ y. By choice.2 of
RE
R
E
RE
❀ we have e e∗ ek | ek vwy ❀ y, but e e∗ ek | ek ≡ e∗ ek , so e∗ ek vwy ❀ y,
ending this part of the proof.
PEG
For subcase choice.1, we have Π(e, Π(e∗ , Gk )) vwy ❀ y, and by Lemma 4
PEG
we have Π(e∗ , Gk ) wy ❀ y. The string v is not empty, so we can use
RE
the induction hypothesis and Lemma 4 again to conclude e∗ ek wy ❀ y.
16
PEG
Then we use the induction hypothesis on Π(e, Π(e∗ , Gk )) vwy ❀ y to
RE
RE
conclude e e∗ ek vwy ❀ y. We can now use rule choice.1 of ❀ to get
RE
RE
e e∗ ek | ek vwy ❀ y, but e e∗ ek | ek ≡ e∗ ek , so e∗ ek vwy ❀ y, ending the
proof.
The following lemma proves that if the second property of equivalence
holds between a regular expression ek and a PEG Gk then it will hold for
e ek and Π(e, Gk ) given a regular expression e.
Lemma 6. Given well-formed regular expressions e and ek and a PEG Gk ,
RE
PEG
where Lemma 5 holds and we have ek wy ❀ y ⇒ Gk wy 6❀ fail, if
RE
PEG
e ek vwy ❀ y then Π(e, Gk ) vwy 6❀ fail.
Proof. By induction on the complexity of the pair (e, vwy). The interesting
case is e∗ . We will use again the equivalence e∗ ek ≡ e e∗ ek | ek . There are
RE
two subcases, choice.1 and choice.2 of ❀.
For subcase choice.1, we have that e matches a prefix of vwy by rule
RE
con.1. As e∗ is well-formed this prefix is not empty, so e∗ ek v ′ wy ❀ y for a
PEG
proper suffix v ′ of v. By the induction hypothesis we have Π(e∗ , Gk ) v ′ wy 6 ❀
PEG
fail, and by induction hypothesis again we get Π(e, Π(e∗ , Gk )) vwy 6 ❀
PEG
fail. This PEG is complete, so we can conclude Π(e∗ , Gk )[p1 | pk ] vwy 6 ❀
PEG
PEG
fail using rule choice.1 of ❀, and then Π(e∗ , Gk ) vwy 6 ❀ fail by rule
var.1, ending this part of the proof.
For subcase choice.2, we can assume that there is no proof tree for the
RE
statement e e∗ ek vwy ❀ y, or we could reduce this subcase to the first one by
using choice.1 instead of choice.2. Because Π(e, Π(e∗ , Gk )) is complete we
PEG
can use modus tollens of Lemma 5 to conclude that Π(e, Π(e∗ , Gk )) vwy ❀
RE
PEG
fail. We also have ek vwy ❀ y, so Gk vwy 6 ❀ fail. Now we can
PEG
PEG
use rule choice.2 of ❀ to conclude G[p1 | pk ] vwy 6 ❀ fail, and then
PEG
Π(e∗ , Gk ) vwy 6 ❀ fail by rule var.1, ending the proof.
The correctness lemma for Π is a corollary of the two previous lemmas:
Lemma 7. Given well-formed regular expressions e and ek and a PEG Gk ,
where ek and Gk are equivalent, then Π(e, Gk ) and e ek are equivalent.
Proof. The proof that first property of equivalence holds for Π(e, Gk ) and e ek
follows from the first property of equivalence for ek and Gk plus Lemma 5.
The proof that the second property of equivalence holds follows from the
17
first property of equivalence for Π(e, Gk ) and e ek , the second property of
equivalence for ek and Gk , plus Lemma 6.
A corollary of the previous lemma combined with Lemma 3 is that L(e $) =
L(Π(e, $)), proving that our transformation can yield a PEG that recognizes
the same language as any well-formed regular expression e just by using
an end-of-input marker, even if the language of e does not have the prefix
property.
It is interesting to see whether the axioms of system F1 (presented on page
7) are still valid if we transform both sides using Π with ε as the continuation
PEG, that is, if l is the left side of the equation and r is the right side then
PEG
PEG
Π(l, ε) xy ❀ y if and only if Π(r, ε) xy ❀ y. This is straightforward for
axioms 1, 2, 4, 6, 7, 8, and 9; in fact, it is easy to prove that these axioms
will be valid for any PEG, not just PEGs obtained from our transformation.
Applying Π to both sides of axiom 2, e1 (e2 e3 ) and (e1 e2 )e3 , makes them
identical; they both become Π(e1 , Π(e2 , Π(e3 , Gk ))). The same thing happens with axiom 5, (e1 | e2 )e3 = e1 e3 | e2 e3 ; the transformation of the left
side, Π((e1 | e2 )e3 , Gk ), becomes Π(e1 , Π(e3 , Gk )) | Π(e2, Π(e3 , Gk )) via the intermediate expression Π(e1 | e2 , Π(e3 , Gk )), while the transformation of the
right side, Π(e1 e3 | e2 e3 , Gk ), also becomes Π(e1 , Π(e3 , Gk )) | Π(e2 , Π(e3 , Gk )),
although via the intermediate expression Π(e1 e3 , Gk ) | Π(e2 e3 , Gk ).
The transformation of axiom 3, e1 | e2 = e2 | e1 , will not be valid; the
left side becomes the PEG Π(e1 , Gk ) | Π(e2, Gk ) and the right side becomes
the PEG Π(e2 , Gk ) | Π(e1 , Gk ), but ordered choice is not commutative in the
general case. One case where this choice is commutative is if the language
of e1 | e2 has the prefix property. We can use an argument analogous to the
argument of Lemma 3 to prove this, which is not surprising, as this lemma
together with Lemma 2 implies that this axiom should hold for expressions
with languages that have the prefix property.
Axiom 10, e∗ = ε | e∗e, needs to be rewritten as e∗ = e∗ e | ε or it is
trivially not valid, as the right side will always match just ε. Again, this is
not surprising, as ε | e∗e does not have the prefix property, and this is the
same behavior of regex implementations. Rewriting the axiom as e∗ = e∗ e | ε
makes it valid when we apply Π to both sides, as long as e∗ is well-formed.
The left side becomes the PEG A where A → Π(e, A) | ε, while the right
side becomes B | ε where B → Π(e, B) | Π(e, ε). If Π(e, A) fails it means that
Π(e, ε) would also fail, and so do Π(e, B), then B fails and the B | ε succeeds
by choice.2. Analogous reasoning holds for the other side, if B fails. If
18
Π(e, A) succeeds then Π(e, ε) matches a non-empty prefix, and A matches
the rest, and we can assume that B | ε matches this rest by induction on the
length of the matched string. We can use this to conclude that B | ε also
succeeds. Again, analogous reasoning holds for the converse.
The right side of axiom 11, (ε | e)∗, is not well-formed, and applying Π
to it would lead to a left-recursive PEG with no possible proof tree. We
still need to show that any regular expression can be made well-formed without changing its language. This is the topic of the next section, where we
give a transformation that rewrites non-well-formed repetitions so they are
well-formed with minimal changes to the structure of the original regular
expression. Applying this transformation to the right side of axiom 11 will
make it identical to the left side, making the axiom trivially valid.
3.1. Transformation of Repetitions e∗ where ε ∈ L(e)
A regular expression e that has a subexpression e∗i where ei can match the
empty string is not well-formed. As ei can succeed without consuming any
input one outcome of e∗i is to stay in the same place of the input indefinitely.
Regex libraries that rely on backtracking may enter an infinite loop with nonwell-formed expressions unless they take measures to avoid it, using ad-hoc
rules to detect and break the resulting infinite loops [17].
When e is not well-formed, the PEG we obtain through transformation Π
is not complete. A PEG that is not complete can make a PEG library enter
an infinite loop. To show an example on how a non-well-formed regular
expression leads to a PEG that is not complete, let us transform (a | ε)∗ b
using Π. Using ε as continuation yields the following PEG A:
A → aA | A | b
The PEG above is left recursive, so it is not complete. In fact, this PEG
does not have a proof tree for any input, so it is not equivalent to the regular
expression (a | ε)∗ b.
Transformation Π is not correct for non-well-formed regular expressions,
but we can make any non-well-formed regular expression well-formed by
rewriting repetitions e∗i where ε ∈ L(ei ) as e′i ∗ where ε 6∈ L(e′i ) and L(e′i ∗ ) =
L(e∗i ). The regular expression above would become a∗ | b, which Π transforms
into an equivalent complete PEG.
This section presents a transformation that mechanically performs this
rewriting. We use a pair of functions to rewrite an expression, fout and fin .
19
empty(ε) = true
empty(a) = false
empty(e∗ ) = empty(e)
empty(e1 e2 ) = empty(e1 ) ∧ empty(e2 )
empty(e1 | e2 ) = empty(e1 ) ∧ empty(e2 )
null(ε) = true
null(a) = false
null (e∗ ) = true
null (e1 e2 ) = null(e1 ) ∧ null (e2 )
null(e1 | e2 ) = null(e1 ) ∨ null (e2 )
Figure 4: Definition of predicates empty and null
Function fout recursively searches for a repetition that has ε in the language
of its subexpression, while fin rewrites the repetition’s subexpression so it
is well-formed, does not have ε in its language, and does not change the
language of the repetition. Both fin and fout use two auxiliary predicates,
empty and null , that respectively test if an expression is equal to ε (if its
language is the singleton set {ε}) and if an expression has ε in its language.
Figure 4 has inductive definitions for the empty and null predicates.
Function fout is simple: for the base expressions it is the identity, for the
composite expressions fout applies itself recursively to subexpressions unless
the expression is a repetition where the repetition’s subexpression matches ε.
In this case fout transforms the repetition to ε if the subexpression is equal
to ε (as ε∗ ≡ ε), or uses fin to rewrite the subexpression. Figure 5 has the
inductive definition of fout . It obeys the following lemma:
Lemma 8. If fin (ek ) is well-formed, ε 6∈ L(fin (ek )), and L(fin (ek )∗ ) = L(e∗k )
for any ek with ε ∈ L(ek ) and L(ek ) 6= ε then, for any e, fout (e) is well-formed
and L(e) = L(fout (e)).
Proof. By structural induction on e. Inductive cases follow directly from the
induction hypothesis, except for e∗ where ε ∈ L(e), where it follows from the
properties of fin .
20
fout (e) = e, if e = ε or e = a
fout (e1 e2 ) = fout (e1 ) fout (e2 )
fout (e1 | e2 ) = fout (e1 ) | fout (e2 )
fout (e)∗ if ¬null (e)
∗
ε
if empty(e)
fout (e ) =
∗
fin (e)
otherwise
Figure 5: Definition of Function fout
fin (e1 e2 ) = fin (e1 | e2 )
fin (e2 )
fout (e2 )
fin (e1 )
fout (e1 )
fin (e1 | e2 ) =
fout (e1 ) | fin (e2 )
f (e ) | fout (e2 )
in 1
fin (e1 ) | fin (e2 )
fin (e) if null (e)
fin (e∗ ) =
fout (e) otherwise
if empty(e1 ) and null (e2 )
if empty(e1 ) and ¬null (e2 )
if null (e1 ) and empty(e2 )
if ¬null (e1 ) and empty(e2 )
if ¬null (e1 ) and ¬empty(e2 )
if ¬empty(e1 ) and ¬null (e2 )
otherwise
Figure 6: Definition of Function fin (e), where ¬empty(e) and null(e)
Function fin does the heavy lifting of the rewriting, it is used when fout
finds an expression e∗ where ¬empty(e) and null (e). If its argument is a
repetition it throws away the repetition because it is superfluous. Then fin
applies fout or itself to the subexpression depending on whether it matches
ε or not. If the argument is a choice fin throws away one of the sides if its
equal to ε, as it is superfluous because of the repetition, and rewrites the
remaining side using fout or fin depending on whether it matches ε or not.
In case both sides are not equal to ε fin rewrites both. If the argument is a
concatenation fin rewrites it as a choice and applies itself to the choice.
Transforming a concatenation into a choice obviously is not a valid transformation in the general case, but it is safe in the context of fin ; fin is working
21
inside a repetition expression, and its argument has ε in its language, so we
can use an identity involving languages and the Kleene closure that says
(AB)∗ = (A ∪ B)∗ if ε ∈ A and ε ∈ B. Figure 6 has the inductive definition
of fin . It obeys the following lemma:
Lemma 9. If fout (ek ) is well-formed and L(fout (ek )) = L(ek ) for any ek then,
for any e with ε ∈ L(e) and L(e) 6= {ε}, ε ∈
/ L(fin (e)), L(e∗ ) = L(fin (e)∗ ),
and fin (e) is well-formed.
Proof. By structural induction on e. Most cases follow directly from the
induction hypothesis and the properties of fout . The subcases of choice
where the result is also a choice use the Kleene closure property (A ∪ B)∗ =
(A∗ ∪ B ∗ )∗ together with the induction hypothesis and the properties of fout .
Concatenation becomes to a choice using the property mentioned above this
lemma.
As an example, let us use fout and fin to rewrite the regular expression
(b c | a∗ (d | ε))∗ into a well-formed regular expression. We show the sequence
of steps below:
fout ((b c | a∗ (d | ε))∗) = (fin (b c | a∗ (d | ε)))∗ = (fout (b c) | fin (a∗ (d | ε)))∗
= (fout (b) fout (c) | (fin (a∗ ) | fin (d | ε)))∗
= (b c | (fout (a) | fout (d)))∗ = (b c | (a | d))∗
The idea is for rewriting to be automated, and transparent to the user of
regex libraries based on our transformation, unless the user wants to see how
their expression can be simplified. Notice that just the presence of ε inside a
repetition does not mean that a regular expression is not well-formed. The
(b c | a (d | ε))∗ expression looks very similar to the previous one, but is wellformed and left unmodified by fout .
4. Optimizing Search and Repetition
A common application of regexes is to search for parts of a subject that
match some pattern, but our formal model of regular expressions and PEGs
is anchored, as our matches must start on the first symbol of the subject
instead of starting anywhere. It is easy to build a PEG that will search for
another PEG (V, T, P, S), though, we just need to add a new non-terminal
S ′ as starting pattern, with S ′ → S | .S ′ as a new production, where . is
22
a shortcut for a regular expression that matches any terminal. If trying to
match S from the beginning of the subject fails, then the PEG skips the first
symbol and tries again on the second.
The search pattern works, but can be very inefficient if the PEG engine
always has to use backtracking to implement ordered choice, as advancing
to the correct starting position may involve a large amount of advancing
and then backtracking. A related problem occurs when converting regex
repetition into PEGs, as the PEG generated from the regular expression e∗1 e2
will greedily try to consume as much of the subject with e1 as possible, then
try e2 and backtrack each match of e1 until e2 succeeds or the whole pattern
fails. In the rest of this section we will show how we can use properties of the
expressions we are trying to search or match in conjunction with syntactic
predicates to reduce the amount of backtracking necessary in both searches
and repetition expressions.
4.1. Search
The search pattern for a PEG tries to match the PEG then advances
one position in the subject and tries again if the match fails. A simple way
to improve this is to advance several positions if possible, skipping starting
positions that have no chance of a successful match. If we know that a
successful match always consumes part of the subject and begins with a
symbol in a set F then we can skip a failing starting position with the pattern
![F ]., where [F ] is a characterset pattern that matches any symbol in the
set. We can skip a string of failing symbols with the pattern (![F ] .)∗ . The
new search expression for the PEG with starting pattern p can be written as
follows:
S → (![F ] .)∗ (p | .S)
The set F for a pattern p derived from a regular expression e is just the
RE
F IRST set of e, which has a simple definition in terms of the ❀ relation
below:
RE
FIRST (e) = {a ∈ T | ∃x, y e axy ❀ y, x, y ∈ T ∗ }
It is easy to prove that the two search expressions are equivalent by
induction on the height of the corresponding derivation trees. The tricky
case, where (![F ] .)∗ , just uses the definition of F IRST to build a tree of
successive applications of rule ord.2 until we can use the induction hypothesis
in its rightmost leaf.
23
4.2. Repetition
If two regular expressions e1 and e2 have disjoint F IRST sets then it
is safe to match e∗1 e2 using possessive repetition. This means that we can
transform e∗1 e2 into the PEG p∗1 p2 where p1 and p2 are the PEGs we get
from transforming e1 and e2 . Formally, we can define Π(e∗1 e2 , Gk ) when
F IRST (e1 ) ∩ F IRST (e2) = ∅ as follows, where we use G2 [ε] as the continuation for transforming e1 just to avoid collisions on the names of nonterminals:
Π(e∗1 e2 , Gk ) = (V1 , T, P1 , p∗1 p2 )
where (V1 , T, P1 , p1 ) = Π(e1 , G2 [ε])
and G2 = (V2 , T, P2 , p2 ) = Π(e2 , Gk )
The easiest way to prove the correctness of the new rule is by proving
the equivalence of the PEGs we get from Π(e∗1 e2 , Gk ) using the old and new
rule. This is a straightforward induction on the height of the proof trees for
these PEGs, using the fact that disjoint F IRST sets for e1 and e2 implies
disjointedness of their equivalent PEGs.
In the general case, where the F IRST sets of e1 and e2 are not disjoint,
we can still avoid some amount of backtracking on e∗1 e2 by being possessive
whenever there is no chance of e2 doing a successful match, as backtracking
to a point where e2 cannot match is useless. The idea is to use a predicated
repetition of p1 before doing the choice p1 A | p2 that guarantees that the
PEG will backtrack to a point where p2 matches, if possible. We can use the
F IRST set of e2 as an approximation to the set of possible matches of e2 ,
and the PEG for e∗1 e2 becomes A → (![F IRST (e2 )] p1 )∗ (p1 A | p2 ). The full
rule for Π(e∗1 e2 , Gk ) becomes as follows:
Π(e∗1 e2 , Gk ) = (V1 ∪ {A}, T, P1 ∪ {A → (![F IRST (e2 )] p1 )∗ (p1 A | p2 )}, A)
where (V1 , T, P1 , p1 ) = Π(e1 , G2 [ε])
with A ∈
/ V1 and G2 = (V2 , T, P2 , p2 ) = Π(e2 , Gk )
Again, the easiest way to prove that this new rule is correct is by proving
the equivalence of the PEGs obtained from the old and the new rule, by
induction on the height of the corresponding proof trees.
4.3. Combining Search and Repetition
We can further optimize the case where we want to search for the pattern
∗
∗
e1 e2 or e1 e∗1 e2 (we will use e+
1 as a shorthand for e1 e1 ), and all strings in L(e1 )
24
have length one. We can safely skip the prefix of the subject that matches a
possessive repetition of e1 before trying again, because if the pattern would
match from any of these positions then it would not have failed in the first
place. We can combine this with our first search optimization to yield the
following search pattern:
S → (![F ] .)∗ (p | p∗1 S)
In the pattern above, p is the starting expression of a PEG equivalent
to the regular expression we are searching and p1 is the starting expression
of a PEG equivalent to e1 with an empty continuation. Set F is still the
F IRST set of the whole regular expression. If the F IRST sets of e1 and e2
are disjoint we can further optimize our search by breaking up p and using
the following search expression to search for e∗1 e2 :
S → (![F ] .)∗ p∗1 (p2 | S)
The special case searching for e+
1 e2 just uses the search expression S →
∗ +
(![F ] .) p1 (p2 | S). Proofs that these optimizations are correct are straightforward, by proving that these search expressions are equivalent to S →
(![F ] .)∗ (p | .S) by induction on the height of the derivation trees.
5. Benchmarks
This section presents some benchmarks that compare a regex engine based
on an implementation of our transformation with the resulting PEGs executed with LPEG, a fast backtracking PEG engine [11]. We compare this
engine with PCRE, a backtracking regex engine that performs ad-hoc optimizations to cut the amount of backtracking needed [3], and with RE2, a
non-backtracking (automata-based) regex engine that nevertheless also incorporates ad-hoc optimizations [8].
We tested our search and repetition optimizations with a series of benchmarks that search for the first successful match of a regular expression inside
a large subject, the Project Gutenberg version of the King James Bible [18].
Our first benchmark searches for a single literal word in the subject, and
serves as a simple test of the search optimization. Table 1 shows the results.
We can see that the optimization is very effective, as LPEG optimizes the
repetition in the search pattern to a single instruction of its parsing machine
that scans the subject checking each character against a bitmap that encodes
25
Word
Geshurites
worshippeth
blotteth
sprang
RE2
1
3
3
7
PCRE
1
3
3
9
Unoptimized
12
25
33
47
Search
1
4
6
11
Line
19936
42140
60005
80000
Table 1: Time in milliseconds to search for a word
Words
Adam - Eve
Israel - Samaria
Jesus - John
Jesus - Judas
Jude - Jesus
Abraham - Jesus
RE2
1
2
2
2
3
5
PCRE
0
2
3
4
4
5
Unopt
0
32
73
81
94
96
Search
0
3
6
6
7
8
Repetition
0
3
6
6
7
8
Line
261
31144
76781
84614
98311
no match
Table 2: Time in milliseconds to search for two words in the same period
the character set. RE2 and PCRE use ad-hoc optimizations to find the string
and are still faster in some of the cases [8].
Our second benchmark searches for two literal words in the same period (separated by letters, spaces or commas), and we test the search and
repetition optimizations, but cannot apply the combined optimization of Section 4.3 because the expression does not have the necessary structure. Table 2
shows the results, and we separate the optimizations to show the contribution
of each one in the final result. The runtime is still dominated by having to
find where in the subject the match is, so optimizing the repetition inside the
pattern does not yield any gains. The pattern starts with a literal, so RE2
and PCRE are using ad-hoc optimizations to find where the match begins.
The third benchmark searches for a literal word that follows any other
word plus a single space (a regular expression [a−zA−Z]+ w, using character
class notation and for the empty space symbol). This pattern falls in
the case where the F IRST sets of the repeated pattern and the pattern
following the repetition are disjoint. We can apply the combined search
and repetition optimization for this pattern, and compare it with the basic
search and repetition optimizations. Table 3 shows the results. Now even
the unoptimized PEG defeats a backtracking regex matcher, but the DFAbased RE2 is much faster. The F IRST set of the pattern includes most
of terminals, and the search optimization is not effective. The biggest gain
26
Word
Geshurites
worshippeth
blotteth
sprang
RE2
5
7
10
12
PCRE
74
156
208
285
Unopt
57
121
159
222
Search
59
126
167
227
Rep
38
86
121
147
Combined
8
18
24
32
Line
19995
42140
60005
80000
Table 3: Time in milliseconds to search for a word following another
Words
Adam - Eve
Israel - Samaria
Jesus - John
Jesus - Judas
Jude - Jesus
Abraham - Jesus
RE2
2
6
12
13
15
15
PCRE
4
504
1134
1246
1446
1470
Unopt
6
752
1710
1884
2176
2214
Search
6
750
1718
1892
2188
2220
Rep
0
126
278
306
364
362
Comb
0
8
18
20
24
24
Line
261
31144
76781
84614
98311
no match
Table 4: Time in milliseconds to search for a period containing two words
comes from the combined optimization, as it lets the PEG skip large portions
of the subject in case of failure, yielding a result that is much closer to RE2.
The fourth and final benchmark extends the second benchmark by bracketing the pattern with a pattern that matches the part of the period that
precedes and follows the two words we are searching, yielding the pattern
[a − zA − Z, ]∗ w1 [a − zA − Z, ]∗ w2 [a − zA − Z, ]∗ . There is overlap in
the F IRST sets of [a − zA − Z, ], w1 , and w2 , so we need to use the more
general form of the repetition optimization. We can also apply the combined
optimization. As in the third benchmark, we compare this optimization with
the basic search and repetition optimizations. Table 4 shows the results. The
effect of the repetition optimization is bigger in this benchmark, but what
brings the performance close to a DFA-based regex matcher, and much better
than a backtracking regex matcher, is still the combined optimization.
Our benchmarks show that without optimizations our PEG-based engine
performs on par with PCRE on more complex patterns. The optimizations
bring it to a factor of 1 to 3 of the performance of RE2, a very efficient
and well-tuned regex implementation that cannot implement common regex
extensions due to its automata-based implementation approach.
27
Π(?ie1 , Gk ) = (V1 , T, P1 , p1 pk ), where (V1 , T, P1 , p1 ) = Π(e1 , Gk [ε])
∗
Π(e∗+
1 , Gk ) = Π(?ie , Gk )
Π(e∗?
1 , Gk ) = G , where G = (V1 , T, P1 ∪ {A → pk | p1 }, A) ,
(V1 , T, P1 , p1 ) = Π(e1 , (Vk ∪ {A}, T, Pk , A)), and A ∈
/ Vk
Π(?!e1 , Gk ) = (V1 , T, P1 , !p1 pk ), where (V1 , T, P1 , p1 ) = Π(e1 , Gk [ε])
Π(?= e1 , Gk ) = (V1 , T, P1 , !!p1 pk ), where (V1 , T, P1 , p1 ) = Π(e1 , Gk [ε])
Figure 7: Adapting Function Π to Deal with Regex Extensions
6. Transforming Regex Extensions
Regexes add several ad-hoc extensions to regular expressions. We can easily adapt transformation Π to deal with some of these extensions, and this
section shows how to use Π with independent expressions, possessive repetitions, lazy repetitions, and lookahead. An informal but broader discussion
of regex extensions in the context of translation to PEGs was published by
Oikawa et al. [19].
The regex ?ie1 is an independent expression (also known as atomic grouping). It matches independently of the expression that follows it, so a failure
when matching the expression that follows ?ie1 does not force a backtracking
regex matcher to backtrack to ?ie1 ’s alternative matches. This is the same
behavior as a PEG, so to transform ?ie1 we first transform it using an empty
continuation, then concatenate the result with the original continuation.
The regex e∗+
1 is a possessive repetition. It always matches as most as
possible of the input, even if this leads to a subsequent failure. It is the same
as ?ie∗ if the longest-match rule is used. The semantics of Π guarantees
longest match, so it uses this identity to transform e∗+
1 .
The regex e∗?
is
a
lazy
repetition.
It
always
matches
as little of the input
1
as necessary for the rest of the expression to match (shortest match). The
transformation of this regex is very similar to the transformation of e∗1 , we
just flip p1 and pk in the production of non-terminal A. Now the PEG tries
to match the rest of the expression first, and will only try another step of
the repetition if the rest fails.
The regex ?!e1 is a negative lookahead. The regex matcher tries to match
the subexpression; it it fails then the negative lookahead succeeds without
28
consuming any input, and if the subexpression succeeds the negative lookahead fails. Negative lookahead is also an independent expression. Transforming this regex is just a matter of using PEGs negative lookahead, which
works in the same way, on the result of transforming the subexpression as
an independent expression.
Finally, the regex ?= e1 is a positive lookahead, where the regex matcher
tries to match the subexpression and fails if the subexpression fails and succeeds if the subexpression succeeds, but does not consume any input. It
is also an independent expression. We transform a positive lookahead by
transforming the subexpression as an independent expression and then using
PEGs negative lookahead twice.
None of these extensions has been formalized before, as they depend on
the behavior of backtracking-based implementations of regexes instead of
the semantics of regular expressions. We decided to formalize them in terms
of their conversion to PEGs instead of trying to rework our semantics of
regular expressions to accommodate them, as these extensions map naturally
to concepts that are already part of the semantics of PEGs.
The well-formedness rewriting of Section 3.1 needs to accommodate the
new extensions. It is not possible to rewrite all non-well-formed expressions
with these extensions while keeping their behavior the same, as these extensions make it possible to write expressions that cannot give a meaningful
result, such as (?i(ε | a))∗ or (?= a(d | ε))∗. Other expressions can work with
some subjects and not work with others, such as (?i(a | ε | b))∗ or (?= a(a | ε))∗ .
Our approach will be to rewrite problematic expressions so they give the
same result for the subjects where they do not cause problems, but also give
a result for other subjects, that is, they will match a superset of the strings
that the original expression matches. For example, the four expressions above
will be respectively rewritten to (?ia)∗ , d∗ , (?i(a | b))∗ , and a∗ .
The empty predicate is true for ?!e1 and ?= e1 expressions, and empty(e1 )
for the other extensions. This is a conservative definition, as expressions such
as ?i(ε | e) can also be replaced by ε given the informal semantics of regexes.
The null predicate is true for all the extensions except ?ie1 , where it is
null (e1 ).
Figure 8 gives the definitions of fout and fin for the extensions. Function
fout just applies itself recursively for ?ie1 , ?!e1 and ?= e1 , but it needs to
rewrite the repetitions using fin if their bodies can match ε. Function fin
applies itself recursively to atomic groupings, keeping them atomic, but it
strips repetitions. A repetition being rewritten by fin is used directly inside
29
fout (?ie1 ) = ?ifout (e1 )
fout (e1 )∗+
ε
fout (e∗+
1 ) =
fin (e1 )∗+
fout (e1 )∗?
∗?
fout (e1 ) =
ε
fin (e1 )∗?
if ¬null (e1 )
if empty(e1 )
otherwise
if ¬null (e1 )
if empty(e1 )
otherwise
fout (?!e1 ) = ?!fout (e1 )
fout (?= e1 ) = ?= fout (e1 )
fin (?ie1 ) = ?ifin (e1 )
fin (e1 )
∗+
fin (e1 ) =
fout (e1 )
fin (e1 )
fin (e∗?
1 ) =
fout (e1 )
if null (e1 )
otherwise
if null (e1 )
otherwise
Figure 8: Definition of Functions fout and fin for regex extensions
another repetition, so it does not matter if it is possessive, lazy, or a regular
greedy repetition, it is the outer repetition that will govern how much of the
subject will be matched.
We do not need to define fin for negative and positive lookaheads, as
pathological uses of these expressions are eliminated by the fout case that
rewrites repetitions with an empty body and the fin cases that rewrite choices
with empty alternatives.
The extensions do not impact the optimizations of Section 4 if we provide
a way of computing a F IRST set for them, as the optimizations do not
depend on the structure of the subexpressions they use. We obviously cannot
∗?
∗
apply the repetition optimization on e∗+
1 e2 , e1 e2 , or ?i(e1 ) e2 , but applying
it on e∗1 e2 where extensions appear inside e1 or e2 is not a problem. The
repetition optimizations are turning repetitions into possessive repetitions
where possible, so not being able to optimize expressions such as the ones
above is not a loss, as they will already exhibit good backtracking behavior.
Figure 9 gives an inductive definition for the F IRST sets of extended
30
F IRST (ε) = ∅
F IRST (a) = {a}
F IRST (e1 )
if ¬null (e1 )
F IRST (e1 e2 ) =
F IRST (e1 ) ∪ F IRST (e2 ) if null (e1 )
F IRST (e1 | e2 )
F IRST (e∗)
F IRST (?ie)
F IRST (e∗+ )
=
=
=
=
F IRST (e1 ) ∪ F IRST (e2 )
F IRST (e)
F IRST (e)
F IRST (e)
F IRST (e∗? ) = F IRST (e)
F IRST (?!e) = ∅
F IRST (?= e) = ∅
Figure 9: Definition of F IRST sets for regexes
regexes. For completeness, we also give cases for the standard regexes. In
RE
our definition of F IRST sets in terms of relation ❀ the F IRST sets cannot
include ε, so expressions that never consume any prefix of the subject have
empty F IRST sets. The F IRST sets of atomic groupings are conservative,
as they may be a proper superset of the first characters that the expression
actually consumes; for example, F IRST (?i(ε | a)) is {a} instead of the more
precise ∅.
7. Conclusion
We presented a new formalization of regular expressions that uses natural
semantics and a transformation Π that converts a given regular expression
into an equivalent PEG, that is, a PEG that matches the same strings that
the regular expression matches in a Perl-compatible regex implementation.
If the regular expression’s language has the prefix property, easily guaranteed by using an end-of-input marker, the transformation yields a PEG that
recognizes the same language as the regular expression.
We also have shown how our transformation can be easily adapted to
accommodate several ad-hoc extensions used by regex libraries: independent expressions, possessive and lazy repetition, and lookahead. Our trans31
formation gives a precise semantics to what were informal extensions with
behavior specified in terms of how backtracking-based regex matchers are
implemented.
We show that, for some classes of regular expressions, we produce PEGs
that perform better by reasoning about how the PEG’s limited backtracking
and syntactical predicates work to control the amount of backtracking that
a PEG will perform. The same reasoning that we apply for large classes of
expressions can be applied to specific ones to yield bigger performance gains
where necessary, although our benchmarks show that simple optimizations
are enough to perform close to optimized regex matchers, while having a
much simpler implementation: both regex engines we used have over ten
times the amount of code of the PEG engine.
Another approach to establish the correspondence between regular expressions and PEGs was suggested by Ierusalimschy [11]. In this approach we
convert Deterministic Finite Automata (DFA) into right-linear LL(1) grammars. Medeiros [14] proves that an LL(1) grammar has the same language
when interpreted as a CFG and as a PEG. But this approach cannot be used
with regex extensions, as they cannot be expressed by a DFA.
The transformation Π is a formalization of the continuation-based conversion presented by Oikawa et al. [19]. That work only presents an informal
discussion of the correctness of the conversion, while we proved our transformation correct with regards to the semantics of regular expressions and
PEGs.
We can also benefit from the LPEG parsing machine [12, 11], a virtual
machine for executing PEGs. We can use the cost model of the parsing
machine instructions to estimate how efficient a given regular expression or
regex is. The parsing machine has a simple architecture with just nine basic
instructions and four registers, and implementations of our transformation
coupled with implementations of the parsing machine can be the basis for
simpler implementations of regex libraries.
References
1. Wall,
L..
Apocalypse 5: Pattern matching.
2002.
http://www.perl6.org/archive/doc/design/apo/A05.html.
URL
2. Conway,
D., Randal,
A., Michaud,
P., Wall,
Lenz,
M..
Synopsis 5: Regexes and rules.
2002.
http://feather.perl6.nl/syn/S05.html.
L.,
URL
32
3. Hazel, P.. PCRE - Perl Compatible Regular Expressions. 2011. URL
http://www.pcre.org/.
4. Aho, A.V.. Algorithms for Finding Patterns in Strings. Cambridge,
MA, USA: MIT Press. ISBN 0-444-88071-2; 1990, p. 255–300.
5. Abigail. Reduction of 3-CNF-SAT to Perl regular expression matching.
http://perl.plover.com/NPC/NPC-3SAT.html; 2001.
6. Friedl, J.. Mastering Regular Expressions. O’Reilly Media, Inc.; 2006.
ISBN 0596528124.
7. Fowler, G.. An interpretation of the POSIX regex standard. Tech. Rep.;
Information and Software Systems Research, AT&T Labs; New Jersey,
USA; 2003.
8. Cox, R.. Regular expression matching in the wild.
http://swtch.com/~ rsc/regexp/regexp3.html.
2010.
URL
9. Reps,
T..
“Maximal-munch” tokenization in linear time.
ACM Transactions on Programming Languages and Systems 1998;
20(2):259–273.
doi:\bibinfo{doi}{10.1145/276393.276394}.
URL
http://doi.acm.org/10.1145/276393.276394.
10. Ford, B.. Parsing Expression Grammars: A recognition-based syntactic
foundation. In: POPL ’04: Proceedings of the 31st ACM SIGPLANSIGACT Symposium on Principles of Programming Languages. New
York, USA: ACM. ISBN 1-58113-729-X; 2004, p. 111–122.
11. Ierusalimschy, R.. A text pattern-matching tool based on Parsing Expression Grammars. Software - Practice & Experience 2009;39(3):221–
258.
12. Medeiros, S., Ierusalimschy, R.. A parsing machine for PEGs. In: DLS
’08: Proceedings of the 3rd Dynamic Languages Symposium. New York,
USA: ACM. ISBN 978-1-60558-270-2; 2008, p. 1–12.
13. Kahn, G.. Natural semantics. In: STACS ’87: Proceedings of the 4th
Symposium on Theoretical Aspects of Computer Science. London, UK:
Springer-Verlag. ISBN 3-540-17219-X; 1987, p. 22–39.
33
14. Medeiros, S.. Correspondência entre PEGs e Classes de Gramáticas
Livres de Contexto. Ph.D. thesis; PUC-Rio; 2010.
15. Hopcroft, J., Ullman, J.. Introduction to Automata Theory, Languages,
and Computation. Boston, USA: Addison-Wesley Longman Publishing
Co., Inc.; 1979. ISBN 0321462254.
16. Salomaa,
A..
Two complete axiom systems for the algebra of regular events.
Journal of the ACM 1966;13(1):158–
169.
doi:\bibinfo{doi}{10.1145/321312.321326}.
URL
http://doi.acm.org/10.1145/321312.321326.
17. The Perl Foundation. Core documentation for Perl 5 version 14.1 –
regular expressions. http://perldoc.perl.org/perlre.html; 2005.
18. Anonymous. The Bible, Both Testaments, King James Version; vol. 10.
P.O. Box 2782, Champaign, IL 61825-2782, USA: Project Gutenberg;
1989.
19. Oikawa, M., Ierusalimschy, R., Moura, A.. Converting regexes to
Parsing Expression Grammars. In: SBLP ’10: Proceedings of the 14th
Brazilian Symposium on Programming Languages. 2010, .
34
| 6cs.PL
|
KAZHDAN SETS IN GROUPS AND EQUIDISTRIBUTION
PROPERTIES
by
arXiv:1601.04289v4 [math.GR] 3 Jun 2017
Catalin Badea & Sophie Grivaux
Abstract. — Using functional and harmonic analysis methods, we study Kazhdan sets in
topological groups which do not necessarily have Property (T). We provide a new criterion for
a generating subset Q of a group G to be a Kazhdan set; it relies on the existence of a positive
number ε such that every unitary representation of G with a pQ, εq-invariant vector has a
finite dimensional subrepresentation. Using this result, we give an equidistribution criterion
for a generating subset of G to be a Kazhdan set. In the case where G “ Z, this shows that
if pnk qkě1 is a sequence of integers such that pe2iπθnk qkě1 is uniformly distributed in the unit
circle for all real numbers θ except at most countably many, then tnk ; k ě 1u is a Kazhdan
set in Z as soon as it generates Z. This answers a question of Y. Shalom from [B. Bekka,
P. de la Harpe, A. Valette, Kazhdan’s property (T), Cambridge Univ. Press, 2008]. We
also obtain characterizations of Kazhdan sets in second countable locally compact abelian
groups, in the Heisenberg groups and in the group Aff` pRq. This answers in particular a
question from [B. Bekka, P. de la Harpe, A. Valette, Kazhdan’s property (T), op. cit.].
1. Introduction
A unitary representation of a topological group G on a Hilbert space H is a group
morphism from G into the group U pHq of all unitary operators on H which is strongly
/ πpgqx is continuous from G into H for all
continuous, i. e. such that the map g ✤
vectors x P H. As all the representations we consider in this paper are unitary, we will
often drop the word “unitary” and speak simply of representations of a group G on a
Hilbert space H. In this paper the Hilbert spaces will always be supposed to be complex,
and endowed with an inner product x ¨ , ¨ y which is linear in the first variable and antilinear
in the second variable.
2000 Mathematics Subject Classification. — 22D10, 22D40, 37A15, 11K069, 43A07, 46M05.
Key words and phrases. — Kazhdan sets, topological groups, Property (T), equidistributed sequences
in groups, tensor products of unitary representations, weakly mixing representations, abstract Wiener
theorem, Heisenberg groups.
We would like to thank an anonymous referee for very useful remarks and comments; the present version
has been completely rewritten after receiving his/her report. This work was supported in part by the
Labex CEMPI (ANR-11-LABX-0007-01) and by the EU IRSES grant AOS (PIRSES-GA-2012-318910).
2
CATALIN BADEA & SOPHIE GRIVAUX
Definition 1.1. — Let Q be a subset of a topological group G, ε a positive real number,
and π a unitary representation of G on a Hilbert space H. A vector x P H is said to be
pQ, εq-invariant for π if
sup ||πpgqx ´ x|| ă ε||x||.
g PQ
A pQ, εq-invariant vector for π is in particular non-zero. A G-invariant vector for π is a
vector x P H such that πpgqx “ x for all g P G.
The notions of Kazhdan sets and Kazhdan pairs will be fundamental in our work.
Definition 1.2. — A subset Q of a topological group G is a Kazhdan set in G if there
exists ε ą 0 such that the following property holds true: any unitary representation π of
G on a complex Hilbert space H with a pQ, εq-invariant vector has a non-zero G-invariant
vector. In this case, the pair pQ, εq is Kazhdan pair, and ε is a Kazhdan constant for Q.
A group G has Property (T), or is a Kazhdan group, if it admits a compact Kazhdan set.
Property (T) is a rigidity property of topological groups which has been introduced by
Kazhdan in [26] for locally compact groups, and which has spectacular applications to
many fields. For instance, the groups SLn pRq and SLn pZq have Property (T) if and only
if n ě 3. We refer the reader to the monograph [6] by Bekka, de la Harpe, and Valette
for a comprehensive presentation of Kazhdan’s Property (T) and its applications (see also
[19]).
The aim of this paper is to identify and study Kazhdan sets in topological groups. For
discrete groups with Property (T) the Kazhdan sets are known. Recall first the following
definition.
Definition 1.3. — If Q is a subset of a group G, we denote by xQy the smallest subgroup
of G containing Q, i. e. the set of all elements of the form g1˘1 . . . gn˘1 , where n ě 1 and
g1 , . . . , gn belong to Q. We say that Q generates G, or is generating in G, if xQy “ G.
Locally compact groups with Property (T) are compactly generated. In particular, discrete groups with Property (T) are finitely generated and it is known (see [6, Prop. 1.3.2])
that the Kazhdan subsets of a discrete group with Property (T) are exactly the generating
subsets of the group. More generally [6, Prop. 1.3.2], a generating set of a locally compact
group which has Property (T) is a Kazhdan set and, conversely, a Kazhdan set which has
non-empty interior is necessarily a generating set.
For groups without Property (T) the results about Kazhdan
? sets and Kazhdan pairs
are very sparse. It is known (see [6, Prop. 1.1.5]) that pG, 2q is a Kazhdan pair for
every topological group G, so G is always a (“large”) Kazhdan subset of itself. The main
motivations for the present paper are two questions from [6, Sec. 7.12]. The first one is
due to Y. Shalom:
Question 1.4. — [6, Sec. 7.12] “The question of knowing if a subset Q of Z is a Kazhdan
set is possibly related to the equidistribution of the sequence pe2iπnθ qn PQ for θ irrational,
in the sense of Weyl.”
We refer the reader to the classical book [28] by Kuipers and Niederreiter for more
information about equidistributed (sometimes called uniformly distributed) sequences.
Recall that the Weyl Criterion ([28, Th. 2.1]) states that if pxk qkě1 is a sequence of real
ř
2iπhxk tends to 0 as N
numbers, pe2iπxk qkě1 is equidistributed in T if and only if N1 N
k“1 e
tends to infinity for every non-zero integer h. Hence if pnk qkě1 is a sequence of elements
KAZHDAN SETS AND EQUIDISTRIBUTION PROPERTIES
3
řN
of Z, pe2iπnk θ qkě1 is equidistributed in T for every θ P RzQ if and only if N1 k“1 e2iπnk θ
tends to 0 as N tends to infinity for every θ P RzQ. If χθ denotes, for every θ P R, the
ř
character on Z associated to θ, this means that N1 N
k“1 χθ pnk q tends to 0 as N tends to
infinity for every θ P RzQ.
The first remark about Question 1.4 is that it concerns Kazhdan sets and equidistributed sequences; notice that a rearrangement of the terms of a sequence can destroy
its equidistribution properties. It is known [28, p. 135] that given a sequence of elements
of the unit circle T, there exists a certain rearrangement of the terms which is is equidistributed if and only if the original sequence is dense in T. The second remark is that,
as mentioned before, Kazhdan sets of Z are necessarily generating, while there are nongenerating subsets Q of Z, like Q “ pZ with p ě 2, for which the sequences pe2iπpkθ qk PZ
are equidistributed for all irrational θ’s. So Question 1.4 may be rephrased as follows:
Question 1.5. — (a) Let Q be a Kazhdan subset of Z. Does a certain rearrangement
pnk qk ě1 of the elements of Q exist such that pe2iπnk θ qk ě1 is equidistributed in T for every
θ P RzQ? Equivalently, is the sequence pe2iπnθ qn PQ dense in T for every θ P RzQ?
(b) Let Q “ tnk ; k ě 1u be a generating subset of Z. Suppose that the sequence
pe2iπnk θ qk ě1 is equidistributed in T for every θ P RzQ. Is Q a Kazhdan set in Z?
We will prove in this paper that Question 1.5 (a) has a negative answer, a counterexample being provided by the set Q “ t2k ` k ; k ě 1u (see Example 6.4). On the other
hand, one of the aims of this paper is to show that Question 1.5 (b) has a positive answer. Actually, we will consider Question 1.5 (b) in the more general framework of Moore
groups, and answer it in the affirmative (Theorem 2.1).
The second question of [6, Sec. 7.12] runs as follows:
Question 1.6. — [6, Sec. 7.12] “More generally, what are the Kazhdan subsets of Zk ,
Rk , the Heisenberg group, or other infinite amenable groups?”
We shall answer Question 1.6 in Section 6 by giving a complete description of Kazhdan
sets in many classic groups which do not have Property (T), including the groups Zk and
Rk , k ě 1, the Heisenberg groups of all dimensions, and the group Aff` pRq of orientationpreserving affine homeomorphisms of R.
2. Main results
Let us now describe our main results in more detail.
2.1. Equidistributed sets in Moore groups. — In order to state Question 1.5 (b) for
more general groups, we first need to define equidistributed sequences. There are several
possible ways of doing this. If pgk qkě1 is a sequence of elements in a locally compact
group G, uniform distribution of pgk qkě1 in any of these senses requires a certain form of
convergence, as N tends to infinity, of the means
(2.1)
N
1 ÿ
πpgk q
N k“1
to the orthogonal projection Pπ on the subspace of invariant vectors for π, for a certain
class of unitary representations π of G. Veech [37], [38] calls pgk qkě1 uniformly distributed
in G if the convergence of the means (2.1) holds in the weak operator topology for all
4
CATALIN BADEA & SOPHIE GRIVAUX
unitary representations of G (or, equivalently, for all irreducible unitary representations
of G, provided G is supposed to be second countable). Unitary uniform distribution in
the sense of Losert and Rindler [29], [17] requires the convergence in the strong operator
topology of the means (2.1) for all irreducible unitary representations of the group, while
Hartman uniform distribution only requires convergence in the strong operator topology
for all finite dimensional unitary representations.
In this paper we deal with the following natural extension to general locally compact
groups G of the equidistribution condition of Question 1.5 (b): if pgk qkě1 is a sequence
of elements of G, we require the sequence of means (2.1) to converge to 0 in the weak
topology for all finite dimensional irreducible unitary representations of G except those
belonging at most countably many equivalence classes of irreducible representations. In
the case of the group Z, sequences pnk qkě1 of integers such that pe2iπθnk qkě1 is uniformly
distributed in T for all θ P R except countably many are said to be of first kind (see for
instance [21]). The class of groups we will consider in relation to Question 1.5 (b) is the
class of second countable Moore groups. Recall that G is said to be a Moore group if all
irreducible representations of G are finite dimensional. Locally compact Moore groups are
completely described in [31]: a Lie group is a Moore group if and only if it has a closed
subgroup H such that H modulo its center is compact, and a locally compact group is a
Moore group if and only if it is a projective limit of Lie groups which are Moore groups. See
also the survey [33] for more information concerning the links between various properties
of topological groups, among them the property of being a Moore group. Of course all
locally compact abelian groups are Moore groups.
Here is the first main result of this paper.
Theorem 2.1. — Let G be a second countable locally compact Moore group. Let pgk qkě1
be a sequence of elements of G. Suppose that pgk qkě1 satisfies the following equidistribution
assumption:
for all (finite dimensional) irreducible unitary representations π of G on a
Hilbert space H, except those belonging to at most countably many equivalence
classes,
(2.2)
N
1 ÿ
/ 0 for every x, y P H.
xπpgk qx, yy
/ `8
N k“1
N
– If Q “ tgk ; k ě 1u generates G (in which case G has to be countable), then Q is a
Kazhdan set in G.
– If Q is not assumed to generate G, Q becomes a Kazhdan set when one adds to it
a suitable “small” perturbation.
More precisely, if pWn qně1 is an increasing sequence of
Ť
subsets of G such that ně1 Wn “ G, there exists n ě 1 such that Wn Y Q is a Kazhdan
set in G.
The equidistribution property (2.2) of the sequence pgk qkě1 takes a more familiar form
when the group G is supposed to be abelian: it is equivalent to requiring that condition
(2.3) below holds true for all characters χ of the group except possibly countably many.
KAZHDAN SETS AND EQUIDISTRIBUTION PROPERTIES
5
Theorem 2.2. — Let G be a locally compact abelian group, and let pgk qkě1 be a sequence
of elements of G. Suppose that
(2.3)
N
1 ÿ
χpgk q
N k“1
/0
N
/
`8
for all characters χ on G, except at most countably many. If Q “ tgk ; k ě 1u generates
G, then Q is a Kazhdan set in G. If Q is not assumed
to generate G, and if pWn qně1 is
Ť
an increasing sequence of subsets of G such that ně1 Wn “ G, then there exists n ě 1
such that Wn Y Q is a Kazhdan set in G.
Theorem 2.2 can thus be seen as a particular case of Theorem 2.1, except for the fact
that there is no need to suppose that the group is second countable when it is known to
be abelian. The case G “ Z provides a positive answer to Question 1.5 (b) above.
2.2. Kazhdan sets and finite dimensional subrepresentations. — The proof of
Theorem 2.2 relies on Theorem 2.3 below, which gives a new condition for a “small perturbation” of a subset Q of a group G to be a Kazhdan set in G. Theorem 2.3 constitutes the
core of the paper, and has, besides the proofs of Theorems 2.1 and 2.2, several interesting
applications which we will present in Sections 5 and 6.
Theorem 2.3. — Let G be a topological group, and let pWn qně1 be an increasing
Ť sequence
of subsets of G such that W1 is a neighborhood of the unit element e of G and ně1 Wn “
G. Let Q be a subset of G satisfying the following assumption:
there exists a positive constant ε such that every unitary representation
(*)
π of G on a Hilbert space H admitting a pQ, εq-invariant vector has a
finite dimensional subrepresentation.
Then there exists an integer n ě 1 such that Qn “ Wn Y Q is a Kazhdan set in G.
If the group G is locally compact, the same
Ť statement holds true for any increasing
sequence pWn qně1 of subsets of G such that ně1 Wn “ G.
The condition that W1 be a neighborhood of e, which appears in the first part of the
statement of Theorem 2.3, will be used in the proof in order to ensure the strong continuity
of some infinite tensor product representations (see Proposition A.2). When G is locally
compact, this assumption is no longer necessary (see Proposition A.1).
We stress that Theorem 2.3 is valid for all topological groups. We will apply it mainly
to groups which do not have Property (T) and to subsets of such groups which are not
relatively compact, a notable exception being the proof of Theorem 5.1, where we retrieve
a characterization of Property (T) for σ-compact locally compact groups due to Bekka
and Valette [5], see also [6, Th. 2.12.9]. The original proof of this result relies on the
Delorme-Guichardet theorem that such a group has Property (T) if and only if it has
property (FH). See Section 5 for more details.
Theorem 2.3 admits a simpler formulation if we build the sequence pWn qně1 starting
from a set which generates the group:
Corollary 2.4. — Let G be a topological group. Let Q0 be a subset of G which generates
G and let Q be a subset of G. Suppose either that Q0 has non-empty interior, or that G
is a locally compact group. If Q satisfies assumption (*) of Theorem 2.3, then Q0 Y Q is
a Kazhdan set in G.
6
CATALIN BADEA & SOPHIE GRIVAUX
One of the main consequences of Corollary 2.4 is Theorem 2.5 below, which shows in
particular that property (*) of Theorem 2.3 characterizes Kazhdan sets among generating
sets (and which have non-empty interior – this assumption has to be added if the group
is not supposed to be locally compact).
Theorem 2.5. — Let G be a topological group and let Q be a subset of G which generates
G. Suppose either that Q has non-empty interior or that G is locally compact. Then the
following assertions are equivalent:
(a) Q is a Kazhdan set in G;
(b) there exists a constant δ P p0, 1q such that every unitary representation π of G on a
Hilbert space H admitting a vector x P H such that inf gPQ |xπpgqx, xy| ą δ}x}2 has a
finite dimensional subrepresentation;
(c) there exists a constant ε ą 0 such that every unitary representation π of G on a Hilbert
space H admitting a pQ, εq-invariant vector has a finite dimensional subrepresentation.
The assumption that Q generates G cannot be dispensed with in Theorem 2.5: Q “ 2Z
is a subset of Z which satisfies property (c), but Q is clearly not a Kazhdan set in Z.
Condition (b) in Theorem 2.5 is easily seen to be equivalent to condition (c), which is
nothing else than assumption (*) of Theorem 2.3. Its interest will become clearer in
Section 6 below, where it will be used to obtain a characterization of Kazhdan sets in
second countable locally compact abelian groups (Theorem 6.1). In the case of the group
Z, the characterization we obtain (Theorem 6.3) involves a classic class of sets in harmonic
analysis, called Kaufman sets. We give in Section 6 several examples of “small” Kazhdan
sets in Z, describe Kazhdan sets in the Heisenberg groups Hn , n ě 1 (Theorem 6.11), and
also in the group Aff` pRq (Theorem 6.14). These results provide an answer to Question
1.6.
The paper also contains an appendix which reviews some constructions of infinite tensor
product representations on Hilbert spaces, used in the proof of Theorem 2.3.
3. Mixing properties for unitary representations and an abstract version of
the Wiener theorem
3.1. Ergodic and mixing properties for unitary representations. — We first
recall in this section some definitions and results concerning the structure of unitary representations of a topological group G. They can be found for instance in the book [27],
the notes [34], and the paper [8] by Bergelson and Rosenblatt.
Recall that the class WAPpGq of weakly almost periodic functions on G is defined as
follows: if ℓ 8 pGq denotes the space of bounded functions on G, f P ℓ 8 pGq belongs to
WAPpGq if the weak closure in ℓ 8 pGq of the set tf ps´1 q ; s P Gu is weakly compact. For
/ f ps´1 tq on G. By comparison, recall
each s P G, f ps´1 q denotes the function t ✤
8
that f P ℓ pGq is an almost periodic function on G, written f P APpGq, if the norm
closure in ℓ 8 pGq of tf ps´1 q ; s P Gu is compact. If π is a unitary representation of G on
a Hilbert space H, the functions
ˇ
ˇ
ˇ
ˇ2
xπpqx, yy, ˇxπpqx, yyˇ, and ˇxπpqx, yyˇ ,
where x and y are any vectors of H, belong to WAPpGq. For more on weakly almost
periodic functions on a group, see for instance [10] or [16, Ch. 1, Sec. 9]. The interest of
KAZHDAN SETS AND EQUIDISTRIBUTION PROPERTIES
7
the class of weakly almost periodic functions on G in our context is that there exists on
WAPpGq a unique G-invariant mean m. It satisfies
mpf ps´1 qq “ mpf p s´1 qq “ mpf q
for every f P WAPpGq and every s P G. The abstract ergodic theorem then states that
if π is a unitary representation of G on H, mpxπpqx, yyq “ xPπ x, yy for every vectors
x, y P H, where Pπ denotes the projection of H onto the space Eπ “ tx P H ; πpgqx “
x for every g P Gu of G-invariant vectors for π. The representation π is ergodic (i. e.
admits no non-zero G-invariant vector) if and only if mpxπpqx, yyq “ 0 for every x, y P H.
Following [8], let us now recall that the representation π is said to be weakly mixing if
mp|xπpqx, xy|q “ 0 for every x P H, or, equivalently, mp|xπpqx, xy|2 q “ 0 for every x P H.
Then mp|xπpqx, yy|q “ mp|xπpqx, yy|2 q “ 0 for every x, y P H.
We will need the following characterization of weakly mixing representations.
Proposition 3.1. — Let π be a unitary representation of G on a Hilbert space H. The
following assertions are equivalent:
(1) π is weakly mixing;
(2) π admits no finite dimensional subrepresentation;
(3) π b π has no non-zero G-invariant vector.
Here π is the conjugate representation of π. The representation π b π is equivalent to
a representation of G on the space HSpHq of Hilbert-Schmidt operators on H, which is
often more convenient to work with. Recall that HSpHq is a Hilbert space when endowed
with the scalar product defined by the formula xA, By “ trpB ˚Aq for every A, B P HSpHq.
The space H b H, where H is the conjugate of H, is identified to HSpHq by associating
to each elementary tensor product x b y of H b H the rank-one operator x , yy x on H.
/ HSpHq extends into a unitary isomorphism, and we have for
This map Θ : H b H
every g P G and every T P HSpHq
Θ π b πpgq Θ´1 pT q “ πpgq T πpg ´1 q.
We will, when needed, identify π b π with this equivalent representation, and use it in
particular in Section 3.3 to obtain a concrete description of the space Eπbπ of G-invariant
vectors for π b π, which is identified to the subspace of HSpHq
Eπ “ tT P HSpHq ; πpgq T “ T πpgq for every g P Gu¨
3.2. Compact unitary representations. — A companion to the property of weak
mixing for unitary representation is that of compactness: given a unitary representation
π of G on H a vector x P H is compact for π if the norm closure of the set tπpgqx ; g P Gu
is compact in H. The representation π itself is said to be compact if every vector of H
is compact for π. Compact representations decompose as direct sums of irreducible finite
dimensional representations. The general structural result for unitary representations is
given by the following result.
Proposition 3.2. — A unitary representation π of G on a Hilbert space H decomposes
as a direct sum of a weakly mixing representation and a compact representation:
K
H “ Hw ‘ Hc ,
CATALIN BADEA & SOPHIE GRIVAUX
8
where Hw and Hc are both G-invariant closed subspaces of H, πw “ π|Hw is weakly mixing
and πc “ π|Hc is compact. Hence π decomposes as a direct sum of a weakly mixing
representation and finite dimensional irreducible subrepresentations.
See [34, Ch. 1], [6, Appendix M], [8] or [12] (in the amenable case) for detailed proofs
of these results.
Now let π be a compact representation of G on a Hilbert space H, decomposed as
a direct sum of irreducible finite dimensional representations of G. We sort out these
representations by equivalence classes, and index the distinct equivalence classes by an
index j belonging to a set J, which may be finite or infinite (and which is countable if
H is separable). For every j P J, we index by i P Ij all the representations appearing
in the decomposition of π which are in the j-th equivalence class. More precisely, we can
decompose H and π as
˘
`
˘
`
and π “ ‘ ‘ πi, j
H “ ‘ ‘ Hi, j
j PJ i PIj
j PJ i PIj
respectively, where the following holds true:
– for every j P J, the spaces Hi, j , i P Ij , are equal. We denote by Kj this common
space, and by dj its dimension (which is finite). We also write
r j “ ‘ Hi, j ,
H
i PIj
so that
rj ;
H“ ‘ H
j PJ
– for every j P J, there exists an irreducible representation πj of G on Kj such that
πi, j is equivalent to πj for every i P Ij ;
– if j, j 1 belong to J and j ‰ j 1 , πj and πj 1 are not equivalent.
Without loss of generality, we will suppose that πi, j “ πj for every i P Ij . However, we
will keep the notation Hi, j for the various orthogonal copies of the space Kj which appear
in the decomposition of H, as discarding this notation may be misleading in some of the
proofs presented below.
Let A P BpHq. We write A in block-matrix form with respect to the decompositions
˘
`
rj
and H “ ‘ H
H “ ‘ ‘ Hi, j
as
j PJ
j PJ i PIj
˘
`
A “ Au, v k, l P J, u P I
k , v P Il
and
`
˘
rk, l
A“ A
k, l PJ
respectively¨
pjq
For every j P J and every u, v P Ij , we denote by iu, v the identity operator from Hu, j into
Hv, j .
3.3. A formula for the projection Pπ of HSpHq on Eπ .— We now give an explicit
formula for the projection Pπ A of a Hilbert-Schmidt operator A P HSpHq on the following
closed subspace of HSpHq:
Eπ “ tT P HSpHq ; πpgq T “ T πpgq for every g P Gu.
We also compute the norm of Pπ A.
Proposition
3.3.
˘ — Let π be a compact representation of G on H, written in the form
`
π “ ‘ ‘ πj as discussed in Section 3.2 above. For every operator A P HSpHq, we
j PJ i PIj
KAZHDAN SETS AND EQUIDISTRIBUTION PROPERTIES
have
Pπ A “
ÿ 1
d
j PJ j
ÿ
u, v P Ij
`
˘ pjq
tr Au, v iu,
v
||Pπ A||2 “
and
ÿ 1
d
j PJ j
9
ÿ ˇ `
ˇtr A
u, v
u, v PIj
The proof of Proposition 3.3 relies on the following straightforward lemma:
˘ˇ2
ˇ .
Lemma 3.4. — The space Eπ consists of the operators T P HSpHq such that
– for every k, l P J with k ‰ l, Trk, l “ 0;
– for every k P J and every u, v P Ik , there exists a complex number λu, v such that
`
pkq
pkq ˘
Tu, v “ λu, v iu, v . Thus Trk, k “ λu, v iu, v u, v PI .
k
Proof of Lemma 3.4. — Let T P Eπ . For every k, l P J, u P Ik and v P Il , πk pgq Tu, v “
Tu, v πl pgq for every g P G. Thus the operator Tu, v intertwines the two representations πk
and πl . If Tu, v is non-zero, it follows from Schur’s Lemma that Tu, v is an isomorphism.
The representations πk and πl are thus isomorphically (and hence unitarily) equivalent.
Since πk and πl are not equivalent for k ‰ l, it follows that Tu, v “ 0 in this case. If now
pkq
k “ l, Schur’s Lemma again implies that Tu, v “ λu, v iu, v for some scalar λu, v . Thus any
operator T P Eπ satisfies the two conditions of the lemma. The converse is obvious.
The proof of Proposition 3.3 is now easy.
Proof of Proposition 3.3. — Consider, for every j P J and u, v P Ij , the one-dimensional
pjq
pjq
subspace Eu, v of HSpHq spanned by the operator iu, v . These subspaces are pairwise
orthogonal in HSpHq, and by Lemma 3.4 we have
`
˘
pjq
‘ Eu,
Eπ “ ‘
v .
j PJ u, v PIj
Hence, for every A P HSpHq,
ÿ ÿ A
A,
Pπ A “
j PJ u, v PIj
pjq
iu, v
pjq
||iu, v ||HS
E
pjq
iu, v
pjq
||iu, v ||HS
which gives the two formulas we were looking for.
“
ÿ 1
dj
j PJ
ÿ
u, v PIj
`
˘ pjq
tr Au, v iu,
v,
˘
`
Corollary 3.5. — Let π “ ‘ ‘ πj be a compact representation of G on H. Let
j PJ i PIj
˘
˘
`
`
x “ ‘ ‘ xi, j and y “ ‘ ‘ yi, j be two vectors of H, and let A P HSpHq be the
j PJ i PIj
j PJ i PIj
rank-one operator x , yy x. Then
Pπ A “
ÿ 1
dj
j PJ
ÿ
u, v PIj
pjq
xxu, j , yv, j y iu,
v.
˘
`
Proof. — For every j P J and u, v P Ij , Au, v “ x , yv, j y xu, j , so that tr Au, v “
xxu, j , yv, j y. The result then follows from Proposition 3.3.
CATALIN BADEA & SOPHIE GRIVAUX
10
3.4. An abstract version of the Wiener Theorem.— As recalled in Section 3.1, Eπ
is the space of G-invariant vectors for the representation π b π on HSpHq, where for every
x, y P H, x b y is identified with the rank-one operator x , yy x. For every pair px, yq of
vectors of H, denote by b x, y the element of K b K, with K “ ‘ Kj , defined by
bx, y “
ÿ
j PJ
¯
1 ´ÿ
a
xi, j b y i, j .
dj i PI
j PJ
j
It should be pointed out that for a fixed index j P J the vectors xi, j and yi, j are understood
in the formula above as belonging to the same space Kj (and not to the various orthogonal
spaces Hi, j ). So b x, y is a vector of K b K, not of H b H. Thus
ÿ 1 ÿ
xxu, j , xv, j y xyu, j , yv, j y.
||bb x, y ||2 “
d
j PJ j u, v PI
j
Combining Corollary 3.5 with the formula
`
˘
m |xπp qx, yy|2 “ xPπb π x b x, y b y y “ xPπ x , xyx, x , yyyy
yields
˘
`
Corollary 3.6. — Let π “ ‘ ‘ πj be a compact representation of G on H. For
˘j PJ i PIj
˘
`
`
every vectors x “ ‘ ‘ xi, j and y “ ‘ ‘ yi, j of H, we have
j PJ i PIj
(3.1)
`
˘ ÿ 1
m |xπp qx, yy|2 “
d
j PJ j
j PJ i PIj
ÿ
u, v PIj
xxu, j , xv, j y . xyu, j , yv, j y “ ||bb x, y ||2 .
We thus obtain the following abstract version of the Wiener Theorem for unitary representations of a group G:
Theorem 3.7. — Let π “ πw ‘ πc be a unitary representation of G on a Hilbert space
H “ Hw`‘ Hc , where
πw is the weakly mixing part of π and πc its compact part. Writing
˘
πc “ ‘ ‘ πj as above, we have for every vectors x “ xw ‘ xc and y “ yw ‘ yc of H
j PJ i PIj
`
˘
m |xπp qx, yy|2 “ || b xc , yc ||2 .
`
˘
`
˘
Proof. — As we have mp|xπp qx, yy|2 q “ m |xπw p qxw , yw y|2 ` m |xπc p qxc , yc y|2 and
mp|xπw p qxw , yw y|2 q “ 0, this follows from Corollary 3.6.
`
˘
We finally derive an inequality on the quantities m |xπp qx, yy|2 for a compact representation π, which is a direct consequence of Corollary 3.6. This inequality will be a
crucial tool for the proof of our main result, to be given in Section 4. Using the same
notation as in the statement of Corollary 3.6, we denote by x “ ‘j PJ x
rj and y “ ‘j PJ yrj
the respective decompositions of the vectors x and y of H with respect to the decompor j of H. Applying the Cauchy-Schwarz inequality twice to (3.1) yields
sition H “ ‘j PJ H
the following inequalities:
(3.2)
Corollary 3.8. — Let π be a compact representation of G on H. For every vectors x
and y of H, we have
ÿ
`
˘ ÿ 1
||r
xj ||2 . ||r
yj ||2 ď
||r
xj ||2 . ||r
yj ||2 .
m |xπp qx, yy|2 ď
d
j
j PJ
j PJ
KAZHDAN SETS AND EQUIDISTRIBUTION PROPERTIES
11
3.5. Why is (3.2) an abstract version of the Wiener Theorem?— Theorem 3.7
admits a much simpler formulation in the case where G is an abelian group. If π is a
compact representation of G, the formula (3.1) becomes
`
˘ ÿ ÿ
m |xπp qx, yy|2 “
xu, j xv, j y u, j yv, j
j PJ u, v P Ij
where xi, j and yi, j , i P Ij , j P J, are simply scalars. Using the notation of Corollary 3.8,
we have
ˇ2
ÿˇ
ˇ2
`
˘ ÿ ˇˇ ÿ
ˇ
ˇxr
(3.3)
m |xπp qx, yy|2 “
xu, j y u, j ˇ “
xj , yrj yˇ .
ˇ
j PJ u P Ij
j PJ
For every character χ P Γ (where Γ denotes the dual group of G), we denote by Eχ the
subspace of H
Eχ “ tx P H ; πpgqx “ χpgqx for every g P Gu
and by Pχ the orthogonal projection of H on Eχ . Each representation πj , j P J, being
r j with Eχ . Equation
in fact a character χj on the group G, we can identify the space H
j
(3.3) then yields the following corollary:
Corollary 3.9. — Let G be an abelian group, and let π be a representation of G on a
Hilbert space H. Then we have for every x, y P H
ÿ ˇ
ˇ
ˇ
`
˘ ÿˇ
ˇxPEχ x, PEχ yyˇ2.
ˇxPEχ x, PEχ yyˇ2 “
m |xπp qx, yy|2 “
j
j
In particular, if x “ y,
χPΓ
j PJ
ÿ ˇˇ
ˇˇ
`
˘
ˇˇPEχ xˇˇ4.
m |xπp qx, xy|2 “
χPΓ
Specializing Corollary 3.9 to the case where G “ Z yields that for any unitary operator
U on H and any vectors x, y P H,
N ˇ
ÿ
ˇ
1
ˇxU n x, yyˇ2
2N ` 1 n“´N
/
N
/
`8
λPT
In particular, we have
(3.4)
N ˇ
ÿ
ˇ
1
ˇxU n x, xyˇ2
2N ` 1 n“´N
ÿ ˇ
ˇ
ˇxPkerpU ´λId q x, PkerpU ´λId q yyˇ2.
H
H
/
N
/
`8
ÿ ˇˇ
ˇˇ
ˇˇPkerpU ´λId q xˇˇ4.
H
λPT
If σ is a probability measure on the unit circle T, the operator Mσ of multiplication by
eiθ on L2 pT, σq is unitary. Applying (3.4) to U “ Mσ and to x “ 1, the constant function
equal to 1, we obtain Wiener’s Theorem:
(3.5)
N
ÿ
1
|σ
ppnq|2
2N ` 1 n“´N
/
N
/
`8
ÿ
σptλuq2 .
λPT
We refer the reader to [1, 2, 3, 9, 14] and the references therein for related aspects
and generalizations of Wiener’s theorem.
We now have all the necessary tools for the proof of Theorem 2.3, which we present in
the next section.
12
CATALIN BADEA & SOPHIE GRIVAUX
4. Proof of Theorem 2.3
4.1. Notation. — Let pWn qně1 be an increasing sequence of subsets of G satisfying the
assumptions of Theorem 2.3, and let Q be a subset of G. For each n ě 1, we denote by Qn
the set Qn “ Wn YQ. Remark that G is the increasing union of the sets Qn , n ě 1. We also
denote by ε0 a positive constant such that assumption (*) holds true: any representation
of G admitting a pQ, ε0 q-invariant vector has a finite dimensional subrepresentation.
In order to prove Theorem 2.3, we argue by contradiction, and suppose that Qn is a nonKazhdan set in G for every n ě 1. We will then construct for every ε ą 0 a representation
π of G which admits a pQ, εq-invariant vector, but is weakly mixing (which, by Proposition
3.1, is equivalent to the fact that π has no finite dimensional subrepresentation), and this
will contradict (*).
4.2. Construction of a sequence pπn qně1 of finite dimensional representations
of G. — The first step of the proof is to show that assumption (*) combined with the
hypothesis that Qn is a non-Kazhdan set for every n ě 1 implies the existence of sequences
of finite dimensional representations of G with certain properties.
Lemma 4.1. — Let ǫ0 be a positive constant such that assumption (*) holds true and
suppose that Qn is a non-Kazhdan set in G for every n ě 1. For every sequence pεn qně1
of positive real numbers decreasing to zero with ε1 P p0, ε0 s, there exist a sequence pHn qně1
of finite dimensional Hilbert spaces and a sequence pπn qně1 of unitary representations of
G such that, for every n ě 1, πn is a representation of G on Hn and
– πn has no non-zero G-invariant vector;
– πn has a pQn , εn q-invariant unit vector an P Hn : ||an || “ 1 and
sup || πn pgqan ´ an || ă εn .
g P Qn
Proof. — Let n ě 1. Since Qn is a not a Kazhdan set in G, there exists a representation
ρn of G on a Hilbert space Kn which has no non-zero G-invariant vector, but is such that
there exists a unit vector xn P Kn with
sup || ρn pgqxn ´ xn || ă 2´n .
g P Qn
Since 2´n ď ε0 for n large enough, assumption (*) implies that, for such integers n, ρn
has a finite dimensional subrepresentation. By Proposition 3.1, ρn is not weakly mixing.
This means that if we decompose Kn as Kn “ Kn, w ‘ Kn, c and ρn as ρn “ ρn, w ‘ ρn, c ,
where ρn,w and ρn,c are respectively the weakly mixing and compact parts of πn , ρn, c is
non-zero. Since ρn has no non-zero G-invariant vector, neither have ρn, w nor ρn, c .
Decomposing xn as xn “ xn, w ‘ xn, c , we have 1 “ ||xn, w ||2 ` ||xn, c ||2 . We claim that
lim nÑ`8 || xn, c || ą 0. Indeed, suppose that it is not the case. Then lim nÑ`8 ||xn, w || “ 1.
Since ||ρn pgqxn ´ xn ||2 “ ||ρn, w pgqxn, w ´ xn, w ||2 ` ||ρn, c pgqxn, c ´ xn, c ||2 for every g P G,
we have
ˇˇ
xn, w
xn, w ˇˇˇˇ
2´n
ˇˇ
sup ˇˇρn, w pgq
´
ˇˇ ă
||xn, w || ||xn, w ||
||xn, w ||
g P Qn
as soon as xn, w is non-zero. Since lim nÑ`8 ||xn, w || “ 1, this implies that for any δ ą 0
there exists an integer n such that ρn, w has a pQn , δq-invariant vector of norm 1. Applying
this to δ “ ε0 , there exists n0 ě 1 such that ρn0 , w has a pQn0 , ε0 q-invariant vector, hence a
KAZHDAN SETS AND EQUIDISTRIBUTION PROPERTIES
13
pQ, ε0 q-invariant vector. But ρn0 , w is weakly mixing, so has no finite dimensional subrepresentation. This contradicts assumption (*). So we deduce that lim nÑ`8 || xn, c || “ γ ą 0.
The same observation as above, applied to the representation ρn, c , shows that
ˇˇ
xn, c ˇˇˇˇ
xn, c
2´n
ˇˇ
sup ˇˇρn, c pgq
´
ˇˇ ă
||xn, c || ||xn, c ||
||xn, c ||
g P Qn
for every n such that xn, c is non-zero, and thus that
ˇˇ
xn, c
xn, c ˇˇˇˇ 2´pn´1q
ˇˇ
sup ˇˇρn, c pgq
´
ˇˇ ă
||xn, c || ||xn, c ||
γ
g P Qn
for infinitely many integers n. For these integers, ρn, c is a compact representation for which
yn “ xn, c {||xn, c || is a pQn , 2´pn´1q {γq-invariant vector of norm 1. It has no non-zero Ginvariant vector. Decomposing ρn, c as a direct sum of finite dimensional representations,
straightforward computations show that there exists for each such integer n a finite dimensional representation σn of G with a pQn , 2´pn´2q {γq-invariant vector but no non-zero
G-invariant vector. Lemma 4.1 follows immediately by taking a suitable subsequence of
pσn qně1 .
4.3. Construction of weakly mixing representations of G with pQ, εq-invariant
vectors. — Let ε ą 0 be an arbitrary positive number. Our aim is to show that there
exists a weakly mixing representation of G with a pQ, εq-invariant vector. We fix a sequence
pεn qně1 of positive numbers decreasing to zero so fast that the following properties hold:
ř
(i) 0 ă εn ă ε0 for every n ě 1, and ně1 εn ă ε2 {2;
ř2n 2
1
(ii) the sequence p pn`1qε
2
j“n εj qně1 tends to 0 as n tends to infinity.
n
We Â
consider the representation π “ bně1 πn of G on the infinite tensor product space
a
H “
ně1 Hn , where the spaces Hn , the representations πn and the vectors an are
associated to εn for each n ě 1 by Lemma 4.1. We refer to the appendix for undefined
notation concerning infinite tensor products. We first prove the following fact:
Fact 4.2. — Under the assumptions above, π is a strongly continuous representation of
G on H which has a pQ, εq-invariant vector.
Proof of Fact 4.2. — In order to prove that π is well-defined and strongly continuous, it
suffices to check that the assumptions of Proposition A.2 in the appendix hold true. For
every g P G and n ě 1, we have |1 ´ xπn pgqan , an y| ď ||πn pgqan ´ an || so that
sup |1 ´ xπn pgqan , an y| ă εn .
g P Qn
ř
By assumption (i), the series ně1 εn is convergent.
ř Since every element g P G belongs to
all the sets Qn except finitely many, the series ně1 |1 ´ xπn pgqan , an y| is convergent for
every g P G. Moreover, it is uniformly convergent on Q1 , and hence on W1 . The function
ÿ
/
|1 ´ xπn pgqan , an y|
g✤
ně1
is thus continuous on W1 , which is a neighborhood of e. It follows then from Proposition
A.2 that π is strongly continuous on H . If G is locally compact, Proposition A.1 and the
first part of the argument above suffice to show that π is strongly continuous, even when
W1 is not a neighborhood of e.
CATALIN BADEA & SOPHIE GRIVAUX
14
Âa
Next, it is easy to check that the elementary vector
ś a “ bně1 an of ně1 Hn satisfies
a|| “ 1 and supgPQ ||π
π pgqa
a ´ a || ă ε. Indeed ||a
a|| “ ně1 ||an || “ 1, and for every g P Q
||a
we have (using the fact that Q Ď Qn for every n ě 1)
ˇ
ˇ
ź
ˇ
ˇ
π pgqa
a ´ a ||2 “ 2 p1 ´ Rexπ
π pgqa
a , a yq ď 2 ˇ1 ´
||π
xπn pgqan , an yˇ
ď2
ÿ
ně1
ně1
|1 ´ xπn pgqan , an y| ă 2
ÿ
εn .
ně1
π pgqa
a ´ a ||2 ă ε2 , and a is
Assumption (i) on the sequence pεn qně1 implies that supg P Q ||π
thus a pQ, εq-invariant vector for π .
Using the notation of Section 3.2, we now decompose πn and Hn as
¯
´
¯
´
‘ Hi, j, n
and Hn “ ‘
‘ πj, n
πn “ ‘
j P Jn
i P Ij, n
j P Jn
i P Ij, n
respectively. Since Hn is finite dimensional, all the sets Jn and Ij, n , j P Jn , are finite,
and we assume that they are subsets
of N. For
˘ every j P Jn , Hi, j, n “ Kj, n . We also
`
aj, n “ ‘iPIj, n ai, j, n for every
decompose an P Hn as an “ ‘
‘ ai, j, n , and write r
j P Jn . We have
(4.1)
j PJn
¨
||r
aj, n || “ ˝
ÿ
i PIj, n
˛1
2
|| ai, j, n ||2 ‚
i PIj, n
and
||an || “
˜
ÿ
j PJn
2
||r
aj, n ||2
so that ||r
aj, n || ď 1 for every j P Jn . Also,
ÿ ÿ
(4.2)
||πj, n pgqai, j, n ´ ai, j, n ||2
||πn pgqan ´ an ||2 “
j PJn i PIj, n
so that
(4.3)
sup
´ ÿ
g P Qn i PI
j, n
||πj, n pgq ai, j, n ´ ai, j, n ||2
¯1{2
ă εn
¸1
“ 1,
for every g P G,
for every j P Jn .
There are now two cases to consider.
‚ Case 1. We have limnÑ`8 maxjPJn ||r
aj, n || “ 0.
ř
aj, n ||2 “ ||an ||2 “ 1, we obtain that
Using Corollary 3.8 and the fact that j PJn || r
ÿ
ÿ
mp |xπn p qan , an y|2 q ď
aj, n ||2 .
|| r
aj, n ||4 ď max || r
aj, n ||2 .
|| r
aj, n ||2 ď max || r
j PJn
j PJn
j PJn
j PJn
It follows from our assumption that limnÑ`8 mp |xπn p qan , an y|2 q “ 0. So π is weakly
mixing by Proposition A.3. We have thus proved in this case the existence of a weakly
mixing representation of G with a pQ, εq-invariant vector.
‚ Case 2. There exists δ ą 0 such that maxjPJn ||r
aj, n || ą δ for every n ě 1.
Let, for every n ě 1, jn P Jn be such that ||r
ajn , n || ą δ. Set In “ Ijn , n Ď N, σn “ πjn , n ,
Kn “ Kjn , n and bi, n “ ai, jn , n for every i P In . Then σn is a non-trivial irreducible
KAZHDAN SETS AND EQUIDISTRIBUTION PROPERTIES
15
representation of G on the finite dimensional space Kn , and by (4.1) and (4.3) the finite
family pbi, n qi PIn of vectors of Kn satisfies
´ÿ
¯1{2
´ÿ
¯1{2
(4.4)
1ě
||σn pgq bi, n ´ bi, n ||2
ă εn .
|| bi, n ||2
ą δ and sup
If we write
this means that
(4.5)
g P Qn i PI
n
i PIn
r n “ ‘ Kn ,
K
i PIn
1 ě ||rbn || ą δ
rbn “ ‘ bi, n ,
i PIn
and
and σ
rn “ ‘ σn ,
i PIn
sup ||r
σn pgqrbn ´ rbn || ă εn .
g P Qn
Now we again have to consider separately two cases.
– Case 2.a. There exists an infinite subset D of N such that whenever k and l are two
distinct elements of D, σk and σl are not equivalent. Replacing the sequence pσn qně1 by
pσn qnPD , we can suppose without loss of generality that for every distinct integers m and
n, with m, n ě 1, σm and σn are not equivalent.
Consider for every n ě 1 the representation
rn ‘ ¨ ¨ ¨ ‘ K
r 2n ,
ρn “ σ
rn ‘ ¨ ¨ ¨ ‘ σ
r2n of G on Hn “ K
1`
`ř2n
˘
˘
r 2 ´ 2 rbn ‘ ¨ ¨ ¨ ‘ rb2n of Hn , which satisfies ||bn || “ 1. For
and the vector bn “
k“n ||bk ||
every g P Qn we have, since Qn is contained in Qj for every j ě n,
2
||ρn pgqbn ´ bn || “
2n
`ÿ
k“n
||rbk ||2
2n ˇˇ
ˇˇ2
˘´1 ÿ
ˇˇσ
rj pgqrbj ´ r
bj ˇˇ ă
j“n
2n
ÿ
1
ε2
δ2 pn ` 1q j“n j
by (4.5). By assumption (ii) on the sequence pεn qně1 , we obtain that there exists an integer
n0 ě 1 such that supg P Qn ||ρn pgqbn ´ bn || ă εn for every n ě n0 . Let now ρ “ bněn0 ρn
Âa
be the infinite tensor product of the representations ρn on the space H “
něn0 Hn .
An argument similar to the one given in Fact 4.2 shows that ρ is a strongly continuous
representation of G on H which has a pQ, εq-invariant vector. It remains to prove that
ρ is weakly mixing, and for this we will show that mp|xρn p qbn , bn y|2 q tends to zero as n
tends to infinity. Recall that for every n ě 1, the representations σn , . . . , σ2n are mutually
non-equivalent, so that, by Corollary 3.8, we have for every n ě 1
2n
2n
2n ˇˇ ÿ
ÿ
ÿ
˘ 1 ˇˇ4
1
1
ˇˇ`
2 ´ 2 r ˇˇ
2
r
||rbj ||4 ď 4
bj ˇˇ ď 4
||bk ||
mp|xρn p qbn , bn y| q ď
ˇˇ
2
δ
pn
`
1q
δ
pn
` 1q
j“n
j“n k“n
by (4.5). So mp|xρn p qbn , bn y|2 q tends to zero as n tends to infinity. By Proposition A.3,
ρ is weakly mixing. We have proved again in this case the existence of a weakly mixing
representation of G with a pQ, εq-invariant vector.
The other case we have to consider is when there exists an integer n1 ě 1 such that for
every n ě n1 , σn is equivalent to one of the representations σ1 , . . . , σn1 . Indeed, if there
is no such integer, we can construct a strictly increasing sequence pnk qkě1 of integers such
that, for every k ě 1, σnk is not equivalent to one of the representations σ1 , . . . , σnk´1 .
The set D “ tnk ; k ě 1u then has the property that whenever m and n are two distinct
elements of D, σm and σn are not equivalent, and we are back to the setting of Case 2.a.
Without loss of generality, we can suppose that σn is equal to σ1 for every n ě 1.
CATALIN BADEA & SOPHIE GRIVAUX
16
– Case 2.b. For every n ě 1, σn is equal to σ1 . By (4.4), we have
¯1{2
´ÿ
¯1{2
´ÿ
||σ1 pgqbi, n ´ bi, n ||2
ă εn ,
|| bi, n ||2
ą δ and sup
1ě
g P Qn i PI
n
i PIn
where all the vectors bi,n , i P In , belong to H1 . For each n ě 1, set cn “ ‘ bi, n , seen
i PIn
as a vector of the infinite direct sum H “ ‘ H1 by defining its j th coordinate to be zero
jě1
when j does not belong to In . Let also σ be the infinite direct sum σ “ ‘ σ1 of σ1 on
jě1
H. Then we have, for every n ě 1,
1 ě ||cn || ą δ
and
sup ||σpgqcn ´ cn || ă εn .
g P Qn
Let now S be a finite subset of G. There exists an integer nS ě 1 such that S Ď Qn for
every n ě nS , and hence
sup ||σpgqcn ´ cn || ă εn
g PS
for every n ě nS .
It follows that σ has almost-invariant vectors for finite sets: for every δ ą 0 and every
finite subset S of G, σ has an pS, δq-invariant vector. This implies that σ1 itself has
almost-invariant vectors for finite sets (see [34, Lem. 1.5.4] or [27]). Since σ1 is a finite
dimensional representation, it follows that σ1 has almost-invariant vectors. If pvn qně1 is a
sequence of unit vectors of H1 such that
sup ||σpgqvn ´ vn || ă 2´n
g PG
for every n ě 1,
then any accumulation point of pvn qně1 is a non-zero G-invariant vector for σ1 . This
contradicts our initial assumption on σ1 , and shows that the hypothesis of Case 2.b cannot
be fulfilled.
Summing up our different cases, we have thus proved that there exists for every ε ą 0 a
representation of G with a pQ, εq-invariant vector but no finite dimensional subrepresentation. This contradicts assumption (*) of Theorem 2.3, and concludes the proof.
5. Some consequences of Theorem 2.3
We begin this section by proving the two characterizations of Kazhdan sets obtained as
consequences of Theorem 2.3.
5.1. Proofs of Corollary 2.4 and Theorem 2.5. — Let us first prove Corollary 2.4.
Proof of Corollary 2.4. — Let Q0 be a subset of G which has non-empty interior and
which generates G. Denote for each n ě 1 by Q0˘n the set tg1˘1 . . . gn˘1 ; g1 , . . . , gn P Q0 u.
Ť
Then G “ ně1 Q0˘n . Let g0 be an element of the interior of Q0 . Then g0´1 Q0 is a
˘pn `1q
neighborhood of e. There exists n0 ě 1 such that g0´1 belongs to Q0˘n0 , and thus Q0 0
˘pn `nq
is a neighborhood of e. If we set Wn “ Q0 0
for n ě 1, the sequence of sets pWn qně1 is
increasing, W1 is a neighborhood of e, and pWn qně1 satisfies the assumptions of Theorem
2.3. So if Q is a subset of G for which assumption (*) of Theorem 2.3 holds true, there
˘pn`n0 q
exists n ě 1 such that Q0
Y Q is a Kazhdan set in G. Let ε ą 0 be a Kazhdan
constant for this set. Then ε{pn ` n0 q is a Kazhdan constant for Q0 Y Q, and Q0 Y Q
KAZHDAN SETS AND EQUIDISTRIBUTION PROPERTIES
17
is a Kazhdan set in G. If G is locally compact, the same proof holds true without the
assumption that Q0 has non-empty interior.
Proof of Theorem 2.5. — Let ?
us first show that (a) implies (b). Suppose
a that Q is a
Kazhdan set, and let 0 ă ε ă 2 be a Kazhdan constant for Q. Let δ “ 1 ´ ε2 {2 and
consider a representation π of G on a Hilbert space H for which there is a vector x P H
with ||x|| “ 1 such that inf gPQ |xπpgqx, xy| ą δ. Then the representation π b π of G on
H b H verifies
2Rexπ b πpgqx b x, x b xy “ 2|xπpgqx, xy|2 ą 2 ´ ε2
for every g P Q. Hence }π b πpgqx b x ´ x b x} ă ε for every g P Q and π b π has a nonzero G-invariant vector. It follows from Proposition 3.1 that π has a finite dimensional
subrepresentation. Thus pbq is true. That (b) implies (c) is straightforward, and that (c)
implies (a) is a consequence of Corollary 2.4.
5.2. Property (T) in σ-compact locally compact groups. — As a consequence of
Theorem 2.3, we retrieve a characterization of Property (T) due to Bekka and Valette [5],
[6, Th. 2.12.9], valid for σ-compact locally compact groups, which states the following:
Theorem 5.1 ([5]). — Let G be a σ-compact locally compact group. Then G has Property
(T) if and only if every unitary representation of G with almost-invariant vectors has a
non-trivial finite dimensional subrepresentation.
The proof of [5] relies on the equivalence between Property (T) and Property (FH) for
such groups [6, Th. 2.12.4]. As a direct consequence of Theorem 2.3, we will derive a new
proof of Theorem 5.1 which does not involve property (FH).
If Q is a subset of a topological group G, and if π is a unitary representation of G on
a Hilbert space H, we say that π has Q-almost-invariant vectors if it has pQ, εq-invariant
vectors for every ε ą 0. The same argument as in [6, Prop. 1.2.1] shows that Q is a
Kazhdan set in G if and only if every representation of G with Q-almost-invariant vectors
has a non-zero G-invariant vector. As a direct corollary of Theorem 2.5, we obtain the
following characterization of Kazhdan sets which generate the group:
Corollary 5.2. — Let Q be a subset of a locally compact group G which generates G.
Then Q is a Kazhdan set in G if and only if every representation π of G with Q-almostinvariant vectors has a non-trivial finite dimensional subrepresentation.
Proof of Corollary 5.2. — The only thing to prove is that if every representation π of G
with Q-almost-invariant vectors has a non-trivial finite dimensional representation, Q is a
Kazhdan set. For this it suffices to show the existence of an ε ą 0 such that assumption
(*) of Theorem 2.3 holds true. The argument is exactly the same as the one given in [6,
Prop. 1.2.1]: suppose that there is no such ε, and let, for every ε ą 0, πε be a representation
of GÀ
with a pQ, εq-invariant vector but no finite dimensional subrepresentation. Then
π “ εą0 πε has Q-almost-invariant vectors but no finite dimensional subrepresentation
(this follows immediately from [6, Prop. A.1.8]), contradicting our initial assumption.
Proof of Theorem 5.1. — It is clear that Property (T) implies that every representation
of G with almost-invariant vectors has a non-trivial finite dimensional subrepresentation.
Conversely, suppose that every representation of G with almost-invariant vectors has a
non-trivial finite dimensional subrepresentation. Using the same argument as in the proof
of Corollary 5.2, we see that there exists a compact subset Q of G such that assumption
CATALIN BADEA & SOPHIE GRIVAUX
18
(*) of Theorem 2.3 holds
Ť true. Choosing for pWn qně1 an increasing sequence of compact
subsets of G such that ně1 Wn “ G, Theorem 2.3 implies that there exists an n ě 1 such
that Wn Y Q is a Kazhdan set in G. Since Wn Y Q is compact, G has Property (T).
5.3. Equidistribution assumptions: proofs of Theorems 2.2 and 2.1. — Let G
be a second countable locally compact group, and let π be a unitary representation of G on
a separable Hilbert space H. Such a representation can be decomposed as a direct integral
of irreducible unitary representations over a Borel space (see for instance [6, Sec. F.5] or
[15]). More precisely, there exists a finite positive measure µ on a standard Borel space Z,
a measurable field z ÞÑ Hz of Hilbert spaces over Z, and a measurable field of irreducible
representations z ÞÑ πz , where each πz is a representation
of G on H
żz , such that π is
ż
unitarily equivalent to the direct integral πµ “
‘
Z
πz dµpzq on H “
‘
Hz dµpzq. The
Z
Hilbert space H is the set of equivalence classes of square integrable vector fields z ÞÑ xz ,
with xz P Hz , with respect to the measure µ; πµ is the representation of G on H defined
by πµ pgqx “ rz ÞÑ πz pgqxz s for every g P G and x P H .
Proof of Theorem 2.1. — Our aim is to show that, under the hypothesis of Theorem 2.1,
assumption (*) of Theorem 2.3 is satisfied. Let π be a representation of G on a Hilbert
space H. Since G is second countable, we can suppose that H is separable. Suppose that
π admits a pQ, 1{2q-invariant vector x P H and, using the notation and the result recalled
above, write
ż‘
ż‘
/ xz s, and H “
Hz dµpzq.
πz dµpzq, x “ r z ✤
π“
Z
Z
We have for every k ě 1
Re xπpgk qx, xy “ Re
ż
Z
xπz pgk qxz , xz y dµpzq “ 1 ´
1
7
||πpgk qx ´ x||2 ą
2
8
so that
(5.1)
Re
ż
Z
N
1 ÿ
7
xπz pgk qxz , xz y dµpzq ą
N k“1
8
for every N ě 1.
Now, assumption (2.2) of Theorem 2.1 states that there exists a countable set C0 of equivalence classes of irreducible representations such that
(5.2)
N
1 ÿ
xπpgk qx, xy ÝÑ 0
N k“1
as N ÝÑ `8
for every irreducible representation π whose equivalence class rπs does not belong to C0
and every vector x in the underlying Hilbert space. It follows from (5.2) that the set
Z0 “ tz P Z ; rπz s P C0 u satisfies µpZ0 q ą 0, and there exists rπ0 s P C0 such that
µptz P Z ; πz and π0 are equivalentuq ą 0. Hence π0 is a subrepresentation of π. Since all
irreducible representations of G are supposed to be finite dimensional, π has a finite dimensional subrepresentation. So assumption (*) of Theorem 2.3 is satisfied. As Q generates
G, it now follows from Theorem 2.5 that Q is a Kazhdan set in G.
Proof of Theorem 2.2. — The proof of Theorem 2.2 is exactly the same as that of Theorem
2.1, using the fact that if G is a locally compact abelian group (not necessarily second
KAZHDAN SETS AND EQUIDISTRIBUTION PROPERTIES
19
countable), any unitary representation of G is equivalent to a direct integral of irreducible
representations (see for instance [15, Th. 7.36]).
6. Examples and applications
We present in this section some examples of Kazhdan sets in different kinds of groups,
some statements being obtained as consequences of Theorems 2.3 or 2.5. We do not try to
be exhaustive, and our aim here is rather to highlight some interesting phenomena which
appear when looking for Kazhdan sets, as well as the connections of these phenomena
with some remarkable properties of the group. We begin with the simplest case, that of
locally compact abelian (LCA) groups.
6.1. Kazhdan sets in locally compact abelian groups. — Let G be a second countable LCA group, the dual group of which we denote by Γ. If σ is a finite Borel measure
on Γ, recall that its Fourier-Stieljes transform is defined by
ż
σ
ppgq “ γpgq dσpγq for every g P G.
Γ
It is an easy consequence of the spectral theorem for unitary representations that if Q is
a subset of a second countable LCA group G, Q is a Kazhdan set in G if and only if there
exists ε ą 0 such that any probability measure σ on Γ with supgPQ |p
σ pgq ´ 1| ă ε satisfies
σpt1uq ą 0, where 1 denotes the trivial character on G. Using Theorem 2.5 combined with
the spectral theorem for unitary representations again, we obtain the following stronger
characterization of Kazhdan sets which generate the group in any second countable LCA
group.
Theorem 6.1. — Let G be a second countable LCA group, and let Q a subset of G which
generates G. The following assertions are equivalent:
(1) Q is a Kazhdan set in G;
(2) there exists δ P p0, 1q such that any probability measure σ on Γ with inf gPQ |p
σ pgq| ą δ
has a discrete part;
(3) there exists ε ą 0 such that any probability measure σ on Γ with supgPQ |p
σ pgq ´ 1| ă ε
has a discrete part.
Theorem 6.1 becomes particularly meaningful in the case of the group Z, as it yields a
characterization of Kazhdan subsets of Z involving some classic sets in harmonic analysis,
introduced by Kaufman in [24]. They are called w-sets by Kaufman [25], and Kaufman
sets (Ka sets) by other authors, such as Hartman [20], [21].
Definition 6.2. — Let Q be a subset of Z, and let δ P p0, 1q.
‚ We say that Q belongs to the class Ka if there exists a finite complex-valued continuous Borel measure µ on T such that inf nPQ |µ̂pnq| ą 0, and to the class δ-Ka if there
exists a finite complex-valued continuous Borel measure µ on T with µpTq “ 1 such that
inf nPQ |µ̂pnq| ą δ.
‚ We say that Q belongs to the class Ka` if there exists a continuous probability
measure σ on T such that inf nPQ |σ̂pnq| ą 0, and to the class δ-Ka` if there exists a
continuous probability measure σ on T such that inf nPQ |σ̂pnq| ą δ.
Our characterization of Kazhdan subsets of Z is given by Theorem 6.3 below:
CATALIN BADEA & SOPHIE GRIVAUX
20
Theorem 6.3. — Let Q a subset of Z which generates Z. Then Q is a Kazhdan set in Z
if and only if there exists a δ P p0, 1q such that Q does not belong to δ-Ka` .
It is interesting to remark [21] that a set Q belongs to Ka if and only if it belongs to
δ-Ka for every δ P p0, 1q. There is no similar statement for the class Ka` : any sufficiently
lacunary subset of Z, such as Q “ t3k ` k ; k ě 1u, is easily seen to belong to Ka` (it
suffices to consider an associated Riesz product – see for instance [22] for details); but the
same reasoning as in Example 6.4 below shows that this set Q is a Kazhdan subset of Z.
Thus there exists by Theorem 6.3 a δ P p0, 1q such that Q does not belong to δ-Ka` .
We present now some typical examples of Kazhdan sets in Z or R obtained using the
above characterizations. The first one provides a negative answer to Question 1.5 (a).
Example 6.4. — The set Q “ t2k ` k ; k ě 0u is a Kazhdan set in Z and there are irrational numbers θ such that pe2iπnθ qnPQ is not dense in T. In particular, no rearrangement
pmk qk ě1 of the elements of Q exists such that pe2iπmk θ qk ě1 is equidistributed in T for
every irrational number θ.
Proof. — The sequence pnk qkě0 defined by nk “ 2k ` k for every k ě 0 satisfies the
relation 2nk “ nk`1 ` k ´ 1 for every k ě 0. Let σ be a probability measure on T such
that supkě0 |p
σ pnk q ´ 1| ă 1{18. Since, by the Cauchy-Schwarz inequality,
ż
?
σ pkq ´ 1|1{2 for every k P Z,
|p
σ pkq ´ 1| ď |λk ´ 1|dσpλq ď 2 |p
T
we have
ż
ż
|p
σ pk ´ 1q ´ 1| ď 2 |λ ´ 1|dσpλq ` |λnk`1 ´ 1|dσpλq
?T
?T
1{2
σ pnk q ´ 1| ` 2 |p
σ pnk`1 q ´ 1|1{2
ď 2 2 |p
nk
for all k ě 1, so that supkě0 |p
σ pkq ´ 1| ă 1. Since
ż ´ ÿ
N
N
¯
1 ÿ
1
/ σpt1uq as N
σ
ppkq “
λk dσpλq
N k“1
T N k“1
/ ` 8,
we have σpt1uq ą 0. So Q “ tnk ; k ě 0u is a Kazhdan set in Z. But pnk qkě0 being
lacunary, it follows from a result proved independently by Pollington [35] and De Mathan
[30] that there exists a subset A of r0, 1s of Hausdorff measure 1 such that for every θ in
A, the set Qθ “ tnk θ ; k ě 0u is not dense modulo 1. One of these numbers θ is irrational,
and the conclusion follows.
Example 6.5. — The set Q1 “ t2k ; k ě 0u is not a Kazhdan set in Z.
Proof. — The fact that Q1 is not a Kazhdan set in Z relies on the observation that 2k
divides 2k`1 for every k ě 0. Using the same construction as the one of [13, Prop. 3.9],
we consider for any fixed ε ą 0 a decreasing
sequence paj qjě1 of positive real numbers
ř
with a1 ă ε{p2πq such that the series jě1 aj is divergent. Then the infinite convolution
of two-points Dirac measures
`
˘
σ “ ˚ p1 ´ aj qδt1u ` aj δteiπ2´j`1 u
jě1
KAZHDAN SETS AND EQUIDISTRIBUTION PROPERTIES
21
is a well-defined
probability measure on T, which is continuous by the assumption that
ř
the series jě1 aj diverges. For every k ě 0,
ź `
ź`
k´j`1 ˘
k´j`1 ˘
1 ´ aj ` aj eiπ2
“
σ
pp2k q “
1 ´ aj p1 ´ eiπ2
q .
jě1
jěk`1
iπ2k´j`1
q| ď 1, it follows that
ÿ
ÿ
k´j`1
|p
σ p2k q ´ 1| ď
aj |1 ´ eiπ2
| ď π ak`1 2k`1
2´j “ 2πak`1 ă ε
As |1 ´ aj p1 ´ e
jěk`1
for every k ě 0. This proves that
jěk`1
Q1
is not a Kazhdan set in Z.
Example 6.6. — If p is a non-constant polynomial with integer coefficients such that
ppZq is included in aZ for no integer a with |a| ě 2, then Q “ tppkq ; k ě 0u is a Kazhdan
set in Z.
Proof. — Our assumption that ppZq is included in aZ for no integer a with |a| ě 2 implies
that Q generates Z. Since the sequence pλppkq qkě0 is uniformly distributed in T for every
λ “ e2iπθ with θ irrational (see for instance [28, Th. 3.2]), Theorem 2.1 implies that Q is
a Kazhdan set in Z.
Example 6.7. — Let p be a non-constant real polynomial, and let Q “ tppkq ; k ě 0u.
Then p´δ, δq Y Q is a Kazhdan subset of R for any δ ą 0.
ř
Proof. — Write p as ppxq “ dj“0 aj xj , d ě 1, and let r P t1, . . . , du be such that ar ‰ 0. It
is well-known (see for instance [28, Th. 3.2]) that the sequence pe2iπtppkq qkPZ is uniformly
distributed in T as soon as tar is irrational. This condition excludes only countably many
values of t. Set now Wn “ p´n, nq for every integer n ě 1. Thanks to Theorem 2.3, we
obtain that there exists n ě 1 such that p´n, nq Y Q is a Kazhdan set in R. Let ε ą 0
be a Kazhdan constant for this set. Fix δ ą 0. In order to prove that p´δ, δq Y Q is a
Kazhdan set in R, we consider a positive number γ, which
ˇ will be
ˇ fixed later on, and let
pptq ´ 1ˇ ă γ. For any a P N and
σ be a probability measure on R such that suptPp´δ,δqYQ ˇσ
any t P p´δ, δq,
ż
ż
ˇ
ˇ
ˇ2
ˇ2
pptqq
2p1 ´ Re σ
ppatqq “ ˇeiatx ´ 1ˇ dσpxq ď a2 ˇeitx ´ 1ˇ dσpxq ď 2a2 Re p1 ´ σ
R
R
so that suptPpδ,δq p1 ´ Re σ
ppatqq ă a2 γ. If we choose a ą n{δ and γ ă minpε, ε2 {p2a2 qq, we
ˇ
ˇ
pptqˇ ă ε, and since ε is a Kazhdan constant for p´n, nq Y Q,
obtain that suptPp´n,nqYQ ˇ1 ´ σ
σpt0uq ą 0. Hence γ is a Kazhdan constant for p´δ, δq Y Q.
Remark 6.8. — It is necessary to add a small interval to the set Q in order to turn
it into a Kazhdan subset of R, even?when Q generates a dense
subgroup of R. Indeed,
?
consider the polynomial ppxq “ x ` 2. The set Q “? tk ` 2 ; k ě 0u is not a Kazhdan
set in R: for any ε ą 0, let b P N be such that |e2iπb 2 ´ 1|?ă ε. The measure σ defined
as the Dirac mass at the point 2πb satisfies supkě0 |p
σ pk ` 2q ´ 1| ă ε, so that Q is not
a Kazhdan set in R.
We finish this section by exhibiting a link between Kazhdan subsets of Zd and Kazhdan
subsets of Rd , d ě 1. Let Q be a subset of Zd . Seen as a subset of Rd , Q is never a
Kazhdan set. But as a consequence of Theorem 2.3, we see that Q becomes a Kazhdan
set in Rd if we add a small perturbation to it.
CATALIN BADEA & SOPHIE GRIVAUX
22
Proposition 6.9. — Fix
Ť an integer d ě 1, and let pWn qně1 be an increasing sequence of
subsets of Rd such that ně1 Wn “ Rd . Let Q be a Kazhdan subset of Zd . There exists an
n ě 1 such that Wn Y Q is a Kazhdan set in Rd . Also, Bp0, δq Y Q is a Kazhdan subset of
Rd for any δ ą 0, where Bp0, δq denotes the open unit ball of radius δ for the Euclidean
norm on Rd .
Proof. — Let ε ą 0 be a Kazhdan constant for Q, seen as a subset of Zd . Let π be a
representation of Rd on a separable Hilbert space H which admits a pQ, ε2 {2q-invariant
vector x P H. Without loss of generality we can suppose that π is a direct integral on a
Borel space Z, with respect to a finite measure µ on Z, of a family pπz qzPZ of irreducible
representations of Rd . So π is a representation of Rd on L2 pZ, µq. We write elements f of
L2 pZ, µq as f “ pfz qzPZ . We suppose that ||x|| “ 1; our hypothesis implies that
ˇ
ˇ ε2
sup ˇ1 ´ xπpttqx, xyˇ ă ¨
2
t PQ
Each representation πz acts on vectors t “ pt1 , . . . , td q of Rd as πz pttq “ expp2iπxtt, θ z yq for
some vector θ z “ pθ1,z , . . . , θd,z q of Rd . Hence
ż
ˇ ε2
ˇ
ˇ
ˇ
e2iπxtt,θθ z y |xz |2 dµpzqˇ ă ¨
sup ˇ1 ´
2
t PQ
Z
nq f : z ✤
Consider now the representation ρ of Zd on L2 pZ, µq defined by ρpn
d
2
for every n “ pn1 , . . . , nd q P Z and every f P L pZ, µq. We have
ˇ
ˇ
nqx ´ x||2 ď 2 supˇ1 ´ xρpn
nqx, xyˇ ă ε2 ,
sup ||ρpn
n PQ
n ,θθ z y
/ e2iπxn
fz
n PQ
and since ε is a Kazhdan constant for Q as a subset of Zd , ρ has a non-zero Zd -invariant
nqf “ f for every
vector. There exists hence f P L2 pZ, µq with ||f || “ 1 such that ρpn
n P Zd . Fix a representative of f P L2 pZ, µq, and set Z0 “ tz P Z ; fz ‰ 0u. Then
µpZ0 q ą 0. For every z P Z0 we have e2iπxnn,θθ z y “ 1 for every n P Zd , which implies that
θ z P Zd . For each n “ pn1 , . . . , nd q P Zd , let Z n “ tz P Z0 ; θŤ
i,z “ ni for each i P t1, . . . , nuu
and H n “ tf P L2 pZ, µq ; f “ 0 µ-a. e. on ZzZ n u. We have n PZd Z n “ Z0 , so there exists
n 0 P Zd such that µpZ n 0 q ą 0. Each subspace H n is easily seen to be invariant for π, and
ny
/ e2iπxtt ,n
fz for
the representation π n induced by π on H n is given by π n pttq f : z ✤
d
every t P R and every f P H n . So π admits a subrepresentation of dimension 1 as soon as
H n is non-zero, i. e. as soon as µpZ n q ą 0. Since µpZ n 0 q ą 0, π admit a subrepresentation
of dimension 1. An application of Theorem 2.5 now shows that Wn Y Q is a Kazhdan
set in Rd for some n ě 1. If we choose Wn “ Bp0, nq for every n ě 1, and proceed as in
the proof of Example 6.7, we obtain that Bp0, δq Y Q is a Kazhdan set in Rd for every
δ ą 0.
We now move out of the commutative setting, and present a characterization of Kazhdan
sets in the Heisenberg groups Hn .
6.2. Kazhdan sets in the Heisenberg groups Hn . — The Heisenberg group of
dimension n ě 1, denoted by Hn , is formed of triples pt, q , p q of R ˆ Rn ˆ Rn “ R2n`1 .
The group operation is given by
1
pt1 , q 1 , p 1 q ¨ pt2 , q 2 , p 2 q “ pt1 ` t2 ` ppp1 ¨ q 2 ´ p 2 ¨ q 1 q, q 1 ` q 2 , p 1 ` p 2 q,
2
KAZHDAN SETS AND EQUIDISTRIBUTION PROPERTIES
23
where p ¨qq denotes the scalar product of two vectors p and q of Rn . Irreducible unitary representations of Hn are completely classified (see for instance [36, Ch. 2], or [15, Cor. 6.51]):
there are two distinct families of such representations, which we denote respectively by
pF1 q and pF2 q:
– the representations belonging to the family pF1 q are representations of Hn on L2 pRn q.
They are parametrized by an element of R, which we write as ˘λ with λ ą 0. Then
π˘λ pt, q , p q, pt, q , p q P R2n`1 , acts on L2 pRn q as
?
?
/ eip˘λt˘ λqq ¨xx` λ2 q ¨ppq upx
x ` λ pq
π˘λ pt, q , p q u : x ✤
where u belongs to L2 pRn q. These representations have the following important property,
which will appear again in the next subsection:
Fact 6.10. — For every ˘λ P R and every u, v P L2 pRn q,
xπ˘λ pt, q , p q u, vy
/0
as
|pp|
/ `8.
Proof. — This follows directly from the dominated convergence theorem if u and v have
compact support in Rn . It then suffices to approximate u and v by functions with compact
support to get the result.
– the representations belonging to the family pF2 q are one-dimensional. They are
parametrized by elements pyy , η q of R2n : for every pt, q , p q P Hn ,
π y ,ηη pt, q , p q “ eipyy ¨qq `ηη ¨ppq .
We denote by πn the projection pt, q , p q ÞÝÑ pqq , p q of Hn onto R2n . Our main result
concerning Kazhdan sets in Hn is the following:
Theorem 6.11. — Let Q be a subset of the Heisenberg group Hn , n ě 1. The following
assertions are equivalent:
(1) Q is a Kazhdan set in Hn ;
(2) πn pQq is a Kazhdan set in R2n .
Proof. — We set Q0 “ πn pQq. The proof of Theorem 6.11 relies on the same kind of ideas
as those employed in the proof of Proposition 6.9. We start with the easy implication,
which is that p1q implies p2q. Suppose that Q is a Kazhdan set in Hn , and let ε ą 0 be a
Kazhdan constant for Q. Let σ be a probability measure on R2n such that
ˇ ε2
ˇż
ˇ
ˇ
ˇ
ˇ
ˇ
ˇ
eipyy ¨qq `ηη ¨ppq dσpyy , η q ´ 1ˇ ă
ppqq , p q ´ 1 “ sup ˇ
(6.1)
sup σ
2
2n
pqPQ0 R
pqPQ0
pqq ,p
pqq ,p
and consider the representation ρ of Hn on L2 pR2n , σq defined by
ρpt, q , p qf : pyy , η q ✤
/ eipyy ¨qq `ηη ¨ppq f py
y, η q
for every pt, q , pq P Hn and every f P L2 pR2n , σq. Then (6.1) implies that the constant
function 1 is a pQ, εq-invariant vector for ρ. Since pQ, εq is a Kazhdan pair in Hn , it follows
that ρ admits a non-zero Hn -invariant function f P L2 pR2n , σq. Fix a representative of
f , and consider the subset A of R2n consisting of pairs pyy , η q such that f pyy , η q ‰ 0.
Then σpAq ą 0, and for every pqq , p q P R2n , σ-almost every element pyy , η q of A satisfies
y ¨ q ` η ¨ p P 2πZ. Hence σ-almost every element pyy , η q of A has the property that
y ¨ q ` η ¨ p P 2πZ for every pqq , p q P Q2n . By continuity, σ-almost every element pyy , η q
of A has the property that y ¨ q ` η ¨ p belongs to 2πZ for every pqq , p q P R2n , so that
CATALIN BADEA & SOPHIE GRIVAUX
24
pyy , η q “ p00, 0 q. We have thus proved that σpt00, 0 uq ą 0, and it follows that Q0 is a
Kazhdan set in R2n .
Let us now prove the converse implication. Suppose that Q0 is a Kazhdan set in R2n ,
and let 0 ă ε ă 3 be a Kazhdan constant for Q0 . Let π be a unitary representation of Hn
on a separable Hilbert space H, which admits a pQ, 8ε q-invariant vector x P H of norm 1.
ş‘
We write as usual π as a direct integral π “ Z πz dµpzq,
µ is a finite Borel measure
ş where
2
on a standard Borel space Z, and x as pxz qzPZ , with Z ||xz || dµpzq “ 1. We have
ż
ˇ ε
ˇ
ˇ
ˇ
xπz pt, q , p qxz , xz ydµpzqˇ ă ¨
(6.2)
sup ˇ1 ´
8
q
p
Z
pt,q ,p qPQ
For every z P Z, the irreducible representation πz belongs to one of the two families pF1 q
and pF2 q. If πz belongs to pF1 q, we write it as π˘λz for some ˘λz P R, and if π belongs to
pF2 q, as π y z ,ηη z for some pyy z , η z q P R2n . Let, for i “ 1, 2, Zi be the subset of Z consisting
of the elements z P Z such that πz belongs to pFi q. We have xz P L2 pRn q for every z P Z1 ,
and xz P C for every x P Z2 . We now observe the following:
Lemma 6.12. — A Kazhdan subset of R2n contains elements pqq , p q such that the Euclidean norm |pp| of p is arbitrarily large.
Proof. — Let Q1 be a Kazhdan subset of R2n , with Kazhdan constant ε ą 0, and suppose
that there exists a constant M ą 0 such that |pp| ď M for every pqq , p q P Q1 . Let δ ą 0 be
such that 2M δ ă ε and consider the probability measure on R2n defined by
dpp
σ “ δ0 ˆ 1Bp00 ,δq
¨
|Bp00, δq|
For every pqq , p q P Q1 ,
ˇ
ˇż
dss
ˇ
ˇ
p
iss ¨p
´ 1ˇ ď 2δ|pp | ď 2M δ ă ε.
e
|p
σ pqq , p q ´ 1| “ ˇ
0
|Bp0
,
δq|
0 ,δq
Bp0
But σptp00 , 0 quq “ 0, and it follows that Q1 is not a Kazhdan set in R2n , which is a
contradiction.
By Fact 6.10, we have for every x P Z1
xπ˘λz pt, q , p qxz , xz y
/0
as |pp|
/ `8 , pt, q , p q P Hn .
ş
Since |xπ˘λ pt, q , p qxz , xz y| ď ||xz ||2 for every z P Z1 , and Z ||xz ||2 dµpzq “ 1, the dominated convergence theorem implies that
ż
/ 0 as |p
p| / `8 , pt, q , p q P Hn .
xπ˘λz pt, q , p qxz , xz ydµpzq
Z1
By Lemma 6.12, there exists an element pt0 , q 0 , p 0 q of Q with |pp0 | so large that
ˇ ε
ˇż
ˇ
ˇ
q
p
xπ˘λz pt0 , 0 , 0 qxz , xz ydµpzqˇ ă ¨
ˇ
8
Z1
Property (6.2) implies then that
ż
ˇ ε
ˇ
ˇ
ˇ
2
q
p
π y z ,ηη z pt0 , 0 , 0 q|xz | dµpzqˇ ă
ˇ1 ´
4
Z2
KAZHDAN SETS AND EQUIDISTRIBUTION PROPERTIES
from which it follows that
this into (6.2) yields that
ż
ε
|xz |2 dµpzq ą 1 ´ , so that
4
Z2
sup
pqPQ
pt,qq ,p
ş
ż
Z1
||xz ||2 dµpzq ă
25
ε
. Plugging
4
ż
ˇ
ˇ
ˇ1 ´
ˇ 3ε
ˇ
π y z ,ηη z pt, q , p q|xz |2 dµpzqˇ ă ¨
8
Z2
Since Z2 |xz |2 dµpzq ą 1 ´ ε{4 and 0 ă ε ă 3, we can, by normalizing the family pxz qzPZ2 ,
ş
suppose without loss of generality that Z “ Z2 , Z |xz |2 dµpzq “ 1 and that
ż
ˇ
ˇ
ˇ
ˇ
(6.3)
sup ˇ1 ´
eipyy z ¨qq `ηη z ¨ppq |xz |2 dµpzqˇ ă ε.
pqPQ
pt,qq ,p
Z
Consider now the unitary representation ρ of R2n on L2 pZ, µq defined by
ρpqq , p q f : z ✤
/ eipyy z ¨qq `ηη z ¨ppq fz
for every pqq , p q P R2n and every f “ pfz qzPZ P L2 pZ, µq. Then (6.3) can be rewritten as
ˇ
ˇ
ˇ
ˇ
sup ˇ1 ´ xρpqq , p qx, xyˇ ă ε, i.e.
sup ˇ1 ´ xρpqq , p qx, xyˇ ă ε.
pqPQ
pt,qq ,p
pqPQ0
pqq ,p
Since ε is a Kazhdan constant for Q0 , the representation ρ admits a non-zero R2n -invariant
vector f P L2 pZ, µq. Proceeding as in the proof of p1q ùñ p2q, we see that for every
pqq , p q P R2n , eipyy z ¨qq `ηη z ¨ppq f pzq “ f pzq µ-almost everywhere on Z, so that there exists a
subset Z0 of Z with µpZ0 q ą 0 such that f does not vanish on Z0 and, for every z P Z0 ,
y z ¨ q ` η z ¨ p belongs to 2πZ for every pqq , pq P Q2n . By continuity, y z ¨ q ` η z ¨ p belongs
to 2πZ for every z P Z0 and every pqq , p q P R2n , so that pyy z , η z q “ p00, 0 q for every z P Z0 .
So if we set Z01 “ tz P Z ; pyy z , η z q “ p00, 0 qu, we have µpZ01 q ą 0. The function f “ 1Z01
is hence a non-zero element of L2 pZ, µq, which is clearly an Hn -invariant vector for the
representation π. So pQ, 8ε q is a Kazhdan pair in Hn , and Theorem 6.11 is proved.
6.3. Kazhdan sets in the group Aff` pRq. — The underlying space of the group
Aff` pRq of orientation-preserving affine homeomorphisms of R is p0, `8q ˆ R, and the
group law is given by pa, bqpa1 , b1 q “ paa1 , b ` ab1 q, where pa, bq and pa1 , b1 q belong to
p0, `8q ˆ R. As in the case of the Heisenberg groups, the irreducible unitary representations of Aff` pRq are completely classified (see [15, Sec. 6.7]) and fall within two classes:
– the class pF1 q consists of two infinite dimensional representations π` and π´ of
Aff` pRq, which act respectively on the Hilbert spaces L2 pp0, `8q, dsq and L2 pp´8, 0q, dsq.
They are both defined by the formula
?
/ a e2iπbs f pasq
π˘ pa, bqf : s ✤
where pa, bq P p0, `8q ˆ R, f P L2 pp0, `8q, dsq in the case of π` , and f P L2 pp´8, 0q, dsq
in the case of π´ . It is a direct consequence of the Riemann-Lebesgue lemma that the
analogue of Fact 6.10 holds true for the two representations π` and π´ of G:
Fact 6.13. — For every f1 , f2 P L2 pp0, `8q, dsq and every g1 , g2 P L2 pp´8, 0q, dsq, we
have
xπ` pa, bqf1 , f2 y
/0
and
xπ´ pa, bqg1 , g2 y
/0
as
|b|
/ `8.
CATALIN BADEA & SOPHIE GRIVAUX
26
– the representations of Aff` pRq belonging to the family pF2 q are one-dimensional. They
are parametrized by R, and πλ is defined for every λ P R by the formula
πλ pa, bq “ aiλ
for every pa, bq P p0, `8q ˆ R.
Proceeding as in the proof of Theorem 6.11, we characterize the Kazhdan subsets of the
group Aff` pRq in the following way:
Theorem 6.14. — Let Q be a subset of Aff` pRq. The following assertions are equivalent:
(1) Q is a Kazhdan set in Aff` pRq;
(2) the set Q0 “ tt P R ; D b P R pet , bq P Qu is a Kazhdan set in R.
Proof. — The proof is similar to that of Theorem 6.11, and we will not give it in full
detail here. Let us first sketch briefly a proof of the implication p1q ùñ p2q. Suppose that
Q is a Kazhdan set in Aff` pRq, and let ε ą 0 beˇ a Kazhdan
constant for Q. Consider
ˇ
pptq ´ 1ˇ ă ε2 {2. We associate to σ a
a probability measure σ on R such that suptPQ0 ˇσ
representation ρ of Aff` pRq on L2 pR, σq by setting, for every pa, bq P p0, `8q ˆ R and
/ eispln aq f psq. Since
every f P L2 pR, σq, ρpa, bqf : s ✤
ˇ
ˇż `
ˇ
ˇ
eispln aq ´ 1qdσpsqˇ for every pa, bq P p0, `8q ˆ R,
||ρpa, bq1 ´ 1||2 ď 2ˇ
R
we have suptpa,bq ; ln aPQ0 u ||ρpa, bq1 ´ 1|| ă ε, i. e. suppa,bqPQ ||ρpa, bq1 ´ 1|| ă ε. Hence ρ
admits a non-zero Aff` pRq-invariant function f P L2 pR, σq, and the same argument as in
the proof of Theorem 6.11 shows then that σpt0uq ą 0. The converse implication p2q ùñ
p1q is proved in exactly the same way as in Theorem 6.11, using the same modifications
as those outlined above. The group R2n has to be replaced by the multiplicative group
pp0, `8q, ˆq and the analogue of Lemma 6.12 is that Kazhdan subsets of this group contain
elements of arbitrarily large absolute value. If Q0 is a Kazhdan set in R, with Kazhdan
constant ε small enough, the same argument as in the proof of Theorem 6.11 (involving
the same notation) shows that it suffices to prove the following statement: let µ be a finite
Borel
measure on a Borel space Z, x “ pxz qzPZ a scalar-valued şfunction of L2 pZ, µq with
ş
‘
2
Z |xz | dµpzq “ 1, and π a representation of G of the form π “ Z πλz dµpzq with
ż
ˇ
ˇ
ˇ
ˇ
ˇ
ˇ
eipln aqλz |xz |2 dµpzqˇ ă ε.
sup ˇ1 ´ xπpa, bqx, xyˇ “
sup ˇ1 ´
pa,bqPQ
ta ; ln a P Q0 u
Z
Then the set Z0 “ tz P Z ; λz “ 0u satisfies µpZ0 q ą 0. The proof of this statement uses
the same argument as the one employed in the proof of Theorem 6.11. It involves the rep/ eipln aqλz fz
resentation ρ of the group pp0, `8q, ˆq on L2 pZ, µq defined by ρpaqf : z ✤
2
for every a ą 0 and every pfz qzPZ P L pZ, µq, and uses the obvious fact that since Q0 is a
Kazhdan set in R, ta ; ln a P Q0 u is a Kazhdan set in pp0, `8q, ˆq.
Remark 6.15. — Facts 6.10 and 6.13 have played a crucial role in the proofs of Theorems
6.11 and 6.14 respectively, as they allowed us to discard all irreducible representations
except the one-dimensional ones in inequalities of the form (6.2). In groups with the HoweMoore property (see for instance [23], [39] or [7] for the definition and for more about this
property), all non-trivial irreducible representations have the vanishing property of the
matrix coefficients stated in Facts 6.10 or 6.13. It easily follows from this observation that
all subsets with non-compact closure are Kazhdan sets in groups with the Howe-Moore
property, and that if the group is additionally supposed not to have Property (T), the
Kazhdan sets are exactly the sets with non-compact closure. As SL2 pRq is a non-compact
KAZHDAN SETS AND EQUIDISTRIBUTION PROPERTIES
27
connected simple real Lie group with finite center, it has the Howe-Moore property. But
it does not have Property (T), and so we have:
Example 6.16. — The Kazhdan sets in SL2 pRq are exactly the subsets of SL2 pRq with
non-compact closure.
These observations testify of the rigidity of the structure of groups with the Howe-Moore
property, and stand in sharp contrast with all the examples we have presented in the rest
of this section.
Appendix A
Infinite tensor products of Hilbert spaces
We briefly describe in this appendix some constructions of tensor products of infinite
families of Hilbert spaces, and of tensor products of infinite families of unitary representations. These last objects play an important role in the proof of Theorem 2.3. We
review here the properties and results which we need, following the original works of von
Neumann [32] and Guichardet [18].
A.1. The complete and incomplete tensor products of Hilbert spaces. — The
original construction of the complete and incomplete tensor products of a family pHα qαPI
of Hilbert spaces is due to von Neumann [32]. It was later on taken up by Guichardet in
[18] under a somewhat different point of view, and the incomplete tensor products of von
Neumann are rather known today as the Guichardet tensor products of Hilbert spaces.
Although these constructions can be carried out starting from an arbitrary family pHα qαPI
of Hilbert spaces, we will present them here only in the case of a countable family pHn qně1
of (complex) Hilbert spaces.
Â
The complete infinite tensor product
ně1 Hn of the Hilbert spaces Hn is defined in
[32, Part II, Ch. 3] in the following way: the elementary infinite tensor products are the
elements
x “ bně1 xn , where xn belongs to Hn for each n ě 1 and the infinite product
ś
||x
n || is convergent in the sense of [32, Def. 2.2.1], which by
ně1
ř [32, Lem. 2.4.1] is
equivalent to the fact that either xn “ 0 for some n ě 1 or the series ně1 maxp||xn ||´1, 0q
is convergent. Sequences pxn qně1 with this property are called by von Neumann in [32]
C-sequences. A scalar product is then defined on the set of finite linear combinations of
elementary tensor products by setting
ź
x, y y “
xx
xxn , yn y
ně1
for any elementary tensor products x “ bně1 xn and y “ bně1 yn , and extending the
definition by linearity to finite linear combinations of such elements. The product definx, y y for two
ing xx
ś elementary vectors x and y is quasi-convergent in the sense of [32,
Def.
ś 2.5.1], i.e. ně1 |xxn , yn y| is convergent. The value of this quasi-convergent product
is ně1 xxn , yn y if the product is convergent in the usual sense, and 0 if it is not.
That this is indeed a scalar product which turns the set of finite linear combinations of
elementary tensor products into a complex prehilbertian space is proved in [32,
śLem. 3.21
x
and Theorem
II].
For
any
elementary
tensor
product
x
“
b
x
,
||x
||
“
ně1 n
ně1 ||xn ||.
Â
The space ně1 Hn is the completion of this space for the topology induced by the scalar
product. It is always non-separable.
28
CATALIN BADEA & SOPHIE GRIVAUX
The incomplete tensor products are closed subspaces of the complete tensor product.
They are defined by von Neumann using an equivalence relation betweenřsequences
pxn qně1
ˇ
ˇ
ˇ
of vectors with xn P Hn for each n ě 1 and such that the series ně1 1 ´ ||xn ||ˇ is
convergent. Such sequences ś
are called C0 -sequences. They are C-sequences,
Âand if pxn qně1
is a C-sequence such that ně1 ||xn || ą 0 (i. e. pxn qně1 is non-zero in
ně1 Hn ) then
pxn qně1
If pxn qně1 is a C0 -sequence, pxn qně1 is bounded, and the series
ˇ
ˇ is a C02-sequence.
ř
ˇ1 ´ ||xn || ˇ is convergent.
ně1
ˇ
ř ˇ
Two C0 -sequences pxn qně1 and pyn qně1 are equivalent if the series ně1 ˇ1 ´ xxn , yn yˇ
is convergent. If A denotes an equivalence
ÂA class of C0 -sequences for this equivalence
relation,
the incomplete tensor product
ně1 Hn associated to A is the closed linear span
Â
in ně1 Hn of the vectors x “ bně1 xn , where pxn qně1 belongs
ÂA to A [32, Def.
ÂA1 4.1.1]. If
A and A1 are two different equivalence classes, the spaces
H
and
ně1 Hn are
ně1 n
ÂA
orthogonal, and the linear span of the incomplete tensor products
ně1 Hn , where A
runs
over
all
equivalence
classes
of
C
-sequences,
is
dense
in
the
complete
tensor product
0
Â
ně1 Hn .
Â
If A is an equivalence class of C0 -sequences, A
ně1 Hn admits another, more transparent
description, which runs as follows [32, Lem. 4.1.2], see also [18, Rem. 1.1]: let pan qně1 be
a sequence with an P Hn and ||an || “ 1 for every n ě 1, such that the equivalence class of
pan qně1 is A (such a sequence pan qně1 does exist: if pxn qně1 is any non-zero C0 -sequence
belonging to A, xn is non-zero for every n ě 1, and we can define a C0 -sequence pan qně1
by setting an “ xn {||xn || for every n ě 1. It is not difficult to check that pan qně1 is
ÂA
equivalent to px
ně1 Hn coincides with the closed
Ân qně1 , and so belongs to A). Then
linear span in ně1 Hn of vectors x “ bně1 xn , where xn “ an for all but finitely many
integers n ě 1. Denoting the vector bně1 an by a , we write this closed linear span as
Âa
Âa
ÂA
and thus
ně1 Hn (see [18]),
ně1 Hn “
ně1 Hn , where A is the equivalence class
Âa
of a . The space
H
is
usually
called
the
Guichardet tensor product of the spaces
ně1 n
Hn associated to the sequence pan qně1 . Proposition 1.1 of [18] states the following, which
is a direct consequence of the discussion
above: if x “ pxn qně1 is a C0 -sequence which
Âa
is equivalent to a , x belongs to
H
ně1 n . Vectors x of this form are also said to be
decomposable with respect to a , while vectors x “Â
pxn qně1 with xn “ an for all but finitely
a
Hn .
many indices n are called elementary vectors of ně1
Suppose that all the spaces Hn , n ě 1, are separable. For each n ě 1, let pep,n q1ďpďpn be
a Hilbertian basis of Hn , withÂ
1 ď pn ď `8 and e1,n “ an . The family of all elementary
a
vectors e β “ bně1 eβpnq,n of
ně1 Hn , where β is a map from N into itself such that
1 ď βpnq ď pn for every
ě 1 and βpnq “ 1 for all but finitely many
 aintegers n ě 1, forms
Ân
a
a Hilbertian basis of
H
[32,
Lem.
4.1.4].
In
particular,
ně1 n
ně1 Hn is a separable
complex Hilbert space.
A.2. Tensor products of unitary representations. — Let G be a topological group,
and let pHn qně1 be a sequence of complex separable Hilbert spaces. Let pan qně1 be a
sequence of vectors with an P Hn and ||an || “ 1 for every n ě 1. We Â
are looking for
a
conditions under which one can define a unitary representation π of G on ně1
Hn which
satisfies
(A.1)
π pgq bně1 xn “ bně1 πn pgqxn
for every g P G and every decomposable vector x “ bně1 xn with respect to a . Observe
that without any assumption, the equality π pgq bně1 xn “ bně1 πn pgqxn does not make
KAZHDAN SETS AND EQUIDISTRIBUTION PROPERTIES
29
any sense, since pπn pgqxÂ
n qně1 , which is a C0 -sequence, may not be equivalent to a , and
a
thus may not belong to ně1
Hn .
Infinite tensor products of unitary representations have already been studied in various
contexts (see for instance [4] and the references therein). In [4, Prop. 2.3], the following
observation is made: suppose that, for each n ě Â
1, Un is a unitary operator on Hn . Then
a
there exists a unitary operator U “ bně1 Un on ně1
Hn satisfying
˘
`
U bně1 xn “ bně1 Un xn
for every
decomposable
vector x “ bně1 xn with respect to a if and only if the series
ˇ
ˇ
ř
ˇ1 ´ xUn an , an yˇ is convergent (which is equivalent to requiring that the C0 -sequence
ně1
pUn an qně1 be equivalent to pan qně1 , i. e. to the fact that bně1 Un an be a decomposable
vector
 a with respect to a ). It follows from this result that the formula (A.1) makes sense
in ně1
Hn if and only if the series
ÿˇ
ˇ
ˇ1 ´ xπn pgqan , an yˇ
(A.2)
ně1
is convergent
for every g P G. Under this condition π pgq “ bně1 πn pgq is a unitary operator
Âa
Hn for every g P G, and π pghq “ π pgq π phq for every g, h P G.
on ně1
If the group G is discrete, this tensor product representation is of course automatically
strongly continuous. It is also the case if G is supposed to be locally compact.
Proposition
A.1. — Suppose that G is a locally compact group, and that the series
ř
π “ bně1 πn is strongly
ně1 |1 ´ xπn pgqan , an y| is convergent for every g P G. Then
Âa
continuous, and is hence a unitary representation of G on ně1 Hn .
Âa
Proof. — Since all the spaces Hn , n ě 1, are separable,
ně1 Hn is separable too, and
/ xπ
π pgq ξ , ξ y is a measurable map from
by [6, Lem. A.6.2] it suffices Â
to show that g ✤
a
G into C for every vector ξ P ně1
Hn . Since the linear span of the elementary vectors is
Âa
dense in ně1 Hn , standard arguments show that it suffices to prove this for elementary
Âa
/ xπn pgqxn , xn y is continuous on
vectors x “ bně1 xn of ně1
Hn . Since ś
each map g ✤
✤
/
π pgqx
x , x y “ ně1 xπn pgqxn , xn y is measurable on G.
xπ
G, it is clear that g
In the general case one needs to impose an additional condition on the representations
πn and on the vectors an in order that π be a strongly continuous representation of G on
Â
a
ně1 Hn .
ˇ
ˇ
ř
Proposition A.2. — Suppose that the series ně1 ˇ1 ´ xπn pgqan , an yˇ is convergent for
ˇ
ˇ
ř
ˇ1 ´ xπn pgqan , an yˇ is continuous on a
/
every g P G and that the function g ✤
ně1
neighborhood of the identity element e of G. Then
 a π “ bně1 πn is strongly continuous,
and is hence a unitary representation of G on ně1 Hn .
Proof
 a of Proposition A.2. — Since the linear span of the elementary vectors is dense in
Hn , and the operators π pgq, g P G, are unitary, it suffices to prove that the map
✤ně1 /
x is continuous at e for every elementary vector x “ bně1 xn of norm 1 of
g
π pgqx
Âa
H
.
Let
N ě 1 be such that xn “ an for every n ą N . We have for every g P G:
ně1 n
ˆ
˙
ź
xn
xn
π pgqx
x ´ x ||2 “ 2p1 ´ Rexπ
π pgqx
x , x yq “ 2 1 ´
||π
Rexπn pgq
,
y
||xn || ||xn ||
ně1
CATALIN BADEA & SOPHIE GRIVAUX
30
x|| “
since ||x
ź
ně1
||xn ||=1. Thus
π pgqx
x ´ x ||2 ď 2
||π
ÿ ˇˇ
xn
xn ˇˇ
,
yˇ
ˇ1 ´ xπn pgq
||xn || ||xn ||
ně1
N ˇ
ÿ
ÿ
xn ˇˇ
xn
ˇ
,
yˇ ` 2
|1 ´ xπn pgqan , an y|.
ď2
ˇ1 ´ xπn pgq
||xn || ||xn ||
ně1
n“1
π pgqx
x ´ x || ă ε if g lies
If ε is any positive number, it follows from the assumptions that ||π
/ π pgqx
x.
in a suitable neighborhood of e. This proves the continuity of the map g ✤
We finish this appendix Â
by giving a sufficient condition for an infinite tensor product
a
representation on a space
ně1 Hn to be weakly mixing: let, for each n ě 1, Hn be a
separable Hilbert space, an a vector of Hn with ||an || “ 1, and πn a unitary representation
of G on Hn . We suppose that the assumptions of either Proposition A.1 (when G is locally
compact) or Proposition A.2 (inÂthe general case) are satisfied, so that π “ bně1 πn is a
a
unitary representation of G on ně1
Hn . Then
Proposition A.3. — In the case where lim nÑ`8 mp|xπn p qan , an y|2 q “ 0, the representation π “ bně1 πn is weakly mixing.
Proof. — The proof of Proposition A.3 relies on the same idea as that Â
of Proposition
a
A.2: let x “ bně1 xn and y “ bně1 yn be two elementary vectors in
ně1 Hn with
x|| “ ||yy || “ 1. We have
||x
ź ˇˇA
xn
yk Eˇˇ2 ˇˇA
yn Eˇˇ2
xk
π pgq x , y y|2 “
,
,
|xπ
ˇ ď ˇ πn pgq
ˇ
ˇ πk pgq
||xk || ||yk ||
||xn || ||yn ||
kě1
for every n ě 1 and every g P G. But
ˇˇ ˇˇ
ˇˇ
ˇˇ
ˇA
ˇˇ ˇˇ yn
ˇˇ
ˇˇ xn
xn
yn Eˇˇ
ˇ
,
´ an ˇˇˇˇ ` ˇˇˇˇ
´ an ˇˇˇˇ.
ˇ πn pgq
ˇ ď |xπn pgqan , an y| ` ˇˇˇˇ
||xn || ||yn ||
||xn ||
||yn ||
Squaring and taking the mean on both sides we obtain that
ˇˇ
ˇˇ2
ˇˇ2
ˇˇ
ˇ ˇ xn
ˇˇ
ˇˇ
ˇˇ yn
2
2
ˇ
ˇ
ˇ
ˇ
ˇ
ˇ
π p qx
x , y y| q ď 4 mp|xπn p q an , an y| q ` 4 ˇˇ
mp|xπ
´ an ˇˇ ` 4 ˇˇ
´ an ˇˇˇˇ
||xn ||
||yn ||
for every n ě 1. Since lim nÑ`8 mp|xπn p qan , an y|2 q “ 0 and the two other terms are equal
π p qx
x , y y|2 q “ 0. Weak mixing of π now follows from
to zero for n sufficiently large, mp|xπ
standard density arguments.
References
[1] M. Anoussis, A. Bisbas, Continuous measures on compact Lie groups, Ann. Inst.
Fourier 50 (2000), p. 1277–1296.
[2] M. Ballotti, Convergence rates for Wiener’s theorem for contraction semigroups, Houston J. Math. 11 (1985), p. 435–445.
[3] M. Ballotti, J. Goldstein, Wiener’s theorem and semigroups of operators, in: Infinite
dimensional systems (Retzhof, 1983), Lecture Notes in Math. 1076 (1984), Springer, p. 16–
22.
KAZHDAN SETS AND EQUIDISTRIBUTION PROPERTIES
31
[4] E. Bedos, R. Conti, On infinite tensor products of projective unitary representations,
Rocky Mountain J. Math. 34 (2004), p. 467–493.
[5] B. Bekka, A. Valette, Kazhdan’s Property (T) and amenable representations, Math.
Zeit. 212 (1993), p. 293–299.
[6] B. Bekka, P. de la Harpe, A. Valette, Kazhdan’s Property (T), New Mathematical
Monographs 11 (2008), Cambridge University Press.
[7] B. Bekka, M. Mayer, Ergodic theory and topological dynamics of group actions on
homogeneous spaces, London Mathematical Society Lecture Note Series 269 (2000), Cambridge University Press.
[8] V. Bergelson, J. Rosenblatt, Mixing actions of groups, Illinois J. Math. 32 (1998),
p. 65–80.
[9] M. Björklund, A. Fish, Continuous measures on homogenous spaces, Ann. Inst.
Fourier 59 (2009), p. 2169–2174.
[10] R. Burckel, Weakly almost periodic functions on semigroups, Gordon and Breach
(1970).
[11] M. Drmota, R. Tichy, Sequences, discrepancies and applications, Lecture Notes in
Mathematics 1651 (1997), Springer-Verlag.
[12] H. Dye, On the ergodic mixing theorem, Trans. Amer. Math. Soc. 118 (1965), p. 123–130.
[13] T. Eisner, S. Grivaux, Hilbertian Jamison sequences and rigid dynamical systems, J.
Funct. Anal. 261 (2011), p. 2013–2052.
[14] B. Farkas, Wiener’s lemma and the Jacobs-de Leeuw-Glicksberg decomposition, Ann.
Univ. Sci. Budapest, Eötvös Sect. Math. 58 (2015), p. 27–35.
[15] G. Folland, A course in abstract harmonic analysis, Studies in Advanced Mathematics
(1995), CRC Press.
[16] E. Glasner, Ergodic theory via joinings, Mathematical Surveys and Monographs 101
(2003), American Mathematical Society.
[17] H. Gröchenig, V. Losert, H. Rindler, Uniform distribution in solvable groups,
Probability measures on groups VIII (Oberwolfach, 1985), Lecture Notes in Math. 1210
(1986), Springer, p. 97–107.
[18] A. Guichardet, Produits tensoriels infinis et représentations des relations d’anticommutation, Ann. Sci. É.N.S. 83 (1966), p. 1–52.
[19] P. de la Harpe, A. Valette, La propriété (T) de Kazhdan pour les groupes localement compacts (avec un appendice de Marc Burger), Astérisque 175 (1989), Societé
Mathématique de France.
[20] S. Hartman, The method of Grothendieck-Ramirez and weak topologies in C(T), Studia
Math. 44 (1972), p. 181–197.
[21] S. Hartman, On harmonic separation, Colloq. Math. 42 (1979), p. 209–222.
[22] B. Host, J.-F. Méla, F. Parreau, Non-singular transformations and spectral analysis
of measures, Bull. Soc. Math. France 119 (1991), p. 33–90.
[23] R. Howe, C. Moore, Asymptotic properties of unitary representations, J. Funct. Anal.
32 (1979), p. 72–96.
[24] R. Kaufman, Remark on Fourier-Stieltjes transforms of continuous measures, Colloq.
Math. 22 (1971) p. 279–280.
[25] R. Kaufman, Continuous measures and analytic sets, Colloq. Math. 58 (1989), p. 17–21.
[26] D. Kazhdan, Connection of the dual space of a group with the structure of its closed
subgroups, Func. Anal. Appl. 1 (1967), p. 63–65.
[27] D. Kerr, H. Li, Ergodic Theory: Independence and dichotomies, Springer Monographs
in Math., Springer (2016).
[28] L. Kuipers, H. Niederreiter, Uniform distribution of sequences, Pure and Applied
Mathematics, Wiley-Interscience (1974).
32
CATALIN BADEA & SOPHIE GRIVAUX
[29] V. Losert, H. Rindler, Uniform distribution and the mean ergodic theorem, Inv. Math.
50 (1978), p. 65–74.
[30] B. De Mathan, Numbers contravening a condition in density modulo 1, Acta Math.
Acad. Sci. Hungar. 36 (1980), p. 237–241.
[31] C. Moore, Groups with finite dimensional irreducible representations, Trans. Amer. Soc.
166 (1972), p. 401–410.
[32] J. von Neumann, On infinite direct products, Compositio Math. 6 (1938), p. 1–77.
[33] T. Palmer, Classes on nonabelian, noncompact, locally compact groups, Rocky Mountain
J. of Math. 8 (1978), p. 683–741.
[34] J. Peterson, Lecture notes on ergodic theory, available at http://www.math.
vanderbilt.edu/peterson/teaching/Spring2011/ErgodicTheoryNotes.pdf
[35] A. D. Pollington, On the density of sequences pnk ξq, Illinois J. Math 23 (1979),
p. 511–515.
[36] M. Taylor, Noncommutative harmonic analysis, Mathematical Surveys and Monographs
22 (1986), American Mathematical Society.
[37] W. Veech, Some questions of uniform distribution, Ann. of Math. 94 (1971), p. 125–138.
[38] W. Veech, Topological dynamics, Bull. Amer. Math. Soc. 83 (1977), p. 775–830.
[39] R. Zimmer, Ergodic theory and semisimple groups, Monographs in Mathematics 81
(1984), Birkhäuser Verlag.
Catalin Badea, Université Lille 1, Laboratoire Paul Painlevé, CNRS UMR 8524, Bât. M2, 59655
Villeneuve d’Ascq Cedex, France ‚ E-mail : [email protected]
Sophie Grivaux, CNRS, Laboratoire Amiénois de Mathématique Fondamentale et Appliquée, UMR
7352, Université de Picardie Jules Verne, 33 rue Saint-Leu, 80039 Amiens Cedex 1, France
E-mail : [email protected]
| 4math.GR
|
arXiv:1710.05735v1 [math.NA] 12 Oct 2017
Nonnegative Polynomial with no Certificate of
Nonnegativity in the Simplicial Bernstein Basis
Christoffer Sloth
University of Southern Denmark
SDU Robotics
[email protected]
Abstract
This paper presents a nonnegative polynomial that cannot be represented with nonnegative coefficients in the simplicial Bernstein basis by
subdividing the standard simplex. The example shows that Bernstein
Theorem cannot be extended to certificates of nonnegativity for polynomials with zeros at isolated points.
1
Introduction
The positivity and nonnegativity of polynomials play an important role in many
problems such as Lyapunov stability analysis of dynamical systems [5, 7]. Therefore, much research has been conducted to exploit different certificates of positivity [9, 1]. The certificates of positivity include sum of squares that is extensively
used in different applications [5, 10].
The certification of nonnegativity for polynomials with zeros was addressed
in [3, 11], where the paper [3] characterizes polynomials that can be certified to
be nonnegative using Pólya’s Theorem.
This paper exploits a certificate of positivity in the Bernstein basis [4] that
has previously been applied within optimization and control [8, 2, 12]. In particular, a nonnegative polynomial is provided that cannot be represented with
nonnegative coefficients in the simplicial Bernstein basis by subdividing the
standard simplex.
The paper is organized as follows. Section 2 provides preliminaries on polynomials in Bernstein form, which are used in Section 3 that provides the example, and relies on results presented in Appendix A.
2
Preliminaries
This section provides preliminary results and notation on polynomials in the
Bernstein basis.
Let R[X] be the real
ring in n variables. For α = (α0 , . . . , αn ) ∈
Ppolynomial
n
Nn+1 we define |α| := i=0 αi .
We utilize Bernstein polynomials defined on simplices via barycentric coordinates.
1
Definition 1 (Barycentric Coordinates). Let λ0 , . . . , λn ∈ R[X] be affine polynomials and let v0 , . . . , vn ∈ Rn be affinely independent. If
n
X
λi = 1
and
i=0
x = λ0 (x)v0 + · · · + λn (x)vn
∀x ∈ Rn
then λ0 , . . . , λn are said to be barycentric coordinates associated to v0 , . . . , vn .
Let λ0 , . . . , λn be barycentric coordinates associated to v0 , . . . , vn . Then the
simplex generated by λ0 , . . . , λn is
4 := {x =
n
X
λi vi | λi (x) ≥ 0, i = 0, . . . , n} ⊂ Rn .
(1)
i=0
This work represents polynomials in the simplicial Bernstein basis, where
the basis polynomials are defined as follows.
Definition 2 (Bernstein Polynomial). Let d ∈ N0 , n ∈ N, α = (α0 , . . . , αn ) ∈
Nn+1
, and let λ0 , . . . , λn ∈ R[X] be barycentric coordinates associated to a
0
simplex ∆. The Bernstein polynomials of degree d on the simplex ∆ are
d
d
Bα =
λα
(2)
α
for |α| = d where
d!
d
=
α
α0 !α1 ! · · · αn !
and
λα :=
n
Y
i
λα
i .
i=0
Every polynomial P ∈ R[X] of degree no greater than d can be written
uniquely as
X
X
P =
bα (P, d, 4)Bαd =
bdα Bαd = bd B d ,
(3)
|α|=d
|α|=d
where bα (P, d, 4) ∈ R is a Bernstein coefficient of P of degree d with respect to
4. We say that (3) is a polynomial in Bernstein form.
From the definition of Bernstein polynomials on simplex 4 (Definition 2)
and (1) it is seen that for any d ∈ N and |α| = d
Bαd (x) ≥ 0
∀x ∈ 4.
This means that polynomial P in (3) is positive on 4 if bdα > 0 for all α ∈ Nn+1
where |α| = d. We call this a certificate of positivity. Similarly, we say that
bdα ≥ 0 for all α ∈ Nn+1 where |α| = d is a certificate of nonnegativity.
If a polynomial P of degree p cannot be certified to be positive in the Bernstein basis of degree p on simplex 4 then two strategies may be exploited to
refine the certificate. These are called degree elevation and subdivision and will
be outlined next.
To certify the positivity of a polynomial by degree elevation, one may represent the polynomial using Bernstein polynomials of a higher degree as in the
following theorem [6].
2
Theorem 1 (Multivariate Bernstein Theorem). If a nonzero polynomial P ∈
R[X1 , . . . , Xn ] of degree p ∈ N0 is positive on a simplex 4, then there exists
d ≥ p such that all entries of b(P, d, 4) are positive.
A similar result related to subdivision is presented in [6]; however, for simplicity the subdivision approach is explained by the following example. The
example shows how an initial simplex 4 may be subdivided into several simplices to certify the positivity of a polynomial on a given domain.
Example 1. Consider the following polynomial
P = X 2 + Y 2 − XY
on the standard simplex 4 with vertices v0 = (0, 0), v1 = (1, 0), v2 = (0, 1).
Polynomial P has the following nonzero Bernstein coefficients in the Bernstein
basis with respect to 4
b(0,2,0) = 1,
b(0,1,1) = −0.5,
b(0,0,2) = 1.
(4)
The nonnegativity of polynomial P cannot be certified directly by the above Bernstein coefficients since b(0,1,1) is negative; hence, simplex 4 must be subdivided
to obtain a certificate of nonnegativity. Consequently, we define a new vertex
ṽ1 = (1 − θ, θ) and a subdivision of 4 given by two simplices 4̃ and 4̂ with vertices v0 , v1 , ṽ1 and v0 , v2 , ṽ1 respectively. The subdivision is shown in Figure 1
and gives a collection of simplices, all with a common vertex at the zero of P .
ṽ1
v2
4̂
v1
4̃
v0
Figure 1: Simplex 4 with vertices v0 , v1 , v2 , simplex 4̃ with vertices v0 , v1 , ṽ1 ,
and simplex 4̂ with vertices v0 , v2 , ṽ1 .
The nonzero Bernstein coefficients of polynomial P on the two simplices are
(by Lemma 1)
1
(1 − θ)θ +
8
1
= (1 − θ)θ +
8
b̃(0,2,0) = 1,
b̃(0,1,1) = −1.5θ + 1,
b̃(0,0,2) =
b̂(0,2,0) = 1,
b̂(0,1,1) = 1.5θ − 0.5,
b̂(0,0,2)
1
4
1
4
It is seen that the nonnegativity of the polynomial can be certified by letting
θ = 0.5.
It should be noted that there exists nonnegative (sum of squares) polynomials that cannot be represented with nonnegative coefficients in the simplicial
Bernstein basis, i.e., it is not possible to certify the nonnegativity of any nonnegative polynomial by employing subdivisioning. This is exemplified in the
next section.
3
3
Counterexample
In this section, a nonnegative polynomial is presented for which there exists no
certificate of nonnegativity in the simplicial Bernstein basis.
Consider the polynomial
P = 21X14 +24X13 X2 −36X13 +18X12 X22 −24X12 X2 +18X12 +12X1 X23 −12X1 X22 +30X24 .
The graph of P on the standard simplex is shown in Figure 2.
30
20
10
0
1
1
0.5
x2
0.5
0
0
x1
Figure 2: Graph of the polynomial P on the standard simplex 4.
We aim at proving the nonnegativity of P on the standard simplex 4 with
vertices v0 = (0, 0), v1 = (1, 0), and v2 = (0, 1). The polynomial P can be
written as the following sum of squares
X1
18 −18 −12 −6
2
−18 21
12
0
X1 ,
P = X1 X12 X1 X2 X22
−12 12
18
6 X1 X2
X22
−6
0
6
30
|
{z
}
=MP
where the matrix MP is positive definite; hence, the polynomial is positive
everywhere except from being zero at (x1 , x2 ) = (0, 0).
Polynomial P can be represented in the Bernstein basis as
X
P =
bα Bαd ,
|α|=d
where Bαd is the Bernstein basis defined on the standard simplex 4. The nonzero
coefficients of P are
b(2,2,0) = 3,
b(1,2,1) = 1,
b(1,1,2) = −1,
b(0,4,0) = 3,
b(0,0,4) = 30.
It is seen that one coefficient is negative (b(1,1,2) < 0); thus, the nonnegativity
of P cannot be directly certified.
Next we show that there exists no certificate of nonnegativity for the polynomial P in the Bernstein basis, since the coefficient with index (1, 1, 2) will
remain negative on one simplex of any triangulation of 4. In particular, the
coefficient remains negative on any nondegenerate simplex with vertex v0 , a
4
vertex on the line between v0 and v2 , and any third vertex in 4, i.e., for any
Bernstein basis defined on the simplex 4̃ with vertices
ṽ0 = v0
ṽ1 = β0 v0 + β1 v1 + β2 v2
ṽ2 = ρv0 + (1 − ρ)v2
where β0 + β1 + β2 = 1, β0 , β1 , β2 ≥ 0, β1 > 0, and ρ ∈ [0, 1). The simplices 4
and 4̃ are illustrated in Figure 3.
v1
v2
ṽ1
ṽ2
v0
Figure 3: Simplex 4 with vertices v0 , v1 , v2 and simplex 4̃ with vertices
v0 , ṽ1 , ṽ2 .
The polynomial P can be represented in the Bernstein basis as
X
P =
b̃α B̃αd ,
|α|=d
where B̃αd is the Bernstein basis defined on the standard simplex 4̃ and where
by Lemma 2 and Lemma 1 the coefficients bα and b̃γ are related as
b̃γ =
γ2
X
γ2
ργ2 −k (1−ρ)k
γ2 − k
k=0
α −(γ +γ −k)
X
|α| = d
γ0 + γ2 − k ≤ α0
k ≤ α2
γ1 !β0 0 0 2
β1α1 β2α2 −k
bα
(α0 − (γ0 + γ2 − k))!α1 !(α2 − k)!
For the polynomial P , it is seen that
b̃(1,1,2) = β1 (1 − ρ)2 b(1,1,2) ,
which is negative for any admissible β1 and ρ (β1 > 0 and ρ ∈ [0, 1)), since
b(1,1,2) = −1. Therefore, there exists no certificate of nonnegativity of P on 4
in the simplicial Bernstein basis.
References
[1] A. Ahmadi and A. Majumdar. DSOS and SDSOS optimization: LP and
SOCP-based alternatives to sum of squares optimization. In proceedings of
the 48th Annual Conference on Information Sciences and Systems, pages
1–5, March 2014.
5
[2] M. Ben Sassi, R. Testylier, T. Dang, and A. Girard. Reachability analysis
of polynomial systems using linear programming relaxations. In Automated
Technology for Verification and Analysis, pages 137–151. Springer Berlin
Heidelberg, 2012.
[3] M. Castle, V. Powers, and B. Reznick. Pólya’s theorem with zeros. Journal
of Symbolic Computation, 46(9):1039–1048, 2011.
[4] R. T. Farouki. The Bernstein polynomial basis: A centennial retrospective.
Computer Aided Geometric Design, 29(6):379 – 419, 2012.
[5] D. Henrion and A. Garulli, editors. Positive Polynomials in Control, volume
312 of Lecture Notes in Control and Information Science. Springer Berlin
Heidelberg, 2005.
[6] R. Leroy. Certificates of positivity in the simplicial Bernstein basis. hal00589945, 2009.
[7] M. Marshall. Positive Polynomials and Sums of Squares, volume 146.
American Mathematical Society, 2008.
[8] C. Muoz and A. Narkawicz. Formalization of bernstein polynomials and
applications to global optimization. Journal of Automated Reasoning,
51(2):151–196, 2013.
[9] P. A. Parrilo. Semidefinite programming relaxations for semialgebraic problems. Math. Programming, 96(2):293–320, 2003.
[10] V. Powers. Positive polynomials and sums of squares: Theory and practice.
Real Algebraic Geometry, 2011.
[11] M. Schweighofer. Certificates for nonnegativity of polynomials with zeros
on compact semialgebraic sets. manuscripta mathematica, 117(4):407–428,
Aug 2005.
[12] C. Sloth and R. Wisniewski. Control to facet for polynomial systems.
In Proceedings of the 17th International Conference on Hybrid Systems:
Computation and Control, pages 123–132, 2014.
A
Useful Results
Lemma 1. Let P ∈ R[X]. Define a simplex 4 with vertices v0 , v1 , v2 ∈ R2 ,
and define the simplex 4̃ with vertices v0 , v1 , ṽ2 , where
ṽ2 = ρv0 + (1 − ρ)v2
and ρ ∈ [0, 1). Then P can be represented in two Bernstein bases as
X
X
P =
bα Bαd =
b̃γ B̃γd ,
|α|=d
(5)
|γ|=d
where Bαd is the Bernstein basis defined by 4, B̃γd is the Bernstein basis defined
by 4̃, and
γ2
X
γ2
b̃γ =
ργ2 −k (1 − ρ)k b(γ0 +γ2 −k,γ1 ,k) .
γ2 − k
k=0
6
Proof. Let λ0 , λ1 , λ2 ∈ R[X] be barycentric coordinates associated to 4 and
λ̃0 , λ̃1 , λ̃2 ∈ R[X] be barycentric coordinates associated to 4̃. Then
λ0 = λ̃0 + ρλ̃2
λ1 = λ̃1
λ2 = (1 − ρ)λ̃2
and
α2 α2
1
λα = (λ̃0 + ρλ̃2 )α0 λ̃α
1 (1 − ρ) λ̃2 .
By binomial theorem
α
λ =
α0
X
α0
k=0
=
α0
X
α0
k=0
k
k
!
λ̃0α0 −k ρk λ̃k2
α2 α2
1
λ̃α
1 (1 − ρ) λ̃2
1 α2 +k
ρk (1 − ρ)α2 λ̃0α0 −k λ̃α
.
1 λ̃2
Therefore, the Bernstein bases Bγd and B̃γd are related as
Bαd
=
α0
X
α2 + k
k=0
k
d
ρk (1 − ρ)α2 B̃(α
,
0 −k,α1 ,α2 +k)
which implies that coefficients bα and b̃α are related as
γ2
X
γ2
b̃γ =
ργ2 −k (1 − ρ)k b(γ0 +γ2 −k,γ1 ,k) .
γ2 − k
k=0
Lemma 2. Let P ∈ R[X]. Define a simplex 4 with vertices v0 , v1 , v2 ∈ R2 ,
and define the simplex 4̃ with vertices v0 , ṽ1 , v2 ∈ R2 , where
ṽ1 = β0 v0 + β1 v1 + β2 v2
and β0 + β1 + β2 = 1, β0 , β2 ≥ 0, β1 > 0. Then P can be represented in two
Bernstein bases as
X
X
P =
bα Bαd =
b̃γ B̃γd
(6)
|α|=d
|γ|=d
where Bαd is the Bernstein basis defined by 4, B̃γd is the Bernstein basis defined
by 4̃, and
b̃γ =
X
|α| = d
α0 ≥ γ0
α2 ≥ γ2
γ1 !
β α0 −γ0 β1α1 β2α2 −γ2 bα .
(α0 − γ0 )!α1 !(α2 − γ2 )! 0
7
Proof. Let λ0 , λ1 , λ2 ∈ R[X] be barycentric coordinates associated to 4 and
λ̃0 , λ̃1 , λ̃2 ∈ R[X] be barycentric coordinates associated to 4̃. Then
λ0 = λ̃0 + β0 λ̃1
λ1 = β1 λ̃1
λ2 = β2 λ̃1 + λ̃2
and
λα = (λ̃0 + β0 λ̃1 )α0 (β1 λ̃1 )α1 (β2 λ̃1 + λ̃2 )α2 .
By binomial theorem
α0
X
α0
λα =
k=0
X
=
k
|γ| = d
γ0 ≤ α0
γ2 ≤ α2
!
0 −k k k
λ̃α
β0 λ̃1
0
1
β1α1 λ̃α
1
α2
X
α2
j=0
α0
α0 − γ0
j
2 −j
β2j λ̃j1 λ̃α
2
α2
β0α0 −γ0 β1α1 β2α2 −γ2 λ̃γ .
α2 − γ2
Therefore, the Bernstein bases Bγd and B̃γd are related as
Bαd =
X
|γ| = d
γ0 ≤ α0
γ2 ≤ α2
γ1 !
β α0 −γ0 β1α1 β2α2 −γ2 B̃γd ,
(α0 − γ0 )!α1 !(α2 − γ2 )! 0
which implies that coefficients bα and b̃α are related as
b̃γ =
X
|α| = d
α0 ≥ γ0
α2 ≥ γ2
γ1 !
β α0 −γ0 β1α1 β2α2 −γ2 bα .
(α0 − γ0 )!α1 !(α2 − γ2 )! 0
8
| 3cs.SY
|
A Rule-Based Computational Model of Cognitive Arithmetic
Ashis Pati, Kantwon Rogers and Hanqing Zhu
arXiv:1705.01208v1 [cs.AI] 3 May 2017
Georgia Institute of Technology
Atlanta, Georgia, United States 30318
{ashis.pati,kantwonrogers,williamzhu}@gatech.edu
Abstract
Cognitive arithmetic studies the mental processes used
in solving math problems. This area of research explores the retrieval mechanisms and strategies used by
people during a common cognitive task. Past research
has shown that human performance in arithmetic operations is correlated to the numerical size of the problem.
Past research on cognitive arithmetic has pinpointed
this trend to either retrieval strength, error checking,
or strategy-based approaches when solving equations.
This paper describes a rule-based computational model
that performs the four major arithmetic operations (addition, subtraction, multiplication and division) on two
operands. We then evaluated our model to probe its validity in representing the prevailing concepts observed
in psychology experiments from the related works. The
experiments specifically explore the problem size effect, an activation-based model for fact retrieval, backup
strategies when retrieval fails, and finally optimization
strategies when faced with large operands. From our
experimental results, we concluded that our models response times were comparable to results observed when
people performed similar tasks during psychology experiments. The fit of our model in reproducing these
results and incorporating accuracy into our model are
discussed.
Introduction
Performing arithmetic operations is an elementary skill that
most people learn and acquire at an early age and is used extensively in our day to day lives. Primary school students
spend a significant time during their education practicing
and honing this skill. Cognitive arithmetic can be defined
as the study of the mental representations and cognitive processes which occur when people perform arithmetic in their
heads. A very similar definition provided by Ashcraft in [3]
states that cognitive arithmetic is “concerned with the mental representation of number and arithmetic, and the processes and procedures that access and use this knowledge.”
Study of cognitive arithmetic promises to provide useful insights into the working of our brain. It is an important tool
which bridges the theoretical and practical understanding of
memory [7]. Understanding these underlying mental processes can also help in designing more effective school curriculum to teach arithmetic to students. Hence, this area of
research has attracted a lot of interest among the cognitive
science community, more so from cognitive psychologists
who have performed extensive empirical studies involving
human participants. These studies have been instrumental
in unraveling several important phenomena associated with
cognitive arithmetic and have attempted to provide several
theories to explain them.
Although the theories based on observations from psychological experiments are intuitive, they cannot be validated
easily and sometimes it is difficult to reproduce and extend
them. On the other hand, computational models are powerful tools in cognitive science to evaluate and test the validity
of theories. Computational models are generally scalable allowing detailed experimentation with any number of inputs
while simultaneously allowing precise control of test conditions. An accurate computational model can also be used to
predict human behavior and hence, is extremely useful. For
e.g. experiments after “ablating” the model can be used to
predict the behavior of neuro-psychological patients. Hence,
coming up with a computational model which explains the
theories is essential to any area of study in cognitive science.
In this paper, we first provide a brief discussion on the
previous research done in this area. We will explain several commonly observed phenomena in experimental studies
on cognitive arithmetic and the theories proposed to explain
them. We then draw ideas and data from these experiments
to create computational models of cognitive arithmetic. We
create two basic models: a) A overall model for how cognitive arithmetic is performed, b) A secondary model which
tries to explain how common arithmetic facts are stored in
and retrieved from memory. Subsequently, we design a few
experiments which test our model in an attempt to evaluate
and ascertain their validity. Finally, based on the experimental results, we identify several shortcomings in our models
and discuss possible directions for future research endeavors.
Related Work
Although, cognitive arithmetic covers number theory and
math problems, our work strictly deals with basic arithmetic equations. The body of existing research in cognitive arithmetic follows an approach grounded in cognitive
psychology by employing empirical human-subject research
and computational modeling to explain mathematic cogni-
tive processing. Past empirical research test hypotheses regarding the cognitive mental processes and structures by examining participants response times, error rates and also by
asking the participants to explain their thought process. The
data thus obtained from participants solving arithmetic problems can be used to inform the results from computational
models. While our approach did not include human-subjects
research, we were able to leverage this past work and data
to inform our model. In this section we discuss the empirical research considered for constructing our computational
model.
Psychology Studies
Early research by Groen and Parkman in cognitive arithmetic involved studying the performance of adults completing addition problems [9]. Their work identified the phenomena of “tie-effect” in cognitive arithmetic, wherein it
was observed that people are faster at solving problems with
the same operands, e.g. 4 + 4 = 8. Groen and Parkman later
completed a study with first-grade children. The children in
the study were asked to complete small addition problems
(the sum did not exceed 9). The researchers found that as
the sums increased in size, the response times of the children linearly increased as well. This led them to believe that
children use a “minimum counting” strategy i.e. counting
begins from the larger of the two operands and increments
for the value of the smaller number.
Other researchers at that time contested the countingbased model being the only mechanism for solving math
problems. They instead proposed a hypothesis that adults
use direct access or fact retrieval in order to solve arithmetic equations. Other researchers theorized that counting
and direct access were only part of the cognitive model for
solving arithmetic equations and that sequential, independent stages are used. Ashcraft inspected the theory of direct access and sequential stages by completing an addition
study with adults [1]. The study found a nonlinear increasing response time compared to the sum (when ties were removed), which supported Ashcrafts hypothesis that counting could not be the only mechanism for solving these problems. Ashcraft and Stazyk later conducted tests of their data
with the Groen and Parkman work to generalize the increasing response time in relation to the size of the operands as
the “problem-size effect”. They claimed that in general response time and error rates increase as the problem-size increases, where “size” refers to the sum of the addition problem. This finding has led to a large body of work aimed at
understanding and documenting the response time and error
rates of participants when completing arithmetic problems.
In the Miller and Perlmutter work, the researchers compared the performance of adults in completing addition,
multiplication and numerical comparison problems [13].
Miller and Perlmutter found the relationship of the response
times to support a “network” or sequential model for computing the solutions as opposed to a counting model. The
authors also found evidence for a distinction between location and accessibility of information within a direct access model. This means that fact recall could vary based on
factors such as how often or recently a person has retrieved
that arithmetic fact. Research by Campbell and Xue looked
into the strategies and performance of people across cultures
for all four major math operations [4]. The study confirmed
the importance of procedural knowledge when successfully
completing arithmetic problems. This work correlates participants’ skill in arithmetic to their ability to complete complex arithmetic procedures (e.g. borrowing, carrying, place
holding), and not entirely based on retrieval performance.
Researchers have also studied the processes used when recall fails participants. This area of research has hypothesized a number of backups strategies used to indirectly solve
math problems, however a general consensus has not been
reached due to the variety of responses found during the
studies [12, 8, 5]. This research has shown that people actively use backup strategies, such as counting by multiples,
subtraction by addition, and operand transformation when
they cannot directly access an answer.
Computational Modeling
Initial computational models to explain cognitive arithmetic
were developed by Groen and Parkman using information
processing models for counting [9]. As previously mentioned, the authors found that the response times generated
by “minimum counting” best fit their sample data when
compared to four other counting models. Ashcraft also
accompanied their psychology study with a computational
model based on direct access. Ashcraft developed a network
retrieval model that inspected sequential stages of arithmetic processing in combination with fact retrieval [3]. This
work helped recognize that basic fact retrieval was used in
conjunction with the carry operation process to solve more
complex addition problems (e.g. two digit operations) and
opened up research initiatives in network models based on
the activation and priming of arithmetic facts through network models [14, 11]. For example, Lebiere developed an
ACT-R model of cognitive arithmetic based on the findings
from the psychology experiments to simulate a lifetime of
arithmetic learning [11]. Similar to our own model, Lebieres
uses an activation-based process to learn arithmetic facts.
Proposed Model
For the purpose of this paper, we designed a model which
could solve elementary two-operand operations (addition,
subtraction, multiplication and division). The overall model
has three steps: the input parsing, working memory operations, and output generation. Figure 1 show the overall system as a block digram. The input to the model is a string
that contains the first operand then an operator followed by
the second operand. Available operations are addition, subtraction, multiplication, and division. Once the input has
been parsed, the working memory is tasked with the computation of the problem. The primary mode is memory retrieval. The working memory starts by trying to fetch the
result of an operation from a database of entries we call the
arithmetic fact library. This models how a person who has
memorized an operation retrieves it from his or her memory.
If this retrieval fails, the working memory then attempts to
simplify the problem by breaking it down into smaller processes and continues to check if any parts of the simplified
version are found in the database. Simplification processes
are dependent on the operation and based on common algorithms that humans execute when performing arithmetic. If
the problem can no longer be simplified and the operation is
not available in the database, the system resorts to a counting method. Throughout this entire process, overall response
time is calculated per operation based on reported response
times of humans performing the respective arithmetic task.
After an answer has been produced, the system displays the
result and the overall response time.
Figure 2: Model for Arithmetic Facts Retrieval
Figure 1: Overall Arithmetic Computation Model
For operands that exist in the arithmetic fact library, the
total response time for such a problem is the RT stored in
memory. Based on this assumption, the problem-size effect (where larger sums take longer time) can be explained
only if we can come up with a computational model for
these RTs. So we designed a computational model for the
RTs associated with individual operand pairs using the neural activation model proposed in [10]. We conjecture that
RT for an operand pair is inversely related to the strength
of the memory representation associated with that operand
pair. Assuming a neural model for memory representation,
the strength of the representation associated with an operand
pair would be directly proportional to the neural activation.
Neural activation in return depends on the number of times
the operand pair is presented to the user. In other words, RT
of an operand pair is inversely proportional to the frequency
of occurrence of that pair. The complete model for arithmetic fact RT is shown in Figure 2. Since, we assume that
basic arithmetic facts (operations on single digit operands)
are stored in memory, the response time for those will be
the same as memory RT. Thus, we conjecture that the experimental data reported by Miller et al. [13] for e.g. corresponds to the memory RT.
Experiments
To evaluate the representational power of our model, we
conducted three experiments to compare its fit to the findings
of psychology experiments discussed in the related works.
Problem-Size Experiment
To test our model for arithmetic fact retrieval (Figure 2) we
designed an experiment to model the frequency of occurrence of a pair of numbers while learning multiplication.
Since, most of elementary school students learn single-digit
multiplication via the use of multiplication tables [2], we
decided to use the same in order to determine the frequency
of occurrence of a number pair. We assumed that learning
of multiplication tables is a pseudo-random process wherein
for “trial”, the table being memorized (e.g. table of 2 corresponds to 2x0, 2x1,....till 2x9) is selected randomly. We
call this the table index. Also, every time the entire table is
not memorized and the extent to which a table is memorized
is also random. For example, when learning multiplication
by2, in one instance we go from 2x0, 2x1,....till 2x8 and in
another instance we go from 2x0, 2x1,....till 2x5 only. In
these examples 8 and 5 would be the table extent. Next, we
initialized a random number pair (each number is between
0 to 9 drawn from a uniform distribution). The first number
corresponds to the the table index and the second number
corresponds to the table extent. For e.g. if the we get a pair
(2,5), this means we memorize 2x0, 2x1, .... till 2x5 for this
trial. Therefore, in our memory representation corresponding to multiplication, the locations corresponding to (2,0),
(2,1),...till (2,5) get activated. We also conjecture that the
conjugates of these locations i.e. (0,2), (1,2),....till (5,2) also
get activated but to a lessor degree as shown in Figure 3. The
learning factor that we use for conjugate location activation
in our experiment is 0.5. This trial is repeated a large number
of times (10000). The resulting activation for a number pair
is indicative of the strength of it’s memory representation
and the inverse of this models the RT. The RTs of our model
and the response times from the experiments conducted by
Miller et al. [13] are shown in Figure 4. Note that both
matrices are normalized between 0 and 1.
Discussion At first glance, the two plots shown in Figure 4
appear to be different. However, a closer look provides some
useful insights:
1. The problem-size effect is clearly visible in both plots.
As the multiplicands increase, there is an increase in the
corresponding RTs. The reason this happens in our simulation (Figure 4b) is because in our experimental design, the frequency of occurrence of smaller numbers is
much higher than that of larger numbers while memorizing multiplication tables. Hence, problem-size effect
might have its roots in the way arithmetic operations are
memorized and practiced. This conclusion from our simulation model also validates some of the experiments reported by Hamann & Ashkraft [10] in which they they
to model our RTs.
Ablation Experiment
Often, to learn about how a working system operates, one
can remove a component and study how the system compensates, a process called as ablation. Since our cognitive
arithmetic system depends on the retrieval of archived operations, we removed a subset of these operations and determined the effect on response time. Based on prior research, novices resort to a counting strategy when exposed
to an addition or subtraction problem in which they do not
have operations memorized. Therefore, we incorporated
this functionality into our system. Additionally, research by
Ashcraft & Guillaume [2] suggests that novices commonly
forget single-digit operations involving the numbers 5,6,7,8,
and 9. Thus, we chose to ablate all operations where both
operands contained either of these numbers and then measure the response times.
Figure 3: Activation modeling for an input pair (2,5)
Discussion Figure 5 shows the response times for all
single-digit addition operations with and without ablation.
It is evident that ablation and the resulting counting strategy
results in significantly higher response times. This provides
further evidence that counting based strategies are not generally employed for operations involving single digit operands
because if that were the case the response times observed in
experimental data would resemble Figure 5b more closely.
Strategy Optimization Experiment
(a)
(b)
Figure 4: Normalized Retrieval Times: a) From experimental data provided by [13], b) From our model simulation
looked at the addition problems in textbooks from three
publishers for grades K through 3. They also found out
that the frequency of operands 2,3 or 4 was significantly
higher than operands 5-9. Similar results were also reported by Clapp & Heubner [6].
2. The tie-effect wherein the RT for an operand pair with
two identical digits is comparatively lower than its surrounding pairs on the plot is also observed in both plots,
albeit to a greater degree in our simulation model. This
can be seen as the slightly “greener” diagonal line running
from the top-left to the bottom-right corner of Figure 4b.
Hence, even with significant differences our simulation
model is able to account for two important effects which
are observed in experimental data which supports our assumption that for single digit operands, we employ a memory based retrieval process. In the subsequent experiments
we thus use the response times reported by Miller et al. [13]
Because of the problem-size effect wherein problems with
larger digit values tend to have a longer RT, it is possible
for one to create to optimized strategies for certain types of
problem from a metacognition perspective. In other words,
it might be beneficial to transform a typical problem into a
problem with operands which have smaller digit values.
Needless to say, it usually takes one huge amount of practice to find out a better strategy for a certain type of problem.
In this process, a person uses logic to find out these strategies with several attempts to the problem. Then, analogy
is employed so that the strategy can be applied to a similar problem. One example of such strategy is “Division by
5”, where, in certain cases, it is more beneficial to multiply
the numerator and the denominator both by 2 first, which results the denominator to be 10. For example, for 1375
5 , it will
1375∗2
be better to implement a strategy so that 1375
=
=
5
5∗2
2750
=
275.
Another
common
strategy
is
“Fast
Addition”,
10
wherein whenever an operand is close to a number with lots
of zeros at the end (100, 1000, 5000, 40000 etc.), it may
be faster to first round the operand to that number, subtract
the difference from the second operand and then add the two
new operands. For example, for 497 + 38, it will be faster to
have 497 + 38 = (497 + 3) + (38 − 3) = 500 + 35 = 535.
We tried to simulate this strategy in this experiment. Consider n+679, where n is from 19950 to 20000. The response
time from our model for that operation with and without Fast
Addition strategy is plotted in Figure 6.
Discussion It is clear to see that the larger the distance of
the augend is from 20000, the lower the difference between
(a)
(b)
Figure 5: Response Times in Milliseconds a) Without Ablation b) With Ablation
Addition. The results can been seen in Table 1.
Table 1: Comparison between Operations for Fast Addition.
Sum
19992 + 679
19952 + 679
Figure 6: Response Time vs. Value of Operand for n+679.
the response times of the two strategies. For cases where the
augend is far away from 20000, it is sometimes not worth
it to implement Fast Addition at all (red cure above blue
curve). Take 19992+679 as one example. The response time
from our model for that operation is 2.694 seconds without
using Fast Addition and 2.059 seconds with Fast Addition.
On the other hand, for 19952 + 679, it takes 2.566 seconds
without using Fast Addition and 2.341 seconds with Fast
Distance: n to 20000
8
48
Time Saved (s)
0.635
0.225
It can be seen in Table 1 that the time saved by using Fast
Addition is 0.635 seconds for 19992 + 679 and 0.225 seconds for 19952 + 679. The reason for this phenomenon is
that if the distance is a 2-digit number with a large value, it
might not be worth it at all to implement the Fast Addition
strategy since it will not save a lot of time. If we consider
accuracy as well, it is more likely to make mistakes in operations with 2-digit numbers than with 1-digit numbers. Thus,
an optimal threshold for the Fast Addition strategy might be
a distance within 10 to a number with more zeros at the end
(20000 in this case). For example in Figure 7, the actual improvement in response time is shown when the Fast Addition
strategy is employed from 19991.
In general, there exist a large number of time saving
strategies for all operations. Using logic from several attempts to a specific type of problems, it is possible to find out
a pattern and solve the problem with minimal time. Furthermore, the same strategy can be applied to similar problems
with certain patterns by analogy. It is also interesting to note
that the reason these optimizing strategies work is because
Finally, the optimization over strategies can be upgraded to
the next level where the new target is to not only produce a
fast but also an accurate response to any arithmetic operation.
References
[1] Ashcraft, M. H., and Battaglia, J. 1978. Cognitive arithmetic: Evidence for retrieval and decision processes in
mental addition. Journal of Experimental Psychology:
Human Learning and Memory 4(5):527.
[2] Ashcraft, M. H., and Guillaume, M. M. 2009. Mathematical cognition and the problem size effect. Psychology
of learning and motivation 51:121–151.
[3] Ashcraft, M. H. 1992. Cognitive arithmetic: A review
of data and theory. Cognition 44(1):75–106.
Figure 7: Response Time vs. Value of Operand for n+679
with the Optimized Strategy.
of the problem-size effect which in turn might be the result
of the non-uniform way in which we learn arithmetic.
Conclusion & Future Work
In this study, a rule-based computational model is proposed
for cognitive arithmetic, where frequent memory updates
and retrievals for arithmetic rules and facts are needed.
From the experiments performed, it is clear to see the RT
is strongly correlated to how frequently a specific operand
pair is encountered. The more frequently an operand pair
appears, the faster one is able to respond to it. In order to
validate the rule-based model, an ablation experiment was
performed. The ablation of certain arithmetic rules and facts
resulted in a dramatic degradation to the overall performance
of the model, where, in this case, the model resorted to
counting, which is the fundamental method in arithmetic operations. It is also shown that response time depends on the
problem size. A problem with larger numbers tends to result in a slower response. Using this, it is possible for one to
find out better strategies that can to be employed for certain
operations. In other words, it becomes possible to optimize
the response time using a strategy, where operations with
larger numbers are avoided. In our rule-based computational
model, the optimization strategies for certain operations are
successful and perform as expected.
A significant drawback of our model is that we don’t have
any way to model the accuracy of the system. Our model
is always accurate and that is one area where it differs from
an actual human. The future work based on this model is to
take accuracy into consideration in modeling cognitive arithmetic. Accuracy is another important evaluation parameter
besides response time for cognitive arithmetic. The relationship between accuracy and problem size can be studied. Furthermore, the relationship between response time and accuracy can be determined since typically the faster a person responds to a problem the more likely he will make a mistake.
[4] Campbell, J. I. D., and Xue, Q. 2001. Cognitive arithmetic across cultures. Journal of Experimental Psychology: General 130(2):299.
[5] Campbell, J. I. D. 2008. Subtraction by addition. Memory & Cognition 36(6):1094–1102.
[6] Clapp, F. L., and Huebner, B. P. 1924. The Number Combinations, Their Relative Difficulty and the Frequency of Their Appearance in Text-books. Number 2.
[7] Fendrich, D. W.; Healy, A. F.; and Bourne Jr, L. E. 2014.
Mental arithmetic: Training and retention of multiplication skill. In Cognitive Psychology Applied: A Symposium at the 22nd International Congress of Applied Psychology, 111. Psychology Press.
[8] Geary, D. C.; Frensch, P. A.; and Wiley, J. G. 1993.
Simple and complex mental subtraction: Strategy choice
and speed-of-processing differences in younger and older
adults. Psychology and Aging 8(2):242–256.
[9] Groen, G. J., and Parkman, J. M. 1972. A chronometric analysis of simple addition. Psychological review
79(4):329.
[10] Hamann, M. S., and Ashcraft, M. H. 1986. Textbook
presentations of the basic addition facts. Cognition and
Instruction 3(3):173–202.
[11] Lebiere, C. 1999. The dynamics of cognition: An
ACT-R model of cognitive arithmetic. Kognitionswissenschaft 8(1):5–19.
[12] LeFevre, J.-A., and Morris, J. 1999. More on the relation between division and multiplication in simple arithmetic: Evidence for mediation of division solutions via
multiplication. Memory {&} Cognition 27(5):803–812.
[13] Miller, K.; Perlmutter, M.; and Keating, D. 1984. Cognitive arithmetic: Comparison of operations. Journal of
Experimental Psychology: Learning, Memory, and Cognition 10(1):46.
[14] Umilta, C. 2005. Computational modeling of numerical cognition. The handbook of mathematical cognition
67.
| 2cs.AI
|
HOPF ALGEBRAS FOR MATROIDS OVER HYPERFIELDS
arXiv:1712.08903v1 [math.CO] 24 Dec 2017
CHRISTOPHER EPPOLITO, JAIUNG JUN, AND MATT SZCZESNY
Abstract. Recently, M. Baker and N. Bowler introduced the notion of matroids over hyperfields as a unifying theory of various generalizations of matroids. In this paper we generalize the notion of minors and direct sums from
ordinary matroids to matroids over hyperfields. Using this we generalize the
classical construction of matroid-minor Hopf algebras to the case of matroids
over hyperfields.
1.
Introduction
A basic result in algebraic geometry is that the category Affk of affine schemes over a field
k is equivalent to the opposite category Algop
k of the category of commutative k-algebras.
When one enhances Affk to affine group schemes over k, one obtains Hopf algebras as an
enrichment of commutative k-algebras. In fact, Hopf algebras naturally appear not only in
algebraic geometry, but also in various fields of mathematics including (but not limited to)
algebraic topology, representation theory, quantum field theory, and combinatorics. For a
brief historical background for Hopf algebras, we refer the readers to [4]. For a comprehensive
introduction to Hopf algebras in combinatorics, we refer the readers to [12]. In this paper,
our main interest is in Hopf algebras arising in combinatorics, namely those obtained from
matroids (or more generally matroids over hyperfields).
Matroids are combinatorial objects arising in two main ways: (1) as a model of cycle
structures in graphs and (2) as combinatorial abstractions of linear independence properties
in vector spaces. While matroids have their own charms, it is their rich interplay with other
areas of mathematics which makes them truly interesting. For instance, N. Mnëv’s universality
theorem [16] roughly states that any semi-algebraic set in Rn is the moduli space of realizations
of an oriented matroid (up to homotopy equivalence). There is an analogue in algebraic
geometry known as Murphy’s Law by R. Vakil [21] for ordinary matroids. Valuated matroids
(a generalization of matroids by giving certain “weights” to bases) are analogous to “linear
spaces” in the setting of tropical geometry. Indeed a moduli space of valuated matroids (a
tropical analogue of a Grassmannian), called a Dressian, has received much attention.
Hopf algebras arising in combinatorics are usually created to encode the basic operations of
an interesting class of combinatorial objects. Matroids have several basic operations, the most
basic of which are deletion, contraction, and direct sum; an iterated sequence of deletions and
contractions on a matroid results in a minor of the matroid. The Hopf algebra associated
to a set of isomorphism classes of matroids closed under taking minors and direct sums is
called a matroid-minor Hopf algebra. In this paper, we generalize the construction of the
matroid-minor Hopf algebra to the setting of matroids over hyperfields, first introduced by
M. Baker and N. Bowler in [5].
Date: December 27, 2017.
2010 Mathematics Subject Classification. 05E99(primary), 16T05(secondary).
Key words and phrases. matroid, hyperfield, matroid over hyperfields, Hopf algebra, minor, direct sum.
1
Remark 1.1. In fact, many combinatorial objects posses notions of “deletion” and “contraction” and hence one can associate Hopf algeras (or bialgebras in the disconnected case).
In [10], C. Dupont, A. Fink, and L. Moci associate a universal Tutte character to such combinatorial objects specializing to Tutte polynomials in the case of matroids and graphs (among
others) generalizing the work [14] of T. Krajewski, I. Moffatt, and A. Tanasa. See §6.3 in
connection with our work.
Hyperfields were first introduced by M. Krasner in his work [15] on an approximation of
a local field of positive characteristic by using local fields of characteristic zero. Krasner’s
motivation was to impose, for a given multiplicative subgroup G of a commutative ring A,
“ring-like” structure on the set of equivalence classes A/G (G acts on A by left multiplication). Krasner abstracted algebraic properties of A/G and defined hyperrings, in particular,
hyperfields. Roughly speaking, hyperfields are fields with multi-valued addition. For instance,
when A = k is a field and G = k − {0}, one has k/G = {[0], [1]}, where [0] (resp. [1]) is
the equivalence of 0 (resp. 1). Then one defines [1] + [1] = {[0], [1]}. This structure is called
the Krasner hyperfield (see, Example 2.12). After Krasner’s work, hyperfields (or hyperrings, in general) have been studied mainly in applied mathematics. Recently several authors
(including the second author of the current paper) began to investigate hyperstructures in
the context of algebraic geometry and number theory. Furthermore, very recently M. Baker
(later with N. Bowler) employ hyperfields in combinatorics: Baker and Bowler found a beautiful framework which simultaneously generalizes the notion of linear subspaces, matroids,
oriented matroids, and valuated matroids. In light of various fruitful applications of Hopf
algebra methods in combinatorics, one might naturally ask the following:
Question. Can we generalize matroid-minor Hopf algebras to the case of matroids over hyperfields?
We address this question in this paper. We first define minors for matroids over hyperfields
which generalize the definition of minors for ordinary matroids:
Theorem A (§3). Let H be a hyperfield. There are two cryptomorphic definitions (circuits
and Grassmann-Plücker functions) of minors of matroids over H. Furthermore, if M is a
weak (resp. strong) matroid over H, then all minors of M are weak (resp. strong).
Next, we introduce the notion of direct sums of matroids over hyperfields and prove that
direct sums preserve the type (weak or strong) of matroids over hyperfields in the following
sense:
Theorem B (§3). Let H be a hyperfield. There are two cryptomorphic definitions (circuits
and Grassmann-Plücker functions) of direct sums of matroids over H. Furthermore, if M1
and M2 are (weak or strong) matroids over H, then the direct sum M = M1 ⊕ M2 is always
a weak matroid over H and M is strong if and only if both M1 and M2 are strong.
Remark 1.2. We note that our definition of direct sum is a natural generalization of the
direct sum of ordinary matroids. It also meshes nicely with the definition of direct sums for
matroids over fuzzy rings in [9].
By appealing to the above results, we define Hopf algebras for matroids over hyperfields in
§5. Finally, in §6, we explain how our current work can be thought in views of matroids over
fuzzy rings as in Dress-Wenzel theory [9], matroids over partial hyperfields [6], and universal
Tutte characters [10] as well as Tutte polynomials of Hopf algebras [14].
2
Acknowledgments
The first author would like to thank Laura Anderson for many helpful comments and conversations as well as much patience and support. The second author would like to thank Boston
University for their support during his visit. The third author gratefully acknowledges the
support of a Simons Foundation Collaboration Grant.
2.
Preliminaries
In this section, we recall the definitions of the following for readers who are not familiar
with these key notions:
• Matroids.
• Hyperfields and matroids over hyperfields.
• Hopf algebras and matroid-minor Hopf algebras.
In what follows, we will only consider matroids on a finite set unless otherwise stated. For
infinite matroids in the context of the current paper, we refer readers to [8].
Notice that we let N = Z≥0 .
2.1 Matroids. This section is intended as a brief refresher on the basic notions and operations in matroid theory. Readers familiar with matroid theory may skip this section. We
refer readers to [17] and [22] for further details and proofs of the facts from this section.
A matroid is a combinatorial model for the properties of linear independence in a (finite
dimensional) vector space and for the properties of cycles in combinatorial graphs. It is wellknown that there are several “cryptomorphic” definitions for matroids. Chief among these are
the notions of bases and circuits.
Let E be a finite set (the ground set of a matroid). A nonempty collection B ⊆ P(E),
where P(E) is the power set of E, is a set of bases of a matroid when B satisfies the basis
exchange axiom, given below:
(1) For all X, Y ∈ B and all x ∈ X \ Y there is an element y ∈ Y \ X such that
(X \ {x}) ∪ {y} ∈ B.
In the context of (finite dimensional) vector spaces, any pair of bases B1 , B2 of a given subspace
satisfies this property by the Steinitz Exchange Lemma. In the context of finite graphs, this
is a corollary of the Tree Exchange Property satisfied by the edge sets of spanning forests in
the graph.
A collection C ⊆ P(E) is a set of circuits of a matroid on E when C satisfies the following
three axioms:
(1) (Nondegeneracy) ∅ ∈
/ C.
(2) (Incomparability) If X, Y ∈ C and X ⊆ Y , then X = Y .
(3) (Circuit elimination) For all X, Y ∈ C and all e ∈ X ∩ Y , there is a Z ∈ C such that
Z ⊆ (X ∪ Y ) \ {e}.
In the context of (finite) graphs, circuits are precisely the edge sets of cycles in the graph. In
the context of (finite dimensional) vector spaces, circuits correspond with minimal dependence
relations on a finite set of vectors.
Remark 2.1. There is a natural bijection between sets of circuits of a matroid on E and sets
of bases of a matroid on E. Given a set C of circuits of a matroid, we define B C to be the set
of maximal subsets of E not containing any element of C. Likewise, given a set B of bases
of a matroid, we define C B to be the set of minimal nonempty subsets of E which are not
contained in any element of B. It is a standard exercise to show that (1) B C is a set of bases
of a matroid, (2) C B is a set of circuits of a matroid, and (3) both B C B = B and C BC = C.
3
In this sense B and C B carry the same information as C B and B C respectively; these are thus
said to determine the same matroid on E “cryptomorphically.”
Example 2.2. The motivating examples of matroids (hinted at above) are given as follows:
(1) Let V be a finite dimensional vector space and E ⊆ V a spanning set of vectors.
The bases of V contained in E form the bases of a matroid on E, and the minimal
dependent subsets of E form the circuits of a matroid on E. Furthermore, these are
the same matroid.
(2) Let Γ be a finite, undirected graph with edge set E (loops and parallel edges are
allowed). The sets of edges of spanning forests in Γ form the bases of a matroid on
E, and the sets of edges of cycles form the circuits of a matroid on E. Furthermore,
these are the same matroid (called the graphic matroid of Γ).
One can define the notion of isomorphisms of matroids as follows.
Definition 2.3. Let M1 (resp. M2 ) be a matroid on E1 (resp. E2 ) defined by a set B1 (resp. B2 )
of bases. We say that M1 is isomorphic to M2 if there exists a bijection f : E1 → E2 such
that f (B) ∈ B2 if and only if B ∈ B1 . In this case, f is said to be an isomorphism.
Example 2.4. Let Γ1 and Γ2 be finite graphs and M1 and M2 be the corresponding graphic
matroids. Every graph isomorphism between Γ1 and Γ2 gives rise to a matroid isomorphism
between M1 and M2 , but the converse need not hold.
Recall that given any base B ∈ B(M ) and any element e ∈ E \ B, there is a unique circuit
(fundamental circuit ) CB,e of e with respect to B such that CB,e ⊆ B ∪ {e}.
One can construct new matroids from given matroids as follows:
Definition 2.5 (Direct sum of matroids). Let M1 and M2 be matroids on E1 and E2 given
by bases B1 and B2 respectively. The direct sum M1 ⊕ M2 is the matroid on E1 ⊔ E2 given
by the bases B = {B1 ⊔ B2 | Bi ∈ Bi for i = 1, 2}.
Remark 2.6. One can easily check that M1 ⊕ M2 is indeed a matroid on E1 ⊔ E2 .
Definition 2.7 (Dual, Restriction, Deletion, and Contraction). Let M be a matroid on a
finite set EM with the set BM of bases and the set CM of circuits. Let S be a subset of EM .
(1) The dual M ∗ of M is a matroid on EM given by bases
BM ∗ := {EM − B | B ∈ BM } .
(2) The restriction M |S of M to S is a matroid on S given by circuits
C M |S = {D ⊆ S | D ∈ C M } .
(3) The deletion M \ S of S is the matroid M \ S := M |(E \ S).
(4) The contraction of M by S is M/S := (M ∗ \ S)∗ .
It is easy to show that M |S, M ∗ , and M/S are indeed matroids. A minor of a matroid M
is any matroid obtained from M by a series of deletions and/or contractions. Basic properties
of these operations are given below:
Proposition 2.8. Let M be a matroid on E. We have the following for all disjoint subsets
S and T of E:
(1) M/∅ = M = M \ ∅
(2) (M \ S) \ T = M \ (S ∪ T )
(3) (M/S)/T = M/(S ∪ T )
(4) (M \ S)/T = (M/T ) \ S
In particular, the minors of a matroid are in one-to-one correspondence with the ordered pairs
of disjoint subsets of the ground set by deleting the first and contracting the second.
4
2.2 Matroids over hyperfields. In this section, we review basic definitions and properties
for matroids over hyperfields first introduced by Baker and Bowler in [5]. Let’s first recall
the definition of a hyperfield. By a hyperaddition on a nonempty set H, we mean a function
+ : H × H → P ∗ (H) such that +(a, b) = +(b, a) for all a, b ∈ H, where P ∗ (H) is the set of
nonempty subsets of H. We will simply write a + b for +(a, b). A hyperaddition + on H is
associative if the following condition holds: for all a, b, c ∈ H,
(1)
a + (b + c) = (a + b) + c.
S
Note that in general for subsets A and B of H, we write A + B := a∈A,b∈B a + b and hence
the notation in (1) makes sense. Also, we will always write a singleton {a} as a.
Definition 2.9. Let H be a nonempty set with an associative hyperaddition +. We say that
(H, +) is a canonical hypergroup when the following conditions hold:
• ∃! 0 ∈ H such that a + 0 = a for all a ∈ H; existence of identity.
• ∀ a ∈ H, ∃! b (=: −a) ∈ H such that 0 ∈ a + b; existence of inverses.
• ∀ a, b, c ∈ H, if a ∈ b + c, then c ∈ a + (−b); ‘hyper-subtraction’ or reversibility.
We will write a − b instead of a + (−b) for brevity of notation.
Definition 2.10. By a hyperring, we mean a nonempty set H with a binary operation · and
hyperaddition + such that (H, +, 0) is a canonical hypergroup and (H, ·, 1) is a commutative
monoid satisfying the following conditions: for all a, b, c ∈ H,
a · (b + c) = a · b + a · c,
0 · a = 0,
and 1 6= 0.
When (H − {0}, ·, 1) is a group, we call H a hyperfield.
Definition 2.11. Let H1 and H2 be hyperrings. A homomorphism of hyperrings from H1 to
H2 is a function f : H1 → H2 such that f is a monoid morphism with respect to multiplication
satisfying the following conditions:
f (0) = 0 and f (a + b) ⊆ f (a) + f (b), ∀a, b ∈ H1 .
The following are some typical examples of hyperfields found in the literature:
Example 2.12 (K; Krasner hyperfield). Let K := {0, 1} and impose the usual multiplication
0 · 0 = 0, 1 · 1 = 1, and 0 · 1 = 0. Hyperaddition is defined as 0 + 1 = 1, 0 + 0 = 0, and
1 + 1 = K. The structure K is the Krasner hyperfield.
Example 2.13 (S; hyperfield of signs). Let S := {−1, 0, 1} and impose multiplication in
a usual way following the rule of signs; 1 · 1 = 1, (−1) · 1 = (−1), (−1) · (−1) = 1, and
1 · 0 = (−1) · 0 = 0 · 0 = 0. Hyperaddition also follows the rule of signs as follows:
1 + 1 = 1, (−1) + (−1) = (−1), 1 + 0 = 1, (−1) + 0 = (−1), 0 + 0 = 0, 1 + (−1) = S.
The structure S is the hyperfield of signs.
Example 2.14 (P; phase hyperfield). Let P := S 1 ∪ {0}, where S 1 is the unit circle in the
complex plane. The multiplication on P is the usual multiplication of complex numbers.
Hyperaddition is defined by:
{−a, 0, a}
if a = −b (−b as a complex number)
a+b=
the shorter open arc connecting a and b if a 6= −b,
The structure P is the phase hyperfield.
5
Example 2.15 (T; tropical hyperfield). Let G be a (multiplicative) totally ordered abelian
group. Then one can enrich the structure of G to define a hyperfield. To be precise, let
Ghyp := G ∪ {−∞} and define multiplication via the multiplication of G together with the
rule g · (−∞) = −∞ for all g ∈ G. Hyperaddition is defined as follows:
max{a, b} if a 6= b
a+b=
[−∞, a]
if a = b,
where [−∞, a] := {g ∈ Ghyp | g ≤ a} with −∞ the smallest element. Then one can easily
see that Ghyp is a hyperfield. When G = R, the set of real numbers (considered as a totally
ordered abelian group with respect to the usual addition), we let T := Rhyp . The structure T
is the tropical hyperfield.
Remark 2.16. One easily observes that for any hyperfield H, there exists a unique homomorphism ϕ : H → K sending every nonzero element to 1 and 0 to 0. In other words, K is
the final object in the category of hyperfields.
In what follows, let (H, ⊞, ⊙) be a hyperfield, H × = H − {0H }, r a positive integer,
[r] = {1, ..., r}, x an element of E r such that x(i) ∈ E is the ith coordinate of x unless otherwise stated. Now we recall the two notions (weak and strong) of matroids over hyperfields
introduced by Baker and Bowler. These notions are given cryptomorphically by structures
analogous to the bases and circuits of ordinary matroids. Their definition simultaneously generalizes several existing theories of “matroids with extra structure,” evidenced by the following
examples:
Example 2.17. Matroids over the following hyperfields have been studied in the past:
• A (strong or weak) matroid over a field K is a linear subspace.
• A (strong or weak) matroid over the Krasner hyperfield K is an ordinary matroid.
• A (strong or weak) matroid over the hyperfield of signs S is an oriented matroid.
• A (strong or weak) matroid over the tropical hyperfield T is a valuated matroid.
We first recall the generalization of bases to the setting of matroids over hyperfields. This
is done via Grassmann-Plücker functions.
Definition 2.18. Let H be a hyperfield, E a finite set, r a nonnegative integer, and Σr the
symmetric group on r letters with a canonical action on E r (acting on indices).
(1) A function ϕ : E r → H is a nontrivial H-alternating function when:
(G1) The function ϕ is not identically zero.
(G2) For all x ∈ E r and all σ ∈ Σr we have ϕ(σ · x) = sgn(σ)ϕ(x).
(G3) If x ∈ E r has x(i) = x(j) for some i < j, then ϕ(x) = 0H .
(2) A nontrivial H-alternating function ϕ : E r → H is a weak-type Grassmann-Plücker
function over H when:
(WG) For all a, b, c, d ∈ E and all x ∈ E r−2 we have
0H ∈ ϕ(a, b, x)ϕ(c, d, x) − ϕ(a, c, x)ϕ(b, d, x) + ϕ(b, c, x)ϕ(a, d, x).
(3) A nontrivial H-alternating function ϕ : E r → H is a strong-type Grassmann-Plücker
function over H when:
(SG) For all x ∈ E r+1 and all y ∈ E r−1 we have
0H ∈
r+1
X
(−1)k ϕ(x|[r+1]\{k} )ϕ(x(k), y).
k=1
(4) The rank of a Grassmann-Plücker function ϕ : E r → H is r.
(5) Two Grassmann-Plücker functions ϕ, ψ : E r → H are equivalent when there is an
element a ∈ H × with ψ = a ⊙ ϕ.
6
A matroid over H is an H × -equivalence class [ϕ] of a Grassmann-Plücker function ϕ.
Before presenting circuits of matroids over hyperfields, we need the following technical
definition.
Definition 2.19. Let S be a collection of inclusion-incomparable subsets of a set E. A
modular pair in S is a pair of distinct elements X, Y ∈ S such that for all A, B ∈ S, if
A ∪ B ⊆ X ∪ Y , then A ∪ B = X ∪ Y .
Having Definition 2.19, we can now give definitions P
of collections of circuits for matroids
over hyperfields. In what follows, we will simply write
instead of ⊞ if the context is clear.
Definition 2.20. Let E be a finite set, (H, ⊞, ⊙) a hyperfield, H E the set of functions from
E to H, and for any X ∈ H E , we define supp(X) := {a ∈ E | X(a) 6= 0H }.
(1) A collection C ⊆ H E is a family of pre-circuits over H when it satisfies the following
axioms:
(C1) 0 ∈
/C
×
(C2) H ⊙ C = C
(C3) For all X, Y ∈ C, if supp(X) ⊆ supp(Y ), then Y = a ⊙ X for some a ∈ H × .
(2) A pre-circuit set C over H is a weak-type circuit set when it satisfies the following
additional axiom:
(WC) ∀ X, Y in C such that {supp(X), supp(Y )} forms a modular pair in supp(C) :=
{supp(X) | X ∈ C} and for all e ∈ supp(X) ∩ supp(Y ), there is a Z ∈ C such
that
Z(e) = 0H and Z ∈ X(e) ⊙ Y − Y (e) ⊙ X,
i.e., for all f ∈ E, Z(f ) ∈ X(e) ⊙ Y (f ) − Y (e) ⊙ X(f ).
(3) A pre-circuit set C over H is a strong-type circuit set when it satisfies the following
additional axioms:
(SC1) The set supp(C) is the set of circuits of an ordinary matroid MC .
(SC2) For all bases B ∈ B C and all X ∈ C we have
X
X∈
X(e) ⊙ YB,e ,
e∈E\B
where YB,e is the (unique) element of C with YB,e (e) = 1 and supp(YB,e ) is the
fundamental circuit of e with respect to B.
Remark 2.21. The definition of strong-type circuit sets given above is not the original definition; rather this is equivalent to the original (much less intuitive) definition by [5, Theorem
3.8, Remark 3.9]. For our purposes, we shall use the definition given above.
The following result is proved in [5]:
Proposition 2.22. Let H be a hyperfield. The H × -orbits of Grassmann-Plücker functions
over H are in natural one-to-one correspondence with the H-circuits of a matroid, preserving
both ranks and types (weak and strong).
The correspondence is described as follows:
Given a Grassmann-Plücker function ϕ over H, one first shows that the collection of subsets
B ⊆ E for which an ordering B has ϕ(B) 6= 0H forms a set of bases for an ordinary matroid
Mϕ . Next, one can define a set of H-circuits by defining for all ordered bases B of Mϕ and
all e ∈ E \ B a function X = XB,e supported on the fundamental circuit for e by B via the
equality
X(B(i))X(e)−1 = (−1)i ϕ(e, B|[r]\{i} )ϕ(B)−1
(2)
7
This equality uniquely determines X : E → H, up to the multiplicative action of H × . The
collection of all such X is a collection of H-circuits of the same type as ϕ.
Constructing a Grassmann-Plücker function from circuits is more difficult to describe, and
requires the additional notion of dual pairs. An explicit description of this construction is
unnecessary for our purposes; the interested reader is referred to [5].
We now describe the duality operation for matroids over hyperfields in terms of GrassmannPlücker functions and subsequently in terms of circuits. It should be noted that the duality
described in [5] incorporates a notion of conjugation generalizing the complex conjugation.
This changes the duality operation, but the change is equally well described by another
operation (called “pushforward through a morphism”) as noted in a footnote in [5, §6]. Our
treatment will also assume that the conjugation is trivial.
Fix a total ordering ≤ of E. A dual of a Grassmann-Plücker function ϕ over H is defined by
the equation ϕ∗ (B) := sgn≤ (B, E \ B)ϕ(E \ B) for all cobases B of the underlying matroid
of ϕ, using the convention that S denotes the ordered tuple with coordinates the elements
of S arranged according to our fixed total ordering on E and sgn≤ (B, E \ B) denotes the
sign of the permutation given by the word (B, E \ B) with respect to the ordering ≤. The
definition can be uniquely extended to the set E #E−r by alternation and the degeneracy
conditions for Grassmann-Plücker functions over H. It is relatively easy to see that if ϕ is
a Grassmann-Plücker function, then ϕ∗ is a Grassmann-Plücker function of the same type.
Notice that this duality is well-defined up to the chosen ordering; a different ordering will
induce a Grassmann-Plücker function which is multiplied by the sign of the permutation used
to translate between the two orderings. In particular, this notion of duality is constant on
the level of H × -orbits of Grassmann-Plücker functions, and thus sends an H-matroid M to
an H-matroid M ∗ of the same type despite the fact that there is no canonical dual to the
original Grassmann-Plücker function.
The dual of a circuit set requires some more care to define; it is here that the contrast
between weak and strong H-matroids is most stark.
Definition 2.23. Let H be a hyperfield and E a finite set.
(1) The dot product of two functions X, Y : E → H is the following subset of H:
X
X(e)Y (e).
X · Y :=
e∈E
(2) Two functions X and Y are strong orthogonal, denoted X ⊥s Y , when
0H ∈ X · Y.
(3) Two functions X and Y are weak orthogonal, denoted X ⊥w Y , when either X ⊥s Y
or the following condition holds:
#(supp(X) ∩ supp(Y )) > 3.
(4) Let C be a set of strong H-circuits on E, we define the following subset of H E :
C ⊥s := {X : E → H | X ⊥s Y for all Y ∈ C} .
(5) Let C be a set of weak H-circuits, we define the following subset of H E :
C ⊥w := {X : E → H | X ⊥w Y for all Y ∈ C} .
For ease of notation the symbol ⊥ is to be understood in context as either ⊥s or ⊥w .
Definition 2.24. Let H be a hyperfield and M be a weak (resp. strong) H-matroid with the
set C of weak type (resp. strong type) H-circuits. The cocircuits of C, denoted by C ∗ , are the
elements of the perpendicular set C ⊥w (resp. C ⊥s ) with minimal support.
8
Remark 2.25. In [5, §6.6], Baker and Bowler show that this determines an H-matroid with
the properties supp(C ∗ ) = (supp(C))∗ and C ∗∗ = C; in other words, the underlying matroid of
the dual is the dual of the underlying matroid and the double dual is identical to the original
H-matroid.
Remark 2.26. From an algebraic geometric view point, Baker and Bowler’s definition of
matroids over hyperfields can be considered as points of a Grassmannian over a hyperfield
H. Motivated by this observation, in [13], the second author proves that certain topological
spaces (the underlying spaces of a scheme, Berkovich analytificaiton of schemes, real schemes)
are homeomorphic to sets of rational points of a scheme over a hyperfield. Also, recently
L. Anderson and J. Davis defined and investigated hyperfield Grassmannians in connection
to the MacPhersonian (from oriented matroid theory) in [2].
2.3 Matroid-Minor Hopf algebras. In this subsection, we recall the definition of matroidminor Hopf algebras. First we briefly recall the definition of Hopf algebras; interested readers
are referred to [7] for more details.
Definition 2.27. Let k be a field. A commutative k-algebra A is a Hopf algebra if A is
equipped with maps
(1) (Comultiplication) ∆ : A → A ⊗k A,
(2) (Counit) ε : A → k,
(3) (Antipode) S : A → A
such that the following diagrams commute:
A ⊗k A
∆⊗id
id ⊗∆
∆
A
A ⊗k A ⊗k A
∆
A ⊗k A,
A ⊗k A
ε⊗id
k ⊗k A
≃
∆
id
A
A,
A ⊗k A
µ◦(S⊗id)
i
∆
A
A
ε
k,
where µ : A⊗k A → A is the multiplication of A. If A is only equipped with ∆ and ε satisfying
the first two commutative diagrams, then A is a bialgebra.
Definition 2.28. Let (A, µ, ∆, η, ε) be a bialgebra over a field k.
L
(1) A is graded if there is a grading A = i∈N Ai which is compatible with the bialgebra
structure of A, i.e., µ, ∆, η, and ε are graded k-linear maps.
(2) A is connected if A is graded and A0 = k.
Definition 2.29. Let A1 and A2 be Hopf algebras over a field k. A homomorphism of Hopf
algebras is a k-bialgebra map ϕ : A1 → A2 which preserves the antipodes, i.e., SA1 ϕ = ϕSA2 .
The following theorem shows that indeed there is no difference between bialgebra maps
and Hopf algebra maps.
Theorem 2.30. [7, Proposition 4.2.5.] Let A1 and A2 be Hopf algebras over a field k. Let
ϕ : A1 → A2 be a morphism of k-bialgebras. Then ϕ is indeed a homomorphism of Hopf
algebras.
We also introduce the following notation:
Definition 2.31. Let A be a Hopf algebra over a field k and i ∈ Z≥1 .
(1) (Iterated multiplication): µi : A⊗(i+1) → A is defined inductively as
µi := µ ◦ (id ⊗µ(i−1) ).
(2) (Iterated comultiplication): ∆i : A → A⊗(i+1) is defined inductively as
∆i := (id ⊗∆(i−1) ) ◦ ∆.
9
Now, let’s recall the definition of matroid-minor Hopf algebras, first introduced by W. R. Schmitt
in [18]. Let M be a collection of matroids which is closed under taking minors and direct
sums. Let Miso be the set of isomorphism classes of matroids in M. For a matroid M in
M, we write [M ] for the isomorphism class of M in Miso . One easily see that Miso can be
enriched to a commutative monoid with the direct sum:
[M1 ] · [M2 ] := [M1 ⊕ M2 ]
and the identity [∅], the equivalence class of the empty matroid (considered as the unique
matroid associated to the empty ground set). Let A be the monoid algebra k[Miso ] over a
field k.
For any matroid M , let EM denote the ground set of M . Consider the following maps:
• (Comultiplication)
X
∆ : k[Miso ] → k[Miso ] ⊗k k[Miso ], [M ] 7→
[M |S ] ⊗ [M/S].
S⊆EM
• (Counit)
ε : k[Miso ] → k,
[M ] 7→
1 if EM = ∅
0 if EM 6= ∅,
Under the above maps, k[Miso ] becomes a connected bialgebra; k[Miso ] is graded by cardinalities of ground sets. It follows from the result of M. Takeuchi [20] that k[Miso ] has a
unique Hopf algebra structure with a unique antipode S given by:
X
S=
(−1)i µi−1 ◦ π ⊗i ◦ ∆i−1 ,
(3)
i∈N
µ−1
where
is a canonical map from k to k[Miso ], ∆−1 := ε, and π : k[Miso ] → k[Miso ] is the
projection map defined by
id if n ≥ 1
π|An
0 if n = 0,
and extended linearly to k[Miso ], where An is the nth graded piece of A.
3.
Minors and sums of matroids over hyperfields
In this section we explicitly write out the constructions of restriction, deletion, contraction,
and direct sums for matroids over hyperfields. We do this cryptomorphically via both circuits
and Grassmann-Plücker functions in both the weak and strong cases. Primarily, we define
the restriction, and subsequently use our characterization to derive the other cryptomorphic
descriptions of minors. It should be noted that formulas for deletion and contraction in the
case of phirotopes are given in [3] for phased matroids and in [5] for general Grassmann-Plücker
function without proof.1 For completeness, we give full proofs and expand the previous work
by giving formulas for the circuits of these objects as well.
3.1 Circuits of H-Matroid Restrictions. Let H be a hyperfield, E be a finite set, C be
a set of (either weak-type or strong-type) H-circuits on E, and S ⊆ E. Recall that H S is the
set of functions from S to H. We define the following notation:
C |S := X|S ∈ H S X ∈ C and supp(X) ⊆ S .
(4)
We have the following:
1There is an error in [3]; in particular, the authors make the false assumption that the Axiom (WG) implies
Axiom (SG). Thus they fail to handle the weak case separately from the strong case. While [5] fixes this issue,
the authors merely state this result without presenting details.
10
Proposition 3.1. Let C be a set of weak-type (resp. strong-type) H-circuits of a matroid M
over H on a ground set E.
(1) ∀ S ⊆ E, the set C |S is a set of weak-type (resp. strong-type) H-circuits on S.
(2) The underlying matroid of the H-matroid M determined by C |S is precisely the restriction of the underlying matroid supp(M )|S. In other words, the restriction commutes
with the push-forward operation to the Krasner hyperfield K.
Proof. One can easily see that if C is a set of circuits of an H-matroid, then
supp(C |S) = supp(C)|S
and hence supp(C |S) is the set of circuits of the restriction of the ordinary matroid; in
particular, the second statement follows immediately from the first statement.
Now, we prove the first statement. Suppose that X, Y ∈ C have supp(X), supp(Y ) ⊆ S. In
this case, we have that
supp(X|S ) = supp(X) and supp(Y |S ) = supp(Y ).
(5)
We first claim that if C is a set of pre-circuits over H on E, then C |S is also a set of pre-circuits
over H on S. Indeed, since supp(X) ⊆ S and X 6= 0, we have that X|S 6= 0 and
(a ⊙ X|S ) ∈ C |S,
∀a ∈ H × .
Finally, if supp(X|S ) ⊆ supp(Y |S ), then
supp(X) = supp(X|S ) ⊆ supp(Y |S ) = supp(Y )
yields Y = a ⊙ X for some a ∈ H × and hence Y |S = a ⊙ X|S as desired. This proves that
C|S is a set of pre-circuits over H on S.
Next we prove that if C is a set of weak-type H-circuits on E, then C |S is also a set of weaktype H-circuits on S. In fact, if X|S and Y |S form a modular pair in C |S, then (5) implies
immediately that X and Y are a modular pair as well in C. More precisely, in this case, the
condition
A ∪ B ⊆ supp(X) ∪ supp(Y ) ⊆ S, A, B ⊂ supp(C)
implies that A, B ⊆ S. Thus, ∀ e ∈ supp(X) ∩ supp(Y ), there exists Z ∈ C such that
Z(e) = 0 and Z ∈ X(e)Y − Y (e)X.
(6)
On the other hand, if a ∈
/ supp(X) ∪ supp(Y ), then
X(e)Y (a) − Y (e)X(a) = {0}.
Thus, for A ∈ C, A ∈ X(e)Y − Y (e)X implies that
supp(A) ⊆ supp(X) ∪ supp(Y ) ⊆ S.
Hence, supp(Z) ⊆ S and Z|S ∈ C|S . One can easily see that C |S satisfies the Axiom (WC).
This proves that C|S is a set of weak-type H-circuits on S.
Finally, we show that if C is a set of strong-type H-circuits on E, then C |S is also a set of
strong-type H-circuits on S. As we mentioned before, supp(C |S) = supp(C)|S is a set of
circuits of a matroid as these are given by the same formula and supp C is a set of circuits
of an ordinary matroid; in particular C |S satisfies axiom (SC1). Let B C |S (resp. BC ) be the
set of bases of an underlying matroid MC |S (resp. MC ) given by the set supp(C |S) (resp.
supp(C)) of circuits. If B ∈ BC |S , then we have that
B = B̃ ∩ S for some B̃ ∈ BC .
It follows from (SC2), applied to C with B̃ and X, that
X
X∈
X(e) ⊙ YB̃,e .
e∈E\B̃
11
(7)
Now YB̃,e |S = YB,e by incomparability of circuits in ordinary matroids, and thus we see (7)
implies that
X
X|S ∈
X|S (e) ⊙ YB,e .
e∈E\B
It follows that the axiom (SC2) holds for C |S and hence C |S is a strong-type H-circuit set,
as claimed.
Now, thanks to Proposition 3.1, the following definition makes sense.
Definition 3.2. Let M be a matroid over hyperfield H on a ground set E given by weak
(resp. strong) H-circuits C, and let S be a subset of E. The restriction of matroid M to S is
the matroid M |S over H given by weak (resp. strong) H-circuits C |S.
3.2 Grassmann-Plücker Functions of H-Matroid Restrictions. We now describe restriction of H-matroids via Grassmann-Plücker functions. Let H be a hyperfield, E a finite
set, r a positive integer, and ϕ a (weak-type or strong-type) Grassmann-Plücker function over
H on E of rank r. Let Mϕ denote the underlying matroid of ϕ given by bases:
B ϕ = {{b1 , · · · , br } ⊆ E | ϕ(b1 , · · · , br ) 6= 0} .
Recall that for any ordered basis B = {b1 , b2 , · · · , bk } of Mϕ /(E \ S), we let
B = (b1 , b2 , ..., bk ) ∈ E k .
Now, for any subset S ⊆ E and any (ordered) basis B = {b1 , b2 , · · · , bk } of Mϕ /(E \ S), we
define
ϕB : S r−k −→ H,
A 7→ ϕ(A, B).
(8)
Proposition 3.3. Let ϕ be a weak-type (resp. strong-type) Grassmann-Plücker function over
H on E of rank r and let S ⊆ E. For all ordered bases B of Mϕ /(E \ S), the function ϕB is a
weak-type (resp. strong-type) Grassmann-Plücker function. Moreover, all such ϕB determine
the H-circuits C |S of M |S.
Proof. For the notational convenience, we let [n] = {1, 2, · · · , n} and we regard B as a function
B : [k] → E. First, one can observe that ϕB is a nontrivial H-alternating function as ϕB is
a restriction of a nontrivial H-alternating function to a subset containing a base of Mϕ . We
claim that if ϕ is a weak-type Grassmann-Plücker function over H, then ϕB is also a weaktype Grassmann-Plücker function over H. To see this, let a, b, c, d ∈ E and Y : [r −k −1] → E
be given. Applying Axiom (WG) to a, b, c, d ∈ E and x = (Y, B) ∈ E r−2 , we obtain the
following:
0H ∈ ϕ(a, b, Y, B)ϕ(c, d, Y, B) − ϕ(a, c, Y, B)ϕ(b, d, Y, B) + ϕ(a, d, Y, B)ϕ(b, c, Y, B)
= ϕB (a, b, Y)ϕB (c, d, Y) − ϕB (a, c, Y)ϕB (b, d, Y) + ϕB (a, d, Y)ϕB (b, c, Y).
This shows that Axiom (WG) holds for ϕB and hence ϕB is a weak-type Grassmann-Plücker
function over H.
We next show that if ϕ is a strong-type Grassmann-Plücker function over H, then ϕB is
also a strong-type Grassmann-Plücker function over H. Indeed, let X : [r − k + 1] → S and
Y : [r − k − 1] → S be given. Applying Axiom (SG) to x := (X, B) and y := (Y, B), we
12
obtain the following:
X
X
0H ∈
(−1)j ϕ(X|[r−k+1]\{j} , B)ϕ(X(j), Y, B) +
(−1)r−k+1+j ϕ(X, B|[k]\{j} )ϕ(B(j), Y, B)
j∈[r−k+1]
=
X
j∈[k]
(−1)j ϕ(X|[r−k+1]\{j} , B)ϕ(X(j), Y, B) +
j∈[r−k+1]
=
X
X
(−1)r−k+1+j 0H
j∈[k]
(−1)j ϕB (X|[r−k+1]\{j} )ϕB (X(j), Y).
j∈[r−k+1]
This shows that Axiom (SG) holds for ϕB and hence ϕB is a strong-type Grassmann-Plücker
function over H.
′
Finally, we show that ϕB determines the same set of circuits as ϕB for all ordered bases B
and B′ of Mϕ /(E \ S). Indeed, we show that the circuits determined by ϕB are precisely C |S.
Fix an ordered base A of Mϕ |S. Now, y := (A, B) is an ordered base of Mϕ . Moreover, for
all e ∈ S \ A, the fundamental H-circuit X = XA∪B,e satisfies
X|S (A(i))X|S (e)−1 = X(y(i))X(e)−1
= (−1)i ϕ(e, A|[r−k]\{i} , B)ϕ(A, B)−1
= (−1)i ϕB (e, A|[r]\{i} )ϕB (A)−1 .
for all i ∈ [r−k] by the cryptomorphism relating C and ϕ. On the other hand, X|S = XA∪B,e |S
is the fundamental H-circuit for e by the basis A in C |S. Hence C |S is the set of H-circuits
determined by the Grassmann-Plücker function ϕB for all ordered bases B of Mϕ /(E \S).
We summarize our results from this section as follows:
Proposition 3.4. The restriction of an H-matroid to a subset is well-defined, and admits
cryptomorphic description in terms of Grassmann-Plücker functions over H and H-circuits.
Furthermore, this correspondence preserves types and all such restrictions have underlying
matroid the ordinary restriction. Finally, we have the following:
(1) The restriction M |S is given by H-circuits
C |S = {X|S | X ∈ C and supp(X) ⊆ S} .
(2) The restriction M |S is obtained by fixing any base B = (b1 , b2 , · · · , bk ) of Mϕ /(E \ S)
and defining:
ϕB : S r−k −→ H, x 7→ ϕ(x, B).
In particular, the H-matroid M |S is determined by the H × -class [ϕB ] of any such B.
3.3 Deletion and Contraction. As noted previously, deletion and contraction for Hmatroids were defined by Baker and Bowler in [5] by using Grassmann-Plücker functions.
In this section, we also provide a cryptomorphic definition for deletion and contraction via
H-circuits by appealing to the definitions of dual H-matroids and restrictions. Throughout
let H be a hyperfield, E a finite set, r a positive integer, and M be a matroid over H on
ground set E of rank r with circuits C and a Grassmann-Plücker function ϕ
Definition 3.5. Let S be a subset of E.
(1) The deletion M \ S of S from M is the H-matroid M |(E \ S).
(2) The contraction M/S of S from M is the H-matroid (M ∗ \ S)∗ .
Remark 3.6. It follows from Definition 3.5 that if M is a weak-type (resp. strong-type), then
the deletion M \ S and the contraction M/S are also weak-type (resp. strong-type).
Proposition 3.7. Let S be a subset of E.
13
(1) The deletion M \ S is given by H-circuits
C |(E \ S) = X|E\S X ∈ C and S ∩ supp(X) = ∅ .
(9)
(2) The deletion M \ S is obtained by fixing base B = (b1 , b2 , · · · , br−k ) of Mϕ /S and
letting ϕB : (E \ S)k → H : x 7→ ϕ(x, B). The H-matroid M \ S is determined by the
H × -class [ϕB ] for any such B.
Proof. The first statement is immediate from the definition of the deletion and the second
statement directly follows from Proposition 3.4.
A description of contractions is a bit more complicated.
Proposition 3.8. Let S be a subset of E.
(1) The contraction M/S is given by H-circuits C ′′ = (C ∗ |(E \ S))∗ . More explicitly
Z ⊥ X|E\S for all X ∈ H E \ {0}
′′
E\S
C = min Z ∈ H
\ {0}
. (10)
with supp(X) ∩ S = ∅ and X ⊥ Y for all Y ∈ C
(2) The contraction M/S is given by the class of Grassmann-Plücker functions ((ϕ∗ )E\S )∗ .
More explicitly, let B = (b1 , · · · , bk ) be an ordered basis of Mϕ |S. A representative of
the H × -orbit of Grassmann-Plücker functions determining M/S is given by
ϕ′′ : (E \ S)r−k −→ H,
x 7→ ϕ(B, x).
(11)
Proof. We note that the formula for ϕ′′ in (11) is given in [5], with proof deferred to [3]; we
give a new proof here.
Proof of (1): Since M/S := (M ∗ \ S)∗ , the formula C ′′ = (C ∗ |(E \ S))∗ follows from the
duality cryptomorphism and the restriction constructions of Propositions 3.4 and 3.7. Recall
that
C ∗ = min X ∈ H E \ {0} X ⊥ Y for all Y ∈ C ,
where “min” means minimal support. Now, the following shows (10):
(C ∗ |(E \ S))∗ = (min X ∈ H E \ {0} X ⊥ Y for all Y ∈ C |(E \ S))⊥
⊥
E
E\S X ∈ H \ {0} and supp(X) ∩ S = ∅
= min X|E\S ∈ H
and X ⊥ Y for all Y ∈ C
Z ⊥ X|E\S for all X ∈ H E \ {0}
E\S
= min Z ∈ H
\ {0}
with supp(X) ∩ S = ∅ and X ⊥ Y for all Y ∈ C
Proof of (2): We prove that ϕ′′ is the Grassmann-Plücker function determined by C ′′ . Let
B be a base of the ordinary matroid Mϕ |S. Then, for any base A of the ordinary matroid
Mϕ /S, A∪ B is a base of Mϕ . Let e ∈ (E \S)\A be given, and let C̃A∪B,e be the fundamental
circuit of e with respect to A ∪ B in M (see, (2) and the paragraph before it). Now suppose
X ∈ H E \ {0} satisfies the conditions that:
supp(X) ∩ S = ∅,
X ⊥ Y ∀ Y ∈ C.
In particular, X ⊥ C̃A∪B,e since C̃A∪B,e ∈ C. On the other hand, as X(s) = {0} ∀ s ∈ S, we
have that
X|E\S ⊥ C̃A∪B,e |E\S .
Hence C̃A∪B,e |E\S = CA,e is the fundamental H-circuit of e ∈ E \ S with respect to A in
M/S by incomparability of supports of elements in C ′′ and the fact that supp(C̃A∪B,e |E\S )
14
is precisely the fundamental circuit of e by A in the underlying matroid of the contraction.
What remains is a computation:
(−1)i ϕ′′ (e, A|[r−k]\{i} )ϕ′′ (A)−1 = (−1)i ϕ(B, e, A|[r−k]\{i} )ϕ(B, A)−1
= C̃A∪B,e (A(i))C̃A∪B,e (e)−1
= CA,e (A(i))CA,e (e)−1 .
3.4 Elementary Properties of Minors. We summarize the constructions of the preceding sections below for easy reference:
Proposition 3.9. Let H be a hyperfield, E a finite set, r a positive integer, and M be a
matroid over H of rank r with circuits C and a Grassmann-Plücker function ϕ. Let S be a
subset of E.
(1) The restriction M |S is given by H-circuits:
C S = {X|S | X ∈ C and supp(X) ⊆ S} .
(2) The restriction M |S is obtained by fixing an ordered base B = (b1 , b2 , · · · , bk ) of the
underlying matroid Mϕ /(E \ S) and defining:
ϕB : S r−k −→ H,
x 7→ ϕ(x, B).
In particular, the H-matroid M |S is determined by the H × -class [ϕB ] of any such B.
(3) The deletion M \ S is given by H-circuits:
C |(E \ S) = X|E\S X ∈ C and S ∩ supp(X) = ∅ .
(4) The deletion M \ S is obtained by fixing an ordered base B = (b1 , b2 , · · · , br−k ) of
Mϕ /S and defining:
ϕB : (E \ S)k −→ H,
x 7→ ϕ(x, B).
In particular, the H-matroid M \ S is determined by the H × -class [ϕB ] of any B.
(5) The contraction M/S is given by H-circuits C = (C ∗ |(E \ S))∗ . More explicitly,
Z ⊥ X|E\S for all X ∈ H E \ {0}
′′
E\S
C = Z ∈H
\ {0}
with supp(X) ∩ S = ∅ and X ⊥ Y for all Y ∈ C
(6) The contraction M/S is given by the class of Grassmann-Plücker functions ((ϕ∗ )E\S )∗ .
More explicitly, let B = (b1 , · · · , bk ) be an ordered basis of Mϕ |S. A representative of
the H × -orbit of Grassmann-Plücker functions determining M/S is given by
ϕ′′ : (E \ S)r−k −→ H,
x 7→ ϕ(B, x).
Using the constructions above, one can easily verify that the following properties hold (cf.
the properties of minors in ordinary matroids):
Corollary 3.10. Let M be an H-matroid on E with S, T ⊆ E disjoint. We have
(1) M/∅ = M = M \ ∅
(2) (M \ S) \ T = M \ (S ∪ T )
(3) (M/S)/T = M/(S ∪ T )
(4) (M \ S)/T = (M/T ) \ S
Proof. The proof of parts (1)-(3) is clear by choosing an appropriate Grassmann-Plücker
function to represent both sides of the equalities using our characterization in Proposition
3.9. To see (4), one can calculate the H-circuits of these and obtain that they are equal.
Finally, we have the following corollary stating that restriction, deletion, and contraction
commute with pushforwards:
15
Corollary 3.11. The following all hold:
(1) The pushforward of a restriction is the restriction of the pushforward.
(2) The pushforward of a deletion is the deletion of the pushforward.
(3) The pushforward of a contraction is the contraction of the pushforward.
Proof. One trivially verifies that the statements hold on Grassmann-Plücker functions.
Example 3.12. Recall that the Krasner hyperfield K is the final object in the category of
hyperfields. Let H be a hyperfield and f : H → K be the canonical map. For any H-matorid
M , the pushforward f∗ M is just the underlying matroid of M . In this case, one can clearly
see Corollary 3.11.
3.5 Direct sums of Matroids over Hyperfields. For matroids over hyperfields, we
provide two cryptomorphic definitions (Grassmann-Plücker functions and circuits) for direct
sum. Let M and N be H-matroids. To state a precise formula for a sum of GrassmannPlücker functions, we will need some additional notation. Fix a total order ≤ on EM ⊔ EN
such that x < y whenever x ∈ EM and y ∈ EN . Now for every x ∈ (EM ⊔ EN )rM +rN either
x has exactly rM components in EM or not. If so, we let σx denote the unique permutation
of [rM + rN ] such that σx · x is monotone increasing with respect to ≤. Now, we have the
following:
Proposition 3.13. Let H be a hyperfield and M (resp. N ) be H-matroids of rank rM
(resp. rN ) given by a Grassmann-Plücker function ϕM (resp. ϕN ). The function ϕM ⊕ ϕN ,
defined by the following formula (a) is a weak-type Grassmann-Plücker function on (EM ⊔
EN )rM +rN :
ϕM ⊕ ϕN : (EM ⊔ EN )rM +rN → H,
(
0, if x does not have precisely rM components in EM
x 7→
sgn(σx )ϕM ((σx · x)|[rM ] )ϕN ((σx · x)|[rM +rN ]\[rM ] ), otherwise.
(a)
Furthermore, ϕM ⊕ ϕN is of strong-type precisely when both ϕM and ϕN are of strong-type.
Moreover, the H × -class of ϕM ⊕ ϕN depends only on M and N .
Proof. Let ϕ := ϕM ⊕ ϕN . We first show that ϕ is a nondegenerate H-alternating function.
Indeed, the function ϕ is clearly nontrivial. To see that ϕM ⊕ ϕN is H-alternating, we let τ be
an arbitrary permutation of [rM + rN ]; note that x does not have precisely rM components in
EM if and only if τ · x does not have precisely rM components in EM . If x does have precisely
rM components in EM , then στ ·x = σx τ −1 and so the following completes the proof of our
claim:
(ϕM ⊕ ϕN )(τ · x) = sgn(στ ·x )ϕM ((στ ·x · τ · x)|[rM ] )ϕN ((στ ·x · τ · x)|[rM +rN ]\[rM ] )
= sgn(σx · τ −1 )ϕM ((σx · τ −1 · τ · x)|[rM ] )ϕN ((σx · τ −1 · τ · x)|[rM +rN ]\[rM ] )
= sgn(τ ) sgn(σx )ϕM ((σx · x)|[rM ] )ϕN ((σx · x)|[rM +rN ]\[rM ] )
= sgn(τ )(ϕM ⊕ ϕN )(x).
Next, we prove that if ϕM and ϕN are weak-type Grassmann-Plücker functions, then ϕ is
also a weak-type Grassmann-Plücker function; we should show that the 3-term GrassmannPlücker relation (WG) holds. In other words, we have to show that ∀ a, b, c, d ∈ EM ⊔ EN
and y ∈ (EM ⊔ EN )(rM +rN )−2 ,
0H ∈ ϕ(a, b, y)ϕ(c, d, y) − ϕ(a, c, y)ϕ(b, d, y) + ϕ(a, d, y)ϕ(b, c, y).
(WG)
Before proceeding, notice that we may assume a < b < c < d and y is strictly increasing
with respect to the ordering ≤ by alternation and degeneration conditions. We may further
16
assume that none of a, b, c, d are coordinates of y by degeneracy.
Case 1 : Suppose all of a, b, c, d belong to the same part of EM ⊔ EN (either EM or EN ). In
this case, we may assume that a, b, c, d ∈ EM . If y does not have exactly rM − 2 components
in EM , then the relation follows trivially as all terms are zero. Otherwise, notice that σx is
identity on the components with elements from EN and ϕN contributes the same constant
to the relation in each term (namely ϕN (y[rM +rN −2]\[rM −2] )2 ). Thus we may reduce to a
consideration of the terms contributed by ϕM , namely
0H ∈ ϕM (ρa,b · (a, b, y′ ))ϕM (ρc,d · (c, d, y′ ))
−ϕM (ρa,c · (a, c, y′ ))ϕM (ρb,d · (b, d, y′ )) + ϕM (ρa,d · (a, d, y′ ))ϕM (ρb,c · (b, c, y′ )),
where ρp,q = σ(p,q,y)|[rM ] and y′ = y|[rM ] . For each p ∈ {a, b, c, d} let
δp := # i ∈ [rM ] y′ (i) < p .
Now one permutes the coordinates in the expression in the following manner. First permute
d to the front of all terms which contain it; this results in a global change of sign (−1)δd +1 as
d must pass over δd coordinates of y′ and the coordinates c in the first term, b in the second
term, and a in the third term. Next permute c to the front of all terms which contain it;
in each term the sign changes by (−1)δc +1 as c must pass over δc coordinates of y′ and the
coordinates d in the first, a in the second, and b in the third. Next permute b to the front
of all terms which contain it; in each term the sign changes by (−1)δb +1 as b must pass over
δb coordinates of y′ and the coordinates a in the first, d in the second, and c in the third.
Finally permute a to the front of all terms which contain it; in each term the sign changes by
(−1)δa +1 as a must pass over δa coordinates of y′ and the coordinates b in the first, c in the
second, and d in the third. Hence permuting coordinates in this way we arrive at the relation
(WG) for ϕM up to a global sign change of (−1)δa +δb +δc +δd +4 . Hence Axiom (WG) holds in
this case.
Case 2 : Suppose not all of a, b, c, d belong to the same part of EM ⊔ EN . By our arrangement
of a < b < c < d and our choice of order ≤ as above we see that a ∈ EM and d ∈ EN . If
y does not have precisely rM − 1 components in EM , then the relation holds trivially as all
terms are zero. Thus, we may further assume that y has precisely rM − 1 components in EM
(and thus precisely rN − 1 components in EN ). Now if b and c belong to the same part of
EM ⊔ EN , again we see that the relation trivially holds as all terms are zero. Thus we can
reduce to the case that b ∈ EM and c ∈ EN . We must see the following to conclude our
desired result:
ϕ(a, c, y)ϕ(b, d, y) = ϕ(a, d, y)ϕ(b, c, y).
(12)
We now obtain the relation by a similar trick as in the first case, “walking” each of d, c, b, a
back to the first coordinate in that order to obtain the relation by corresponding relations on
ϕM and ϕN .
Next, we prove that ϕ is strong-type only if ϕM and ϕN are strong-type. Suppose ϕP is weaktype but not strong-type for either P = M or P = N . Then there is an (rP + 1)-tuple x and
(rP − 1)-tuple y for which (SG) is violated. Pick any ordered basis z of the other H-matroid
P ′ ; trivially ϕM ⊕ ϕN fails (SG) for (x, z) and (y, z), as this reduces to the ϕP ′ (z)-multiple
of the failing relation for ϕP ; this yields that ϕM ⊕ ϕN is weak-type but not strong-type. On
the other hand, if ϕM and ϕN are both strong-type, then the relations required by (SG) for
ϕM ⊕ ϕN can be rewritten as a constant times an (SG)-relation for M plus a constant times
an (SG)-relation for N . This immediately implies that ϕM ⊕ ϕN is strong-type.
Finally, invariance of the resulting H × -class is immediate from the following:
αϕM ⊕ βϕN = αβ(ϕM ⊕ ϕN ),
17
∀ α, β ∈ H × .
Proposition 3.14. Let M and N be H-matroids of rank rM and rN on disjoint ground sets
EM and EN given by H-circuits C M and C N respectively. Define
either both X|EM ∈ C M and X|EN = 0
C M ⊕ C N = X : EM ⊔ EN → H
or both X|EM = 0 and X|EN ∈ C N
Then, C M ⊕ C N is a set of H-circuits. Furthermore, C M ⊕ C N is of strong-type exactly when
both C M and C N are of strong-type.
Proof. That C M ⊕ C N is a set of pre-circuits over H follows trivially from its definition. Moreover, one can see easily see that
supp(C M ⊕ C N ) = supp(C M ) ⊔ supp(C N )
(13)
and hence the underlying matroid of the H-matroid determined thereby is the direct sum of
the underlying matroids of the summands. It follows that every modular pair in C M ⊕ C N
reduces to two modular pairs, one in C M and one in C N . Thus (WC) holds by noting that
any modular pair with nontrivial intersection is either a modular pair in C M or a modular
pair in C N . If C M and C N are both strong, then by (13), (SC1) holds. Moreover (SC2) holds
by noting that the computation reduces to a computation in precisely one of C M or C N .
The next result shows that the direct sum of H-matroids admits the cryptomorphic descriptions given in this section.
Proposition 3.15. If M is an H-matroid given by H-circuits C M and Grassmann-Plücker
function ϕM on EM and N is an H-matroid given by H-circuits C N and Grassmann-Plücker
function ϕN on EN such that EM ∩EN = ∅. Then, ϕM ⊕ϕN and C M ⊕ C N both determine the
same H-matroid under cryptomorphism. Furthermore, this matroid has underlying matroid
the direct sum of the underlying matroids of M and N .
Proof. We must verify that C M ⊕ C N is cryptomorphically determined by ϕM ⊕ ϕN . Let BM
and BN be any bases of the underlying matroids of M and N , respectively. Notice that for
all e ∈ (EM ⊔ EN ) \ (BM ⊔ BN ), the fundamental circuit XBM ∪BN ,e has support contained
in EM or in EN . Thus, the cryptomorphism relation required reduces to the relation on
the fundamental circuit the part contiaining e ∈ EM ⊔ EN . Hence ϕM ⊕ ϕN and C M ⊕ C N
determine the same H-matroid as desired.
Corollary 3.16. The pushforward of a direct sum of matroids is the direct sum of the pushforwards. In other words, direct sum commutes with pushforwards.
Proof. This is trivially verified on Grassmann-Plücker functions.
Example 3.17. The case when we pushforward to the Krasner hyperfield K, i.e., taking
underlying matroids, is directly proven in terms in H-circuits in Proposition 3.15.
4.
Isomorphisms of matroids over hyperfields
In this section, we introduce a notion of isomorphisms of matroids over hyperfields which
generalizes the definition of isomorphisms of ordinary matroids. We will subsequently use this
definition to construct matroid-minor Hopf algebras for matroids over hyperfields in §5.
Definition 4.1 (Isomorphism via Grassmann-Plücker function). Let E1 and E2 be finite
sets, r be a positive integer, and H be a hyperfield. Let M1 (resp. M2 ) be a matroid on
E1 (resp. E2 ) of rank r over H which is represented by a Grassmann-Plücker function ϕ1
18
(resp. ϕ2 ). We say that M1 and M2 are isomorphic if there is a bijection f : E1 → E2 and
an element α ∈ H × such that the following diagram commutes:
E1r
ϕ1
H
⊙α
fr
E2r
ϕ2
(14)
H
Proposition 4.2. Definition 4.1 is well-defined.
Proof. Let ϕ′1 and ϕ′2 be different representatives of M1 and M2 . In other words, there exist
β, γ ∈ H × such that ϕ′1 = β ⊙ ϕ1 and ϕ′2 = γ ⊙ ϕ2 . In this case, we have that
γ −1 ⊙ ϕ′2 ◦ f r = ϕ2 ◦ f r = α ⊙ ϕ1 = (α ⊙ β −1 ) ⊙ ϕ′1
It follows that ϕ′2 ◦ f r = (γ ⊙ α ⊙ β −1 ) ⊙ ϕ′1 and hence Definition 4.1 is well-defined.
Proposition 4.3. Let H and K be hyperfields and g : H → K be a morphism of hyperfields.
If M1 and M2 are matroids over H which are isomorphic, then the pushforwards g∗ M1 and
g∗ M2 are isomorphic as well.
Proof. Let M1 (resp. M2 ) be represented by a Grassmann-Plücker function ϕ1 (resp. ϕ2 ).
Since M1 and M2 are isomorphic, there exist a ∈ H × and a bijection f : E1 → E2 such
that ϕ2 ◦ f r = a ⊙ ϕ1 . Notice that the pushforward g∗ M1 (resp. g∗ M2 ) is represented by the
Grassmann-Plücker function g ◦ ϕ1 (resp. g ◦ ϕ2 ), we obtain
(g ◦ ϕ2 ) ◦ f r = g ◦ (ϕ2 ◦ f r ) = g ◦ (a ⊙ ϕ1 ) = g(a) ⊙ (g ◦ ϕ1 ).
One notes that in the special case K = K, the underlying matroids of two isomorphic
matroids are isomorphic in the classical sense. Therefore, our definition of isomorphisms
generalizes the definition of isomorphisms of ordinary matroids.
Proposition 4.4. If M and M ′ (resp. N and N ′ ) are isomorphic H-matroids, then M ⊕ N
and M ′ ⊕ N ′ are isomorphic H-matroids.
Proof. Consider Grassmann-Plücker functions ϕM , ϕM ′ , ϕN , and ϕN ′ . By assumption there
′ and constants α , α ∈ H × such that
are bijections fM : EM → EM ′ and fN : EN → EN
M
N
rN
rM
αM ⊙ ϕM = ϕM ′ ◦ fM and αN ⊙ ϕN = ϕN ′ ◦ fN . Let fM ⊔ fN : EM ⊔ EN → EM ′ ⊔ EN ′
denote the obvious bijection. Then, we have
αM αN ⊙ (ϕM ⊕ ϕN ) = (αM ⊙ ϕM ) ⊕ (αN ⊙ ϕN )
rN
rM
)
) ⊕ (ϕN ′ ◦ fN
= (ϕM ′ ◦ fM
= (ϕM ′ ⊕ ϕN ′ ) ◦ (fM ⊔ fN )rM +rN .
Remark 4.5. Although we stick with Definition 4.1 in this paper, any ⊕-congruence relation
could be used in place of “isomorphism” for matroids over hyperfields; indeed, all that we will
need from our notion of isomorphism is that M ∼ M ′ and N ∼ N ′ implies M ⊕ N ∼ M ′ ⊕ N ′
to define Hopf algebras for matroids over hyperfields.
Remark 4.6. Our initial definition for isomorphism was as follows: M1 and M2 are isomorphic if there is a bijection f : E1 → E2 and an automorphism g : H → H such that the
following diagram commutes:
E1r
ϕ1
g
fr
E2r
H
ϕ2
19
H
This definition was inspired by the classical notion of semilinear maps, i.e., linear maps up
to “twist” of scalars by the automorphisms of a ground field. Unfortunately, this is not
well-defined on the level of H × -equivalence classes and hence we use the current definition.
Although we do not pursue this line of thought in this paper, it seems really interesting to
investigate a notion of general linear groups over hyperfields. For instance, a proper notion
of general linear groups over hyperfields is needed to study matroid bundles (a combinatorial
analogue of vector bundles, as in [1]) for matroids over hyperfields.
5.
The matroid-minor Hopf algebra associated to a matroid over a hyperfield
In this section, by appealing to Definition 4.1, Propositions 3.13 and 3.14, we generalize the
classical construction of matroid-minor Hopf algebras to the case of matroids over hyperfields.
Let H be a hyperfield. Let M be a set of matroids over H which is closed under taking direct
sums and minors. Let Miso be the set of isomorphisms classes of elements in M, where the
isomorphism class is defined by Definition 4.1. Then, Miso has a canonical monoid structure
as follows:
· : Miso × Miso → Miso , ([M1 ], [M2 ]) 7→ [M1 ⊕ M2 ].
(15)
Note that (15) is well-defined thanks to Proposition 4.4 and the isomorphism class of the empty
matroid [∅] becomes the identity element. Let k be a field. Then we have the monoid algebra
k[Miso ] over k with the unit map η : k → k[Miso ] sending 1 to [∅] and the multiplication:
µ : k[Miso ] ⊗k k[Miso ] → k[Miso ], generated by [M1 ] ⊗ [M2 ] 7→ [M1 ⊕ M2 ].
Proposition 5.1. Let k be a field and H be a hyperfield. Let (Miso , ·) be the monoid and
k[Miso ] be the monoid algebra over k as above. Then H := k[Miso ] is a bialgebra with the
following maps:
• (Comultiplication)
∆ : H −→ H ⊗k H,
[M ] 7→
X
[M |A ] ⊗k [M/A].
(16)
A⊆E
• (Counit)
ε : H −→ k,
[M ] 7→
1 if EM = ∅
0 if EM 6= ∅,
(17)
Furthermore, H is graded and connected and hence has a unique Hopf algebra structure.
Proof. There is a canonical grading on H via the cardinality of the underlying set of each
element [M ] and this is clearly compatible with the bialgebra structure of H. In this case, [∅]
has a degree 0 and hence H is connected. The last assertion simply follows from the result
of [20].
Example 5.2. Let H be the hyperfield which is obtained from Q as in Example 2.15 (considered as a totally ordered abelian group) and k be a field. Let M = U11 (the uniform rank-1
matroid on one element) and Miso be the free monoid generated by the isomorphism class
[M ] of M . Then one can easily see that the Hopf algebra k[Miso ] is just k[T ], where T is the
isomorphism class of U11 .
Let X be the set of matroids over H whose pushforward is U11 . Let [M1 ], [M2 ] ∈ X. Then
M1 and M2 are isomorphic if and only if there exists q ∈ H × = Q such that a ⊙ ϕ1 = ϕ2 ,
where ϕ1 (resp. ϕ2 ) is a Grassmann-Plücker function for M1 (resp. M2 ). It follows that the
free monoid MH
iso , which is generated by the isomorphisms classes of X, is as follows:
n
j
n1 n2
MH
iso = {Tq1 Tq2 · · · Tqj | ni , j ∈ N, qi ∈ Q}.
20
Hence, the Hopf algebra k[MH
iso ] is just k[Tq ]q∈Q . One then has the following surjection:
π : k[MH
iso ] = k[Tq ]q∈Q −→ k[Miso ] = k[T ],
Tq 7→ T.
Note that the map π is a surjection since any matroid in Miso is realizable over H by some
matroid in MH
iso . One can easily see that Ker(π) is generated by elements of the form Tq1 −Tq2
for qi ∈ Q.
6.
Relations to other generalizations
6.1 Relation to matroids over fuzzy rings. In this section, we investigate the results
in previous sections in a view of matroids over fuzzy rings, introduced by A. Dress in [8] (and
later with W. Wenzel in [9]). We will employ the functor, which is constructed by the second
author together with J. Giansiracusa and O. Lorscheid, from the category of hyperfields to
the category of fuzzy rings for this purpose. We note that the most recent work of Baker and
Bowler [6] generalizes matroids over hyperfields and matroids over fuzzy rings at the same
time. For the brief overview of this approach in connection to our previous work, see §6.2.
We first review the definition of matroids over fuzzy rings. In what follows, we let E
be a finite set, K a fuzzy ring, and K × the group of multiplicatively invertible elements of
K, unless otherwise stated. Roughly speaking a fuzzy ring K is a set, equipped with two
binary operations +, · such that (K, +, 0K ) and (K, ·, 1K ) are commutative monoids (but not
assuming that two binary operations are compatible), together with a distinguished subset
K0 and a distinguished element ε, satisfying certain list of axioms. The element ε of K plays
the role of the additive inverse of 1 and K0 is “the set of zeros”; this is where the term “fuzzy”
came from. For the precise definition of fuzzy rings, we refer the readers to [11, §2.3.].
Remark 6.1. We restrict ourselves to the case that E is a finite set to make an exposition
simpler, although one interesting facet of Dress and Wenzel’s theory is that E does not have
to be finite.
Definition 6.2. Let E be a finite set and (K; +, ·; ε, K0 ) a fuzzy ring.
(1) The unit-support of a function f : E → K is defined by
usupp(f ) := f −1 (K × ).
(2) The inner product of two functions f, g : E → K is defined by
X
hf, gi :=
f (e) · g(e).
e∈supp (f )∩supp (g)
(3) Two functions f, g : E → K are orthogonal, denoted f ⊥ g, when hf, gi is an element
of K0 .
(4) The wedge of f, g : E → K is the function
(
0
if x = y
f ∧ g : E × E → K, (x, y) 7→
(18)
f (x) · g(y) + εf (y) · g(x) otherwise
Clearly, we have usupp(f ) ⊆ supp(f ). The following lemma now directly follows from the
definition:
Lemma 6.3. Let f, g, h : E → K be functions. Suppose that f is orthogonal to both g and h.
If
supp (f ) ∩ (supp (g) ∪ supp (h)) ⊆ usupp (f ),
then for all x ∈ E the function (g ∧ h)|{x}×E is orthogonal to f .
21
We further define for all R ⊆ K E ,
^
R := {f1 ∧ f2 ∧ · · · ∧ fn | n ∈ N and for all i ∈ [n], we have fi ∈ R} .
For each R ⊆ K E , we let
n
[R] := r|(x1 ,x2 ,··· ,xn−1 )×E
r = f1 ∧ f2 ∧ · · · ∧ fn ∈
For S ⊆ E and R ⊆ K E , we define
^
o
R and xi ∈ E for i ∈ [n] .
(19)
(20)
RS := {f ∈ R | supp (f ) ∩ S = usupp (f ) ∩ S} .
Definition 6.4. A matroid over K on E is presented by (X, R) if
(1) X ⊆ P(E) is a set of bases of an ordinary matroid,
(2) R is a subset of (K E ) satisfying the following: for any n ∈ N, f = f1 ∧ f2 ∧ · · · ∧ fn
with fi ∈ R for i = 1, ..., n, and (x1 , ..., xn ) ∈ E n with xn ∈ Y for Y ∈ X such that
f (x1 , ..., xn ) ∈
/ K0 , there exists g ∈ R|Y such that
xn ∈ supp(g) ∩ Y ⊆ supp(f ).
Remark 6.5. Two different data (X, R) and (X ′ , R′ ) on E may present the same matroid
over K (see, [8]). Also, it is not difficult to see that the set of circuits of a matroid over K
(i.e. the set of support minimal elements of R) has a supports set which is the set of circuits
of an ordinary matroid.
We now review the functors constructed in [11] to link matroids over hyperfields and matroids over fuzzy rings. Note that Dress and Wenzel also introduced the cryptomorphic
description of matroids over fuzzy rings by using Grassmann-Plücker functions in [9]. There
are two types of morphisms for fuzzy rings (called weak and strong morphisms in [11]). We let
Hyperfields be the category of hyperfields, FuzzyRingswk the category of fuzzy rings with
weak morphisms, and FuzzyRingsstr the category of fuzzy rings with strong morphisms.
Then, one has the following:
Theorem 6.6. [11, §3] There exists a fully faithful functor from Hyperfields to FuzzyRingswk .
The construction goes as follows. For a hyperfield (H, ⊞, ⊙), we let K := P ∗ (H) and
impose two binary operations + and · as follows: for all A, B ⊆ H,
[
A + B :=
a ⊞ b, A · B := {a ⊙ b | a ∈ A, b ∈ B}.
a∈A,b∈B
Then, (K, +, {0H }) and (K, ·, {1H }) become commutative monoids. It is shown in [11] that
with K0 = {A ⊆ H | 0H ∈ A} and ε = {−1H }, (K, +, ·, ε, K0 ) becomes a fuzzy ring. For any
hyperfield H, we let F(H) be the fuzzy ring defined in this way. For morphisms and more
details, we refer the readers to [11].
Remark 6.7. It is also shown in [11] that there exists a quasi-inverse G of the functor F.
Now, we employ the functor F to yield minors of matroids over fuzzy rings as in Dress and
Wenzel from minors of matroids over hyperfields as in Baker and Bowler. In what follows, all
matroids over hyperfields are assumed to be strong. One has the following:
Theorem 6.8. [11, §7.2] Let E be a finite set, H be a hyperfield, K = F(H) be the fuzzy
ring obtained from H, and r a positive integer. Then a function
ϕ : E r −→ H × = F(H)×
is a Grassmann-Plücker function over the fuzzy ring K = F(H) in the sense of Dress and
Wenzel in [9] if and only if ϕ is a strong-type Grassmann-Plücker function over H.
22
For a matroid M over H we abuse notation and let F(M ) be the corresponding matroid
over the fuzzy ring F(H). One has the following corollary:
Corollary 6.9. Let E be a finite set, H a hyperfield, K = F(H) the fuzzy ring obtained from
H, r a positive integer, and S a subset of E. Then, we have
(1) F(M |S) = F(M )|S.
(2) F(M \S) = F(M )\S.
(3) F(M/S) = F(M )/S.
In particular, if N is a minor of M , then F(N ) is a minor of F(M ).
Proof. This directly follows from the Grassmann-Plücker function characterizations of minors
for matroids over hyperfields in §3 and for matroids over fuzzy rings in [9, §5].
Remark 6.10. By using the quasi-inverse G constructed in [11], for field-like fuzzy rings
(see, [11, §4] for the definition), one can also obtain minors of matroids over hyperfields from
minors of matroids over fuzzy rings.
Remark 6.11. One can use our definition of direct sums of matroids over hyperfields in §3.5
to define direct sums for matroids over fuzzy rings and hence obtain matroid-minor Hopf
algebras for matroids over fuzzy rings.
6.2 Relation to matroids over partial hyperfields. In this section, we review Baker
and Bowler’s more generalized framework, namely matroids over partial hyperfields [6] and
explain how our work can be generalized in this setting.
Definition 6.12. [6, §1] A tract is an abelian group G together with a designated subset
NG of the group semiring N[G] such that
(1) 0N[G] ∈ NG and 1G 6∈ NG .
(2) ∃ ! ε ∈ G such that 1 + ε ∈ NG .
(3) G · NG = NG .
The idea is similar to fuzzy rings; ε plays the role of −1 and NG encodes “non-trivial
dependence” relations; in the case of fuzzy rings, one has a designated subset K0 of “zeros”,
however, by using ε one can always change K0 to NG as above.
For a hyperfield (H, ⊞, ⊙), one can canonically associate a tract (G, NG ); this is very similar
to the functor from the category of hyperfield
to the category of fuzzy rings in §6.1. To be
P
precise, one sets G = H × , and lets f =
ai gi ∈ N[G] be in NG if and only if
0H ∈ ⊞(ai ⊙ gi )
(as elements of H).
(21)
Recall that partial fields are introduced by C. Semple and G. Whittle in [19] to study realizability of matroids. A partial field (G ∪ {0R }, R) consists of a commutative ring R and
a multiplicative subgroup G of R× such that −1 ∈ G and G generates R. Inspired by this
definition (along with hyperfields), Baker and Bowler define the following:
Definition 6.13. [6, §1] A partial hyperfield is a hyperdomain R (a hyperring without zero
divisors) together with a designated subgroup G of R× .
One can naturally associate a tract to a partial hyperfield (G, R)
P in a manner similar to
the previous association of a tract to a hyperfield by stating that
ai gi ∈ N[G] if and only
if (21) holds.
With tracts (or partial hyperfields), Baker and Bowler generalize their previous work on
matorids over hyperfields. Their main idea is that in their proofs for matroids over hyperfields,
one only needs the three conditions of tracts given in Definition 6.12. Therefore, although we
only focus on the case of matroids over hyperfields, one can easily generalize our results to
the case of matroids over partial hyperfields.
23
6.3 Tutte polynomials of Hopf algebras and Universal Tutte characters. The Tutte
polynomial is one of the most interesting invariants of graphs and matroids. In [14], T. Krajewsky, I. Moffatt, and A. Tanasa introduced Tutte polynomials associated to Hopf algebras.
More recently, C. Dupont, A. Fink, and L. Moci introduced universal Tutte characters generalizing [14]. In fact, both [10] and [14] consider the case when one has combinatorial objects
which have notions of “deletion” and “contraction” (e.g. graphs and matroids). In the context
of our work, the following is straightforward.
Proposition 6.14. Let H be a hyperfield. A set Miso of isomorphisms classes of matroids
over H, which is stable under taking direct sums and minors, satisfies the axioms of a minor
system in [14, Definition 2.].
Proof. This directly follows from Corollary 3.10.
The term a minors system is used in [10] to define universal Tutte characters. The following
is an easy consequence of §3.
Proposition 6.15. Let H be a hyperfield and MatH be the set species such that MatH (E) is
the set of matroids over H with an underlying set E. Then MatH is a connected multiplicative
minors system as in [10, Definition 2.6. and 2.8.].
Proof. Let S := MatH . Clearly, S is connected since the empty matroid over H is the only
object of S[∅]. Multiplicative structure of S comes from direct sums. The axioms (M1)-(M3),
(M4′-M8′) can be easily checked as in the ordinary matroids case.
Remark 6.16. It follows from the above observations that the construction in [14, §2] can
be applied to define the Tutte polynomial for k[Miso ]. Furthermore, one can also associate
the universal Tutte characters in our setting.
References
[1] Laura Anderson, Matroid bundles, New perspectives in algebraic combinatorics (1999), 1–21.
[2] Laura Anderson and James F Davis, Hyperfield grassmannians, arXiv preprint arXiv:1710.00016 (2017).
[3] Laura Anderson and Emanuele Delucchi, Foundations for a theory of complex matroids, Discrete & Computational Geometry 48 (2012), no. 4, 807–846.
[4] Nicolás Andruskiewitsch and Walter Ferrer Santos, The beginnings of the theory of Hopf algebras, Acta
applicandae mathematicae 108 (2009), 3–17.
[5] Matthew Baker and Nathan Bowler, Matroids over hyperfields, arXiv preprint arXiv:1601.01204 (2016).
, Matroids over partial hyperstructures, arXiv preprint arXiv:1709.09707 (2017).
[6]
[7] Sorin Dascalescu, Constantin Nastasescu, and Serban Raianu, Hopf algebra: An introduction, CRC Press,
2000.
[8] Andreas Dress, Duality theory for finite and infinite matroids with coefficients, Advances in Mathematics
59 (1986), no. 2, 97–123.
[9] Andreas Dress and Walter Wenzel, Grassmann-Plücker relations and matroids with coefficients, Advances
in Mathematics 86 (1991), no. 1, 68–110.
[10] Clément Dupont, Alex Fink, and Luca Moci, Universal Tutte characters via combinatorial coalgebras,
arXiv preprint arXiv:1711.09028 (2017).
[11] Jeffrey Giansiracusa, Jaiung Jun, and Oliver Lorscheid, On the relation between hyperrings and fuzzy
rings, Beiträge zur Algebra und Geometrie / Contributions to Algebra and Geometry 58, no. 4, 735–764.
[12] Darij Grinberg and Victor Reiner, Hopf algebras in combinatorics, arXiv preprint arXiv:1409.8356 (2014).
[13] Jaiung Jun, Geometry of hyperfields, arXiv preprint arXiv:1707.09348 (2017).
[14] Thomas Krajewski, Iain Moffatt, and Adrian Tanasa, Hopf algebras and Tutte polynomials, arXiv preprint
arXiv:1508.00814 (2015).
[15] Marc Krasner, Approximation des corps valués complets de caractéristique p 6= 0 par ceux de caractéristique 0, Colloque d’algébre supérleure (Bruxelles), 1956.
[16] Nikolai E Mnëv, The universality theorems on the classification problem of configuration varieties and
convex polytopes varieties, Topology and geometry–Rohlin seminar, 1988, pp. 527–543.
[17] James G Oxley, Matroid theory, Vol. 3, Oxford University Press, USA, 2006.
24
[18] William R Schmitt, Incidence Hopf algebras, Journal of Pure and Applied Algebra 96 (1994), no. 3,
299–330.
[19] Charles Semple and Geoff Whittle, Partial fields and matroid representation, Advances in Applied Mathematics 17 (1996), no. 2, 184–208.
[20] Mitsuhiro Takeuchi, Free Hopf algebras generated by coalgebras, Journal of the Mathematical Society of
Japan 23 (1971), no. 4, 561–582.
[21] Ravi Vakil, Murphy’s law in algebraic geometry: badly-behaved deformation spaces, Inventiones mathematicae 164 (2006), no. 3, 569–590.
[22] Dominic Welsh, Matroid theory, Courier Corporation, 2010.
Department of Mathematical Sciences, Binghamton University, Binghamton, NY, 13902,
USA
E-mail address: [email protected]
Department of Mathematical Sciences, Binghamton University, Binghamton, NY, 13902,
USA
E-mail address: [email protected]
Department of Mathematics and Statistics, Bostin University, 111 Cummington Mall, Boston,
USA
E-mail address: [email protected]
25
| 0math.AC
|
Learning to Recognize Actions from Limited Training
Examples Using a Recurrent Spiking Neural Model
Priyadarshini Panda1 and Narayan Srinivasa2,*
1 School
of Electrical & Computer Engineering, Purdue University, West Lafayette, IN, USA 47907
Labs, Hillsboro, OR, USA 97124
* Correspondence: [email protected]
arXiv:1710.07354v1 [cs.NE] 19 Oct 2017
2 Intel
ABSTRACT
A fundamental challenge in machine learning today is to build a model that can learn from few examples. Here, we describe
a reservoir based spiking neural model for learning to recognize actions with a limited number of labeled videos. First, we
propose a novel encoding, inspired by how microsaccades influence visual perception, to extract spike information from raw
video data while preserving the temporal correlation across different frames. Using this encoding, we show that the reservoir
generalizes its rich dynamical activity toward signature action/movements enabling it to learn from few training examples. We
evaluate our approach on the UCF-101 dataset. Our experiments demonstrate that our proposed reservoir achieves 81.3%/87%
Top-1/Top-5 accuracy, respectively, on the 101-class data while requiring just 8 video examples per class for training. Our
results establish a new benchmark for action recognition from limited video examples for spiking neural models while yielding
competetive accuracy with respect to state-of-the-art non-spiking neural models.
Introduction
The exponential increase in digital data with online media, surveillance cameras among others, creates a growing need to
develop intelligent models for complex spatio-temporal processing. Recent efforts in deep learning and computer vision have
focused on building systems that learn and think like humans1, 2 . Despite the biological inspiration and remarkable performance
of such models, even beating humans in certain cognitive tasks3 , the gap between humans and artificially engineered intelligent
systems is still great. One important divide is the size of the required training datasets. Studies on mammalian concept
understanding have shown that humans and animals can rapidly learn complex visual concepts from single training examples4 .
In contrast, the state-of-the-art intelligent models require vast quantities of labeled data with extensive, iterative training to
learn suitably and yield high performance. While the proliferation of digital media has led to the availability of massive raw
and unstructured data, it is often impractical and expensive to gather annotated training datasets for all of them. This motivates
our work on learning to recognize from few labeled data. Specifically, we propose a reservoir based spiking neural model that
reliably learns to recognize actions in video data by extracting long-range structure from a small number of training examples.
Reservoir or Liquid Computing approaches have shown surprising success in recent years on a variety of temporal data
based recognition problems (though effective vision applications are still scarce)5–7 . This can be attributed to populations
of recurrently connected spiking neurons, similar to the mammalian neocortex anatomy8 , which create a high-dimensional
dynamic representation of an input stream. The advantage with such an approach is that the reservoir (with sparse/random
internal connectivity) implicitly encodes the temporal information in the input and provides a unique non-linear description
for each input sequence, that can be trained for read out by a set of linear output neurons. In fact, in our view, the non-linear
integration of input data by the high-dimensional dynamical reservoir results in generic stable internal states that generalize over
common aspects of the input thereby allowing rapid learning from limited examples. While the random recurrent connectivity
generates favorable complex dynamic activity within a reservoir, it poses unique challenges that are generally not encountered
in constructing feedforward/deep learning networks.
A pressing problem with recurrently connected networks is to determine the connection matrix that would be suitable for
making the reservoir perform well on a particular task, because it is not entirely obvious how individual neurons should spike
while interacting with other neurons. To address this, Abbott et. al9 , proposed an elegant method for constructing recurrent
spiking networks, termed as Autonomous (A) models, from continuous variable (rate) networks, called Driven (D) models. The
basic idea in the D/A approach is to construct a network (Auto) that performs a particular task by copying another network
(Driven) statistics that does the same task. The copying provides an estimate of the internal dynamics for the autonomous
network including currents at individual neurons in the reservoir to produce the same outputs. Here, we adopt the D/A based
reservoir construction approach and modify it further by introducing approximate delay apportioning to develop a recurrent
spiking model for practical action recognition from limited video data. Please note, we use Auto/Autonomous interchangeably
in the remainder of the text.
Further, in order to facilitate limited example training in a reservoir spiking framework, we propose a novel spike based
encoding/pre-processing method to convert the raw pixel valued videos in the dataset into spiking information that preserves
the temporal statistics and correlation across different frames. This approach is inspired by how MicroSaccades (MS) cause
retinal ganglion cells to fire synchronously corresponding to contrast edges after the MS. The emitted synchronous spike volley
thus rapidly transmits the most salient edges of the stimulus, which often constitute the most crucial information10 . In addition,
our encoding eliminates irrelevant spiking information due to ambiguity such as, noisy background activity or jitter (due to
unsteady camera movement), further making our approach more robust and reliable. Hence, our encoding, in general, captures
the signature action or movement of a subject across different videos as spiking information. This, in turn, enables the reservoir
to recognize/generalize over motion cues from spiking data, to enable learning various types of actions from few video samples
per class. Our proposed spiking model is much more adept at learning object/activity types, due in part to our ability to analyze
activity dynamically as it takes place over time.
Reservoir Model: Framework & Implementation
Model architecture
Let us first discuss briefly the Driven/Autonomous model approach9 . The general architecture of the Driven/ Autonomous
reservoir model is shown in Fig. 1 (a). Both driven/auto models are 3-layered networks. The input layer is connected to a
reservoir that consists of nonlinear (Leaky-Integrate-and-Fire, LIF), spiking neurons recurrently connected to each other in a
random sparse manner (generally with a connection probability of 10%), and the spiking output produced by the reservoir is
then projected to the output/readout neurons. In case of driven, the output neurons are rate neurons that simply integrate the
spiking activity of the reservoir neurons. On the other hand, the auto model is an entirely spiking model with LIF neurons at the
output as well as within the reservoir. Please note, each reservoir neuron here is excitatory in nature and the reservoir does not
have any inhibitory neuronal dynamics in a manner consistent with conventional Liquid or Reservoir Computing.
Essentially, the role of the driven network is to provide targets ( fout ) for the auto model, which is the spiking network we
are trying to construct for a given task. To achieve this, the driven network is trained on an input ( fD ) that is a high pass filtered
version of the desired output ( fout ),
fD = fout + τ f ast
d fout
dt
(1)
The connections from the reservoir to the readout neurons (w) of the driven model are then trained to produce desired
activity using the Recursive Least Square (RLS) rule11, 12 . Once trained, the driven network connections (J f ast ) are imported
into the autonomous model and a new set of network connections (Jslow ) are added for each fast connection. The combination
of the two sets of synapses, J f ast and Jslow , allows the auto model to match its internal dynamics to that of the driven network
that is already tuned to producing desired fout . In fact, the connections u and Jslow in the auto model are derived from the driven
network as, u = uD uR and J = uD Bw, where uR and B are randomly chosen. Note, the time constants of the two sets of synapses
is different i.e. τslow ∼ 10 ∗ τ f ast that allows the autonomous network to slowly adjust its activity to produce fout during training.
For further clarification, please refer to9 to gain more details and insights on D/A approach. After the autonomous model is
constructed, we train its output layer (that consist of LIF spiking neurons), w0 , using a supervised Spike Timing Dependent
Plasticity (STDP) rule13 based on the input fin .
The motif behind choosing fD as a filtered version of fout is to compensate for the synaptic filtering at the output characterized
by time constant, τ f ast 9, 14 . However, it is evident that the dependence of fD on fout from Eqn. 1 imposes a restriction on the
reservoir construction, that is, the number of input and output neurons must be same. To address this limitation and extend the
D/A construction approach to variable input-output reservoir models, we propose an approximate definition of the driven input
using a delay apportioning method as,
( fD i ) j = ( fout ) j + τ f ast i
(d fout ) j
dt
(2)
Here, i varies from 1 to X, j varies from 1 to Y, where X/Y are the number of input/output neurons in our reservoir
model respectively and τ f ast = ∑Xi=1 τ f ast i . Fig. 1 (b, Top) shows the architecture of the D/A model with the proposed delay
apportioning. For each output function (( fout ) j ), we derive the driven inputs ( fD i=1...X ) j by distributing the time constant across
different inputs. We are essentially compensating the phase delay due to τ f ast in an approximate manner. Nevertheless, it turns
2/17
Original D/A Model
(b)
Proposed D/A Model
Firing rate
(a)
Figure 1. (a) Structure of Driven (Autonomous) networks with rate (spiking) output neurons and fast (fast-slow) synaptic
connections. Here, fast connections have equivalent delay time constant. (b) (Top) Structure of Driven-Autonomous model
with the proposed τ f ast apportioning to construct variable input(denoted as X)/output (denoted as Y ) reservoir models. Here,
each fast connection has a different delay time constant. (Bottom) Desired output produced by an autonomous model and its
corresponding Driven model, that is constructed using approximate τ f ast apportioning. Please note, input spiking activity (blue
curve) and output (red/green curve) in driven are similar since input is a filtered version of the output, while autonomous model
works on random input data.
out this approximate version of fD still drives the driven network to produce the desired output and derive the connection matrix
of the autonomous model that is then trained on the real input (video data in our case). Fig. 1 (b, Bottom) illustrates the desired
output genereated by a D/A model of 2(input)×400(reservoir neurons)×1(output) configuration constructed using Eqn. 2 .
Note, for the sake of convenience in representation and comparative purposes, the output of both driven/auto models are shown
as rate instead of spiking activity. It is clearly seen in the driven case that the input spiking activity for a particular input neuron
fD i , even with τ f ast apportioning, approximates the target quite well. Also, the autonomous model derived from the driven
network in Fig. 1 (b, Bottom), produces the desired spiking activity in response to random input upon adequate training of w0
connections.
Supervised STDP & One-Hot Encoding
Reservoir framework relaxes the burden of training by fixing the connectivity within the reservoir and that from input to the
reservoir. Merely, the output layer of readout neurons (w, w0 in Fig. 1 (a)) is trained (generally in a supervised manner) to
match the reservoir activity with the target pattern. In our D/A based reservoir construction, readout of the driven network is
trained using standard RLS learning12 while we use a modified version of the supervised STDP rule proposed in13 to conduct
autonomous model training. An illustration of our STDP model is shown in Fig. 2 (a). For a given target pattern, the synaptic
weights are potentiated or depressed as
∆w = xtrace (ttarget − tactual )
(3)
Where tactual /ttarget denotes the time of occurrence of actual/desired spiking activity at the post-synaptic neuron during the
simulation and xtrace , namely the presynaptic trace, models the pre-neuronal spiking history. Every time a pre-synaptic neuron
fires, xtrace , increases by 1, otherwise it decays exponentially with τ pre that is of the same order as τ f ast . Fundamentally, as
illustrated in Fig. 2 (a), Eqn. 3 depresses (or potentiates) the weights at those time instants when actual (or target) spike activity
occurs. This ensures that actual spiking converges toward the desired activity as training progresses. Once the desired/actual
activity become similar, the learning stops, since at time instants where both desired and actual spike occur simulataneously, the
weight update value as per Eqn. 3 becomes 0.
Since the main aim of our work is to develop a spiking model for action recognition in videos, we model the desired spiking
activity at the output layer based on one-hot encoding. As with standard machine learning applications, the output neuron
assigned to a particular class is trained to produce high spiking activity while the remaining neurons are trained to generate zero
3/17
(a)
(b)
OutputNeuron 1
Class 1
Class 2
OutputNeuron 2
Class 3
OutputNeuron 3
Figure 2. (a) Supervised STDP rule (defined by Eqn. 3) using pre-synaptic trace to perform potentiation/depression of
weights. t1 ,t2 ...t7 are the time instants at which target/actual spiking activity occurs resulting in a weight change. The extent of
potentiation or depression of weights ∆w is proportional to the value of the pre-synaptic trace x1 , x2 ...x7 at a given time instant.
(b) One-hot spike encoding activity for a 3-class problem: Based on the class of the training input, the corresponding output
neuron is trained to produce high activity while remaining neurons yield near-zero activity.
activity. Fig. 2 (b) shows the one-hot spike encoding for a 3-class problem, wherein the reservoir’s readout has 3 output neurons
and based on the training input’s class, the desired spiking activity (sort of ‘bump’) for each output neuron varies. Again, we
use the supervised STDP rule described above to train the output layer of the autonomous model. Now, it is obvious that since
the desired spiking activity of the auto model ( fout referring to Fig. 1) follows such one-hot spike encoding, the driven network
(from which we derive our auto model) should be constructed to produce equivalent target output activity. To reiterate, the auto
model (entirely spiking) works on the real-world input data and is trained to perform required tasks such as classification. The
connection matrix (Jslow ) of the auto model is derived from a driven network (with continuous rate-based output neurons) that,
in turn, provides the target activity for the output neurons of the auto model.
Input Processing
In this work, we use the UCF101 dataset15 to perform action recognition with our reservoir approach. UCF101 consists of
13320 realistic videos (collected from YouTube) that fall into 101 different categories or classes, 2 of which are shown in Fig. 3.
Since the reservoir computes using spiking inputs, we convert the pixel valued action information to spike data in steps outlined
below.
Pixel Motion Detection
Unlike static images, video data are a natural target in event-based learning algorithms due to their implicit time-based
processing capability. First, we track the moving pixels across different frames so that the basic action/signature is captured.
This is done by monitoring the difference of the pixel intensity values (Pdi f f ) between consecutive frames and comparing the
difference against some threshold (ε, that is user-defined) to detect the moving edges (for each location (x,y) within a frame), as
Pdi f f (x, y) = Frame j − Frame j−1 ; (where j = 2 to NumberOfFrames)
If Pdi f f (x, y) ≥ ε, Spikedi f f (x, y) = 1
∑ εi Spikedi f f (x, y)
W Spikedi f f (x, y) = i
; (where i=1 to N)
∑i εi
(4)
The thresholding detects the moving pixels and yields an equivalent spike pattern (Spikedi f f ) for each frame. It is evident
that the conversion of the moving edges into spikes preserves the temporal correlation across different frames. While Spikedi f f
encodes the signature action adequately, we further make the representation robust by varying the threshold (εi = {1,2,4,8,16,32};
N=6 in our experiments) and computing the weighted summation of the spike differences for each threshold as shown in the
last part of Eqn. 4. The resultant weighted spike difference W Spikedi f f (x, y) is then binarized to yield the final spike pattern for
each frame. This kind of pixel-change based motion detection to generate spike output is now possible with a new breed of
energy efficient sensors inspired by the retina16 .
4/17
Weighted
Spike
Spike
Difference
Weighted
Spike
Time
Spike
Difference
Background
Figure 3. Fixed threshold spiking input (denoted by Spike Difference) and variable threshold weighted input data (denoted by
Weighted spike) shown for selected frames for two videos of the UCF101 dataset. As time progresses, the correlation across
spiking information is preserved. Encircled regions show the background captured in the spiking data.
Fig. 3 shows the spiking inputs captured from the moving pixels based on fixed threshold spike input and varying threshold
weighted input for selected frames. It is clearly seen that the weighted input captures more relevant edges and even subtle
movements, for instance, in the PlayingGuitar video, the delicate body movement of the person is also captured along with the
significant hand gesture. Due to the realistic nature of the videos in UCF101, there exists a considerable amount of intra class
variation along with ambiguous noise. Hence, a detailed input pattern that captures subtle gestures facilitates the reservoir in
differentiating the signature motion from noise. It should be noted that we convert the RGB pixel video data into grayscale
before performing motion detection.
Scan based Filtering
We observe from Fig. 3, for the BabyCrawling case, that along with the baby’s movement, some background activity is also
captured. This background spiking occurs on account of the camera movement or jitter. Additionally, we see that majority of
the pixels across different frames in Fig. 3 are black since the background is largely static and hence does not yield any spiking
activity. To avoid the wasteful computation due to non-spiking pixels as well as circumvent insignificant activity due to camera
motion, we create a Bounding Box (BB) around the Centre of Gravity (CoG) of spiking activity for each frame and scan across
5 directions as shown in Fig. 4. The CoG (xg , yg ) is calculated as the average of the active pixel locations within the BB. In
general, the dimensionality of the video frames in the UCF101 dataset is ∼ 200 × 300. We create a BB of size 41 × 41 centered
at the CoG, capture the spiking activity enclosed within the BB, in what we refer to as the Center (C) scan, and then stride
the BB along different directions (by number of pixels equal to ∼ half the distance between the CoG and the edge of the BB),
namely, Left (L), Right (R), Top (T), Bottom (B), to capture the activity in the corresponding directions. These five windows of
spiking activity can be interpreted as capturing the salient edge pixels (derived from pixel motion detection above) after each
microsaccade10 . In our model, the first fixation is at the center of the image. Then, the fixation is shifted from the center of the
image to the CoG using the first microsaccade. The rest of the four microsaccades are assumed to occur in four directions (i.e.,
top, bottom, left and right) with respect to CoG of the image by the stride amount. It should be noted that this sequence of four
miscrosaccades is assumed to be always fixed in its direction. The stride amount can, however, vary based on the size of the BB.
Fig. 4 (a) clearly shows that the CRLTB scans only retain the relevant region of interest across different frames of the
BabyCrawling video while filtering out the irrelevant spiking activity. Additionally, we also observe that in the PlayingGuitar
video shown in Fig. 4 (b), different scans capture diverse aspects of the relevant spiking activity, for instance, hand gesture in C
scan while guitar shape in B scan shown in the 3rd row. This ensures that all significant movements for a particular action in a
given frame are acquired in a holistic manner.
In addition to the scan based filtering, we also delete certain frames to further eliminate the effect of jitter or camera
movement on the spike input data. We delete the frames based on two conditions:
• If the overall number of active pixels for a given spiking frame are greater/lesser than a certain maximum/minimum
threshold, then, that frame is deleted. This eliminates all the frames that have flashing activity due to unsteady camera
movement. In our experiments, we use a min/max threshold value of 8%/75% of the total number of pixels in a given
5/17
(a)
Bounding
Box
CoG
*
(b)
*
*
Original
Center
Left
Right
Top
Bottom
Original
Center
Left
Right
Top
Bottom
Figure 4. Scan based filtering from the original spike data resulting in Center, Left, Right, Top, Bottom scans per frame for (a)
BabyCrawling video (b) PlayingGuitar video. Each row shows the original frame as well as the different scans captured.
frame. Note, we perform this step before we obtain the CRLTB scans.
• Once the CRLTB scans are obtained, we monitor the activity change for a given frame in RLTB as compared to C (i.e.
for Framei check Ci − {Ri , Li , Ti , Bi }) . If the activity change (say, Ci − Ri ) is greater than the mean activity across Ci ,
then, we delete that particular frame (Ri ) corresponding to the scan that yields high variance.
It is noteworthy to mention that the above methods of filtering eliminate the ambiguous motion or noise in videos with
minimal background activity or jitter motion (such as PlayingGuitar) to a large extent. However, the videos where the
background has significant motion (such as HorseRacing) or those where multiple human subjects are interacting (such as
Fencing, IceDancing), the extracted spike pattern does not encode a robust and reliable signature motion due to extreme
variation in movement. For future implementations, we are studying using a DVS camera16, 17 to extract reliable spike data
despite ego motion. Nonetheless, our proposed input processing method yields reasonable accuracy (shown later in Results
section) with the D/A model architecture for action recognition with limited video examples. Supplementary Video 1,2 give
better visualization of the spiking data obtained with the pixel motion detection and scan based filtering scheme.
Experimental Methodology
Network Parameters
The proposed reservoir model and learning was implemented in MATLAB. The dynamics of the reservoir neurons in the
Driven/Autonomous models are described by
τmem
τmem
dV
= Vrest −V + (J f ast f (t) + uD fD ); Driven
dt
dV
= Vrest −V + (J f ast f (t) + Jslow s(t) + u fin ); Auto
dt
(5)
where τmem = 20 ms, fin denotes the real input spike data, fD is the derived version of the target output (refer to Eqn. 2), uD /u
are the weights connecting the input to the reservoir neurons in the driven/auto model respectively. Each neuron fires when its
membrane potential, V , reaches a threshold Vth = -50 mV and is then reset to Vreset = Vrest = -70 mV. Following a spike event,
the membrane potential is held at the reset value Vreset for a refractory period of 5 ms. The parameters f (t), s(t) correspond
to fast/slow synaptic currents observed in the reservoir due to the inherent recurrent activity. In the driven model, only fast
synaptic connections are present while auto model has both slow/fast connections. The synaptic currents are encoded as traces,
that is, when a neuron in the reservoir spikes, f (t) (and s(t) in case of auto model) is increased by 1, otherwise it decays
exponentially as
τslow
ds(t)
= −s(t);
dt
τ f ast
d f (t)
= − f (t)
dt
(6)
where τslow = 100 ms, τ f ast = 5 ms. Both Jslow and J f ast recurrent connections within the reservoir of D/A model are fixed
during the course of training. J f ast of the driven model are randomly drawn from a normal distribution. Jslow and u (refer to
6/17
Fig. 1 (a)) of the auto network are derived from the driven model using uR = 1, B = random number drawn from a uniform
distribution in the range [0, 0.15]. We discuss the impact of variation of B on the decoding capability of the autonomous
reservoir in a later section. If the reservoir outputs a spike pattern, say x(t), the weights from the reservoir to the readout (w) are
trained to produce the desired activity, fout . The continuous rate output neurons in driven network are trained using RLS such
that w ∗ x(t) ∼ fout (t). The spiking output neurons of the auto model (with similar neuronal parameters as Eqn. 5) integrate the
net current w0 ∗ x(t) to fire action potentials, that converge toward the desired spiking activity with supervised STDP learning.
D/A model for Video Classification
The input processing technique yields 5 different CRLTB scan spike patterns per frame for each video. In the D/A based
reservoir construction, each of the 5 scans for every video is processed by 5 different autonomous models. Fig. 5 shows
an overview of our proposed D/A construction approach for video classification. Each scan is processed separately by the
corresponding autonomous model. An interesting observation here is that since the internal topology across all auto models is
equivalent, they can all be derived from a single driven model as shown in Fig. 5. In fact, there is no limit to the number of
autonomous models that can be extracted from a driven network, which is the principal advantage of our approach. Thus, in
case, more scans based on depth filtering or ego motion monitoring are obtained during input processing, our model can be
easily extended to incorporate more autonomous models corresponding to the new scans.
Figure 5. Proposed Driven/Autonomous construction scheme for processing the CRLTB scans obtained from the scan based
filtering
During training, the driven network is first trained to produce the target pattern (similar to the one-hot spike encoding
described in Fig. 2 (b)). Then, each autonomous model is trained separately on the corresponding input scan to produce the
target pattern. In all our experiments, we train each auto model for 30 epochs. In each training epoch, we present the training
input patterns corresponding to all classes sequentially (for instance, Class1→Class 2→Class 3) to produce target patterns
similar to Fig. 2 (b). Each input pattern is presented for a time period of 300 ms (or 300 time steps) . In each time step, a
particular spike frame for a given scan is processed by the reservoir. For patterns where total number of frames is less than 300,
the network still continues to train (even in absence of input activity) on the inherent reservoir activity generated due to the
recurrent dynamics. In fact, this enables the reservoir to generalize its’ dynamical activity over similar input patterns belonging
to the same class while discriminating patterns from other classes. Before presenting a new input pattern, the membrane
potential of all neurons in the reservoir are reset to their resting values.
After training is done, we pass the test instance CRLTB scans to each auto model simultaneously and observe the output
activity. The test instance is assigned a particular class label based on the output neuron class that generated the highest spiking
response. This assignment is done for each auto model. Finally, we take a majority vote across all the assignments (or all auto
models) for each test input to predict its class. For instance, if majority of the auto models (i.e. ≥ 3) predict a class (say Class
1), then predicted output class from the model is Class 1. Now, a single auto model might not have sufficient information to
recognize the test input correctly. The remaining scans or auto models, in that case, can compensate for the insufficient or
inaccurate output prediction from the individual models. While we reduce the computational complexity of training by using
the scan based filtering approach that partitions the input space into smaller scans, the parallel voting during inference enhances
the collective decision making of the proposed architecture, thereby improving the overall performance.
The reservoir (both driven/auto) topology in case of the video classification experiments consist of X(InputNeurons) −
N(ReservoirNeurons) − Y (Out putNeurons) with p% connectivity within the reservoir. X is equivalent to the size of the
bounding box used during scan based filtering (in our case X = 41 × 41 = 1681). The number of output neurons vary based
on the classification problem, that is, Y = 5 for a 5-class problem. In most experiments below, we use p = 10% connectivity
to maintain the sparseness in the reservoir. In the Results section below, we further elaborate on the role of connectivity on
the performance of the reservoir. The number of neurons N are varied based upon the complexity of the problem in order to
achieve maximum classification accuracy.
7/17
(a)
Training
Testing (Correct)
Testing (Incorrect)
GolfSwing
(b)
BenchPress
Figure 6. Training/Testing video clips of the 2-class problem (a) GolfSwing (b) BenchPress: Column 1 shows the training
video, Column 2/3 show the test video examples correctly/incorrectly classified with our D/A model.
Classification Accuracy on UCF101
First, we considered a 2-class problem of recognizing BenchPress/GolfSwing class in UCF101 to test the effectiveness of our
overall approach. We specifically chose these 2 classes in our preliminary experiment since majority of videos in these categories
had a static background and the variation in action across different videos was minimal. We simulated an 800N-2Output, 10%
connectivity D/A reservoir (N:number of reservoir neurons) and trained the autonomous models with a single training video
from each class as shown in Fig. 6. It is clearly seen that even with single video training, the reservoir correctly classifies a host
of other examples irrespective of the diverse subjects performing the same action. For instance, in the golf examples in Fig. 6
(a), the reservoir does not discriminate between a person wearing shorts or trousers while recognizing the action. Hence, we
can gather that the complex transient dynamics within a reservoir latches onto a particular action signature (swing gesture in
this case) that enables it to classify several test examples even under diverse conditions. Investigating the incorrect predictions
from the reservoir for both classes, we found that the reservoir is sensitive toward variation in views and inconsistencies in
action, which results in inaccurate prediction. Fig. 6 shows that the incorrect results are observed when the view angle of
the training/test video changes from frontal to adjacent views. In fact, in case of BenchPress, even videos with similar view
as training are incorrectly classified due to disparity in motion signature (for instance, number of times the lifting action is
performed). Note, we only show a selected number of videos classified correctly/incorrectly in Fig. 6.
To account for such variations, we increased the number of training examples from 1 to 8. In the extended training set, we
tried to incorporate all kinds of view/signature based variations for each action class as shown in Fig. 7 (a). For the 2-class
problem, with 8 training examples for each class, we simulated a 2000N-10% connectivity reservoir that yields an accuracy
of 93.25%. Fig. 7 (b) shows the confusion matrix for the 2-class problem. BenchPress has more misses than GolfSwing as
the former has more variation in action signature across different videos. Now, it is well-known that the inherent chaotic or
spontaneous activity in a reservoir enables it to produce a wide variety of complex output patterns in response to an input
stimulus. In our opinion, this behavior enables data efficient learning as the reservoir learns to generalize over a small set of
training patterns and produce diverse yet converging internal dynamics that transform different test inputs toward the same
output. Note, testing is done on all the remaining videos in the dataset that were not selected for training.
Next, we simulated a reservoir of 4K/6K/8K/9K neurons (with 10% connectivity) to learn the entire 101 classes in the
UCF101 dataset. Note, for each of the 101 classes, we used 8 different training videos that were randomly chosen from the
dataset. To ensure that more variation is captured for training, we imposed a simple constraint during selection, that each
training example must have a different subject/person performing the action. Fig. 8 (a) illustrates the Top-1, Top-3, Top-5
accuracy observed for each topology. We also report the average number of tunable parameters in each case. The reservoir has
fixed connectivity from the input to reservoir as well as within the reservoir that are not affected during training. Hence, we only
compare the total number of trainable parameters (comprising the weights from reservoir to output layer) as the reservoir size
varies in Fig. 8 (a). We observe that the reservoir with 8K neurons yields maximum accuracy (Top−1 : 80.2%, Top−5 : 86.1%).
Also, the classification accuracy improves as the reservoir size increases from 4K → 8K neurons. This is apparent since
larger reservoirs with sparse random connectivity generate richer and heterogeneous dynamics that overall boosts the decoding
8/17
(a)
(b)
Figure 7. (a) Clips of the 8 training videos for each class (BenchPress/GolfSwing) incorporating multiple views/variation in
action signature (b) Confusion Matrix showing the overall as well as class-wise accuracy for 2-class BenchPress/GolfSwing
classification
capability of the network. Thus, for the same task difficulty (in this case 101-class problem), increasing the reservoir size
improves the performance. However, beyond a certain point, the accuracy saturates. Here, the reservoir with 9K neurons
achieves similar accuracy as 8K. In fact, there is a slight drop in accuracy from 8K → 9K. This might be a result of some sort of
overfitting phenomenon as the reservoir (with more parameters than necessary for a given task) might begin to assign, instead
of generalizing, its dynamic activity toward a certain input. Note, all reservoir models discussed so far have 10% connectivity.
Hence, the reservoirs in Fig. 8 (a) have increasing number of connections per neuron (1 : 400 → 1 : 900) as we increase the
number of neurons in the reservoir from 4000 to 9000. Please refer to the supplementary information (see Supplementary Fig.
S1, S2) that details out the confusion matrix for the 101-class problem with 8K neurons and 10% connectivity.
(a)
(b)
Reservoir Model
(10% Connectivity)
# Tunable
Parameters
Top-1
(%)
Top-3
(%)
Top-5
(%)
4000
16.2M
70.2
74.3
75.8
6000
24.2M
77.1
80.9
82.3
8000
32.3M
80.2
84.5
86.1
9000
36.4M
80
84.1
85.8
Increasing Sparseness
C: Connections
Figure 8. (a) Top-1/3/5 accuracy on 101-classes obtained with 8-training example (per class) learning for D/A models of
different topologies (b) Effect of variation in sparsity (for fixed 1 : 800 number of connections per neuron across different
reservoir topologies) on the Top-1 accuracy achieved by the reservoir. Note, all results are performed on the D/A reservoir
scheme with 8 training videos per class processed with the input spike transformation technique mentioned earlier.
Impact of Variation of Reservoir Sparsity on Accuracy
Since sparsity plays an important role in determining the reservoir dynamics, we simulated several topologies in an isoconnectivity scenario to see the effect of increasing sparseness on the reservoir performance. Fig. 8 (b) shows the variation
in Top-1 accuracy for 101-classes as the reservoir sparseness changes. From Fig. 8 (a), we observe that the 8000N-10%
connectivity reservoir (with 1 : 800 number of connections per neuron) gave us the best results. Hence, in the sparsity analysis,
we fixed the number of connections per neuron to 1 : 800 across different reservoir topologies. Initially, for a 2000N-40%
connectivity reservoir, the accuracy observed is pretty low (60.1%). This is evident as lesser sparsity limits the richness in the
complex dynamics that can be produced by the reservoir that further reduces its’ learning capability. Increasing the sparseness
improves the accuracy as expected. We would like to note that the 4000N-20% connectivity reservoir yields lesser accuracy
than that of the sparser 4000N-10% connectivity reservoir of Fig. 8 (a). Another interesting observation here is that the
10,000N-8% reservoir yields higher accuracy (∼ 1% more) than that of the 8000N-10% reservoir. This result supports the
fact that in an iso-connectivity scenario, more sparseness yields better results. However, the accuracy again saturates with a
risk of overfitting or losing generalization capability beyond a certain level of sparseness as observed in Fig. 8 (b). Hence,
there is no fixed rule to determine the exact connectivity/sparsity ratio for a given task. We empirically arrived at the results
by simulating a varying set of reservoir topologies. However, in most cases, we observed that for an N-reservoir model 10%
connectivity generally yields reasonable results as seen from Fig. 8 (a). It is worth mentioning that while the 10K neuron
reservoir yields the highest accuracy of 81.3%, it consists of larger number of tunable parameters (total connections/weights
9/17
from reservoir to output) 40.4M as compared to 32.3M observed with the 8K-10% connectivity reservoir (Fig. 8 (a)). In
fact, the best accuracy achieved with our approach corresponds to the 10K-8% connectivity reservoir with Top-1/Top-3/Top-5
accuracy as 81.3%/85.2%/87%, respectively. Hence, depending upon the requirements, the connectivity/size of the reservoir
can be varied to achieve a favorable efficiency-accuracy tradeoff. Please note that the total number of tunable parameters in the
reservoir model include all 5-autonomous models.
EigenValue Spectra Analysis of Reservoir Connections
We analysed the EigenValue (EV) spectra of the synaptic connections (both Jslow and J f ast of the autonomous models) to study
the complex dynamics of the reservoir. EV spectra has been shown to characterize the memory/learning ability of a dynamical
system18 . Rajan et. al18, 19 have shown that stimulating a reservoir with random recurrent connections activates multiple modes
where neurons start oscillating with individual frequencies. These modes then superpose in a highly non-linear manner resulting
in the complex chaotic/ persistent activity observed from a reservoir. The EVs of the synaptic matrix typically encode the
oscillatory behavior of each neuron in the reservoir. The real part of an EV corresponds to the decay rate of the associated
neuron, while frequency is given by the imaginary part. If the real part of a complex EV exceeds 1 (or Re(EV)) > 1), the
activated mode leads to long lasting oscillatory behavior. If there are a few modes with Re(EV)) > 1 , the network is at a fixed
point, meaning the reservoir exhibits the same pattern of activity for a given input. It is desirable to construct a reservoir with
multiple fixed points so that each one can be used to retain a different memory. In a recognition scenario, each of the fixed
points latch onto a particular input pattern or video in our case, thereby yielding a good memory model. As the number of
modes with Re(EV) > 1 increases, the reservoir starts generating chaotic patterns, that are complex and non-repeating. In order
to construct a reservoir model with good memory, we must operate in a region between singular fixed point activity (Re(EV)
<< 1) and complete chaos (Re(EV) >> 1).
Figure 9. (a) Change in EigenValue spectra of the J f ast and Jslow connections of an autonomous network . Note, the J f ast
spectra (corresponding to the connections imported from driven model) remains constant, while the Jslow spectra changes with
varying B values. (b) Impact of variation in B, representative of stable/chaotic reservoir dynamics, on the accuracy for 2-class
(BenchPress/GolfSwing) 1-training example problem (Fig. 6). The minimum and maximum accuracy observed across 5
different trials, where each trial uses a different video for training, is also shown.
Since our autonomous models have two kinds of synapses with different time constants, we observe two different kinds
of EV spectra that contribute concurrently to the dynamical state of the auto network. While the overall reservoir dynamics
are dictated by the effective contribution of both connections, Jslow (due to larger time constant of synaptic decay) will
have a slightly more dominating effect. Fig. 9 (a) illustrates the EV spectra corresponding to the bottom scan autonomous
800N-reservoir model simulated earlier for the 2-class problem (BenchPress/GolfSwing) with just 1 training video (refer Fig.
6). Now, the EV of J f ast have considerable number of modes with Re(EV) > 1 that is characteristic of chaotic dynamics. Hence,
as discussed above, we need to compensate for the overwhelming chaotic activity with modes that exhibit fixed points to obtain
a reliable memory model. This can be attained by adjusting the Jslow connections. From Fig. 9 (a), we see that as the range
of B (random constant used to derive the Jslow connections from the driven model) increases, the EV spectral radius of the
Jslow connections grows. For B = [0 − 0.01], the EVs are concentrated at the center. Such networks have almost no learning
capability as the dynamics converge to a single point that produces the same time-independent pattern of activity for any input
stimulus. As the spectral circle enlarges, the number of modes with Re(EV) > 1 also increases. For B = [0 − 0.15], the number
of fixed point modes reaches a reasonable number that in coherence with the chaotic activity due to J f ast results in generalized
10/17
reservoir learning for variable inputs. Increasing B further expands the spectral radius for Jslow (with more Re(EV) >1) that
adds onto the chaotic activity generated due to J f ast . In such scenarios, the reservoir will not be able to discern between inputs
that look similar due to persistent chaos. This is akin to the observations about the need for the reservoir to strike a balance
between pattern approximation and separability18 .
Fig. 9 (b) demonstrates the accuracy variation for the same 2-class 1-training example learning model with different
values of B. We repeated the experiment 5 times wherein we chose a different training video per class in each case. As
expected, the accuracy for a fixed state reservoir with B = [0 − 0.01] is very low. As the number of modes increase, the accuracy
improves justifying the relevance of stable and chaotic activity for better learning. As the chaotic activity begins to dominate
with increasing B, the accuracy starts declining. We also observe that the overall accuracy trend (for the 2-class 1-training
example model) represented by the min-max accuracy bar is consistent even when the training videos are varied. This indicates
that the reservoir accuracy is sensistive to the B-initialization and is not biased towards any particular training video in our
limited training example learning scheme. While the empirical result corroborates our theoretical analysis and justifies the
effectiveness of using the D/A approach in reservoir model construction, further investigation needs to be done to gauge the
other advantages/disadvantages of the proposed approach. Please refer to18, 19 for more clarification on random matrices and
EV spectra implication on dynamical reservoir models.
Comparison with State-of-the-Art Recognition Models
Deep Learning Networks (DLNs) are the current state-of-the-art learning models that have made major advances in several
recognition tasks2, 20 . While these large-scale networks are very powerful, they inevitably require large training data and
enormous computational resources. Spiking networks offer an alternative solution by exploiting event-based data-driven
computations that makes them attractive for deployment on real-time neuromorphic hardware where power consumption
and speed become vital constraints21, 22 . While research efforts in spiking models have soared in the recent past23–26 , the
performance of such models are not as accurate as compared to their artificial counterparts (or DLNs). From our perspective,
this work provides a new standard establishing the effectiveness of spiking models and their inherent timing-based processing
for action recognition in videos.
Fig. 10 compares our reservoir model to state-of-the-art recognition results27–29 . We simulated a VGG-1630 DLN
architecture that only processes the spatial aspects of the video input. That is, we ignore the temporal features of the video data
and attempt to classify each clip by looking at a single frame. We observe that the spatial VGG-16 model yields reduced accuracy
of 66.2% (42.9%) when presented with the full training dataset (8-training examples per class), implying the detrimental effect
of disregarding temporal statistics in the data. Please refer to the supplementary information (see Supplementary Fig. S3)
that details out the efficiency vs. accuracy tradeoff between our proposed reservoir and the spatial VGG-16 model. The static
hierarchical/feedforward nature of DLNs serves as a major drawback for sequential (such as video) data processing. While
deep networks extract good flexible representations from static images, addition of temporal dimension in the input necessitates
the model to incorporate recurrent connections that can capture the underlying temporal correlation from the data.
Recent efforts have integrated Recurrent Neural Networks (RNNs) and Long Short-Term Memory (LSTMs) with DLNs to
process temporal information28, 31 . In fact, the state-of-the-art accuracy on UCF-101 is reported as 92.5%29 as illustrated in
Fig. 10, where a two-stream convolutional architecture with two VGG-16 models, one for spatial and another for temporal
information processing is used. While the spatial VGG model is trained on the real-valued RGB data, the temporal VGG model
uses optical flow stacking to capture the correlated motion. Here, both the models are trained on the entire UCF-101 dataset
that significantly increases the overall training complexity. All models noted in Fig. 10 (including the LSTM model) use the
temporal information in the input data by computing the flow features and subsequently training the deep learning/ recurrent
models on such inputs. The overall accuracy shown in Fig. 10 is the combined prediction obtained from the spatial RGB and
temporal flow models. The incorporation of temporal information drastically improves a DLN’s accuracy.
Fig. 10 also quantifies the benefits with our reservoir approach as compared to the state-of-the-art models from a
computational efficiency perspective. We quantify efficiency in terms of total number of resources utilized during training a
particular model. The overall resource utilization is defined as the product of the Number of (tunable) Parameters × Number of
training Data Examples. The number of data examples are calculated as Number of Frames per video × Number of Training
videos (≡ 180 × 9537 for full training data). Referring to the results for spatial VGG-16 model, we observe that the reservoir
(10K-10% connectivity) is 2.4× more efficient, while yielding significantly higher accuracy, in the iso-data scenario where both
the models are shown 8 video samples per class. In contrast, the efficiency improves extensively to 28.5× in the iso-accuracy
scenario wherein the reservoir still uses 8 video samples per class (808 total videos), while the VGG-16 model requires the entire
training dataset (9537 total videos) during training. Comparing with the models that include temporal information, our reservoir
(10K-10% connectivity), although with lower accuracy, continues to yield significantly higher benefits of approximately 54×
and 7× as compared to Two-stream architecture and Temporal Convolutional network, respectively, on account of being trained
on limited data. However, in limited training example scenario, the reservoir (81.3% accuracy) continues to outperform an
11/17
LSTM based machine learning model (67% accuracy) while requiring lesser training data. It is worth mentioning that our
reservoir due to the inherent temporal processing capability requires only 30 epochs for training. On the other hand, the spatial
VGG-16 model needs 250 epochs. As the number of epochs required for training increase, the overall training cost (can be
defined as, Number of Epochs × Resource Utilization) would consequently increase. This further ascertains the potential of
reservoir computing for temporal data processing.
Model
# Training Accuracy
#
Videos
Parameters
# Data
Examples
Resource
Utilization (B)
Model
Efficiency
Spatial VGG-16 Model
Full
Training
66.2 %
97.5M
1716660
167374.4
0.035
Temporal
Convolutional Net [27]
Full
Training
83.7 %
23.5M
1716660
40341.5
0.146
Composite LSTM
Model [28]
Full
Training
84.3 %
-
1716660
-
-
Two-Stream
Convolutional Net [29]
Full
Training
92.5 %
185M
1716660
317582.1
0.019
Spatial VGG-16 Model
8 videos
per class
42.9 %
97.5M
145440
14180.4
0.414
Composite LSTM
Model [28]
10 videos
per class
67 %
-
181800
-
-
Ours
8K-10%Connectivity
8 videos
per class
80.2%
32.3M
145440
4697.7
1.251
Ours
10K-10%Connectivity
8 videos
per class
81.3%
40.4M
145440
5875.8
1
Figure 10. Accuracy and Efficiency comparison with state-of-the-art recognition models. Resource Utilization (B) is the
product of Number of Parameters × Number of Data Examples, where B denotes billion. The model efficiency is calculated by
normalizing the net resource utilization of each model against our reservoir model (10K-10% connectivity). Full Training (in
column corresponding to # Training Videos) denotes the entire training dataset that comprises of 9537 total videos across 101
classes.
Discussion
In this paper, we demonstrated the effectiveness of a spiking-based reservoir computing architecture for learning from limited
video examples on UCF101 action recognition task. We adopted the Driven/Autonomous reservoir construction approach
originally proposed for neuroscientific use and gave it a new dimension to perform multi input-output transformation for
practical recognition tasks. In fact, the D/A approach opens up novel possibilities for multi-modal recognition. For instance,
consider a 5-class face recognition scenario. Here, the driven network is trained to produce 5 different desired activity. The auto
models derived from the driven network will process the facial data to perform classification. Now, the same driven model can
also be used to derive autonomous models that will work on, say voice data. Combining the output results across an ensemble
of autonomous models will boost the overall confidence of classification. Apart from performance, the D/A model also has
several hardware oriented advantages. Besides sparse power-efficient spike communication, the reservoir internal topology
across all autonomous models for a particular problem remains equivalent. This provides a huge advantage of reusability of
network structures while scaling from one task to another.
Our analysis using eigenvalue spectra results on the reservoir provide a key insight about D/A based reservoir construction
such that the complex dynamics can converge to fixed memory states. We believe that this internal stabilization enables it to
generalize its memory from a few action signatures. We also compared our reservoir framework with state-of-the-art models
and reported the cost/accuracy benefits gained from our model. We would like to note that our work delivers a new benchmark
for action recognition from limited training videos in spiking scenario. We believe that we can further improve the accuracy
with our proposed approach by employing better input encoding techniques. The scan-based filtering approach discussed in this
paper does not account for depth variations and is susceptible to extreme spatial/rotational variation. In fact, we observe that
our reservoir model yields higher accuracy for classes with more consistent action signatures and finds it difficult to recognize
classes with more variations in movement (refer to Supplementary Fig. S1 for action specific accuracies). Using an adequate
spike processing technique (encompassing depth based filtering, motion based tracking or a fusion of these) will be key to
achieving improved accuracy. Another intriguing possibility is to combine our spike based model that can rapidly recognize
simpler and repeatable actions (e.g., GolfSwing) with minimal samples for training with more other DLN models for those
actions that have several variations (e.g., Knitting) or the actions that appear different dependent on the view point or both. We
12/17
are also studying the extension of our SNN model to a hierarchical model that could possibly capture more complex and subtle
aspects of the spatio-temporal signatures in these actions.
Finally, our reservoir model, although biologically inspired, only has excitatory components. The brain’s recurrent circuitry
consists of both excitatory and inhibitory neurons that operate in a balanced regime giving it the staggering ability to make
sense of a complex and ever-changing world in the most energy-efficient manner32 . In the future, we will concentrate on
incorporating the inhibitory scheme that will further augment the sparseness in reservoir activity thereby improving the overall
performance. An interesting line of research in this regard can be found here7, 33 . Although there have been multiple efforts in
the spiking domain exploring hierarchical feedforward and reservoir architectures with biologically plausible notions, most of
them have been focused on static images (and some voice recognition) that use the entire training dataset to yield reasonable
results34, 35 . To that effect, this work provides a new perspective on the power of temporal processing for rapid recognition
in cognitive applications using spiking networks. In fact, we are currently exploring the use of this model for learning at the
edge (such as cell phones and sensors where energy is of paramount importance) using a recently developed neural chip called
the Loihi36 that enables energy efficient on-chip spike timing based learning. We believe that results from this work could be
leveraged to enable Loihi based applications such as video analytics, video annotation and rapid searches for actions in video
databases.
References
1. LeCun, Y., Bengio, Y. & Hinton, G. Deep learning. Nature 521, 436–444 (2015).
2. Mnih, V. et al. Playing atari with deep reinforcement learning. arXiv preprint arXiv:1312.5602 (2013).
3. Silver, D. et al. Mastering the game of go with deep neural networks and tree search. Nature 529, 484–489 (2016).
4. Lake, B., Salakhutdinov, R., Gross, J. & Tenenbaum, J. One shot learning of simple visual concepts. In Proceedings of the
Cognitive Science Society, vol. 33 (2011).
5. Maass, W. Liquid state machines: motivation, theory, and applications. Computability in context: computation and logic in
the real world 275–296 (2010).
6. Lukoševičius, M. & Jaeger, H. Reservoir computing approaches to recurrent neural network training. Computer Science
Review 3, 127–149 (2009).
7. Srinivasa, N. & Cho, Y. Unsupervised discrimination of patterns in spiking neural networks with excitatory and inhibitory
synaptic plasticity. Frontiers in computational neuroscience 8, 159 (2014).
8. Wehr, M. & Zador, A. M. Balanced inhibition underlies tuning and sharpens spike timing in auditory cortex. Nature 426,
442–446 (2003).
9. Abbott, L., DePasquale, B. & Memmesheimer, R.-M. Building functional networks of spiking model neurons. Nature
neuroscience 19, 350 (2016).
10. Masquelier, T., Portelli, G. & Kornprobst, P. Microsaccades enable efficient synchrony-based coding in the retina: a
simulation study. Scientific reports 6, 24086 (2016).
11. Haykin, S. S. Adaptive filter theory (Pearson Education India, 2008).
12. Sussillo, D. & Abbott, L. F. Generating coherent patterns of activity from chaotic neural networks. Neuron 63, 544–557
(2009).
13. Ponulak, F. & Kasiński, A. Supervised learning in spiking neural networks with resume: sequence learning, classification,
and spike shifting. Neural Computation 22, 467–510 (2010).
14. Eliasmith, C. A unified approach to building and controlling spiking attractor networks. Neural computation 17, 1276–1314
(2005).
15. Soomro, K., Zamir, A. R. & Shah, M. Ucf101: A dataset of 101 human actions classes from videos in the wild. arXiv
preprint arXiv:1212.0402 (2012).
16. Hu, Y., Liu, H., Pfeiffer, M. & Delbruck, T. Dvs benchmark datasets for object tracking, action recognition, and object
recognition. Frontiers in Neuroscience 10, 405 (2016). URL http://journal.frontiersin.org/article/
10.3389/fnins.2016.00405.
17. Li, H., Liu, H., Ji, X., Li, G. & Shi, L. Cifar10-dvs: An event-stream dataset for object classification. Frontiers in
Neuroscience 11 (2017).
18. Rajan, K. Spontaneous and stimulus-driven network dynamics (Columbia University, 2009).
13/17
19. Rajan, K. & Abbott, L. Eigenvalue spectra of random matrices for neural networks. Physical review letters 97, 188104
(2006).
20. Cox, D. D. & Dean, T. Neural networks and neuroscience-inspired computer vision. Current Biology 24, R921–R929
(2014).
21. Han, S., Pool, J., Tran, J. & Dally, W. Learning both weights and connections for efficient neural network. In Advances in
Neural Information Processing Systems, 1135–1143 (2015).
22. Merolla, P. A. et al. A million spiking-neuron integrated circuit with a scalable communication network and interface.
Science 345, 668–673 (2014).
23. Maass, W. Energy-efficient neural network chips approach human recognition capabilities. Proceedings of the National
Academy of Sciences 113, 11387–11389 (2016).
24. Masquelier, T. & Thorpe, S. J. Unsupervised learning of visual features through spike timing dependent plasticity. PLoS
computational biology 3, e31 (2007).
25. Srinivasa, N. & Cho, Y. Self-organizing spiking neural model for learning fault-tolerant spatio-motor transformations.
IEEE transactions on neural networks and learning systems 23, 1526–1538 (2012).
26. Panda, P. & Roy, K. Unsupervised regenerative learning of hierarchical features in spiking deep networks for object
recognition. In Neural Networks (IJCNN), 2016 International Joint Conference on, 299–306 (IEEE, 2016).
27. Simonyan, K. & Zisserman, A. Two-stream convolutional networks for action recognition in videos. In Advances in neural
information processing systems, 568–576 (2014).
28. Srivastava, N., Mansimov, E. & Salakhudinov, R. Unsupervised learning of video representations using lstms. In
International Conference on Machine Learning, 843–852 (2015).
29. Feichtenhofer, C., Pinz, A. & Zisserman, A. Convolutional two-stream network fusion for video action recognition. In
Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 1933–1941 (2016).
30. Simonyan, K. & Zisserman, A. Very deep convolutional networks for large-scale image recognition. arXiv preprint
arXiv:1409.1556 (2014).
31. Donahue, J. et al. Long-term recurrent convolutional networks for visual recognition and description. In Proceedings of
the IEEE conference on computer vision and pattern recognition, 2625–2634 (2015).
32. Herculano-Houzel, S. The remarkable, yet not extraordinary, human brain as a scaled-up primate brain and its associated
cost. Proceedings of the National Academy of Sciences 109, 10661–10668 (2012).
33. Brendel, W., Bourdoukan, R., Vertechi, P., Machens, C. K. & Denéve, S. Learning to represent signals spike by spike.
arXiv preprint arXiv:1703.03777 (2017).
34. Kheradpisheh, S. R., Ganjtabesh, M. & Masquelier, T. Bio-inspired unsupervised learning of visual features leads to robust
invariant object recognition. Neurocomputing 205, 382–392 (2016).
35. Diehl, P. U. et al. Fast-classifying, high-accuracy spiking deep networks through weight and threshold balancing. In Neural
Networks (IJCNN), 2015 International Joint Conference on, 1–8 (IEEE, 2015).
36. Morris, J. http://www.zdnet.com/article/why-intel-built-a-neuromorphic-chip/ (2017).
Acknowledgements
We would like to acknowledge the support of Priyadarshini Panda through a summer internship in the Microarchitecture
Research Labs at Intel Corporation. We would also like to thank Michael Mayberry and Daniel Hammerstrom for their helpful
comments and suggestions.
Supplementary Information
Fig. S1 shows the class-wise accuracy obtained with our reservoir approach. We can deduce that certain classes are learnt
more accurately than the others. This is generally seen for those classes that have more consistent action signatures and
lesser variation such as GolfSwing, PommelHorse among others. On the other hand, classes that have more variations in
either action or views or both, such as TennisSwing, CliffDiving etc yield less accuracy. In fact, classes that consist of many
moving subjects such as MilitaryParade, HorseRacing exhibit lower accuracy. In such cases, the input processing technique
may not be able to extract motions from all relevant moving subjects. Here, irrelevant clutter or jitter due to camera motion
14/17
gets captured in the spiking information as a result of which the reservoir is unable to recognize a particular signature. Such
limitations can be eliminated with an adequate processing technique that incorporates depth and (or) motion based tracking.
Also, data augmentation with simple transformations can further improve the accuracy. For instance, for a given training video,
incorporating both left/right handed views by applying a 180 deg rotational transformation to all the frames of the original video
will yield a different view/action for the same video. Such data augmentation will account for more variations in views/angles
enabling the reservoir to learn better.
Fig. S2 shows the confusion matrix of the 101-class problem for the same reservoir topology as Fig. S1.
1: Babycrawling 65%
2: BalanceBeam 89%
3: BenchPress 87.5%
4: CliffDiving 58.2%
5: Diving 85.4%
6: GolfSwing 97%
7: HammerThrow 94.2%
8: HandStandPushups 89.8 %
9: LongJump 94.2%
10: BodyWeightSquats 58.3%
11: JumpingJack 94.8%
12: JumpRope 65%
13: CleanandJerk 94.5%
14: PlayingCello 88%
15: PlayingGuitar 63.4%
16: PlayingPiano 95%
17: PlayingFlute 79.6%
18: PlayingDhol 61.8%
19: ParallelBars 89%
20: PlayingViolin 91.1%
21: PommelHorse 97%
22: Pullups 80%
23: Pushups 78%
24:UnevenBars 89.6%
25:TennisSwing 59.4%
26: ThrowDiscus 91.6%
27: Typing 95.8%
28: WallPushUps 94.4%
29: HandStandWalking 80.3%
30: Bowling 91.2%
31: Punch 56%
32: Billiards 89%
33: HorseRace 56.1%
34: TrampolineJumping 62%
35: PlayingTabla 78%
36: BasketBallDunk 63%
37: IceDancing 77%
38: Skijet 86.3%
39: VolleyballSpiking 62%
40: SkyDiving 66.4%
41: BoxingPunchingBag 88%
42: HorseRiding 93.7%
43: SalsaSpin 48%
44: WritingonBoard 79.2%
45: PlayingSitar 92.3%
46: Knitting 61.1%
47: BoxingSpeedBag 82%
48: Swing 79%
49: StillRings 80.8%
50: SkateBoarding 62.5%
51: Biking 83.4%
52: FloorGymnastics 94.5%
53: PoleVault 94.6%
54: ShotPut 91%
55: SoccerPenalty 96.5%
56: BlowDryHair 61%
57: Surfing 94.3%
58: Fencing 70.1%
59: Skiing 79.8%
60: Rowing 78.4%
61: TableTennisShot 77.3%
62: FrontCrawl 94.3%
63: SumoWrestling 67.3%
64: YoYo 64.1%
65: BandMarching 79.4%
66: FrisbeeCatch 68.3%
67: RockClimbingIndoor 78.4%
68: Mixing 79.5%
69: JugglingBalls 81.4%
70: MilitaryParade 58.3%
71: Kayaking 91.3%
72: FieldHockeyPenalty 93%
73: PlayingDaf 95%
74: ApplyEyeMakeup 77.2%
75: SoccerJuggling 57.2%
76: CricketBowling 53.9%
77: HulaHoop 92.1%
78: Rafting 58%
79: CricketShot 74.1%
80: BreastStroke 78.9%
81: Drumming 79.2%
82: HeadMassage 55.8%
83: BrushingTeeth 71.1%
84: BaseballPitch 79.3%
85: BlowingCandles 77%
86: TaiChi 94.7%
87: WalkingWithDog 73.7%
88: ApplyLipstick 78.3%
89: CuttingInKitchen 92.9%
90: RopeClimbing 93.4%
91: ShavingBeard 72.5%
92: PizzaTossing 89.4%
93: Hammering 84%
94: Archery 91.3%
95: BasketBall 96.2%
96: Nunchuks 88.6%
97: MoppingFloor 86.1%
98: Lunges 94%
99: JavelinThrow 93.2%
100: HighJump 93%
101: HairCut 73.4%
Figure S1. Class-wise accuracy obtained with 8000 neuron (10% connectivity) model on UCF-101 dataset when trained with 8
training videos per class. Few select classes with high/low accuracy are highlighted with green/red font to quantify the classes
with more/less consistent videos.
Here, we further quantify the advantages with our reservoir approach against the spatial VGG-16 model described in the
main manuscript by gauging the efficiency vs. accuracy tradeoff obtained from both the models in equivalent training scenarios.
Fig. S3 (a) illustrates the accuracy obtained for different training scenarios. For scenario 1, wherein the entire training dataset
is used, the Top-1 accuracy obtained is 66.2%, thus implying the detrimental effect of disregarding temporal statistics in the
data. Then, we tested the VGG model for two different limited example scenarios, one equivalent to our reservoir learning with
8 training examples per class and other with 40 training examples. In both cases, we see that the accuracy drops drastically
with the 8-training example scenario yielding merely 42.9% accuracy. Increasing the number of training examples to 40 does
improve the accuracy to 50.02%, but the performance is much lower than that of the reservoir model.
A noteworthy observation in Fig. S3 (a) is that the Top-5 accuracy for the VGG-16 model (89.8%) trained with the entire
dataset and the 8000N-10% connectivity reservoir (86.1%) trained on 8 videos per class (refer to Fig. 8 (a) in main manuscript)
is comparable. In contrast, for limited training scenarios, the Top-5 accuracy with VGG is drastically reduced as compared
to the reservoir models shown in Fig. 8 (a) of the main manuscript. This is indicative of the DLN’s inferior generalization
capability on temporal data.
Fig. S3 (b) shows the normalized benefits observed with the different topologies of reservoir (Fig. 8 (a) in main manuscript)
as compared to spatial VGG-16. Here, we quantify efficiency in terms of total number of trainable parameters (or weights) in
a given model. The VGG-16 model for 101-class has 97.5M tunable parameters, against which all other models in Fig. S3
(b) are normalized. We observe ∼ 6× improvement in efficiency with the reservoir model where both VGG and the reservoir
15/17
0.9
0.8
Target Output Class
0.7
0.6
0.5
0.4
0.3
0.2
0.1
Predicted Output Class
Figure S2. Confusion Matrix for 101-class problem corresponding to Fig. S1 above. The X/Y axis represent the class numbers
with equivalent notations as that of Fig. S1. The colormap indicates the accuracy for a given class label with highest intensity
(red) corresponding to 100% match between target & output class and lowest intensity (blue) corresponding to 0% match
between target & output class.
(4000-N reservoir with 20% connectivity) have almost similar accuracy ∼ 67%. A noteworthy observation here is that VGG-16
in this case was trained with the entire UCF101 training data, while our reservoir had only 8 samples per class for training. This
ascertains the potential of reservoir computing for temporal data processing. In an iso-data scenario (i.e. 8 videos per class for
both VGG and reservoir models), we observe that the reservoirs yielding maximal accuracy of 81.3%(80.2%) continue to be
more efficient by 2.2 × (3×) than the VGG model.
# Training Examples
Top-1
(%)
Top-3
(%)
Top-5
(%)
Full Training Data
66.2
81.6
89.8
8 videos per class
42.9
49.7
53.4
40 videos per class
50.02
56.3
65.2
4K
10%C
10K
8K 8%C
10%C
6K
10%C
Accuracy
4K
20%C
VGG
66.2%
VGG
68.5%
VGG-16 (# Tunable Parameters 97.5M )
C: Connectivity
42.9%
(b)
77.1%
80.2%
81.3%
70.2%
(a)
Efficiency comparison for near iso-accuracy
Efficiency comparison for iso-training videos
Figure S3. (a) Top-1/3/5 accuracy obtained with the VGG-16 deep learning model on 101-classes for different training
scenarios (b) Efficiency (quantified as total # of tunable parameters) comparison between VGG-16 model and D/A reservoir
model of different topologies (from Fig. 8 (a) in main manuscript) for iso-accuracy and iso-data (same number of training
examples) scenario. The accuracy/topology for each model is noted on the graph.
For implementing the VGG-16 convolutional network model, we used the machine learning tool, Torch1 . Also, we used
standard regularization Dropout2 technique to optimize the training process for yielding maximum performance. In case of the
DLN model, we feed the raw pixel valued RGB data as inputs to the DLN for frame-by-frame training thereby providing the
model with all information about the data. The input video frames originally of 200 × 300 dimensions are resized to 224 × 224.
16/17
The macro-architecture of the VGG-16 model was imported from3 and then altered for 101-class classification. To ensure
sufficient number of training examples for the deep VGG-16 model in the limited example learning cases, we augmented the
data via two random transformation schemes: rotation in the range of (0 − 40 deg), width/ height shifts within 20% of total
width/height. For the 8/40-training example scenario, each frame of a training video was augmented 10/2 times, respectively
using any of the above transformations. The VGG-16 model in each of the training example scenarios described above is
trained for 250 epochs. Here, in each epoch, the entire limited/full training dataset (that is the individual frames corresponding
to all classes) are presented to the VGG model. Note, UCF101 provides three train/test split recommendations. We use just
split #1 for all of our experiments.
References
1. Collobert, R., Kavukcuoglu, K. & Farabet, C. Torch7: A matlab-like environment for machine learning. In BigLearn, NIPS
Workshop, EPFL-CONF-192376 (2011).
2. Srivastava, N., Hinton, G. E., Krizhevsky, A., Sutskever, I. & Salakhutdinov, R. Dropout: a simple way to prevent neural
networks from overfitting. Journal of machine learning research 15, 1929–1958 (2014).
3. Zagoruyko, S. http://torch.ch/blog/2015/07/30/cifar.html (2015).
17/17
| 9cs.NE
|
ON THE CONSTRUCTION OF CONVERGING HIERARCHIES
FOR POLYNOMIAL OPTIMIZATION
BASED ON CERTIFICATES OF GLOBAL POSITIVITY∗
arXiv:1709.09307v1 [math.OC] 27 Sep 2017
AMIR ALI AHMADI† AND GEORGINA HALL‡
Abstract. In recent years, techniques based on convex optimization and real algebra that
produce converging hierarchies of lower bounds for polynomial optimization problems (POPs) have
gained much popularity. At their heart, these hierarchies rely crucially on Positivstellensätze from
the late 20th century (e.g., due to Stengle, Putinar, or Schmüdgen) that certify positivity of a
polynomial on an arbitrary basic semialgebraic set. In this paper, we show that such hierarchies
could in fact be designed from much more limited Positivstellensätze dating back to the early 20th
century that only certify positivity of a polynomial globally. More precisely, we show that any inner
approximation to the cone of positive homogeneous polynomials that is arbitrarily tight can be turned
into a converging hierarchy for general polynomial optimization problems with compact feasible sets.
This in particular leads to a semidefinite programming-based hierarchy that relies solely on Artin’s
solution to Hilbert’s 17th problem. We also use a classical result of Polyá on global positivity of
even forms to construct an “optimization-free” converging hierarchy for general POPs with compact
feasible sets. This hierarchy only requires polynomial multiplication and checking nonnegativity of
coefficients of certain fixed polynomials. As a corollary, we obtain new linear programming and
second-order cone programming-based hierarchies for POPs that rely on the recently introduced
concepts of dsos (diagonally dominant sum of squares) and sdsos (scaled diagonally dominant sum
of squares) polynomials.
Key words. Positivstellensätze, polynomial optimization, convex optimization
AMS subject classifications. 14Q99, 90C05, 90C26, 90C22, 90-03
1. Introduction. A polynomial optimization problem (POP) is an optimization
problem of the form
inf
(1)
x∈Rn
s.t.
p(x)
gi (x) ≥ 0, i = 1, . . . , m,
where p, gi , i = 1, . . . , m, are polynomial functions in n variables x := (x1 , . . . , xn )
and with real coefficients. It is well-known that polynomial optimization is a hard
problem to solve in general. For example, simply testing whether the optimal value
of problem (1) is smaller than or equal to some rational number k is NP-hard already
when the objective is quadratic and the constraints are linear [14]. Nevertheless, these
problems remain topical due to their numerous applications throughout engineering,
operations research, and applied mathematics (see, e.g., [10, 6, 2]). In this paper, we
are interested in obtaining lower bounds on the optimal value of problem (1). We focus
on a class of methods which construct hierarchies of tractable convex optimization
problems whose optimal values are lowerbounds on the optimal value of (1), with
convergence to it as the sequence progresses. This implies that even though the
original POP is nonconvex, one can obtain increasingly accurate lower bounds on its
optimal value by solving convex optimization problems. One method for constructing
∗ This
work is partially funded by the DARPA Young Faculty Award, the Young Investigator
Award of the AFOSR, the CAREER Award of the NSF, the Google Faculty Award, and the Sloan
Fellowship.
† Department of Operations Research and Financial Engineering, Princeton, NJ 08540
([email protected], http://aaa.princeton.edu).
‡ Department of Operations Research and Financial Engineering, Princeton, NJ 08540
([email protected], http://scholar.princeton.edu/ghall).
1
2
these hierarchies of optimization problems that has gained attention in recent years
relies on the use of Positivstellensätze (see, e.g., [11] for a survey). Positivstellensätze
are algebraic identities that certify infeasibility of a set of polynomial inequalities, or
equivalently1 , positivity of a polynomial on a basic semialgebraic set. (Recall that
a basic semialgebraic set is a set defined by finitely many polynomial inequalities.)
These Positivstellensätze can be used to prove lowerbounds on POPs. Indeed, if we
denote the feasible set of (1) by S, the optimal value of problem (1) is equivalent to
sup
γ
γ
(2)
s.t.
p(x) − γ ≥ 0, ∀x ∈ S.
Hence if γ is a strict lower bound on (1), we have that p(x) − γ > 0 on S, a fact
that can be certified using Positivstellensätze. At a conceptual level, hierarchies that
provide lower bounds on (1) are constructed thus: we fix the “size of the certificate” at
each level of the hierarchy and search for the largest γ such that the Positivstellensätze
at hand can certify positivity of p(x) − γ over S with a certificate of this size. As the
sequence progresses, we increase the size of the certificates allowed, hence obtaining
increasingly accurate lower bounds on (1).
Below, we present three of the better-known Positivstellensätze, given respectively
by Stengle [25], Schmüdgen [21], and Putinar [19]. These all rely on sum of squares
certificates. We recall that a polynomial is a sum of squares (sos) if it can be written
as a sum of squares of other polynomials. We start with Stengle’s Positivstellensatz,
which certifies infeasibility of a set of polynomial inequalities. It is sometimes referred
to as “the Positivstellensatz” in related literature as it requires no assumptions, contrarily to Schmüdgen and Putinar’s theorems which can be viewed as refinements of
Stengle’s result under additional assumptions.
Theorem 1.1 (Stengle’s Positivstellensatz [25]). The basic semialgebraic set
S = {x ∈ Rn | g1 (x) ≥ 0, . . . , gm (x) ≥ 0}
is empty if and only if there exist sum of squares polynomials s0 (x),s1 (x),. . ., sm (x),
s12 (x), s13 (x),. . ., s123...m (x) such that
X
X
sij (x)gi (x)gj (x) + . . . + s123...m (x)g1 (x) . . . gm (x).
si (x)gi (x) +
−1 = s0 (x) +
i
{i,j}
The next two theorems, due to Schmüdgen and Putinar, certify positivity of a polynomial p over a basic semialgebraic set S. They impose additional compactness assumptions comparatively to Stengle’s Positivstellensatz.
Theorem 1.2 (Schmüdgen’s Positivstellensatz [21]). Assume that the set
S = {x ∈ Rn | g1 (x) ≥ 0, . . . , gm (x) ≥ 0}
is compact. If a polynomial p is positive on S, then
X
X
sij (x)gi (x)gj (x) + . . . + s123...m (x)g1 (x) . . . gm (x),
p(x) = s0 (x) +
si (x)gi (x) +
i
{i,j}
where s0 (x),s1 (x),. . ., sm (x), s12 (x), s13 (x),. . ., s123...m (x) are sums of squares.
that the set {x ∈ Rn | g1 (x) ≥ 0, . . . , gm (x) ≥ 0} is empty if and only if −g1 (x) > 0 on
the set {x ∈ Rn | g2 (x) ≥ 0, . . . , gm (x) ≥ 0}.
1 Note
CONVERGING HIERARCHIES FOR POLYNOMIAL OPTIMIZATION
3
Theorem 1.3 (Putinar’s Positivstellensatz [19]). Let
S = {x ∈ Rn | g1 (x) ≥ 0, . . . , gm (x) ≥ 0}
and assume that {g1 , . . . , gm } satisfy the Archimedean property, i.e., there exists
N ∈ N such that
X
x2i = σ0 (x) + σ1 (x)g1 (x) + . . . + σm (x)gm (x),
N−
i
where σ1 (x), . . . , σm (x) are sums of squares. If a polynomial p is positive on S, then
p(x) = s0 (x) + s1 (x)g1 (x) + . . . + sm (x)gm (x),
where s1 (x), . . . , sm (x) are sums of squares.
Note that these three Positivstellensätze involve in their expressions sum of squares
polynomials of unspecified degree. To construct hierarchies of tractable optimization problems for (2), we fix this degree: at level r, we search for the largest γ
such that positivity of p(x) − γ over S can be certified using the Positivstellensätze
where the degrees of all sos polynomials are taken to be less than or equal to 2r.
Solving each level of these hierarchies is then a semidefinite program (SDP). This is
a consequence of the fact that one can optimize over (or test membership to) the
set of sum of squares polynomials of fixed degree using semidefinite programming
[16, 15, 9]. Indeed, a polynomial p of degree 2d and in n variables is a sum of squares
if and only if there exists a symmetric matrix Q 0 such that p(x) = z(x)T Qz(x),
where z(x) = (1, x1 , . . . , xn , . . . , xdn )T is the standard vector of monomials in n variables and of degree less than or equal to d. We remark that the hierarchy obtained
from Stengle’s Positivstellensatz was proposed and analyzed by Parrilo in [16]; the
hierarchy obtained from Putinar’s Positivstellensatz was proposed and analyzed by
Lasserre in [9]. There have been more recent works that provide constructive proofs
of Schmüdgen and Putinar’s Positivstellensätze; see [5, 22, 24]. These proofs rely on
other Positivstellensätze, e.g., a result by Polyá (see Theorem 1.6 below) in [22, 24],
and the same result by Polyá, Farkas’ lemma, and Stengle’s Positivstellensatz in [5].
There has further been an effort to derive complexity bounds for Schmüdgen and
Putinar’s Positivstellensätze in recent years; see [13, 23].
On a historical note, Stengle, Schmüdgen, and Putinar’s Positivstellensätze were
derived in the latter half of the 20th century. As mentioned previously, they all
certify positivity of a polynomial over an arbitrary basic semialgebraic set (modulo
compactness assumptions). By contrast, there are Positivstellensätze from the early
20th century that certify positivity of a polynomial globally. Perhaps the most wellknown Positivstellensatz of this type is due to Artin in 1927, in response to Hilbert’s
17th problem. Artin shows that any nonnegative polynomial is a sum of squares of
rational functions. Here is an equivalent formulation of this statement:
Theorem 1.4 (Artin [4]). For any nonnegative polynomial p, there exists an sos
polynomial q such that p · q is a sum of squares.
To the best of our knowledge, in this area, all converging hierarchies of lower bounds
for POPs are based off of Positivstellensätze that certify nonnegativity of a polynomial
over an arbitrary basic semialgebraic set. In this paper, we show that in fact, under
compactness assumptions, it suffices to have only global certificates of nonnegativity
(such as the one given by Artin) to produce a converging hierarchy for general POPs.
4
As a matter of fact, even weaker statements that apply only to globally positive (as
opposed to globally nonnegative) forms are enough to derive converging hierarchies
for POPs. Examples of such statements are due to Habicht [8] and Reznick [20]. With
such an additional positivity assumption, more can usually be said about the structure
of the polynomial q in Artin’s result. Below, we present the result by Reznick.
Theorem 1.5
P (Reznick [20]). For any positive definite form p, there exists r ∈ N
such that p(x) · ( i x2i )r is a sum of squares.
We show in this paper that this Positivstellensatz also gives rise to a converging
hierarchy for POPs with a compact feasible set similarly to the one generated by
Artin’s Positivstellensatz.
Through their connections to sums of squares, the two hierarchies obtained using
the theorems of Reznick and Artin are semidefinite programming-based. In this paper,
we also derive an “optimization-free” converging hierarchy for POPs with compact
feasible sets where each level of the hierarchy only requires that we be able to test
nonnegativity of the coefficients of a given fixed polynomial. To the best of our
knowledge, this is the first converging hierarchy of lower bounds for POPs which
does not require that convex optimization problems be solved at each of its levels.
To construct this hierarchy, we use a result of Polyá [17], which just like Artin’s and
Reznick’s Positivstellensätze, certifies global positivity of forms. However this result is
restricted to even forms. Recall that a form p is even if each of the variables featuring
in its individual monomials has an even power. This is equivalent (see [7, Lemma 2])
to p being invariant under change of sign of each of its coordinates, i.e.,
p(x1 , . . . , xn ) = p(−x1 , . . . , xn ) = · · · = p(x1 , . . . , −xn ).
Theorem 1.6 (Polyá
P [17]). For any positive definite even form p, there exists
r ∈ N such that p(x) · ( i x2i )r has nonnegative coefficients.2
Our aforementioned hierarchy enables us to obtain faster-converging linear programming (LP) and second-order cone programming (SOCP)-based hierarchies for general
POPs with compact feasible sets that rely on the concepts of dsos and sdsos polynomials. These are recently introduced inner approximations to the set of sos polynomials
that have shown much better scalability properties in practice [3].
As a final remark, we wish to stress the point that the goal of this paper is
first and foremost theoretical, i.e., to provide methods for constructing converging
hierarchies of lower bounds for POPs using as sole building blocks certificates of
global positivity. We do not make any claims that these hierarchies can outperform
the popular existing hierarchies due, e.g., to Lasserre [9] and Parrilo [16]. We do
believe however that the optimization-free hierarchy presented in Section 4.1 could
potentially be of interest in large-scale applications where the convex optimization
problems appearing in traditional hierarchies are too cumbersome to solve.
1.1. Outline of the paper. The paper is structured as follows. In Section 2, we
show that if one can inner approximate the cone of positive definite forms arbitrarily
well (with certain basic properties), then one can produce a converging hierarchy
of lower bounds for POPs with compact feasible sets (Theorem 2.4). This relies
2 A perhaps better-known but equivalent formulation of this theorem is the following: for any
P
form h that is positive on the standard simplex, there exists r ∈ N such that h(x) · ( i xi )r has
nonnegative coefficients. The two formulations are equivalent by simply letting p(x) = h(x2 ).
CONVERGING HIERARCHIES FOR POLYNOMIAL OPTIMIZATION
5
on a reduction (Theorem 2.1) that reduces the problem of certifying a strict lower
bound on a POP to that of proving positivity of a certain form. In Section 3, we
see how this result can be used to derive semidefinite programming-based converging
hierarchies (Theorems 3.2 and 3.4) from the Positivstellensätze by Artin (Theorem
1.4) and Reznick (Theorem 1.5). In Section 4, we derive an optimization-free hierarchy
(Theorem 4.1) from the Positivstellensatz of Polyá (Theorem 1.6) as well as LP and
SOCP-based hierarchies which rely on dsos/sdsos polynomials (Corollary 4.8). We
conclude with a few open problems in Section 5.
1.2. Notation and basic definitions. We use the standard notation A 0
to denote that a symmetric matrix A is positive semidefinite. Recall that a form is
a homogeneous polynomial, i.e., a polynomial whose monomials all have the same
degree. We denote the degree of a form f by deg(f ). We say that a form f is
nonnegative (or positive semidefinite) if f (x) ≥ 0, for all x ∈ Rn (we write f ≥ 0). A
form f is positive definite (pd) if f (x) > 0, for all nonzero x in Rn (we write f > 0).
Throughout the paper, we denote the set of forms (resp. the set of nonnegative
forms) in n variables and of degree d by Hn,d (resp Pn,d ). We denote the ball of
radius R and centered at the origin by B(0, R) and the unit sphere in x-space, i.e.,
{x ∈ Rn | ||x||2 = 1}, by Sx . We use the shorthand f (y 2 − z 2 ) for y, z ∈ Rn to
denote f (y12 − z12 , . . . , yn2 − zn2 ). We say that a scalar γ is a strict lower bound on (1)
if p(x) > γ, ∀x ∈ S. Finally, we ask the reader to carefully read Remark 2.3 which
contains the details of a notational overwriting occurring before Theorem 2.4 and
valid from then on throughout the paper. This overwriting makes the paper much
simpler to parse.
2. Constructing converging hierarchies for POP using global certificates of positivity. Consider the polynomial optimization problem in (1) and denote its optimal value by p∗ . Let d be such that 2d is the smallest even integer larger
than or equal to the maximum degree of p, gi , i = 1, . . . , m. We denote the feasible
set of our optimization problem by
S = {x ∈ Rn | gi (x) ≥ 0, i = 1, . . . , m}
and assume that S is contained within a ball of radius R. From this, it is easy to
provide (possibly very loose) upper bounds on gi (x) over the set S: as S is contained
in a ball of radius R, we have |xi | ≤ R, for all i = 1, . . . , n. We then use this to
upper bound each monomial in gi and consequently gi itself. We use the notation ηi
to denote these upper bounds, i.e., gi (x) ≤ ηi , for all i = 1, . . . , m and for all x ∈ S.
Similarly, we can provide an upperbound on −p(x). We denote such a bound by β,
i.e., −p(x) ≤ β, ∀x ∈ S.
The goal of this section is to produce a method for constructing converging hierarchies of lower bounds for POPs if we have access to arbitrarily accurate inner
approximations of the set of positive definite forms. The first theorem (Theorem 2.1)
connects lower bounds on (1) to positive definiteness of a related form. The second
theorem (Theorem 2.4) shows how this can be used to derive a hierarchy for POPs.
Theorem 2.1. Consider the general polynomial optimization problem in (1) and
recall that d is such that 2d is the smallest even integer larger than or equal to the
maximum degree of p, gi , i = 1, . . . , m. Suppose S ⊆ B(0, R) for some positive scalar
R. Let ηi , i = 1, . . . , m (resp. β) be any finite upper bounds on gi (x), i = 1, . . . , m
(resp. −p(x)).
Then, a scalar γ is a strict lower bound on (1) if and only if the homogeneous
6
sum of squares polynomial
(3)
fγ (x, s, y) := γy 2d − y 2d p(x/y) − s20 y 2d−2
+
(R +
m
X
i=1
2
+
m
X
i=1
y 2d gi (x/y) − s2i y 2d−2
m
n
X
X
s2i )d − s2d
x2i +
ηi + β + γ)d y 2d − (
m+1
i=1
i=0
2
!2
of degree 4d and in n + m + 3 variables (x1 , . . . , xn , s0 , . . . , sm , sm+1 , y) is positive
definite.
Proof. It is easy to see that γ is a strict lower bound on (1) if and only if the set
X
x2i ≤ R}
T := {x ∈ Rn | γ − p(x) ≥ 0; gi (x) ≥ 0, i = 1, . . . , m;
i
is empty. Indeed, if T is nonempty, then there exists a point x̌ ∈ S such that p(x̌) ≤ γ.
This implies that γ cannot be a strict lower bound on (1). Conversely, if T is empty,
the intersection of S with {x | γ − p(x) ≥ 0} is empty, which implies that ∀x ∈ S,
p(x) > γ.
We now define the set:
(4)
Ts = {(x, s) ∈ Rn+m+2 | γ − p(x) = s20 ; gi (x) = s2i , i = 1, . . . , m;
m
n
m
X
X
X
s2i )d − s2d
x2i +
ηi + β + γ)d − (
(R +
m+1 = 0}.
i=1
i=1
i=0
Note that Ts is empty if and only if T is empty. Indeed, if Ts is nonempty, then there
exists x̂ ∈ Rn and ŝ ∈ Rm+2 such that the three sets of equations are satisfied. This
obviously implies
P that γ − p(x̂) ≥ 0 and that gi (x̂) ≥ 0, for all i = 1, . . . , m. It further
implies that i x̂2i ≤ R as by assumption, if x̂ ∈ S, then x̂ is in a ball of radius R.
Conversely, suppose now that TPis nonempty. There exists x̃ such that γ − p(x̃) ≥ 0,
gi (x̃) ≥ 0 for i = 1, . . . , m, and i x̃i 2 ≤ R. Hence, there exist s˜0 , . . . , s˜m such that
γ − p(x̃) = s˜0 2 and gi (x̃) = s˜i 2 , i = 1, . . . , m.
P
Combining the fact that i x̃i 2 ≤ R and the fact that ηi , i = 1, . . . , m (resp. γ + β)
are upperbounds on gi (resp. γ − p(x̃)), we obtain:
R+
m
X
i=1
ηi + β + γ ≥
n
X
i=1
x̃i 2 +
m
X
s˜i 2 .
i=0
By raising both sides of the inequality to the power d, we show the existence of s̃m+1 .
We now show that Ts is empty if and only if fγ (x, s, y) is positive definite. Suppose
that Ts is nonempty, i.e., there exists (x̆, s̆) ∈ Rn+m+2 such that the equalities given
in (4) hold. Note then that fγ (x̆, s̆, 1) = 0. As (x̆, s̆, 1) is nonzero, this implies that
fγ (x̆, s̆, y̆) is not positive definite.
For the converse, assume that fγ (x, s, y) is not positive definite. As fγ (x, s, y) is
a sum of squares and hence nonnegative, this means that there exists nonzero (x̄, s̄, ȳ)
such that f (x̄, s̄, ȳ) = 0. We proceed in two cases. If ȳ 6= 0, it is easy to see that
(x̄/ȳ, s̄/ȳ) ∈ Ts and Ts is nonempty. Consider now the case where ȳ = 0. The third
square in fγ being equal to zero gives us:
m
X
X
2
s̄2i )d = s̄2d
x̄i +
−(
m+1 .
i
i=0
CONVERGING HIERARCHIES FOR POLYNOMIAL OPTIMIZATION
7
This implies that s̄m+1 = 0 and that x̄1 = . . . = x̄m = s¯0 = . . . = s̄m = 0 which
contradicts the fact that (x̄, s̄, ȳ) is nonzero.
Remark 2.2. Note that Theorem 2.1 implies that testing feasibility of a set of
polynomial inequalities is no harder than checking whether a homogeneous polynomial
that is sos has a zero. Indeed, as mentioned before, the basic semialgebraic set
{x | g1 (x) ≥ 0, . . . , gm (x) ≥ 0}
is empty if and only if γ = 0 is a strict lower bound on the POP
inf
− g1 (x)
s.t.
g2 (x) ≥ 0, . . . , gm (x) ≥ 0.
x
In principle, this reduction can open up new possibilities for algorithms for testing
feasibility of a basic semialgebraic set. For example, the work in [1] shows that positive
definiteness of a form f is equivalent to global asymptotic stability of the polynomial
vector field ẋ = −∇f (x). One could as a consequence search for Lyapunov functions,
as is done in [1, Example 2.1.], to certify positivity of forms. Conversely, simulating
trajectories of the above vector field can be used to minimize f and potentially find its
nontrivial zeros, which, by our reduction, can be turned into a point that belongs to
the basic semialgebraic set at hand.
We further remark that one can always take the degree of the sos form fγ in
(3) whose positivity is under consideration to be equal to four. This can be done by
changing the general POP in (1) to only have quadratic constraints and a quadratic
objective via an iterative introduction of new variables and new constraints in the
following fashion: xij = xi xj .
Remark 2.3 (Notational remark). As a consequence of Theorem 2.1, we now
know that certifying lower bounds on (1) is equivalent to proving positivity of the form
fγ that appears in (3). To simplify notation, we take this form to have n variables
and be of degree 2d from now on (except for our Positivstellensätze in Corollaries 3.5
and 4.5 which stand on their own). To connect back to problem (1) and the original
notation, the reader should replace every occurrence of n and d in the future as follows:
n ← n + m + 3,
d ← 2d.
Recall that n was previously the dimension of the decision variable of problem (1), d
was such that 2d is the smallest even integer larger than or equal to the maximum
degree of gi and p in (1), and m was the number of constraints of problem (1).
Our next theorem shows that, modulo some technical assumptions, if one can
inner approximate the set of positive definite forms arbitrarily well (conditions (a)
and (b)), then one can construct a converging hierarchy for POPs.
r
Theorem 2.4. Let Kn,2d
be a sequence of sets (indexed by r) of homogeneous
polynomials in n variables and of degree 2d with the following properties:
r
0
(a) Kn,2d
⊆ Pn,2d , ∀r, and there exists a pd form sn,2d ∈ Kn,2d
.
r
(b) If p > 0, then ∃r ∈ N such that p ∈ Kn,2d .
r+1
r
(c) Kn,2d
⊆ Kn,2d
, ∀r.
r
r
(d) If p ∈ Kn,2d
, then ∀ǫ ∈ [0, 1], p + ǫsn,d ∈ Kn,2d
.
Recall the definition of fγ (z) given in (3). Consider the hierarchy of optimization
8
problems indexed by r:
lr := sup
γ
γ
(5)
s.t.
1
r
fγ (z) − sn,2d (z) ∈ Kn,2d
.
r
Then, lr ≤ p∗ for all r, {lr } is nondecreasing, and limr→∞ lr = p∗ .
Proof. We first show that the sequence {lr } is upperbounded by p∗ . Suppose that
a scalar γ satisfies
1
r
.
fγ (z) − sn,2d (z) ∈ Kn,2d
r
We then have fγ (z)− r1 sn,2d (z) ∈ Pn,2d using (a). This implies that fγ (z) ≥ 1r sn,2d (z),
and hence fγ is pd as sn,2d is pd. From Theorem 2.1, it follows that γ has to be a
strict lower bound on (1). As any such γ satisfies γ < p∗ , we have that lr ≤ p∗ for all r.
We now show monotonicity of the sequence {lr }. Let γ be such that
1
r
.
fγ (z) − sn,2d (z) ∈ Kn,2d
r
We have the following identity:
fγ (z) −
1
1
1
sn,2d (z) = fγ (z) − sn,2d (z) +
sn,2d (z).
r+1
r
r(r + 1)
Now, using the assumption and properties (c) and (d), we conclude that
fγ (z) −
1
r+1
sn,2d (z) ∈ Kn,2d
.
r+1
r+1
1
r
This implies that {γ | fγ (z) − 1r sn,2d (z) ∈ Kn,2d
} ⊆ {γ | fγ (z) − r+1
sn,2d (z) ∈ Kn,2d
}
and that lr ≤ lr+1 .
Note that as the sequence {lr } is upperbounded and nondecreasing, it converges.
Let us show that the limit of this sequence is p∗ . To do this, we show that for any strict
lower bound γ on (1), there exists a positive integer r such that fγ (z) − 1r sn,2d (z) ∈
r
Kn,2d
. By Theorem 2.1, as γ is a strict lower bound, fγ (z) is positive definite. Hence,
by continuity, there exists a positive integer r′ such that fγ (z) − r1′ sn,2d (z) is positive
definite. Using (b), this implies that there exists a positive integer r′′ such that
(6)
fγ (z) −
1
r ′′
sn,2d (z) ∈ Kn,2d
.
r′
We now proceed in two cases. If r′′ ≤ r′ , we take r = r′ and use property (c) to
conclude. If r′ ≤ r′′ , we have
fγ (z) −
1
r′′ − r′
1
s
(z)
=
f
(z)
−
s
(z)
+
sn,2d (z).
n,2d
γ
n,2d
r′′
r′
r′ · r′′
We take r = r′′ and use (6) and properties (c) and (d) to conclude.
Remark 2.5. Note that condition (d) is subsumed by the more natural condition
r
that Kn,d
be a convex cone for any n, d, and r. However, there are interesting and relevant cones which we cannot prove to be convex though they trivially satisfy condition
(d) (see Theorem 3.2 for an example).
CONVERGING HIERARCHIES FOR POLYNOMIAL OPTIMIZATION
9
3. Semidefinite programming-based hierarchies obtained from Artin’s
and Reznick’s Positivstellensätze. In this section, we construct two different
semidefinite programming-based hierarchies for POPs using Positivstellensätze derived by Artin (Theorem 1.4) and Reznick (Theorem 1.5). To do this, we introduce
two sets of cones that we call the Artin and Reznick cones.
Definition 3.1. We define the Reznick cone of level r to be
r
Rn,2d
:= {p ∈ Hn,2d | p(x) ·
n
X
i=1
x2i
!r
is sos}.
Similarly, we define the Artin cone of level r to be
Arn,2d := {p ∈ Hn,2d | p(x) · q(x) is sos for some sos form q of degree 2r}.
We show that both of these cones produce hierarchies of the type discussed in
Theorem 2.4. Recall that p∗ is the optimal value of problem (1) and that fγ is defined
as in (3) with the change of notation discussed in Remark 2.3.
Theorem 3.2. Consider the hierarchy of optimization problems indexed by r:
lr := sup
γ
γ
(7)
n
1 X 2 d
r
z ) ∈ Rn,2d
.
fγ (z) − (
r i=1 i
s.t.
Then, lr ≤ p∗ for all r, {lr } is nondecreasing, and limr→∞ lr = p∗ .
r
Proof. It suffices to show that the Reznick cones Rn,2d
satisfy properties (a)-(d)
result
will
then
follow
from
that
theorem.
For property (a),
in Theorem 2.4. The
P
P
it is clear that, as ( i x2i )r > 0 and p(x) · ( i x2i )r is a sum of squares and hence
r
nonnegative, p(x) must be nonnegative, so Rn,2d
⊆ Pn,2d . Furthermore, the form
P 2 d
0
sn,2d := ( i xi ) belongs to Rn,2d and is positive definite. Property (b) is verified
P 2 r
as a consequence
P 2 r+1 of Theorem 1.5. For (c), note that if p(x) · ( i xi ) is sos, then
is sos since the product of two sos polynomials is sos. Finally, for
p(x) · ( i xi )
r
property (d), note that Rn,2d
is a convex cone. Indeed, for any λ ∈ [0, 1],
X
X
X
x2i )r
x2i )r + (1 − λ)q(x)(
x2i )r = λp(x)(
(λp(x) + (1 − λ)q(x)) · (
i
i
i
r
r
is sos if p and q are in Rn,2d
. Combining the fact that Rn,2d
is a convex cone and the
P 2 d
r
fact that ( i xi ) ∈ Rn,d , we obtain (d).
Remark 3.3. To solve a fixed level r of the hierarchy given in Theorem 3.2, one
must proceed by bisection on γ. Bisection here would produce a sequence of upper
bounds {Uk } and lower bounds {Lk } on lr as follows. At iteration k, we test whether
k
k
γ = Uk +L
is feasible for (7). If it is, then we take Lk+1 = Uk +L
and Uk+1 = Uk .
2
2
Uk +Lk
If it is not, we take Uk+1 =
and Lk+1 = Lk . We stop when |Ukǫ − Lkǫ | < ǫ,
2
where ǫ is a prescribed accuracy, and the algorithm returns lr,ǫ = Lkǫ . Note that
lr − ǫ ≤ lr,ǫ ≤ lr and that to obtain lr,ǫ , one needs to take a logarithmic (in 1ǫ ) number
of steps using this method.
10
Hence, solving the rth level of this hierarchy using bisection can be done by
semidefinite programming. Indeed, for a fixed r and γ given by the bisection algorithm, one simply needs to test membership of
!
X
1 X 2 d
fγ (z) − (
zi2 )r
zi ) · (
r i
i
to the set of sum of squares polynomials. This amounts to solving a semidefinite
program. We remark that all semidefinite programming-based hierarchies available
only produce an approximate solution to the optimal value of the SDP solved at level
r in polynomial time. This is independent of whether they use bisection (e.g., such as
the hierarchy given in Theorem 3.2 or the one based on Stengle’s Positivstellensatz)
or not (e.g., the Lasserre hierarchy).
PnOur next theorem improves on our previous hierarchy by freeing the multiplier
( i=1 zi2 )r and taking advantage of our ability to search for an optimal multiplier
using semidefinite programming.
Theorem 3.4. Recall the definition of Artin cones from Definition 3.1. Consider
the hierarchy of optimization problems indexed by r:
lr := sup
γ
γ,q
(8)
s.t.
n
1 X 2 d
z ) ∈ Arn,2d .
fγ (z) − (
r i=1 i
Then, lr ≤ p∗ for all r, {lr } is nondecreasing, and limr→∞ lr = p∗ .
Proof. Just as the previous theorem, it suffices to show that the Artin cones Arn,2d
satisfy properties (a)-(d) of Theorem 2.4. The proof of property (a) follows the proof
given for Theorem 3.2. Property (b) is satisfied as a (weaker) consequence of Artin’s
sos for some sos
result (see Theorem 1.4). For (c), we have
P that if p(x) · q(x) isP
polynomial of degree 2r, then p(x) · q(x) · ( i x2i ) is sos, and q(x) · ( i x2i ) has degree
2(r + 1). Finally, for (d), suppose that p ∈ Arn,2d . Then there exists an sos form q
such that p(x) · q(x) is sos. We have
!
X
X
2 d
x2i )d · q(x),
xi ) · q(x) = p(x) · q(x) + ǫ(
p(x) + ǫ(
i
i
which is sos as the product (resp. sum) of two sos polynomials is sos.
Note that again, for any fixed r, the level r of the hierarchy can be solved using
bisection which leads to a sequence of semidefinite programs.
Our developments in the past two sections can be phrased in terms of a Positivstellensatz.
Corollary 3.5 (A new Positivstellensatz).
Consider the basic semialgebraic
set
S := {x ∈ Rn | gi (x) ≥ 0, i = 1, . . . , m}
and a polynomial p := p(x). Suppose that S is contained within a ball of radius R.
Let ηi and β be any finite upperbounds on gi (x) and, respectively, −p(x) over the set
S.3 Let d be such that 2d is the smallest integer larger than or equal to the maximum
3 As
discussed at the beginning of Section 2, such bounds are very easily computable.
11
CONVERGING HIERARCHIES FOR POLYNOMIAL OPTIMIZATION
degree of p, gi , i = 1, . . . , m. Then, p(x) > 0 for all x ∈ S if and only if there exists a
positive integer r such that
r
2d
m+1
m+1
n
n
X
X
X
X
1
s2j + y 2
s2j + y 2 ·
x2i +
x2 +
h(x, s, y) −
r i=1 i
j=0
j=0
i=1
is a sum of squares, where the form h in variables (x1 , . . . , xn , s0 , . . . , sm+1 , y) is as
follows:
h(x, s, y) := y 2d p(x/y) + s20 y 2d−2
+
(R +
m
X
i=1
2
+
m
X
i=1
y 2d gi (x/y) − s2i y 2d−2
2
m
n
X
X
s2i )d − s2d
x2i +
ηi + β)d y 2d − (
m+1
i=1
i=0
!2
.
Proof. This is an immediate corollary of arguments given in the proof of Theorem
2.1 and in the proof of Theorem 3.2 for the case where γ = 0.
4. Polyá’s theorem and hierarchies for POPs that are optimizationfree, LP-based, and SOCP-based. In this section, we use a result by Polyá on
global positivity of even forms to obtain new hierarchies for polynomial optimization
problems. In Section 4.1, we present a hierarchy that is optimization-free, in the sense
that each level of the hierarchy only requires multiplication of two polynomials and
checking if the coefficients of the resulting polynomial are nonnegative. In Section 4.2,
we use the previous hierarchy to derive linear programming and second-order cone
programming-based hierarchies with faster convergence. These rely on the recently
developed concepts of dsos and sdsos polynomials (see Definition 4.7 and [3]), which
are alternatives to sos polynomials that have been used in diverse applications to
improve scalability; see [3, Section 4].
4.1. An optimization-free hierarchy of lower bounds for POPs. The
main theorem in this section presents an optimization-free hierarchy of lower bounds
for general POPs with compact feasible sets:
Theorem 4.1. Recall the definition of fγ (z) as given in (3), with z ∈ Rn and
deg(fγ ) = 2d. Let (v, w) ∈ R2n and define
(9)
!
n
X
X
2
1 X 4
4 d
r
2
2
vi2 +
P oln,2d := {p ∈ Hn,2d | p(v − w ) + ( (vi + wi )) · (
wi2 )r
2r i=1
i
i
has nonnegative coefficients }.
Consider the hierarchy of optimization problems indexed by r:
lr := sup
γ
γ
(10)
s.t.
n
1 X 2 d
r
z ) ∈ P oln,2d
.
fγ (z) − (
r i=1 i
Let mr = maxi=1,...,r li . Then mr ≤ p∗ for all r, {mr } is nondecreasing, and
limr→∞ mr = p∗ .
12
As before, we use bisection to obtain the optimal value lr of the rth level of the
hierarchy up to a fixed precision ǫ (see Remark 3.3). At each step of the bisection
algorithm, one simply needs to multiply two polynomials together and check nonnegativity of the coefficients of the resulting polynomial to proceed to the next step. As
a consequence, this hierarchy is optimization-free as we do not need to solve (convex)
optimization problems at each step of the bisection algorithm. To the best of our
knowledge, no other converging hierarchy of lower bounds for general POPs dispenses
altogether with the need to solve convex subprograms. We also provide a Positivstellensatz counterpart to the hierarchy given above (see Corollary 4.5). This corollary
implies in particular that one can always certify infeasibility of a basic semialgebraic
set by recursively multiplying polynomials together and simply checking nonnegativity
of the coefficients of the resulting polynomial.
We now make a few remarks regarding the techniques used in the proof of Theor
rem 4.1. Unlike Theorems 3.2 and 3.4, we do not show that P oln,d
satisfies properties
(a)-(d) as given in Theorem 2.4 due to some technical difficulties. It turns out however that we can avoid showing properties (c) and (d) by using a result by Reznick
and Powers [18] that we present below. Regarding properties (a) and (b), we show
that a slightly modified version of (a) holds and that (b), which is the key property
in Theorem 2.4, goes through as is. We note though that obtaining (b) from Polyá’s
result (Theorem 1.6) is not as immediate as obtaining (b) from Artin’s and Reznick’s
results. Indeed, unlike the theorems by Artin and Reznick (see Theorems 1.4 and
1.5) which certify global positivity of any form, Polyá’s result only certifies global
positivity of even forms. To make this latter result a statement about general forms,
we work in an appropriate lifted space. This is done by replacing any form p(z) in
variables z ∈ Rn by the even form p(v 2 − w2 ) in variables (v, w) ∈ R2n . This lifting
operation preserves nonnegativity, but unfortunately it does not preserve positivity:
even if p(z) is pd, p(v 2 − w2 ) always has zeros (e.g., when v = w). Hence, though we
now have access to an even form, we still cannot use Polyá’s property as p(v 2 − w2 )
is not positive. This
P is what leads us to consider the slightly more complicated form
1
( i vi4 + wi4 )d in (9).
p(v 2 − w2 ) + 2r
Theorem 4.2 (Powers and Reznick [18]). Let α = (α1 , . . . , αn ) ∈ Nn , xα =
n
. . . xα
n , and write |α| = α1 +. . .+αn . Denote the standard simplex by ∆n . Assume
that f is a form of degree 2d that is positive on ∆n and let
1
xα
1
λ = λ(f ) := min f (x).
x∈∆n
Define c(α) =
(2d)!
α1 !...αn ! .
We have:
f (x) =
X
|α|=2d
aα xα =
X
bα c(α)xα .
|α|=2d
Let ||f (x)|| := max|α|=2d |bα |.4
Then, the coefficients of
f (x1 , . . . , xn ) · (x1 + . . . + xn )N
are nonnegative for N > d(2d − 1) ||f (x)||
− 2d.
λ
4 As
defined, ||f || is a submultiplicative norm; see [23].
CONVERGING HIERARCHIES FOR POLYNOMIAL OPTIMIZATION
13
Note that here the bound is given in the case where one considers the alternative
(but equivalent) formulation of Polyá’s Positivstellensatz to the one given in Theorem
1.6, i.e., when one is concerned with positivity of a form over the simplex. The result
can easily be adapted to the formulation where one considers global positivity of an
even form as shown below.
Lemma 4.3. Let p := p(x) be an even form of degree 2d that is positive definite.
Let β > 0 be its minimum on Sx . Then,
X
x2i )N
p(x1 , . . . , xn ) · (
i
√
has nonnegative coefficients for N > d(2d − 1) ||p( β x)|| − 2d.
√
√
Proof. Let f (x1 , . . . , xn ) = p( x1 , . . . , xn ). Since p(x) ≥ β on Sx , then f (x) ≥
β on ∆n . Indeed, by contradiction, suppose
√ that there existsPx̂ ∈ ∆n such that
fP(x̂) = β − ǫ (where ǫ > 0) and let y = x̂. Note that as
i x̂i = 1, we have
2
y
=
1.
Furthermore,
p(y)
=
f
(x̂)
=
β
−
ǫ
which
contradicts
the assumption.
i i
√
||p( x)||
− 2d,
Hence, using Theorem 4.2, we have that when N > d(2d − 1) β
X
xi )N
f (x)(
i
has nonnegative coefficients. Hence,
X
X
yi2 )N
yi2 )N = p(y)(
f (y 2 )(
i
i
also has nonnegative coefficients.
Before we proceed with the proof of Theorem 4.1, we need the following lemma.
Lemma 4.4. Let
(11)
n
n
1 X
1 X
pγ,r (v, w) := fγ (v 2 − w2 ) − ( (vi2 − wi2 )2 )d + ( (vi4 + wi4 ))d ,
r i=1
2r i=1
where fγ is defined as in (3) and let
N (r) = d(2d − 1) ·
√ √
||pγ,r ( v, w)||
− 2d.
minSv,w pγ,r (v, w)
If fγ (z) is positive definite, there exists r̂ such that r2 ≥ N (r), for all r ≥ r̂.
Proof. P
As fγ (z) is positive definite, there exists a positive integer r0 such that
fγ (z) − r1 ( i zi2 )d is positive definite for all r ≥ r0 and hence
(12)
1 X
fγ (v 2 − w2 ) − ( (vi2 − wi2 )2 )d
r i
P
is nonnegative for all r ≥ r0 . Recall now that ||x||p = ( i xpi )1/p is a norm for p ≥ 1
and that
√
||x||2 ≤ n||x||4 .
14
This implies that
X
X
wi4 )d ≥
vi4 +
(
i
i
X
X
1
wi2 )2d
vi2 +
(
2d
(2n)
i
i
and hence in view of (12) and the definition of pγ,r , we have
pγ,r (v, w) ≥
X
X
1
2
wi2 )2d , ∀r ≥ r0 .
v
+
(
i
22d+1 n2d r i
i
This enables us to conclude that
(13)
min pγ,r (v, w) ≥
Sv,w
1
, for any r ≥ r0 .
22d+1 n2d r
Further, notice that using properties of the norm, we have the following chain of
inequalities for any positive integer r:
√ √
1 X
1 X
||pγ,r ( v, w)|| ≤ ||fγ (v − w)|| + ||( (vi − wi )2 )d || + ||( (vi2 + wi2 ))d ||
r
2r
i
i
X
X
2
2 d
vi + wi2 )d || =: cγ .
≤ ||fγ (v − w)|| + ||( (vi − wi ) ) || + ||(
i
i
As a consequence, combining this with the definition of N (r) and (13), we have
N (r) ≤ d(2d − 1)22d+1 rn2d cγ , ∀r ≥ r0 .
Now taking r̂ = max(r0 , ⌈d(2d − 1)22d+1 n2d cγ ⌉), we have r2 ≥ N (r), ∀r ≥ r̂.
We now proceed with the proof of Theorem 4.1.
Proof of Theorem 4.1. By definition, the sequence {mr } is nondecreasing. We
show that it is upperbounded by p∗ by showing that if γ is such that
1 X 2 d
r
fγ (z) − (
z ) ∈ P oln,2d
,
r i i
for some r, then fγ must be positive definite. Then Theorem 2.1 gives us that γ is
a strict lower bound on (1). As p∗ > γ for any such γ, we have that lr ≤ p∗ , ∀r and
hence mr ≤ p∗ , ∀r.
Assume that γ is such that
1 X 2 d
r
fγ (z) − (
z ) ∈ P oln,2d
r i i
P
P
2
r
for some r. By definition of P oln,2d
and as ( i vi2 + i wi2 )r is nonnegative, we get
that the form
1 X
1 X 4
fγ (v 2 − w2 ) − ( (vi2 − wi2 )2 )d + (
v + wi4 )d
r i
2r i i
is nonnegative. This implies that
(14)
X
X
1 X
1
fγ (v 2 − w2 ) − ( (vi2 − wi2 )2 )d ≥ −
wi4 = 1},
vi4 +
for (v, w) ∈ {(v, w) |
r i
2r
i
i
CONVERGING HIERARCHIES FOR POLYNOMIAL OPTIMIZATION
15
which gives
1 X 2 d
1
fγ (z) − (
z ) ≥ − , ∀z ∈ Sz .
r i i
2r
(15)
Indeed, suppose that there exists ẑ ∈ Sz such that (15) does not hold. Then, let
−
+
−
ẑ + = max(ẑ, 0) and
√ ẑ = max(−ẑ,
√ 0). Note that both ẑ and ẑ are nonnegative so
+
−
we
take
have that as ẑ ∈ Sz and ẑ = v̂ 2 − ŵ2 ,
P v̂ 4= ẑ and ŵ = ẑ . We2further
P can
2
4
i ŵi = 1. Substituting ẑ by v̂ − ŵ in (15) then violates (14). Using (15),
i v̂i +
we conclude that
1
fγ (z) ≥ , ∀z ∈ Sz
2r
and that fγ is positive definite.
We now show that the hierarchy converges, i.e., that limr→∞ mr = p∗ . To do
this, we show that if γ is a strict lower bound on (1), or equivalently from Theorem
2.1, if fγ (z) is positive definite, then there exists r′ such that
fγ (z) −
1 X 2 d
r′
z ) ∈ P oln,2d
.
(
r′ i i
Pn
Since fγ is pd, there exists a positive integer r0 such
that fγ (z) − r1 ( i=1 zi2 )d is pd
P
1
for any r ≥ r0 . This implies that fγ (v 2 − w2 ) − r ( i (vi2 − wi2 )2 )d is nonnegative and
1 X
1 X
fγ (v 2 − w2 ) − ( (vi2 − wi2 )2 )d + ( (vi4 + wi4 ))d
r i
2r i
is positive definite for r ≥ r0 . Using Lemma 4.3 and the definition of N (r) in Lemma
4.4, for any r ≥ r0 , we have that
!
X
X
1 X 4
1 X 2
2 2 d
4 d
2
2
wi2 )⌈N (r)⌉
vi2 +
fγ (v − w ) − ( (vi − wi ) ) + ( (vi + wi )) · (
r i
2r i
i
i
has nonnegative coefficients. From Lemma 4.4, there exists r̂ such that r ≥ r̂ implies
r2 ≥ N (r). Taking r′ = max{r0 , r̂} and considering pγ,r′ as defined in (11), we get
that
X
X
′2
wi2 )r
vi2 +
pγ,r′ (v, w)(
i
i
X
X
X
X
′2
′
′
wi2 )r −⌈N (r )⌉
vi2 +
wi2 )⌈N (r )⌉ · (
vi2 +
= pγ,r′ (v, w)(
i
i
i
i
has nonnegative coefficients, which is the desired result. This is because
X
X
′
wi2 )⌈N (r )⌉
vi2 +
pγ,r′ (v, w)(
i
i
has nonnegative coefficients as r′ ≥ r0 , and
X
X
′2
′
wi2 )r −⌈N (r )⌉
vi2 +
(
i
i
has nonnegative coefficients as r′ ≥ r̂, and that the product of two polynomials with
nonnegative coefficients has nonnegative coefficients.
16
Corollary 4.5 (An optimization-free Positivstellensatz).
semialgebraic set
S := {x ∈ Rn | gi (x) ≥ 0, i = 1, . . . , m}
Consider the basic
and a polynomial p := p(x). Suppose that S is contained within a ball of radius R. Let
ηi and β be any finite upperbounds on gi (x) and, respectively, −p(x) over the set S.5
Let d be such that 2d is the smallest even integer larger than or equal to the maximum
degree of p, gi , i = 1, . . . , m. Then, p(x) > 0 for all x ∈ S if and only if there exists a
positive integer r such that
!
n+m+3
n+m+3
1 X
1 X
2
2
2
2 2 d
4
4 d
h(v − w ) − (
(vi − wi ) ) + (
(vi + wi ))
r i=1
2r i=1
!r 2
n+m+3
n+m+3
X
X
2
2
wi
vi +
·
i=1
i=1
has nonnegative coefficients, where the form h := h(z) in variables
(z1 , . . . , zn+m+3 ) := (x1 , . . . , xn , s0 , . . . , sm+1 , y)
is as follows:
h(x, s, y) := y 2d p(x/y) + s20 y 2d−2
+
(R +
m
X
i=1
2
+
d 2d
ηi + β) y
m
X
i=1
y 2d gi (x/y) − s2i y 2d−2
2
m
n
X
X
s2i )d − s2d
x2i +
−(
m+1
i=1
i=0
!2
.
Proof. This is an immediate corollary of arguments given in the proof of Theorem 2.1 and in the proof of Theorem 4.1 for the case where γ = 0.
4.2. Linear programming and second-order cone programming-based
hierarchies for POPs. In this section, we present a linear programming and a
second-order cone programming-based hierarchy for general POPs which by construction converge faster than the hierarchy presented in Section 4.1. These hierarchies are
based on the recently-introduced concepts of dsos and sdsos polynomials [3] which we
briefly revisit below to keep the presentation self-contained.
Definition 4.6. A symmetric matrix M
P is said to be
• diagonally dominant (dd) if Mii ≥ j6=i |Mij | for all i.
• scaled diagonally dominant (sdd) if there exists a diagonal matrix D, with
positive diagonal entries, such that DAD is dd.
We have the following implications as a consequence of Gershgorin’s circle theorem:
(16)
M dd ⇒ M sdd ⇒ M 0.
Requiring M to be dd (resp. sdd) can be encoded via a linear program (resp. a
second-order cone program) (see [3] for more details). These notions give rise to the
concepts of dsos and sdsos polynomials.
5 Once
again, as discussed at the beginning of Section 2, such bounds are very easily computable.
CONVERGING HIERARCHIES FOR POLYNOMIAL OPTIMIZATION
17
Definition 4.7 ([3]). Let z(x) = (xd1 , xd−1
x2 , . . . , xdn )T be the vector of mono1
mials in (x1 , . . . , xn ) of degree d. A form p ∈ Hn,2d is said to be
• diagonally-dominant-sum-of-squares (dsos) if it admits a representation
p(x) = z T (x)Qz(x), where Q is a dd matrix.
• scaled-diagonally-dominant-sum-of-squares (sdsos) if it admits a representation
p(x) = z T (x)Qz(x), where Q is a sdd matrix.
The following implications are a consequence of (16):
(17)
p(x) dsos ⇒ p(x) sdsos ⇒ p(x) sos ⇒ p(x) nonnegative.
Given the fact that our Gram matrices and polynomials are related to each other via
linear equalities, it should be clear that optimizing over the set of dsos (resp. sdsos)
polynomials is an LP (resp. SOCP).
We now present our LP and SOCP-based hierarchies for POPs.
Corollary 4.8. Recall the definition of fγ (z) as given in (3), with z ∈ Rn and
deg(f ) = 2d, and let pγ,r be as in (11). Consider the hierarchy of optimization
problems indexed by r:
lr := sup
γ
γ,q
(18)
s.t.
pγ,r (v, w) · q(v, w) is s/dsos
q(v, w) is s/dsos and of degree 2r2 .
Let mr = maxi=1,...,r li . Then, mr ≤ p∗ for all r, {mr } is nondecreasing, and we have
limr→∞ mr = p∗ .
Proof. This is an immediate consequence of the fact that any even form p ∈ Hn,2d
with nonnegative coefficients can be written as p(x) = z(x)T Qz(x) where Q is diagonal
and has nonnegative (diagonal) entries. As such a Q is dd (and also sdd), we conclude
that p is dsos (and also sdsos). The corollary then follows from Theorem 4.1.
Note that similarly to our previous hierarchies, one must proceed by bisection on
γ to solve the level r of the hierarchy. At each step of the hierarchy, we solve a linear
program (resp. second-order cone program) that searches for the coefficients of q that
make q dsos (resp. sdsos) and pγ,r · q dsos (resp. sdsos).
There is a trade-off between the hierarchies developed in this subsection and the
one developed in the previous subsection: the hierarchy of Section 4.1 is optimizationfree whereas those of Section 4.2 use linear or second-order cone programming. Hence
the former hierarchy is faster to run at each step. However, the latter hierarchies could
potentially take fewer levels to converge. This is similar to the trade-off observed
between the hierarchies presented in Theorems 3.2 and 3.4.
5. Open problems. To conclude, we present two open problems spawned by
the writing of this paper. The first one concerns the assumptions needed to construct
our hierarchies.
Open problem 1. Theorems 2.1 and 2.4 require that the feasible set S of the POP
given in (1) be contained in a ball of radius R. Can these theorems be extended to
the case where there is no compactness assumption on S?
The second open problem is linked to the Artin and Reznick cones presented in
Definition 3.1.
18
r
Open problem 2. As mentioned before, Reznick cones Rn,2d
are convex for all r.
r
We are unable to prove however that Artin cones An,2d are convex (even though they
satisfy properties (a)-(d) of Theorem 2.4 like Reznick cones do). Are Artin cones
convex for all r? We know that they are convex for r = 0 and for r large enough as
they give respectively the sos and psd cones (see [12] for the latter claim). However,
we do not know the answer already for r = 1.
Acknowledgments. We are grateful to Pablo Parrilo for very insightful comments, particularly as regards Section 4 and the observation that any form can be
made even by only doubling the number of variables and the degree.
REFERENCES
[1] A. A. Ahmadi, On the difficulty of deciding asymptotic stability of cubic homogeneous vector
fields, in Proceedings of the American Control Conference, 2012.
[2] A. A. Ahmadi and A. Majumdar, Some applications of polynomial optimization in operations research and real-time decision making, Optimization Letters, 10 (2016), pp. 709–729.
[3] A. A. Ahmadi and A. Majumdar, DSOS and SDSOS optimization: more tractable alternatives to sum of squares and semidefinite optimization, (2017). Available on arXiv at
https://arxiv.org/abs/1706.02586.
[4] E. Artin, Über die Zerlegung definiter Funktionen in Quadrate, in Abhandlungen aus dem
mathematischen Seminar der Universität Hamburg, vol. 5, Springer, 1927, pp. 100–115.
[5] G. Averkov, Constructive proofs of some Positivstellensätze for compact semialgebraic subsets
of Rd , Journal of Optimization Theory and Applications, 158 (2013), pp. 410–418.
[6] G. Blekherman, P. A. Parrilo, and R. R. Thomas, Semidefinite Optimization and Convex
Algebraic Geometry, SIAM, 2012.
[7] E. de Klerk, M. Laurent, and P. Parrilo, On the equivalence of algebraic approaches
to the minimization of forms on the simplex, Positive Polynomials in Control, (2005),
pp. 580–580.
[8] W. Habicht, Über die Zerlegung strikte definiter Formen in Quadrate, Commentarii Mathematici Helvetici, 12 (1939), pp. 317–322.
[9] J. B. Lasserre, Global optimization with polynomials and the problem of moments, SIAM
Journal on Optimization, 11 (2001), pp. 796–817.
[10] J. B. Lasserre, Moments, Positive Polynomials and their Applications, vol. 1, World Scientific, 2009.
[11] M. Laurent, Sums of squares, moment matrices and optimization over polynomials, in Emerging applications of algebraic geometry, Springer, 2009, pp. 157–270.
[12] H. Lombardi, D. Perrucci, and M.-F. Roy, An elementary recursive bound for effective
Positivstellensatz and Hilbert 17-th problem, arXiv preprint arXiv:1404.2338, (2014).
[13] J. Nie and M. Schweighofer, On the complexity of Putinar’s Positivstellensatz, Journal of
Complexity, 23 (2007), pp. 135–150.
[14] P. M. Pardalos and S. A. Vavasis, Quadratic programming with one negative eigenvalue is
NP-hard, Journal of Global Optimization, 1 (1991), pp. 15–22.
[15] P. A. Parrilo, Structured semidefinite programs and semialgebraic geometry methods in robustness and optimization, PhD thesis, California Institute of Technology, May 2000.
[16] P. A. Parrilo, Semidefinite programming relaxations for semialgebraic problems, Mathematical Programming, 96 (2003), pp. 293–320.
[17] G. Pólya, Über positive Darstellung von Polynomen, Vierteljschr. Naturforsch. Ges. Zürich,
73 (1928), pp. 141–145.
[18] V. Powers and B. Reznick, A new bound for Pólya’s theorem with applications to polynomials positive on polyhedra, Journal of pure and applied algebra, 164 (2001), pp. 221–229.
[19] M. Putinar, Positive polynomials on compact semi-algebraic sets, Indiana University Mathematics Journal, 42 (1993), pp. 969–984.
[20] B. Reznick, Uniform denominators in Hilbert’s 17th problem, Math Z., 220 (1995), pp. 75–97.
[21] K. Schmüdgen, The k-moment problem for compact semi-algebraic sets, Mathematische Annalen, 289 (1991), pp. 203–206.
[22] M. Schweighofer, An algorithmic approach to Schmüdgen’s Positivstellensatz, Journal of
Pure and Applied Algebra, 166 (2002), pp. 307–319.
[23] M. Schweighofer, On the complexity of Schmüdgen’s Positivstellensatz, Journal of Complexity, 20 (2004), pp. 529–543.
CONVERGING HIERARCHIES FOR POLYNOMIAL OPTIMIZATION
19
[24] M. Schweighofer, Optimization of polynomials on compact semialgebraic sets, SIAM Journal
on Optimization, 15 (2005), pp. 805–825.
[25] G. Stengle, A Nullstellensatz and a Positivstellensatz in semialgebraic geometry, Mathematische Annalen, 207 (1974), pp. 87–97.
| 3cs.SY
|
arXiv:1601.01679v2 [math.GR] 14 Jan 2016
MORE ON REGULAR SUBGROUPS OF THE AFFINE GROUP
M.A. PELLEGRINI AND M.C. TAMBURINI BELLANI
Abstract. This paper is a new contribution to the study of regular subgroups
of the affine group AGLn (F), for any field F. In particular we associate to any
partition λ 6= (1n+1 ) of n + 1 abelian regular subgroups in such a way that
different partitions define non-conjugate subgroups. Moreover, we classify the
regular subgroups of certain natural types for n ≤ 4. Our classification is
equivalent to the classification of split local algebras of dimension n + 1 over F.
Our methods, based on classical results of linear algebra, are computer free.
1. Introduction
Let F be any field. We identify the affine group AGLn (F) with the subgroup of
GLn+1 (F) consisting of the matrices having (1, 0, . . . , 0)T as first column. With this
notation, AGLn (F) acts on the right on the set A = {(1, v) : v ∈ Fn } of affine points.
Clearly, there exists an epimorphism π : AGLn (F) → GLn (F) induced by the action
of AGLn (F) on Fn . A subgroup (or a subset) R of AGLn (F) is called regular if it
acts regularly on A, namely if, for every v ∈ Fn , there exists a unique element in R
c n (F) R,
having (1, v) as first row. Thus R is regular precisely when AGLn (F) = GL
c n (F) ∩ R = {In+1 }, where GL
c n (F) denotes the stabilizer of (1, 0, . . . , 0).
with GL
A subgroup H of AGLn (F) is indecomposable if there exists no decomposition of
Fn as a direct sum of non-trivial π(H)-invariant subspaces. Clearly, to investigate
the structure of regular subgroups, the indecomposable ones are the most relevant,
since the other ones are direct products of regular subgroups in smaller dimensions.
So, one has to expect very many regular subgroups when n is big. Actually, in
Section 6 we show how to construct at least one abelian regular subgroup, called
standard, for each partition λ 6= (1n+1 ) of n + 1, in such a way that different
partitions produce non-conjugate subgroups. Several of them are indecomposable.
The structure and the number of conjugacy classes of regular subgroups depend
on F. For instance, if F has characteristic p > 0, every regular subgroup is unipotent
[14, Theorem 3.2], i.e., all its elements satisfy (t − In+1 )n+1 = 0. A unipotent group
is conjugate to a subgroup of the group of upper unitriangular matrices (see [10]):
in particular it has a non-trivial center. By contrast to the case p > 0, AGL2 (R)
contains 2|R| conjugacy classes of regular subgroups with trivial center, hence not
unipotent (see Example 2.5). So, clearly, a classification in full generality is not
realistic.
Since the center Z(R) of a regular subgroup R is unipotent (see Theorem 2.4(a))
if Z(R) is non-trivial one may assume, up to conjugation, that R is contained in the
centralizer of a unipotent Jordan form (see Theorem 4.4). But even this condition
is weak.
2010 Mathematics Subject Classification. 15A21, 20B35, 16L99.
Key words and phrases. Regular subgroup; local algebra; Jordan canonical form.
1
2
M.A. PELLEGRINI AND M.C. TAMBURINI BELLANI
Before introducing a stronger hypothesis, which allows to treat significant cases,
we need some notation. We write every element r of R as
1
v
1
v
1
v
(1)
r=
=
=
= µR (v),
0 π(r)
0 τR (v)
0 In + δR (v)
where µR : Fn → AGLn (F), τR : Fn → GLn (F) and δR := τR − id : Fn → Matn (F).
The hypothesis we introduce is that δR is linear. First of all, if R is abelian, then
δR is linear (see [1]). Moreover, if δR is linear, then R is unipotent by Theorem
2.4(b), but not necessarily abelian. One further motivation for this hypothesis
is that δR is linear if and only if L = FIn+1 + R is a split local subalgebra of
Matn+1 (F). Moreover, two regular subgroups R1 and R2 , with δRi linear, are
conjugate in AGLn (F) if and only if the corresponding algebras L1 and L2 are
isomorphic (see Section 3). In particular, there is a bijection between conjugacy
classes of abelian regular subgroups of AGLn (F) and isomorphism classes of abelian
split local algebras of dimension n + 1 over F. This fact was first observed in [1]. It
was studied also in connection with other algebraic structures in [3, 4, 5, 6] and in
[2], where the classification of nilpotent associative algebras given in [7] is relevant.
In Section 7 we classify, up to conjugation, certain types of regular subgroups
U of AGLn (F), for n ≤ 4, and the corresponding algebras. More precisely, for
n = 1 the only regular subgroup is the translation subgroup T , which is standard.
For n = 2, 3, we assume that δU is linear. If n = 2 all subgroups U , over any F,
are standard (Table 1). For n = 3 the abelian regular subgroups are described in
Table 2. The non abelian ones are determined in Lemmas 7.2 and 7.3: there are
|F| conjugacy classes when char F = 2, |F| + 1 otherwise (see also [7]). If n = 4
we assume that U is abelian. The conjugacy classes, when F has no quadratic
extensions, are shown in Tables 3 and 4.
In particular, by the reasons mentioned above, we obtain an independent classification of the split local algebras of dimension n ≤ 4 over any field. We obtain
also the classification of the commutative split local algebras of dimension 5 over
fields with no quadratic extensions. Actually, regular subgroups arise from matrix
representations of these algebras. In the abelian case and using the hypothesis
that F is algebraically closed, the same classification, for n ≤ 6, was obtained
by Poonen [12], via commutative algebra. Namely he presents the algebras as
quotients of the polynomial ring F[t1 , . . . , tr ], r ≥ 1. The two approaches are
equivalent but, comparing our results with those of Poonen, we detected an inaccuracy1. Namely for n = 5 and char F = 2, the two algebras defined, respectively,
by F[x, y, z]/hx2 , y 2 , xz, yz, xy + z 2 i and F[x, y, z]/hxy, xz, yz, x2 + y 2 , x2 + z 2 i are
isomorphic (see Remark 7.6).
Our methods, based on linear algebra, are computer independent.
2. Some basic properties and examples of regular subgroups
Lemma 2.1. A regular submonoid R of AGLn (F) is a subgroup.
Proof. For any v ∈ Fn , the first row of µR −vτR (v)−1 µR (v) is the
same as the
−1
first row of In+1 . From the regularity
of
R
it
follows
µ
−vτ
(v)
µR (v) = In+1 ,
R
R
whence µR (v)−1 = µR −vτR (v)−1 ∈ R. Thus R is a subgroup.
1See [13] for a revised version of [12].
MORE ON REGULAR SUBGROUPS OF THE AFFINE GROUP
3
If δR ∈ HomZ (Fn , Matn (F)), i.e. δR is additive, direct calculation gives that a
regular subset R of AGLn (F) containing In+1 is a submonoid, hence a subgroup, if
and only if:
(2)
δR (vδR (w)) = δR (v)δR (w),
for all v, w ∈ Fn .
Also, given v, w ∈ Fn ,
(3)
µR (v)µR (w) = µR (w)µR (v)
if and only if
vδR (w) = wδR (v).
For the next two theorems we need a basic result, that we recall below for the
reader’s convenience. A proof can be found in any text of linear algebra.
Lemma 2.2. Let g ∈ GLm (F) have characteristic polynomial χg (t) = f1 (t)f2 (t).
If (f1 (t), f2 (t)) = 1, then g is conjugate to h = diag(h1 , h2 ), with χhi (t) = fi (t),
i = 1, 2. Also, CMatm (F) (h) consists of matrices of the same block-diagonal form.
Lemma 2.3. Let z be an element of a
unipotent then, up to conjugation of R
1
(4)
z = 0
0
regular subgroup R of AGLn (F). If z is not
under AGLn (F), we may suppose that
w1 0
A1 0 ,
0 A2
where A1 is unipotent and A2 does not have the eigenvalue 1.
1 w
, where z0 = In + δR (w). Since z0 has the eigenvalue 1 [14,
Proof. Let z =
0 z0
Lemma 2.2], the characteristic polynomial χz0 (t) of z0 factorizes in F[t] as
χz0 (t) = (t − 1)m1 g(t),
g(1) 6= 0,
m1 ≥ 1,
deg g(t) = m2 ≥ 1.
c n (F), we may set z0 = diag(A1 , A2 ) with:
By Lemma 2.2, up to conjugation in GL
χA1 (t) = (t − 1)m1 ,
χA2 (t) = g(t).
and w2 ∈ Fm2 . Since A2 does not have the
Write w = (w1 , w2 ) with w1 ∈ F
eigenvalue 1, the matrix A
2 − Im2 is invertible. Conjugatingby the translation
1
0
w2 (A2 − Im2 )−1
1 w1 0
0 Im1
we may assume w2 = 0, i.e. z = 0 A1 0 .
0
0
0
Im2
0 0 A2
m1
We observe that if δR ∈ HomF (Fn , Matn (F)), i.e. δR is linear, then there is a
b of AGLn (F)
b for any extension
natural embedding of R into a regular subgroup R
b
F of F. Namely:
1
v̂
b=
b ≤ AGLn (F),
b
R
: v̂ ∈ Fn ⊗F F
0 In + δRb (v̂)
P
P
b
λ̂v
where δRb
i =
b is linear.
i λ̂i δR (vi ), λ̂i ∈ F. Clearly δR
i
We obtain the following consequences.
Theorem 2.4. Let R be regular subgroup of AGLn (F). Then the following holds:
(a) the center Z(R) of R is unipotent;
(b) if δR is linear, then R is unipotent.
4
M.A. PELLEGRINI AND M.C. TAMBURINI BELLANI
Proof. (a) Our claim is clear if Z(R) = {1}. So let 1 6= z ∈ Z(R) and assume,
by contradiction, that z is not unipotent. By Lemma 2.3 up to conjugation z has
shape (4). Lemma 2.2 gives the contradiction that its centralizer is not transitive
on the affine vectors.
b be the algebraic closure of F. By what observed above, substituting
(b) Let F
b if necessary, we may assume F algebraically closed. By contradicR with R,
tion, suppose that z ∈ R is not unipotent. Up to conjugation we may suppose z as in the statement of Lemma 2.3. In the same notation, δR (w1 , 0) =
diag(A1 − Im1 , A2 − Im2 ). Let ξ ∈ F be an eigenvalue of A2 − Im2 and 0 6=
v ∈ Fm2 be a corresponding eigenvector. We have ξ 6=
0 and, by linearity,
δR (−ξ −1 w1 , 0) = diag −ξ −1 (A1 − Im1 ) , −ξ −1 (A2 − Im2 ) . It follows that the
first row of the product µR (0, v)µR (−ξ −1 w1 , 0) is equal to the first row of the
second factor µR (−ξ −1 w1 , 0). From the regularity of R we get µR (0, v) = In+1 . In
particular v = 0, a contradiction. We conclude that R is unipotent.
The following examples show how the linearity of δR seems to be necessary to
manage a classification of the regular subgroups of AGLn (F), even when unipotent.
Example 2.5. Let B be a basis of R over Q. For every subset S of B denote by
fS : B → R the function such that fS (v) = v if v ∈ S, fS (v) = 0 otherwise.
Let fbS be its extension by linearity to R. In particular fbS ∈ HomZ (R, R) and
Im (fbS ) = SpanR (S), the subspace generated by S.
Consider the regular subgroup RS of the affine group AGL2 (R) defined by:
x
y
1
RS = 0 efbS (y) 0 : x, y ∈ R .
0
0
1
The set of eigenvalues of the matrices in RS is {er : r ∈ SpanR (S)}. It follows that
S1 6= S2 gives RS1 not conjugate to RS2 under GL3 (R). A fortiori RS1 and RS2 are
not conjugate under AGL2 (R). Thus in AGL2 (R) there are as many non conjugate
regular subgroups as possible, namely 2|R| .
Example 2.6. Let F 6= Fp be a field
1
R = 0
0
of characteristic p > 0. Then the set
x1 x2
1 xp1 : x1 , x2 ∈ F
0
1
is a non abelian unipotent regular subgroup of AGL2 (F) such that δR is not linear.
Example 2.7 (Hegedűs, [8]). Let n ≥ 4 and F = Fp (p odd). Take the matrices
1 2 −2
A = diag(A3 , In−4 ) and J = diag(J3 , In−4 ) of GLn−1 (Fp ), where A3 = 0 1 −2
00 1
0 0 1
and J3 = 0 1 0 . Then A has order p, its minimum polynomial has degree 3 and
100
AJAT = J. The subset R of order p4 , defined by
T
1 v h + vJv
2
R = 0 Ah Ah Jv T : v ∈ Fp n−1 , h ∈ Fp ,
0 0
1
is a regular submonoid of AGLn (Fp ), hence a regular subgroup. Moreover R ∩ T =
{1}. Note that δR is not linear: suppose the contrary and observe that the elements
MORE ON REGULAR SUBGROUPS OF THE AFFINE GROUP
5
µR (vn ) and µR (2vn ) correspond to h = 1 and h = 2 respectively (vn = (0, . . . , 0, 1)).
Thus, in order that δR (2vn ) = 2δR (vn ) we should have A2 − In−1 = 2A − 2In−1 ,
i.e. (A − In−1 )2 = In−1 , against our choice of A.
3. Algebras and regular subgroups
In this section we highlight the connections between regular subgroups and finite
dimensional split local algebras over a field F. To this purpose we recall that an
F-algebra L with 1 is called split local if L/J(L) is isomorphic to F, where J(L)
denotes the Jacobson radical of L. In particular L = F 1 + J(L), where the set
F 1 = {α1L : α ∈ F} is a subring of L contained in its centre Z(L). Note that
L \ J(L) is the set L∗ of the invertible elements of L. If J(L), viewed as an Fmodule, has finite dimension, we say that L is finite dimensional.
If ψ is an isomorphism between two local split F-algebras L1 , L2 , then ψ(J(L1 )) =
J(L2 ) and ψ(α 1L1 ) = α1L2 for all α ∈ F.
Theorem 3.1. Let L be a finite dimensional split local F-algebra. In the above
notation, set n = dimF (J(L)). Then, with respect to the product in L, the subset
R = 1 + J(L) = {1 + v : v ∈ J(L)}
is a group, isomorphic to a regular subgroup of AGLn (F) for which δR is linear.
Proof. Clearly R is closed under multiplication. Moreover any r ∈ R has an inverse
in L. From r − 1 ∈ J(L) we get r−1 − 1 ∈ J(L), whence r−1 ∈ R. We conclude
that R is a group. Consider the map R → AGLn (F) such that, for all v ∈ J(L):
1
vB
,
(5)
1 + v 7→
0 In + δR (v)
where vB and δR (v) are, respectively, the coordinate vector of v and the matrix
of the right multiplication by v with respect to a fixed basis B of J(L), viewed as
F-module. In particular, considering the right multiplication by w:
(6)
(vw)B = vB δR (w),
for all v, w ∈ J(L).
The map (5) is injective and we claim that it is a group monomorphism.
Set δ = δR for simplicity. For all v, w ∈ J(L) we have δ(vw) = δ(v)δ(w) by the
associativity law and δ(v + w) = δ(v) + δ(w) by the distributive laws. Now:
1
vB + wB + (vw)B
(1 + v)(1 + w) = 1 + v + w + vw 7→
.
0 In + δ(v) + δ(w) + δ(vw)
On the other hand, considering the images of 1 + v and 1 + w:
1
vB
1
wB
1
wB + vB + vB δ(w)
=
.
0 In + δ(v)
0 In + δ(w)
0 In + δ(v) + δ(w) + δ(v)δ(w)
Thus (5) is a homomorphism if and only if (6) holds. This proves our claim.
Notice that Theorem 3.1 shows how to construct a regular subgroup starting
from the presentation of a split local algebra.
Example 3.2. As in [12], consider the split local algebra
L=
F[t1 , t2 , t3 ]
.
ht1 2 + t2 2 , t1 2 + t3 2 , t1 t2 , t1 t3 , t2 t3 i
6
M.A. PELLEGRINI AND M.C. TAMBURINI BELLANI
In this case J(L) ∼
= F4 has a basis given by {t1 , t1 2 , t2 , t3 } and considering the
right multiplication by an element of this basis we get δ(t1 ) = E1,2 , δ(t1 2 ) = 0,
δ(t2 ) = −E3,2 , δ(t3 ) = −E4,2 , and so the associated regular subgroup is
1 x1 x2 x3 x4
x1
0
0
0 1
1
0
0 : x1 , x2 , x3 , x4 ∈ F .
R = 0 0
0 0 −x3 1
0
0 0 −x4 0
1
Conversely we have the following result.
Theorem 3.3. Let R = µR (Fn ) be a regular subgroup of AGLn (F). Set
0
v
and LR = FIn+1 + V.
V = R − In+1 =
| v ∈ Fn
0 δR (v)
(a) If δR is additive, then V is a subring (without identity) of Matn+1 (F);
(b) the function δR is linear if and only if LR is a split local subalgebra of
Matn+1 (F) with J(LR ) = V .
Proof. Set µ = µR , δ = δR , I = In+1 and L = LR .
(a) Since δ is additive, V is an additive subgroup. From (µ(v) − I)(µ(w) − I) =
(µ(v)µ(w) − I) − (µ(v) − I) − (µ(w) − I) it follows that V is a subring.
(b) Suppose first that δ is linear. Using (a) we have that L is an additive subgroup.
By the linearity, (FI)V = V (FI) = V . It follows that LL = L, hence L is a
subalgebra of Matn+1 (F). Again linearity gives αI + (µ(αv) − I) = αµ(v) for all
α ∈ F, v ∈ Fn . Thus L \ V = F∗ R consists of elements with inverse in L. We
conclude that V = J(L) and L is a split local F-algebra.
Vice-versa, let L be a split local subalgebra with J(L) = V . In particular V is
an additive subgroup, whence δ(v + w) = δ(v) + δ(w) for all v, w ∈ Fn . Since V is
an ideal we get δ(αv) = αδ(v) for all α ∈ F, v ∈ Fn .
Our classification of the regular subgroups of AGLn (F) is based on the following
c n (F) is defined as in the Introduction.
proposition (see [1, Theorem 1]), where GL
Proposition 3.4. Assume that R1 , R2 are regular subgroups of AGLn (F) such that
δR1 and δR2 are linear maps. Then the following conditions are equivalent:
c n (F);
(a) R1 and R2 are conjugate in GL
(b) R1 and R2 are conjugate in AGLn (F);
(c) the algebras LR1 and LR2 are isomorphic.
Proof. Set δ1 = δR1 , L1 = LR1 , δ2 = δR2 , L2 = LR2 , µ1 = µR1 and I = In+1 .
(a) =⇒ (b) =⇒ (c) is clear. Let us prove (c) =⇒ (a). By Theorem 3.3, L1 and L2 are
split local algebras with Jacobson radicals R1 − I and R2 − I, respectively. Suppose
that ψ : L1 → L2 is an algebra isomorphism. In particular ψ(R1 − I) = R2 − I
and ψ induces an F-automorphism of Fn . Let P ∈ GLn (F) be the matrix of this
automorphism with respect to the canonical basis of Fn . Then
0
v
0
vP
ψ
=
.
0 δ1 (v)
0 δ2 (vP )
MORE ON REGULAR SUBGROUPS OF THE AFFINE GROUP
7
For all v, w ∈ Fn we have ψ(µ1 (w)µ1 (v)) = ψ(µ1 (w))ψ(µ1 (v)). This implies
δ1 (v)P = P δ2 (vP ), whence δ2 (vP ) = P −1 δ1 (v)P for all v ∈ Fn . We conclude:
−1
0
v
0
vP
1 0
0
v
1 0
ψ
=
=
.
0 δ1 (v)
0 δ2 (vP )
0 P
0 δ1 (v)
0 P
4. Centralizers of unipotent elements
Our classification of unipotent regular subgroups of AGLn (F) is connected to
the classical theory of canonical forms of matrices. For the reader’s convenience we
recall the relevant facts. For all m ≥ 2, the conjugacy classes of unipotent elements
in GLm (F) are parametrized by the Jordan canonical forms
X
mi ,
(7)
J = diag(Jm1 , . . . , Jmk ), m =
where each Jmi is a Jordan block of size mi , namely a matrix having all 1’s on the
main diagonal and the diagonal above it, and 0’s elsewhere. A Jordan block Jm
has minimal polynomial (t − 1)m and its centralizer is the m-dimensional algebra
x0 x1 . . . xm−2 xm−1
0 x0 . . . xm−3 xm−2
.
..
.
:
x
∈
F
(8)
Tm,m (F) :=
i
.
.
0
0 ...
x0
x1
0
0 ...
0
x0
generated by Jm . To study the centralizer of J in (7) we write c ∈ Matm (F) as:
C1,1 . . . C1,k
.. , C ∈ Mat
..
(9)
c = ...
i,j
mi ,mj (F).
.
.
Ck,1
...
Ck,k
Clearly c centralizes J if and only if
(10)
Jmi Ci,j = Ci,j Jmj ,
for all i, j.
Lemma 4.1. Take J as in (7) and assume further that m1 ≥ · · · ≥ mk . Then the
element c above centralizes J if and only if
Ci,j ∈ Tmi ,mj (F),
for all i, j,
where each Tmi ,mi (F) is defined as in (8) with m = mi and, for mi > mj :
Tmj ,mj (F)
0
Tmi ,mj (F) :=
, Tmj ,mi (F) := 0 . . . 0 Tmj ,mj (F) .
.
..
0
Lemma 4.2. Take J as in (7). If m1 < mi for some i ≥ 2, then the group
CMatm (F) (J) ∩ AGLm−1 (F) is not transitive on affine row vectors (1, x1 , . . . , xm−1 ).
Proof. Consider c ∈ CMatm (F) (J) and decompose it as in (9). Application of (10)
and elementary matrix calculation give that, whenever m1 < mi , the first row of
the matrix C1,i must be the zero vector. Our claim follows immediately.
8
M.A. PELLEGRINI AND M.C. TAMBURINI BELLANI
Lemma 4.3. Let J be as in (7). If k > 1, assume further that m1 ≥ mi for all
i ≥ 2. Then, for every v ∈ Fm , there exists c ∈ CMatm (F) (J) having v as first row.
If the first coordinate of v is non zero, such c can be chosen nonsingular.
Proof. Both claims are direct consequences of the description of Tm,m (F) preceding
(8) and Lemma 4.1.
Theorem 4.4. Let R be a regular subgroup of AGLn (F) and 1 6= z be an element
of the center Z(R) of R. Then, up to conjugation of R under AGLn (F), we may
suppose that z = Jz , where Jz = diag (Jn1 , . . . , Jnk ) is the Jordan form of z having
Jordan blocks of respective sizes ni ≥ ni+1 for all i ≥ 1.
Proof. Let t be the number of non-trivial invariant factors of z.
Case t = 1, i.e. Jz = Jn+1 . Let g ∈ GLn+1 (F) be such that g −1 zg = Jz . Since
must be the
is the eigenspace of Jz (acting on the left) we have that geT
eT
0
0
T
eigenspace of z (acting on the left). From z ∈ AGLn (F) it follows ze0 = eT
0 , hence
−1
T
geT
g ∈ AGLn (F) conjugates z to Jz .
0 = λe0 . We conclude that λ
c n (F) that conjugates z to
Case t > 1. By the unipotency of z, there exists g ∈ GL
1 w1 . . . wh
0 Jm1
z′ = .
.
..
..
.
0
0
...
Jm h
We claim that the first coordinate of wi ∈ Fni cannot be 0 for all i ≥ 1. Indeed, in
this case, there exists ui ∈ Fni such that wi = ui (Ini − Jni ) for all i ≥ 1. Setting
u = (u1 , . . . , uh ) we have
−1
1 u ′ 1 u
1 0
′′
z :=
z
=
, J = diag (Jm1 , . . . , Jmh ) .
0 I
0 I
0 J
It follows that CAGLn (F) (z ′′ ) is not transitive on the affine vectors by Lemma 4.2,
noting that J 6= In : a contradiction. So there exists some t, 1 ≤ t ≤ h, such that
the first coordinate of wt is non zero. Up to conjugation by an obvious permutation
c n (F) we may assume t = 1, i.e., wt = w1 .
matrix in GL
By Lemma 4.3 there exists p ∈ GLn (F) which centralizes diag (Jm1 , . . . , Jmh )
and has v = (w1 , . . . , wh ) as first row. It follows that vp−1 = (1, 0, . . . , 0). Thus
0
1 0 ′ 1
z ′′′ =
= diag(J1+m1 , Jm2 , . . . , Jmh ).
z
0 p−1
0 p
Again by Lemma 4.2 we must have m1 + 1 ≥ mi for all i ≥ 2. A final conjugation,
c n (F) allows to arrange the blocks of z ′′′
if necessary, by a permutation matrix in GL
in non-increasing sizes, i.e. allows to conjugate z ′′′ to Jz as in the statement.
5. Some useful parameters
We introduce some parameters that will be used mainly to exclude conjugacy
among regular subgroups. Let H be any unipotent subgroup of Gn := AGLn (F).
For each h− In+1 ∈ H − In+1 we may consider the degree of its minimal polynomial
over F, denoted by deg minF (h − In+1 ), and its rank, denoted by rk(h − In+1 ). E.g.,
deg minF (h − In+1 ) = n + 1 if and only if h is conjugate to a unipotent Jordan block
Jn+1 of size n + 1. Note that rk(Jn+1 − In+1 ) = n.
MORE ON REGULAR SUBGROUPS OF THE AFFINE GROUP
9
Hence, we may set:
d(H) = max{deg minF (h − In+1 ) | h ∈ H};
r(H) = max{rk(h − In+1 ) | h ∈ H};
k(H) = dimF {w ∈ Fn : wπ(h) = w}.
If H is a subgroup of a regular subgroup U such that δU is linear, then k(H) =
dimF Ker (δU |H ).
Clearly if two unipotent subgroups H1 , H2 are conjugate, then:
d(H1 ) = d(H2 ),
d(Z(H1 )) = d(Z(H2 )),
r(H1 ) = r(H2 ),
k(H1 ) = k(H2 ),
r(Z(H1 )) = r(Z(H2 )),
k(Z(H1 )) = k(Z(H2 )).
Lemma 5.1. Let H be a unipotent subgroup of AGLn (F) and assume that Fn is
the direct sum of non-trivial π(H)-invariant subspaces V1 , . . . , Vs . Then k(H) ≥ s.
In particular, if k(H) = 1, then H is indecomposable.
Proof. H induces on each Vi a unipotent group Hi . So, in each Vi , there exists a
non-zero vector wi fixed by all elements of Hi (see [10, Theorem 17.5 page 112]). It
follows that w1 , . . . , ws are s linearly independent vectors of Fn fixed by π(H).
Notation. For sake of brevity, we write
x1 x2 . . . xn
1
(11)
U=
0 τU (x1 , x2 , . . . , xn )
to indicate the regular subgroup
1 x1 x2 . . . xn
U=
: x1 , . . . , xn ∈ F .
0 τU (x1 , x2 , . . . , xn )
For all i ≤ n, we denote by Xi the matrix of U − In+1 obtained taking xi = 1 and
xj = 0 for all j 6= i.
The set {v1 , . . . , vn } is the canonical basis of Fn .
We recall that the center of a unipotent group is non-trivial.
Lemma 5.2. Let U be a unipotent regular subgroup of Gn . If d(Z(U )) = n + 1
then, up to conjugation, U = CGn (Jn+1 ). Moreover U is abelian.
Proof. Up to conjugation Jn+1 ∈ Z(U ), whence U ≤ CGn (Jn+1 ). Since this group
is regular, we have U = CGn (Jn+1 ). It follows that U is abelian.
Lemma 5.3. Let U be a regular subgroup of Gn such that δU is linear. If r(U ) = 1,
then U is the translation subgroup T , which is an abelian normal subgroup of Gn .
Furthermore, d(T ) = 2.
Proof. By the unipotency, we may always assume that U is upper unitriangular.
Then, the rank condition gives δ(vi ) = 0, for 1 ≤ i ≤ n, whence our claim.
Lemma 5.4. Let U be a regular subgroup of Gn , n ≥ 2, such that δU is linear. If
d(Z(U )) = n then, up to conjugation, for some fixed α ∈ F:
1 x1 x2 ... xn−2 xn−1 xn
0 1 x1 ... xn−3
0 1 ... xn−4
(12)
0
U = Rα = ...
0
0
0 0
0 0
..
.
0 ...
0 ...
0 ...
..
.
1
0
0
0
0
xn−2
xn−3
0
1
0
x1
αxn−1
1
..
.
..
.
.
10
M.A. PELLEGRINI AND M.C. TAMBURINI BELLANI
In particular U is abelian and r(U ) = n − 1. Furthermore, R0 and Rα are not
conjugate for any α 6= 0,
If n ≥ 3, an epimorphism Ψ : F[t1 , t2 ] → FIn+1 + R0 is obtained setting
(13)
Ψ(t1 ) = X1 ,
Ψ(t2 ) = Xn−1 .
In this case we have Ker (Ψ) = ht1 n , t1 2 , t1 t2 i.
If n ≥ 4 is even, then Rα is conjugate to R1 for any α 6= 0 and an epimorphism
Ψ : F[t1 , t2 ] → FIn+1 + Rα is obtained setting
(14)
Ψ(t1 ) = αX1 ,
Ψ(t2 ) = α
n−2
2
Xn−1 .
In this case Ker (Ψ) = ht1 n−1 − t2 2 , t1 t2 i.
If n is odd, write α 6= 0 as α = λε2 , λ, ε ∈ F∗ . Then Rα is conjugate to Rλ and
an epimorphism Ψ : F[t1 , t2 ] → FIn+1 + Rα is obtained setting
(15)
Ψ(t1 ) = αX1 ,
Ψ(t2 ) = λ
n−3
2
εn−2 Xn−1 .
In this case Ker (Ψ) = ht1 n−1 − λt2 2 , t1 t2 i. In particular Rα and Rβ (α, β ∈ F∗ )
are conjugate if and only if β/α is a square in F∗ .
Proof. Up to conjugation we may suppose that z
= diag(Jn , J1 ) ∈ Z(U ). The
X 0
subalgebra generated by z coincides with the set
: X ∈ CGn−1 (Jn ) .
0 1
This information gives the values of δ(vi ) for 1 ≤ i ≤ n − 1. From Lemma 4.1, we
get δ(vn ) = αEn,n−1 . Applying (3), it follows that U is abelian. Conjugating by
the permutation matrix associated to the transposition (n, n + 1) we obtain that U
is conjugate to Rα .
The subgroups R0 and Rα are not conjugate when α 6= 0, since k(R0 ) = 2 and
k(Rα ) = 1. The presentations of the corresponding algebras can be verified by
matrix calculation.
Assume now that n is odd. If α = βε2 6= 0, then the subgroups Rα and Rβ are
conjugate in virtue of (15) and Proposition 3.4. Conversely, suppose that α, β ∈ F∗
c n (F) (see Proposition 3.4). According to
and that Q−1 Rα Q = Rβ for some Q ∈ GL
(12), write
1
X
xn
1
Y
yn
Rα = 0 In−1 + DX AX T , Rβ = 0 In−1 + DY BY T ,
0
0
1
0
0
1
where X = (x1 , . . . , xn−1 ), Y = (y1 , . . . , yn−1 ), A = diag(antidiag(1, . . . , 1), α)
and B = diag(antidiag(1, . . . , 1), β). Hence det(A) = ζα and det(B) = ζβ, where
n+1
ζ = (−1) 2 . We may assume that vn Q = λvn since hvn i is the subspace fixed
pointwise by both subgroups. Thus, for some N ∈ Fn−1 and λ ∈ F∗ , Q = Q1 Q2 ,
where
1
0
0
1 0 0
Q1 = 0 In−1 λ−1 N , Q2 = 0 M 0 , det(M ) 6= 0, λ 6= 0.
0
0
1
0 0 λ
e
Now, Q−1
1 Rα Q1 = Rα , where
1
1
X
x̃n
eα = 0 In−1 + DX λ−1 DX N + AX T = 0
R
0
0
1
0
X
In−1 + DX
0
x̃n
e T ,
AX
1
MORE ON REGULAR SUBGROUPS OF THE AFFINE GROUP
11
e = ζα. From R
eα Q2 − Q2 Rβ we get, in particular, Y = XM and
with det(A)
e T , whence M BM T X T = λAX
e T , for all X ∈ Fn−1 . It follows that
M BY T = λAX
e and taking the determinant of both sides, we obtain ζβ(det(M ))2 =
M BM T = λA
n−1
ζλ
α. We conclude that β/α is a square.
6. Standard regular subgroups
The aim of this section is to define, for every partition λ of n + 1 different from
(1
), one or two abelian regular subgroups Sλ , Sλ♯ of AGLn (F) so that different
partitions define non-conjugate subgroups.
To this purpose we start by identifying the direct product AGLm1 (F)×AGLm2 (F)
with the stabilizer of Fm1 and Fm2 in AGLm1 +m2 (F), namely with the subgroup:
1 v w
0 A 0 : v ∈ Fm1 , w ∈ Fm2 , A ∈ GLm1 (F), B ∈ GLm2 (F) .
0 0 B
n+1
Clearly, in this identification, if Ui are respective regular subgroups of AGLmi (F)
for i = 1, 2 then U1 × U2 is a regular subgroup of AGLm1 +m2 (F).
Here, and in the rest of the paper, we denote by S(1+n) the centralizer in AGLn (F)
of a unipotent Jordan block of size n + 1, namely:
S(1+n) = CAGLn (F) (J1+n ).
Moreover we write τ1+n for τS(1+n) and δ1+n for δS(1+n) . By Lemma 5.2, a regular
subgroup of AGLn (F ) is conjugate to S(1+n) if and only if d(Z(S(1+n) )) = n + 1.
Observe that S(1+n) is abelian and indecomposable by Lemma 5.1. More generally,
for a partition λ of n + 1 such that:
(16)
λ = (1 + n1 , n2 , . . . , ns ),
s ≥ 1,
ni ≥ n1+i ≥ 1,
1 ≤ i ≤ s − 1,
we define an abelian regular subgroup Sλ of AGLn (F). Namely we set:
Sλ = S(1+n1 ,...,ns ) =
s
Y
S(1+nj ) .
j=1
Notice that Sλ is indecomposable only for s = 1. In particular, if λ = (2, 1n−1 ),
then Sλ = T .
Lemma 6.1. Given a partition (1 + n1 , n2 ) of n + 1, with 1 + n1 ≥ n2 > 1, set:
u
v
1
♯
0 τ1+n1 (u) w ⊗ DuT : u ∈ Fn1 , v ∈ Fn2 ,
S(1+n
=
,n
)
1
2
0
0
τn2 (v)
♯
where w = (0, . . . , 0, 1) ∈ Fn2 , D = antidiag(1, . . . , 1) ∈ GLn1 (F). Then S(1+n
1 ,n2 )
♯
is an indecomposable regular subgroup of AGLn (F) with d(S(1+n
) = n1 + 2.
1 ,n2 )
♯
Proof. Routine calculation with matrices shows that S(1+n
is closed under mul1 ,n2 )
tiplication, hence a subgroup. Moreover it is indecomposable by Lemma 5.1. Again
♯
by matrix calculation one can see that d(S(1+n
) = n1 + 2. To check that
1 ,n2 )
♯
S(1+n
is a subgroup it is useful to note that:
1 ,n2 )
(a) all components of u(w ⊗ DuT ), except possibly the last one, are zero;
12
M.A. PELLEGRINI AND M.C. TAMBURINI BELLANI
(b) δn2 (v) does not depend on the last component of v;
(c) (w ⊗ DuT )δn2 (v) = 0;
n1
(d) w ⊗ D(u1 δ1+n1 (u2 ))T = δ1+n1 (u1 )(w ⊗ DuT
2 ) for all u1 , u2 ∈ F .
Next, consider a partition µ of n + 1 such that:
(17) µ = (1+n1 , n2 , . . . , ns ), s ≥ 2, 1+n1 ≥ n2 > 1, ni ≥ n1+i ≥ 1, 2 ≤ i ≤ s−1.
We define the abelian regular subgroup Sµ♯ in the following way:
Sµ♯
=
♯
S(1+n
1 ,n2 ,n3 ...,ns )
=
♯
S(1+n
1 ,n2 )
×
s
Y
S(1+nj ) .
j=3
The regular subgroups Sλ , Sµ♯ associated to partitions as above will be called
standard regular subgroups. As already mentioned Sλ and Sµ♯ are always abelian.
Remark 6.2. Let λ = (1 + n1 , . . . , ns ) be a partition as in (16). Then Fn is a direct
sum of s indecomposable modules of respective dimensions ni for which d = ni + 1.
Furthermore,
F[t1 , t2 , . . . , ts ]
(18)
Lλ = FIn+1 + Sλ ∼
.
= ni +1
hti
: 1 ≤ i ≤ s; ti tj : 1 ≤ i < j ≤ si
An epimorphism Ψ : F[t1 , t2 , . . . , ts ] → Lλ is obtained by setting
(19)
Ψ(ti ) = Xi .
Let now µ = (1 + n1 , . . . , ns ) be a partition as in (17). Then Fn is a direct
sum of s − 2 indecomposable modules of respective dimension ni , i ≥ 3, for which
d = ni + 1, and a single indecomposable module of dimension n1 + n2 for which
d = n1 + 2. Furthermore,
F[t1 , t2 , . . . , ts ]
.
(20) L♯µ = FIn+1 + Sµ♯ ∼
= n1 +1
n2
ni +1
ht1
− t2 ; ti
: 3 ≤ i ≤ s; ti tj : 1 ≤ i < j ≤ si
An epimorphism Ψ : F[t1 , t2 , . . . , ts ] → L♯µ is obtained by setting
(21)
Ψ(ti ) = Xi .
Theorem 6.3. Let λ1 = (1 + n1 , . . . , ns ), λ2 = (1 + m1 , . . . , mt ) be two partitions of
n + 1 as in (16) and let µ1 = (1 + a1 , . . . , ah ), µ2 = (1 + b2, . . . , bk ) be two partitions
of n + 1 as in (17). Then
(a) Sλ1 is not conjugate to Sµ♯ 1 ;
(b) Sλ1 is conjugate in AGLn+1 (F) to Sλ2 if and only if λ1 = λ2 ;
(c) Sµ♯ 1 is conjugate in AGLn+1 (F) to Sµ♯ 2 if and only if µ1 = µ2 .
Proof. (a) There exists an isomorphism ϕ : Lλ1 → L♯µ1 . Setting vu := vϕ(u)
for all u ∈ Lλ1 , v ∈ J(L♯µ1 ) we may consider J(L♯µ1 ) as an Lλ1 -module. Clearly
J(Lλ1 ) ∼
= Fn as Lλ1 -modules. By construction, Fn is a direct sum of
= J(L♯µ1 ) ∼
indecomposable Lλ1 -modules. By Remark 6.2, for J(Lλ1 ) each direct summand
has dimension ni and d = ni + 1 and for J(L♯µ1 ) there is a direct summand of
dimension a1 + a2 and d = a1 + 2 < a1 + a2 + 1. Hence, by Krull-Schmidt Theorem
(e.g., see [11, page 115]), Sλ1 is not conjugate to Sµ♯ 1 .
(b) Arguing as before, we may consider J(Lλ2 ) as an Lλ1 -module. In this case,
by construction, Fn is a direct sum of indecomposable Lλ1 -modules of dimension
MORE ON REGULAR SUBGROUPS OF THE AFFINE GROUP
13
ni such that d = ni + 1 (see Remark 6.2(1)). By Krull-Schmidt Theorem they are
conjugate if and only if s = t and ni = mi .
(c) Again, we may consider J(L♯µ2 ) as an L♯µ1 -module. In this case, by construction, Fn is a direct sum of indecomposable L♯µ1 -modules, one of dimension a1 + a2
such that d = a1 + 2 < a1 + a2 + 1, the others of dimension ni such that d = ai + 1
(see Remark 6.2(2)). By Krull-Schmidt Theorem they are conjugate if and only if
h = k and ai = bi .
Remark 6.4. Note that for n ≥ 3, the regular subgroups R0 and R1 of Lemma 5.4
♯
coincide, respectively, with S(n,1) and S(n−1,2)
.
7. Regular subgroups with linear δ
By [14, Lemma 5.1],
the only regular subgroup of AGL1 (F) is the translation
x0 x1
. However, already for n = 2, a description
subgroup T = S(2) =
0 x0
becomes much more complicated. As seen in Example 2.5 there are 2|R| conjugacy
classes of regular subgroups of AGL2 (R) with trivial center. Also, restricting to
the unipotent case, one may only say that every unipotent regular subgroup is
conjugate to:
1 x1
x2
0 1 σ(x1 )
0 0
1
where σ ∈ HomZ (F, F) ([14, Lemma 5.1]). See also Examples 2.6 and 2.7.
Thus, from now on, we restrict our attention to regular subgroups U of AGLn (F)
such that the map δU defined in (1) is linear. U is unipotent by Theorem 2.4 and
so there exists 1 6= z ∈ Z(U ). By Theorem 4.4, up to conjugation, we may assume
that z is a Jordan form. For the notation we refer to Section 5. Our classification
is obtained working on the parameters d, r and k, considered in this order.
In the tables of the next subsections the indecomposability of the regular subgroups follows from Lemma 5.1, since k(U ) = 1, except for the subgroup U14 of
Table 3, for which we refer to Lemma 7.4. Also, we describe the kernel of the
epimorphism
Ψ : F[t1 , . . . , ts ] → FIn+1 + U.
7.1. Case n = 2. If d(Z(U )) = 3, then U is abelian and conjugate to S(3) by
Lemma 5.2. If d(Z(U )) = 2, then U is abelian and r(U ) = 1 by Lemma 5.4. Thus
U is conjugate S(2,1) by Lemma 5.3. Table 1 summarizes these results.
U
FIn+1 + U
Ψ Ker (Ψ)
x0 x1 x2
ht1 3 i
indec.
S(3) 0 x0 x1 (19)
0
0 x0
x0 x1 x2
S(2,1) 0 x0 0 (19) ht1 , t2 i2
0
0 x0
Table 1. Representatives for the conjugacy classes of regular subgroups U of AGL2 (F) with linear δ, for any field F.
14
M.A. PELLEGRINI AND M.C. TAMBURINI BELLANI
7.2. Case n = 3. To obtain the full classification, we need some preliminary results.
Lemma 7.1. If d(Z(U )) = r(Z(U )) = 2, then U is abelian, conjugate to
1 x1 x2 x3
0 1
0 x2
U13 =
0 0
1 x1
0 0
0
1
and char F = 2. An epimorphism Ψ : F[t1 , t2 ] → FI4 + U is obtained by setting
(22)
Ψ(t1 ) = X1 ,
2
Ψ(t2 ) = X2 .
2
In this case, we have Ker (Ψ) = ht1 , t2 i.
Proof. We may assume z = diag(J2 , J2 ) ∈ Z(U ), whence δ(v1 ) = E2,3 . From
Lemma 4.1 and the unipotency of U we obtain δ(v2 ) = E1,3 + αE2,1 + βE2,3 . It
follows v1 δ(v2 ) = v3 . Now, we apply (2) to v1 , v2 , which gives δ(v3 ) = δ(v1 )δ(v2 ) =
0. Direct calculation shows that U is abelian. Hence d(Z(U )) = d(U ) = 2. In
particular, (µ(v2 ) − I4 )2 = 0 gives α = β = 0. Finally (µ(v1 + v2 ) − I4 )2 = 0 gives
char F = 2.
It is convenient to denote by V (α2 , α3 , β2 , β3 ) the regular subgroup defined by:
1
x1
x2 x3
0
1
0
0
.
(23)
V (α2 , α3 , β2 , β3 ) =
0 α2 x2 + α3 x3 1
0
0 β2 x2 + β3 x3 0
1
Notice that V (α2 , α3 , β2 , β3 ) is abelian if and only if α3 = β2 .
We need the cosquare A− T A of a nonsingular matrix A. If A, B are congruent,
i.e., B = P AP T for a nonsingular P , their cosquares are conjugate (e.g., see [9]).
Lemma 7.2. Let β, γ ∈ F with βγ 6= 0. The subgroups Vβ = V (1, 1, 0, β) and
Vγ = V (1, 1, 0, γ) are conjugate in AGL3 (F) if and only if β = γ.
c 3 (F) (see Proposition 3.4).
Proof. Suppose that Q−1 Vβ Q = Vγ for some Q ∈ GL
We may assume that v1 Q = λv1 since hv1 i is the subspace of F3 fixed pointwise by
both subgroups. Thus Q has shape:
1 0
0
0
0 λ
0
0
p1,1 p1,2
Q=
, λ 6= 0, P =
nonsingular.
0 q1 p1,1 p1,2
p2,1 p2,2
0 q2 p2,1 p2,2
q1
λ E3,2
−
q2
λ E4,2
normalizes V
β . Hence,
substituting
Qwith
1 1
1 1
KQ, we may suppose q1 = q2 = 0. Setting B =
and C =
we
0 β
0 γ
have:
1
x1
X
1
y1
Y
Vβ = 0
1
0 X = (x2 , x3 ), Vγ = 0
1
0 Y = (y2 , y3 ).
0 BX T I2
0 CY T I2
The matrix K = I4 −
From Vβ Q = QVγ we get Y = XP and λBX T = P CY T , whence
λBX T = P CP T X T ,
for all X ∈ F2 .
MORE ON REGULAR SUBGROUPS OF THE AFFINE GROUP
15
It follows that λB = P CP T , i.e., the matrices λB and C are congruent. So their
cosquares must be conjugate. But the characteristic polynomials of the cosquares
are respectively t2 + (β −1 − 2)t + 1 and t2 + (γ −1 − 2)t + 1, whence β = γ.
Lemma 7.3. Suppose d(Z(U )) = 2 and r(Z(U )) = 1. If U is abelian, then U =
S(2,12 ) . Otherwise, U is conjugate to exactly one of the following subgroups:
1 x1 x2 x3
0 1
0
0
, if k(U ) = 2;
(a) N1 =
0 0
1 x1
0
1
0 0
1 x1 x2 x3
0 1
0 −x2
, if k = 1, d(U ) = 2 and char F 6= 2;
(b) N2 =
0 0
1
x1
0 0
0
1
1 x1 x2
x3
0 1
0 x1 + x2
, λ ∈ F∗ , if k = 1 and d(U ) = 3.
(c) N3,λ =
0 0
1
λx2
0 0
0
1
The algebras L1 = FI4 + N1 , L2 = FI4 + N2 and L3,λ = FI4 + N3,λ have the
following presentation:
L1 = SpanF (t1 , t2 ),
L2 = SpanF (t1 , t2 ),
L3,λ = SpanF (t1 , t2 ),
where
where
where
t1 2 = t2 2 = t1 t2 = 0;
t1 2 = t2 2 = t1 t2 + t2 t1 = 0;
t2 2 − λt1 2 = t2 t1 = t1 2 − t1 t2 = 0,
λ ∈ F∗ .
Proof. If U is abelian, then r(U ) = r(Z(U )) = 1 and by Lemma 5.3, U = S(2,12 ) .
So, suppose that U is not abelian. We may assume z = diag(J2 , J1 , J1 ) ∈ Z(U ).
Now z = µ(v1 ) gives δ(v1 ) = 0 and,
(3), we get that the first row of δ(v) is
from
zero for all v ∈ F3 . Hence δ(v) =
0 0 0
α y1 y2
β y3 −y1
with y12 + y2 y3 = 0, by the unipotency
of U . Now fix v = (0, x2 , x3 ) 6= 0 in hv2 , v3 i. Conjugating by g = diag(I2 , P ) with
a suitable P ∈ GL2 (F), we may assume either (i) y1 = y2 = 0 and y3 = 1 or (ii)
y1 = y2 = y3 = 0.
0 00
In case (i) if x3 = 0, we may suppose v = v2 . So δ(v2 ) = α2 0 0 and δ(v3 ) =
β2 1 0
0 0 0
α3 γ1 γ2
. Applying (2) to v2 , v2 and to v3 , v2 we obtain respectively α2 = 0
β3 γ3 −γ1
and γ1 = −1, γ2 = 0, which contradicts
the unipotency
of U . On the other hand,
−1
0
x
we may assume v = v3 , hence
if x3 6= 0, conjugating by diag I2 , −x 3x−2 x−1
2 3
3
0 0 0
0 0 0
δ(v3 ) = α3 0 0 . Applying (2) to v3 , v3 we obtain δ(v2 ) = δ(v3 )2 = 0 0 0 .
α3 0 0
β3 1 0
However, in this case U is abelian.
In case (ii), up to a further conjugation
shape of g, we
0 0 0by
a matrix of the
0same
0 0
α
γ
γ
3
1
2
may suppose v = v2 , hence δ(v2 ) = α2 0 0 . Set δ(v3 ) =
. Now, (2)
β2 0 0
β3 γ3 −γ1
applied to v2 , v3 gives γ1 δ(v2 ) + γ2 δ(v3 ) = δ(v2 )δ(v3 ) = 0. In particular, γ22 = 0,
whence γ2 = 0. It follows γ1 = 0 by the condition γ12 + γ2 γ3 = 0. Replacing, if
necessary, v3 by a scalar multiple, we get either γ3 = 1 or γ3 = 0. In the first case,
(2) applied to v3 , v3 gives α3 = β2 , α2 = 0 and U is abelian. In the second one, U
is conjugate V (α2 , α3 , β2 , β3 ) with α3 6= β2 , from the non-abelianity.
16
M.A. PELLEGRINI AND M.C. TAMBURINI BELLANI
Let ∆ = α2 β3 − α3 β2 , with α3 6= β2 . If ∆ = 0, then k(U ) = 2. An isomorphism
Ψ : L1 → L = FI4 + V (α2 , α3 , β2 , β3 ) is obtained by setting
Ψ(t1 ) = β2 X2 − α2 X3
if α2 6= 0 take
;
Ψ(t2 ) = α3 X2 − α2 X3
Ψ(t1 ) = − αβ33 X2 + X3
;
if α2 = β2 = 0 take
Ψ(t2 ) = X2
Ψ(t1 ) = X2
if α2 = α3 = 0 take
.
Ψ(t2 ) = − ββ32 X2 + X3
By Proposition 3.4, the subgroup U is conjugate to N1 .
Suppose now that ∆ 6= 0 (which implies k(U ) = 1). If d(U ) = 2, then char F =
6
2, α2 = β3 = 0 and β2 = −α3 6= 0. An isomorphism Ψ : L2 → L = FI4 +
V (α2 , α3 , β2 , β3 ) is obtained by setting
Ψ(t1 ) = X2 ,
Ψ(t2 ) = X3 .
By Proposition 3.4, the subgroup U is conjugate to N2 .
If d(U ) = 3, an isomorphism Ψ : L3,λ → L = FI4 + V (α2 , α3 , β2 , β3 ) is obtained
2 β3 −α3 β2
by setting λ = α(α
and
2
3 −β2 )
if β3 6= 0 take
if β3 = 0, α2 6= 0 take
if α2 = β3 = 0 (α3 6= −β2 ) take
(
Ψ(t1 ) = X3
3
X2 +
Ψ(t2 ) = − α3β−β
2
Ψ(t1 ) = X2
2
Ψ(t2 ) = − α3β−β
X2 +
2
α3
α3 −β2 X3
α2
α3 −β2 X3
;
;
2
X3
Ψ(t1 ) = X2 + α3α−β
3
.
β2
Ψ(t2 ) = − α3 −β2 X2 + X3
By Proposition 3.4, the subgroup V (α2 , α3 , β2 , β3 ) is conjugate to N3,λ , λ 6= 0. The
statement now follows from Lemma 7.2
We can now classify the regular subgroups U of AGL3 (F) having linear δ, including the non-abelian ones, arising from d(Z(U )) = 2 and r(Z(U )) = 1.
If d(Z(U )) = 4, then U is conjugate to S(4) , by Lemma 5.2. If d(Z(U )) = 3, then
U is one of the subgroups Rα described in Lemma 5.4: namely, for k(U ) = 2, U is
conjugate to R0 and, for k(U ) = 1, U is conjugate to Rλ , where λ can be chosen in
a transversal F of (F∗ )2 in F∗ . As observed in Remark 6.4, R0 and R1 coincide,
♯
respectively, with S(3,1) and S(2,2)
. If d(Z(U )) = 2 we have two possibilities. When
r(Z(U )) = 2 we apply Lemma 7.1, that gives char F = 2 and U is conjugate to U13 .
When r(Z(U )) = 1 we apply Lemma 7.3: when abelian U = S(2,12 ) , otherwise U is
conjugate to one of the subgroups N1 , N2 , N3,λ , λ ∈ F∗ .
A complete set of representatives of the abelian regular subgroups of AGL3 (F)
is given in Table 2. The conjugacy classes of non-abelian regular subgroups of
AGL3 (F) are described in Lemma 7.3.
7.3. Case n = 4. Once again, to obtain the full classification of the abelian regular
subgroups of AGL4 (F) we need some preliminary results.
MORE ON REGULAR SUBGROUPS OF THE AFFINE GROUP
U
17
FI4 + U
char F Ψ
Ker (Ψ)
x0 x1 x2 x3
0 x0 x1 x2
any
(19)
ht1 4 i
indec.
S(4)
0
0 x0 x1
0
0
0 x0
x0 x1 x2 x3
0 x0 x1 0
S(3,1)
any
(13)
ht1 3 , t2 2 , t1 t2 i
0
0 x0 0
0
0
0 x0
x0 x1 x2 x3
0 x0 0
x1
Rλ , λ ∈ F
any
(15) ht1 2 − λt2 2 , t1 t2 i indec.
♯
0
)
(R1 =S(2,2)
0 x0 λx2
0
0
0
x0
x0 x1 x2 x3
0 x0 0 x2
U13
2
(22)
ht1 2 , t2 2 i
indec.
0
0 x0 x1
0
0
0 x0
x0 x1 x2 x3
0 x0 0
0
S(2,12 )
any
(19)
ht1 , t2 , t3 i2
0
0 x0 0
0
0
0 x0
Table 2. Representatives for the conjugacy classes of abelian regular subgroups U of AGL3 (F), for any field F.
Lemma 7.4. Let U be a regular subgroup of AGL4 (F) such that δ is linear. If
d(Z(U )) = r(Z(U )) = 3, then U is abelian and is conjugate to
1 x1 x2 x3
x4
0 1
x1
0
x3
,
1
0
0
R(α, β) = 0 0
α, β ∈ F.
0 0 βx3 1 x1 + αx3
0 0
0
0
1
Furthermore, if F has no quadratic extensions, there are exactly two conjugacy
classes of such subgroups, whose representatives are, for instance, R(0, 0) and
R(1, 0), which is conjugate to S(3,2) . Finally, U14 = R(0, 0) is indecomposable.
Proof. We may suppose that z = diag(J3 , J2 ) ∈ Z(U ). From z, z 2 ∈ U we obtain
δ(v1 ) = E1,2 + E3,4 and δ(v2 ) = 0. By Lemma 4.1 an the unipotency of U it follows
that δ(v3 ) = E1,4 + αE3,4 + βE3,2 + γ(E3,1 + E4,2 ). Applying (2) to v3 , v1 we obtain
δ(v4 ) = δ(v3 )δ(v1 ) = γE3,2 . In particular U is abelian. Now (µ(v3 ) − I5 )3 = 0 gives
γ = 0. We conclude that U is conjugate to R(α, β).
Now, assume that F has no quadratic extensions. An epimorphism Ψ : F[t1 , t2 ] →
FIn+1 + R(α, β) is obtained in the following way. If α2 + 4β 6= 0, take
p
p
(24) Ψ(t1 ) = (α + α2 + 4β)X1 − 2X3 , Ψ(t2 ) = (α − α2 + 4β)X1 − 2X3
when char F 6= 2 and take
(25)
Ψ(t1 ) = rX1 + X3 ,
Ψ(t2 ) = (r + α)X1 + X3 ,
18
M.A. PELLEGRINI AND M.C. TAMBURINI BELLANI
when char F = 2 (here r ∈ F is such that r2 + αr + β = 0). In both cases,
Ker (Ψ) = ht1 3 , t2 3 , t1 t2 i. Comparison with the presentation of FI5 + S(3,2) given
in (18) shows that R(α, β) is conjugate to S(3,2) by Proposition 3.4.
If α2 + 4β = 0, take
(26)
Ψ(t1 ) = X1 ,
when char F 6= 2 and take
(27)
Ψ(t1 ) = X1 ,
Ψ(t2 ) = αX1 − 2X3
Ψ(t2 ) =
p
βX1 + X3
when char F = 2. In both cases Ker (Ψ) = ht1 3 , t1 2 t2 , t2 2 i.
The algebras defined by the two presentations above are not isomorphic, as the
subspaces consisting of elements whose square is zero (namely ht1 2 , t2 2 i in the first
case, ht1 2 , t2 , t1 t2 i in the second case) have different dimensions. By Proposition
3.4, there are exactly two conjugacy classes of subgroups R(α, β), depending on the
nullity of α2 + 4β.
Noting that k(U14 ) = 2, direct computation shows that U14 is indecomposable.
Lemma 7.5. Let U be an abelian regular subgroup of AGL4 (F). If d(U ) = 3 and
r(U ) = 2, then U is conjugate to
1 x1
x2
x3 x4
0 1
x1
0
0
.
0
0
1
0
0
R(α, β, γ) =
0 0 αx3 + βx4 1
0
0 0 βx3 + γx4 0
1
If β 2 − αγ 6= 0, then k(U ) = 1. If β 2 − αγ = 0 with (α, β, γ) 6= (0, 0, 0), then
k(U ) = 2 . Finally, if α = β = γ = 0, then k(U ) = 3. Furthermore, assuming that
every element of F is a square, there are exactly three conjugacy classes of such
subgroups. Their representatives are, for instance,
U24 = R(0, 0, 1) (if k = 1),
R(0, 1, 0) (if k = 2),
R(0, 0, 0) = S(3,1,1) (if k = 3).
♯
Observe that R(0, 1, 0) is conjugate to S(2,2,1)
.
Proof. We may assume z = diag(J3 , J1 , J1 ) ∈ U . From z, z 2 ∈ U we obtain δ(v1 ) =
E1,2 and δ(v2 ) = 0. By Lemma 4.1 and the unipotency of U , for any v ∈ hv3 , v4 i
0 0 0 0
we have δ(v) = 00 α0 0ξ η0 , where ξ 2 + ηǫ = 0. Now, rk(µ(v1 + v3 + v4 ) − I5 ) = 3
0 β ǫ −ξ
implies δ(v3 ) = α3 E3,2 + β3 E4,2 and δ(v4 ) = α4 E3,2 + β4 E4,2 . Since U is abelian,
(3) applied to v3 , v4 implies α4 = β3 . Hence U is conjugate to R(α, β, γ).
Suppose first that ∆ = β 2 − αγ 6= 0. In this case k(U ) = 1. If char F 6= 2 define
(28)
if α 6= 0
Ψ(t1 ) = X1√
,
β+ ∆
Ψ(t2 ) =
2∆ X3 −
Ψ(t3 )
=
√
−β+ ∆
X3
α
and if char F = 2, define
(29)
Ψ(t1 )
Ψ(t2 )
Ψ(t3 )
if α = 0
α
2∆ X4 ,
+ X4 ,
Ψ(t1 ) =
Ψ(t2 ) =
Ψ(t3 ) =
√
√
γ
= 4 ∆X1 + √
X +
4
∆ 3
√
αX1 + X3 ,
=
√
γX1 + X4 .
=
βX1 ,
− γ2 X3 + βX4 ,
X3 ,
√
α
√
X ,
4
∆ 4
MORE ON REGULAR SUBGROUPS OF THE AFFINE GROUP
19
We get Ker (Ψ) = ht1 2 − t2 t3 , t2 2 , t3 2 , t1 t2 , t1 t3 i.
Next suppose β 2 − αγ = 0 and (α, β, γ) 6= (0, 0, 0). Then k(U ) = 2. Define
(30)
if α 6= 0
Ψ(t1 ) =
Ψ(t2 ) =
Ψ(t3 ) =
X1 ,√
− β √α+1
X3 + αX4 ,
α
−βX3 + αX4 ,
if α = 0 and γ 6= 0
√
γX1 ,
Ψ(t1 ) =
Ψ(t2 ) =
X4 ,
Ψ(t3 ) =
X3 .
In both cases, Ker (Ψ) = ht1 2 − t2 2 , t3 2 , t1 t2 , t1 t3 , t2 t3 i. Comparison with the pre♯
♯
sentation of FI5 + S(2,2,1)
given in (20) shows that U is conjugate to S(2,2,1)
.
Finally, if α = β = γ = 0, then R(0, 0, 0) = S(3,1,1) and k(U ) = 3.
Remark 7.6. Consider the algebra L of Example 3.2 and the corresponding regular
subgroup R = R(−1, 0, −1) of the previous Lemma. Since k(R) = 1, under the
assumption that −1 is a square in F, we obtain that the algebras
F[t1 , t2 , t3 ]
F[t1 , t2 , t3 ]
and
ht1 2 + t2 2 , t1 2 + t3 2 , t1 t2 , t1 t3 , t2 t3 i
ht1 2 − t2 t3 , t2 2 , t3 2 , t1 t2 , t1 t3 i
are isomorphic in any characteristic. Actually, if char F = 2, an isomorphism can
also be obtained directly via the change of variables t′1 = t1 + t2 + t3 , t′2 = t2 + t3 ,
t′3 = t1 + t3 . This fixes an inaccuracy of [12], corrected in [13].
Lemma 7.7. Let U be an abelian regular subgroup of AGL4 (F). If d(U ) = r(U ) =
2, then char F = 2 and U is conjugate to U13 × S(1) , where U13 is defined in Table 2.
Proof. We may suppose z = diag(J2 , J2 , J1 ) ∈ U . From z ∈ U we get δ(v1 ) = E2,3 .
Lemma 4.1 and the condition (µ(v2 ) − I5 )2 = 0 give δ(v2 ) = E1,3 + α2 E4,3 . Now,
(µ(v1 + v2 ) − I5 )2 = 0 implies char F = 2. Applying (2) to v2 , v1 we get δ(v3 ) =
δ(v2 )δ(v1 ) = 0. From (µ(v2 +v4 )−I5 )2 = 0 we obtain δ(v4 ) = (α2 +α4 )E2,3 +α4 E4,3 ,
but (µ(v4 ) − I5 )2 = 0 gives α4 = 0. In conclusion, U is conjugate to
1 x1 x2
x3
x4
0 1
0
x2
0
0
0
1
x
+
αx
0
R(α) =
1
4
.
0 0
0
1
0
0 0
0
αx2
1
An epimorphism Ψ : F[t1 , t2 , t3 ] → FIn+1 + R(α) is obtained by setting
(31)
Ψ(t1 ) = X1 ,
Ψ(t2 ) = X2 ,
Ψ(t3 ) = αX1 + X4 .
We have Ker (Ψ) = ht1 2 , t2 2 , t3 2 , t1 t3 , t2 t3 i. Considering the presentation of U13
given in Table 2, we have that U is conjugate to U13 × S(1) .
We now classify the abelian regular subgroups U of AGL4 (F). If d(U ) = 5,
then U is conjugate to S(5) by Lemma 5.2. If d(U ) = 4, then by, Lemma 5.4, U
♯
is conjugate to R0 = S(4,1) when k(U ) = 2, and to R1 = S(3,2)
when k(U ) = 1.
Suppose d(U ) = 3. If r(U ) = 3, then U is conjugate either to U14 or to S(3,2) by
♯
or to S(3,1,1)
Lemma 7.4. If r(U ) = 2, then U is conjugate either to U24 or to S(2,2,1)
by Lemma 7.5. Finally suppose d(U ) = 2. If r(U ) = 1 then U is conjugate to
S(2,13 ) by Lemma 5.3. If r(U ) = 2, then char F = 2 and U is conjugate to U13 × S(1)
by Lemma 7.7.
When F has no quadratic extensions, a complete set of representatives of the
conjugacy classes of abelian regular subgroups is given in Tables 3 and 4.
20
M.A. PELLEGRINI AND M.C. TAMBURINI BELLANI
U
S(5)
♯
S(3,2)
U14
U24
FI5 + U
char F
Ψ
Ker (Ψ).
x0 x1 x2 x3 x4
0 x0 x1 x2 x3
0
0 x0 x1 x2
any
(19)
ht1 5 i
0
0
0 x0 x1
0
0
0
0 x0
x0 x1 x2 x3 x4
0 x0 x1 0 x2
0
0 x0 0 x1
any
(21)
ht1 3 − t2 2 , t1 t2 i
0
0
0 x0 x3
0
0
0
0 x0
x0 x1 x2 x3 x4
0 x0 x1 0 x3
6= 2
(24)
0
0
x
0
0
ht1 3 , t1 2 t2 , t2 2 i
0
0
0
0 x0 x1
2
(25)
0
0
0
0 x0
x0 x1 x2 x3 x4
0 x0 0
6= 2
(28)
0 x1
0
0 x0 0 x3
ht1 2 − t2 t3 , t2 2 , t3 2 , t1 t2 , t1 t3 i
0
0
0 x0 x2
2
(29)
0
0
0
0 x0
Table 3. Representatives for the conjugacy classes of indecomposable abelian regular subgroups U of AGL4 (F), when F has no
quadratic extensions.
Acknowledgments. We are grateful to Marco Degiovanni for a very useful suggestion, which led to Example 2.5.
References
[1] A. Caranti, F. Dalla Volta and M. Sala, Abelian regular subgroups of the affine group and
radical rings, Publ. Math. Debrecen 69 (2006), no. 3, 297–308.
[2] F. Catino, I. Colazzo and P. Stefanelli, On regular subgroups of the affine group, Bull. Aust.
Math. Soc. 91 (2015), no. 1, 76–85.
[3] F. Catino, I. Colazzo and P. Stefanelli, Regular subgroups of the affine group and asymmetric
product of radical braces, preprint.
[4] F. Catino and R. Rizzo, Regular subgroups of affine group and radical circle algebras, Bull.
Aust. Math. Soc. 79 (2009), 103–107.
[5] L.N. Childs, Elementary abelian Hopf Galois structures and polynomial formal groups, J.
Algebra 283 (2005), 292–316.
[6] L.N. Childs, On abelian Hopf Galois structures and finite commutative nilpotent rings, New
York J. Math. 21 (2015), 205–229.
[7] W. De Graaf, Classification of nilpotent associative algebras of small dimension,
arXiv:1009.5339, 27 Sep 2010.
[8] P. Hegedűs, Regular subgroups of the affine group, J. Algebra 225 (2000), no. 2, 740–742.
[9] R.A. Horn and V.V. Sergeichuk, Canonical forms for complex matrix congruence and
∗ congruence, Linear Algebra Appl. 416 (2006), no. 2-3, 1010–1032.
[10] J.E. Humphreys, Linear algebraic groups, Graduate Texts in Mathematics, No. 21. SpringerVerlag, New York-Heidelberg, 1975.
[11] N. Jacobson, Basic Algebra II, W.H. Freeman and Company, New York, (1989).
MORE ON REGULAR SUBGROUPS OF THE AFFINE GROUP
21
U
FI5 + U
char F Ψ
Ker (Ψ)
x0 x1 x2 x3 x4
0 x0 x1 x2 0
0
0 x0 x1 0
S(4,1)
any
(19)
ht1 4 , t2 2 , t1 t2 i
0
0
0 x0 0
0
0
0
0 x0
x0 x1 x2 x3 x4
0 x0 x1 0
0
0 x0 0
0
S(3,2)
any
(21)
ht1 3 , t2 3 , t1 t2 i
0
0
0
0 x0 x3
0
0
0
0 x0
x0 x1 x2 x3 x4
0 x0 x1 0
0
0
0
x
0
0
S(3,1,1)
any
(19) ht1 3 , t2 2 , t3 2 , t1 t2 , t1 t3 , t2 t3 i
0
0
0
0 x0 0
0
0
0
0 x0
x0 x1 x2 x3 x4
0 x0 0 x1 0
♯
0
0 x0 x2 0
S(2,2,1)
any
(21) ht1 2 − t2 2 , t3 2 , t1 t2 , t1 t3 , t2 t3 i
0
0
0 x0 0
0
0
0
0 x0
x0 x1 x2 x3 x4
0 x0 0 x2 0
3
0 x0 x1 0
U1 × S(1)
2
(31)
ht1 2 , t2 2 , t3 2 , t1 t3 , t2 t3 i
0
0
0
0 x0 0
0
0
0
0 x0
x0 x1 x2 x3 x4
0 x0 0
0
0
0 x0 0
0
S(2,13 )
any
(19)
ht1 , t2 , t3 , t4 i2
0
0
0
0 x0 0
0
0
0
0 x0
Table 4. Representatives for the conjugacy classes of decomposable abelian regular subgroups U of AGL4 (F), when F has no quadratic extension.
[12] B. Poonen, Isomorphism types of commutative algebras of finite rank over an algebraically
closed field, in Computational arithmetic geometry, 111–120, Contemp. Math., 463, Amer.
Math. Soc., Providence, RI, 2008.
[13] http://www-math.mit.edu/∼poonen/papers/dimension6.pdf, 11 January 2016.
[14] M.C. Tamburini Bellani, Some remarks on regular subgroups of the affine group, Int. J.
Group Theory 1 (2012), no. 1, 17–23.
Dipartimento di Matematica e Fisica, Università Cattolica del Sacro Cuore, Via
Musei 41, I-25121 Brescia, Italy
E-mail address: [email protected]
Dipartimento di Matematica e Fisica, Università Cattolica del Sacro Cuore, Via
Musei 41, I-25121 Brescia, Italy
E-mail address: [email protected]
| 4math.GR
|
Automated Tumor Segmentation and Brain Mapping for
the Tumor Area
Pranay Manocha*1, Snehal Bhasme*1, Tanvi Gupta2, B.K Panigrahi2,
Tapan K. Gandhi2
1
Indian Institute of Technology- Guwahati, India 781039
{m.pranay, b.walmik}@iitg.ernet.in
2 Indian Institute of Technology-Delhi, India 110016
{[email protected], [email protected],
[email protected]}
Abstract. Magnetic Resonance Imaging (MRI) is an important diagnostic
tool for precise detection of various pathologies. Magnetic Resonance (MR) is
more preferred than Computed Tomography (CT) due to the high resolution in
MR images which help in better detection of neurological conditions. Graphical
user interface (GUI) aided disease detection has become increasingly useful due
to the increasing workload of doctors. In this proposed work, a novel two steps
GUI technique for brain tumor segmentation as well as Brodmann area detection of the segmented tumor is proposed. A data set of T2 weighted images of
15 patients is used for validating the proposed method. The patient data incorporates variations in ethnicities, gender (male and female) and age (25-50), thus
enhancing the authenticity of the proposed method.
The tumors were segmented using Fuzzy C Means Clustering and Brodmann
area detection was done using a known template, mapping each area to the
segmented tumor image. The proposed method was found to be fairly accurate
and robust in detecting tumor.
Keywords: MRI, T2 Weighted, Fuzzy C Means, Brodmann Area, Tumor Segmentation.
1
Introduction
Diagnosing neurological disorders is a challenging domain. This is because of highly
complex anatomical structure of the brain along with some other factors like variations in size, age, previous history, gender, ethnicity and most importantly variations
of physiology and also pathology. This requires a comprehensive clinical diagnosis of
the patient which may require multiple sequences to be run. Manual detection by doctors is time consuming and prone to human errors and hence automatic detection
techniques could aide in providing fast and accurate healthcare.
Tumors are an abnormal growth which can vary in size, location and type. The patient may have a single or multiple tumors at different locations in the brain. MRI
*marked authors have equal contribution
(Magnetic Resonance Imaging) is one of the most advanced technology for diagnosing neurological abnormalities like tumors. It is preferred over CT Scan, X-Ray and
Ultrasound for obtaining high-resolution images. There are several MR sequences like
T1, T1 contrast, T2, FLAIR, PD which are used for detection of various disorders like
stroke, cysts, tumor as well as diagnosing of various neurodevelopmental disorders
like Parkinson’s and Alzheimer’s. As each patient data consists of a large number of
sequences which requires time to analyses, therefore, automation would aide diagnosticians in reducing the time taken to diagnose the patient as well as make a more accurate diagnosis.
In the past, various methods have been proposed for tumor segmentation. Kraus et
al. developed an algorithm on automatic template based segmentation, which has used
brain atlas for statistical classification process that divided the image into different
parts based on the signal intensity value [1]. The automatic segmentation accuracy
varied only 0.6% compared to the manual segmentation accuracy. The algorithm was
tested for 20 patients. An atlas is used for the segmentation process and is therefore
dependent on an existing data base. Corso et al. developed a brain tumor segmentation
technique using Bayesian formulation for calculating affinities & using it in Multilevel segmentation by weighted aggregation algorithm. They have used T1 contrast,
T2 and FLAIR sequences for 20 patient data. The accuracy was found to be 70% [2].
The segmentation is done but the area in which the tumor appears is not identified by
name. Moon et al. used the expectation maximization (EM) algorithm to modify the
probabilistic brain atlas using the spatial priors of the tumor which are obtained using
T1 weighted & T1 post contrast difference images. The dataset consisted of T1, T2,
T1 post-contrast MR images [3]. Pereira et al. used a 3X3 patch kernel for training
convolutional neural network (CNN). Bias field correction, patch and intensity normalization were used for pre-processing. Stochastic Gradient Descent is used to minimize the CNN loss function. The idea of data augmentation in brain tumor segmentation was also explored in that work. For testing they used the BRATS 2013 & 2015
databases. Each patient had T1, T1 contrast, T2 and FLAIR MR sequences [4].
Vergleichende et al. proposed a method where the cerebral cortex was divided in
47 distinct regions as visible in the cell-body stained histological sections. This cytoarchitectonic map of the human brain was revolutionary in itself as it helped doctors
understand the human brain like never before [5]. Amunts et al. suggested a 3 level
concept which included repetitive modular structure and micro and miso maps [6].
Fischl et al. proposed a new way to label the brain areas. The method used the Bayesian framework to label the brain as it allowed explicit incorporation of prior information [7]. Thompson et al. created a probability space of random transformations,
based on the theory of anisotropic Gaussian Random Fields to represent the probabilistic brain atlas with an aim to detect any changes in the normal anatomy of the human brain [8]. Shattuck et al. proposed a new method for creating a probabilistic brain
atlas from a set of T1-weighted MRI volumes of 40 patients. The atlas was constructed from manually delineated MRI data. Delineation was performed using the trained
rafters into 56 structures. Then 3 normalization algorithms were used to create 3 brain
templates. Average calculated at each voxel location was used to estimate the probability of that voxel belonging to one of the 56 structures [9].
In this paper, a novel GUI based tumor segmentation and probable mapping to its
corresponding Brodmann area is proposed. This proposed algorithm is tested on a
dataset of T2 sequences of 15 patients, taking into consideration all possible variations in age, gender and ethnicity. This paper is organized as follows: Section 2 discusses the methodology used for tumor segmentation and GUI development for
Brodmann area identification. Section 3 analyses and discusses the results obtained.
Section 4 concludes the work done.
2
Methodology
Our work has been divided into two sections to comprehensively segment tumor and
find the corresponding Brodmann areas and anatomical description of the tumor regions. The first section segments the tumor by forming a neat boundary over the tumor and the second section uses this segmented image and displays the Brodmann
areas of the brain in which the tumor exists. The proposed method is depicted in Fig
1.
Fig. 1. Proposed Method
2.1
Tumor Segmentation
Normalization
The patient data was normalized using T2.nii template on the SPM toolbox in
MATLAB. A 10mm gap between the slices was taken and so 16 slices per patient
were obtained.
3
Fuzzy C Means
Fuzzy clustering or soft clustering is a clustering technique in which each data
point can belong to more than one clusters. This is in comparison to K means clustering technique or hard clustering which assigns a data point to a single cluster.
Clustering is a method of assigning group or cluster to the different data points,
such that items in the same class or cluster are as similar as possible, while items belonging to different classes are as dissimilar as possible.
The steps for Fuzzy Clustering are as follows:
Fig. 2. Fuzzy Clustering Method
• Assumption: Initially choose a number of clusters, c.
• Initialization: Randomly assign each point a cluster coefficient, compute the probability that each data point xi is a member of a given cluster k.
• Iteration: Repeat until the algorithm has converged by:
─ Computing the centroid for each cluster.
𝐶𝑘 =
∑𝑥 𝑤𝑘 (𝑥)𝑚 𝑥
(1)
∑𝑥 𝑤𝑘 (𝑥)𝑚
─ Computing the coefficient matrix is for each point.
𝑎𝑟𝑔𝑚𝑖𝑛𝐶 ∑𝑛𝑖=1 ∑𝑐𝑗=1 𝑤𝑖𝑗 𝑛 ‖𝑥𝑖 − 𝑐𝑗 ‖
2
(2)
1
𝑤𝑖𝑗 =
∑𝑐𝑘=1(‖
𝑥𝑖 −𝑐𝑗
𝑥𝑖 −𝑐𝑘
‖)
2
𝑚−1
(3)
• Termination: Iteration is carried out until the difference in coefficients between
two iterations is no more than a given sensitivity threshold.
2.2
Brodmann Area and Anatomical Brain Area Detection
Brodmann Database creation.
A database consisting of 47 Brodmann areas was organized for right and left hemispheres. For each hemisphere a database of different Brodmann areas was created for
each slice by masking each Brodmann area using a known template.
Brodmann Area Detection.
A 3D matrix using the above database of Brodmann areas was created with dimensions 79x95xn where n is the number of Brodmann Areas in a particular slice which is
fixed for each slice for each hemisphere.
The matrix addition of the binary segmented tumor image and Brodmann Area images was done and the coordinates whose intensity exceeded ‘1’ were taken into consideration and the slice number was noted.
Creating of GUI
GUI provides a user-friendly and an interactive tool for finding out which Brodmann area has been affected by tumor and the anatomical description of the tumor
regions so that it becomes easier for the doctor to use appropriate treatment methodology for the patient.
The GUI provides the user a wide variety of options which include: •
•
•
•
•
Loading the MRI images stored anywhere in the computer.
Viewing each slice of the MRI image.
Tumor segmentation for each slice of MRI image using fuzzy-c- means method.
Determining the best segmented tumor among the 5 segmented tumor images.
Displaying the Brodmann areas and the anatomical description of the tumor region
for both right and left hemisphere separately.
5
3
Results
3.1
Data Used
Data was acquired for 15 patients using a 3T PHILIPS MRI machine. The acquired
data had varying contrasts in ethnicity, age (25-50), tumor locations and sizes. All the
images were normalized with 10mm thickness slice gap.
3.2
Tumor Segmentation
The user inputs the respective normalized nifti file in the GUI interface. Then the slice
for which the segmented tumor wants to be seen is selected. As the number of clusters
taken in Fuzzy C Means Algorithm is 5, hence 5 different segmented clusters are obtained as shown in figure 3. Then the most appropriate of the five segmented images
was selected.
The tumor was delineated from the MR image. The tumor segmentation was found
to be effective.
Fig. 3. Fuzzy gives 5 output clusters
3.3
Brodmann Area Description of the tumor
This step takes in the best selected tumor segmented image from the previous step and
tells us the Brodmann areas and anatomical regions of the brain where the tumor exists as shown in figure 4.
Fig. 4. Segmentation and Brodmann Area Detection
3.4
Discussion
Automatic Tumor Detection has gained importance as various research groups are
currently developing methodologies for the same. This GUI can aid the doctors in
correctly and accurately diagnosing tumor. Cobzas et al. proposed a new method
where they defined a set of multidimensional features based on templates and atlases
and then used them to segment tumor. An argument that region based segmentation
methods based on texture features are not suitable for segmentation as they do not
differentiate between tumor and patient effectively was presented [10]. Joshi et al.
proposed a 2 step process for tumor segmentation. In the first stage they used the gabor filter to extract the features and then contour based methods are used to detect the
different objects followed by the discrete level sets for segmentation. They then found
their results to be better than the existing methods [11]. Khotanlou et al. used the concept that the brain has to be symmetric about the central plane and hence presence of
tumor will disturb the symmetry. This is followed by Fuzzy C- means classification
7
(FCM) and Probabilistic Fuzzy C Means Classification which segments the tumor
from the MR image. Testing was done on a dataset of 20 patients having T1 weighted
MR images and decent segmentation accuracy was reported by them [12].Manasa et
al. used Canny Edge Detection algorithms to segment tumor. They argued Canny
Edge Detection is best way to find all the contours irrespective of the edges of white
and grey matter as compared to other methods [13]. Zhan et. al made use of multiple
modalities in MRI to detect features of tumor which is accompanied by the use of
Multiple Classifier system for calculating the brain tumor probability of the image
[14]. Xie et al. used the region and boundary information in a hybrid level set(HLS)
algorithm to segment the MRI image. The author used 256 axial tumor MR images of
10 patients for tumor segmentation. The method was tested against the manual segmentation method and a mean accuracy of 85.67% was achieved [15].
Tiede et al. described the anatomy of brain in two levels. First level is the volume
level which contains info information about the voxel intensity and its relation to the
basic regions Second level is the knowledge level which describes the relation between different basic regions in First level. The authors found their method to be superior that the other existing methods [16]. Lancaster et al. proposed a method to determine brain labels from the Talairach atlas. 160 regions were marked in the atlas
using volume and sub-volume components. The atlas divided the brain into 5 levels
with each structure being divided into substructures in the next level [17]. Mazziotta
et al. explained the need of a probabilistic atlas of human brain stating the fact that a
probabilistic atlas is tolerant to changes and will preserve variation in brain anatomy.
The accuracy of such atlas in clearly labelling the areas improves as the data is increased. This map allowed the user to see the anatomy of the brain and also provided
them with a database of each and every region they asked about [18]. MacDonald et
al. proposed deformable models for cortex identification. For this purpose, they have
used the anatomic segmentation using Proximities (ASP) Algorithm. The advantage
of ASP in segmenting human cortex is that it uses simple geo- metric constraints
based on anatomical knowledge [19]. Christensen et. al. used a transformation function to accommodate the differences of the brain shape with respect to the standard
brain images dictionary. Grenander’s shape models were used by them to represent
the structure and variation of brain [20]. Van Lemeput et. al. proposed the idea of
using a digital brain atlas for prior location of tissues and used MRF (Markov Random Field) to classify brain and non-brain tissues [21].
In this paper, a widely varied dataset of 15 patients was used. T2 weighted sequences were analyzed. Each data was normalized to ensure symmetry between different images of different patients. Fuzzy C Means and K Means Clustering were analyzed for all patients. It was found that Fuzzy C Means was found to give better segmentation accuracy. This method is better as compared to other existing state of the
art methods as no training of the data or template for comparison is required. This
method takes care of all types of test cases which sometimes are not properly segmented by training methods like Convolutional Neural Network(CNN).
The segmented tumor image is then sent for Brodmann area detection. A database
of the Brodmann and anatomical areas of the brain was created and then mapping
between the segmented tumor image and the database was done. This method is better
as compared to other existing methods as the database created was very vast taking
care of all possible Brodmann Areas.
The GUI provides the doctor an easy, user friendly tool to aid them in tumor detection. It also provides the doctor the exact areas where the tumor exists and thus helps
them to plan the exact course of treatment for the patient.
4
Conclusion
The algorithm is comprehensive in segmenting as well as detecting the Brodmann
areas of the tumor. This was tested on a dataset of 15 very widely variant test patients
and very good results were achieved. The creation of a GUI makes the whole process
very user friendly.
5
References
1. Kaus, Michael R., Simon K. Warfield, Arya Nabavi, Peter M. Black, Ferenc A. Jolesz, and
Ron Kikinis. "Automated segmentation of MR images of brain tumors 1." Radiology 218,
no. 2 (2001): 586-591
2. Corso, Jason J., Eitan Sharon, Shishir Dube, Suzie El-Saden, Usha Sinha, and Alan Yuille.
"Efficient multilevel brain tumor segmentation with integrated bayesian model classification." IEEE transactions on medical imaging 27, no. 5 (2008): 629-640.
3. Moon, Nathan, Elizabeth Bullitt, Koen Van Leemput, and Guido Gerig. "Model-based
brain and tumor segmentation." In Pattern Recognition, 2002. Proceedings. 16th International Conference on, vol. 1, pp. 528-531. IEEE, 2002.
4. Pereira, Sérgio, Adriano Pinto, Victor Alves, and Carlos A. Silva. "Brain Tumor Segmentation Using Convolutional Neural Networks in MRI Images." IEEE transactions on medical imaging 35, no. 5 (2016): 1240-1251.
5. Brodmann, Korbinian. Vergleichende Lokalisationslehre der Grosshirnrinde in ihren
Prizipien dargestellt auf Grund des Zellenbaues. Barth, 1909.
6. Amunts, Katrin, and Karl Zilles. "Architectonic mapping of the human brain beyond
Brodmann." Neuron 88, no. 6 (2015): 1086-1107.
7. Fischl, Bruce, David H. Salat, Evelina Busa, Marilyn Albert, Megan Dieterich, Christian
Haselgrove, Andre Van Der Kouwe et al. "Whole brain segmentation: automated labeling
of neuroanatomical structures in the human brain." Neuron 33, no. 3 (2002): 341-355.
8. Thompson, Paul M., and Arthur W. Toga. "Detection, visualization and animation of abnormal anatomic structure with a deformable probabilistic brain atlas based on random
vector field transformations." Medical image analysis 1, no. 4 (1997): 271-294.
9. Shattuck, David W., Mubeena Mirza, Vitria Adisetiyo, Cornelius Hojatkashani, Georges
Salamon, Katherine L. Narr, Russell A. Poldrack, Robert M. Bilder, and Arthur W. Toga.
"Construction of a 3D probabilistic atlas of human cortical structures." Neuroimage 39,
no.3 (2008): 1064-1080.
9
10. Cobzas, Dana, Neil Birkbeck, Mark Schmidt, Martin Jagersand, and Albert Murtha. "3D
variational brain tumor segmentation using a high dimensional feature set." In 2007 IEEE
11th International Conference on Computer Vision, pp. 1-8. IEEE, 2007.
11. Joshi, Anjali, V. Charan, and Shanthi Prince. "A novel methodology for brain tumor detection based on two stage segmentation of MRI images." In Advanced Computing and Communication Systems, 2015 International Conference on, pp. 1-5. IEEE, 2015.
12. Khotanlou, Hassan, Olivier Colliot, Jamal Atif, and Isabelle Bloch. "3D brain tumor segmentation in MRI using fuzzy classification, symmetry analysis and spatially constrained
deformable models." Fuzzy sets and systems 160, no. 10 (2009): 1457-1473.
13. Manasa, N., G. Mounica, and B. Divya Tejaswi. "Brain Tumor Detection Based on Canny
Edge Detection Algorithm and its area calculation." Brain (2016).
14. Zhan, Tianming, Yongzhao Zhan, Yao Ji, Shenghua Gu, Jin Wang, and Lei Jiang. "Brain
Tumor Segmentation in Multi-modality MRIs Using Multiple Classifier System and Spatial Constraint." In 2015 3rd International Conference on Computer, Information and Application, pp. 18-21. IEEE, 2015.
15. Xie, Kai, Jie Yang, Z. G. Zhang, and Y. M. Zhu. "Semi-automated brain tumor and edema
segmentation using MRI." European Journal of Radiology 56, no. 1 (2005): 12-19.
16. Tiede, Ulf, Michael Bomans, Karl Heinz Höhne, Andreas Pommert, Martin Riemer, Th
Schiemann, Rainer Schubert, and Werner Lierse. "A computerized three-dimensional atlas
of the human skull and brain." In Neuroimaging I, pp. 185-197. Springer US, 1996.
17. Lancaster, J. L., L. H. Rainey, J. L. Summerlin, C. S. Freitas, P. T. Fox, A. C. Evans, A.
W. Toga, and J. C. Mazziotta. "Automated labeling of the human brain: a preliminary report on the development and evaluation of a forward-transform method." Human brain
mapping 5, no. 4 (1997): 238.
18. Mazziotta, John C., Arthur W. Toga, Alan Evans, Peter Fox, and Jack Lancaster. "A probabilistic atlas of the human brain: Theory and rationale for its development: The international consortium for brain mapping (icbm)." Neuroimage 2, no. 2 (1995): 89-101.
19. MacDonald, David, Noor Kabani, David Avis, and Alan C. Evans. "Automated 3-D extraction of inner and outer surfaces of cerebral cortex from MRI." NeuroImage 12, no. 3
(2000): 340-356.
20. Christensen, Gary E., Richard D. Rabbitt, and Michael I. Miller. "3D brain mapping using
a deformable neuroanatomy." Physics in medicine and biology 39, no. 3 (1994): 609.
21. Van Leemput, Koen, Frederik Maes, Dirk Vandermeulen, and Paul Suetens. "Automated
model-based tissue classification of MR images of the brain." IEEE transactions on medical imaging 18, no. 10 (1999): 897-908.
| 1cs.CV
|
Wireless Link Capacity under Shadowing and Fading
arXiv:1706.05269v1 [cs.IT] 16 Jun 2017
Magnús M. Halldórsson
Tigran Tonoyan
ICE-TCS, School of Computer Science
Reykjavik University
[email protected],[email protected]
June 19, 2017
Abstract
We consider the following basic link capacity (a.k.a., one-shot scheduling) problem in wireless
networks: Given a set of communication links, find a maximum subset of links that can successfully transmit simultaneously. Good performance guarantees are known only for deterministic
models, such as the physical model with geometric (log-distance) pathloss. We treat this problem under stochastic shadowing under general distributions, bound the effects of shadowing
on optimal capacity, and derive constant approximation algorithms. We also consider temporal fading under Rayleigh distribution, and show that it affects non-fading solutions only by a
constant-factor. These can be combined into a constant approximation link capacity algorithm
under both time-invariant shadowing and temporal fading.
1
Introduction
Efficient use of networks requires attention to the scheduling of the communication. Successful
reception of the intended signal requires attention to the interference from other simultaneous
transmissions, and both are affected crucially by the vagaries in the propagation of signals through
media. We aim to understand the fundamental capacity question of how much communication can
coexist, and the related algorithmic aspect of how to select large sets of successfully coexisting links.
The hope is to capture the reality of signal propagation, while maintaining the fullest generality:
arbitrary instances, and minimal distributional assumptions.
The basic property of radio-wave signals is that they attenuate as they travel. In free space, the
attenuation (or “pathloss”) grows with the square of the distance. In any other setting, there are
obstacles, walls, ceilings and/or the ground, in which the waves can go through complex transforms:
reflection, refraction (or shadowing), scattering, and diffraction. The signal received by a receiver is
generally a combination of the multiple paths that it can travel, that are phase-shifted, resulting in
patterns of constructive and destructive interference. The general term for variation of the received
strength of the signal from the free-space expectation is fading.
The fading of signals can be a function of time, location, frequency and other parameters, of
which we primarily focus on the first two. A distinction is often made between large-scale fading, the
effects of larger objects like buildings and trees, and the small-scale fading at the wavelength scale
caused by multiple signal-paths. We primarily distinguish between temporal fading, that varies
randomly within the time frame of communication, and shadowing, that is viewed as invariant
within the time horizon of consideration. Temporal fading is typically experienced at the smallscale as a combination of multi-path propagation and movement or other environment changes.
The most common way of modeling true fading is stochastic fading. To each point in space-time,
we associate a random variable drawn from a distribution. Typically, this is given by a distribution
in the logarithmic dBm scale, so on an absolute scale the distributions are exponential. There is a
general understanding that log-normal shadowing (LNS), which is Gaussian on the dBm scale, is
the most faithful approximation known of medium-large scale fading or even all atemporal fading
[46, 8, 11]. Empirical models often add variations depending on the environment, the heights of
the sender/receiver from the ground, and whether there is a line-of-sight (e.g., [30]). The most
prominent among the many models proposed for small-scale and temporal fading are the ones of
Rayleigh and Rice [39], with the former (latter) best suited when there is (is no) line-of-sight,
respectively. The Rayleigh distribution mathematically captures the case when the signal is highly
scattered and equally likely to arrive at any angle. Though probabilistic models are known to be far
from perfect, they are generally understood to be highly useful for providing insight into wireless
systems, and certainly more so than the free-space model alone.
Stochastic fading is the norm in generational models, such as for simulation purposes. For
instance, LNS is built into the popular NS-3 simulator. It is also commonly featured in stochastic
analysis, e.g. [29]. Worst-case analysis of algorithms has, however, nearly always involved deterministic models, either the geometric free-space model, or extensions to more general metric spaces,
e.g., (cite various). One might expect such analysis to treat similarly arbitrary or “any-case” fading,
but that quickly invokes the ugly specter of computational intractability [16].
Problems and setting. In the Link Capacity problem, we are given a set L of links, each of
which is a sender-receiver pair of nodes on the plane. We seek a maximum feasible subset of links
in L, where a set is feasible in the physical (or SINR) model if, for each link, the strength of the
signal at the receiver is β times larger than total strengths of the interferences from the other links.
We consider arbitrary/any-case positions of links, aiming for algorithms with good performance
guarantees, as well as characterizations of optimal solutions.
1
We treat Link Capacity in extensions of the standard physical model to stochastic fading. We
separate the fading into temporal and atemporal (or time-invariant) aspects, which we refer to as
temporal fading and shadowing, respectively. We generally assume independence across space in
time-invariant distributions and across time in the temporal distributions. This is a simplification,
aimed to tackle most pronounced aspects; where possible, we relax the independence assumptions,
sometimes allowing for arbitrary (worst-case) values. Observe that the two forms can be arbitrarily
correlated: the temporal results hold under arbitrary time-invariant fading.
Our results. We give a comprehensive treatment of link capacity under stochastic fading models.
We give constant-factor approximation algorithms for both time-invariant and temporal stochastic
models. These are complementary and can be multiplexed into algorithms for both types of fading.
For (time-invariant) shadowing, we allow for essentially any reasonable stochastic distribution.
We show that shadowing never decreases the optimal link capacity (up to a constant factor), but can
significantly increase it, where the prototypical case is that of co-located links. We give algorithms
for general instances, that achieve a constant factor approximation, assuming length diversity is
constant.
For temporal fading, we treat arbitrary instances that can have arbitrary pathloss/shadowing.
We show that algorithms that ignore the temporal fading given by Rayleigh distribution achieve a
constant factor approximation. The links can additionally involve weights and can be of arbitrary
length distribution.
Besides the specific results obtained, our study leaves us with a few implications that may be
of general utility for algorithm and protocol designers. One such lesson is that to achieve good
performance for shadowing,
algorithms can concentrate on the signal strengths of the links,
and can largely ignore the strength of the interference between the links. Another useful lesson is
that
algorithms can base decisions on time-invariant shadowing alone,
since the temporal fading will even out.
These appear to be the first any-case analysis of scheduling problems in general stochastic
models. In particular, ours appears to be the first treatment of approximation algorithm for
scheduling problems under shadowing or time-invariant fading.
Related Work. Gupta and Kumar [18] introduced the physical model, which corresponds to
our setting with no fading. Their work spawned off a large number of studies on “scaling laws”
regarding throughput capacity in instances with stochastic input distributions. First algorithms
with performance guarantees in the physical model were given by Moscibroda and Wattenhofer
[31]. Constant approximation for the Link Capacity problem were given for uniform power [16],
linear power [14, 38], fixed power assignments [21], and arbitrary power control [24]. This was
extended to a distributed setting [12, 2], admission control in cognitive radio [22], link rates [25],
multiple channels [6, 41], spectrum auction [23], changing spectrum availability [9], and MIMO
[43]. NP-hardness was established in [17]. Numerous works on heuristics are known, as well as
exponential time exact algorithms (e.g., [35]).
The Link Capacity problem has been fundamental to various other scheduling problems, appearing as a key subroutine for shortest link schedule [42, 16, 20], maximum multiflow [40, 4],
weighted link capacity [42, 25], and capacity region stability [3, 26].
Numerous experimental results have indicated that simplistic range-based models of wireless
reception are insufficient, e.g., [15, 28, 46]. Significant experimental literature exists that lends
2
support for stochastic models [33], especially log-normal shadowing, e.g., [46, 8, 11]. Analytic
results on stochastic fading are generally coupled with stochastic assumptions on the inputs, such
as point processes in stochastic geometry [29, 19, 44]. Most stochastic fading models though do
not lend themselves to closed-form formulation; Rayleigh fading is a rare exception [7]. Log-normal
shadowing has been shown to result in better connectivity [37, 32] and throughput capacity [36],
but this may be artifact of the i.i.d. assumption [1].
The only work on Link Capacity with any-case instances in fading models is by Dams et al.
[10], who showed that temporal Rayleigh fading does not significantly affect the performance of
LinkCapacity algorithms, incurring only a O(log∗ n)-factor increase in performance for link capacity algorithms. We improve this here to a constant factor. Rayleigh fading has also been
considered in distributed algorithms for local broadcast [45].
The non-geometric aspects of signal propagation have been modeled non-stochastically in various ways. One simple mechanism is to vary the pathloss constant α [18]. A more general approach
is to view the variation as deforming the plane into a general metric space [13, 21]. Also, the
pairwise pathlosses can be obtained directly from measurements, inducing a quasi-metric space
[6]. All of these, however, lead to very weak performance guarantees in the presence of the huge
signal propagation variations that are seen in practice (although some of that can be ameliorated
by identifying parameters with better behavior, like “inductive independence” [23]).
2
2.1
Models and Formulations
Communication Model
The main object of our consideration is a set L of communication links, numbered from 1 to n = |L|.
Each link i ∈ L represents a unit-demand communication request between a sender node si and a
receiver node ri , both point-size wireless nodes located on the plane.
We assume the links all work in the same channel, and all (sender) nodes use the same transmission power level P (unless stated otherwise). We consider the following basic question, which
is called the LinkCapacity problem: what is the maximum number of links in L that can successfully communicate in a single time slot? We will refer to a set of links that can successfully
communicate in a single time slot as feasible.
When a subset S of links transmit at the same time, a given link i will succeed if its signal
(the power of the transmission of si when measured at ri ) is larger than β times the total (sum)
interference from other transmissions, where β ≥ 1 is a threshold parameter, and the interference
of link j on link i is the power of transmission of sj when measured at ri . We will denote by Si the
received signal power of link i and by Iji the interference of link j on link i. In this notation, link
i transmits successfully if1
Si
> β.
(1)
SIR(S, i) = P
j∈S\i Iji
2.2
Geometric Path-Loss
The Geometric Path-Loss model or GPL for short, defines the received signal strength between
nodes u and v as P/d(u, v)α , where P is the power used by the sender u, α > 2 is the path-loss
exponent and d denotes the Euclidean distance. In particular, the signal strength/power of a link
1
In general, there should also be a Gaussian noise term in the success condition, which is omitted for simplicity of
exposition. It may be noted that in expectation, the success of only a fraction of links will be affected by the noise.
3
i and the interference of a link j on link i are, respectively,
G
SiG = P/liα and Iji
= P/d(sj , ri )α ,
where li = d(si , ri ) denotes the length of link i and d(sj , ri ) is the distance from the sender node of
link j to the receiver node of link i.
If the links in a set S transmit simultaneously, the formula determining the success of the
transmission on link i is similar to (1), but we will use the slightly modified notation SIRG (S, i) > β
to indicate that GPL model is considered.
2.3
Shadowing
One of the effects that GPL ignores (or models only by appropriate change of the exponent α), is
signal obstruction by objects, or shadowing. In generic networks shadowing is often modeled by a
Stochastic Shadowing model, or SS for short, such as the Log-Normal Shadowing model, or LNS for
short. In this case, there is a parametrized probability distribution D, such that the signal strength
SiD of a link i at ri is assumed to have been sampled from the distribution D and E[SiD ] = SiG ,
D is sampled from D and E[I D ] = I G . We
and similarly, for any two links i, j, the interference Iij
ij
ij
assume that signals and interferences do not change in time (due to shadowing), at least during
the time period when LinkCapacity needs to be solved. In this model too, signal reception is
characterized by the signal to interference ratio, but we will use the notation SIRD (S, i) > β to
indicate that SS model is considered.
We shall be assuming independence among the random variables. This may lead to artifacts
that are contrary to experience. It is nevertheless valuable to examine closely this case that might
be considered the most extreme.
2.4
Temporal Fading
Another effect that is not described by the models above is the temporal variations in the signal,
due to a combination of movement (of either transceivers or people/objects in the environment)
and the scattered multipath components of the signal. We will concentrate on Rayleigh fading,
where the signal power SiF is distributed according to an exponential distribution with mean Si ,
F is distributed according to an exponential
i.e. E[SiF ] = Si , and similarly, the interference power Iij
F ] = I , where S and I are the signal and interference values,
distribution with mean Iij , i.e., E[Iij
ij
i
ij
not necessarily from SS or GPL. Again, the success of transmission is described by the signal to
interference ratio. Note, however, that in this case the success is probabilistic: the same set of links
can be feasible in one time slot and non-feasible in another.
2.5
Computational Aspects
There is a striking difference between GPL and shadowing on one side, and temporal fading on the
other side, from the computational point of view. This difference stems from the spatial nature
of GPL and shadowing, and the time-variant nature of temporal fading. In the former case, an
algorithm can be assumed to have access to, e.g., the signal strengths of links, which could be
obtained by measurements. This, however, is impossible or impractical under temporal fading,
which forces the algorithms to be probabilistic and base the actions solely on the expected values
of signal strengths (w.r.t. fading distribution), and the performance ratio of algorithms is measured
accordingly (see Sec. 5 for details).
4
2.6
Technical Preliminaries
Throughout this text, by “constants” we will mean fixed values, independent of the network size
and topology (e.g. distances). Some examples are parameters α, β, and the constants under big O
notation.
Affectance. In order to describe feasibility of a set of links with arbitrary signal and interference
values, we will use the notion of affectance, which is more convenient than (but equivalent to) the
I
1
= Sijj and extend this definition to
signal to noise ratio. For two links i, j we let ai (j) = SIR(i,j)
P
P
subsets: If S ⊆ L is a set of links, then aS (j) = i∈S\j ai (j) and ai (S) = j∈S\i ai (j). Then, a
set S of links is feasible if and only if aS (j) < 1/β holds for every link j ∈ S. When considering
a particular SS distribution D or GPL we will use superscripts D and G respectively, as before.
We will use the following result of [5], which shows that feasibility is robust with respect to the
threshold value β.
Lemma 1. If a set S of links and number β ′ > 0 are such that aS (i) ≤ 1/β ′ for each link i ∈ S,
then S can be partitioned into at most ⌈2β/β ′ ⌉ feasible subsets.
Smooth Shadowing Distributions. We will use quantiles of an SS distribution. Consider an
SS distribution D and assume links use uniform power assignment. For a probability p ∈ (0, 1) and
a link i with E[SiD ] = SiG , let S̄ip denote the 1 − p-quantile, i.e., a value x such that Pr[SiD > x] = p.
For a given number p ∈ (0, 1), the distribution D is called p-smooth, if there is a constant c > 0 such
that S̄ip ≥ c · SiG holds for each link i. Note that all major distributions used to model stochastic
shadowing satisfy such a smoothness condition.
3
Comparing Shadowing to GPL
We start by comparing the optimal solutions of LinkCapacity under SS and GPL for any given
set L of links. However, a particular instance drawn from an SS distribution D is arguably not
informative and can be hard to solve. Instead, we will be more interested in the gap between GPL
optimum and a “typical” SS optimum, in the sense of expectation.
We denote by OP T G (L) the size (number of links) of the optimal solution to LinkCapacity
for a set L of links under GPL. Similarly, we denote by OP T D (L) the size of the optimal solution to
LinkCapacity under SS model with distribution D. We assume that for every link i, the variables
D (for all j) are independent, unless specified otherwise.
SiD and Iij
We will compare the expected value E[OP T D (L)] with OP T G (L), for a given set L of links,
D for all links i, j. In
where the expectation is over the distributions of random variables SiD and Iji
particular, we prove that the expected SS optimum is never worse than a constant factor of the
GPL optimum. On the other hand, due to the presence of links with high signal strength that can
appear as a result of shadowing, the capacity can considerably increase.
3.1
SS Does not Decrease Capacity
First, we show that E[OP T D (L)] = Ω(OP T G (L)), i.e., a typical optimum under SS is not worse
than the optimum under GPL.
Theorem 1. Let D be a p-smooth SS distribution with a constant p > 0, and let L be any set of
links. Then E[OP T D (L)] = Ω(OP T G (L)).
5
Proof. Let S be a maximum cardinality subset of L that is feasible under
P GPL,G and let
P us fixD a
G
D
link i ∈ S. Recall that Si = E[Si ] and, by additivity of expectation, j∈S Iji = E[ j∈S Iji ].
Also, by smoothness assumption, there are constants c > 0 and p ∈ (0, 1) such that S̄ip ≥ c · SiG and
Pr[SiD > S̄ip ] = p. On the other hand, by Markov’s inequality,
X
X
D
G
Iji
≤2
Iji
≥ 1/2.
Pr
j∈S\i
j∈S\i
D for each i, j ∈ L. Thus,
Recall that we assumed that the random variable SiD is independent from Iji
P
P
p
G
D ≤ 2
we have that with probability at least p/2, both SiD > S̄i > cSiG and j∈S\i Iji
j∈S\i Iji
P
D
hold, implying that SiD > βc
j∈S\i Iji . By additivity of expectation, it follows that the expected
2
P
p
D
size of a subset S ′′ of S with SiD > βc
j∈S\i Iji for each link i is at least 2 · |S|. On the other
2
hand, such a set S ′′ can be partitioned into at most 4/c feasible (under SS) subsets, by Lemma 1.
G
This implies that E[OP T D (L)] ≥ pc
8 · OP T (L).
In the particular case of Log-Normal Shadowing, even independence is not necessary for the
result above to hold, as shown in th next theorem. We use the standard notation X ∼ ln N (µ, σ 2 )
and Y ∼ N (µ, σ 2 ) to denote log-normally and normally distributed random variables, respectively.
In Log-Normal Shadowing model, we assume that for all links i, j, Si ∼ ln N (µi , σ 2 ) and Iij ∼
ln N (µij , σ 2 ) for appropriate positive values µi , µij and σ. In particular, the second parameter σ is
constant.
A log-normally distributed variable X ∼ ln N (µ, σ 2 ) can be seen as X = eZ , where Z ∼ N (µ, σ 2 )
2
is a normal random variable. We will use the fact that E[X] = eµ+σ /2 . We will also use the following
basic fact.
Fact 1. If X ∼ N (µ1 , σ 2 ) and Y ∼ N (µ2 , σ 2 ) are normal random variables, then E[eX−Y ] ≤
2
eµ1 −µ2 +2σ .
Proof. By Cauchy-Schwartz,
X−Y
E[e
X
−Y
] = E[e · e
]≤
q
E[e2X ]E[e−2Y ].
Since 2X ∼ N (2µ1 , 4σ 2 ) and −2Y ∼ N (−2µ2 , 4σ 2 ), using the formula for the expectation of a
log-normal variable gives
p
2
E[eX−Y ] ≤ e2µ1 +2σ2 · e−2µ2 +2σ2 = eµ1 −µ2 +2σ .
Theorem 2. For any set L of links under Log-Normal shadowing D, E[OP T D (L)] = Ω(OP T G (L)),
even if the signal and interference distributions are arbitrarily correlated.
Proof. Let S ⊆ L be a feasible subset of L under GPL. It is enough to show that the expected
size of an optimal feasible subset of S under LNS is Ω(|S|). Consider an arbitrary link i ∈ S. The
affectance of i under LNS is:
aD
S (i)
=
D
X Iji
j∈S\{i}
SiD
6
=
X eZji
,
eZi
j∈S\{i}
where log-normal random variables eZi and eZji represent the signal of link i and interference caused
G
by link j, respectively. Recall that Iji
= E[eZji ] and SiG = E[eZi ]. Let us denote µi = E[Zi ] and
µji = E[Zji ] and note that the variables Zi , Zji have variance σ 2 . Hence, using the expectation
2
2
G
formula for log-normal variables, we can observe that SiG = eµi +σ /2 and Iji
= eµji +σ /2 . Using
this observation together with Fact 1, we obtain that for every pair i, j,
2
E[aD
j (i)]
Zji −Zi
= E[e
=
G
Iji
SiG
2
µji −µi +2σ2
]≤e
2
eµji +σ /2
= µ +σ2 /2 · e2σ
i
e
2
· e2σ = e2σ aGj (i).
Using the fact that S is feasible, and linearity of expectation, we have, for every i ∈ S, that
2σ2 · aG (i) ≤ e2σ2 /β. Thus, using Markov’s inequality, we obtain that
E[aD
S (i)] ≤ e
S
2
2
2σ
2σ
< 1/2.
/β] < β · E[aD
P[aD
S (i)]/2e
S (i) > 2e
2
2σ /β is less than |S|/2.
The latter implies that the expected number of links i ∈ S with aD
S (i) > 2e
2
It remains to note that by Lemma 1, a 1/(4e2σ )-th fraction of the remaining set of links will be
feasible under LNS, i.e., we will have that aD
S (i) < 1/β for those links i.
3.2
SS Can Increase Capacity
Next we show that, perhaps surprisingly, there are instances L for which E[OP T D (L)] ≫ OP T G (L):
the typical optima under SS can be much better than the optimum under GPL. The intuition is
that shadowing will create many links with higher signal strength than the expectation, which will
be the main contributors to the increase in capacity.
In the remainder of this section we consider a set L of links of the same length li = ℓ, and assume
that all sender nodes are located at one point (for all) and all receivers are located at another point.
We call such links co-located. Note that under GPL, any feasible subset of co-located links contains
a single link. We show below that under SS, the capacity can significantly increase. Let us fix a
p-smooth shadowing distribution D for a constant p > 0. Since L is a set of co-located links of
G
G
length ℓ, we have that for each i, j ∈ L, SiG = SjG = Iij
= Iji
= ℓPα which we denote S̄ for short.
In a dense or co-located set of links, the only hope for an increase in capacity are links of signal
strength higher than S̄. Intuitively, if there is a feasible subset S ⊆ L made of links of strength kS̄,
then |S| = O(k), since the total interference on each link is likely to be k S̄. The following definition
essentially captures the maximum size of such a set of strong links there can be in L.
For each link i, denote f (t) = Pr[SiD > tS̄] the probability that link i has signal strength at
′
least t times
i is expected. For each integer n > 0, there is a maximal value gn ≥ 0 such that
h ′ what
′
f (gn′ ) ∈ gnn , 2gnn , because f (t) is a non-increasing function of t, and f (t) → 0 when t → ∞. We
will use the following slightly
different definition: gn = max{1, gn′ }. For the case of log-normal
√
distribution, gn = Θ(e2σ ln n ), as shown in Cor. 1 below.
The following lemma (Lemma 2) gives an upper bound on the SS optimum, when the signal
strengths of links are fixed and there are few links with “strong signal”. Essentially, it indicates
that the main contribution to the capacity is by links with strong signal, and interestingly, using
power control cannot change this.
We will use the following result from linear algebra.
P √
Fact 2. [34, 27] Let A = (aij ) be an n × n non-negative real matrix. Then r(A) ≥ n1 i,j aij aji ,
where r(A) is the largest eigenvalue of A.
7
Lemma 2. Let L be a set of co-located links. Assume that the signal strengths Si are fixed, but
D are drawn from a p-smooth SS distribution D, for a number p > 0. For a number
interferences Iij
n}
for
s > 0, assume that there are at most m links i ∈ S with Si ≥ s · S̄, where m > C max{s,log
p2
D
a large enough constant C > 0. Then OP T (L) < 4m holds w.h.p. with respect to interference
distributions, even when links use power control.
Proof. Let m′ = 4m. Let E∃f eas denote the event that L contains a feasible (under SS) subset of
size m′ . Let ESf eas denote the event that a set S is feasible. By the union bound, we have that
′
P
en m
Pr[E∃f eas ] ≤ S⊆L,|S|=m′ Pr[ESf eas ]. The sum is over mn′ ≤ m
subsets, so it is enough to
′
′
′
−m
.
prove that for each subset S ⊆ L of size m , Pr[ESf eas ] < (en)
′
Let us fix a subset S of size m = 4m for the rest of the proof. Note that S contains a subset S ′
D are identically
of t = 3m links such that Si < s · S̄. Since D is a p-smooth distribution and all Iji
D > S̄ p ] = p for every
distributed, there is a constant c > 0 s.t. S̄ p ≥ cS¯ where S̄ p is such that Pr[Iij
′
pair i, j ∈ L. For each pair of links i, j ∈ S , let Bij denote the binary random variable that is 1 iff
D , I D } > S̄ p . Note that B are i.i.d. variables and Pr[B = 1] = p2 > 0. Consider the sum
min{Iij
ij
ij
P ji
X = i,j∈S ′ Bij . We have that E[X] = p2 · t(t − 1)/2. By a standard Chernoff bound, we have
2
that Pr[X < p2 · t(t − 1)/6] < e−p ·t(t−1)/9 . Recall that p is constant. We choose the constant C so
· log(en), which gives Pr[X < p2 · t(t − 1)/4] < (en)−m .
as to have t ≥ 3m ≥ 1 + 12
p2
It remains to prove that if X ≥ p2 · t(t − 1)/6, then the set S is not feasible with any power
assignment. Let A = (aij ) denote the normalized gain matrix of the set S ′ , where for any pair of
D /S and a = 1. As shown in [47], the largest SIR ratio that
links i, j ∈ S ′ we denote aij = Iij
j
ii
1
can be achieved with power control is r(A)−1
, where r(A) is the largest eigenvalue2 of A. Thus,
if we show that r(A) > 1/β + 1, then the set S is not feasible
P √even with power control. To that
end, we will use the bound given in Fact 2: r(A) > 1t i,j aij aji . We restrict our attention
only to the terms with Bij = 1, as those will have sufficient contribution to the sum. Indeed,
D , I D } > S̄ p ≥ cS̄. Since it also holds
recall that for each pair i, j with Bij = 1, we have min{Iij
ji
q
√
(S¯p )2
′
for each link i ∈ S , that Si < s · S̄, we have:
≥ c/s. Since we also have
aij aji >
(sS̄)2
|{i, j ∈ S ′ : Bij = 1}| > p2 t(t − 1)/6, we obtain the bound:
r(A) >
1
c
cp2 (t − 1)
· |{i, j ∈ S ′ : Bij = 1}| · ≥
.
t
s
6s
Hence, for any fixed β, we choose the constant C so as to have t ≥ 3m > 1 +
case r(A) > 1/β + 1 and the set S is infeasible. This completes the proof.
6s(1/β+1)
,
cp2
in which
Theorem 3. Let L be a set of co-located links under a p-smooth SS distribution D with associated
sequence {gn }. There are constants c1 , c2 > 0, such that
c1 gn < E[OP T D (L)] < c2 (gn + log n),
where the expectation is taken w.r.t. interference and signal strength distributions. In particular, if
gn = Ω(log n), then E[OP T D (L)] = Θ(gn ). The upper bound holds even if power control is used.
Proof. We begin by showing the first inequality. Note that the case gn = 1 trivially holds, so we
assume gn > 1.
2
It is important here that we defined the normalized gain matrix in terms of interferences and signal strengths
D
with respect to uniform power assignment (i.e., the power level P is “cancelled” in the ratio Iij
/Sj , leaving the gain
ratio, since links i and j use the same power level).
8
Let us call a link strong if SiD > gn S̄. By the definition of values gn and since gn > 1, the
probability that any fixed link i is strong is at least gn /n and at most 2gn /n. Let S be the
subset of strong links. The observation above readily implies, by the linearity of expectation, that
gn ≤ E[|S|] ≤ 2gn .
D ] = S̄ for all links i, j ∈ L. Thus, the expected affectance on each link i ∈ S is
Recall that E[Iji
D
X
X Iji
S̄ · E[|S|]
D
≤ 1 E
≤ 2,
Iji
E
=
Si
gn S̄
gn S̄
j∈S\{i}
j∈S\{i}
D . Thus, using Markov’s
where the expectation is taken over the distribution of interferences Iji
inequality, we get that the expected number of links in S with aD
S (i) ≤ 4 is Ω(|S|). This, together
with Lemma 1 and the fact that E[|S|] ≥ gn , proves that E[OP T D (L)] = Ω(gn ).
Now, let us demonstrate the second inequality of the claim. Let us assume that power control
is allowed, and, as before, let Si denote the signal strength of link i and Iij denote the interference
n
′
from link i to link j when all links use uniform power P . Denote c = c′ · (1 + log
gn ), where c > 8
is a large enough constant. We will show that E[OP T D (L)] = O(cgn ). Let Estr be the event that
there are at least cgn /4 strong links. Recall that the expected number of strong links is at most
2gn . Since c′ > 8, applying standard Chernoff bound gives Pr[Estr ] < e− log(2n) < 1/n, i.e., with
high probability, Estr holds: There are at most cgn /4 strong links.
Given that Estr holds, Lemma 2, applied with m = cgn /4 and s = gn (assuming the constant c′ is
suitably large), tells us that gives that with high probability, OP T D (L) = O(cgn ). This completes
the proof.
Corollary 1. For a set L of n co-located equal length links with Log-Normal Shadowing distribution
D,
√
E[OP T D (L)] = Θ(e2σ ln n ).
Proof. First, let us estimate gn . Let µ and σ be the parameters associated with the LNS distribution
of each link i. Since the links have equal lengths, those parameters are the same across all links
2
in L. Recall that S̄ = E[SiD ] = eµ+σ /2 for each link i ∈ L. We will use the fact that the tail
,
probability of a log-normal variable X with parameters µ, σ is as follows: P [X > t] = Q ln t−µ
σ
2
R ∞ −x2 /2
−x /2
where Q(x) = x e
dx is the tail probability of the standard normal distribution φ(x) = e √2π .
There is no closed form expression for Q(x), but it can be approximated as follows for all x > 0:
x
· φ(x) < Q(x) < x1 · φ(x).
x2 +1
Using these formulas, we obtain: f (t) = P[SiD > tS̄] = Q(ln t/σ + σ/2). Denote x = gn /σ + σ/2
−x2 /2
for some n > 1. Then, by the definition of gn , we must have that f (gn ) = Q(x) = Θ e x
=
√
Θ(gn /n). A simplification gives gn = Θ(e2σ ln n ). It remains to show that LNS is a smooth
distribution, i.e., for some constant p, S̄ p = Ω(S̄). To this end, note that the mean of a log normal
2
variable X with parameters σ, µ is eµ+σ /2 and the median is eµ , so assuming σ is fixed, we can
take p = 1/2.
Remark. Theorem 3 can be extended in two ways. First, we can assume that the sender nodes
and receiver nodes are not in exactly the same location, but are within a region of diameter smaller
compared to the link length. Second, we can assume that the links do not have exactly equal
lengths, but the lengths differ by small constant factors. These modifications incur only changes
9
in constant factors. This follows from Lemma 1. That is, the theorem predicts increased capacity
due to shadowing not only for co-located sets of links, but also sets that contain dense parts, i.e.,
have subsets that are “almost co-located”.
The results above assert that under SS, the main contributor to the capacity increase in a
co-located set of links are the “strong” links. The following result demonstrates that significant
capacity increase (though not as dramatic as above) can happen even when all links have signal
strength fixed to the GPL value. This phenomenon is due to interference distributions.
Theorem 4. Let L be a set of co-located links, all of length ℓ. Assume that the signal strengths are
fixed and equal to Si = SiG = S̄ but the interferences are drawn from Log-Normal shadowing model
D. Then,
p
log n/ log log n .
E[OP T D (L)] = Ω
Proof. Assume, for simplicity, that β = 1. Consider any fixed subset S ⊆ L of size |S| = t > eσ
for a fixed t to be specified later, and let i, j ∈ S. From the definition of LNS we have,
!
2
ln(eµ+σ /2 /t) − µ
D
Pr[Iij < S̄/t] = 1 − Q
σ
2
σ /2 − ln t
=1−Q
σ
2
ln t − σ /2
=Q
σ
2
≥ exp(− ln2 t/2σ 2 + ln t/2 − ln ln t + c),
for a constant c, where we used the fact that Q(x) = 1 − Q(−x). Denote φ(t) = − ln2 t/2σ 2 +
D < S̄/t holds for all pairs i, j ∈ S, is et(t−1)φ(t) .
ln t/2 − ln ln t + c. Thus, the probability that Iji
Note that if the latter event happens then S is a feasible set. Let us partition L into ⌊n/t⌋ subsets
of size t. From the discussion above we have that each of those subsets is feasible with probability
at least et(t−1)φ(t) . Since those subsets are disjoint, the probability that none of them is feasible is
at most (1 − e−t(t−1)φ(t) )n/t < exp(− nt · e−t(t−1)φ(t) ). In order to have the latter probability smaller
q
log n
than 1/2, it is sufficient to set n/t > e−t(t−1)φ(t) , which holds when t = O
log log n
The following result demonstrates the limitations of Thm. 3. Namely, it shows that the O(log n)
slack in the upper bound cannot be replaced with constant in general: We show that there is an SS
distribution
with gn = O(1), such that the gap between SS and GPL capacities is at least a factor
p
of Ω( log n/ log log log n) for co-located equal length links. It also shows that gn = Ω(log n) is not
necessary for increase in capacity compared to GPL.
Theorem
5. There
is an SS distribution D with gn = O(1) for each n, such that E[OP T D (L)] =
q
log n
Ω
log log log n holds for a set L of n co-located links.
Proof. Let S denote the expected signal strength of each link, as before. Consider the following
6
discrete SS distribution D, where for each link i and t ≥ 1, Pr[SiD = cS
2t ] = π 2 t2 , where c =
−1
P
1
6
. It is easy to show that E[SiD ] = S and gn = O(1) for each n. Consider an
t≥1 t2 2t
π2
arbitrary subset S of size m and fix a link i ∈ S. It follows from the definition of D, that Pr[SiD ≥
1
D ] = S and Pr[I D < S ] ≥
S] ≥ 1/2. For each link j ∈ S \ {i}, E[Iji
= 2 log12 m . Thus, by
ij
β·m
2β log2 m
10
m−1
P
1
D < S ] ≥ Pr[I D < S , for all j ∈ S \ i] ≥
independence of interferences, Pr[ j∈S\i Iji
ji
β
βm
2 log2 m
and
m
X
1
S D
D
D
D
.
Pr[aS (i) < 1/β] ≥ Pr
Iji < |Si ≥ S · Pr[Si ≥ S] ≥
β
2 log2 m
j∈S\i
m2
. Split L into n/m disjoint subsets
Thus, set S is feasible with probability at least 2 log12 m
S1 , S2 , . . . , Sn/m of size m and let Et denote the event that St is feasible. Note that the events Et
m2
are independent, and Pr[Et ] ≥ 2 log12 m
. By the union bound and independence, Pr[∃t, Et ] ≥
q
m2
log n
n
1
·
.
Thus,
in
order
to
have
Pr[∃t,
E
]
≥
1/2,
it
suffices
to
take
m
≥
t
m
3 log log log n .
2 log2 m
4
Computing Capacity under Shadowing
In this section we study the algorithmic aspect of LinkCapacity under SS. Namely, our aim
is to design algorithms that perform well in expectation under SS, compared with the expected
optimum under SS. In the first part, we handle (nearly) co-located links, while the second part
discusses the more general case when the links are arbitrarily placed on the plane. In both cases,
we obtain constant factor approximations for links of bounded length diversity, under general SS
distributions satisfying weak technical assumptions. This holds against an optimum that can use
arbitrary power control.
Let us start with several definitions. For any set L of links, let ∆ = ∆(L) = maxi,j∈L {li /lj }
denote the max/min link length ratio in L. A set L is called equilength if ∆(L) < 2.
An equilength set S of links is called cluster if there is a square of side length ℓ/2 that contains
the sender nodes of all links in S, where ℓ = mini∈S {li }.
4.1
Clusters
Consider a cluster L of links with minimum length ℓ. The algorithm is based on Thm. 3, which
suggests that choosing only the strong links is sufficient for a constant factor approximation to the
expected optimum. We prove that a similar result holds in a more general setting, where the signal
strengths Si of links i ∈ L are fixed and arbitrary, while interferences are drawn from a p-smooth
SS distribution D for a constant p > 0. The algorithm is as follows.
Algorithm. Let set S be constructed by iterating over the set L in a decreasing order of link
¯ Output the set
strength, and adding each link i to S if its strength Si satisfies Si > 2β|S|S.
1
D
ALG = {i ∈ S : aS (i) < β } of successful links in S.
The theorem below shows that this strategy results in only additive O(log n) expected error,
and yields a fully constant factor approximation when signal strengths are drawn from an SS
distribution with gn = Ω(log n), such as LNS.
Theorem 6. Let L be a cluster of n links. Assume that the signal strengths of links are fixed and
arbitrary, but interferences are drawn from a p-smooth SS distribution D for a constant p > 0.
Then,
Eint [OP T D (L)] = O(Eint [|ALG|] + log n),
11
where the expectation is taken only w.r.t. interference distributions. When the signal strengths are
also drawn from D and we further have gn = Ω(log n), then
E[|ALG|] = Θ(E[OP T D (L)]),
where expectation is taken w.r.t. signal and interference distributions.
Proof. P
First, let us note that E[|ALG|] ≥ |S|/2. Indeed, take a linkP
i ∈ S. We have that Si > 2β S̄|S|
D ] ≤ |S|S̄. Thus, with probability at least 1/2,
D
and E[ j∈S\i Iji
j∈S\i Iji ≤ 2S̄|S| < Si /β, i.e.,
i ∈ ALG. By additivity of expectation, this implies that E[|ALG|] ≥ |S|/2. It remains to show
that E[OP T D (L)] = O(|S| + log n). But this simply follows from Lemma 2 with s = 2β|S| and an
appropriate value m = O(max{|S|, log n}). The second part of the theorem follows from Thm. 3,
which asserts that E[|S|] = Θ(gn ) = Ω(log n).
4.2
General Equilength Sets
The algorithm presented in the previous section can be extended to general sets of equilength
links, which are not necessarily clusters. The essential idea is to partition such a set into clusters,
solve each cluster separately, then combine the solutions. This, however, requires some technical
elaboration.
Let L be an arbitrary equilength set. First, we partition L into a constant number of wellseparated subsets, where an equilength set S is called well-separated if S is a disjoint union of
subsets S1 , S2 , . . . such that for each t, St is a cluster and for each s 6= t and links i ∈ Ss and j ∈ St ,
min{d(si , rj ), d(sj , ri )} > ℓ, where ℓ = mini∈S {li }.
Proposition 1. Any equilength set L on the plane can be split into a constant number of wellseparated subsets.
Proof. Let ℓ = mini∈L li . Partition the plane into squares of side ℓ/2 with horizontal and vertical
lines. Consider only the squares intersecting the bounding rectangle of links L. Assign the squares
integer coordinates (x, y) in the following way: if two squares with coordinates (x, y) and (x′ , y ′ )
share a vertical (horizontal) edge then y = y ′ and |x − x′ | = 1 (x = x′ and |y − y ′ | = 1, resp.). Then
split L into 49 subsets Lks , k, s = 0, 1, . . . , 6, where Lk,s = {i ∈ L : si is in a square with (x, y) =
(k, s) mod 7}. Let us fix R = Lk,s for some arbitrary indices s, k and let i, j ∈ R be any links. It
remains to note that if the sender nodes of i and j are in different squares, then d(si , sj ) ≥ 6ℓ/2 = 3ℓ,
which via the triangle inequality (and using the assumption that L is equilength) implies that
d(si , rj ), d(sj , ri ) > ℓ.
Algorithm. Partition L into well-separated subsets, solve the capacity problem for each subset
separately (as described below) and output the best solution obtained.
To process a well-separated subset L′ , observe first that L′ is a disjoint union of clusters
L1 , L2 , . . . , Lm , by definition. Run the algorithm from Sec. 4.1 on each cluster Lt separately,
obtaining a subset St . Denote S = ∪m
t=1 St . Let c > 0 be a constant, as indicated in the proof of
Thm. 7, and let R be the set of all links j in S such that aD
S (j) < 2c. Partition R into at most 2cβ
feasible subsets (under SS) using Lemma 1 and let ALG be the largest of those.
Theorem 7. Let L be a set of n equilength links (arbitrarily placed on the plane) under a p-smooth
SS distribution D, for a constant p > 0, with gn = Ω(log n). Then,
E[|ALG|] = Θ(E[OP T D (L)]).
12
Proof. First, note that performing the step of partitioning into well-separated subsets and selecting
the best one, we lose at most a constant factor against the optimum. So we concentrate on a
well-separated set L′ , consisting of clusters L1 , L2 , . . . Lm . Recall that S1 , S2 , . . . , Sm be the subsets
obtained by the algorithms on the clusters and S = ∪m
t=1 St .
We first show that the expected number of links the algorithm chooses from S is Ω(|S|), where
the expectation is only w.r.t. interference distributions. To that end we show that the out-affectance
from any link in S to all the other links is constant under GPL. That means that the expected
affectance on a link in S, even under D, is constant, which yields the claim after some sparsification
of S.
For any two indices s, t ∈ {1, . . . , m}, s 6= t, define the representative affectance a(s, t) by set Ss
on set St as the largest GPL affectance by a link in Ss on a link in St : a(s, t) = maxi∈Ss ,j∈St aGi (j).
It follows from the definition of the GPL affectance and the assumption that all links have length at
most 2ℓ, that a(s, t) = O(1) · (ℓ/dst )α , where dst = mini∈Ss ,j∈St {d(si , rj )} ≥ ℓ (by well-separation).
P
Claim 1. For every s, m
t=1,t6=s a(s, t) = O(1).
Sketch. This can be shown by an area argument that uses the fact that α > 2. Take any link i ∈ Ss .
Partition the space around the sender node si into concentric rings of width ℓ. By well-separation,
there are no other links within distance ℓ from si . Number the rings starting from si . Let ar
denote the total contribution of sets St that intersect the ring number r. It is easy to see that
ar = O(1/r α ) · zr , where zr is the number of such sets. Since each set St occupies an area O(ℓ2 )
(it is a cluster) and different
P∞ a simple area argument gives that zr = O(r), so
Pm sets are disjoint,
α−2
ar = O(1/r
). Thus, t=1,t6=s a(s, t) = r=1 ar = O(1), because α > 2 and thus the last sum
converges.
Now, fix an index s and a link i ∈ Ss . Let j ∈ St , for t 6= s. Recall, from the definition
P
IG
IG
of St , that Sj = Ω(|St |)SjG . Thus, Sijj = O(1/|St |) · aGi (j) = O(1/|St |) · a(s, t) and j∈St Sijj =
O(1/|St |) · a(s, t) · |St | = O(a(s, t)). A similar argument applied to the links j ∈ Ss gives that
G
G
P
Iij
Iij
G
=
O(1/|S
|)
·
a
(j)
=
O(1/|S
|),
so
s
s
j∈S
\i
i
Sj
Sj = O(1).
s
Summing over all links and using Claim 1, we obtain that
G
X Iij
j∈S\i
Sj
=
G
X Iij
j∈Ss \i
Sj
+
G
X X Iij
t6=s j∈St
= O(1) + O(1) ·
Summing over all links i ∈ S, we obtain
P
X
Sj
a(s, t) = O(1).
t6=s
G
Iij
i,j∈S Sj
= O(|S|). By the pigeonhole principle, there
P
IG
is a subset S ′ ⊆ S with |S ′ | ≥ |S|/2, such that for each link j ∈ S ′ , i∈S Sijj ≤ c for a constant
G
D
P
P
Iij
Iij
=
c > 0. Thus, for each link j ∈ S ′ , Eint
i∈S Sj < c. The latter implies that
i∈S Sj
′
D
Pr[aD
S (j) < 2c] > 1/2 and the expected number of links j ∈ S with aS (j) < 2c is at least
′
|S |/2 ≥ |S|/4. This is exactly the set R (or Rr , for a particular index set Ir ) we are looking for.
Thus, by Lemma 1, R can be split into at most 2βc feasible subsets. The largest of them, which is
output by the algorithm, has expected cardinality at least |S|/(8βc).
It now only remains to bound OP T D (L) in terms of |S|. Recall that OP T D (L) = O(OP T D (L′ )).
We know from Thm. 6 that E[|St |] = Θ(E[OP T D (Lt )]) for t = 1, 2, . . . , m. Combining those
13
observations, we obtain the desired bound:
E[OP T D (L)] = E[OP T D (L′ )] = O
m
X
OP T D (Lt )
t=1
=O
m
X
t=1
4.3
!
(E[|St |])
!
= O (E[|S|]) .
General Sets
Any set of links can be partitioned into at most ⌈log ∆⌉ equilength subsets. Thus, we can solve the
problem for each of the subsets using the algorithm from the previous section and take the largest
solution obtained. That will give us, e.g., O(log ∆) approximation for general sets of links under
LNS. This holds even against an optimum that can use arbitrary power control.
5
The Effect of Temporal Fading
In Rayleigh fading model, due to the temporal variability of signals, it is natural to define a ”typical
optimum” as follows. For a set L of n links (numbered, as before), we consider a transmission
probability vector p̄ = (p1 , p2 , . . . , pn ), representing that link i transmits with probability pi ∈ [0, 1].
Let Qp̄ (i) denote the probability that link i succeeds when all links transmit according to the
vector p̄ (as defined by theP
signal to interference ratio). Then, the expected number of successful
transmissions is w(L, p̄) = i∈L Qp̄ (i). Finally, we define the optimum capacity as the maximum
expected number of successful transmissions with any transmission probability vector p̄; namely,
OP T F (L) = maxp̄ w(L, p̄).
Here we denote the non-fading values of signals and interferences as Si and Iij and assume that
they can take arbitrary positive values. Also, OP T (L) denotes the size of the optimal solution to
LinkCapacity with respect to those values. We will show that OP T (L) and OP T F (L) are only
constant factor apart from each other, namely, the effect of temporal fading is at most a (small)
constant factor.
In fact, we can prove a more general result. Assume that each link i has a positive weight
wi . The weighted variant
Pof LinkCapacity requires to find a feasible subset S ⊂ L of links with
maximum total weight i∈S wi . We let W OP T (L) denote the weight of an optimal solution in
the non-fading setting. Similarly, given a transmission probability vector p̄, let
Pus re-define the
expected weight of the set of successful transmissions when using p̄ as w(L, p̄) = i∈L wi Qp̄ (i) and
let W OP T F (L) = maxp̄ w(L, p̄). Then we have the following result. We assume that for every link
F (for all j) are independent.
i, the variables SiF and Iij
Theorem 8. For any set L of links, W OP T (L) = Θ(W OP T F (L)).
Dams et al. [10] have already shown that W OP T (L) = O(W OP T F (L)). Specifically, given a
feasible set S, if we form the vector p̄ by pi = 1 if li ∈ S and pi = 0 otherwise, then w(L, p̄) ≥ |S|/e.
Thus, it remains to show that W OP T (L) = Ω(W OP T F (L)).
Proof. Let q̄ be the probability
vector corresponding to W OP T F (L), namely, q̄ is such that
P
W OP T F (L) = w(L, q̄) = i∈L wi Qq̄ (i).
Note thatP
within the confines of the argument, q̄ is a deterministic vector, not a random variable.
Let Aq̄ (j) = i∈L qi · ai (j) denote the expected affectance on link j.
14
Dams et al. [10, Lemma 2] applied a characterization of Li and Haenggi [29] to obtain that for
each link i,
qi e−Aq̄ (i) ≤ Qq̄ (i)) ≤ qi e−Aq̄ (i)/2 .
(2)
The idea of our proof is to consider only nodes with small expected affectance under q̄; they
can be sparsified to a feasible set of large size. The key observation is that much of the weight in
the optimal solution with fading is centered on those low-affectance nodes; otherwise, one could
obtain a larger solution by uniformly reducing the probabilities.
Let k = 4 · ln 11/2 = 4(ln 4 + ln 11/8) ∼ 6.819. Let H = {i ∈ L : Aq̄ (i) ≤ k} be the links of
small weighted affectance and let H ′ = L \ H.
Claim 2. w(H, q̄) ≥ 23 w(L, q̄) = 23 W OP T F (L).
Proof. Suppose otherwise, so w(H ′ , q̄) > 23 w(L, q̄). Form a new probability vector p̄ = q̄/4, i.e.,
pi = qi /4, for all i ∈ L. Then, Ap̄ (i) = Aq̄ (i)/4, for all i ∈ L. For i ∈ H ′ , since Aq̄ (i) > k, it holds
from the definition of k that
−
Aq̄ (i)
Aq̄ (i)
− ln 4 ≥ −
+ ln 11/8 .
4
2
(3)
Thus, for i ∈ H ′ ,
qi −Ap̄ (i)
e
= qi e−Aq̄ (i)/4−ln 4
4
11
11 −Aq̄ (i)/2
qi e
≥ Qq̄ (i) ,
≥ qi e−Aq̄ (i)/2+ln 11/8 =
8
8
Qp̄ (i) ≥
using (2), the definition of p̄, (3), rearrangement, and again (2). Thus, w(H ′ , p̄) ≥
follows that
w(L, p̄) = w(H, p̄) + w(H ′ , p̄) ≥
11
′
8 w(H , q̄).
It
11
1
w(H, q̄) + w(H ′ , q̄)
4
8
1
9
= w(L, q̄) + w(H ′ , q̄) > w(L, q̄),
4
8
where the strict inequality uses the supposition. This contradicts the choice of q̄. Hence, the claim
follows.
We continue with the proof of Theorem 8. We use the probabilistic method to show that there
is a feasible subset of H of weight at least W OP T F (L)/(6k).
Let X be a random set of links from H, where link i ∈ H is added to X with probability qi /(2k).
Consider a link t ∈ H and let Yt = aX (t) be the random variable whose value is the affectance of
set X on link t. By definition of H, Aq̄ (t) ≤ k. Observe that
X
E[Yt ] = E[aX (t)] =
Pr[j ∈ X] · aj (t)
j∈H
Aq̄ (t)
1 X
1
=
qj · aj (t) ≤
≤ ,
2k
2k
2
j∈H
using the linearity of expectation and the definition of Aq̄ . Thus, by Markov’s inequality,
Pr[Yt ≤ 1] ≥ Pr[Yt ≤ 2 · E[Yt ]] ≥
15
1
.
2
(4)
Observe that the value of Yt is independent of the event that link t was selected into X. Let
SX = {i ∈ X : aX (i) ≤ 1} be the subset of feasible links in X. It follows that
X
X
E[
wi ] =
wi Pr[i ∈ X and Yi ≤ 1]
i∈SX
i∈L
≥
≥
≥
X
wi Pr[i ∈ X] · Pr[Yi ≤ 1]
X
wi ·
i∈H
i∈H
qi 1
·
2k 2
(Independence)
((4) and defn. of X)
1 X
wi Qq̄ (i)
4k
((2))
i∈H
w(H, q̄)
4k
W OP T F (L)
≥
6k
(Defn. of w)
=
(Claim 2) .
As per the probabilistic method, this implies there exists a feasible set of size Ω(W OP T F (L)).
6
Conclusions
We have compared link capacity in instances with and without stochastic shadowing, as well as
with and without temporal fading. We have also obtained constant-factor approximations in both
cases. Numerous open problems and directions still remain.
There is room to strengthen and generalize our results. This includes extending the approximations for shadowing to links of unbounded length diversity, and extending the temporal fading
analysis to other stochastic models.
Modeling correlations and analyzing its effect on link capacity would be valuable, and the same
holds for analyzing other scheduling problems, including weighted link capacity and shortest length
schedules.
Acknowledgements. This work was supported by grants 152679-05 and 174484-05 from the
Icelandic Research Fund. We thank Michael Neely for pointing out an error in an earlier version
of the paper.
References
[1] P. Agrawal and N. Patwari. Correlated link shadow fading in multi-hop wireless networks.
IEEE Trans. Wireless Commun., 8(8):4024–4036, 2009.
[2] E. Ásgeirsson and P. Mitra. On a game theoretic approach to capacity maximization in wireless
networks. In INFOCOM, 2011.
[3] E. I. Asgeirsson, M. M. Halldórsson, and P. Mitra. A fully distributed algorithm for throughput
performance in wireless networks. In CISS, pages 1–5, 2012.
[4] E. I. Ásgeirsson, M. M. Halldórsson, and P. Mitra. Maximum MIMO flow in wireless networks
under the SINR model. In WiOpt, pages 295–302, 2014.
16
[5] J. Bang-Jensen and M. M. Halldórsson. Vertex coloring edge-weighted digraphs. Inf. Process.
Lett., 115(10):791–796, 2015.
[6] M. Bodlaender and M. M. Halldórsson. Beyond geometry: Towards fully realistic wireless
models. In PODC, 2014.
[7] P. Cardieri. Modeling interference in wireless ad hoc networks. IEEE Commun. Surveys Tuts.,
12(4):551–572, 2010.
[8] Y. Chen and A. Terzis. On the implications of the log-normal path loss model: an efficient
method to deploy and move sensor motes. In SenSys, pages 26–39. ACM, 2011.
[9] J. Dams, M. Hoefer, and T. Kesselheim. Sleeping experts in wireless networks. In Distributed
Computing, DISC, pages 344–357. Springer, 2013.
[10] J. Dams, M. Hoefer, and T. Kesselheim. Scheduling in wireless networks with Rayleigh-fading
interference. IEEE Trans. Mobile Comput., 14(7):1503–1514, 2015.
[11] B. Dezfouli, M. Radi, S. A. Razak, T. Hwee-Pink, and K. A. Bakar. Modeling low-power
wireless communications. J Netw. Comput. Appl., 51:102–126, 2015.
[12] M. Dinitz. Distributed algorithms for approximating wireless network capacity. In INFOCOM,
pages 1–9. IEEE, 2010.
[13] A. Fanghänel, T. Kesselheim, H. Räcke, and B. Vöcking. Oblivious interference scheduling. In
PODC, pages 220–229, August 2009.
[14] A. Fanghänel, T. Kesselheim, and B. Vöcking. Improved algorithms for latency minimization
in wireless networks. In ICALP, pages 447–458, 2000.
[15] D. Ganesan, B. Krishnamachari, A. Woo, D. Culler, D. Estrin, and S. Wicker. Complex
behavior at scale: An experimental study of low-power wireless sensor networks. Technical
report, UCLA/CSD-TR 02, 2002.
[16] O. Goussevskaia, M. M. Halldórsson, and R. Wattenhofer. Algorithms for wireless capacity.
IEEE/ACM Trans. Netw., 22(3):745–755, 2014.
[17] O. Goussevskaia, Y.-A. Oswald, and R. Wattenhofer. Complexity in geometric SINR. In
MobiHoc, pages 100–109, 2007.
[18] P. Gupta and P. R. Kumar. The capacity of wireless networks. IEEE Trans. Inf. Theory,
46(2):388–404, 2000.
[19] M. Haenggi, J. G. Andrews, F. Baccelli, O. Dousse, and M. Franceschetti. Stochastic geometry
and random graphs for the analysis and design of wireless networks. IEEE J. Sel. Areas
Commun., 27(7):1029–1046, 2009.
[20] M. M. Halldórsson. Wireless scheduling with power control. ACM TALG, 9(1):7, 2012.
[21] M. M. Halldórsson and P. Mitra. Wireless Capacity with Oblivious Power in General Metrics.
In SODA, 2011.
[22] M. M. Halldórsson and P. Mitra. Wireless connectivity and capacity. In SODA, pages 516–526,
2012.
17
[23] M. Hoefer and T. Kesselheim. Secondary spectrum auctions for symmetric and submodular
bidders. ACM TEAC, 3(2):9, 2015.
[24] T. Kesselheim. A constant-factor approximation for wireless capacity maximization with power
control in the SINR model. In SODA, pages 1549–1559, 2011.
[25] T. Kesselheim. Approximation algorithms for wireless link scheduling with flexible data rates.
In ESA, pages 659–670, 2012.
[26] T. Kesselheim. Dynamic packet scheduling in wireless networks. In PODC, pages 281–290,
2012.
[27] L. Y. Kolotilina. Lower bounds for the Perron root of a nonnegative matrix. Linear Algebra
Appl., 180:133–151, 1993.
[28] D. Kotz, C. Newport, R. S. Gray, J. Liu, Y. Yuan, and C. Elliott. Experimental evaluation of
wireless simulation assumptions. In MSWiM, pages 78–82. ACM, 2004.
[29] X. Liu and M. Haenggi. Throughput analysis of fading sensor networks with regular and
random topologies. EURASIP J. Wirel. Commun. Netw., 2005(4):554–564, 2005.
[30] B. Mondal, T. A. Thomas, E. Visotsky, F. W. Vook, A. Ghosh, Y.-H. Nam, Y. Li, J. Zhang,
M. Zhang, Q. Luo, et al. 3d channel model in 3gpp. IEEE Commun. Mag., 53(3):16–23, 2015.
[31] T. Moscibroda and R. Wattenhofer. The complexity of connectivity in wireless networks. In
INFOCOM, pages 1–13, 2006.
[32] T. Muetze, P. Stuedi, F. Kuhn, and G. Alonso. Understanding radio irregularity in wireless
networks. In SECOND, pages 82–90. IEEE, 2008.
[33] H. Nikookar and H. Hashemi. Statistical modeling of signal amplitude fading of indoor radio
propagation channels. In Proc. 2nd IEEE International Conference on Universal Personal
Communications, volume 1, pages 84–88, 1993.
[34] A. J. Schwenk. Tight bounds on the spectral radius of asymmetric nonnegative matrices.
Linear Algebra Appl., 75:257–265, 1986.
[35] Y. Shi, Y. T. Hou, S. Kompella, and H. D. Sherali. Maximizing capacity in multihop cognitive
radio networks under the SINR model. IEEE Trans. Mobile Comput., 10(7):954–967, 2011.
[36] P. Stuedi and G. Alonso. Log-normal shadowing meets SINR: A numerical study of capacity
in wireless networks. In SECON, pages 550–559, 2007.
[37] P. Stuedi, O. Chinellato, and G. Alonso. Connectivity in the presence of shadowing in 802.11
ad hoc networks. In WCNC, volume 4, pages 2225–2230. IEEE, 2005.
[38] T. Tonoyan. On the capacity of oblivious powers. In ALGOSENSORS, volume 7111 of Lecture
Notes in Comput. Sci., pages 225–237. Springer, 2011.
[39] D. Tse and P. Viswanath. Fundamentals of wireless communication. Cambridge University
Press, 2005.
[40] P. Wan. Multiflows in multihop wireless networks. In MobiHoc, pages 85–94, 2009.
18
[41] P.-J. Wan. Joint selection and transmission scheduling of point-to-point communication requests in multi-channel wireless networks. In MobiHoc, pages 231–240, 2016.
[42] P.-J. Wan, O. Frieder, X. Jia, F. Yao, X. Xu, and S. Tang. Wireless link scheduling under
physical interference model. In INFOCOM, pages 838–845. IEEE, 2011.
[43] P.-J. Wan, B. Xu, O. Frieder, S. Ji, B. Wang, and X. Xu. Capacity maximization in wireless
mimo networks with receiver-side interference suppression. In MobiHoc, pages 145–154, 2014.
[44] L. Wang, P. Wan, and W. Washington. Connectivity of multihop wireless networks with
log-normal shadowing. Wireless Networks, 21(7):2279–2292, 2015.
[45] Y. Wang, D. Yu, Q. Liu, and F. C. Lau. Inductive coloring: Implementing basic communication
primitives with Rayleigh-fading interference. In INFOCOM, pages 1–9. IEEE, 2016.
[46] M. Z. Zamalloa and B. Krishnamachari. An analysis of unreliability and asymmetry in lowpower wireless links. ACM TOSN, 3(2):7, 2007.
[47] J. Zander. Performance of optimum transmitter power control in cellular radio systems. IEEE
Trans. Veh. Technol, 41(1):57–62, 1992.
19
| 7cs.IT
|
arXiv:1706.05201v1 [cs.IT] 16 Jun 2017
THE MONTENEGRIN ACADEMY OF SCIENCES AND ARTS
PROCEEDINGS OF THE SECTION OF NATURAL SCIENCES, 22, 2017.
Conditions for Unique Reconstruction of Sparse
Signals Using Compressive Sensing Methods
Ljubiša Stanković, Miloš Daković, Srdjan Stanković, Irena Orović
Abstract
A signal is sparse in one of its representation domain if the number of nonzero coefficients in that domain is much smaller than the total number of coefficients. Sparse signals can be reconstructed from a
very reduced set of measurements/observations. The topic of this paper are conditions for the unique reconstruction of sparse signals from
reduced set of observations. After the basic definitions are introduced,
the unique reconstruction conditions are reviewed using the spark, restricted isometry, and coherence of the measurement matrix. Uniqueness of the reconstruction of signals sparse in the discrete Fourier domain (DFT), as the most important signal transformation domain, is
considered as well.
1
Introduction
A discrete-time signal can be transformed into other domains using different
signal transformations. Some signals that cover the whole considered interval in one domain could be sparse in a transformation domain, i.e., could
be located within a few nonzero coefficients. An observation or measurement is a linear combination of sparsity domain coefficients. Since the signal
samples are linear combinations of the signal transformation coefficients they
could be considered as the observations of a sparse signal in the transformation domain. Compressive sensing is a field dealing with a model for data
acquisition including the problem of sparse signal recovery from a reduced
set of observations [1–13]. A reduced set of observations can be a result of a
desire to sense a sparse signal with the lowest possible number of measurements/observations (compressive sensing). It can also be a result of a physical or measurement unavailability to take a complete set of observations [3].
The authors are with the University of Montenegro, 81000 Podgorica, Montenegro.
Conditions for Unique Reconstruction
2
In applications it could happen that some arbitrarily positioned samples of a
signal are so heavily corrupted by disturbances that it is better to omit them
and consider as unavailable in the analysis and to try to reconstruct the signal
with a reduced set of samples [14–16]. Although the reduced set of observations/samples appears in the first case as a result of user strategy to compress
the information, while in the next two cases the reduced set of samples is not
a result of user intention, all of them can be considered within the unified
framework. Under some conditions, a full reconstruction of a sparse signal
can be performed with a reduced set of observations/samples, as in the case
if a complete set of samples/observations were available [4, 17–20]. A priori information about the nature of the analyzed signal, i.e., its sparsity in
a known transformation domain, must be used in this analysis. Sparsity is
the main requirement that should be satisfied in order to efficiently apply the
compressive sensing methods for sparse signal reconstruction.
Compressive sensing methods are successfully applied to many fields,
including radar signal processing [21–26], time-frequency analysis [21,27–29],
L-statistics [16, 30], data hiding [31], communications [32], image processing
[33, 34], etc.
Topic of this paper are conditions for the unique reconstruction of sparse
signals from reduced set of observations/samples. The basic idea for unique
reconstruction will be introduced through an illustrative and simple example in the next section. Then the unique reconstruction condition will be
explained within the spark, restricted isometry, and coherence framework.
A special case of the signals sparse in the discrete Fourier domain (DFT), as
the most important signal transformation domain, will be considered at the
end. A simple uniqueness criterion will be presented and illustrated on an
example.
2
Illustrative Examples
Consider a large set of N numbers X (0), X (1),...,X ( N − 1). Assume that only
one of them is nonzero. We do not know either its position or its value. The
aim is to find the position and the value of this number. This case can be
related to many real life examples when we have to find one sample which
differs from other N − 1 samples. The problem can easily be reformulated to
the case when only one number differs from the expected and known value,
and all other assume their expected-known values.
The nonzero value at an position i will be denoted by X (i ). A direct way
to find the position i of nonzero sample would be to perform up to N mea-
2
3
ILLUSTRATIVE EXAMPLES
surements and compare each X (m) with zero. However, if N is very large
and there is only one nonzero sample we can get the result with just a few
observations/measurements. A procedure for the reduced number of observations/measurements is described next.
Take random numbers as weighting coefficients ai , i = 0, 1, 2, ..., N − 1,
for each coefficient. Measure the total value of all N weighted coefficients,
with weights ai . Since only one of them is different from the known expected
values mi (or from zero) we will get the total measured value
G = a1 m1 + a2 m2 + ... + ai (mi + X (i )) + ... + a N m N .
Next we will subtract the expected value GT = a1 m1 + a2 m2 + ... + a N m N
from G. The obtained observation/measurement, denoted by y(0), is
N −1
y ( 0 ) = G − GT =
∑
a k X ( k ) = a i X ( i ),
k =0
since the nonzero value in the space of X (0), X (1),...,X ( N − 1) is at one position only, X (k ) = X (i )δ(k − i ), k = 0, 1, . . . , N − 1.
As an illustration consider a set of N bags with coins. Assume that only
one bag contains false coins of a weight mi + X (i ). It is different from the
known weights mi of true coins in bag i. The goal is to find the position and
the difference in weight of false coins. From each of N bags we will take
ai , i = 1, 2, ...N, coins, respectively. Number of coins taken from the ith bag
is denoted by ai . The total measured weight of all coins from N bags is M,
Fig.1.
After the expected value is subtracted the observation/measurement y(0)
is obtained
N −1
y (0) =
∑
X (k )ψk (0),
(1)
k =0
where the weighting coefficients for this measurement are denoted by ψk (0) =
ak , k = 0, 1, ..., N − 1. In the space of unknowns (variables) X (0), X (1),...,X ( N −
1) this equation represents an N-dimensional hyperplane. We know that only
one unknown X (k) is nonzero at an unknown position k = i. The intersection of hyperplane (1) with any of the coordinate axes could be a solution
of our problem.
Assuming that a single X (k ) is nonzero, a solution will exist for any k.
Thus, one measurement would produce a set of N possible single nonzero
values equal to
X (k ) = y(0)/ψk (0),
ψk (0) 6= 0, k = 0, 1, 2, ..., N − 1.
Conditions for Unique Reconstruction
4
1
2
a1
N
3
a2
a3
aN
+
G−GT= a1 m1+a2 m2+...+ai (mi+X(i))+...+aN mN
−(a1 m1 +a2 m2 +...+ai mi+...+aN mN) = ai X(i)
i=?, X(i)=?
Figure 1: There are N bags with coins. One of them, at an unknown position, contains false coins. False coins differ from the true ones in mass for
unknown X (i ) = ∆m. The mass of the true coins in the ith bag is mi .
As expected, from one measurement we are not able to solve the problem and
to find the position and the value of nonzero sample.
For N = 3 possible solutions are illustrated with circles in Fig.2a), denoting intersections of measurements hyperplane with coordinate axes.
If we perform one more measurement y(1), with another set of weighting coefficients ψk (1), k = 0, 1, ..., N − 1, and get measured value y(1) =
X (i )ψi (1) the result will be a hyperplane
N −1
y (1) =
∑
X (k)ψk (1).
k =0
This measurement will produce a new set of possible solutions for each X (k )
defined by
X (k ) = y(1)/ψk (0), k = 0, 1, 2, ..., N − 1.
If these two hyperplanes (sets of solutions) produce only one common value
X (i ) = y(0)/ψi (0) = y(1)/ψi (1).
2
5
ILLUSTRATIVE EXAMPLES
Figure 2: The solution illustration for N = 3, K = 1, and various possible
cases: (a) Three possible solutions for one measurement plane. (b) Unique
solution for two measurement planes. (c) Two possible solutions for two
measurement planes.
then it is the solution of our problem.
In a matrix form these two measurements can be written as
X (0)
y (0)
ψ0 (0) ψ1 (0) ... ψN −1 (0)
X (1)
=
y (1)
ψ0 (1) ψ1 (1) ... ψN −1 (1)
...
X ( N − 1)
y = AX
where A is the matrix of coefficients (measurement matrix)
ψ0 (0) ψ1 (0) ... ψN −1 (0)
A=
ψ0 (1) ψ1 (1) ... ψN −1 (1)
and y are observations/measurements of sparse variable X.
Common value for two measurements X (i ) = y(0)/ψi (0) and X (i ) =
y(1)/ψi (1) is unique if
ψi (0)ψk (1) − ψi (1)ψk (0) 6= 0
for any i 6= k.
In order to prove this statement assume that two different solutions X (i )
and X (k ), for the case of one nonzero coefficient, satisfy the same measurement hyperplane equations
ψi (0) X (i ) = y(0), ψi (1) X (i ) = y(1)
Conditions for Unique Reconstruction
6
and
ψk (0) X (k ) = y(0),
ψk (1) X (k ) = y(1).
Then
ψi (0) X (i ) = ψk (0) X (k )
and
ψi (1) X (i ) = ψk (1) X (k ).
If we divide these two equations we get
ψi (0)/ψi (1) = ψk (0)/ψk (1)
or ψi (0)ψk (1) − ψi (1)ψk (0) = 0. This is contrary to the assumption that
ψi (0)ψk (1) − ψi (1)ψk (0) 6= 0.
The same conclusion can be made considering matrix form relations for
X (i ) and X (k ). If both of them may satisfy the same two measurements then
y (0)
y (1)
y (0)
y (1)
ψi (0) ψk (0)
ψi (1) ψk (1)
X (i )
0
ψi (0) ψk (0)
ψi (1) ψk (1)
0
X (k)
=
=
.
(2)
Subtraction of the previous matrix equations results in
ψi (0) ψk (0)
ψi (1) ψk (1)
X (i )
− X (k)
= 0.
For ψi (0)ψk (1) − ψi (1)ψk (0) 6= 0 follows X (i ) = X (k ) = 0. Therefore two
different nonzero solutions X (i ) and X (k) in this case cannot exist. This concludes the proof that the solution is unique if
ψi (0)ψk (1) − ψi (1)ψk (0) = det
ψi (0) ψk (0)
ψi (1) ψk (1)
6= 0
for any i 6= k. It also means that rank (A2 ) = 2 for any A2 being a 2 × 2
submatrix of the matrix of coefficients (measurement matrix) A.
Let us consider M measurements in this example. Since we have assumed
that only one coefficient X (i ) is nonzero it will satisfy all measurements
ψi (0) X (i ) = y(0), ψi (1) X (i ) = y(1), . . . , ψi ( M − 1) X (i ) = y( M − 1).
2
7
ILLUSTRATIVE EXAMPLES
The solution will not be unique if there is another coefficient X (k ), k 6= i
satisfying
ψk (0) X (k) = y(0), ψk (1) X (k ) = y(1), . . . , ψk ( M − 1) X (k ) = y( M − 1).
Then the corresponding coefficients of the measurement matrix satisfy
ψi (0)
ψ (1)
ψ ( M − 1)
= i
= ... = i
.
ψk (0)
ψk (1)
ψk ( M − 1)
In this case measurement matrix is
ψ0 (0)
ψ1 (0)
ψ
(
1
)
ψ1 (1)
0
A =
..
..
.
.
ψ0 ( M − 1) ψ1 ( M − 1)
...
...
..
.
ψ N −1 (0 )
ψ N −1 (1 )
..
.
...
ψ N −1 ( M − 1 )
The solution is not unique if any two columns are linearly dependent. The
uniqueness requires that all two column submatrices A2 of A are of rank 2.
The determinant for all A2T A2 is nonzero.
In numerical and practical applications we would not be satisfied, if for
example det(A2T A2 ) 6= 0 but det(A2T A2 ) = ε close to zero. In this case the theoretical condition for a unique solution would be satisfied, however the analysis and possible inversion would be highly sensitive to any kind of noise,
including quantization noise. Thus, a practical requirement is that the determinant is not just different from zero, but that it sufficiently differs from zero
so that an inversion stability and robustness to a noise is achieved. Inversion stability for a matrix B = A2T A2 is commonly described by the condition
number of matrix
λmax
cond {B} =
λmin
where λmax and λmin are the largest and the smallest eigenvalue of matrix
B. The inversion stability worsens as λmin approaches to zero (when λmin is
small as compared to λmax ). For stable and robust calculations a requirement
λmax
≤ 1+δ
λmin
is imposed, with a nonnegative constant δ being sufficiently small. In our
example this condition should hold for all submatrices A2 .
Conditions for Unique Reconstruction
8
As a next example consider a signal described by a weighted sum of K
harmonics from a set of possible oscillatory functions e j2πkn/N , k = 0, 1, 2,
..., N − 1,
x (n) = A1 e j2πk1 n/N + A2 e j2πk2 n/N + ... + AK e j2πkK n/N ,
with K N. In the DFT domain this signal will be sparse with X (k ) =
DFT { x (n)} having only few nonzero values at k = k i , i = 1, 2, ..., K. According to the sampling theorem the sampling of this kind of signals should
be adjusted to the maximal expected signal frequency k = max{k1 , k2 , ..., k K }.
For an arbitrary set of frequencies, it means that we should adjust sampling to
the maximal possible frequency k = N − 1 and to use the full set of N signal
values/measurements at n = 0, 1, 2, ..., N − 1 in order to avoid aliasing.
However, if we know that the signal consists of only K N functions
with unknown amplitudes, then regardless of their frequencies, the signal
can be fully reconstructed from a reduced set of samples. Samples can be
considered as weighted measurements of the sparse function X (k ),
N −1
y (0) = x ( n1 ) =
∑
X (k )ψk (n1 ),
k =0
with the weighting coefficients ψk (n1 ) = exp( j2πn1 k/N )/N. The previous
relation is the IDFT. Now a similar analysis like in the previous illustrative
example can be performed, assuming for example K = 1 or K = 2. We
can find the position and the value of nonzero X (k ) using just a few signal
samples y(i ).
This model corresponds to many signals in real life. For example, in the
Doppler-radar systems the speed of a radar target is transformed into a frequency of a sinusoidal signal [23, 24]. Since the returned signal contains only
one or just a few targets, the signal representing target velocity is a sparse
signal in the DFT domain. It can be reconstructed from fewer samples than
the total number of radar return signal samples N, Fig.3.
In signal processing the DFT as the domain of signal sparsity is commonly
used, since it plays the central role in engineering applications [3, 20]. Note
that in the compressive sensing theory random measurement matrices are
mainly used. The compressive sensing results and algorithms are used as a
tool to solve the problems involving sparse signals.
3
9
DEFINITIONS
60 X(k)
2 x(n)
Two target velocities
within 64 bins range
1
40
0
20
−1
0
Target velocities transformed
into a dense signal
−2
0
20
40
60 (a)
2 y(n)
0
60 Y(k)
1
20
40
60 (b)
DFT of y(n)
before reconstruction
40
0
20
−1
0
Mesurements of x(n)
−2
0
20
40
60 (c)
0
20
40
60 (d)
Figure 3: (a) Signal in the frequency domain, where it is sparse (for example, velocities of two targets in Doppler radar signal). (b) Signal in the time
domain, where it is dense. (c) Reduced set of measurements (samples) and
(d) its DFT before reconstruction, calculated using the available samples only.
Real parts of signals are presented.
3
Definitions
A big set of discrete-time data x (n), n = 0, 1, . . . , N − 1, with a large number
of samples N is considered. Its coefficients in a transformation domain are
denoted as
X = [ X (0), X (1), ..., X ( N − 1)] T ,
where T represents the transpose operation. We consider a signal to be sparse
in this transformation domain if the number of nonzero coefficients K is much
smaller than the number of the original signal samples N, i.e., if X (k) = 0 for
k ∈
/ K = {k1 , k2 , ..., k K } and K N. The number of nonzero coefficients is
commonly denoted by kXk0
kXk0 = card {K} = K,
where card {K} is the cardinality of set K. It is equal to the number of elements in K. It is called the `0 -norm (norm-zero) or the `0 -pseudo-norm of
vector X although it does not satisfy the norm properties.
Conditions for Unique Reconstruction
10
The observations/measurements are defined as linear combinations of
signal coefficients in the sparsity domain
N −1
y(m) =
∑
X (k )ψk (m),
(3)
k =0
where m = 0, 1, . . . , M − 1 is the measurement index and ψk (m) are the
weighting coefficients. The vector form of the measurement signal is denoted
by y
y = [y(0), y(1), ..., y( M − 1)] T .
The measurements defined by Eq.(3) can be written as a undetermined
system of M < N equations
y (0)
y (1)
..
.
ψ0 (0)
ψ0 (1)
..
.
=
y ( M − 1)
ψ1 (0)
ψ1 (1)
..
.
···
···
..
.
ψ0 ( M − 1) ψ1 ( M − 1) · · ·
ψ N −1 (0 )
ψ N −1 (1 )
..
.
ψ N −1 ( M − 1 )
X (0)
X (1)
..
.
X ( N − 1)
or using matrix notation
y = AX
where A is the measurement matrix of size M × N.
The fact that the signal is sparse with X (k ) = 0 for k ∈
/ K = {k1 , k2 , ..., k K }
is not included in the measurement matrix A since the positions of the nonzero
values are unknown. If the knowledge that X (k ) = 0 for k ∈
/ K were included
then a reduced system would be obtained as
y (0)
y (1)
..
.
y ( M − 1)
=
ψk1 (0)
ψk1 (1)
..
.
ψk2 (0)
ψk2 (1)
..
.
···
···
..
.
ψk1 ( M − 1) ψk2 ( M − 1) · · ·
ψkK (0)
ψkK (1)
..
.
ψkK ( M − 1)
X (k1 )
X (k2 )
..
.
X (k K )
with a reduced M × K measurement matrix AK defined as
y = AK XK .
(4)
This is an overdetermined system of equation, K < N. Matrix AK would
be formed if we assumed/knew the positions of nonzero samples k ∈ K.
It would follow from the measurement matrix A by omitting the columns
corresponding to the zero-valued coefficients in X.
3
11
DEFINITIONS
3.1
Common Measurement Matrices
Some common measurement matrices used in practical applications and theoretical considerations will be presented here.
Randomness of measurement matrices is a favorable property in compressive sensing and matrices with random elements are often used. The
most common is the measurement matrix with zero-mean unity variance
Gaussian distributed numbers as elements
1
φk (n) ∼ √ N (0, 1)
M
√
normalized with 1/ M so that the energy of each column is one.
In signal processing the most common transform is the DFT. The coefficients of its direct transform matrix Φ are defined as
φk (n) = exp(− j2πnk/N ).
The inverse DFT matrix coefficients are ψk (n) = N1 exp( j2πnk/N ). Commonly the measurements are the signal samples y(m − 1) = x (nm ) for m =
1, . . . , M where
nm ∈ M = {n1 , n2 , . . . , n M } ⊂ {0, 1, . . . , N − 1},
and
y ( m − 1) = x ( n m ) =
1
N
N −1
∑
X (k )e j2πnm k/N .
k =0
Therefore, the measurement matrix is obtained by keeping the rows of the
inverse DFT matrix corresponding to the samples at nm ∈ {0, 1, . . . , N − 1},
for the measurements m = 1, 2, . . . , M,
e j2πn1 /N
e j2πn2 /N
..
.
···
···
..
.
1 e j2πn M /N
···
1
1 1
A=
.
N ..
e j2πn1 ( N −1)/N
e j2πn2 ( N −1)/N
.
..
.
(5)
e j2πn M ( N −1)/N
This is a partial inverse DFT matrix. In compressive sensing theory it is common to normalize the measurement matrix so that the energy of its columns
(diagonal elements of AH√
A matrix) is equal to one. Then the factor 1/N in A
should be replaced by 1/ M.
12
Conditions for Unique Reconstruction
In order to increase randomness in the Fourier transform matrix, the measurements may be taken at any random instant. Then the measurement vector elements are y(m − 1) = x (tm ) where tm , m = 1, 2, . . . , M are random
instants within the considered time interval T. The measurement matrix follows then from the Fourier series definition x (t) = ∑kN=−01 X (k ) exp( j2πkt/T ).
It has been assumed that the Fourier series coefficients are within 0 ≤ k ≤
N − 1. The measurements matrix is
1 e j2πt1 /T · · · e j2πt1 ( N −1)/T
1 e j2πt2 /T · · · e j2πt2 ( N −1)/T
(6)
A = .
..
..
..
..
.
.
.
1 e j2πt M /T · · · e j2πt M ( N −1)/T
√
with a possible normalization factor 1/ M. This measurement matrix is a
partial random inverse Fourier transform matrix.
4
Reconstruction Problem Formulation
The signal can be reconstructed from its measurements defined by vector y by
finding the sparsest vector X that corresponds to the measurements y. Hence,
by introducing the notation for the number of components based on the `0 norm K = kXk0 , the fundamental minimization problem can be formulated
as:
min kXk0 subject to y = AX.
(7)
In general, the `0 -norm is not very suitable for most minimization methods. However, a class of algorithms is based on the minimization of the number of coefficients K = kXk0 in an implicit way. For instance, in certain applications we can predict the number of components or we are able to estimate
the position of non-zero coefficients. Thus, compared to the direct search
method, the computational complexity will be significantly reduced if we are
able to estimate positions of nonzero coefficients and solve the problem with
the minimal possible number of nonzero coefficients. The important class of
this algorithms are matching pursuit (MP) algorithms.
Minimization of the number of non-zero coefficients using the `0 -norm
is a nonconvex optmization problem that cannot be solved using well developed iterative algorithms and linear programming methods [35–42]. To
avoid dealing with NP-hard problems, significant efforts have been undertaken to replace the nonconvex and discontinuous `0 -norm with a convex
and continuous norm that would be more appropriate for optimization. As
4
RECONSTRUCTION PROBLEM FORMULATION
13
a result, the `1 -norm (norm-one) has been commonly employed in many signal reconstruction approaches [43–46]. It has been shown that, under certain
conditions, minimization of the `1 -norm produces the same solution as the
minimization of the `0 -norm.
In the `1 -norm based reconstructions the problem is formulated as
min kXk1
subject to y = AX
where
N −1
k X k1 =
∑
| X (k)| .
k =0
This is the so-called basis pursuit (BP) approach to sparse signal reconstruction.
4.1
4.1.1
Conditions for Unique Reconstruction
Spark
The spark of a matrix A is defined as the smallest number of linearly dependent columns of A. In other words if spark{A} = K, then any collection of
K1 < K columns of A are linearly independent.
Spark can also be defined as a minimal number of nonzero entries in a
vector X 6= 0 such that AX = 0
min kXk0
such that AX = 0
If matrix A is of size N × K with N > K and spark{A} = K + 1 then all
K × K submatrices of matrix A are nonsingular, i.e., with nonzero determinant.
The analysis of a signal with an arbitrary sparsity K is similar to the analysis for K = 1. To get the first set of possible solutions for K nonzero coefficients (of sparsity K) we need K measurements. For any combination of K
(out of N) nonzero coefficients X (k ), k ∈ {k1 , k2 , ..., k K }, we will get a possible
solution. There exist ( N
K ) such possible combinations/solutions. Additional K
measurements will be used to produce another set of ( N
K ) possible solutions.
The intersection of these two sets is then the solution of our problem.
Consider the case when the number of measurements M is twice higher than the
sparsity K, M = 2K. The K-sparse solution is unique if the determinants of all A2K
submatrices of matrix A are different from zero.
This statement will be proven by contradiction. Assume that M = 2K
measurements are available within the vector y. Assume that two different
Conditions for Unique Reconstruction
14
solutions for X of sparsity K exist. Denote the nonzero parts of the solutions
(1)
(2)
by XK and XK . Both of them satisfy the measurements equation,
(1) (1)
AK XK = y
and
(2) (2)
AK XK = y,
(1)
(2)
where AK and AK are two different submatrices of matrix A of size M × K
(1)
(2)
corresponding to the elements in XK and XK . If we rewrite these equations
by adding zeros
"
#
"
#
i 0
i (1)
h
h
K
XK
(2)
(1)
(2)
(1)
(8)
=y
and
AK AK
AK AK
(2) = y
XK
0K
and subtract them we get
h
(1)
AK
(2)
AK
i
"
(1)
#
(1)
XK
(2) = 0.
−XK
(9)
(2)
There are
i solutions for XK and XK if the determinant of matrix
h no nonzero
A2K = A(K1) A(K2) is nonzero. If all possible submatrices A2K (including
all lower order submatrices) of measurement matrix A are nonsingular then
two solutions of sparsity K cannot exist, and the solution is unique. Note that
N
there are (2K
) submatrices A2K .
Based on the previous analysis, the solution for a K sparse problem is
unique if
spark{A} > 2K.
h
i
(1)
(2)
For M > 2K the matrix A2K = AK AK dimension is M × 2K. Again
T A ) = 2K system (9) does not have a nonzero soluif rank(A2K ) = rank(A2K
2K
T A ) =
tion. It means that the reconstruction is unique. If rank(A2K ) = rank(A2K
2K
2K for all submatrices A2K then spark{A} > 2K.
If the vector X is of sparsity K, with kXk0 = K then if
K<
1
spark {A}
2
the solution X is unique.
In order to prove this statement (that has been already explained) consider a measurement matrix A whose spark is spark {A}. Then for a sparse
4
RECONSTRUCTION PROBLEM FORMULATION
15
vector X of sparsity K = spark {A} obviously there exists such a combination
of nonzero elements in X so that they coincide with the dependent columns.
Then we can obtain
AX = 0.
This property is used for the spark definition as well.
Note that for any X of sparsity K < spark {A} the relation AX = 0 will
not hold, since corresponding independent columns of A multiplied with
nonzero elements of X cannot produce a zero result. Since K < spark {A} it
means that all sets of K columns from A are independent.
The proof of the previous statement will be based on the contradiction.
Assume that X is a solution of AX = y and that its sparsity satisfies K <
1
2 spark { A }. Assume also that there is another solution H such that AH = y
and H is also sparse with sparsity lower than 12 spark {A}. Since
AH = AX = y
A(H − X) = 0
then
spark {A} < kH − Xk0
or
spark {A} < kH − Xk0 ≤ kHk0 + kXk0
spark {A} − kHk0 ≤ kXk0 .
The inequality follows from the fact that two nonzero elements, at the same
position in H and X, can produce a zero element in H − X, while two zero elements in these vectors cannot produce a nonzero element in H − X. If there is
another solution H such that kHk0 < 12 spark {A} then from the last inequality follows that kXk0 > 12 spark {A} . This is a contradiction to the assumption
that both solutions H and X have sparsity lower than 12 spark {A}.
4.1.2
Restricted Isometry Property
Note that for any square matrix its determinant is equal to the product of its
eigenvalues
T
det{A2K
A2K } = λ1 λ2 · . . . · λ2K .
T A
The condition that the solution is unique if the determinant of A2K
2K for all
A2K submatrices of matrix A are different from zero can be rewritten as
min |λi | > 0.
i
Conditions for Unique Reconstruction
16
In numerical and practical applications we would not be satisfied, if any
of the determinants is very close to zero. In this case the theoretical condition
for a unique solution would be satisfied, however the analysis and possible
inversion would be highly sensitive to any kind of noise in measurements.
Thus, a practical requirement is that the determinant is not just different from
zero, but that it sufficiently differs from zero so that an inversion stability and
noise robustness is achieved.
From the matrix theory it is known that the norm of a matrix A2K satisfies
λmin ≤
T AT A X
X2K
kA2K X2K k22
2K 2K 2K
=
≤ λmax ,
2
T
X
kX2K k2
2K X2K
(10)
where λmin and λmax are the minimal and the maximal eigenvalue of the ma2
T A
2
2
trix A2K
2K and k X k2 = | X (0)| + . . . | X ( N − 1)| is the squared `2 -norm
(norm-two).
The isometry property for a linear transformation matrix A holds if
kAXk22 = kXk22 or
kAXk22
= 1.
kXk22
The restricted isometry property (RIP) for a matrix A2K and a 2K-sparse
vector X2K holds if
1 − δ2K ≤
kA2K X2K k22
≤ 1 + δ2K ,
kX2K k22
(11)
where 0 ≤ δ2K < 1 is the isometric constant. From Eqs.(10) and (11) we can
write
δ2K = max{1 − λmin , λmax − 1}.
Commonly, isometric constant is defined by λmax − 1 and it is calculated
T A
as maximal eigenvalue of matrix A2K
2K − I. Normalized energies of the
T A ) are assumed. Othercolumns of matrix A (diagonal elements of A2K
2K
wise, the normalization factors should be added. For complex-valued matriH A .
ces Hermitian transpose should be used in A2K
2K
For a K-sparse vector X and a measurement matrix A the RIP is satisfied if
relation (11) holds for all submatrices AK with 0 ≤ δK < 1. The solution for
K-sparse vector is unique if the measurement matrix satisfy the RIP for 2K-sparse
vector X with 0 ≤ δ2K < 1.
Note that if the RIP is satisfied then λmin > 0. It means that there is no
H A
A2K submatrix of A such that A2K
2K is a singular matrix. The uniqueness
proof reduces to the previous one.
4
17
RECONSTRUCTION PROBLEM FORMULATION
Restricted isometry property for small δ2K is closer to the isometry property and improves the solution stability. It can be related to the matrix condiT A
tional number. The conditional number of a matrix A2K
2K is defined as the
ratio of its maximal and minimal eigenvalues
n
o
λmax
T
cond A2K
A2K =
.
λmin
If a matrix A2K satisfies the restricted isometry property with δ2K then
n
o
1 + δ2K
T
cond A2K
A2K ≤
.
1 − δ2K
With small values of δ2K the conditional number is close to one, meaning
stable invertibility and low sensitivity to the input noise (small variations of
the input signal (measurements) do not cause large variations of the result).
Common requirement for this constant is
√
0 ≤ δ2K < 2 − 1.
The restricted isometry constant within this range will also guarantee the
equivalence of the solutions obtained in the reconstruction based on the `0 norm and the `1 -norm minimization [17, 18].
4.1.3
Coherence
The mutual coherence (coherence index) of a matrix A is defined as the maximal absolute value of the normalized scalar product of its two columns
µ = max |µmk | , for m 6= k
where
µmk = q
−1
∗
∑iM
=0 α m ( i ) α k ( i )
2
−1
M −1
∑iM
=0 | αm (i )| ∑i =0 | αk (i )|
2
(12)
2
−1
and αk (i ) are the elements of the kth column of matrix A. If ∑iM
=0 | αk (i )| =
2
−1
∑iM
=0 | αm (i )| then
1
∗
∑ M−
0 αm (i ) αk (i )
µmk = i=M
.
(13)
2
1
∑i=−
0 | αk (i )|
Note that µmk , m 6= k, are the off-diagonal elements of matrix A H A normalized with the corresponding diagonal elements.
Conditions for Unique Reconstruction
18
This index plays an important role in the analysis of measurement matrices. The coherence index should be as small as possible, or in other words the
incoherence is a desirable property. With smaller values of coherence index
the matrix A H A is closer to the identity matrix.
The condition that all eigenvalues of matrix A H A are nonzero can be written in terms of the coherence index. In general, the eigenvalue relation for
matrix A H A reads
(A H A)u = λu
where u denotes an eigenvector. For an eigenvector we can always choose
that its maximal coordinate is um = maxk (uk ) = 1 and uk ≤ 1 for k 6= m.
Now we can write the general eigenvalue relation in the form
∑ µmk uk = λum = λ
k
or
∑
µmk uk = λ − µmm
k,k 6=m
From this relation we can conclude
|λ − µmm | ≤
∑
|µmk uk | ≤
k,k 6=m
∑
|µmk |.
k,k 6=m
Considering eigenvalue λ as a variable and µmk as constants we conclude
that the last inequality describes a disc area with the center at µmm and radius
∑k,k6=m |µmk |. It does not include point λ = 0 if
µmm >
∑
|µmk |.
(14)
k,k 6=m
Therefore the matrix A H A will be nonsingular if the above condition is met.
This is the Gershgorin circle (disk) theorem.
For normalized matrix A H A we have µmm = 1 and
µ = max |µmk | .
m6=k
We have already concluded that the solution for K-sparse vector will be
H A
unique if for all possible submatrices A2K the matrices A2K
2K are nonsinH
gular. Note that the off-diagonal elements of A2K A2K are a subset of the offdiagonal elements of matrix A H A. The same holds for the diagonal elements.
It mean that the coherence µ of matrix A will be always greater than or equal
to the coherence of any submatrix A2K .
4
19
RECONSTRUCTION PROBLEM FORMULATION
H A , and unique solution for
The nonsingularity condition for all matrices A2K
2K
a K sparse vector X, is achieved if
1 > (2K − 1)µ
or
K<
1
1
(1 + ).
2
µ
The proof follows from (14) for normalized matrix A H A. The inequality
2K
1 = µmm >
∑
|µmk |
k=1,k 6=m
is satisfied if 1 > (2K − 1)µ since ∑2K
k =1,k 6=m | µmk | < (2K − 1) µ.
The coherence index cannot be arbitrarily small for an M × N matrix A
(M < N). The Welch upper bound relation holds
s
N−M
.
(15)
µ≥
M ( N − 1)
The Gershgorin circle (disk) theorem can be used to determine the spark
lower bound. If the relation K < (1 + 1/µ)/2 holds for a given K then it
H A
holds for any order lower than K. It means that all submatrices A2K
2K of
2K and lower order are nonsingular. Therefore the spark of such a matrix is
greater than 2K
spark(A) > 2K
(16)
or
K<
1
spark(A)
2
if K < (1 + 1/µ)/2. It means that
spark(A) ≥ 1 +
4.2
1
.
µ
Numerical Example
Consider 5 × 8 measurement matrix
0.1
0.1
0.3 −0.7
0.7
0.4 −0.8 −0.4 −0.1
0.3
0.3
0.5
−
0.5
0.4
0.5
A=
−0.7 −0.3
0.1
0.3
0.4
−0.5
0.1 −0.7 −0.5 −0.1
−0.1
0.1
0.3
0.3
−0.7
0.1
−0.5
0.5
−0.4 −0.8
0.3
−0.5
−0.4
−0.7
−0.1
.
(17)
Conditions for Unique Reconstruction
20
Columns of this matrix are normalized. The norm of each column is equal
to one. The matrix dimensions are small so we can solve NP-hard problems
and calculate the spark and the restricted isometry property constants, by
checking all possible combinations, in a reasonable time.
4.2.1
Spark Calculation
For the spark calculation we first concluded that there is no all zero column,
meaning that spark{A} > 1. Then we have check linear dependence of each
pairs of columns. In total (82) combinations are checked by calculating rank
of each 2 × 5 submatrix. In all cases rank was equal to 2 meaning that all
pair of two columns are linearly independent. It means that spark{A} > 2.
Next all (83) possible combinations of three columns are considered. For all
submatrices we concluded that their rank is 3 meaning that there is no a set
of three linearly dependent columns in the measurement matrix A. Therefore spark{A} > 3. Calculation is repeated for all combinations of four and
five columns with the same result. The final conclusion is that the spark of
this matrix is spark{A} = 6 meaning that all combinations of five and less
columns are linearly independent. The uniqueness condition based on matrix
spark state that sparsity K limit is
K<
1
spark{A} = 3.
2
According to the spark based uniqueness condition, the reconstruction is
unique for K = 1 and K = 2. We may conclude that if we find a sparse
vector X in the reconstruction with sparsity K ≤ 2 then this is the sparsiest
possible solution of our problem.
4.2.2
Coherence Calculation
Coherence of the considered matrix is calculated a maximal absolute value of
the off-diagonal element of A T A. If the diagonal elements were not normalized then this maximal value should be normalized with the diagonal values
of this matrix. For the considered measurement matrix A we get
µ = 0.49
resulting in the sparsity limit
K<
1
1
(1 + ) ≈ 1.5204.
2
µ
4
RECONSTRUCTION PROBLEM FORMULATION
21
The worst case, that determine value of µ, was coherence between 5th and
7th column of the considered matrix. The reconstruction is unique only for
K = 1. Note that in contrast to the spark limit this condition will guarantee
that the same unique solution is obtained using `1 -norm and `0 -norm. This is
the reason why this limit is more conservative.
For a measurement matrix of order 5 × 8 the smallest possible value of the
the coherence index is
s
r
N−M
3
µ≥
=
= 0.2928
M ( N − 1)
5×7
with maximal possible bound K < 2.2078. The matrix with minimal coherence index is quite specific and it is called the equiangular tight frame (ETF).
In practice many optimization approaches are based on finding the measurement matrix with coherence as low as possible (as close to the ETF as possible).
4.2.3
Restricted Isometry Property Constant Calculation
Restricted isometry property (RIP) constants of orders 1, 2, 3, 4, and 5 are
calculated. For the calculation of the RIP constant with assumed sparsity
K = 1 all possible submatrices A1 are formed. These are 5 × 1 matrices.
There are 8 of them. The matrices A1T A1 are formed. All of them are scalars
equal to 1 with λ = 1, resulting in
δ1 = max{1 − λmin , λmax − 1} = 0.
Next the sparsity K = 2 of the resulting X is assumed. All possible measurement submatrices A2 corresponding to this sparsity are formed. There
are (82) of them. The matrices A2T A2 are formed. Then their eigenvalues are
calculated. The RIP constant δ2 is obtained as a maximal value of
δ2 = max{1 − λmin , λmax − 1} = 0.49.
over all possible submatrices A2 .
The calculation is repeated for assumed sparsity K = 3, 4, 5 by forming
corresponding submatrices A3 , A4 , and A5 , respectively. The obtained numerical values for these sparsities are
δ3 = 0.9406
δ4 = 1.2063
δ5 = 1.3368
22
Conditions for Unique Reconstruction
We can conclude that matrix A satisfy the restricted isometry property
0 ≤ δK < 1
for sparsity 1, 2, and 3. The uniqueness condition require that for sparsity
K measurement matrix satisfies restricted isometry property for 2K meaning
that the uniqueness is guarantied only for K = 1. For K = 2 the condition
should be satisfied for δ4 , what is not the case.
The minimization for K sparse vector X √
using `1 -norm will produce the
same result as if `0 -norm were used if δ2K < 2 − 1. It means that there is no
guarantee that `1 norm minimization could be used in the reconstruction for
sparsity K = 1. Note that different bounds have been derived in literature
for this equivalence. One of the derived bounds is that δ2K < 0.493. The
considered measurement matrix A would produce a unique solution with `1 norm based minimization, according to this bound since δ2 = 0.49 < 0.493.
From this example we can see that uniqueness conditions produce different limits, and that they are very restrictive.
4.3
Uniqueness of the DFT of Sparse Signals
In general, the reconstructed signal uniqueness is guarantied if the restricted
isometry property is used and checked. However, two problems exist in the
implementation of this approach. For a specific measurement matrix it produces quite conservative bounds. In addition, uniqueness check with the
restricted isometry property requires a combinatorial approach, which is an
NP hard problem.
In some reconstruction methods the missing measurements are considered as the minimization variables. The available measurements/samples
are known and fixed. The number of variables in the minimization process
is equal to the number of missing samples/measurements in the observation
domain. This approach is possible when the common signal transforms are
the domains of signal sparsity [20, 48–54]. Then the missing and available
samples/measurements form a complete set of samples/measurements.
The DFT is such a signal sparsity domain. The solution uniqueness is
defined in the sense that the variation of the missing sample values cannot
produce another signal of the same sparsity. In the case when the signal is
already reconstructed then the uniqueness is checked in the sense that there
is no other signal of the same or lower sparsity with the same set of available
samples [20].
Consider a signal x (n) with n ∈ N = {0, 1, 2, ...., N − 1}. Assume that Q
of its samples at the positions qm ∈ NQ = {q1 , q2 , ...., qQ } are missing/omit-
4
RECONSTRUCTION PROBLEM FORMULATION
23
ted. The signal is sparse in the DFT domain, with sparsity K. The reconstruction goal is to get x (n), for all n ∈ N using available samples at n ∈ M =
N\NQ . A new signal of the form
x a (n) = x (n) + z(n)
will be analyzed here. For the available signal positions n ∈ M the value of
z(n) is fixed z(n) = 0, while z(n) may take arbitrary value at the positions of
missing samples n = qm ∈ NQ = {q1 , q2 , ...., qQ }. If x (n) is a K sparse signal
then the DFT of x a (n) is
Xa (k ) = X (k ) + Z (k )
K
= N ∑ Ai δ(k − k0i ) +
i =1
Q
∑
z(qm )e− j2πqm k/N .
m =1
Positions of nonzero values in X (k ) are k0i ∈ K = {k01 , k02 , ...., k0K } with amplitudes X (k0i ) = N Ai . The values of missing samples of x a (n) = x (n) + z(n)
for n ∈ NQ are considered as variables. The goal of reconstruction process is
to get x a (n) = x (n), or z(n) = 0 for all n ∈ N. This goal should be achieved
by minimizing a sparsity measure of the signal transform Xa (k ) [3, 47]. Existence of the unique solution of this problem depends on the number of missing samples, their positions, and the signal form.
If a signal with the transform X (k ) of sparsity K is obtained using a reconstruction method, with a set of missing samples, then the reconstruction X (k )
is unique if there is no other signal of the same or lower sparsity that satisfies
the same set of available samples (using the same set of missing samples as
variables).
Consider a signal x (n) that is sparse in the DFT domain with unknown sparsity.
Assume that the signal length is N = 2r samples and that Q samples are missing
at the instants qm ∈ NQ . Assume that the reconstruction is performed and that the
DFT of reconstructed signal is of sparsity K. The reconstruction result is unique if
the inequality
n
o
K< N−
max
h=0,1,...,r −1
2 h ( Q 2h − 1 ) − K
holds. Integers Q2h are calculated as
Q 2h =
max
b=0,1,...,2h −1
{card{q : q ∈ NQ and mod(q, 2h ) = b}}
(18)
For example, consider a signal with N = 25 = 32 and Q = 9 missing
samples at
qm ∈ NQ = {2, 3, 8, 13, 19, 22, 23, 28, 30}.
Conditions for Unique Reconstruction
24
Using the presented we will find the sparsity limit K when we are able to
claim that the reconstructed sparse signal is unique for any signal form.
• For h = 0 we use Q20 = Q and get 20 ( Q20 − 1) − 1 = ( Q − 1) − 1 = 9.
• For h = 1, the number Q21 is the greater value of
card{q : q ∈ NQ and mod(q, 2) = 0} = card{2, 8, 22, 28, 30} = 5
card {q : q ∈ NQ and mod(q, 2) = 1} = card{3, 13, 19, 23} = 4,
i.e., the maximal number of even or odd positions of missing samples.
Thus Q21 = max {5, 4} = 5 with 21 ( Q21 − 1) = 8.
• Next Q22 is calculated as the maximal number of missing samples whose
distance is multiple of 4. For various initial counting positions b =
0, 1, 2, 3 the numbers of missing samples with distance being multiple
of 4 are 2, 1, 3, and 3, respectively. Then Q22 = max {2, 1, 3, 3} = 3 with
22 ( Q2h − 1) = 8.
• For Q23 the number of missing samples at distances being multiple of
8 are found for various b = 0, 1, 2, 3, 4, 5, 6, 7. The value of Q23 is 2 with
23 ( Q23 − 1) = 8.
• Finally we have two samples at distance 16 (samples at the positions
q2 = 3 and q5 = q2 + N/2) producing Q24 = Q16 = 2 with 24 (2 − 1) =
16.
The reconstructed signal of sparsity K is unique if
K< N−
max
h=0,1,2,3,4
n
o
2 h ( Q2h − 1 ) − K
K < 32 − max {9, 8, 8, 8, 16} − K
K < 32 − 16 − K
or
K < 8.
An extended discussion about the DFT uniqueness, within the framework of
the missing samples as variables, can be found in [20].
5
5
CONCLUSION
25
Conclusion
Sparse signals can be reconstructed from a very reduced set of observations,
through compressive sensing. This property has found applications in many
fields. The topic of this paper was to introduce the basic definitions in compressive sensing. The conditions for exact and unique reconstruction of original signals are of crucial importance in theory and applications. These conditions are reviewed and related in this paper.
Appendix
MATLAB R functions for spark calculation (Algorithm 1), restricted isometry
constant calculation (Algorithm 2) and uniqueness test for partial DFT matrix
(Algorithm 4) are provided. Auxiliary function nextcomb used for generation
of all possible columns combinations of the measurement matrix used in Algorithms 1 and 2 is given in Algorithm 3.
Algorithm 1 Measurement matrix spark calculation
1 function s = spark(A)
2 % Matrix spark calculation
3 [M,N] = size(A);
4 s = M+1;
5 for k = 1:M
6
kk = nchoosek(N,k);
7
p = 1:k;
8
for m = 1:kk
9
A1 = A(:,p);
10
if rank(A1) < k
11
s = k; break
12
end
13
p = nextcomb(p,N);
14
end
15
if s < M+1, break, end
16 end
Conditions for Unique Reconstruction
26
Algorithm 2 Restricted isometry constant calculation
1
2
3
4
5
6
7
8
9
10
11
12
function d = RIP_calc(A,K)
% Restricted isometry constant caluclation
[M,N] = size(A);
d = 0;
kk = nchoosek(N,K);
p = 1:K;
for m = 1:kk
A1 = A(:,p);
l = eig(A1’∗A1);
d = max([d, 1−min(l), max(l)−1]);
p = nextcomb(p,N);
end
Algorithm 3 Auxiliary function for generation of all combinations
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
function p = nextcomb(p,N)
% Generate next combination (in lexicographical order )
% Input: previous combination p and number of elements N
% Output: next combination or []
i = length(p);
K = N;
while i > 0 && p(i) == K
i = i −1;
K = K−1;
end
if i > 0
p(i ) = p(i )+1;
for k = i+1:length(p)
p(k) = p(k−1)+1;
end
else
p = [];
end
REFERENCES
27
Algorithm 4 Sparsity limit for partial DFT measurement matrix
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
function Kt = DFT_check(N,Nq)
% Sparsity limit for partial DFT matrix
% Inputs:
% N − total number of samples (must be power of two)
% Nq − set of missing sample positions
r = log2(N);
if r −round(r) ~= 0, error(’N must be power of two’), end
Kt = N;
for h = 0: r −1
p = rem(Nq,2^h);
Q = zeros(1,2^h);
for s = 0:2^h−1
Q(s+1) = sum(p==s);
end
Q2h = max(Q);
if Kt > (N−2^h∗(Q2h−1))/2
Kt = (N−2^h∗(Q2h−1))/2;
end
end
References
[1] D. L. Donoho, “Compressed sensing,” IEEE Transactions on Information
Theory, vol. 52, no. 4, 2006, pp. 1289–1306.
[2] E.J. Candès, M. Wakin, "An Introduction to Compressive Sampling",
IEEE Signal Processing Magazine, vol. 25, no. 2, March 2008, pp. 21-30.
[3] L. Stanković, M. Daković, S. Stanković, and I. Orović, "Sparse Signal
Processing - Introduction ," Wiley Encyclopedia of Electrical and Electronics
Engineering, John Wiley & Sons, 2017.
[4] E. J. Candès, J. Romberg, T. Tao, “Robust uncertainty principles: Exact
signal reconstruction from highly incomplete frequency information,”
IEEE Transactions on Information Theory, vol. 52, no. 2, 2006, pp. 489-509.
[5] Y. C. Eldar, G. Kutyniok, Compressed Sensing: Theory and Applications,
Cambridge University Press, 2012.
28
Conditions for Unique Reconstruction
[6] R. Baraniuk, “Compressive sensing,” IEEE Signal Processing Magazine,
vol. 24, no. 4, 2007, pp. 118–121.
[7] L. Stanković, S. Stanković, M. G. Amin, “Missing Samples Analysis in
Signals for Applications to L-Estimation and Compressive Sensing”, Signal Processing, Elsevier, vol. 94, Jan. 2014, pp. 401–408.
[8] L. Stanković, Digital Signal Processing with Applications: Adaptive Systems,
Time-Frequency Analaysis, Sparse Signal Processing, CreateSpace Independent Publishing Platform, 2015.
[9] S. Stanković, I. Orović, E. Sejdić, Multimedia Signals and Systems: Basic and
Advanced Algorithms for Signal Processing, Springer International Publishing, second edition, 2015.
[10] S. Stanković, L. Stanković, and I. Orović, “A Relationship between the
Robust Statistics Theory and Sparse Compressive Sensed Signals Reconstruction," IET Signal Processing, 2014.
[11] I. Orovic, and S. Stankovic, "Improved Higher Order Robust Distributions based on Compressive Sensing Reconstruction," IET Signal Processing, vol. 8, no. 7, 2014, pp. 738 - 748.
[12] M. Elad, Sparse and Redundant Representations: From Theory to Applications
in Signal and Image Processing, Springer, 2010.
[13] I. Orović, V. Papić, C. Ioana, X. Li, and S. Stanković, “Compressive Sensing in Signal Processing: Algorithms and Transform Domain Formulations," Mathematical Problems in Engineering, 2016.
[14] L. Stankovic, M. Dakovic, S. Vujovic, "Reconstruction of Sparse Signals
in Impulsive Disturbance Environments," Circuits, Systems and Signal
Processing, vol. 2016. pp. 1-28.
[15] I. Stankovic, I. Orovic, M. Dakovic, and S. Stankovic, "Denoising of
Sparse Images in Impulsive Disturbance Environment," Multimedia Tools
and Applications, pp.1-21, 2017.
[16] S. Stankovic, I. Orovic, and M. Amin, "L-statistics based Modification
of Reconstruction Algorithms for Compressive Sensing in the Presence
of Impulse Noise," Signal Processing, vol.93, no.11, November 2013, pp.
2927-2931.
REFERENCES
29
[17] E. J. Candès, "The restricted isometry property and its implications for
compressed sensing", Comptes Rendus Mathematique, Vol.346, Issues 9-10,
May 2008, pp 589-592.
[18] T. Zhang, "Sparse Recovery with Orthogonal Matching Pursuit Under
RIP," IEEE Trans. on Information Theory, 57(9), 2011, pp. 6215-6221.
[19] D. L. Donoho, M. Elad, V. Temlyakov, “Stable recovery of sparse overcomplete representations in the presence of noise,” IEEE Transactions on
Information Theory, vol. 52, 2006, pp. 6–18.
[20] L. Stanković, M. Daković, "On the Uniqueness of the Sparse Signals Reconstruction Based on the Missing Samples Variation Analysis," Mathematical Problems in Engineering, vol. 2015, Article ID 629759, 14 pages,
doi:10.1155/2015/629759
[21] L. Stankovic, M. Dakovic, T. Thayaparan, Time-Frequency Signal Analysis
with Applications, Artech House, Boston, March 2013.
[22] L. Stankovic, "On the ISAR Image Analysis and Recovery with Unavailable or Heavily Corrupted Data," IEEE Transactions on Aerospace and Electronic Systems, Vol.51, no.3, pp.2093-2106, July 2015
[23] S. Costanzo, A. Rocha, M. D. Migliore, "Compressed Sensing: Applications in Radar and Communications", The Scientific World Journal, Volume 2016, Article ID 5407415, 2 pages, Editorial, 2016.
[24] J. Ender, "On compressive sensing applied to radar", Signal Processing,
Volume 90, Issue 5, May 2010, Pages 1402-1414.
[25] L. Stankovic, S. Stankovic, T. Thayaparan, M. Dakovic, and I. Orovic,
"Separation and Reconstruction of the Rigid Body and Micro-Doppler
Signal in ISAR Part II - Statistical Analaysis," IET Radar, Sonar and Navigation, vol.9, no.9, pp.1155-1161, 2015 doi: 10.1049/iet-rsn.2015.0137
[26] L. Stankovic, S. Stankovic, T. Thayaparan, M. Dakovic, and I. Orovic,
"Separation and Reconstruction of the Rigid Body and Micro-Doppler
Signal in ISAR Part I- Theory," IET Radar, Sonar and Navigation, vol.9,
no.9, pp.1147-1154, 2015 doi: 10.1049/iet-rsn.2015.0096
[27] P. Flandrin, P. Borgnat, "Time-Frequency Energy Distributions Meet
Compressed Sensing," IEEE Transactions on Signal Processing, vol.58,
no.6, pp.2974, 2982, June 2010.
30
Conditions for Unique Reconstruction
[28] L. Stankovic, I. Orovic, S. Stankovic, and M. Amin, "Compressive Sensing Based Separation of Non-Stationary and Stationary Signals Overlapping in Time-Frequency," IEEE Transactions on Signal Processing, Vol. 61,
no. 18, pp. 4562 - 4572, Sept. 2013.
[29] I. Orović, S. Stanković, and T. Thayaparan, "Time-Frequency Based Instantaneous Frequency Estimation of Sparse Signals from an Incomplete
Set of Samples," IET Signal Processing, Special issue on Compressive Sensing
and Robust Transforms, 2014.
[30] S. Stankovic, I. Orovic, M. Amin, "L-statistics based Modification of Reconstruction Algorithms for Compressive Sensing in the Presence of Impulse Noise," Signal Processing, vol.93, no.11, November 2013, pp. 29272931 , 2013.
[31] G. Hua, Y. Hiang, G. Bi, "When Compressive Sensing meets Data Hiding", IEEE Signal Processing Letters, Vol. 23, No. 4, April 2016.
[32] J. Bazerque, G. Giannakis, "Distributed spectrum sensing for cognitive
radio networks by exploiting sparsity," IEEE Trans. Signal Process., vol.
58, no. 3, pp. 1847-1862, Mar. 2010.
[33] S. Stanković, and I. Orović, “An Approach to 2D Signals Recovering in
Compressive Sensing Context," Circuits, Systems and Signal Processing,
doi 10.1007/s00034-016-0366-8
[34] I. Stanković, I. Orović, M. Daković, and S. Stanković, "Denoising of
Sparse Images in Impulsive Disturbance Environment," Multimedia Tools
and Applications, 2017, doi:10.1007/s11042-017-4502-7
[35] M. Fornsaier, H. Rauhut, "Iterative thresholding algorithms", Applied and
Computational Harmonic Analysis, vol. 25, no. 2, Sept. 2008, pp. 187- 208.
[36] S. Mallat, Z. Zhang, "Matching pursuits with time-frequency dictionaries", IEEE Transactions on Signal Processing, vol. 41, no. 12. pp. 3397-3415,
1993.
[37] J. A. Tropp, A. C. Gilbert, "Signal Recovery From Random Measurements Via Orthogonal Matching Pursuit," IEEE Transactions on Information Theory, Vol. 53, np. 12, pp. 4655-4666, 2007.
[38] D. Needell, J. A. Tropp, "CoSaMP: Iterative signal recovery from
noisy samples", Applied and Computational Harmonic Analysis, 2008. DOI:
10.1016/j.acha.2008.07.002.
REFERENCES
31
[39] R. Tibshirani, "Regression Shrinkage and Selection via the LASSO". Journal of the Royal Statistical Society. Series B (methodological) 58 (1). Wiley:
267-88, 1996.
[40] S. Stanković, I. Orović, and L. Stanković, “An Automated Signal Reconstruction Method based on Analysis of Compressive Sensed Signals in
Noisy Environment," Signal Processing, vol. 104, Nov 2014, pp. 43 - 50,
2014.
[41] S. G. Mallat, Z. Zhang, “Matching pursuits with time-frequency dictionaries,” IEEE Transactions on Signal Processing, vol. 41, no. 12, 1993, pp.
3397–3415.
[42] I. Daubechies, M. Defrise, C. De Mol "An iterative thresholding algorithm for linear inverse problems with a sparsity constraint," Communications on Pure and Applied Mathematics, Volume 57, Issue 11, November
2004, Pages 1413-1457.
[43] T. Blumensath, M. E. Davies, "Gradient Pursuits," IEEE Transactions on
Signal Processing, vol.56, no.6, pp.2370-2382, June 2008,
[44] M. A. Figueiredo, R. D. Nowak, S. J. Wright, “Gradient projection for
sparse reconstruction: Application to compressed sensing and other inverse problems,” IEEE Journal of Selected Topics in Signal Processing, vol. 1,
no. 4, 2007, pp. 586–597.
[45] E. Candès, J. Romberg, “`1 -magic : Recovery of Sparse Signals via
Convex Programming, ” Caltech, http://users.ece.gatech.edu/justin/
l1magic/downloads/l1magic.pdf, Oct.2005.
[46] L. Stanković, M. Daković, S. Vujović, “Adaptive Variable Step Algorithm
for Missing Samples Recovery in Sparse Signals,”IET Signal Processing,
vol. 8, no. 3, 2014. pp. 246 -256.
[47] L. Stanković, “A measure of some time–frequency distributions concentration,” Signal Processing, vol. 81, 2001, pp. 621–631.
[48] S. Stanković, I. Orović, and L. Stanković, "Polynomial Fourier Domain
as a Domain of Signal Sparsity," Signal Processing, Vol. 130, Jan. 2017, pp.
243-253.
[49] S. Stankovic, L. Stankovic, and I. Orovic, "Compressive sensing
approach in the Hermite transform domain," Mathematical Problems
32
Conditions for Unique Reconstruction
in Engineering, Volume 2015 (2015), Article ID 286590, 9 pages
http://dx.doi.org/10.1155/2015/286590.
[50] A. Sandryhaila, J. Kovacevic, M. Puschel, "Compression of QRS complexes using Hermite expansion", IEEE int. Conference on Acoust., Speech
and Signal Process., ICASSP, Prague, 2011, pp. 581-584.
[51] M. Brajović, I. Orović, M. Daković, and S. Stanković, "On the Parameterization of Hermite Transform with Application to the Compression of
QRS Complexes," Signal Processing, vol. 131, pp. 113-119, February 2017.
[52] M. Brajović, I. Orović, M. Daković, and S. Stanković, "Gradient-based
signal reconstruction algorithm in the Hermite transform domain," Electronics Letters, Vol. 52, no. 1, pp.41-43, 2016.
[53] S. Stanković, I. Orović, and A. Krylov, "Video Frames Reconstruction
based on Time-Frequency Analysis and Hermite projection method,"
EURASIP Journal on Advances in Signal Processing, Special Issue on TimeFrequency Analysis and its Application to Multimedia signals, Vol. 2010, Article ID 970105, 11 pages, 2010.
[54] I. Orović, S. Stanković, T. Chau, C. M. Steele, and E. Sejdić, "Timefrequency analysis and Hermite projection method applied to swallowing accelerometry signals," EURASIP Journal on Advances in Signal Processing, Vol. 2010, Article ID 323125, 7 pages, 2010.
| 7cs.IT
|
Adapting Graph Application Performance Via
Alternate Data Structure Representations
Amlan Kusum
Iulian Neamtiu
Rajiv Gupta
University Of California,
Riverside
University Of California,
Riverside
University Of California,
Riverside
arXiv:1412.8120v1 [cs.PL] 28 Dec 2014
[email protected]
[email protected]
[email protected]
ABSTRACT
Graph processing is used extensively in areas from social
networking mining to web indexing. We demonstrate that
the performance and dependability of such applications critically hinges on the graph data structure used, because a
fixed, compile-time choice of data structure can lead to poor
performance or applications unable to complete. To address this problem, we introduce an approach that helps
programmers transform regular, off-the-shelf graph applications into adaptive, more dependable applications where
adaptations are performed via runtime selection from alternate data structure representations. Using our approach,
applications dynamically adapt to the input graph’s characteristics and changes in available memory so they continue to run when faced with adverse conditions such as low
memory. Experiments with graph algorithms on real-world
(e.g., Wikipedia metadata, Gnutella topology) and synthetic
graph datasets show that our adaptive applications run to
completion with lower execution time and/or memory utilization in comparison to their non-adaptive versions.
Keywords
runtime data structure selection, space-time trade-off
1.
INTRODUCTION
Graph processing continues to increase in popularity with
the emergence of applications such as social network mining, real-time network traffic monitoring, etc. Due to their
data-intensive nature, the performance and dependability of
such applications depends upon how well the choice of runtime data structure matches the input data characteristics
and availability of memory (low memory can prevent the
applications from completing).
Input Data Characteristics. Programmers often choose specific, fixed data structures when developing graph applications. The memory used by the data structure can be greatly
influenced by the input data characteristics. Thus, it is possible that the characteristics of data may not match the
choice of the data structure. This is particularly problematic
when the application is expected to encounter a wide range
of input data characteristics, and these characteristics may
change during the course of execution. For example, matrices can be represented in the Compressed Column Storage
(CCS) format, appropriate for sparse matrices, or the array
representation, appropriate for dense matrices. An application, e.g., matrix multiplication, programmed to use the
sparse CCS format, could take longer to complete when presented with a dense input. Similarly, evolving graphs [33],
where nodes or edges are added during execution, are another example of changes in input data characteristics. The
data structure selection based on input pre-analysis will fail
under such scenario. Therefore, in our approach, adaptive
applications tailor the choice of data structure to match input data characteristics at runtime.
Availability of Memory. Since real-world applications often do not run in isolation, they share the available memory resources with other applications. There could be times
where the application experiences a resource crunch, caused
by other running programs. In this scenario the performance of the application may be degraded, or the application may even be prematurely terminated. Therefore, in
our approach, adaptive applications tailor the choice of data
structure to match availability of memory at runtime.
It is well known that for data-intensive applications, the
choice of data structure is critical to memory usage and execution time. There has been previous work on data structure
identification [11], as well as data structure prediction and
selection [5, 6, 17, 22]. While these prior approaches help
in data structure selection, none of them support switching from one data structure to another as the application
executes. There has also been work on dynamically adapting the representation of individual data items for impacting
memory usage and performance—employing data compression [29] or replacing float data with int data [20]. These
techniques are orthogonal to our work that switches between
alternate high level data structures. Other approaches dynamically switch between implementations. Elastin [20] allows a program to switch between versions using dynamic
software update techniques [21, 9]; however, it does not consider switching between alternate high level data structures.
IBM’s K42 Operating System [1, 3] supports hot-swapping
classes as a mechanism for performing dynamic updates.
Scenario Based Optimization [19], a binary level online optimization technique dynamically changes the course of execution through a route meant for a particular runtime scenario as predefined by developer. Wang et al. [32] proposed
dynamic resource management techniques based on userspecific, application-specific and hardware-specific management policies. In contrast, our objective is to simultaneously
support alternate data structures and switch between them.
In this paper we consider several widely-used graph applications and study how data structure representations impact
execution time and memory consumption on a range of input
graphs (Section 2). The input graphs consist of both realworld graphs such as Wikipedia metadata, Gnutella network topology (from the SNAP library [31]), and synthetic
graphs. Based upon the observations from our study, we
design a concrete adaptation system that supports switch-
ing between alternate representations of the data in memory (Section 3). We demonstrate that the cost of performing the runtime adaptations is quite small in comparison
to the benefits of adaptation (Section 4). Moreover, the
lightweight monitoring we employ to detect adaptation opportunities imposes acceptable overhead even when no adaptations are triggered at runtime. Thus, our adaptive versions
have nearly the same performance as the most appropriate
non-adaptive versions for various input characteristics. We
compare our approach with related work in Section 5, and
in Section 6 we conclude.
2.
Table 1: Relative performance ranges.
Application
MSSP
BC
MST-K
BFS
MST-B
PP
Table 2: Density ranges where each data structure prevails.
Application
ADJLIST
best space &
time
MSSP
BC
MST-K
BFS
MST-B
PP
<
<
<
<
<
<
A STUDY OF GRAPH APPLICATIONS
In this section we study the execution time and memory
usage behavior of a set of graph applications. The goal of
this study is two fold. First, we want to quantify how input
data characteristics and the choice of data structures used
to represent the graphs impact memory usage and execution
time. Second, we would like to develop a simple characterization of program behavior that can be used to guide data
structure selection at runtime.
We considered six graph algorithms: Muliple Source Shortest Path (MSSP) finds the shortest path from all the nodes
to every other node; Betweenness Centrality (BC) computes
the importance of a node in a network; Breadth First Search
(BFS) traverses the graph with each node as root per iteration; Boruvka’s Algorithm (MST-B) and Kruskal’s Algorithm (MST-K), finds the minimum spanning tree; Preflow
Push (PP), finds out the maximum flow in a network starting with each individual node as source. The core data
structure used in these applications is a graph. We consider two different representations of graphs: Adjacency List
(ADJLIST); and Adjacency Matrix (ADJMAT). When the graph
is sparse, it is expected that ADJLIST will use less memory
than ADJMAT. On the other hand, for highly dense graphs
ADJMAT may use less memory than ADJLIST. Determining
whether a pair of nodes is connected by an edge can be
done in constant time using ADJMAT while it may require
searching through a list with ADJLIST. Thus, the runtime
memory usage and execution time depend upon the sparsity, or conversely the density, of the input graph. The input
graphs with relevant properties and densities were generated
to study program behavior.
To observe the trade-offs of using the alternative representations of graphs, we executed each of the programs using
the two representations. The programs were run on inputs
consisting of randomly-generated graphs with varying den|E|
sity which is computed as |V ||(V
, where |V | and |E| are
−1)|
number of nodes and edges in the graph. The inputs were selected such that the trade-offs could be exposed easily. The
results of these executions are summarized as follows:
Impact of data structure selection on memory usage and execution time. We present the relative memory usage and execution time of program versions in Table 1. In particular, we
computed the ratios of memory usages and execution times
for ADJLIST and ADJMAT versions across all graph densities
considered. The minimum and maximum values of observed
ratios is given in Table 1. As we can see, in terms of both
memory usage and execution time, the relative performances
vary a great deal. Moreover, neither representation gives the
best memory usage or execution time performance across all
graph densities. Hence, it is crucial to select the data structure at runtime, based upon the input data characteristics.
ADJLIST / ADJMAT
Memory Usage Execution Time
0.68 - 8.02
0.40 - 4.00
0.40 - 4.00
0.59 - 2.88
0.72 - 2.44
0.35 - 3.83
0.54 - 5.47
0.72 - 3.14
0.71 - 1.67
0.16 - 7.21
0.60 - 5.40
0.50 - 3.53
9%
10%
25%
8%
10%
2%
ADJLIST
best
space,
ADJMAT
best time
9% - 25%
10% - 25%
25% - 37%
8% - 25%
10% - 40%
2% - 34%
ADJMAT
best space &
time
>
>
>
>
>
>
25%
25%
37%
25%
40%
34%
Characterization of application behavior. For the purpose of
runtime data structure selection, we characterize the behavior of each application as shown in Table 2. Note that graph
densities are divided into three subranges. In the first range
(e.g., < 9% for MSSP) the ADJLIST is both more memoryand time-efficient than ADJMAT. In the second range (e.g.,
9% − 25%) ADJLIST is more memory-efficient while ADJMAT
is more time-efficient. Thus, the selection can be made at
runtime based upon memory availability. Finally, in the
third range (e.g., > 25% for MSSP) ADJMAT is both more
memory and time efficient than ADJLIST.
3. ADAPTIVE APPLICATIONS
&''()*)+,$"(-./+$0(,+$
!"#$
!"%$
&,*123+$&1145/*2('$
"(-./+$/(,+$:5);$
).*'852('$4(<5/$
!"#$
!"%$
&,*1)*2('$6(,-4+$
7.*'852('$
1(45/5+8$
9-'26+$
6('5)(.8$
Figure 1: High level overview of our approach.
We now present our approach for building adaptive applications; an overview is shown in Figure 1. The starting
point is the annotated source code: in the source code, programmers add annotations to identify the alternative data
structures, e.g., DS1 and DS2 , and functions operating on
them. The compiler takes heed of these annotations and
generates the source code with transition logic, that is capable of dynamically switching among alternative data structure representations. The transitions are allowed at selected
program points where the processing of an input item has
just completed and that of another item is about to begin. Lastly, the adaptation module consists of the runtime
monitors for tracking input data characteristics and memory usage as well as the code that implements the transition
policy that triggers the switch from one data structure representation to another. The adaptation can be triggered by
a mismatch between the input data characteristics and the
data structure currently in use. To discover this mismatch
the characterization of application behavior as performed in
the previous section is used. The adaptation can also be
triggered by the system during high memory usage.
Programming for adaptation. To enable adaptation, the
programmer implements the alternate data structures. In
addition, a compute-intensive function during whose execution adaptation may be performed, must be coded as follows.
First, it should contain a variable that tracks the progress
in terms of processing steps defined as either the amount of
input processed or results produced. Second, it should be
written so that it can commence execution from any point
between two processing steps. The latter is needed because
we allow execution to switch from one data representation to
another at these points. We used a set of pragmas in our approach to identify alternate data structure representations,
enable generation of code that transfers code from one representation to another, and identify program points where
transitions may be performed. First, the programmer identifies the data structure to the compiler. The programmer
annotates the alternate representation of data structures
in multiple files with #pragma ADP(<SRC_FILENAME>, "data1_def").
<SRC_FILENAME>’s presence clearly differentiates the alternate
representation of the data structure in multiple files. If there
are multiple data structures with alternate representations
in different files, then they could be annotated with a different index, e.g., #pragma ADP(<SRC_FILENAME>, "data2_def"). Second, the programmer uses several pragmas to identify the
key methods (insert, delete, traverse, and fetch) that manage data stored in the data structure. Another pragma allows access to the initialization parameters which must be
migrated from one data structure to another. All of this information is used to generate the code for data and function
migration when we switch between data structures.
Triggering adaptations. The adaptation module decides
whether or not to switch between data structures based
upon the input from runtime monitors and the transition
policy. Since the adaptation could be program-triggered or
system-triggered, there are two kinds of monitors which are
required by the adaptation module. The input data monitor
captures input data characteristics and the memory monitor reports the available system memory. The transition
policy defines which data structure representation is better for what range of input data characteristics in terms of
execution time and memory consumption. Its specification
consist of three parts, as illustrated below:
/* EXECUTION TIME */
DS1 [0,9)
DS2 [9,100]
/*MEMORY*/
DS1 [0,25)
DS2 [25,100]
/*THRESHOLD*/
MEMORY 100
The first part indicates the ranges for which a particular
data structure representation is best in terms of execution
time: under EXECUTION TIME in the figure, the input data property for which ADJLIST (DS1) is better is denoted by directives DS1, which means that ADJLIST is favorable in terms
of execution time if the input data property or density of
the graph (in case of MSSP) is in between 0% and 9%. The
second part consists of the ranges of the input data property for which a particular data structure representation is
better in terms of memory. According to the figure, under
MEMORY, we see that ADJLIST (DS1) is better when the density
of the input graph is between 0% and 25% while ADJMATRIX
(DS2) is better when the density of the graph is between
26% and 100%. The third part is the threshold for memory,
defined by the programmer to notify the system that if the
available memory is below this threshold then, regardless
of input data characteristics always use the representation
requiring least memory; in the figure (under THRESHOLD) the
threshold is set to 100MB.
dataMigrationDS1DS2(void* DS1, void* DS2)
{
initializationParameters* ip;
ip = getInitializationParameter(DS1);
initializeDS2(&DS2,ip);
transferDataDS1DS2(&DS1,&DS2)
deleteDS1(&DS1);
}
transferDataDS1DS2(void** DS1, void** DS2)
{
i = 0; void* dataValue;
for(i = 0;i< **DS1->maxData;i++) {
dataValue = fetchDataDS1(i,*DS1);
if(dataValue != NULL) {
insertDataDS2(*DS2, dataValue, i);deleteDataDS1(i,*DS1);
}}}
Figure 2: Data migration.
Switching between data structure representations. The data
structure transition logic is inserted into the source files by
the compiler, guided by the pragmas. This transition logic
carries out on-the-fly transitions from one data structure
representation to another whenever required. To accomplish
the transition, the in-memory data must be transformed
from one representation to another, along with the functions operating on them. The transition logic handles this
by function migration and in-memory data methods contained in the logic. When the code for transition logic is
inserted, appropriate header files are also inserted such that
source code after modification compiles and links properly.
To avoid recomputation of already-computed results, the result transfer logic (injected into code along with the transition logic) will transfer the already-computed results from
one representation to the other representation.
An example data migration function is shown in Figure 2.
The code in the figure transfers the data from the data structure representation DS1 to another representation DS2. It
begins with initialization of the DS2 data structure representation. The initialization parameters are fetched from
DS1 and they consist of standard parameters that are invariant in both DS1 and DS2. For example, in the MSSP
benchmark the invariant data is the number of nodes. In
the PP benchmark the invariant data consists of number
of nodes, the height, capacity and flow of each node. The
transferData function is generated from traverseData function
of DS1 as provided by the developer. This function traverses
through the data by reading each data value, migrating it
to DS2 representation using insertDataDS2 and also deleting
that data from DS1 using deleteDataDS1 thus releasing memory. The deleteDS1 clears memory which contains the data
regarding the initialization parameters.
The transition between implementations, i.e., switching
from one set of functions operating on representation DS1
to functions operating on representation DS2 must be carefully orchestrated. The developer denotes an operation with
a directive such as #pragma ADP("DS1","data1_op1"), which informs the compiler that the function is compute-intensive,
as shown in Figure 3. Any call to that function is replaced by our customized method, which checks and executes operations with the suitable data structure. In this
example computeMSSP_DS1 is replaced by callOP1. The additional parameter, startDS, denotes the type of the current
data structure representation in memory. The other three
parameters are the data structure, a progress gauge, and
the result set for storing the result. For example in the
#pragma ADP("DS1",
"ds1_op1")
void computeMSSP_DS1(
void* graph, void*
rs,
int* progress);
...
computeMSSP_DS1(graph,
rs, progress);
...
//#pragma ADP("DS1",
"ds1_op1")
void computeMSSP_DS1(
void* graph,void*
rs,
int* progress);
...
callOP1(graph,
rs,progress,
startDS);
...
void computeMSSP_DS1(
void* graph,
void* rs,
int* progress){
...
#pragma ADP("DS1",
"ds1_op1_safe")
...
void computeMSSP_DS1(
void* graph,
void* rs,
int* progress){
...
//#pragma ADP("DS1"
,"ds1_op1_safe")
if(checkChangeStatus()==1)
{
*progress = curProgress;
return;
}
}
}
void callOP1(void* ds, void* rs, int progress, currentDS){
extern int changeReq; void* newDS; void* newRS;
while(progress < 100){
if(changeReq == 1){ switch(currentDS) {
case 1:
currentDS = 2; dataMigrationDS1DS2(ds, newDS);
resultMigrationRS1RS2(rs, newRS);
ds = newDS; newDS = NULL; rs = newRS; newRS = NULL;
computeMSSPDS2(ds, rs, progress);
break;
case 2:
currentDS = 1; dataMigrationDS2DS1(ds, newDS);
resultMigrationRS2RS1(rs, newRS);
ds = newDS; newDS = NULL; rs = newRS; newRS = NULL;
computeMSSPDS1(ds, rs, progress);
break;
}}
else { switch(currentDS) {
case 1: computeMSSPDS1(ds, rs, progress); break;
case 2: computeMSSPDS2(ds, rs, progress); break;
}}}}
Figure 4: Switching between implementations.
case of MSSP, a method that finds MSSP has the signature
void computeMSSP_DS1(void* graph, void* rs ,int* progress). The
first parameter is the input graph and the second parameter rs stands for the result set and its declaration must
be annotated by the programmer with #pragma ADP("DS1",
"data1_res1"). The last parameter identifies the progress,
which is the iteration number of the outer most long running
loop. For example, if the method is called with a progress
value 10, then the execution is started from progress value 10
and continuously updated with the loop iteration number.
The detailed function selection and migration activity is
shown in Figure 4—for MSSP benchmark. An external variable changeReq, set by the adaptation module, is checked
(line 4). If a transition has been requested, then first the
data is migrated from one data structure representation to
another (lines 6 and 12). Next, if needed, the result is migrated from one representation to another (lines 7 and 13).
Finally, the corresponding MSSP function for that data structure is called (lines 9 and 15) and the operation is resumed
from the progress point. If there is a change request from
the adaptation module, then operation is paused and it returns back to callOP1. This process continues until the MSSP
computation completes.
The question arises where ongoing MSSP computations
should be interrupted to check if the adaptation module has
requested a change or not. To solve this problem, we rely
on the programmers to use the directive #pragma ADP("DS1",
"ds1_op1_safe") to indicate the safe transition points in operation1
as shown in Figure 5. This directive notifies our framework
that, if the operation is paused and the transformation is
performed at that point, then there is minimal recomputation of result. This is typically the end of an iteration in
long-running loops. Since the programmer is well aware of
the long running loops in the compute-intensive function, it
is best to have the programmer mark the points appropri-
Figure 5: Adaptation module interrupt before compilation
(left) and after compilation (right).
Table 3: Comparison of execution times of non-adaptive versions with adaptive version under program triggered adaptations, on the original p2p-Gnutella graph. Note that ADJLIST is the better representation.
App.
Non-Adaptive Ex.
Time (sec)
Adaptive:
ADJMAT→
ADJLIST (sec)
Ex.
Transition
Time
Latency
1,408
3.00
1,383
2.73
397
3.17
1,457
2.56
1,465
3.15
87
3.2
ADJLIST ADJMAT
MSSP
BC
MST-K
BFS
MST-B
PP
1,386
1,362
389
1,434
1,454
81
2,489
2,565
956
2,594
2,114
256
Benefit
Realized
(%)
98.05
98.32
98.55
98.05
98.35
96.35
ate for the insertion of adaptation module interrupts. The
directive is replaced by an interrupt which checks if there is
a change required and thus returns back to callOP1.
ADAPTIVE VERSION
ADJMAT
BC
EXECUTION
TIME (SEC)
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
ADJLIST
BFS
3000
2000
1000
0
MSSP
3000
2000
1000
0
8
9
10
20
3000
2000
1000
0
8
% DENSITY
9
10
20
2000
1000
0
30
40
% DENSITY
10
20
PP
1000
80
500
40
0
20
9
% DENSITY
MST-K
3000
9
8
% DENSITY
MST-B
EXECUTION
TIME (SEC)
Figure 3: Function migration method before compilation
(left) and after compilation (right).
0
10
20
30
% DENSITY
40
1
2
3
4
% DENSITY
Figure 6: Adaptive vs. non-adaptive performance.
4. EVALUATION
In this section we evaluate the performance of adaptive
versions of graph algorithms and compare them with corresponding non-adaptive versions of the applications. The
goals of these experiments are as follows. First, we evaluate the efficiency of our approach by measuring its benefits
and overhead. Second, we consider the benefits of adaptation under two scenarios: adaptation triggered by the input characteristic, i.e., graph density; and system triggered
adaptation. All experiments were run on a 24-core machine
(4 six-core AMD OpteronT M 8431 processors) with 32GB
RAM. The system ran Ubuntu 10.04, Linux kernel version
2.6.32-21-server. The sources were compiled with Gcc 4.4.3.
Real World Data-sets: We evaluate our system on some of
the real-world graphs from the SNAP graph library [31]. The
first graph, wiki-Vote, contains the who-votes-for-whom
graph in Wikipedia administrator elections. This graph has
7,115 nodes and 103,689 edges. The second graph, p2pGnutella, is a snapshot of Gnutella, a decentralized peer to
peer file sharing network from August 9, 2002. This graph
Table 5: Breakdown of adaptation overhead.
Table 4: Programming effort.
BC MST-K BFS
9
9
8
12
10
8
MST-B PP
9
8
14
6
has 8,114 nodes representing hosts and 26,013 edges representing the connections between these hosts. For experiments, in cases where a more dense graph was needed, we
added edges in both the graphs to raise the required density.
4.1 Programming Effort
The programmers need to add annotations to transform
off-the-shelf applications to adaptive ones. In addition to
this, programmers also need to modify the compute-intensive
methods so they can be executed in incrementalized fashion.
The number of pragmas added and the number of additional
lines of code added to modify the methods are shown in Table 4. As we can see, these numbers are fairly modest.
4.2 Input Triggered Adaptation
In this scenario we study how adaptive applications respond to the mismatch between the data structure representation fixed a priori at compile time and the density of
the input graph. We compute the benefit realized by our approach for various applications. In particular, we start the
program by using the ADJMAT representation and select a real
world graph (p2p-Gnutella) which is 0.004% dense, which
makes ADJLIST the ideal representation. Therefore, when
the adaptive application is run, it dynamically switches from
the ADJMAT to the ADJLIST representation.
In Table 3 we present the execution times of the nonadaptive (ADJLIST and ADJMAT representations) and adaptive (ADJMAT→ADJLIST) versions of the applications. For the
latter version, we also present the transition latency which
is the execution time after which the program has completed
the transition to the ADJLIST representation. From the results in Table 3, we observe the following. The execution
time of the adaptive version, on average, is 2.49% higher
than the non-adaptive ADJLIST version; but 48.09% lower
than the non-adaptive ADJMAT version. For example, for
MSSP, the execution of the adaptive version is 1,408 seconds
which is 1.54% higher than the execution time of the nonadaptive ADJLIST version (1386 seconds) and 56.55% lower
than the execution time of the non-adaptive ADJMAT version
(2,489 seconds). In addition, we observe that the transition
latency of the adaptive version is small in comparison to the
total execution time. For example, for MSSP, the transition
latency of 3 seconds is approximately 0.21% of the total execution time of 1,408 seconds. That is, the adaptation is performed quickly (low transition latency) and efficiently (low
transition overhead). Thus, nearly all the benefits of using
ADJLIST over ADJMAT are realized by the adaptive version.
We quantify the benefit realized by our approach as follows. The maximum possible benefit is given by the difference in the execution times of the non-adaptive ADJMAT and
non-adaptive ADJLIST versions. The benefit our approach
realizes is the difference between the execution times of the
non-adaptive ADJLIST version and the adaptive version. The
realized benefit, as a percentage of maximum possible benefit, is given in the last column of Table 3. As we can see,
the realized benefit is over 96% for these applications.
The additional execution time taken by the adaptive version over the non-adaptive ADJLIST version can be divided
into three categories: time spent on converting from one
Application
DS Conversion
Monitoring
&
Transition Logic
Suboptimal Mode
MSSP
1.56
10.78
9.07
BC MST-K BFS
1.44
1.98
1.4
10.11
5.89
12.74
8.62
300
250
200
150
100
50
0
300
250
200
150
100
50
0
0.3
BC
8.41
2 4 6 8 10
TIME (SEC)
0
MSSP
2 4 6 8 10
TIME (SEC)
0
2 4 6 8 10
TIME (SEC)
MST-K
300
250
200
150
100
50
0
2 4 6 8 10
TIME (SEC)
1.48
300
250
200
150
100
50
0
MST-B
0
1.71
ADJLIST
BFS
300
250
200
150
100
50
0
0
MST-B PP
1.98
1.98
7.19
2.92
TRANSITION
ADJMAT
MEMORY
CONSUMPTION (MB)
MSSP
8
9
MEMORY
CONSUMPTION (MB)
Application
# pragmas
Additional
LOC
PP
300
250
200
150
100
50
0
0
2 4 6 8 10
TIME (SEC)
0
2 4 6 8 10
TIME (SEC)
Figure 7: Program-triggered adaptation.
data structure representation to another; time spent on runtime monitoring and transition logic to trigger adaptation;
and the time lost due to running the application in suboptimal mode, i.e., with the ADMAT data structure. The breakdown of the extra execution time into the three categories is
shown in Table 5. As we can see, the majority of the time is
spent on runtime monitoring and transition logic. The next
significant component is the time spent due to running the
program in the suboptimal configuration before the transition occurs. Note that the time spent on converting one
data structure into another (column 2) is the least.
An intuitive way to visualize adaptation is to plot how
the memory used by applications varies before, during, and
after adaptation. In Figure 7 we show how memory (yaxis) varies over time (x-axis) when starting the application
in the ADJMAT representation and then through adaptation,
the application transitions to ADJLIST. The charts point out
several aspects. First, since we are using sparse graphs, as
expected, the memory used is reduced significantly (tens of
megabytes) when we switch from the ADJMAT to ADJLIST
representation. Second, the switch from one data structure
to the other takes place fairly early in the execution of the
program. Third, the time to perform adaptation and the
extra memory used during adaptation are very low.
In Figure 6 we show the execution time of the adaptive
version for varying input densities over the range where we
expect the adaptive application to switch from the ADJLIST
to the ADJMAT representation. For these experiments, we
have used graph size of 4000 nodes and varied densities.
The execution times of the non-adaptive versions that use
fixed representations (ADJLIST and ADJMAT) are also shown.
As we can see, the performance of the adaptive application
is very close to the best of the two non-adaptive versions.
4.3 System Triggered Adaptation
In this section we study the second scenario, i.e., when the
adaptation is triggered by the system. The graph used for
these experiments was p2p-Gnutella at 20% density. However, we select ADJMAT as the initial data structure representation so that no adaptation was triggered due to the
mismatch between the data structure and graph density. Instead we provided the program with a system trigger that
forces the program to reduce its memory consumption. This
NORMALIZED EXECUTION TIME
1ADMAT ONLY
0
075% ADJMAT − 25% ADJLIST
1
0
1
150% ADJMAT − 50% ADJLIST
0
1
0.8
0.6
0.4
0.2
0
4.4 Limitations of Our Approach
ADJMAT − 75% ADJLIST
011025%
ADJLIST ONLY
01
1
001
11
0
1
0
00
11
00
11
0
0
1
0
1
00
11
0
1
0
1
00
11
00
11
00
11
0
1
0
1
00
11
00
11
0
1
0
1
0
1
00
11
0
1
0
1
0
1
0
1
0
1
00
11
00
11
00
11
0
1
0 1
1
00
11
0
1
0
1
0
1
00
11
00
11
0
0
1
0
1
00
11
0
1
00
11
00
11
0
1
0
1
0
1
0
1
0
1
00
11
00
11
01
1
01
1
00
11
0
1
0 1
1
00
11
00
11
00
11
0
1
01
1
0
1
0
1
00
11
0
00
11
00
11
0
0
00 11
1
0
1
0
1
00
11
00
11
0
1
0
1
0
1
0
1
0
1
0
1
00
11
00
01
1
01
1
00
11
0
1
0 1
1
00
11
00
11
0
1
00
11
0
1
00
01
1
0
0
1
00
11
0
1
00
11
00
11
0
0
0011
1
00 1
1
00
11
0
1
00
11
00
11
0
1
0
1
0
1
00
11
0
1
0
1
0
1
0
1
00
11
0
1
0
1
0
1
00
11
00
11
00
011
1
0
1
00
00
11
00
0
1
00
0
1
0
1
0
1
00
11
011
1
011
0 10101010101011
00
11
00
0
1
0
1
01
1
01
1
00
11
011
1
01
1
00
11
00
11
0
1
00
11
0
0
1
011
1
0
1
0
1
00
11
0
1
00
11
01
1
0
1
00
11
00
0
1
00
11
0
1
0
1
00
11
00
11
0
1
00
11
00
11
0
1
0
1
0
1
00
11
011
1
00
11
0
00
11
00
0
1
0
1
0011
01
1
00
11
00
11
01
1
01
1
00
11
00
11
0
1
0
1
00
0
0
1
011
1
0
1
0
1
0
1
00
11
0
1
00
11
01
1
0
1
00
11
00
11
0
1
00
00
11
0
1
0
1
00
11
00
11
0
1
0
1
00
11
00
11
0
1
0
1
0
1
0
1
00
11
0
1
00
11
011
0
1
00
11
00
11
0
1
0
1
0
1
0
1
00
11
00
11
0
1
0
1
00
11
00
11
0
1
0
1
00
0
1
0
1
0
1
0
1
0
1
0
1
00
11
0
1
00
11
01
1
0
1
00
11
00
11
0
1
00
11
00
11
0
1
0
1
00
11
00
11
0
1
0
1
00
11
00
11
0
1
01
1
0
1
0
1
00
11
011
1
00
11
0
00
11
00
0
1
0
0011
01
00
11
00
11
01
1
01
1
00
11
00
11
0
1
0
1
0
1
0
011
1
0
1
0
1
00
11
0
1
0
1
00
11
0 00
1
0
1
00
11
00
11
011
1
00
0
1
00
0
1
00
11
00
11
0
1
0
1
00
11
00
11
0
1
0 10BC 1
0MST−K
00 BFS
11
00MST−B
11
0 PP 1
0
1
MSSP1
01
1
Figure 8: Comparison of adaptive vs. non adaptive normalized execution time in system-triggered adaptation.
TRANSITION
MEMORY
CONSUMPTION (MB)
MEMORY
CONSUMPTION (MB)
ADJMAT
BC
MSSP
300
300
300
250
250
250
200
200
200
150
150
0
2500
5000
TIME (SEC)
150
0
MST-B
300
2500
5000
TIME (SEC)
200
150
0
2500
5000
TIME (SEC)
0
MST-K
300
250
200
150
100
250
5. RELATED WORK
ADJLIST
BFS
2500
5000
TIME (SEC)
PP
300
250
200
150
100
0
500 1000 1500
TIME (SEC)
0
200 400
TIME (SEC)
First, our approach is only useful when the alternative
data structures offer a significant trade-off between memory
usage and execution time. For example, for the agglometric clustering benchmark, when we tried using two alternate
data structures of kd-tree and r-tree, we observed no significant trade-off between memory usage and execution time.
Since there is a need to bulk load the data, the kd-tree always outperforms the r-tree. Second, our approach is only
useful when the application is sufficiently compute and data
intensive to justify the cost of runtime monitoring and transition logic. For example, in the case of the Max Cardinality
Bipartite Matching benchmark, although the trade-off exists, the benchmark is not sufficiently compute-intensive to
justify the adaptation cost.
600
Figure 9: System-triggered adaptation for wiki-Vote.
causes adaptation to be triggered, and the program to switch
from ADJMAT to ADJLIST representation to save memory. As
expected, the execution takes longer. Since the conversion
from one representation to another can be triggered at any
time during a program’s execution, in this study we present
data for different trigger points – after 25%, 50%, and 75% of
total processing. We controlled the trigger point by tracking
the amount of processing that has been completed.
The results are presented in Figure 8. The execution times
of the following versions are presented: non-adaptive version
in the ADJMAT representation (leftmost bar); three adaptive
versions with different trigger points (middle three bars);
and non-adaptive ADJLIST (rightmost bar). All times are
normalized with respect to the time for non-adaptive ADJLIST. As we can see, the execution time of the adaptive version is always greater than the non-adaptive ADJMAT version
and less than the non-adaptive ADJLIST version. In other
words, if large amounts of memory are available for longer
duration, the adaptive version yields greater reduction in
execution time over the non-adaptive ADJLIST version.
To study the behavior of our approach when there are
multiple transitions, we ran experiments on wiki-Vote at
10% density in the following scenario. For each benchmark,
the execution was started with ADJMAT and then switched to
ADJLIST and vice versa after 20 %, 40%, 60% and 80%. We
controlled the triggers for memory changes from the system
by tracking the amount of processing that has been completed. We present the results in Figure 9. We can clearly
see that, during a resource crunch when available memory
decreases, our applications adapt to decrease their memory
requirements accordingly, hence running slower; after the
resource crunch is over, our applications re-assume the uncompressed representation and their performance increases.
There is a large body of work on program transformations applied at compile-time or runtime to enhance program
performance, which also influences resource usage. Some of
these techniques can be used to support adaptation. ContextErlang [4] supports the construction of self-adaptive software using different call back modules. Compiler-enabled
adaptation techniques include altering of the contentiousness of an application [14, 16], which enables co-location
of applications without interfering with their performance;
data spreading [18] migrates the application across multiple cores; adaptive loop transformation [7] allows a program
to execute in more than one way during execution based on
runtime information. Multiple applications that are running
on multicore systems can significantly impact each other’s
performance as they must share hardware resources (e.g.,
last level cache, access paths to memory) [27]. The impact
of interference on program performance can be predicted and
estimated [10, 15], and contention management techniques
guided by last level shared cache usage and lock contention
have been developed [2, 25, 24, 13, 26, 12].
Huang et al. proposed Self Adaptive Containers [8] where
they provide the developer with a container library which
adjusts the underlying data structure associated with the
container to meet Service Level Objectives (SLO); adaptation occurs during SLO violations. Similarly, CoCo [28] allows adaptation by switching between Java collections during execution depending on the size of collection. These
methods are orthogonal to our approach as they do not have
scope for user-defined data structures, and the space-time
tradeoff is not taken into consideration.
6. CONCLUSION
Graph applications have resource requirements that vary
greatly across runs due to differences in graph characteristics; moreover, the required memory might not be available due to pressure from co-located applications. We have
observed that data structure choice is crucial for allowing
the application to get the best out of available resources.
We propose an approach that uses programming and runtime support to allow graph applications to be transformed
into adaptive applications by choosing the most appropriate
data structure. Experiments with graph-manipulating applications which adapt by switching between data structure
representations show that our approach is easy to use on offthe-shelf applications, is effective at performing adaptations,
and imposes very little overhead.
Acknowledgments
This work was supported in part by NSF grants CCF-0963996
and CCF-1149632. This research was sponsored by the
Army Research Laboratory and was accomplished under Cooperative Agreement Number W911NF-13-2-0045 (ARL Cyber Security CRA). The views and conclusions contained in
this document are those of the authors and should not be
interpreted as representing the official policies, either expressed or implied, of the Army Research Laboratory or the
U.S. Government. The U.S. Government is authorized to
reproduce and distribute reprints for Government purposes
notwithstanding any copyright notation here on.
7.
REFERENCES
[1] A. Baumann, J. Appavoo, D. D. Silva, J. Kerr,
O. Krieger, and R. W. Wisniewski. Providing dynamic
update in an operating system. USENIX ATC’05.
[2] M. Bhadauria and S. A. McKee. An approach to
resource-aware co-scheduling for cmps. ICS’10.
[3] C. A. N. Soules, J. Appavoo, D. D. Silva,
M. Auslander, G. R. Ganger, M. Ostrowski, and et al..
System support for online reconfiguration. USENIX
ATC’03.
[4] C. Ghezzi, M. Pradella, and G. Salvaneschi.
Programming language support to context-aware
adaptation: a case-study with erlang. SEAMS’10.
[5] C. Jung, S. Rus, B. P. Railing, N. Clark, and S. Pande.
Brainy: effective selection of data structures. PLDI’11.
[6] E. Schonberg, J. T. Schwartz, and M. Sharir. An
automatic technique for selection of data
representations in setl programs. ACM TOPLAS’81.
[7] R. Gupta and R. Bodik. Adaptive loop
transformations for scientific programs. IPDPS’95.
[8] W.-C. Huang and W. J. Knottenbelt. Self-adaptive
containers: Building resource-efficient applications
with low programmer overhead. SEAMS 2013.
[9] I. Neamtiu, M. Hicks, G. Stoyle, and M. Oriol.
Practical dynamic software updating for C. PLDI’06.
[10] J. Mars, L. Tang, R. Hundt, K. Skadron, and M. L.
Soffa. Bubble-up: increasing utilization in modern
warehouse scale computers. In MICRO-44 ’11.
[11] C. Jung and N. Clark. Ddt: design and evaluation of a
dynamic program analysis for optimizing data
structure usage. MICRO-42 ’09.
[12] K. Kumar Pusukuri, R. Gupta, and L. N. Bhuyan.
Adapt: A framework for coscheduling multithreaded
programs. In ACM TACO’13.
[13] K. Pusukuri, R. Gupta, and L. Bhuyan. No more
backstabbing... a faithful scheduling policy for
multithreaded programs. PACT’11.
[14] L. Tang, J. Mars, and M. L. Soffa. Compiling for
niceness: mitigating contention for qos in warehouse
scale computers. CGO’12.
[15] L. Tang, J. Mars, N. Vachharajani, R. Hundt, and
M. Soffa. The impact of memory subsystem resource
sharing on datacenter applications. ISCA’11.
[16] L. Tang, J. Mars, W. Wang, T. Dey, M. Soffa. Reqos:
Reactive static/dynamic compilation for qos in
warehouse scale computers. ASPLOS’13.
[17] L. Liu and S. Rus. Perflint: A context sensitive
performance advisor for C++ programs. CGO’09.
[18] M. Kamruzzaman, S. Swanson, and D. M. Tullsen.
Software data spreading: leveraging distributed caches
to improve single thread performance. PLDI’10.
[19] J. Mars and R. Hundt. Scenario based optimization:
A framework for statically enabling online
optimizations. In CGO, pages 169–179, 2009.
[20] I. Neamtiu. Elastic executions from inelastic
programs. SEAMS’11
[21] I. Neamtiu and M. W. Hicks. Safe and timely updates
to multi-threaded programs. PLDI’09.
[22] O. Shacham, M. Vechev, and E. Yahav. Chameleon:
adaptive selection of collections. PLDI’09.
[23] K. Pingali, D. Nguyen, M. Kulkarni, M. Burtscher,
M. A. Hassaan, R. Kaleem, T.-H. Lee, A. Lenharth,
R. Manevich, M. Méndez-Lojo, D. Prountzos, and
X. Sui. The tao of parallelism in algorithms. PLDI’11.
[24] R. Knauerhase, P. Brett, B. Hohlt, T. Li, and
S. Hahn. Using os observations to improve
performance in multicore systems. Micro’93.
[25] S. Blagodurov, S. Zhuravlev, A. Fedorova, and
A. Kamali. A case for numa-aware contention
management on multicore systems. PACT’10.
[26] S. Zhuravlev, S. Blagodurov, and A. Fedorova.
Addressing shared resource contention in multicore
processors via scheduling. ASPLOS’10.
[27] W. Wang, T. Dey, J. Mars, L. Tang, J. Davidson, and
M. Soffa. Performance analysis of thread mappings
with a holistic view of the hardware resources.
ISPASS’12.
[28] Guoqing Xu. Coco: Sound and Adaptive replacement
of java collections. In ECOOP, pages 1–26, June 2013.
[29] Y. Zhang and R. Gupta. Data compression
transformations for dynamically allocated data
structures. LNCS 2304 ’02.
[30] J. Eastep, D. Wingate and A. Agarwal. Smart Data
Structures: An Online Machine Learning Approach to
Multicore Data Structures. In ICAC, pages 11–20,
2011.
[31] J. Leskovec et al. Stanford Network Analysis Platform.
http://snap.stanford.edu/snap/index.html
[32] Wei Wang et al. REEact: A Customizable Virtual
Execution Manager for Multicore Platforms. VEE’12.
[33] Kunegis, Jérôme. KONECT: The Koblenz Network
Collection. WWW ’13.
| 6cs.PL
|
arXiv:1712.07426v1 [math.GR] 20 Dec 2017
Actions of E−dense semigroups and an application to
the discrete log problem
James Renshaw
Mathematical Sciences
University of Southampton
Southampton, SO17 1BJ
England
[email protected]
June 2017
Abstract
We describe the structure of E−dense acts over E−dense semigroups in an analogous
way to that for inverse semigroup acts over inverse semigroups. This is based, to a large
extent, on the work of Schein on representations of inverse semigroups by partial oneto-one maps. We consider an application to the discrete log problem in cryptography
as well as an application to the same problem using completely regular semigroups.
Key Words Semigroup, monoid, E−dense, E−inversive, completely regular, semigroup
acts, E−dense acts, discrete logarithm, cryptography
2010 AMS Mathematics Subject Classification 20M30, 20M50, 20M99.
1
Introduction and Preliminaries
Let S be a semigroup. By a left S−act we mean a non-empty set X together with an action
S × X → X given by (s, x) 7→ sx such that for all x ∈ X, s, t ∈ S, (st)x = s(tx). If S is a
monoid with identity 1, then we normally require that 1x = x for all x ∈ X. A right S−act
is defined dually. If X is a left S−act then the semigroup morphism ρ : S → T (x) given
by ρ(s)(x) = sx is a representation of S. Here T (X) is the full transformation semigroup
on X consisting of all maps X → X. Conversely, any such representation gives rise to an
action of S on X. If X is both a left S−act and a right T −act for semigroups/monoids S
and T and if in addition (sx)t = s(xt) then X is said to be an (S, T )−biact. Throughout
this paper, unless otherwise stated, all acts will be left S−acts. We refer the reader to [7]
for basic results and terminology in semigroups and monoids and to [2] and [8] for those
concerning acts over monoids. If S is an inverse semigroup then we can replace T (X) by
I(X), the inverse semigroup of partial one-to-one maps. A comprehensive theory of these
types of representations was given by Boris Schein in the early 1960’s and an account of
that work can be found in [1] and [7]. Here we wish to emulate that approach for E−dense
semigroups and do so in section 2. In section 3, we apply some of these results to the discrete
log problem found in cryptography (see for example [9]).
Recall that an idempotent in a semigroup S is an element s ∈ S such that s2 = s. A band is
a semigroup consisting entirely of idempotents whilst a semilattice is a commutative band.
We shall denote the idempotents of a semigroup S as E(S) or more generally E. Let S
1
be a semigroup and let W (s) = {s′ ∈ S|s′ ss′ = s′ } be the set of weak inverses of s and
V (s) = {s′ ∈ S|s′ ∈ W (s), s ∈ W (s′ )} be the set of inverses of s. If S is a group then clearly
W (s) = V (s) = {s−1 } for all s ∈ S, whilst if S is a rectangular band, that is to say a band
in which xyx = x for all x, y ∈ S, then W (s) = V (s) = S for each s ∈ S. Notice that if
s′ ∈ W (s) then s′ s, ss′ ∈ E. Moreover, if e ∈ E then e ∈ W (e). It may of course be the case
that for a given element s ∈ S, W (s) = Ø. We do however have
Lemma 1.1 ([20, Corollary 3.3]) Let S be a semigroup in which E 6= Ø. Then E is a
band if and only if for all s, t ∈ S, W (st) = W (t)W (s).
From the proof of [5, Lemma 7.14] we can deduce
Lemma 1.2 Let S be a semigroup with band of idempotents E. Then for all s ∈ S, s′ ∈
W (s), e ∈ E it follows that ses′ , s′ es ∈ E.
If the conclusions of Lemma 1.2 hold, we say that S is weakly self conjugate. We shall make
frequent use of both the previous properties of semigroups in which E is a band without
further reference. Notice also that if s′ ∈ W (s) then ss′ s ∈ V (s′ ) ⊆ W (s′ ), a fact that we
shall also use frequently. In particular, s′ is a regular element of S.
It was shown by Mitsch [12] that the following is a natural partial order on any semigroup
S
a ≤M b if and only if there exists x, y ∈ S, a = xb = by, xa = ay = a.
Notice that if there exist idempotents e and f such that a = eb = bf then it follows that
a = ea = af and so a ≤M b. If a is a regular element of S then it is easy to check that
e = aa′ x ∈ E and f = ya′ a ∈ E for any a′ ∈ V (a), and that a = eb = bf . Hence if a is
regular then
a ≤M b if and only if there exists e, f ∈ E, a = eb = bf.
In particular, this is true if a ∈ E. It is also worth noting here that if E is a semilattice then
the restriction of ≤M to E is compatible with multiplication, a fact that we shall use later.
Let A be a subset of a semigroup S and define
AωM = {s ∈ S|a ≤M s for some a ∈ A}.
If A = {a} then we will write AωM as aωM . Notice that (AωM )ωM = AωM and that if
A ⊆ B then AωM ⊆ BωM . Also, if A ⊆ BωM then AωM ⊆ BωM . We call AωM the
(ωM −)closure of A and say that A is (ωM −)closed if A = AωM . Notice that if A ⊆ B with
B being (ωM −)closed, then AωM ⊆ B.
If T is a subset of a semigroup S then we say that T is left (resp. right) dense in S if for all
s ∈ S there exists s′ ∈ S such that s′ s ∈ T (resp. ss′ ∈ T ). We say that T is dense in S if
it is both left and right dense in S. We are particularly interested in the case where T = E
the set of idempotents of S and we shall refer to semigroups in which E is dense in S as
E−dense or E−inversive semigroups. This concept was originally studied by Thierrin [19]
and subsequently by a large number of authors (see Mitsch [15] for a useful survey article,
but note that the term E−dense has a slightly different meaning there). Included in this
class of semigroup are the classes of all regular semigroups, inverse semigroups, groups,
eventually regular semigroups (that is to say every element has a power that is regular),
periodic semigroups (every element is of finite order) and finite semigroups.
Let S be a semigroup, let L(s) = {s′ ∈ S|s′ s ∈ E}. Then it is well known that W (s) ⊆ L(s).
Moreover, for each s′ ∈ L(s), s′ ss′ ∈ W (s) and so W (s) 6= Ø if and only if L(s) 6= Ø. The
following is then immediate.
2
Lemma 1.3 Let S be an E−dense semigroup. Then for all s ∈ S there exists s′ ∈ S such
that s′ s, ss′ ∈ E.
Let S be an E−dense semigroup with a band of idempotents E and define a partial order
on S by
s ≤ t if and only if either s = t or there exists e, f ∈ E with s = te = f t
and note that ≤ ⊆ ≤M and that if E is a semilattice then ≤ is compatible with multiplication
by weak inverses. If s is regular (in particular idempotent) then s ≤ t if and only if s ≤M t.
If A is a subset of S then define
Aω = {s ∈ S|a ≤ s for some a ∈ A}
and notice that A ⊆ Aω ⊆ AωM . It is also clear that (Aω)ω = Aω. Note from above that
if A ⊆ E then AωM = Aω. We shall make use of ω in Section 2.
Weak inverses of elements will not in general be unique and in section 2 we will often need
to deal with more than one weak inverse of a given element. The following useful result is
easy to establish.
Lemma 1.4 Let S be an E−dense semigroup with a semilattice of idempotents E.
1. If s′ ∈ W (s) and e, f ∈ E then es′ f ∈ W (s).
2. If s′ , s∗ ∈ W (s) then s′ ss∗ ∈ W (s) and s′ ss∗ = s′ ∧ s∗ .
3. If s′ ∈ W (s) and s′∗ ∈ W (s′ ) then s′∗ = s′∗ s′ s = ss′ s′∗ and for all e ∈ E, es′∗ ≤ s. In
particular s′∗ ≤ s.
4. If s′ ∈ W (s) then W (s′ ) = sW (s)s and V (s′ ) = {ss′ s}. In particular if s∗ ∈ W (s)
then W (s′ ) = W (s∗ ) and ss′ ss∗ s ∈ W (s′ ).
5. Let W = {s′ ∈ W (s)|s ∈ S}. Then W is an inverse subsemigroup of S.
6. For all s ∈ S, W (W (W (s))) = W (s).
Proof.
Let E and S be as stated.
1. Suppose that e, f ∈ E and s′ ∈ W (s). Then (es′ f )s(es′ f ) = es′ f ss′ f = es′ f and so
es′ f ∈ W (s)
2. This is straightforward on noting that (s′ ss∗ )s(s′ ss∗ ) = s′ ss′ ss∗ ss∗ = s′ ss∗ . Note that
s′ ss∗ = s∗ ss′ . It is clear that s′ ss∗ ≤ s′ , s∗ , so suppose that t ≤ s′ , s∗ . Then there exist
e, f, g, h ∈ E(S) such that t = es′ = s′ f = gs∗ = s∗ h. Hence since t = s′ st = (s′ ss∗ )h
and t = tss′ = gs∗ ss′ = g(s′ ss∗ ) then t ≤ s′ ss∗ as required.
3. If s′∗ ∈ W (s′ ) then s′∗ = s′∗ s′ s′∗ = s′∗ s′ ss′ s′∗ = ss′ s′∗ s′ s′∗ = ss′ s′∗ = s′∗ s′ s′∗ s′ s =
s′∗ s′ s. Finally notice that for all e ∈ E,
s(s′ s′∗ s′ es′∗ s′ s) = (ss′ s′∗ s′ es′∗ s′ )s = s′∗ s′ es′∗ = es′∗ s′ s = es′∗
and so es′∗ ≤ s.
4. Clearly sW (s)s ⊆ W (s′ ). Let s′∗ ∈ W (s′ ) so that by part (3), s′∗ = ss′ s′∗ = ss′ s′∗ s′ s
and s′ s′∗ s′ ∈ W (s).
Now let s′∗ ∈ V (s′ ) ⊆ W (s′ ). Since s′ s′∗ = s′ s′∗ s′ s = s′ s then s′∗ = ss′ s′∗ = ss′ s.
If s∗ ∈ W (s) then W (s∗ ) = sW (s)s = W (s′ ) and since s′ ss∗ ∈ W (s) then ss′ ss∗ s ∈
W (s′ ).
3
5. Clearly W 6= Ø. Let s′ , t′ ∈ W with s′ ∈ W (s), t′ ∈ W (t) then it is easy to check that
s′ t′ ∈ W (ts) ⊆ W and by part (4), |V (s′ )| = 1 for each s′ ∈ W and so W is an inverse
subsemigroup of S. Alternatively, note that W = Reg(S) the set of regular elements
of S.
6. Let s′ ∈ W (s), s′∗ ∈ W (s′ ), s′∗ ′ ∈ W (s′∗ ). Then by part (3), s′∗ ′ ss′∗ = s′∗ ′ ss′ s′∗ s′∗ ′ =
′
′
′
s′∗ s′∗ s′∗ = s′∗ . On the other hand s′ = s′ ss′ = s′ (ss′ s)s′ ∈ W (W (s′ )) and the result
follows easily.
′
Lemma 1.5 ([13, Proposition 2]) A semigroup S is a group if and only if for every element s ∈ S, |L(s)| = 1.
It is also easy to see that
Lemma 1.6 Let S be an E−dense monoid. Then S is a group if and only if |E| = 1.
A subset A of a semigroup S is called unitary in S if whenever sa ∈ A or as ∈ A it necessarily
follows that s ∈ A. If E is a unitary subset of S then we shall refer to S as an E−unitary
semigroup.
Lemma 1.7 ([17], [15, Theorem 6.8]) Let S be an E−dense semigroup.
E−unitary if and only if E is a band and Eω = E.
Then S is
Lemma 1.8 ([3, Proposition 1.2]) Let S be an E−unitary semigroup. For all s ∈ S, if
s′ ∈ L(s) then s ∈ L(s′ ).
2
E−dense actions of E−dense semigroups
In this section we take inspiration from the theory of inverse semigroup actions, which in
turn is based on Schein’s representation theory of inverse semigroups by partial one-to-one
maps (see [1] and [7]).
Let S be an E−dense semigroup, let X be a non-empty set and let φ : S × X → X be a
partial map with the property that φ(st, x) exists if and only if φ(s, φ(t, x)) exists and then
φ(st, x) = φ(s, φ(t, x)).
We will, as is usual, denote φ(s, x) as sx and simply write (st)x as stx when appropriate.
By a partial map we of course mean that not every element of S need act on every element
of X. A more formal definition can be found in [7]. We say that φ is an E−dense action of
S on X, and refer to X as an E−dense S−act, if
1. the action is cancellative; meaning that whenever sx = sy then x = y;
2. the action is reflexive; that is to say, for each s ∈ S, if sx exists then there exists
s′ ∈ W (s) such that s′ (sx) exists.
The domain of an element s ∈ S is the set
DsX = {x ∈ X|sx exists}.
We shall denote DsX as simply Ds when the context is clear. We shall denote the domain
of an element x ∈ X by
Dx = {s ∈ S|sx exists}.
4
Clearly x ∈ Ds if and only if s ∈ Dx . Notice also that it follows from the definition that
x ∈ Ds if and only if x ∈ Ds′ s for some s′ ∈ W (s).
If S is a group then an E−dense act X is simply an S−set, while if S is an inverse semigroup
then an E−dense act is an inverse semigroup act defined by the Wagner-Preston representation ρ : S → I(X) where sx = ρ(s)(x) and Ds = dom(ρ(s)) (see Example 2.2 below for a
generalisation).
Let X be an E−dense S−act and let x ∈ X. We define the stabilizer of an element x as the
set Sx = {s ∈ S|sx = x}. The following is easy to establish.
Lemma 2.1 Let S be an E−dense semigroup and X an E−dense S−act. Let s, t ∈ S, x, y ∈
X. Then
1. E ∩ Dx ⊆ Sx ,
2. if s′ ∈ W (s) then x ∈ Ds′ if and only if x ∈ Dss′ ,
3. if s ∈ Dx then sx = y if and only if there exists s′ ∈ W (s) ∩ Dy such that x = s′ y,
4. if s, t ∈ Dx then sx = tx if and only if there exists s′ ∈ W (s) such that s′ t ∈ Sx . In
addition, any such s′ necessarily satisfies s′ ∈ Dsx ,
Example 2.2 (Wagner-Preston action) Let S be an E−dense semigroup with semilattice of
idempotents E and X a set on which S acts (on the left) via the representation ρ : S → T (X).
In other words the action on X is a total action. For each s ∈ S define
Ds = {x ∈ X| there exists s′ ∈ W (s), x = s′ sx} = {s′ sx|x ∈ X, s′ ∈ W (s)}
and define an E−dense action of S on X by s ∗ x = sx for all x ∈ Ds .
To see that ∗ really is an E−dense action suppose that x ∈ Dst so that there exists (st)′ ∈
W (st) such that x = (st)′ (st)x. By Lemma 1.1 there exists s′ ∈ W (s), t′ ∈ W (t) such that
(st)′ = t′ s′ and so x = t′ s′ stx. Then t′ tx = t′ tt′ s′ stx = t′ s′ stx = x and so x ∈ Dt . In
addition, s′ stx = s′ stt′ s′ stx = tt′ s′ stx = tx and tx ∈ Ds . Conversely, suppose that x ∈ Dt
and tx ∈ Ds . Then there exists t′ ∈ W (t), s′ ∈ W (s) such that x = t′ tx and tx = s′ stx and
so x = t′ s′ stx ∈ Dst . Clearly, (st) ∗ x = s ∗ (t ∗ x). Finally, if x, y ∈ Ds and s ∗ x = s ∗ y
then there exists s′ ∈ W (s), s∗ ∈ W (s) such that x = s′ sx, y = s∗ sy and such that sx = sy.
Hence
x = s′ sx = s′ sy = s′ ss∗ sy = s∗ ss′ sy = s∗ ss′ sx = s∗ sx = s∗ sy = y.
In addition, if x ∈ Ds then x = s′ sx for some s′ ∈ W (s), and so letting (s′ s)′ = s′ s ∈ W (s′ s)
then
(s′ s)′ (s′ s)x = s′ sx = x
and x ∈ Ds′ s as required. Hence ∗ satisfies the conditions of an E−dense action.
In particular, we can take X = S, or indeed any left ideal of S, with (total) action given by
the multiplication in S.
A element x of X is said to be effective if Dx 6= Ø. An E−dense S−act X is effective if all
its elements are effective. An E−dense S−act is transitive if for all x, y ∈ X, there exists
s ∈ S with y = sx. Notice that this is equivalent to X being locally cyclic in the sense
that for all x, y ∈ X there exists z ∈ X, s, t ∈ Dz with x = sz, y = tz. We shall consider
transitive acts in more detail in Section 2.2.
If X is an E−dense S−act and Y is a subset of X then we shall say that Y is an E−dense
S−subact of X if for all s ∈ S, y ∈ DsX ∩ Y ⇒ sy ∈ Y . Notice that this makes Y an E−dense
S−act with the action that induced from X and DsY = DsX ∩ Y for all s ∈ S.
5
Let X and Y be two E−dense S−acts. A function f : X → Y is called an (E−dense)
S−map if for all s ∈ S, x ∈ DsX if and only if f (x) ∈ DsY and then f (sx) = sf (x).
For example, if Y is an S−subact of an E−dense S−act X, then the inclusion map ι : Y → X
is an S−map.
Let x ∈ X and define the S−orbit of x as
Sx = {sx|s ∈ Dx } ∪ {x}.
Notice that if x is effective, then there exists s ∈ Dx and so for any s′ ∈ W (s) ∩ Dsx , x =
s′ sx ∈ {sx|s ∈ Dx } = Sx. However if x is not effective then {sx|s ∈ Dx } = Ø and Sx = {x}.
Notice also that Sx is an E−dense S−subact of X (the subact generated by x) and that the
X
action is such that, for all tx ∈ Sx and all s ∈ S, tx ∈ DsSx if and only if x ∈ Dst
and in
which case s(tx) = (st)x. Then we have
Lemma 2.3 For all x ∈ X, if x is effective then so is Sx, in which case Sx is a transitive
E−dense S−act. Conversely, if an E−dense S−act is effective and transitive then it has
only one S−orbit.
Proof.
Suppose that x is effective. Then let s ∈ Dx so that sx ∈ Sx, and notice that
there exists s′ ∈ W (s) ∩ Dsx . Therefore ss′ (sx) = sx ∈ Sx and hence Sx is effective. If
y = s1 x and z = s2 x then put t = s1 s′2 , where s′2 ∈ W (s2 ) ∩ Dz , to get y = tz (if x = y 6= z
then take t = s′2 ; if x = z 6= y then take t = s1 while if x = y = z take t = s′ s where
s ∈ Dx , s′ ∈ W (s) ∩ Dsx ).
The converse is easy. Note that in this case Sx = {sx|s ∈ Dx }.
Notice that Sx = Sy if and only if y ∈ Sx and so the orbits partition X.
Recall that Green’s L−relation is given by aLb if and only if S 1 a = S 1 b. As is normal, we
shall denote the L−class containing a as La .
Proposition 2.4 Let S be an E−dense semigroup with semilattice of idempotents E and
consider S as an E−dense S−act with the Wagner-Preston action.
1. If e ∈ E then Se = eω and Se = Le .
2. For all s ∈ S and for all s′ ∈ W (s), Ss ⊆ (ss′ )ω and Ss ⊆ Ls . In addition Ss = (ss′ )ω
for some s′ ∈ W (s) if and only if s is regular, in which case Ss = Ls and we can
assume that s′ ∈ V (s).
3. For all se ∈ Se (the orbit of e), Sse = (ses′ )ω for some s′ ∈ W (s) and Sse = Lse .
4. For all s ∈ S, s′ ∈ W (s) it follows that Ss′ = (s′ s)ω and Ss′ = Ls′ .
Proof.
1. If t ∈ Se then there exists t′ ∈ W (t) such that e = t′ te = t′ e. Hence
′
e = e(t t) = (tet′ )t and so e ≤ t and t ∈ eω.
Conversely, if e ≤ t then there exists f, g ∈ E such that e = f t = tg and it is easy to
check that e = et = te. Since e ∈ W (e) then there exist e′ ∈ W (e), t′ ∈ W (t) such
that e = e′ t′ and so t′ te = t′ tee′ t′ = ee′ t′ = e and t ∈ De and since te = e then t ∈ Se
as required.
If te ∈ Se then there exists t′ ∈ W (t) such that e = t′ te. Hence teLe. On the other
hand, if sLe then there exist u, v ∈ S 1 such that us = e, ve = s, from which we deduce
that se = s. If s = e then obviously s ∈ Se, otherwise note that s′ = eu ∈ W (s) and
so since s′ se = euse = e then s ∈ De and s = se ∈ Se and hence the orbit of e and
L−class containing e coincide.
6
2. Let t ∈ Ss so that there exists t′ ∈ W (t) such that s = t′ ts and ts = s. Then
ss′ = tss′ = t(t′ tss′ ) = (tss′ t′ )t
and so t ∈ (ss′ )ω. If rs ∈ Ss then there exists r′ ∈ W (r) such that s = r′ rs and so
rsLs and Ss ⊆ Ls .
If (ss′ )ω ⊆ Ss then in particular ss′ ∈ Ss and so s′ ∈ Ds . Hence there exists s′∗ ∈
W (s′ ) such that s = s′∗ s′ s and so s = s′∗ ∈ W (s′ ) which means that s is regular and
s′ ∈ V (s). Conversely, if s is regular then there exists s′ ∈ V (s) and so ss′ ∈ Ss .
Hence (ss′ )ω ⊆ Ss since Ss is closed. In this case, since L is a right congruence, then
for any s′ ∈ V (s)
′
ts ∈ Ss ⇐⇒ t ∈ Ds ⇐⇒ t ∈ Dss ⇐⇒ tss′ ∈ Sss′ ⇐⇒ tss′ Lss′ ⇐⇒ tsLs
Hence Ss = Ls .
3. This follows from part (2) since se is regular.
4. Since s′ is regular then from part(2) there exists s′∗ ∈ V (s′ ) such that Ss′ = (s′ s′∗ )ω
and Ss′ = Ls′ . But from Lemma 1.4, s′∗ = ss′ s and the result follows.
Let x ∈ X and set E x = E ∩ Dx . In analogy with group theory, and following [6], we shall
say that an E−dense S−act X is locally free if for all x ∈ X, Sx = (E x )ω.
Theorem 2.5 Let S be an E−dense semigroup with semilattice of idempotents E and let
X be an E−dense S−act. Then X is locally free if and only if for all x ∈ X, s, t ∈ Dx ,
whenever sx = tx there exists e ∈ Sx such that se = te.
Proof.
Suppose that S acts locally freely on X and that sx = tx for some x ∈ X, s, t ∈
Dx . Then there exists s′ ∈ W (s) with s′ t ∈ Sx and so there exist f, g ∈ E, e ∈ E x with
(s′ t)g = f (s′ t) = e. Since e ∈ W (e) then there exist t′ ∈ W (t), s′∗ ∈ W (s′ ), f ′ ∈ W (f ) with
e = t′ s′∗ f ′ . Hence since ge = eg = e, t′ te = e and ss′ s′∗ = s′∗ then
se = sf s′ tt′ s′∗ f ′ = tt′ sf s′ tt′ s′∗ f ′ = tt′ ss′ tgt′ s′∗ f ′ = tt′ tgt′ s′∗ f ′ = tt′ tge = te.
Conversely, suppose that s ∈ Sx so that sx = x. Then there exists s′ ∈ W (s) such that sx =
s′ sx. By assumption there exists e ∈ E x such that se = s′ se = es′ s. But ses′ = es′ ss′ = es′
and so se = (ses′ )s ∈ E x and hence s ∈ (E x )ω and X is locally free.
2.1
Graded actions
Let S be an E−dense semigroup with semilattice of idempotents E. We can consider E as
an E−dense S−act with action given by the Munn representation on E. In more detail, let
e ∈ E and let [e] denote the order ideal generated by e. This is the set
[e] = {s ∈ S|s ≤ e} = {s ∈ E|s = es = se} = eE.
The second equality is easy to establish on observing that [e] ⊆ E (see [14, Lemma 2.1]).
Lemma 2.6 Let S be an (E−dense) semigroup with semilattice of idempotents E. Then
for all e ∈ E, [e] = W (e).
7
Proof.
If s ∈ [e] then s = es = se and so ses = s2 = s. Hence [e] ⊆ W (e). Conversely,
if s ∈ W (e) then ses = s and so es, se ∈ E. Hence s = ses = sees ∈ E. Consequently,
s = se = es ∈ [e] and so [e] = W (e).
S
The action of S on E is given as follows. For each s ∈ S define Ds = s′ ∈W (s) [s′ s] and for
each x ∈ Ds ⊆ E define an action s ∗ x = sxs′ where x ∈ [s′ s] with s′ ∈ W (s). Notice that if
x ∈ [s∗ s] ∩ [s′ s] and s′ , s∗ ∈ W (s) then x = s′ sx = xs′ s and x = s∗ sx = xs∗ s. Consequently,
sxs′ = sxs∗ ss′ = ss′ sxs∗ = sxs∗ .
So the action is well-defined. Notice then that if x ∈ Dst then x ≤ (st)′ (st) for some (st)′ ∈
W (st). Since, by Lemma 1.1, W (st) = W (t)W (s) then there exists s′ ∈ W (s), t′ ∈ W (t)
such that x = t′ s′ stx = xt′ s′ st. Hence xt′ t = t′ tx = t′ tt′ s′ stx = t′ s′ sx = x and so x ∈ Dt .
In addition s′ s(txt′ ) = (txt′ )s′ s = txt′ s′ stt′ = txt′ and so t ∗ x ∈ Ds .
Conversely, suppose that x ∈ Dt and t ∗ x ∈ Ds so that x = t′ tx = xt′ t for some t′ ∈ W (t)
and that txt′ = s′ stxt′ = txt′ s′ s for some s′ ∈ W (s). Then
xt′ s′ st = t′ s′ stx = t′ s′ stxt′ t = t′ txt′ t = x
and so x ∈ Dst .
Now, if x ∈ Dst then for some (st)′ ∈ W (st), s′ ∈ W (s), t′ ∈ W (t) we have
(st) ∗ x = (st)x(st)′ = stxt′ s′ = s ∗ (txt′ ) = s ∗ (t ∗ x).
If s∗ x = s∗ y then x = s′ sx = xs′ s, y = s∗ sy = ys∗ s and sxs′ = sys∗ for some s′ , s∗ ∈ W (s).
Hence x = s′ sxs′ s = s′ sys∗ s and so x ≤ y. Dually y ≤ x and so x = y.
Finally, if x ∈ Ds then there exists s′ ∈ W (s) such that x = s′ sx = xs′ s. Since s′ s ∈ W (s′ s)
then it easily follows that x ∈ Ds′ s . Consequently we have established that E is an E−dense
S−act with action given as above.
Let X be an E−dense S−act. Following [18] we say that the action is graded if there exists
a function p : X → E such that for all e ∈ E, De = p−1 ([e]), and refer to p as the grading.
Lemma 2.7 Let S be an E−dense semigroup with semilattice of idempotents E, and X
a graded E−dense S−act. Then X is effective and for all x ∈ X, p(x) is the minimum
idempotent in Sx .
Proof.
Suppose that X is graded with grading p : X → E and let x ∈ X. Then as
x ∈ p−1 ([p(x)]) = Dp(x) for all x ∈ X it follows that X is effective. Notice also that
p(x) ∈ Sx ∩ E. Suppose that there exists e ∈ Sx ∩ E. Then x ∈ De = p−1 ([e]) and so
p(x) ∈ [e]. Hence p(x) ≤ e as required.
The following is fairly clear.
Proposition 2.8 Let S be an E−dense semigroup with semilattice of idempotents E, and
X a graded E−dense S−act with grading p : X → E. Then X is locally free if and only if
for all x ∈ X, Sx = p(x)ω.
Conversely, if X is an E−dense S−act with the property that for all x ∈ X there exists
ex ∈ E with Sx = ex ω, then X is locally free and graded with grading p : X → E given by
p(x) = ex .
8
Proof.
Suppose that X is locally free. If s ∈ (E x )ω then there exists e ∈ E x such that
e ≤ s. Then since e ∈ Sx it follows that p(x) ≤ e ≤ s. On the other hand, it is clear that
p(x)ω ⊆ (E x )ω ⊆ Sx and so Sx = p(x)ω.
If Sx = p(x)ω then clearly Sx ⊆ (E x )ω. But (E x )ω ⊆ Sx and so X is locally free.
The converse follows easily from Lemma 2.7.
Notice that it follows from Lemma 2.7 that the grading function p is unique. Notice also
that if p(x)′ ∈ W (p(x)) ∩ Dp(x)x then p(x)′ p(x) ∈ Sx and so p(x)′ ∈ Sx . Consequently
p(x)p(x)′ ∈ Sx . Moreover p(x) ≤ p(x)′ p(x), p(x)p(x)′ from which we easily deduce that
p(x) = p(x)′ p(x) = p(x)p(x)′ . But then p(x)′ = p(x)′ p(x)p(x)′ = p(x)p(x)′ = p(x).
Lemma 2.9 Let S be an E−dense semigroup with semilattice of idempotents E and X a
graded E−dense S−act with grading p. Then for x ∈ Ds , if s′ s = p(x) for s′ ∈ W (s) then
ss′ = p(sx).
Proof.
Suppose that s′ s = p(x). Then x ∈ Ds′ s and so x ∈ Ds . In addition, sx ∈ Ds′
′
and so ss ∈ Ssx which means that p(sx) ≤ ss′ . Now s′ s = s′ ss′ s ≥ s′ p(sx)s (since E is a
semilattice). But since s′ p(sx)s ∈ Sx ∩ E then by Lemma 2.7, p(x) = s′ s = s′ p(sx)s and
so ss′ = sp(x)s′ = ss′ p(sx)ss′ , or in other words ss′ ≤ p(sx). But as ss′ ≥ p(sx) then
p(sx) = ss′ as required.
Corollary 2.10 Let S be an E−dense semigroup with semilattice of idempotents E and X a
graded E−dense S−act with grading p. Let s ∈ S and x ∈ Ds . Then for all s′ ∈ W (s)∩Dsx ,
p(sx) = sp(x)s′ .
Proof.
Let t = sp(x) and let s′ ∈ W (s) ∩ Dsx . Then t′ = p(x)s′ ∈ W (sp(x)) = W (t).
′
Hence t t = p(x)s′ sp(x) = s′ sp(x) = p(x) as s′ s ∈ Sx . In addition, tx = sp(x)x = sx and so
by Lemma 2.9, p(sx) = p(tx) = tt′ = sp(x)p(x)s′ = sp(x)s′ as required.
Proposition 2.11 (Cf. [18, Proposition 1.1]) Let S be an E−dense semigroup with semilatticeS
of idempotents E and X a graded E−dense S−act
with grading p and let s ∈ S. Then
S
Ds = s′ ∈W (s) p−1 ([s′ s]) and sX = {sx|x ∈ Ds } = s′ ∈W (s) p−1 ([ss′ ]).
Let s ∈ S, x ∈ Ds , s′ ∈ W (s) ∩ Dsx . Then x ∈ Ds′ s = p−1 ([s′ s]) and so Ds =
([s′ s]). Since p(sx) = sp(x)s′ = (sp(x)s′ )(ss′ ) = (ss′ )(sp(x)s′ ) ≤ ss′ then
s′ ∈W (s) p
′
p(sx) ∈ [ss ] and so sx ∈ p−1S
([ss′ ]). Conversely, if y ∈ p−1 ([ss′ ]) = Dss′ then y = ss′ y = sx
′
where x = s y. Hence sX = s′ ∈W (s) p−1 ([ss′ ]).
Proof.
S
−1
Theorem 2.12 Let S be an E−dense semigroup with semilattice of idempotents E and X
an E−dense S−act. The following are equivalent.
1. X is a graded E−dense S−act,
2. there exists an E−dense S−map f : X → E,
3. X is an effective E−dense S−act and for all x ∈ X, Sx contains a minimum idempotent.
9
X
′
sx
Proof.
(1) =⇒ (2). If x ∈ D
Ss then from Corollary 2.10, for all s ∈ W (s)∩D , p(sx) =
sp(x)s′ = s ∗ p(x) and p(x) ∈ s′ ∈W (s) [s′ s]. Hence p(x) ∈ DsE . Conversely, if p(x) ∈ DsE
S
then p(x) ∈ s′ ∈W (s) [s′ s] and so there exists s′ ∈ W (s) such that x ∈ p−1 ([s′ s]) ⊆ DsX . In
addition s ∗ p(x) = sp(x)s′ = p(sx) and it follows that p is an S−map.
(2) =⇒ (3). Suppose that X is an S−act with an E−dense S−map f : X → E and let
x ∈ X. Then as f (x) ∈ DfE(x) , it follows that x ∈ DfX(x) and X is effective. Notice also that
f (x) ∈ Sx ∩ E. Suppose then that there exists e ∈ Sx ∩ E. Then x ∈ DeX so f (x) ∈ DeE = [e]
and so f (x) ≤ e as required.
(3) =⇒ (1). If X is an effective E−dense S−act and for all x ∈ S, Sx contains a minimum
idempotent, say ex , then define a function p : X → E by p(x) = ex . Suppose then that
e ∈ E and x ∈ DeX . Then e ∈ Sx ∩ E and so p(x) ≤ e. Hence p(x) ∈ [e] or in other words
x ∈ p−1 ([e]) and so DeX ⊆ p−1 ([e]). On the other hand, if x ∈ p−1 ([e]) then p(x) ∈ [e] and
so p(x) = p(x)e and since x ∈ Dp(x) then x ∈ De as well. Hence De = p−1 ([e]) and p is a
grading.
It is easy to check that E is a graded E−dense S−act with grading 1E : E → E, the identity
function. The following is clear.
Corollary 2.13 If X is an E−dense S−act and E is finite then X is graded.
Let (X, p) and (Y, q) be graded E−dense S−acts with grading functions p and q. A graded
morphism is an E−dense S−map f : X → Y such that qf = p. It is clear that graded
E−dense S−acts and graded morphisms form a category and that (E, 1E ) is a terminal
object in this category.
2.2
Transitive S−acts
An E−dense S−act is called indecomposable if it cannot be written as the coproduct (i.e.
disjoint union) of two other E−dense S−acts. In particular, a transitive S−act is easily seen
to be indecomposable. Conversely, if X is indecomposable, then suppose that Y = X \ Sx 6=
Ø for some x ∈ X. Then Y cannot be a subact of X as X is indecomposable, so there
exists y ∈ Y, s ∈ S with sy ∈ Sx and hence y ∈ Sx, a contradiction. Therefore X = Sx is
transitive. The transitive S−acts are therefore the ‘building blocks’ of E−dense S−acts. In
this section, we restrict our attention, in the main, to those E−dense semigroups where E
is a semilattice.
Suppose that S is an E−dense semigroup and that H is a subsemigroup of S. If for all
h ∈ H, W (h) ∩ H 6= Ø then we will refer to H as an E−dense subsemigroup of S. For
example, if E is a band then E is an E−dense subsemigroup of S.
Lemma 2.14 Let S be an E−dense semigroup with semilattice of idempotents E and let H
be an E−dense subsemigroup of S. Then Hω is an E−dense subsemigroup of S.
Proof.
Suppose that x, y ∈ Hω so that there exist a, b ∈ H such that a ≤ x, b ≤ y. In
addition, there exists a′ ∈ W (a) ∩ H, b′ ∈ W (b) ∩ H. Hence there exists e, f, g, h ∈ E such
that a = xe = f x, b = yg = hy. Let x′ ∈ W (x), f ′ ∈ W (f ), y ′ ∈ W (y), h′ ∈ W (h) be such
that a′ = x′ f ′ ∈ W (a), b′ = y ′ h′ ∈ W (b). Then
(xy)(y ′ h′ hx′ f ′ f xy) = (xyy ′ h′ hx′ f ′ f )(xy) = (f ′ f )(xhyy ′ h′ x′ )(xhy) =
(f ′ f )(f xhyy ′ h′ x′ )(xhy) = (f xhyy ′ h′ x′ )(f ′ f xhy) = abb′ a′ ab ∈ H
10
and so xy ∈ Hω and Hω is a subsemigroup of S. Now suppose that x ∈ Hω so that there
exists h ∈ H and e, f ∈ E such that h = ex = xf . Suppose also that h′ ∈ W (h) ∩ H so that
there exists x′ ∈ W (x), f ′ ∈ W (f ) such that h′ = f ′ x′ . Then
x′ (xf f ′ x′ ) = (x′ xf f ′ )x′ = f f ′ x′ xf f ′ x′ = f ′ x′ xf f f ′ x′ = h′ hh′ = h′ ∈ H
and so x′ ∈ Hω and Hω is an E−dense subsemigroup of S.
Lemma 2.15 Let S be an E−dense semigroup with semilattice of idempotents E and let H
be an E−dense subsemigroup of S. Let x, y ∈ S, x′ ∈ W (x), y ′ ∈ W (y), e ∈ E. Then
1. if x′ ex ∈ Hω then x′ x ∈ Hω;
2. if x′ ey, y ′ y ∈ Hω then x′ y ∈ Hω.
Proof.
Let x, y ∈ S, x′ ∈ W (x), y ′ ∈ W (y), e ∈ E. Notice that by Lemma 2.14, Hω is an
E−dense subsemigroup of S.
1. By assumption there exists f, g ∈ E, a ∈ H such that a = (x′ ex)f = g(x′ ex). Consequently
a = x′ exf = (x′ x)(x′ exf ) = (x′ exf )(x′ x)
and so a ≤ x′ x and x′ x ∈ Hω.
2. By assumption there exists f, g ∈ E, a ∈ H such that a = (x′ ey)f = g(x′ ey). Consequently
ay ′ y = x′ eyf y ′ y = x′ eyy ′ yf = (x′ y)(y ′ eyf ) = x′ xx′ eyf y ′ y = (x′ eyf y ′ x)(x′ y)
and so ay ′ y ≤ x′ y and x′ y ∈ Hω as required.
Proposition 2.16 Let S be an E−dense semigroup with semilattice of idempotents E and
let H be an E−dense subsemigroup of S. Then the following are equivalent
1. H is ω−closed in S;
2. H is unitary in S;
3. H is ωM −closed in S.
Proof.
(1) =⇒ (2). Suppose that H is ω−closed in S and suppose that hs = h1 for
some s ∈ S, h, h1 ∈ H. Then there exists h′ ∈ W (h) ∩ H, h′1 ∈ W (h1 ) ∩ H and so there exist
s′ ∈ W (s), h∗ ∈ W (h) such that h′1 = s′ h∗ ∈ W (hs). Then
s(s′ h′ hh∗ hs) = (ss′ h′ hh∗ h)s = h′ hss′ h∗ h1 = h′ h1 h′1 h1 ∈ H
and so s ∈ Hω = H. Consequently H is left unitary in S. The right unitary property
follows in a similar way.
(2) =⇒ (3). Suppose H is unitary in S and that s ≥M h for h ∈ H. Then there
exist x, y ∈ S with h = xs = sy, xh = hy = h. Let h′ ∈ W (h) ∩ H and notice that
h′ hyh′ h = h′ hh′ h = h′ h ∈ H. Therefore y ∈ H and so s ∈ H and H is ωM −closed in S.
(3) =⇒ (1). As H ⊆ Hω ⊆ HωM then this is clear.
11
In view of the above result, we shall simply say that a set A is closed if it is ω−closed.
We briefly review Schein’s theory of partial congruences when applied to E−dense semigroups which have a semilattice of idempotents (see [1, Chapter 7] or [7, Chapter 5] for
more details of the case for inverse semigroups).
Let T ⊆ S be sets and suppose that ρ is an equivalence on T . Then we say that ρ is a partial
equivalence on S with domain T . It is easy to establish that ρ is a partial equivalence on
S if and only if it is symmetric and transitive. If now T is an E−dense subsemigroup of
an E−dense semigroup S and if ρ is left compatible with the multiplication on S (in the
sense that for all s ∈ S, (u, v) ∈ ρ either su, sv ∈ T or su, sv ∈ S \ T and (su, sv) ∈ ρ in the
former case) then ρ is called a left congruence on S and the set T /ρ of ρ−classes will often
be denoted by S/ρ.
Theorem 2.17 Let H be a closed E−dense subsemigroup of an E−dense semigroup S and
suppose that E is a semilattice. Define
πH = {(s, t) ∈ S × S|∃s′ ∈ W (s), s′ t ∈ H}.
Then πH is a left partial congruence on S and the domain of πH is the set DH = {s ∈
S|∃s′ ∈ W (s), s′ s ∈ H}.
The (partial) equivalence classes are the sets (sH)ω for s ∈ DH . The set (sH)ω is the
equivalence class that contains s and in particular H is one of the πH −classes.
Proof.
It is clear that πH is reflexive on DH . Notice first that if there exists s′ ∈ W (s)
such that s′ t ∈ H then s′ ss′ t ∈ H and so since H is unitary, s′ s ∈ H. Suppose then
that (s, t) ∈ πH . Then there exists s′ ∈ W (s), t′ ∈ W (t) such that s′ s, t′ t, s′ t ∈ H. Let
t∗ ∈ W (t), s′∗ ∈ W (s′ ) be such that t∗ s′∗ ∈ W (s′ t) ∩ H. Then let x = t′ s and x′ =
(s′ t)(t′ t) ∈ W (x) so that x′∗ = (t′ t)(t∗ s′∗ ) ∈ W (x′ ) ∩ H. By Lemma 1.4 x′∗ = (t′ t)x′∗ ≤ x
and so t′ s = x ∈ H and hence πH is symmetric.
Now suppose that (s, t), (t, r) ∈ πH . Then there exists s′ ∈ W (s), t′ ∈ W (t), r′ ∈ W (r) such
that s′ s, s′ t, t′ t, t′ r, r′ r ∈ H. Consequently
(s′ r)(r′ tt′ ss′ r) = (s′ rr′ tt′ s)(s′ r) = s′ tt′ rr′ ss′ r = s′ tt′ rr′ r ∈ H
and so s′ r ∈ Hω = H and πH is transitive.
Suppose that (s, t) ∈ πH and that r ∈ S and suppose further that rs, rt ∈ DH . Then there
exists s′ ∈ W (s), t′ ∈ W (t) such that s′ s, t′ t, s′ t ∈ H. Further, there exists s∗ ∈ W (s), t∗ ∈
W (t), r′ , r∗ ∈ W (r) such that s∗ r′ rs, t∗ r∗ rt ∈ H. From Lemma 1.4, s∗ ss′ ∈ W (s) and
so (rs)′ (rt) = s∗ ss′ r′ rt = s∗ r′ rss′ t ∈ H. Hence (rs, rt) ∈ πH and πH is a left partial
congruence on S.
Now suppose that s ∈ (tH)ω where t′ t ∈ H. Then there exists h ∈ H such that th ≤ s and so
there exists e, f ∈ E such that th = se = f s. Hence t′ th = t′ se = t′ f s = t′ tt′ f s = (t′ f t)t′ s
and so t′ th ≤ t′ s and t′ s ∈ Hω = H. Consequently s ∈ [t]πH . On the other hand, if
s ∈ [t]πH then there exists s′ ∈ W (s), t′ ∈ W (t) such that s′ s, t′ t, s′ t ∈ H. Hence there exists
t∗ ∈ W (t), s′∗ ∈ W (s′ ) such that t∗ s′∗ ∈ W (s′ t) ∩ H. Now by Lemma 1.4, tt∗ s′∗ ≤ s and
hence s ∈ (tH)ω.
Finally, if s ∈ DH then there exists s′ ∈ W (s) such that s′ s ∈ H and so s(s′ s) = (ss′ )s ∈ sH
and hence s ∈ (sH)ω. In particular, for all h1 , h2 ∈ H we see that h1 πH h2 and so H =
Hω = (hH)ω for all h ∈ H is an equivalence class.
The sets (sH)ω, for s ∈ DH , are called the left ω−cosets of H in S. The set of all left
ω−cosets is denoted by S/H. Notice that (sH)ω is a left ω−coset if and only if there exists
s′ ∈ W (s) such that s′ s ∈ H. The following is then immediate.
12
Proposition 2.18 Let H be a closed E−dense subsemigroup of an E−dense semigroup S
in which E is a semilattice, and let (aH)ω, (bH)ω be left ω−cosets of H. Then the following
statements are equivalent:
1. (aH)ω = (bH)ω;
2. aπH b that is, there exists b′ ∈ W (b), b′ a ∈ H;
3. a ∈ (bH)ω;
4. b ∈ (aH)ω.
Lemma 2.19 With H and S as in Theorem 2.17,
1. precisely one left ω−coset, namely H, contains idempotents,
2. each left ω−coset is closed,
3. πH is left cancellative i.e. xaπH xb implies that aπH b,
4. ((st)H)ω is an ω−coset if and only if (tH)ω and (s((tH)ω))ω are ω−cosets and then
(s((tH)ω))ω = ((st)H)ω.
Proof.
1. If e is an idempotent contained in an ω−coset then there exists e′ ∈ W (e)
′
with e e ∈ H. As e′ e ≤ e then e ∈ Hω = H. As H is an E−dense subsemigroup of S
then for each h ∈ H there exists h′ ∈ W (h) ∩ H and so h′ h ∈ H and hence E(H) 6= Ø.
2. This is clear.
3. Suppose that (xa, xb) ∈ πH so that there exists (xa)′ ∈ W (xa), (xb)′ ∈ W (xb) such
that (xa)′ (xa), (xa)′ (xb) ∈ H. Then there exists x′ ∈ W (x), a′ ∈ W (a) such that
a′ x′ xb ∈ H and a′ x′ xa ∈ H. It follows from Lemma 2.15 that a′ a, a′ b ∈ H. Hence
aπH b.
4. Suppose that ((st)H)ω is an ω−coset, so that there exists (st)′ ∈ W (st) such that
(st)′ st) ∈ H. Then there exist s′ ∈ W (s), t′ ∈ W (t) such that t′ s′ = (st)′ . Since
t′ s′ st ∈ H and H is closed then it follows from Lemma 2.15 that t′ t ∈ H and so
(tH)ω is an ω−coset. If x ∈ (s((tH)ω))ω then there exists y ∈ (tH)ω such that
sy ≤ x and so there exists h ∈ H such that th ≤ y. Hence there exist idempotents
e1 , e2 , f1 , f2 such that sy = e1 x = xf1 and th = e2 y = yf2 . Now let h′ ∈ W (h)
then h′ t′ ∈ W (th) = W (yf2 ) and so there exists f2′ ∈ W (f2 ) and y ′ ∈ W (y) such that
h′ t′ = f2′ y ′ . But y ′ s′ ∈ W (sy) = W (xf1 ) and so there exist f1′ ∈ W (f1 ) and x′ ∈ W (x)
such that y ′ s′ = f1′ x′ . Hence
x(f1 f2 f2′ f1′ f1 f2 x′ x) = (xf1 f2 f2′ f1′ f1 f2 x′ )x
= sthf2′ f1′ x′ xf1 f2 = sthh′ t′ s′ sth ∈ (st)H
and so x ∈ ((st)H)ω.
On the other hand, suppose that x ∈ ((st)H)ω so that there exists e, f ∈ E, h ∈ H such
that ex = xf = sth. Then s(th) ≤ x and th ∈ tH ⊆ (tH)ω and hence x ∈ (s((tH)ω))ω.
Conversely, if (tH)ω and (s((tH)ω))ω are ω−cosets then as t ∈ (tH)ω it follows
that st ∈ s((tH)ω) ⊆ (s((tH)ω))ω. Which means that (s((tH)ω))ω is the ω−coset
containing st and so equals ((st)H)ω.
13
Notice that Sx is a closed E−dense subsemigroup of S for every x ∈ X.
Theorem 2.20 For all x ∈ X, Sx is either empty or a closed E−dense subsemigroup of S.
Proof.
Assume that Sx 6= Ø. If s, t ∈ Sx then x = sx = s(tx) = (st)x and so Sx is a
subsemigroup. Also sx = x implies that x = s′ x for any s′ ∈ W (s) ∩ Dsx and so Sx is an
E−dense subsemigroup of S. Let s ≤ h with s ∈ Sx . Then there exist e, f ∈ E such that
s = he = f h. Consequently e ∈ Dx and h ∈ Dex = Dx and so hx = hex = sx = x which
means that h ∈ Sx . Hence Sx is ω−closed and so closed.
From Lemma 2.19 we can easily deduce the following important result.
Theorem 2.21 If H is a closed E−dense subsemigroup of an E−dense semigroup S with
semilattice of idempotents E then S/H is a transitive E−dense S−act with action given by
s · X = (sX)ω whenever X, sX ∈ S/H. Moreover, it is easy to establish that SHω = H.
Proof.
Let X = (rH)ω be an ω−coset and suppose that s, t ∈ S and that X ∈ Dst .
Then by Lemma 2.19, ((st)X)ω = (st(rH)ω)ω is an ω−coset and
(st) · X = ((st)X)ω = ((str)H)ω = s · (t · X).
In addition if s · X = s · Y then (sX)ω = (sY )ω and so Lemma 2.19 X = Y . Now suppose
that X ∈ Ds . We are required to show that there exists s′ ∈ W (s) ∩ DsX . Suppose then
that X = (tH)ω so that s · X = ((st)H)ω. Then there exists (st)′ ∈ W (st) such that
(st)′ (st) ∈ H. Hence there exist s′ ∈ W (s), t′ ∈ W (t) such that t′ s′ st ∈ H. Consequently,
t′ s′ (ss′ s)s′ st ∈ H and since t′ s′ (ss′ s) ∈ W (s′ st) then there exists (s′ st)′ ∈ W (s′ st) such that
(s′ st)′ (s′ st) ∈ H and so ((s′ st)H)ω is an ω−coset of H and s′ ∈ W (s) ∩ DsX as required.
If (sH)ω and (tH)ω are ω−cosets then there exist s′ ∈ W (s), t′ ∈ W (t) such that s′ s, t′ t ∈ H.
Now as s′ (ss′ s)t′ = s′ st′ ∈ W (ts′ s) and as s′ st′ ts′ s = t′ ts′ s ∈ H then ((ts′ s)H)ω is an
ω−coset. Moreover, as (ts′ s)t′ t = tt′ (ts′ s) ∈ tH then ts′ s ∈ (tH)ω and so (tH)ω =
((ts′ s)H)ω = (ts′ ) · ((sH)ω) and S/H is transitive.
Finally SHω = {s ∈ S|(sH)ω = Hω}. Hence sπH h for any h ∈ H and so s ∈ H as H is an
ω−coset.
The converse of Theorem 2.21 is also true.
Theorem 2.22 Let S be an E−dense semigroup with semilattice of idempotents E, let X
be an effective, transitive E−dense S−act, let x ∈ X and let H = Sx . Then X is isomorphic
to S/H. If K is a closed E−dense subsemigroup of S and if X is isomorphic to S/K then
there exists x ∈ X such that K = Sx .
Proof.
Let y ∈ X and notice that since X is transitive then there exists s ∈ S such that
y = sx. Notice then that there exists s′ ∈ W (s) ∩ Dsx such that s′ s ∈ Sx = H. Moreover
if y = tx for some t ∈ S then sx = tx and so s′ t ∈ H and hence (sH)ω = (tH)ω. Therefore
we have a well-defined map φ : X → S/H given by
φ(y) = (sH)ω.
Since φ(sx) = (sH)ω for all (sH)ω ∈ S/H then φ is onto. If (sH)ω = (tH)ω then s′ t ∈
H = Sx and so sx = tx and φ is a bijection. Finally, suppose φ(y) = (sH)ω and t ∈ S.
Then y ∈ Dt if and only if t ∈ Dy if and only if ts ∈ Dx if and only if there exists
(ts)′ ∈ W (ts) such that (ts)′ (ts) ∈ Sx = H if and only if ((ts)H)ω is an ω−coset if and only
if φ(y) = (sH)ω ∈ Dt . In this case it is clear that φ(ty) = tφ(y) and φ is an isomorphism.
14
By assumption there is an isomorphism θ : S/K → X. Let x = θ(Kω) so that sx = θ((sK)ω)
for all s ∈ DK . Notice also that since θ is an S−map then s ∈ Dx if and only if s ∈ DK . If
s ∈ Sx then θ(Kω) = θ((sK)ω) and so s ∈ K as θ is an isomorphism. On the other hand,
if s ∈ K then sx = θ((sK)ω) = θ(Kω) = x and so s ∈ Sx as required.
Recall that Le denotes the L−class containing e.
Theorem 2.23 Let S be an E−dense semigroup with a semilattice of idempotents E and
let X be a locally free, transitive, graded E−dense S−act with grading p. Then there exists
e ∈ E such that X ∼
= Se ∼
= Le . Conversely, if e ∈ E then the orbit Se of e in the E−dense
S−act S (with the Wagner-Preston action) is a locally free, transitive, graded E−dense
S−act.
Proof.
If X is transitive then X ∼
= Sx for some (any) x ∈ X. Using a combination of
Theorem 2.21, Proposition 2.8 and Proposition 2.4, we deduce
X∼
= S/Sx = S/p(x)ω ∼
= S/Sp(x) = Sp(x) = Lp(x) .
Conversely, the orbit Se is clearly a transitive E−dense S−act. By Proposition 2.4, Ste =
(tet′ )ω and so by Proposition 2.8 Se is locally free and graded with grading p : Se → E
given by p(te) = tet′ .
Lest X be a graded S−act and let x ∈ X. If p(x) ∈ DsS then there exists s′ ∈ W (s) such
that p(x) = s′ sp(x) and so it follows that x ∈ DsX . Conversely, if x ∈ DsX then there exists
s′ ∈ W (s) ∩ Dsx and so p(x) ≤ s′ s. Consequently p(x) = s′ sp(x) and p(x) ∈ DsS . Hence
the map Sp(x) → Sx given by sp(x) 7→ sx is an S−map which is clearly onto. We have
therefore demonstrated
Proposition 2.24 Let S be an E−dense semigroup with a semilattice of idempotents E
and let X be a graded E−dense S−act. Then X is a quotient of a locally free graded S−act.
The question now arises as to when two transitive E−dense S−acts are isomorphic.
Lemma 2.25 Let H be a closed E−dense subsemigroup of an E−dense semigroup S with a
semilattice of idempotents E. Let (sH)ω be a left ω−coset of H so that there exists s′ ∈ W (s)
such that s′ s ∈ H. Then sHs′ ⊆ S(sH)ω .
Proof.
Let h ∈ H and consider (shs′ ) · ((sH))ω. First notice that (shs′ s)Hω is an
ω−coset as for any h′ ∈ W (h) ∩ H, s′ sh′ s′ ∈ W (shs′ s) and (s′ sh′ s′ )(shs′ s) ∈ H. So
(shs′ ) · ((sH))ω = ((shs′ s)H)ω = (sH)ω and so sHs′ ⊆ S(sH)ω .
If H and K are two closed E−dense subsemigroups of an E−dense semigroup S with
semilattice of idempotents E, then we say that H and K are conjugate if S/H ∼
= S/K (as
E−dense S−acts).
Theorem 2.26 Let H and K be closed E−dense subsemigroups of an E−dense semigroup
S with semilattice of idempotents E. Then H and K are conjugate if and only if there exist
s ∈ S, s′ ∈ W (s) such that
s′ Hs ⊆ K and sKs′ ⊆ H.
Moreover, any such element s necessarily satisfies ss′ ∈ H, s′ s ∈ K.
15
Proof.
Suppose that H and K are conjugate. Then by Theorem 2.21 there is an ω−coset,
(sK)ω say, such that S(sK)ω = H. So by Lemma 2.25 there exists s′ ∈ W (s) such that
s′ s ∈ K and sKs′ ⊆ H. Hence ss′ ∈ H. In addition, for each h ∈ H, (hsK)ω = (sK)ω and
so hsπK s. Consequently s′ hs ∈ K and so s′ Hs ⊆ K as required.
Conversely suppose there exist s ∈ S, s′ ∈ W (s) such that s′ Hs ⊆ K and sKs′ ⊆ H. Then
ss′ Hss′ ⊆ sKs′ ⊆ H. If e ∈ E(H) then ss′ ess′ = ess′ = ss′ e ∈ H and since H is unitary
in S then ss′ ∈ H from which we deduce that s′ s ∈ K. Therefore (sK)ω is an ω−coset of
K in S. Now suppose that t ∈ S(sK)ω . Then ((ts)K)ω = (sK)ω and so tsπK s. Therefore
s′ ts ∈ K and so ss′ tss′ ∈ H and since H is unitary in S we deduce that t ∈ H. Conversely
if t ∈ H then s′ ts ∈ K and so sπK ts or in other words ((ts)K)ω = (sK)ω and t ∈ S(sK)ω .
Hence H = S(sK)ω . Define φ : S/H → S/K by φ((tH)ω) = (tsK)ω and notice that φ is a
well-defined morphism. To see this note that there exists t′ ∈ W (t) with t′ t ∈ H. It follows
that s′ (t′ t)s ∈ K and since s′ t′ ∈ W (ts) then ((ts)K)ω is an ω−coset. If (tH)ω = (rH)ω
then there exists r′ ∈ W (r) such that r′ r, r′ t ∈ H = S(sK)ω and so (rsK)ω = (tsK)ω.
Finally, as H = S(sK)ω then φ is injective and as S/K is transitive then φ is onto and so an
isomorphism as required.
In fact we can go a bit further
Theorem 2.27 Let H and K be closed E−dense subsemigroups of an E−dense semigroup
S. Then H and K are conjugate if and only if there is exist s ∈ S, s′ ∈ W (s)
(s′ Hs)ω = K and (sKs′ )ω = H.
Moreover, any such element s necessarily satisfies ss′ ∈ H, s′ s ∈ K.
Proof.
From Theorem 2.26, if H and K are conjugate, then there exists s ∈ S, s′ ∈ W (s)
such that
s′ Hs ⊆ K and sKs′ ⊆ H.
Now it is clear that (s′ Hs)ω ⊆ K so let k ∈ K, k ′ ∈ W (k) ∩ K and let l = skk ′ ks′ ∈ H. Now
put m = s′ ls = s′ skk ′ ks′ s ∈ s′ Hs and notice that m ≤ k and so k ∈ (s′ Hs)ω as required.
Notice that if ss′ ∈ H then s′ Hs is an E−dense subsemigroup of H. To see this note that
it is clearly a subsemigroup and that s′ h′ ss′ s ∈ W (s′ hs) ∩ s′ Hs for any h′ ∈ W (h) ∩ H. In
particular from Theorem 2.20 we immediately deduce
Proposition 2.28 Let S be an E−dense semigroup with a semilattice of idempotents E
and let X be an E−dense S−act. Let s ∈ S and x ∈ Ds . Then sSx s′ is an E−dense
subsemigroup of S for any s′ ∈ W (s) ∩ Dsx .
Theorem 2.29 Let S be an E−dense semigroup with semilattice of idempotents E and let
X be an E−dense S−act. Let s ∈ S and x ∈ Ds . Then Sx and Ssx are conjugate.
Proof.
Since Sx = Ssx the result follows from Theorem 2.22. In fact, we have that
(sSx s′ )ω = Ssx for any s′ ∈ W (s) ∩ Dsx .
If H is a closed E−dense subsemigroup of an E−dense semigroup S with semilattice of
idempotents E, then we say that H is self-conjugate if H is only conjugate to itself.
Proposition 2.30 Let H be a closed E−dense subsemigroup of an E−dense semigroup S
with semilattice of idempotents E. Then H is self-conjugate if and only if for all s ∈ S and
all s′ ∈ W (s) such that s′ s ∈ H then sHs′ ⊆ H.
16
Proof.
If s ∈ S and s′ ∈ W (s) with s′ s ∈ H then by Lemma 2.25, sHs′ ⊆ S(sH)ω . By
Theorem 2.22, S(sH)ω is conjugate to H and so since H is self-conjugate, sHs′ ⊆ H.
Conversely, suppose that K is a closed E−dense subsemigroup of S and that K is conjugate
to H. Then by Theorem 2.21 there is an ω−coset, (sH)ω say, such that S(sH)ω = K. Hence
there exists s′ ∈ W (s) such that s′ s ∈ H. If k ∈ K then sπH ks and so s′ ks ∈ H and in
addition, since s′ s ∈ H and sHs′ ⊆ H then ss′ = s(s′ s)s′ ∈ H. Hence s(s′ ks)s′ ∈ sHs′ ⊆ H
and so K ⊆ H as H is unitary in S. Since ss′ = (ss′ s)s′ and since ss′ s ∈ W (s′ ) then
s′ H(ss′ s) ⊆ H and so s′ Hs ⊆ H as H is unitary. Consequently, for all h ∈ H, sπH hs and
so H ⊆ S(sH)ω = K.
An alternative characterisation of self-conjugacy is given by
Proposition 2.31 Let H be a closed E−dense subsemigroup of an E−dense semigroup S
with semilattice of idempotents E. Then H is self-conjugate if and only if for all s, t ∈
S, st ∈ H implies ts ∈ H.
Proof.
If st ∈ H then there exists t′ ∈ W (t), s′ ∈ W (s) such that t′ s′ ∈ W (st) and
′ ′
t s st ∈ H. Hence (t′ t)(t′ s′ st) ∈ H and so t′ t ∈ H. Since tHt′ ⊆ H then tt′ ∈ H. But then
ts(tt′ ) = tstt′ tt′ = t(stt′ t)t′ ∈ tHt′ ⊆ H and so ts ∈ H as H is unitary in S.
Conversely, suppose that s ∈ S, s′ ∈ W (s) with s′ s ∈ H and let h ∈ H. Then s′ (shs′ s) =
(s′ s)h(s′ s) ∈ H and so shs′ = (shs′ s)s′ ∈ H and H is self-conjugate.
If H is self-conjugate then S/H has a richer structure. First notice that
Lemma 2.32 Let H be a self-conjugate closed E−dense subsemigroup of an E−dense semigroup S with semilattice of idempotents E. Then DH is a closed E−dense subsemigroup of
S.
Proof.
Let s, t ∈ DH and s′ ∈ W (s), t′ ∈ W (t) with s′ s, t′ t ∈ H. By Proposition 2.31
′
tt ∈ H and since tt′ t ∈ W (t′ ) then t′ Htt′ t ⊆ H. Then t′ s′ ∈ W (st) and t′ s′ st = t′ s′ stt′ t ∈
t′ Htt′ t ⊆ H so that st ∈ DH . Further, as ss′ ∈ H and as ss′ s ∈ W (s′ ) then s′ ∈ DH .
Hence DH is an E−dense subsemigroup of S. Now suppose that s ≤ r with r ∈ S so that
there exist e, f ∈ E such that s = re = f r. Hence there exists f ′ ∈ W (f ), r′ ∈ W (r) such
that s′ = r′ f ′ Consequently, r′ rs′ s = r′ rr′ f ′ f r = r′ f ′ f r = s′ s ∈ H and so r′ r ∈ H as H is
unitary in S. Hence DH is a closed E−dense subsemigroup of S.
As a consequence we can deduce the following interesting result.
Theorem 2.33 Let H be a self-conjugate closed E−dense subsemigroup of an E−dense
semigroup S with semilattice of idempotents E. Then S/H is a group under the multiplication
((sH)ω)((tH)ω) = ((st)H)ω.
Proof.
The multiplication given is well defined as if (s1 H)ω = (s2 H)ω and (t1 H)ω =
(t2 H)ω with s1 , s2 , t1 , t2 ∈ DH then there exist s′1 ∈ W (s1 ), t′1 ∈ W (t1 ) such that s′1 s2 , t′1 t2 ∈
H and so by Proposition 2.31 we have t2 t′1 ∈ H and s′1 s2 t2 t′1 ∈ H and so t′1 s′1 s2 t2 ∈ H.
Hence ((s1 t1 )H)ω = ((s2 t2 )H)ω as required. Multiplication is clearly associative and it is
easy to see that Hω is the identity. It is also clear that (s′ H)ω ∈ W ((sH)ω) and so S/H
is an E−dense monoid. Let (sH)ω ∈ E(S/H) so that s′ s2 ∈ H. Since s′ s ∈ H and H is
unitary in S then s ∈ H and so |E(S/H)| = 1. Hence by Lemma 1.6, S/H is a group.
In particular, if H is self-conjugate and DH = S then πH is a group congruence on S.
17
Proposition 2.34 Let H be a self-conjugate closed E−dense subsemigroup of an E−dense
semigroup S with semilattice of idempotents E. Then for each s ∈ DH ρs : S/H → S/H
given by ρs (X) = (sX)ω is a bijection. The map ρ : DH → Sym(S/H) given by ρ(s) = ρs
is a semigroup homomorphism with ker(ρ) = πH .
Proof.
If s, t ∈ DH then it is clear that ρst = ρs ρt and so ρ is a homomorphism.
Let X ∈ S/H so that X = (tH)ω for some t ∈ DH . If s ∈ DH then st ∈ DH and
so ρs (X) = (sX)ω = ((st)H)ω ∈ S/H. In addition there exists s′ ∈ DH ∩ W (s) and
so s′ s ∈ DH and ρs′ s (X) = ((s′ s)X)ω = ((s′ st)H)ω = (tH)ω = X. In a similar way
ρss′ (X) = ((ss′ )X)ω = X and so ρs′ is the inverse of ρs .
If (s, t) ∈ ker(ρ) then in particular ρs (Hω) = ρt (Hω) and so (s, t) ∈ πH . Conversely
if (s, t) ∈ πH then (sH)ω = (tH)ω and so since S/H is a group then for any r ∈ DH ,
(srH)ω = (trH)ω or in other words (sX)ω = (tX)ω for any X ∈ S/H. Therefore (s, t) ∈
ker(ρ).
3
Semigroup acts and the discrete log problem
Many modern cryptographic applications make implicit use of the inherent difficulty of
solving the discrete log problem. In this section we consider the problem from an abstract
perspective focussing on the (total) action of semigroups on sets (see [9] for more details of
this approach).
Let S be a semigroup and X a (total) left S−act. Suppose also that the action on X is
cancellative in the sense that for all s ∈ S and all x, y ∈ X if sx = sy then x = y. For each
s ∈ S we shall call the pair (X, s) an S−cryptosystem with encryption (function) x 7→ sx.
We refer to x as the plaintext, sx as the ciphertext and s as the cipher key. Our problem is
then to find a decrypt key t such that (ts)x = x.
If for s ∈ S, x, y ∈ X we have y = sx then we shall refer to s as the discrete log of y to the
base x. The discrete log problem is then to compute s given sx and x. In general of course,
the discrete log of sx may not be unique.
As an example, let S = Up−1 be the group of units of the ring Zp−1 and X = Up the group
of units of Zp with p a prime. For n ∈ S, x ∈ X define n · x = xn mod p. By Fermat’s
little theorem, if x is a unit modulo p, then xp−1 ≡ 1 mod p and since n is coprime to p − 1
then there is a positive integer m such that mn ≡ 1 mod p − 1 and hence xmn ≡ x mod p.
Consequently m is a decrypt key for n. The usefulness of this system lie in the fact that we
know of no efficient, non-quantum algorithms, to solve this particular discrete log problem.
More generally, we can let X be a finite group of order r and let S be the group of units
of the ring Zr . Then the action S × X → X given by (n, x) 7→ xn is the basis of an
S−cryptosystem, in which the inverse of any key n ∈ S can easily be computed using the
Euclidean algorithm. The case when r = pq with p and q being distinct primes, forms the
basis of the RSA public-key encryption system.
There are in fact a number of well-know algorithms or protocols for public key encryption
which depend on the difficulty of solving the discrete log problem. For example
Example 3.1 Massey-Omura
Let S be a commutative semigroup that acts on a set X and suppose that for each s ∈ S
there is an inverse element s−1 with the property that s−1 sx = x for all x ∈ X. Suppose
now that Alice wants to send Bob a secure message x. She chooses a secret random element
of the semigroup s, say and sends Bob the value sx. Bob also chooses a secret random
element of the semigroup t, say and sends Alice the value t(sx). Alice then computes
18
tx = (s−1 s)(tx) = s−1 (t(sx)) and sends this to Bob. Bob then computes x = t−1 (tx) as
required.
We can in fact remove the need for S to be commutative if we assume that X is an
(S, S)−biact. In this case, Alice sends Bob the value sx and Bob sends Alice the value
(sx)t = s(xt). Alice then computes xt = (s−1 s)(xt) = s−1 (s(xt)) and Bob then proceeds as
normal.
The beauty of such a scheme is that the values of s and t are chosen at random, do not need
to be exchanged in advance and do not need to be re-used.
Example 3.2 Generalised ElGamal encryption.
In this system, we again assume that S is a (not necessarily commutative) semigroup that
acts on a set X and that a shared secret key, s ∈ S, has previously (or concurrently)
been exchanged. Alice chooses a secret random value c ∈ S, while Bob chooses a secret
random value d ∈ S and publishes sd as his public key. Alice then sends the pair of values
((c(sd))x, cs) to Bob, who computes (cs)d = c(sd) and hence (c(sd))−1 and so recovers x.
Again the values c and d do not have to be re-used.
It is clear that if S is a group, the inverse element s−1 will always exist, namely the group
inverse. For semigroups in general however this may not always be the case. We require
that the stabilizers Sx be left dense in S in order to guarantee that the inverse key will exist
for all s ∈ S.
Proposition 3.3 Let S be a semigroup and X an S−act. The following are equivalent;
1. for all x ∈ X, Sx is left dense in S,
2. for all x ∈ X, Sx is a transitive S−act and x ∈ Sx,
3. every locally cyclic S−subact of X is transitive and for all x ∈ X, x ∈ Sx.
Proof.
(1) =⇒ (2). For all s, r ∈ S there exists t ∈ S such that tsx = x and so
(rt)sx = rx. Hence Sx is transitive and x ∈ Sx.
(2) =⇒ (3). Let Y be a locally cyclic S−subact of X and let x, y ∈ Y . Then there exists
z ∈ Y such that x, y ∈ Sz and so since Sz is transitive then there exists s ∈ S such that
y = sx as required.
(3) =⇒ (1). Let x ∈ X so that by assumption x ∈ Sx. It is clear that Sx is locally cyclic
and so transitive. Consequently, for all s ∈ S there exists t ∈ S such that t(sx) = x and Sx
is left dense in S.
We also require that X is a cancellative S−act. The following result is straightforward to
prove, but note that we only require S to be E−dense in order to justify (4) =⇒ (1).
Lemma 3.4 Let S be an E−dense semigroup and X an S−act. The following are equivalent
1. X is cancellative,
2. for all x ∈ X, E ⊆ Sx ,
3. for all x ∈ X, Eω ⊆ Sx ,
4. for all s ∈ S, s′ ∈ L(s), x ∈ X then s′ s ∈ Sx .
19
Notice from property (4) that if X is a cancellative S−act then for all x ∈ X, Sx is left dense
in S. So if S is an E−dense semigroup then all cancellative cyclic acts are automatically
transitive. So the question arises as to how we can construct a cancellative S−act over an
E−dense semigroup. We do know the structure of E−dense transitive acts over E−dense
semigroups and these are automatically cancellative. In fact it is then clear that if S is
E−dense, then a total S−act X is cancellative if and only if it is an E−dense S−act in
which for each s ∈ S, Ds = X.
Let S be an E−dense semigroup, let (X, s) be an S−cryptosystem and let s′ , s′′ ∈ L(s).
Then for any x ∈ X we see that
s′ x = (s′′ s)(s′ x) = s′′ (ss′ x) = s′′ x.
As with E−dense S−acts we have
Lemma 3.5 Let S be an E−dense semigroup and let X be a cancellative S−act. Then for
all x ∈ X, Sx is ω−closed.
If K(s, x) = {t ∈ S|ts ∈ Sx }, the decrypt key space, then we know that W (s) ⊆ L(s) ⊆
K(s, x).
Theorem 3.6 Let S be an E−dense semigroup, let (X, s) be an S−cryptosystem and let
x ∈ X. Then
1. K(s, x) is ωM −closed,
2. (Sx W (s)Ssx ) ωM ⊆ K(s, x),
3. If E is a band then (Sx W (s)Ssx ) ω = K(s, x),
4. If S is an inverse semigroup then K(s, x) = Sx s−1 ω.
Proof.
Let S, s and x be as in the statement of the theorem.
1. If t ∈ K(s, x) and if t ≤M r then there exist a, b ∈ S such that t = ar = rb, at = tb = t.
Hence if b′ ∈ W (b) then rsx = r(bb′ sx) = tb′ sx = tbb′ sx = tsx = x and so r ∈ K(s, x).
2. Let t ∈ Sx , s′ ∈ W (s) and let r ∈ Ssx . Then (ts′ r)(sx) = ts′ sx = tx = x and so
ts′ r ∈ K(s, x) and the result then follows by part (1).
3. Let t ∈ K(s, x) and notice that for any s′ ∈ W (s) and any t′ ∈ W (t) it follows that
tss′ t′ t ∈ Sx W (s)Ssx . But ss′ t′ t ∈ E since E is a band and tss′ t′ ∈ E since S is weakly
self-conjugate. Hence tss′ t′ t ≤ t and so K(s, x) ⊆ (Sx W (s)Ssx )ω.
4. If t ∈ Sx then ts−1 = ts−1 ss−1 ∈ Sx L(s)Ssx and so (Sx s−1 )ω ⊆ K(s, x). Conversely,
let t ∈ K(s, x) so that tsx = x. Then tss−1 ∈ Sx s−1 and since tss−1 ≤ t the result
follows.
In particular, if S is a group then K(s, x) = Sx s−1 and so |K(s, x)| = |Sx |. A group S is
said to act freely on a set X if for all x ∈ X, Sx = {1}. Clearly in this case, for each key s
there is then a unique decrypt key s−1 . However if the action is not free then there will be
more than one decrypt key for at least one s ∈ S. Notice that for the classic discrete log
cipher Up−1 × Up → Up , (n, x) 7→ xn , the action is indeed a free action. Notice also that for
any E−dense semigroup S and for all x ∈ X, Eω ⊆ Sx . As with E−dense acts, we shall
say that a cancellative S−act X is locally free if for all x ∈ X, Sx = Eω. This is a different
definition from the usual concept of freeness in S−acts (see [8]).
20
Example 3.7 Let S be an E−dense semigroup with a band of idempotents E and let I be
a left ideal of S. Then I is a locally free S−act.
To see this suppose that s ∈ S, x ∈ I and that sx = x. Then for x′ ∈ W (x), s′ ∈ W (s)
it follows that sxx′ s′ s = xx′ s′ s ∈ E since E is a band. However, sxx′ s′ s = s(xx′ s′ s) =
(sxx′ s′ )s and since xx′ s′ s, sxx′ s′ ∈ E it follows that s ≥ sxx′ s′ s so that Sx ⊆ Eω.
From the point of view of decryption, ideally we need a group acting freely on a set. However,
the point of the discrete log problem is not that it is impossible to solve, but rather that it is
hard to solve. Perhaps if finding one needle in a haystack is hard, then finding two, or at least
a relatively small number, is equally hard. Having said that, we probably wish to minimise
the size of K(s, x) and so if S is an E−dense semigroup then we may wish to consider
those semigroups for which Eω = E, which in the case of those E−dense semigroups with
a band of idempotents is, by Lemma 1.7, an E−unitary semigroup. We shall refer to such
semigroups as E−unitary dense semigroups. Notice that in this case, if X is locally free,
K(s, x) = {t|ts ∈ Sx } = {t|ts ∈ E} = L(s). Notice also that by Lemma 1.5, if |L(s)| = 1
then S is a group.
Proposition 3.8 Suppose that S is an E−unitary dense semigroup with a semilattice of
idempotents E and suppose that X is a locally free cancellative S−act. Then for all s ∈
S, x ∈ X, K(s, x) = (W (s))ω.
Proof.
By Theorem 3.6, K(s, x) = (EW (s)E)ω and by Lemma 1.4, EW (s)E ⊆ W (s).
Hence K(s, x) ⊆ (W (s))ω. But if s′ ≤ t for some s′ ∈ W (s), t ∈ S then there exist e, f ∈ E
such that s′ = et = f t. Consequently f ts = s′ s ∈ E and so ts ∈ E as S is E−unitary.
Hence (W (s))ω ⊆ L(s) = K(x, s) and the result follows.
There have been many results concerning the structure of E−unitary dense semigroups
based on the celebrated results of McAlister ([10], [11]) and we present here a version of
the one first given in [3]. First notice that if S is a semigroup and if 1 S is the monoid
obtained from S by adjoining an identity element 1 (regardless of whether S already has
an identity), then S is an E−unitary dense semigroup if and only if 1 S is an 1 E−unitary
dense monoid. This observation allows us to present the construction for E−unitary dense
monoids, without much loss of generality. We use, for the most part, the terminology of [5].
Let C be a small category, considered as an algebraic object, with a set of objects, Obj C and
a disjoint collection of sets, Mor(u, v) of morphisms, for each pair of objects u, v ∈ Obj C.
The collection of all morphisms of C is denoted by Mor C, for each object u ∈ Obj C
the identity morphism is denoted by 0u and composition of morphisms, denoted by p + q
for p, q ∈ Mor C, is considered as a partial operation on Mor C. Notice that, despite the
notation, we do not assume that + is commutative. For each object u ∈ Obj C the set
Mor(u, u) is a monoid under composition and is called the local monoid of C at u. We shall
say that C is locally idempotent if each local monoid Mor(u, u) is a band, and that C is
strongly connected if for every u, v ∈ Obj C, Mor(u, v) 6= Ø.
Let G be a group. An action of the group G on a category C, is given by a group action on
Obj C and Mor C such that
1. if p ∈ Mor(u, v) then gp ∈ Mor(gu, gv),
2. g(p + q) = gp + gq for all g ∈ G, p, q ∈ Mor C, (whenever both sides are defined),
3. g0u = 0gu for all g ∈ G, u ∈ Obj C.
21
The action is said to be transitive if for all objects u, v ∈ Obj C there exists g ∈ G, gu = v,
and free if the action on the objects if a free action (i.e. Su = {1} for all u ∈ Obj C). Notice
that if the action is both transitive and free then for each pair u, v ∈ Obj C there exists a
unique g ∈ G with gu = v.
Now suppose that C is a strongly connected, locally idempotent category and that the group
G acts transitively and freely on C. Let u ∈ Obj C and let
Cu = {(p, g)|g ∈ G, p ∈ Mor(u, gu)}.
Then Cu is a monoid with multiplication defined by
(p, g)(q, h) = (p + gq, gh).
Theorem 3.9 ([3, Proposition 3.2, Theorem 3.4]) Let S be a monoid with band of
idempotents E. Then S is E−unitary dense if and only if there exists a strongly connected,
locally idempotent category C and a group G that acts transitively and freely on C and S is
isomorphic to Cu for some (any) u ∈ Obj C.
Notice that the idempotents of S correspond to the elements of the form (p, 1). Also,
as K(s, x) = L(s), we see that K((p, g), x) = {(q, g −1 ) ∈ S} and so |K((p, g), x)| =
|Mor(u, g −1 u)|. Consequently we see by Lemma 1.5 that S is a group if and only if for
all g ∈ G, |Mor(u, gu)| = 1. In fact we see from Lemma 1.6 that in order for G not to be a
group we require |E| > 1 (this would not be true if S is not a monoid).
Define the support of the category C to be the underlying graph of C. Now consider the
following category. Let Obj C = S and for u, v ∈ Obj C define Mor(u, v) = {(u, s, v)|s ∈
S, v = su}. This is called the derived category of the monoid S. The support of C is often
called the left Cayley graph of S.
For a specific example of the above construction of an E−unitary dense monoid, let G be a
group and let C be the derived category of G with the induced action of G on C. That is
to say g(u, s, v) = (gu, gsg −1, gv). Then Cu is an E−unitary dense monoid, C is a locally
idempotent category on which G acts transitively and freely and Cu ∼
= G. Notice that
= C1 ∼
in this case every morphism in C is an isomorphism and so C is a groupoid.
01
1
(h, h−1 , 1)
0h
h
(1, h, h)
(g, g −1 , 1)
(1, g, g)
g
0g
If we wish to work with E−unitary dense semigroups rather than monoids, we can simply
remove the need for an identity element in Mor(u, u) (see [3] for more details).
A simple modification of the previous example can provide us with an E−unitary dense
semigroup that is not a group. Let G be a group and in the derived category of G, replace
Mor(u, u) = {0u } with the 2-element band {0u , eu }. Now extend the composition of maps
so that we form a category. In other words for each u ∈ Obj C, g ∈ G add in the maps
eu + (u, g, gu), (u, g, gu) + egu , eu + (u, g, gu) + egu . In addition we can extend the action of
G accordingly so that geu = egu .
22
01
(1, h, h)
eh
e1
1
(1, g, g)
(h, h−1 , 1)(g, g −1 , 1)
h
g
0g
0h
eg
Notice then that (u, g, gu)+egu +(gu, g −1, u) ∈ Mor(u, u) and so must be either equal to 0u or
eu . If it were equal to 0u then we can add (gu, g −1 , u) to the left and (u, g, gu) to the right to
deduce that eu = 0u which is obviously a contradiction. Hence (u, g, gu)+egu +(gu, g −1, u) =
eu and so for all u ∈ Obj C, g ∈ G we have
(u, g, gu) + egu = eu + (u, g, gu).
It then follows that Mor(u, gu) = {(u, g, gu), eu + (u, g, gu)} and C1 is an E−unitary dense
monoid with 2 idempotents and |K(s, x)| = 2 for all s ∈ S and x ∈ X. Notice that we can
view this monoid in the following way. Let G be a group and e a symbol not in G and let
eG = {eg|g ∈ G} be a set in 1-1 correspondence with G. Let S = G ∪˙ eG and extend the
multiplication on G to S by setting e2 = e, eg = ge for all g ∈ G and all other products
defined by associativity or the multiplication in G. Then S ∼
= C1 and the isomorphism
is given by g 7→ ((1, g, g), g), eg 7→ (e1 + (1, g, g), g) = (e1 , 1)((1, g, g), g). The element e
corresponds to (e1 , 1) ∈ C1 .
By replacing Mor(u, u) by a band of any given size, we should be able to construct an
E−unitary dense monoid with any finite number of idempotents.
The above construction gives us a mechanism to build a suitable E−unitary dense semigroup
S. However we need X to be a locally free cancellative S−act, so let us revisit the theory
of E−dense S−acts. If S is finite (or at least E is finite) and E is a semilattice, then every
E−dense actṠis graded and so by Theorem 2.23, X is a locally free E−dense S−act if and
only if X ∼
= Sei for some idempotents ei , where the action is that given in Example 2.2.
As previously observed, if X is a cancellative total act then it is automatically reflexive and
hence an E−dense act. Consequently, if X is locally free then as every idempotent acts on
ei , we can deduce that for each i, ei = f , the minimum idempotent in S. Conversely if f is
the minimum idempotent in S then Sf ∼
= S/f ω is a locally free transitive cancellative total
S−act. We have therefore shown
Theorem 3.10 Let S be a finite E−dense semigroup with semilattice of idempotents E,
let s ∈ S and let f be the minimum idempotent in S. Then (X, s) is a locally free
Ṡ
S−cryptosystem if and only if X ∼
= Sf . In addition, if S is E−unitary then for each
x ∈ X, |K(s, x)| = |(W (s))ω|.
In the above example where S = G ∪˙ eG, the minimum idempotent is e and X = eG = Ge
is a locally free cancellative S−act and for each x ∈ X, |K(s, x)| = 2. In the classic discrete
log cipher, Up−1 acts freely on Up by exponentiation, the minimm idempotent is 1 ∈ Up−1
Ṡ
and in fact Up ∼
= |Up | Up−1 .
23
3.1
Completely Regular Semigroups
In the classic discrete log cipher, a group acts freely on a group by exponentiation. We now
briefly consider a group acting freely on a semigroup by exponentiation. It is clear that the
semigroup needs to be periodic as every element will need to have finite order.
A semigroup S is called completely regular if every element of S belongs to a subgroup of S.
A particular example of such a semigroup is a completely simple semigroup, which by Rees’
Theorem ([7, Theorem 3.2.3]), can be shown to be isomorphic to a Rees Matrix Semigroup.
Indeed a semigroup is completely regular if and only if it is isomorphic to a semilattice
of completely simple semigoups ([7, Theorem 4.1.3]). A semigroup S = M[G; I, Λ; P ] is a
called a Rees Matrix Semigroup if
S =I ×G×Λ
and P = (pλi ) is a Λ × I matrix with entries in the group G, and where multiplication is
given by
(i, g, λ)(j, h, µ) = (i, gpλj h, µ).
If follows that for n ∈ N, (i, g, λ)n = (i, (gpλi )n−1 g, λ). Notice that S is not in general
commutative, even if G is abelian.
Suppose now that S is a completely simple semigroup, considered as a Rees matrix semigroup
M[G; I, Λ; P ] and suppose also that G is finite, of order r so that g r = 1 for all g ∈ G. Define
an action of Ur , the group of units in Zr , on S by n · x = xn , so that if x = (i, g, λ) then
n · x = (i, (gpλi )n−1 g, λ). This action is clearly a free action and group actions are always
cancellative.
Suppose now that n is coprime to r and that mn ≡ 1 mod r. Then
−1
xmn = (i, (gpλi )mn−1 g, λ) = (i, (gpλi )mn p−1
λi , λ) = (i, (gpλi )pλi , λ) = (i, g, λ) = x.
Consequently if we know n, xn and P , then we can compute xmn and so recover x. We can
in fact compute xmn in an efficient manner, as we can deduce the values of i and λ from xn
and so we can compute pλi . Then
m
(gpλi )mn−1 g = (gpλi )mn p−1
(gpλi )n−1 g pλi p−1
λi .
λi =
Suppose now we know x, xn and G. Can we compute n? If we also know P then we know pλi
and so (gpλi )n . Consequently, the discrete log problem in this case is equivalent to that in
the classic discrete log problem. Suppose however that P is secret. We know (gpλi )n−1 g and
we know g and so we can compute (gpλi )n−1 but we don’t know pλi and so can’t obviously
recover the classic discrete log problem from this. However, according to [4], the discrete
log problem here, can be reduced, in polynomial time, to the classic discrete log problem in
a subgroup of S, namely the kernel of the element x.
An alternative strategy might be to utilise the matrix P to form a kind of “Vigenère” version
of the discrete log cipher. Here, the data to be enciphered would be encoded using the group
G alone as with the classic discrete log cipher, and an additional keyword would be used to
generate a large sequence of values (i, λ)j with the j th such pair used to encipher the j th
plaintext block using the scheme above. In principle different data blocks, even if containing
the same value, would produce different ciphertext blocks, thereby potentially increasing the
diffusion.
References
[1] A.H. Clifford and G.B. Preston, The Algebraic Theory of Semigroups II, American
Mathematical Society, Mathematical Surveys 7, AMS, 1967.
24
[2] Ahsan, Javed and Liu Zhongkui, A Homological Approach to the Theory of Monoids,
Science Press, Beijing, (2008).
[3] J. Almeida, J.E. Pin and P. Weil, Semigroups whose idempotents form a subsemigroup,
Math. Proc. Camb. Phil. Soc. (1992), 111241–253.
[4] Matan Banin, Boaz Tsaban, A reduction of Semigroup DLP to Classic DLP, Designs,
Codes and Cryptography, (2016), Volume 81, Issue 1, 75–82.
[5] John Fountain, Jean-Eric Pin, Pascal Weil, Covers for monoids, Journal of Algebra,
271 (2004) 529–586.
[6] Jonathon Funk and Pieter Hofstra, Topos theoretic aspects of semigroup actions, Theory
and applications of Categories, Vol 24, No. 6, 2010, pp. 117–147.
[7] J.M. Howie, Fundamentals of Semigroup Theory, London Mathematical Society Monographs, (OUP, 1995).
[8] Kilp, Mati, Knauer, Ulrich and Alexander V. Mikhalev, Monoids, Acts and Categories,
De Gruyter Expositions in Mathematics (29), (Walter de Gruyter, Berlin, New York,
2000).
[9] Maze, G., Monico, C., Rosenthal, J., Public key cryptography based on semigroup
actions, Adv. Math. Commun., 1(4), 489–507 (1996).
[10] D. B. McAlister. Groups, semilattices and inverse semigroups. I. Trans. Amer. Math.
Soc. 192 (1974) 227-244.
[11] D. B. McAlister. Groups, semilattices and inverse semigroups. II. Trans. Amer. Math.
Soc. 196 (1974) 351-370.
[12] H. Mitsch, A Natural Partial Order for Semigroups, Proceedings of the American Mathematical Society, Vol. 97, No. 3 (Jul., 1986), pp. 384-388.
[13] H. Mitsch, Subdirect Products Of E−inversive Semigroups, J. Austral. Math. Soc.
(Series A) 48 (1990), 66–78
[14] H. Mitsch, Semigroups and their natural order, Mathematica Slovaca, Vol 44 (1994),
No 4, 445-462.
[15] H. Mitsch, Introduction to E−inversive semigroups, in “Proceedings of the International
Conference on Semigroups. Ed. Paula Smith, Emilia Giraldes, Paula Martins, World
Scientific, 1999, 114–135.
[16] J. Renshaw, Inverse semigroups acting on graphs in “Proceedings of the Workshop Semigroups and Languages. Ed. Isabel M. Araújo, Mário J.J. Branco, Vitor H. Fernandes,
Graconda M.S. Gomes, World Scientific, 2004, 212–239.
[17] S. Reither, Die nat urliche Ordnung auf Halbgruppen, University of Vienna, PhD-Thesis
(1994).
[18] Benjamin Steinberg, A note on amalgams of inverse semigroups Journal of the Australian Mathematical Society (2001), 70: 71-75.
[19] G. Thierrin, Demigroupes inversés et rectangularies, Bull. Cl. Sci. Acad. Roy. Belgique,
(1955) 41, 83–92.
25
[20] B. Weipoltshammer, On classes of E−inversive semigroups and semigroups whose idempotents form a subsemigroup, Comunications in Algebra, Vol 32, No 8, pp. 2929–2948
(2004).
26
| 4math.GR
|
Under review as a workshop contribution at ICLR 2015
I MPROVING APPROXIMATE RPCA WITH A K - SPARSITY
PRIOR
arXiv:1412.8291v1 [cs.NE] 29 Dec 2014
Maximilian Karl & Christian Osendorfer
Fakultät für Informatik
Technische Universität München
{karlma,osendorf}@in.tum.de
A BSTRACT
A process centric view of robust PCA (RPCA) allows its fast approximate implementation based on a special form of a deep neural network with weights shared
across all layers. However, empirically this fast approximation to RPCA fails to
find representations that are parsemonious. We resolve these bad local minima
by relaxing the elementwise L1 and L2 priors and instead utilize a structure inducing k-sparsity prior. In a discriminative classification task the newly learned
representations outperform these from the original approximate RPCA formulation significantly.
1
I NTRODUCTION
In this work an efficient implementation of sparse coding is evaluated. Sparse coding is an optimisation problem, where it is to find a sparse representation of given data. Not only is it diffcult to find
a mapping from the sparse code to the data but the search for the perfect sparse code for a single
datapoint is a single extra optimisation procedure. This process is very time intensive because you
need to optimize two problems one after another. The idea behind an efficient implementation of this
sparse coding problem comes from ? ?. A gradient descent algorithm optimizing the sparse code
with many iteration is transformed into a neural network with very few layers, each representing one
iteration of the gradient descent algorithm. This network is then trained using the same objective
function as used for creating the gradient descent iterates creating an efficient version of the initial
optimisation procedure.
The Robust Principal Component Analysis (RPCA or Robust PCA)? ? version of such an efficient
sparse coding network is evaluated. An own optimised version of this algorithm producing much
sparser latent codes is presented and evaluated.
2
ROBUST PCA
The motivation behind Robust PCA is the decomposition of a large matrix into a low rank matrix
and a sparse matrix ?. The sparse matrix is also called outlier matrix, therefore also the name Robust
PCA. This decomposition can be formulated as follows:
M = L0 + S0
(1)
where M is the large data matrix, L0 is the low rank matrix and S0 is the sparse outlier matrix.
In the often used Principal Component Analysis (PCA) a similar problem is solved. However normal
PCA features no outlier matrix. So it tries to minimize kM − LkF subject to rank(L) ≤ k where
only small disturbances are allowed. Single large disturbances could render the low rank matrix
different from the true low rank matrix. Through introducing an outlier matrix these corruptions
could be eliminated helping the low rank matrix capture the information of the real data.
1
Under review as a workshop contribution at ICLR 2015
3
E FFICIENT S PARSE C ODING
The efficient RPCA algorithm? uses the same objective function as the original RPCA formulation?:
1
λ∗
kX − DS − Ok2F +
(kDk2F + kSk2F ) + λkOk1
2
2
This objective function is transformed into a neural network by first deriving the proximal descent
iterations. This means computing the gradient of the smooth part of the objective function wrt S and
O and computing a proximal operator out of the non-smooth part. The proximal operator is defined
as followed ? ?:
λ
1
π λ ψ (z) = argmin ku − zk22 + ψ(u)
α
α
u∈Rm 2
(2)
where ψ(u) is the non-smooth part of the objective function.
Constructing the proximal descent of this objective function results in the following algorithm:
Algorithm 3 RPCA Proximal Descent. With πt (b) = sign(b)max(0, |b| − t). Taken from ?. x is
the input, D is the dictionary, l and o are the low-rank approximation and the outlier
DT
D0 + λ∗ I DT
1
0
0
Define H = I − α
,
D0
I
T
0
D0
W = α1
.
, and t = αλ
1
I
Initialize z0 = 0, b0 = Wx.
for k = 1, 2, . . . until convergence do
zk+1 = πt (bk )
bk+1 = bk + H(zk+1 − zk )
Split zk+1 = (s; o) and output l = D0 s.
Because this iterative algorithm is costly we need an efficient implementation of it. This is done by
unrolling the loop and building a neural network with a fixed size out of it? ?. Each layer of the
neural network represents one iteration of the proximal splitting algorithm. The matrices H and W
can be interpreted as weight matrices. The parameters W , H, t and D can now be trained using
standard optimization techniques from neural networks. This fine-tuning creates iterations that are
more efficient than the original proximal splitting method. One could either train all parameter
at once or constrain H, W and t to train only the dictionary D. Another possibility is to train
different H and W for every layer to create a more powerful model. The focus of this work was
put on first training the dictionary and fixing all other parameters to the proximal splitting algorithm
initialisation.
4
I NSTABILITIES
During evaluating this efficient RPCA network on MNIST some problems arised. The objective
function consists of a reconstruction term, a sparsity term and an outlier term. Optimizing this
RPCA Network resulted in a decrease of this objective function. At the same time the reconstruction
error in this objective function increased which implies a bad reconstruction from the sparse code.
However the output of the network featured every detail of the desired output. This came from the
fact that the network saved all information in the outlier matrix. The sparse code was therefore
completely blank. Changing the parameters to stabilize this problem resulted in non-sparse codes
and good reconstruction which is also not desired.
The problem lies in the regularizer for the sparse code. Here for the sparse code the l2-norm and for
the sparse outlier the l1-norm was used. Both of them only act on single elements of the sparse code
2
Under review as a workshop contribution at ICLR 2015
and outlier. A regularizer selecting some of these elements and applying a regular l1-norm to only
them would solve this problem.
5
K -S PARSE
R EGULARIZER
The solution to this problem is to use the k-sparse function from k-sparse autoencoders? and taking
it as a base for the new regularizer. The k-sparse function selects the k-largest elements of an array
and sets all other elements to zero. This makes it an ideal candidate for building a regularizer which
only applies a l1-norm to some of these elements. The new norm is defined as follows:
kS − kSparse(S, k∗ )k1
It is a l1-norm between the k-sparse operator applied to the sparse code and the sparse code itself. S
is the sparse code and k∗ the parameter regulating the number of non-zero elements. This regularizer
now protects all k-largest elements of the sparse code from the l1-norm.
6
E FFICIENT K -S PARSE C ODING
Instead of just applying the l1-norm to the outliers we now use the k-sparse norm. This allows a
fixed amount of information to be stored in O. This amount of information can be controlled by
the parameter k. This prevents the network from stroring all information in the outlier matrix and
leaving the sparse code empty. To further improve the sparsiness of the sparse code the norm was
also applied to S. Using this k-sparse prior the overall sparse coding objective function changes to:
1
kX − DS − Ok2F + λ∗ kS − kSparse(S, k∗ )k1 + λkO − kSparse(O, k)k1
2
Of course the optimal parameter from RPCA may not be the perfect parameters for the k-sparse instance but it has shown that this new setting is much more robust against variations in the parameters.
Also kDk2F is not present anymore because its minimisation was entangled with the minimisation
of kSk2F since they represent together the minimisation of the rank of DS ?.
When this k-sparse prior is used in the objective function and processed using the proximal descent
framework something interesting happens. Instead of just applying the shrinkage function to every
element now the k-largest values are protected from the shrinkage function. Instead of applying
the k-sparse operator directly on the sparse code as in the k-Sparse Autoencoder setting ? here the
k-sparse function is applied as some kind of soft manner.
The derivation of the proximal operator for the k-sparse coding case with ψ(u) = λ∗ kS −
kSparse(S, k∗ )k1 + λkO − kSparse(O, k)k1 can be splitted in two separate proximal operators
since S and O are independent parts of the vector u. The derivation of the proximal operator for one
single of these vector parts:
π α1 ψ (z)
=
0
=
0
=
z
=
1
1
argmin ku − zk22 + ψ(u)
m
2
α
u∈R
1
λ
∇u ( ku − zk22 + |u − kSparse(u, k)|1 )
2
α
λ
u − z + ∇u |u − kSparse(u, k)|1
α
λ
u + sgn(u − kSparse(u, k))
α
This function needs to be inverted. For elements for which u = kSparse(u, k) applies the proximal
operator is the identity function. In the other case kSparse(u, k) = 0 the proximal function is the
same as in the original RPCA case. This soft k-sparse shrinkage function derived from the objective
function looks like this:
3
Under review as a workshop contribution at ICLR 2015
κt,k (b) = τt,k (b) − τt (kSparse(b, k)) + kSparse(b, k)
where τt,k (b) is the original soft shrinkage function applied at every iteration. kSparse(b, k) is the
original k-sparse function from ?. The complete algorithm looks very similiar to the RPCA case:
Algorithm 4 k-Sparse Proximal Descent. With πt,k∗ ,k (b) =
κt,k∗ (S)
κt,k (O)
. Taken from ? and
modified to match the k-Sparse Proximal Descent.
DT
D0 DT
1
0
0
Define H = I − α
,
D0
I
λ∗
DT
0
, and t = α1
.
W = α1
λ
I
Initialize z0 = 0, b0 = Wx.
for k = 1, 2, . . . until convergence do
zk+1 = πt,k∗ ,k (bk )
bk+1 = bk + H(zk+1 − zk )
Split zk+1 = (s; o) and output l = D0 s.
The differences to the RPCA algorithm is not only the change in the activation function but also in
the matrix H. This matrix does not include the parameter λ∗ anymore. This comes from the fact
that the norm of the sparse code is not part of the smooth part of the objective function but now of
the non-smooth function which only affects the proximal operator. Instead t now incorperates λ∗
since t is derived from the non-smooth part of the objective function.
7
E XPERIMENTS
For the experiments the MNIST1 Dataset was used. The dataset contained no outliers. We were
only interested in the relative performance of the two algorithms. The efficient RPCA and efficient
k-sparse coding model were both trained unsupervised on this dataset. To be able to compare the
quality of these different representations the classification error was choosen. For each representation a supervised logistic regressor was trained to classify the correct type of digit. The errors for
this experiment are shown in 1. They represent the number of falsely classified digits in percent.
The k-sparse coding model is producing suprisingly lower errors compared to RPCA. This shows
k-sparse is producing hidden representations better suited for classification using linear classifiers.
The change of the k parameter shows small changes in classification error and allows some finetuning. Very small values of k results in high error rates since less information can be stored in this
small number of non-zero hidden values. Whereas very high values would result in similar error
rates to the RPCA case since then the objective function is then more similar to the one of k-sparse
coding.
The learned filters of these two unsupervised models are also very interesting. Results from the
RPCA network using 1000 elements large sparse codes are shown in 1. Randomly selected entries
from the learned dictionary are shown in this picture. One can see typical filters just like one would
expect it from standard PCA. In 2 and 3 the dictionaries of two k-sparse coding networks are shown,
one with k = 20 and the other with k = 40. In contrast to the RPCA case now one does not see
global filters, but instead, local filters representing line segments of digits. Larger segments in the
k = 20 case, and smaller ones in the k = 40 case. These filters are very similar to those produced
by the k-sparse Autoencoder from ?.
1
yann.lecun.com/exdb/mnist/
4
Under review as a workshop contribution at ICLR 2015
Figure 1: Some entries from the dictionary for RPCA with 1000 hidden units
Figure 2: Some entries from the dictionary using the new k-sparse prior with 1000 hidden units and
k = 20
Figure 3: Some entries from the dictionary using the new k-sparse prior with 1000 hidden units and
k = 40
5
Under review as a workshop contribution at ICLR 2015
k=100 k=40 k=20
RPCA 7.80
k-sparse
2.87 3.27 3.45
Table 1: Results for logistic regression on the sparse codes of the efficient RPCA and k-sparse
implementation.
8
C ONCLUSION AND FUTURE WORK
The classification quality of an efficient version of RPCA has been presented. An additional addon
was presented solving several problems that arised during the usage of RPCA. This solution consists
of changing the regularizer from a l1-norm to a completely new prior using the k-sparse function.
Due to the mathematical derivation of the network structure from the objective function this new
prior automatically incoperates itself inside the transfer function. Now the sparse code has a much
sparser structure but also the parameter decision got much more stable. This new k-sparse coding
model resulted in much lower classification errors than the original efficient RPCA version.
Future work consists of testing this new k-sparse norm as prior also for regular sparse coding or
non-negative matrix factorization. Another application could be to use it as regularizer for any other
machine learning algorithm.
6
| 9cs.NE
|
arXiv:1312.0325v1 [math.AC] 2 Dec 2013
On the (n, d)th f -Ideals
∗
Jin Guo†, Tongsuo Wu‡
Department of Mathematics, Shanghai Jiaotong University
Abstract. A square-free monomial ideal I is called an f -ideal, if both
δF (I) and δN (I) have the same f -vector, where δF (I) (δN (I), respectively) is the facet (Stanley-Reisner, respectively) complex related to I.
In this paper, we introduce the concepts of perfect set containing k and
perfect set without k. We study the (n, d)th perfect sets and show that
V (n, d) 6= ∅ for d ≥ 2 and n ≥ d + 2. Then we give some algorithms
to construct (n, d)th f -ideals and show an upper bound of the (n, d)th
perfect number.
Key Words and phrases: perfect set containing k; perfect set
without k; f -ideal; unmixed f -ideal; perfect number
2010 Mathematics Subject Classification: 13P10, 13F20,
13C14, 05A18.
1
Introduction
Throughout the paper, for a set A, we use Ad to denote the set of the subsets of A with
cardinality d. For a monomial ideal I of S, let sm(I) be the set of square-free monomials
in I. As we know, there is a natural bijection between sm(S) and 2[n] , denoted by
σ : xi1 xi2 · · · xik 7→ {i1 , i2 , . . . , ik },
where [n] = {1, 2, . . . , n} for a positive integer n. A square-free monomial u is called
covered by a square-free monomial v, if u | v holds. For other concepts and notations, see
references [2, 5, 7, 8, 10, 11].
∗
This research is supported by the National Natural Science Foundation of China (Grant No.
11271250).
†
Corresponding author. [email protected]
‡
[email protected]
Constructing free resolutions of a monomial ideal is one of the core problems in commutative algebra. A main approach to the problem is by taking advantage of the properties of
a simplicial complex, so it is important to have a research on the properties of the complex
corresponding to the related ideals, see for example, references [4, 6, 9, 12]. There is an
important class of ideals called f -ideals, whose facet complex δF (I) and Stanley-Reisner
complex δN (I) have the same f -vector, where δF (I) is generated by the set σ(G(I)), and
δN (I) = {σ(g) | g ∈ sm(S) \ sm(I)}. Note that the f -vector of a complex δN (I), which
is not easy to calculate in general, is essential in the computation of the Hilbert series of
S/I. Since the correspondence of the complex δF (I) and the ideal I is direct and clear,
it is more easier to calculate the f -vector of δF (I). So, it is convenient to calculate the
Hilbert series and study other corresponding properties of S/I while I is an f -ideal.
The formal definition of an f -ideal first appeared in [1], and it was then studied in [3].
In [7], the authors characterized the f -ideals of degree d, as well as the f -ideals in general
case. They introduced a bijection between square-free monomial ideals of degree 2 and
simple graphs, and showed that V (n, 2) 6= ∅ for any n ≥ 4, where V (n, d) is the set of
f -ideals of degree d in S = K[x1 , . . . , xn ]. The structure of V (n, 2) was determined, and
the characterization of the unmixed f -ideals is also studied in [7].
In this paper, we give another characterization of unmixed f -ideals in part two. In
part three, we generalize the aforementioned result of [7] by showing that V (n, d) 6= ∅ for
general d ≥ 3 and n ≥ d + 2. In part four, we introduce some algorithms to construct
(n, d)th f -ideals, and we show an upper bound of the (n, d)th perfect number in part five.
In part six, we show some examples of nonhomogeneous f -ideals, which is still open in
[7].
The following propositions are needed in this paper.
Proposition 1.1. ([7] Theorem 2.4) Let S = K[x1 , . . . , xn ], and let I be a square-free
monomial ideal of S of degree d with the minimal generating set G(I). Then I is an
f -ideal if and only if G(I) is (n, d)th perfect and |G(I)| = 12 Cnd holds true.
Proposition 1.2. ([7] Proposition 3.3) V (n, 2) 6= ∅ if and only if n = 4k or n = 4k + 1
for some positive integer k.
Proposition 1.3. ([7] Proposition 5.3) Let S = K[x1 , . . . , xn ]. If I is an f -ideal of S of
degree d, then I is unmixed if and only if sm(S)d \ G(I) is lower perfect.
In [7], a method for finding an (n, 2)th perfect set with the smallest cardinality is
provided, namely, first, decompose the set [n] into a disjoint union of two subsets B and
C uniformly, i.e., ||B| − |C|| ≤ 1 holds; then set A = {xi xj | i, j ∈ B, or i, j ∈ C}. Finally,
A is an (n, 2)th perfect set whose cardinality is equal to the (n, 2)th perfect number N(n,2) ,
where
k 2 − k, if n = 2k;
N(n,2) =
(1)
k 2 ,
if n = 2k + 1.
Note that any set D with A ⊆ D ⊆ sm(S)2 is also an (n, 2)th perfect set.
2
(n, d)th unmixed f -ideals
For a positive integer d greater than 2, an (n, d)th f -ideal may be not unmixed, see Example 5.1 of [7] for a counterexmple. So, it is interesting to characterize the unmixed
f -ideals. In this section, we show a characterization of unmixed f -ideals by the corresponding simplicial complex, by taking advantage of the bijection σ between square-free
monomial ideals and simplicial complexes.
Recall that a simplicial complex is a d-flag complex if all of its minimal non-faces
contain d elements. Recall that ∆∨ denotes the Alexander dual of a simplicial complex
∆, see [8] for details.
Proposition 2.1. Let S = K[x1 , . . . , xn ], and let I be a square-free monomial ideal of S
of degree d. I is an (n, d)th unmixed f -ideal if and only if the followings hold:
(1) |G(I)| = Cnd /2;
(2) dim δF (I)∨ = n − d − 1;
(3) hσ(u) | u ∈ sm(S)d \ G(I)i is a d-flag complex.
P roof. We claim that the following two results hold true: First, the condition (2) holds
if and only if G(I) is lower perfect. Second, the condition (3) holds if and only if G(I) is
upper perfect and sm(S)d \ G(I) is lower perfect. If the above two results hold true, then
it is easy to see that the conclusion holds by Proposition 1.1 and Proposition 1.3.
For the first claim, if G(I) is lower perfect, then for each minimal non-face F of δF (I),
|F | ≥ d holds. By the definition of the Alexander dual, G is a face of δF (I)∨ if and
only if [n] \ G is a non-face of δF (I). So, for each facet L of δF (I)∨ , |L| ≤ n − d. Since
|G(I)| =
6 Cnd , there exists some non-face of δF (I) with cardinality d, so there exists some
facet of δF (I)∨ with cardinality n − d. Thus dim(δF (I)∨ ) = n − d − 1.
Conversely, assume dim(δF (I)∨ ) = n − d − 1. By a similar argument, one can see that
the smallest cardinality of non-faces of δF (I) is d, hence G(I) is lower perfect.
For the second claim, if sm(S)d \ G(I) is lower perfect, then for the complex ∆ =
hσ(u) | u ∈ sm(S)d \ G(I)i, the cardinality of a non-face is not less than d. Since G(I) is
upper perfect, for each non-face F of ∆, there exists v ∈ G(I) such that σ(v) ⊆ F . Note
that σ(v) is a non-face of ∆, so all the minimal non-faces of ∆ have cardinality d. Hence
∆ is a d-flag complex.
Conversely, assume that ∆ = hσ(u) | u ∈ sm(S)d \ G(I)i is a d-flag complex. In a
similar way, one can see that G(I) is upper perfect and sm(S)d \ G(I) is lower perfect.
3
Existence of (n, d)th f -ideals
′
For a subset M of sm(S)d , denote M = {σ −1 (A) | A = [n] \ σ(u) for some u ∈ M}. The
following lemma is essential in the proof of our main result in this section.
′
Lemma 3.1. M is a perfect subset of sm(S)d if and only if M is a perfect subset of
sm(S)n−d .
P roof. For the necessary part, if M is a subset of sm(S)d , then it follows from definition
′
′
that M is a subset of sm(S)n−d . In order to check that M is upper perfect, we will show
′
for each monomial u ∈ sm(S)n−d+1 that u ∈ ⊔(M ) holds. This is equivalent to showing
′
that there exists some v ∈ M , such that σ(v) ⊆ σ(u) holds. In fact, since M is lower
′
perfect, for the monomial u = σ −1 ([n] \ σ(u)) ∈ sm(S)d−1 , there exists some w ∈ M such
′
′
that u | w holds. Hence σ(u ) ⊆ σ(w) holds. Now let v = σ −1 ([n] \ σ(w)) and then it is
′
′
easy to see that σ(v) = [n] \ σ(w) ⊆ [n] \ σ(u ) = σ(u) hold. This shows that M is upper
′
perfect. In a similar way, one can prove that M is lower perfect.
The sufficient part is similar to prove, and we omit the details.
By the proof of the above lemma, one can see that M is an upper (lower, respectively)
′
perfect subset of sm(S)d if and only if M is a lower (upper, respectively) perfect subset
of sm(S)n−d .
Corollary 3.2. If I is a square-free monomial ideal of S of degree d, then I is an (n, d)th
′
f -ideal if and only if |G(I)| = Cnd /2 and G(I) is a perfect subset of sm(S)n−d .
Denote sm(S{ǩ})d = {u ∈ sm(S)d | xk ∤ u}, and sm(S{k})d = {u ∈ sm(S)d | xk |u}.
For a subset X = {i1 , . . . , ij } of [n], denote
sm(S{X̌})d = {u ∈ sm(S)d | xk ∤ u f or every k ∈ X},
and let sm(S{X})d = {u ∈ sm(S)d | xk | u for every k ∈ X}.
Definition 3.3. For a subset M of sm(S{ǩ})d , if sm(S{ǩ})d+1 ⊆ ⊔(M) holds, then M
is called upper perfect without k. Dually, a subset M of sm(S{ǩ})d is called lower perfect
without k, if sm(S{ǩ})d−1 ⊆ ⊓(M) holds. A subset M of sm(S{k})d is called upper
perfect containing k, if sm(S{k})d+1 ⊆ ⊔(M) holds; a subset M of sm(S{k})d is called
lower perfect containing k, if sm(S{k})d−1 ⊆ ⊓(M) holds. If M is not only upper but
also lower perfect without k, then M is called perfect without k. Similarly, if M is both
upper and lower perfect containing k, then M is called perfect containing k.
For a subset X of [n], we can define the upper perfect (lower perfect, perfect, respectively) set without X (containing X) similarly. For a subset A of sm(S)d , let
A{X̌} = A ∩ sm(S{X̌})d , and let A{X} = A ∩ sm(S{X})d.
Proposition 3.4. Let A be a subset of sm(S)d , and let X = {i1 , . . . , ij } be a subset of
[n]. Then the following statements hold:
(1) A{X̌} = A{iˇ1 }{iˇ2 } . . . {iˇj }, and A{X} = A{i1 }{i2 } . . . {ij };
(2) If A is upper perfect, then A{X̌} is upper perfect without X;
(3) If A is lower perfect, then A{X} is lower perfect containing X;
′
(4) If A is upper (lower, respectively) perfect without X, then A is lower (upper,
respectively) perfect containing X. Furthermore, the converse also holds true.
P roof. (1) and (2) are easy to see by the corresponding definitions.
In order to prove (3), it is sufficient to show that A{k} is a lower perfect set containing
k for each k ∈ [n]. In fact, since A is lower perfect, for each monomial u ∈ sm(S{k})d−1 ,
there exists a monomial v in A such that u | v. Note that xk | u holds, so xk | v also holds,
which implies that v ∈ sm(S{k})d holds. Hence A{k} is a lower perfect set containing k.
′
For (4), we only show that A is lower perfect containing k when A is upper perfect
without k, and the remaining implications are similar to prove. In fact, for each monomial
′
′
u ∈ sm(S{k})n−d−1 ⊆ sm(S)n−d−1 , u ∈ sm(S)d+1 , note that xk | u implies xk ∤ u holds
′
true, hence u ∈ sm(S{ǩ})d+1 also hold. Since A is upper perfect without k, there exists
′
′
′
′
a monomial v ∈ A such that v | u holds, hence u | v holds, where v ∈ A . This completes
the proof.
Remark 3.5. For a perfect subset A of sm(S)d , A{X̌} needs not to be a lower perfect
set without X, and A{X} needs not to be an upper perfect set containing X, see the
following for counter-examples:
Example 3.6. Let S = K[x1 , . . . , x6 ], let
A = {x1 x2 x3 , x1 x2 x4 , x1 x2 x5 , x3 x4 x5 , x1 x2 x6 , x1 x3 x6 , x2 x3 x6 , x4 x5 x6 },
and let B = A \ {x1 x2 x6 }. It is easy to see
A{6̌} = B{6̌} = {x1 x2 x3 , x1 x2 x4 , x1 x2 x5 , x3 x4 x5 },
A{6} = {x1 x2 x6 , x1 x3 x6 , x2 x3 x6 , x4 x5 x6 }, and B{6} = {x1 x3 x6 , x2 x3 x6 , x4 x5 x6 }. Also, it
is direct to check that both A and B are perfect sets, and that both A{6̌} and B{6̌} are
perfect sets without 6. Note that A{6} is a perfect set containing 6, but B{6} is not
upper perfect.
′
′
By Proposition 3.4, we have the following example by mapping A, B to A , B respectively.
Example 3.7. Let S = K[x1 , . . . , x6 ], and let
′
A = {x1 x2 x3 , x1 x4 x5 , x2 x4 x5 , x3 x4 x5 , x1 x2 x6 , x3 x4 x6 , x3 x5 x6 , x4 x5 x6 },
′
′
and B = A \ {x3 x4 x5 }. It is easy to see that
′
′
A {6̌} = {x1 x2 x3 , x1 x4 x5 , x2 x4 x5 , x3 x4 x5 }, B {6̌} = {x1 x2 x3 , x1 x4 x5 , x2 x4 x5 },
′
′
and A {6} = B {6} = {x1 x2 x6 , x3 x4 x6 , x3 x5 x6 , x4 x5 x6 }. It is direct to check that both
′
′
′
′
A and B are perfect sets, and that both A {6} and A {6} are perfect sets containing 6.
′
′
Note that A {6̌} is a perfect set without 6, but B {6̌} is not lower perfect.
In order to obtain the main result of this section, we need a further fact and we omit
the verification.
Lemma 3.8. Let S = K[x1 , . . . , xn ], and let A be a subset of sm(S)d . If A{ǩ} is a perfect
subset of sm(S{ǩ})d without k, and A{k} is a perfect subset of sm(S{k})d containing k
for some k ∈ [n], then A is a perfect subset of sm(S)d .
Theorem 3.9. For any integer d ≥ 2 and any integer n ≥ d + 2, there exists an (n, d)th
perfect set with cardinality less than or equal to Cnd /2.
P roof. We prove the result by induction on d.
If d = 2, the conclusion holds true for any integer n ≥ 4 by Proposition 1.2. In the
following, assume d > 2.
Assume that the conclusion holds true for any integer less than d. For d, we claim that
the conclusion holds true for any integer n ≥ d + 2. We will show the result by induction
on n.
If n = d + 2, then Cnd = Cn2 . Note that for any integer n ≥ 4, there exists an (n, 2)th
′
perfect set M, such that |M| ≤ Cn2 /2. By Lemma 3.1, M is an (n, d)th perfect set. Note
′
that |M | = |M| ≤ Cn2 /2 = Cnd /2 holds.
Now assume that the conclusion holds true for any integer less than n. Then by Lemma
3.8, it will suffice to show that there is a perfect subset A of sm(S{ň})d without n and
d
a perfect subset B of sm(S{n})d containing n, such that |A| ≤ |sm(S{ň})d |/2 = Cn−1
/2
d−1
and |B| ≤ |sm(S{n})d |/2 = Cn−1 /2 hold.
Let L = K[x1 , . . . , xn−1 ]. Then clearly, sm(S{ň})d = sm(L)d holds. By induction on
d
n, there exists an (n − 1, d)th perfect subset A of sm(L)d , such that |A| ≤ Cn−1
/2. It is
easy to see that A is a perfect subset of sm(S{ň})d without n. By induction on d, there
d−1
exists an (n − 1, d − 1)th perfect subset B1 of sm(L)d−1 , such that |B1 | ≤ Cn−1
/2 holds.
−1
Let B = {σ (D) | D = σ(u) ∪ {n} for some u ∈ B1 }. It is easy to see that B is a perfect
d−1
/2.
subset of sm(S{n})d containing n, and |B| = |B1 | ≤ Cn−1
Finally, by Lemma 3.8, A ∪ B is a perfect subset of sm(S)d , and |A ∪ B| = |A| + |B| ≤
d−1
d
Cn−1 /2 + Cn−1
/2 = Cnd /2. This completes the proof.
By Proposition 1.1 and Theorem 3.9, the following corollary is clear.
Corollary 3.10. For any integer d ≥ 2 and any integer n ≥ d + 2, V (n, d) 6= ∅ if and
only if 2 | Cnd .
4
Algorithms for constructing examples of (n, d)th f ideals
In this section, we will show some algorithms to construct (n, d)th f -ideals. We discuss
the following cases:
Case 1: d = 2. An (n, 2)th f -ideal is easy to construct by [7]. For readers convenience,
we repeat it as the following: Decompose the set [n] into a disjoint union of two subsets
B and C uniformly, namely, ||B| − |C|| ≤ 1. Then set A = {xi xj | i, j ∈ B, or i, j ∈ C}
to obtain an (n, 2)th perfect set. Note that |A| = N(n,2) ≤ Cn2 /2, choose a subset D of
sm(S)2 \ A randomly, such that |D| = Cn2 /2 − N(n,2) holds. It is easy to see that A ∪ D is
still a perfect set, and |A ∪ D| = Cn2 /2. By Proposition 1.1, the ideal generated by A ∪ D
is an (n, 2)th f -ideal. Note that each (n, 2)th f -ideal can be obtained in this way except
C5 by [7].
Case 2: d > 2 and n = d + 2.
Algorithm 4.1. In order to build an f -ideal I ∈ V (d + 2, d), we obey the following steps:
d
d
2
Step 1: Calculate Cd+2
/2. Note that Cd+2
/2 = Cd+2
/2.
2
Step 2: As in the case 1, find a perfect subset B of sm(S)2 such that |B| ≤ Cd+2
/2,
where S = K[x1 , . . . , xd+2 ].
′
Step 3: Let A = B . Then A is a perfect subset of sm(S)d by Lemma 3.1, and
2
d
|A| = |B| ≤ Cd+2
/2 = Cd+2
/2.
d
Step 4: Choose a subset D of sm(S)d \ A randomly, such that |D| = Cd+2
/2 − |A|
d
holds. It is easy to see that M = A ∪ D is still a perfect set, and |A ∪ D| = Cd+2 /2.
Step 5: Let I be the ideal generated by A ∪ D, By Proposition 1.1 again, I is an
(d + 2, d)th f -ideal.
Note that in this way, we constructed almost all (d + 2, d)th f -ideals.
Example 4.2. Show an f -ideal I ∈ V (8, 6).
Note that 8 = 6 + 2, we obey the Algorithm 4.1.
Note that C86 /2 = 14. Find a perfect subset B of sm(S)2 such that |B| ≤ C82 /2 = 14,
where S = K[x1 , . . . , x8 ]. It is easy to see that
B = {x1 x2 , x1 x3 , x1 x4 , x2 x3 , x2 x4 , x3 x4 , x5 x6 , x5 x7 , x5 x8 , x6 x7 , x6 x8 , x7 x8 }
is a perfect subset of sm(S)2 , with |B| = 12. Let
′
A = B = {x3 x4 x5 x6 x7 x8 , x2 x4 x5 x6 x7 x8 , x2 x3 x5 x6 x7 x8 , x1 x4 x5 x6 x7 x8 ,
x1 x3 x5 x6 x7 x8 , x1 x2 x5 x6 x7 x8 , x1 x2 x3 x4 x7 x8 , x1 x2 x3 x4 x6 x8 ,
x1 x2 x3 x4 x6 x7 , x1 x2 x3 x4 x5 x8 , x1 x2 x3 x4 x5 x7 , x1 x2 x3 x4 x5 x6 }.
A is a perfect subset of sm(S)6 . Choose D = {x1 x2 x3 x5 x6 x7 , x1 x2 x4 x5 x6 x8 }, then the
ideal I generated by A ∪ D is an (8, 6)th f -ideal.
Case 3: d > 2 and n > d + 2. Let S [k] = K[x1 , . . . , xk ], and let S = S [n] =
K[x1 , . . . , xn ].
Algorithm 4.3. For an integer n > d + 2, we construct an (n, d)th f -ideal by using the
following steps:
Step 1: Let t = n, l = d and E = ∅. Set B = {B(t, l, E)}.
Step 2: Assign C = B, and denote i = |C|.
Step 3: Choose each B(t, l, E) ∈ C one by one, deal with each one obeying the following
rules:
If l = 2 or t = l + 2, don’t change anything.
If l 6= 2 and t > l + 2, then cancel B(t, l, E) from B, and add B(t − 1, l, E) and
B(t − 1, l − 1, E ∪ {t}) into B.
After i times, i.e., when B(t, l, E) goes through all the element of C, make a judgement:
If l = 2 or t = l + 2 for each B(t, l, E) ∈ B, then go to step 4, else return to step 2.
Step 4: Choose B(t, l, E) ∈ B one by one, deal with each one obeying the following
rules:
If l = 2, assign B(t, l, E) a perfect subset of sm(S [t] )l as case 1.
If l 6= 2 and t = l + 2, assign B(t, l, E) a perfect subset of sm(S [t] )l as case 2.
Step 5: For each B(t, l, E) ∈ B, denote B ∗ (t, l, E) = {uxE | u ∈ B(t, l, E)}, where
Q
xE = j∈E xj . Denote B∗ = ∪B(t,l,E)∈B B ∗ (t, l, E). It is direct to check that B∗ is a perfect
subset of sm(S)d , and |B∗ | ≤ Cnd /2. Choose a subset D of sm(S)d \ B∗ randomly, such
that |D| = Cnd /2 − |B∗ | holds.
Step 6: Let I be the ideal generated by B∗ ∪ D. By Proposition 1.1 again, I is an
(n, d)th f -ideal.
Example 4.4. Show a (6, 3)th f -ideal.
Let S = K[x1 , . . . , x6 ]. By the above algorithm, we will choose a perfect subset
B(5, 3, ∅) of sm(S 5 )3 and a perfect subset B(5, 2, {6}) of sm(S 5 )2 . Set B(5, 3, ∅) =
{x3 x4 x5 , x2 x4 x5 , x1 x4 x5 , x1 x2 x3 }, and set B(5, 2, {6}) = {x1 x2 , x1 x3 , x2 x3 , x4 x5 }. Correspondingly, B ∗ (5, 3, ∅) = B(5, 3, ∅) and
B ∗ (5, 2, {6}) = {x1 x2 x6 , x1 x3 x6 , x2 x3 x6 , x4 x5 x6 }.
Hence
B∗ = {x3 x4 x5 , x2 x4 x5 , x1 x4 x5 , x1 x2 x3 , x1 x2 x6 , x1 x3 x6 , x2 x3 x6 , x4 x5 x6 }
is a perfect subset of sm(S)3 . Note that C63 /2 = 10, and |B∗ | = 8. Set D = {x1 x2 x4 , x1 x2 x5 }.
The ideal I generated by B∗ ∪ D is a (6, 3)th f -ideal.
Note that the (6, 3)th f -ideal given in the above example is not unmixed. In fact,
consider the simplicial complex σ(sm(S)3 \ G(I)), and note that {1, 2} is a non-face of
σ(sm(S)3 \ G(I)), which implies that σ(sm(S)3 \ G(I)) is not a 3-flag complex. So, I is
not unmixed by Proposition 2.1.
5
An upper bound of the perfect number N(n,d)
For a positive integer k and a pair of positive integers i ≤ j, denote by Qk[i,j] the set
of square-free monomials of degree k in the polynomial ring K[xi , xi+1 , . . . , xj ]. Note
that Qk[i,j] = ∅ holds for i > j. For a pair of monomial subsets A and B, denote by
A • B = {uv | u ∈ A, v ∈ B}. If B = ∅, then assume A • B = A. The following theorem
gives an upper bound of the (n, d)th perfect number for n > d + 2.
Theorem 5.1. Given a integer d > 2, and a integer n ≥ d + 2. The following statements
about the perfect number N(n,d) hold:
(1) If n = d + 2, then
k 2 − k, if n = 2k;
N(n,d) = N(n,2) =
(2)
k 2 ,
if n = 2k + 1.
(2) If n > d + 2, then
N(n,d) ≤
n−d+2
X
i=5
d−3
N(i,2) Cn−i−1
+
d
X
d−j
N(j+2,2) Cn−j−3
,
(3)
j=3
where C00 = 1.
P roof. By Lemma 3.1 and the equation 1 in the first section, (1) is clear.
In order to prove (2), it will suffice to show that there exists a perfect set with cardiP
Pn−d+2
d−j
d−3
N(i,2) Cn−i−1
+ dj=3 N(j+2,2) Cn−j−3
.
nality t = i=5
th
Let P(i,2) be an (i, 2) perfect set with cardinality N(i,2) for 5 ≤ i ≤ n − d + 2, and let
P(j+2,j) be a (j + 2, j)th perfect set with cardinality N(j+2,j) for 3 ≤ j ≤ d. We claim that
the set
d−j
n−d+2
d
M = (∪i=5
P(i,2) • xi+1 • Qd−3
[i+2,n] ) ∪ (∪j=3 P(j+2,j) • Q[j+4,n] )
is an (n, d)th perfect set, with cardinality t. It is easy to check that the cardinality of M
is t. It is only necessary to prove that M is perfect.
For each w ∈ sm(S)d+1 , denote by nk (w) the cardinality of the set {xi | i ≤ k and
xi | w}. If n5 (w) ≥ 4, then choose the smallest k such that nk+3 (w) = nk+2 (w) =
k + 1. Clearly, 3 ≤ k ≤ d. It is direct to check that w is divided by some monomial
in P(k+2,k) • Qd−k
[k+4,n] . If n5 (w) ≤ 3, then choose the smallest k such that nk (w) = 3 and
nk+1 (w) = 4. Clearly, 5 ≤ k ≤ n − d + 2. It is not hard to check that w is divided by
some monomial in P(k,2) • xk+1 • Qd−3
[k+2,n] . Hence M is upper perfect.
For each w ∈ sm(S)d−1 , if n5 (w) ≥ 2, then choose the smallest k such that nk+3 (w) =
nk+2 (w) = k − 1. Clearly, 3 ≤ k ≤ d. It is direct to check that w is covered by
some monomial in P(k+2,k) • Qd−k
[k+4,n] . If n5 (w) ≤ 1, then choose the smallest k such that
nk (w) = 1 and nk+1 (w) = 2. Clearly, 5 ≤ k ≤ n − d + 2 holds. It is not hard to check
that w is covered by some monomial in P(k,2) • xk+1 • Qd−3
[k+2,n] . Hence M is lower perfect.
The following figure may help to interpret the above theorem intuitively.
.(d, n)
t ✻l = 2 ...l = 3
. 3
......❄
.
t = .l...+
..
.
✠
.. ...
..
.
.
.
.
..
.
..
..
..
...✠ ✠ ....
... t = l + 2
.
.
.
.
..
.
. ❄ ❄.
...
... ......✠ ❄ ...
.. ...
.. ......
....
.. ..
...... ✠ ✠ .... .....
.
.
.. .. ❄ ❄ .....
..✠ ✠ ... ..
..
.
..❄ ❄ ......
..
...
.
✠ ... ✠.....
.
.. .❄
.. .. ..
..... ❄
(3,. 6)
(2, 4)
✲
O
l
Figure 1. Upper Bound
In this figure, there is a boundary consisting of the line l = 2 and the line t = l + 2.
From the point (d, n) to a point of the boundary, every directed chain C denotes a set of
monomials M(C) by the following rules:
(1) Every arrow of C is from (l, t) to either (l, t − 1) or (l − 1, t − 1).
(2) If the arrow is from (l, t) to (l, t − 1), then each monomial in M(C) is not divided
by xt . Correspondingly, if it is from (l, t) to (l − 1, t − 1), then each monomial in M(C)
is divided by xt .
(3) Each point (l, t) of the boundary is a (t, l)th perfect set.
Actually, the figure shows us a class of (n, d)th perfect sets. For each point (l, t) of the
boundary, if we choose the corresponding perfect set to be a (t, l)th perfect set with cardiPn−d+2
d−3
nality N(t,l) , then the cardinality of the (n, d)th perfect set is exactly i=5
N(i,2) Cn−i−1
+
Pd
d−j
j=3 N(j+2,2) Cn−j−3 .
Example 5.2. Calculation of the (6, 3)th perfect number.
Let A be a (6, 3)th perfect set. By Proposition 3.4(2), A{6̌} is an upper perfect set
without 6. Hence |A{6̌}| ≥ N(5,3) = 4. By Proposition 3.4(3), A{6} is a lower perfect set
containing 6. Note that for the monomials of {x1 , x2 , x3 , x4 , x5 }, each monomial in A{6}
covers at most two of them. So, |A{6}| ≥ 3. Hence |A| ≥ |A{6̌}| + |A{6}| ≥ 7. Actually,
as showed in Example 3.6, there exists a (6, 3)th perfect set
B = {x1 x2 x3 , x1 x2 x4 , x1 x2 x5 , x3 x4 x5 , x1 x3 x6 , x2 x3 x6 , x4 x5 x6 }
with cardinality 7. Thus N(6,3) = 7. Note that the upper bound given by Proposition
5.1(2) is 8, and is not bad for the perfect number in the case.
6
Nonhomogeneous f -ideal
In [7], a characterization of f -ideals in general case is shown, but it is still not easy to
show an example of nonhomogeneous f -ideal. In fact, the interference from monomials
of different degree makes the computation complicated. Anyway, we worked out the
following examples:
Example 6.1. Let S = K[x1 , x2 , x3 , x4 , x5 ], and let
I = hx1 x2 , x3 x4 , x1 x3 x5 , x2 x4 x5 i.
It is direct to check that
δF (I) = h{1, 2}, {3, 4}, {1, 3, 5}, {2, 4, 5}i
and
δN (I) = h{1, 3}, {2, 4}, {1, 4, 5}, {2, 3, 5}i.
It is easy to see they have the same f -vector, and hence I is an f -ideal, which is clearly
nonhomogeneous.
In fact, there are a lot of nonhomogeneous f -ideals. We will show another example to
end this section.
Example 6.2. Let S = K[x1 , x2 , x3 , x4 , x5 , x6 ], and let
I = hx1 x2 , x2 x3 , x1 x3 , x4 x5 , x1 x4 x6 , x1 x5 x6 , x2 x4 x6 i.
Note that
δN (I) = h{1, 4}, {1, 5}, {1, 6}, {2, 4}, {2, 5, 6}, {3, 4, 6}, {3, 5, 6}i.
It is direct to check that I is also a nonhomogeneous f -ideal.
References
[1] G. Q. Abbasi, S. Ahmad, I. Anwar, W. A. Baig, f -Ideals of degree 2, Algebra Colloquium, 19(Spec1)(2012), 921 − 926.
[2] M.F. Atiyah and I.G. MacDonald, Introduction to Commutative Algebra, AddisonWesley, Reading, MA, 1969.
[3] I. Anwar, H. Mahmood, M. A. Binyamin and M. K. Zafar, On the Characterization
of f -Ideals, Comm. Algebra (accepted), see also arXiv: 1309.3765 (Sep. 15, 2013 ).
[4] E. Connon and S. Faridi. Chorded complexes and a necessary condition for a monomial ideal to have a linear resolution. Journal of Combinatorial Theory, Series A,
120(2013), 1714 − 1731.
[5] D. Eisenbud, Commutative Algebra with a View Toward Algebraic Geometry. Springer
Science + Business Media, Inc, 2004.
[6] S. Faridi, The facet ideal of a simplicial complex, Manuscripta Mathematica,
109(2002), 159 − 174.
[7] J. Guo, T. S. Wu and Q. Liu, Perfect sets and f-Ideals, preprint.
[8] J. Herzog and T. Hibi, Monomial Ideals. Springer-Verlag London Limited, 2011.
[9] J. Herzog, T. Hibi, X. Zheng, Diracs theorem on chordal graphs and Alexander
duality, European J. Combin. 25 (2004)949 − 960.
[10] R. H. Villarreal, Monomial Algebra. Marcel Dekker, Inc, New York, 2001.
[11] O. Zariski and P. Samuel, Commutative Algebra. Vol.1 Reprints of the 1958 − 60
edition. Springer-Verlag New York, 1979.
[12] X. Zheng, Resolutions of facet ideals, Commun. Algebra 32 (6) (2004) 2301 − 2324.
| 0math.AC
|
arXiv:1401.6579v1 [math.AC] 25 Jan 2014
ISOMORPHISMS BETWEEN JACOBSON GRAPHS
A. AZIMI, A. ERFANIAN AND M. FARROKHI D. G.
Abstract. Let R be a commutative ring with a non-zero identity and JR be
its Jacobson graph. We show that if R and R′ are finite commutative rings,
then JR ∼
= JR′ if and only if |J(R)| = |J(R′ )| and R/J(R) ∼
= R′ /J(R′ ). Also,
for a Jacobson graph JR , we obtain the structure of group Aut(JR ) of all
automorphisms of JR and prove that under some conditions two semi-simple
rings R and R′ are isomorphic if and only if Aut(JR ) ∼
= Aut(JR′ ).
1. introduction
Let Γ and Γ′ be two simple graphs. An isomorphism between Γ and Γ′ is a
bijection α : V (Γ) −→ V (Γ′ ) such that
{u, v} ∈ E(Γ) ⇐⇒ {α(u), α(v)} ∈ E(Γ′ ).
Two graphs Γ and Γ′ are isomorphic whenever there is an isomorphism between
them, and the isomorphism is denoted by Γ ∼
= Γ′ . If Γ′ = Γ, then an isomorphism
′
between Γ and Γ is said to be an automorphism of Γ. The set of all automorphisms
of Γ is denoted by Aut(Γ).
Let R and R′ be commutative rings with non-zero identities. The zero-divisor
graph Γ(R) of R is a graph whose vertices are non-zero zero-divisors and two distinct
vertices x and y are adjacent if and only if xy = 0. Anderson, Frazier, Lauve and
Livingston [2] proved that if R and R′ are finite reduced rings which are not fields,
then Γ(R) ∼
= Γ(R′ ) if and only if R ∼
= R′ . Later, Akbari and Mohammadian [1]
generalized this result and proved that if R is a finite reduced ring which is not
isomorphic to Z2 ⊕Z2 and Z6 and R′ is a ring such that Γ(R) ∼
= Γ(R′ ), then R ∼
= R′ .
Anderson and Livingston [3] have shown that Aut(Γ(Zn )) is a direct product of
some symmetric groups when n ≥ 4 is a non-prime integer. In the case of noncommutative rings, Han [6] shows that Aut(Γ(R)) ∼
= Sp+1 is the symmetric group
of degree p + 1 when R = Mat2 (Zp ) (p prime), and Park and Han [11] generalized
Han’s result and proved that Aut(Γ(R)) ∼
= S|F |+1 when R = Mat2 (F ) with F a
finite field.
Let R be a commutative ring with non-zero identity. The Jacobson graph of R,
denoted by JR , is a graph whose vertex set is R\J(R) such that two distinct vertices
x, y ∈ V (JR ) are adjacent whenever 1 − xy 6∈ U (R), in which U (R) is the group of
units of R. It is easy to see that the subgraph of JR induced by all vertices x such
that 1 − x2 is a unit, is the co-normal product of some graphs, each of which is a
union of isomorphic complete bipartite graphs. Recall that the Jacobson radical
J(R) of R is the intersection of all maximal ideals of R and it has the property that
2000 Mathematics Subject Classification. Primary 05C25, 05C60, Secondary 16P10, 13H99,
16N20.
Key words and phrases. Jacobson graph, isomorphism, automorphism, wreath product.
1
2
A. AZIMI, A. ERFANIAN AND M. FARROKHI D. G.
1 − xr ∈ U (R) whenever x ∈ J(R) and r ∈ R. A ring with trivial Jacobson radical
is called a semi-simple ring.
The Jacobson graphs first introduced in [4] by the authors and several graph theoretical properties as well as numerical invariants of such graphs is discovered. In
this paper, we shall concentrate on isomorphisms between Jacobson graphs. Also,
we shall determine the structure of automorphism group of Jacobson graphs and
give some conditions under which two rings are isomorphic whenever the automorphism group of their Jacobson graphs are isomorphic. In this paper, all rings will
be commutative rings with non-zero identity.
Let Γ be a graph. In what follows, NΓ (x) and NΓ [x] or simply N (x) and N [x]
stand for the open and closed neighborhood of x, respectively, for every vertex x
of Γ. Also, degΓ (x) or simply deg(x) denotes the degree of x and δ(Γ) denotes the
minimum vertex degree of Γ.
2. Isomorphic Jacobson graphs
We begin with recalling some results from ring theory. Let R be a finite ring.
Then, by [9, Theorem VI.2], R = R1 ⊕ · · · ⊕ Rn , where (Ri , mi ) are local rings with
associated fields Fi . Moreover,
J(R) = J(R1 ) ⊕ · · · ⊕ J(Rn )
and
U (R) = U (R1 ) ⊕ · · · ⊕ U (Rn ).
Clearly,
|U (R)| =
n
Y
i=1
|U (Ri )| = |J(R)|
n
Y
(|Fi | − 1) .
i=1
In what follows, 1 stands for the element (1, . . . , 1) with all entries equal to 1 and ei
is the element with 1 on its ith entry and 0 elsewhere, for i = 1, . . . , n. Also, e(R)
and o(R) denote the number of Fi such that |Fi | is even and odd, respectively.
Note
Sn
that, by [4, Theorem 5.8], for a unit element x ∈ R, degJR (x) = | i=1 Mi | − εx ,
where εx = 0 if 1 − x2 ∈ U (R), εx = 1 if 1 − x2 ∈
/ U (R) and M1 , . . . , Mn are the
maximal ideals of R. In particular,
degJR (1) =
n
[
i=1
Mi − 1 = |R| − |J(R)|
n
Y
(|Fi | − 1) − 1.
i=1
Now, we shall study isomorphisms between Jacobson graphs. It is easy to see that
Z4 and Z2 [x]/(x2 ) have the same Jacobson graphs while they are non-isomorphic
rings. The following theorem gives necessary and sufficient conditions under which
two rings have isomorphic same Jacobson graphs.
Theorem 2.1. Let R and R′ be finite rings. Then JR ∼
= JR′ if and only if |J(R)| =
|J(R′ )| and R/J(R) ∼
= R′ /J(R′ ).
Proof. Let ϕ : JR −→ JR′ be a graph isomorphism. Since R and R′ are finite rings,
we may write R = R1 ⊕ · · · ⊕ Rm and R′ = R1′ ⊕ · · · ⊕ Rn′ , where Ri and Rj′ are
local rings with associated fields Fi and Fj′ , respectively. We further assume that
|F1 | ≤ · · · ≤ |Fm | and |F1′ | ≤ · · · ≤ |Fn′ |. It is easy to see that the unit vertices
have largest degrees among all other vertices and that there is two possible values
for the degree of such elements (see [4, Theorem 5.8]). Hence ϕ maps the units of
JACOBSON GRAPH
3
R onto the units of R′ and consequently |U (R)| = |U (R′ )|. In particular, we have
deg(1R ) = deg(1R′ ). Therefore,
|J(R)|
m
Y
(|Fi | − 1) = |J(R′ )|
|R| − |J(R)|
m
Y
|Fj′ | − 1
j=1
i=1
and
n
Y
(|Fi | − 1) − 1 = |R′ | − |J(R′ )|
n
Y
j=1
i=1
′
|Fj′ | − 1 − 1,
from which it follows that |R| = |R | and consequently |J(R)| = |J(R′ )| as the
graphs have the same number of vertices.
If |Fm | ≤ 3 and |Fn′ | ≤ 3, then
2e(R) 3o(R) =
′
′
R
R′
=
= 2e(R ) 3o(R ) ,
′
J(R)
J(R )
which implies that R/J(R) ∼
= R′ /J(R′ ). Now, suppose that |Fm | > 3. It is easy to
see that
|R|
δ(JR ) = degJR (em ) =
−1
|Fm |
and
δ(JR′ ) = degJR′ (e′n ) =
|R′ |
− 1.
|Fn′ |
∼ F ′ . Let S = R1 ⊕· · ·⊕
Since δ(JR ) = δ(JR′ ) we obtain |Fm | = |Fn′ | and hence Fm =
n
′
′
′
Rm−1 and S = R1 ⊕· · ·⊕Rn−1 . By [4, Lemma 2.1(2) and Theorem 2.2], there exist
elements am ∈ Rm \ J(Rm ) and bn ∈ Rn′ \ J(Rn′ ) such that 1 − a2m ∈ U (Rm ) and
1 − b2n ∈ U (Rn′ ). Without loss of generality, we may assume that ϕ(am em ) = bn e′n .
Since ϕ(NJR (am em )) = NJR′ (bn e′n ), the map
φ : JS
x
−→ JS ′
′
7−→ ϕ(x + a−1
m em ) · (1R′ − en )
is an isomorphism between JS and JS ′ . Using induction we may assume that
S/J(S) ∼
= S ′ /J(S ′ ), from which it follows that R/J(R) ∼
= R′ /J(R′ ). The converse
is clear.
Corollary 2.2. Let R and R′ be finite semi-simple rings. Then JR ∼
= JR′ if and
only if R ∼
= R′ .
Remark. As infinite fields with the same characteristic and cardinality have isomorphic Jacobson graphs, finiteness is a necessary condition in the above theorem.
3. Automorphisms of Jacobson graphs
In what follows, the support of an element x = (x1 , . . . , xn ) in a semi-simple ring
R = F1 ⊕ · · · ⊕ Fn , where Fi are fields, is the set of all indices i such that xi 6= 0
and it is denoted by Supp(x). Our results are based on the following two lemmas.
Lemma 3.1. Let R be a finite semi-simple ring such that R ∼
6= Z2 ⊕ Z2 and R is
not a field. If x, y ∈ V (JR ), then N (x) = N (y) if and only if x = y.
4
A. AZIMI, A. ERFANIAN AND M. FARROKHI D. G.
Proof. Let R = F1 ⊕· · ·⊕Fn be the decomposition of R into finite fields Fi . Suppose
that x = (x1 , . . . , xn ) and y = (y1 , . . . , yn ) are vertices of JR such that N (x) = N (y)
but x 6= y. Then there exists i ∈ {1, . . . , n} such that xi 6= yi and without loss of
generality xi 6= 0. If xi 6= ±1, then x−1
i ei ∈ N (x) so that yi = xi , a contradiction.
Hence xi = ±1. If |Supp(x)| ≥ 2, then since xi ei ∈ N (x) we have that yi = xi , a
contradiction. Thus Supp(x) = {i} and x = ±ei . Similarly y = ±ej . Now a simple
verification shows that N (x) = N (y) only if R = Z2 ⊕ Z2 , which is a contradiction.
The converse is obvious.
Lemma 3.2. Let R be a finite semi-simple ring that is not a field. If x, y ∈ V (JR ),
then N [x] = N [y] if and only if x = y.
Proof. Let R = F1 ⊕ · · · ⊕ Fn be the decomposition of R into finite fields Fi .
Suppose that x = (x1 , . . . , xn ) and y = (y1 , . . . , yn ) are vertices of JR such that
N [x] = N [y] but x 6= y. Clearly, x and y are adjacent. If Supp(x) = {i} is a
singleton, then x = xi ei and yi = xi−1 . If y 6= xi−1 ei and j ∈ Supp(y) \ {i}, then
yj−1 ej ∈ N [y] \ N [x], which is impossible. Thus y = x−1
i ei . If xi 6= ±1 and j 6= i,
then xi ei + ej ∈ N [y] \ N [x], which is impossible. Thus xi = ±1 and hence x = y,
a contradiction. Thus |Supp(x)| > 1 and similarly |Supp(y)| > 1. Suppose that
xi 6= yi and without loss of generality xi 6= 0. Then x−1
i ei ∈ N (x), which implies
that x−1
e
∈
N
(y).
Hence
y
=
x
,
which
is
a
contradiction.
The converse is
i
i
i
i
obvious.
Let R be a finite ring and θ : Aut(JR ) −→ Aut(JR/J(R) ) be the map give by
θ(α) = α such that α(x + J(R)) = α(x) + J(R) for every x ∈ R \ J(R). With
this definition, θ gives rise to an epimorphism from Aut(JR ) onto Aut(JR/J(R) ).
Therefore
Aut(JR ) ∼
R
.
= Aut J J(R)
Ker(θ)
Moreover,
Ker(θ) = S|J(R)| × · · · × S|J(R)|
is the direct product of |R/J(R)|−1 copies of the symmetric group on |J(R)| letters.
So, in the remainder of this section, we just consider the case of finite semi-simple
rings.
Theorem 3.3. Let R = F1 ⊕ · · · ⊕ Fn be the decomposition of semi-simple ring
R into finite fields Fi and |F1 | ≥ · · · ≥ |Fn |. Also, let Ai = {j : |Fj | = |Fi |},
Π ≤ Sn be the set of all permutations preserving Ai ’s, and Σi ≤ S|Fi | be the set
of all permutations fixing 0 and preserving the sets {±1} and {fi , fi−1 } for all
fi ∈ Fi \ {0, ±1}. Then
)
(
n
n
X
X
σi (ai )eπ(i) , π ∈ Π, σi ∈ Σi .
ai ei ) =
Aut(JR ) = α : JR −→ JR : α(
i=1
i=1
Proof. Let Inn(JR ) be the set of automorphisms in the theorem. We show that
Aut(JR ) = Inn(JR ). We proceed by induction on n. Clearly, the result holds
for n = 1. Hence we assume that n > 1 and the result holds for n − 1. Let
α ∈ Aut(JR ), and a, b ∈ F1 \ {0}. Then deg(ae1 ), deg(be1 ) = |R|/|F1 | − 1 or
|R|/|F1 | according to the cases when a, b = ±1 or not, respectively. On the other
hand, for a vertex x ∈ V (JR ), deg(x) = |R|/|F1 | − 1 or |R|/|F1 | only if x = fk ek
for some k ∈ A1 . Hence α(ae1 ) = fi ei and α(be1 ) = fj ej for some i, j ∈ A1 .
JACOBSON GRAPH
5
If i 6= j, then NJR (fi ei ) ∩ NJR (fj ej ) 6= ∅ while NJR (ae1 ) ∩ NJR (be1 ) = ∅, a
contradiction. Thus i = j and we may assume modulo Inn(JR ) that α(f e1 ) = f e1
for all f ∈ F1 \ {0}. In particular, α(NJR (f e1 )) = NJR (f e1 ) for all f ∈ F1 \ {0}.
Let S = F2 ⊕ · · · ⊕ Fn . Since NJR (f e1 ) = f −1 e1 + S for all f ∈ F1 \ {0}, it follows
that α(S) = S. Let α|S : J(S) −→ J(S) be the restriction of α on S \ {0}. By
assumption, α|S ∈ Inn(JS ). Hence we may further assume modulo Inn(JR ) that
α(x) = x for all x ∈ S \ {0}. Let Sf = f e1 + S for all f ∈ F1 \ {0}. Since α fixes
Sf , for each x ∈ S, there exists αf (x) ∈ S such that α(f e1 + x) = f e1 + αf (x).
Let f ∈ F \ {0, ±1}. If x, y ∈ S \ {0}, then x ∼ y if and only if f e1 + x ∼ f e1 + y
if and only if f e1 + αf (x) ∼ f e1 + αf (y) if and only if αf (x) ∼ αf (y). Hence αf
induces an automorphism on JS . On the other hand, if x, y ∈ S \ {0}, then x ∼ y
if and only if x ∼ f e1 + y if and only if x ∼ f e1 + αf (y) if and only if x ∼ αf (y)
or x = αf (y). Now it is easy to see that, if x 6∼ αf (x), then αf (NJS (x)) = NJS (x)
and if x ∼ αf (x), then αf (NJS [x]) = NJS [x]. Hence, by Lemmas 3.1 and 3.2,
αf (x) = x except probably when αf (NJS (x)) = NJS (x) and S ∼
= Z2 ⊕ Z2 , or S is
field. If S ∼
= Z2 ⊕ Z2 , then since F + ei ⊆ NJR [ei ], it follows that α(F + ei ) ⊆
(F + ei ) ∪ (F + e2 + e3 ), for i = 2, 3. But degJR (x) < degJR (y) for all x ∈ F + ei
and y ∈ F + e2 + e3 , which implies that α(F + ei ) = F + ei , for i = 2, 3. Moreover,
α(F + e2 + e3 ) = F + e2 + e3 . Thus α stabilizes NJR [f e1 + e2 ], NJR [f e1 + e3 ] and
NJR [f e1 +e2 +e3 ], from which by Lemma 3.2, it follows that α fixes all the elements
f e1 + e2 , f e1 + e3 and f e1 + e2 + e3 , as required. Now, suppose that S is a field
and αf (x) 6= x. If αf (NJS [x]) = NJS [x], then αf (x) = x−1 . But then αf (x−1 ) = x
and since x ∼ f e1 + x−1 , we should have x ∼ f e1 + αf (x−1 ), a contradiction.
Thus αf (NJS (x)) = NJS (x), which implies that x = ±1 and αf (x) = −x. Since
x ∼ f e1 + x it follows that x ∼ f e1 + αf (x), which is a contradiction. Therefore,
in all cases αf (x) = x that is α(f e1 + x) = f e1 + x for all x ∈ S.
Finally, suppose that f = ±1. Let
2
NJS (x), 1 − x ∈ U (S),
N JS (x) =
/ U (S),
NJS [x], 1 − x2 ∈
for each x ∈ S \ {0}. Then
NJR (x) = NJS (x) ∪
[
f ′ ∈F \{0}
f ′ e1 + N JS (x) .
Since α(x) = x, it follows that α(NJR (x)) = NJR (x). Hence
(1)
α(f ′ e1 + N JS (x)) = f ′ e1 + N JS (x)
for each f ′ ∈ F1 \ {0}. Now, we have
NJR [f e1 + x] = (f e1 + S) ∪
[
f ′ ∈F1 \{f }
f ′ e1 + N JS (x)
so that, by using (1), α(NJR [f e1 + x]) = NJR [f e1 + x]. Thus NJR [α(f e1 + x)] =
NJR [f e1 + x] and by Lemma 3.2, α(f e1 + x) = f e1 + x. Therefore, α = I is the
identity automorphism and the proof is complete.
For each finite field F , we let εF = 0 if |F | is even and εF = 1 if |F | is odd. By [4,
Theorem 2.2], JF is a union of εF +1 isolated vertices together with (|F |−2−εF )/2
disjoint edges. Hence, we have the following result.
6
A. AZIMI, A. ERFANIAN AND M. FARROKHI D. G.
Corollary 3.4. Let R = F1m1 ⊕ · · · ⊕ Fkmk be the decomposition of the semi-simple
ring R into finite fields Fi , where F1 , . . . , Fk are distinct fields. Then
Aut(JR ) = Aut(JF1 ) ≀ Sm1 × · · · × Aut(JFk ) ≀ Smk ,
in which
Aut(JFi ) =
Z2 ≀ S |Fi |−2−εFi × ZεFi +1
2
if |Fi | ≥ 4, and Aut(Z3 ) = Z2 and Aut(Z2 ) = 1 is the trivial group.
Remark. Let R = Z2 ⊕ · · · ⊕ Z2 (n times). Then, by Corollary 3.4, Aut(JR ) ∼
= Sn .
Note that JR is the complement of the zero-divisor graph of R and it is known, say
from [10], that Aut(JR ) ∼
= Sn . Hence JR is an example of graphs with symmetric
automorphism group, in which the number of vertices exceeds that of [8].
Remark. Let R = F ⊕ · · · ⊕ F (n times), where F is a field of order 3 or 4. Then,
by Corollary 3.4, Aut(JR ) ∼
= Z2 ≀ Sn . Hence, by [7], Aut(JR ) ∼
= Aut(Qn ), where Qn
denotes the hypercube of dimension n.
In the remainder of this section, we shall seek for conditions under which two
finite rings are isomorphic if and only if the automorphism group of their Jacobson
graphs are isomorphic.
Lemma 3.5. Let G be a finite group and W = G ≀ Sn . Then W is decomposable
if and only if G has a non-trivial central direct factor H such that (|H|, n) = 1, in
which case H is a direct factor of W .
Proof. Suppose W is decomposable. By [5, Theorem 2.8], G has a non-trivial
Abelian direct factor H such that its elements each of which has unique nth root.
Suppose on the contrary that gcd(|H|, n) 6= 1 and p be a common prime divisor
of |H| and n. Let x be a p-element of largest order pk in H and y be the unique
nth root of x in H. Then p |y|, which implies that |y n | < pk , a contradiction. The
converse follows from [5, Theorem 2.8].
In the remaining of this section, Fm stands for the finite field of order m and
m
b denotes the number (m − 2 − εm )/2 for each prime power m, in which εm =
m − 2[m/2] denotes the parity of m. If p is a prime, then ep (n) is the exponent
of p in n, that is pep (n) |n but pep (n)+1 ∤ n. Bertrand’s postulate states that there
always exists a prime p such that n < p < 2n − 2 for all n > 3. Also, a weaker
formulation of Bertrand’s postulate states that there always exists a prime p such
that n < p < 2n for all n > 1. Hence, if pn denotes the largest prime not exceeding
n, then epn (n!) = 1. Also, it is known that the following equality holds for each
natural number n and prime p,
ep (n!) =
n − Sp (n)
,
p−1
where Sp (n) denotes the sum of digits of n when n is written in base p.
By Lemma 3.5, we have
× Hm ,
Aut(JFm ) ∼
= Zεm +1 × Zεm
c +1
JACOBSON GRAPH
7
b
m
c m!,
where |Hm | = 2m−ε
b and
m = 2,
Sn ,
Aut(JFm ) ≀ Sn ∼
m = 3, 4,
= Zεn +1 × Kn ,
εn
Zεm
× Zεεnm +1 × Lm,n , m ≥ 5,
c +1
n
b
m
c +εm ) 2m+ε
m!
b
n!.
where |Kn | = 2n−εn n! and |Lm,n | = 2−εn (εm
Lemma 3.6. Let m, m′ , n, n′ , k be two natural numbers greater than one. Then
b = (4, 3),
(1) Sk ∼
= Hm , if and only if (k, m)
(2) Sk ∼
= Kn , if and only if (k, n) = (4, 3),
(3) Sk ∼
6 Lm,n ,
=
b =m
b ′,
(4) Hm ∼
= Hm′ , if and only if m
∼
′
b = n′ ,
(5) Hm = Kn , if and only if m
∼
′
′
(6) Hm =
6 Lm ,n ,
(7) Kn ∼
= Kn′ , if and only if n = n′ ,
(8) Kn ∼
6 Lm′ ,n′ ,
=
(9) Lm,n ∼
= Lm′ ,n′ , if and only if (m, n) = (m′ , n′ ).
b
m
c m!.
Proof. (1) If Sk ∼
b Then k > m,
b which implies that
= Hm , then k! = 2m−ε
k−1−εk−1
k−2
m
b = k − 1. Thus k = 2
=2
, which holds only if k = 4. On the other
hand, Z2 ≀ S3 ∼
= Z2 × S4 and the result follows.
(2) It is similar to case (1).
b n n! for l = (m
b + εm )n − εn (εm
(3) If Sk ∼
= Lm,n , then k! = 2l m!
b + εm ), which
implies that k > n. Since epk (k!) = 1, it follows that epk (m)
b = 0 and epk (n!) = 1.
(2l m!
b n n!), we
(k!) = epm
Hence m
b < pk ≤ n < k ≤ 2n. If m
b > 2, then since epm
c
c
have
(n)
(k) n − Spm
k − Spm
c
c
−
.
(n!) =
(k!) − epm
n = epm
c
c
pm
−
1
p
−
1
b
m
b
Thus
(n),
(k) < n + Spm
(n) − Spm
(pm
b − 1)n = k − n + Spm
c
c
c
(n), a contradiction. If m
b = 2, then
which implies that n ≤ (pm
b − 2)n < Spm
c
k! = 2(3+εm )n−εm εn n!. Hence n = k − 1 and k = 2(3+εm )k−3−2εm ≥ 23k−5 , which
is impossible since k > 2. Also, if m
b = 1, then m = 5 and k! = 22n−2εn n!. Hence
2k−4
n = k − 1 and k = 2
, which is impossible.
b
b ′ −εm
m
c m!
c′ m
b = 2m
b ′ !. Suppose on the contrary that
(4) If Hm ∼
= Hm′ , then 2m−ε
′
′
′
m
b 6= m
b and m
b >m
b , say. Then m
b =m
b − 1 and we have
b
b
m
c −1 )−(m−ε
m
c)
m
b = 2(m−1−ε
= 2−2 ,
which is not an integer, a contradiction.
(5) It is similar to case (4).
b > n′ and the same as in (3)
(6) If Hm ∼
= Lm′ ,n′ , then it is easy to see that m
′
b
m
c m!
we can show that m
b ′ ≤ 2. If m
b ′ = 2, then 2m−ε
b = 2(3+εm′ )n −εm′ εn′ n′ !. Hence
n′ = m
b − 1 and
b
b
m
c)
c −1 )−(m−ε
m′ εm
m
b = 2((3+εm′ )(m−1)−ε
b
m′
= 2(2+εm′ )m−3−2ε
b
≥ 22m−5
,
8
A. AZIMI, A. ERFANIAN AND M. FARROKHI D. G.
b
m
c m!
which is impossible since m
b > 2. Also, if m
b ′ = 1, then m′ = 5 and 2m−ε
b =
2n′ −2εn′ ′
′
2
n !. Hence n = m
b − 1 and
b
b
b
m
c)
m
c −1 )−(m−ε
= 2m−4
,
m
b = 2(2(m−1)−2ε
which is impossible.
(7) It is similar to case (4).
(8) It is similar to case (6).
′
′
b n n! = 2l m
b ′ !n n′ !, where l = (m
b + εm )n − εn (εm
(9) If Lm,n ∼
= Lm′ ,n′ , then 2l m!
b +
′
εm ) and l′ = (m
b ′ + εm′ )n′ − εn′ (εm
b ′ + εm′ ). If n = n , then
′
b
b +εm′ )n−εn (εm
m )n−εn (εm
c +εm )
c ′ +εm′ ) m
2(m+ε
m!
b n = 2 (m
b ′ !n .
If m
b 6= m
b ′ and m
b >m
b ′ , say, then m
b′ = m
b − 1 and we have
b
b
m )n−εn (εm
c −1 +εm′ ))−((m+ε
c +εm ))
m′ )n−εn (εm
m
b n = 2((m−1+ε
= 2(εm′ −εm −1)n−εn (1+εm′ −εm ) ≤ 1,
which is a contradiction. Thus m
b =m
b ′ and consequently m = m′ . Thus we may
′
assume that n 6= n and without loss of generality n < n′ . First suppose that
′
′
(2l m
b ′ !n n′ !), we obtain
(2l m!
b n n!) = epm
m
b > 2. Since epm
c
c
(m
b ′ !).
(n!) ≥ n′ epm
(n′ !) − epm
(m
b ′ !) + epm
n = n′ epm
c
c
c
c
(m
b ′ !) = 0. Hence m
b ′ < pm
b
(m
b ′ !) ≤ n/n′ < 1 and consequently epm
Thus epm
b ≤ m.
c
c
Now, we have
(n)
(n′ ) n − Spm
n′ − Spm
c
c
−
.
(n!) =
(n′ !) − epm
n = epm
c
c
pm
−
1
p
−
1
b
m
b
Thus
(n) > (pm
(n′ ) + n − Spm
n′ = (pm
b − 1)n ≥ 2n.
b − 1)n + Spm
c
c
′
′
On the other hand, epn′ (2l m!
b n n!) = epn′ (2l m
b ′ !n n′ !) and by invoking Bertrand’s
′
postulate, if follows that nepn′ (m!)
b = n epn′ (m
b ′ !) + 1. Thus epn′ (m!)
b 6= 0 and
′
consequently epn′ (m
b !) 6= 0. Also, by Bertrand’s postulate, we have
Thus
m
b − Spn′ (m)
b
epn′ (m!)
b
n′
m
b
=
>
> pm
.
b −1≥
m
b ′ − Spn′ (m
b ′)
epn′ (m
b ′ !)
n
2
2(m
b − Spn′ (m))
b
Sp ′ (m)
b
=2−2 n
< 2.
m
b
m
b
Hence m
b ′ − Spn′ (m
b ′ ) ≤ 1. Since m
b ′ − Spn′ (m
b ′ ) = (pn′ − 1)epn′ (m
b ′ !), it follows that
epn′ (m
b ′ !) = 0, which is a contradiction.
Thus m
b ≤ 2. If m
b = 2, then
m
b ′ − Spn′ (m
b ′) <
′
′
′
b +εm′ )n −εn′ (εm
c ′ +εm′ ) m
2(3+εm )n−εm εn n! = 2(m
b ′ !n n′ !.
′
Hence n = n′ − 1 and m
b ′ ≤ 2. If m
b ′ = 2, then n′ = 2(εm −εm′ )n −3−2εm , which
′
implies that εm = 1 and εm′ = 0. Hence n′ = 2n −5 , which holds only if n′ = 8 and
consequently n = 7. Then
(Z2 ≀ S2 × Z2 ) ≀ S7 ∼
= Z2 × (Z2 ≀ S2 ) ≀ S8 ,
which is impossible by Jordan-Hölder’s theorem for the simple sections of the left
hand side group are Z2 and A7 while the simple sections of the right hand side
′
′
group are Z2 and A8 . Also, if m
b ′ = 1, then n′ = 2(2+εm −εm′ )n −3−2εm ≥ 22n −5 ,
which is impossible.
JACOBSON GRAPH
9
Finally, suppose that m
b = 1. Then m = 5 and we have
′
′
′
b +εm′ )n −εn′ (εm
c ′ +εm′ ) m
22n−2εn n! = 2(m
b ′ !n n′ !.
Hence n = n′ − 1 and m
b ′ = 1. Thus m′ = 5 so that n′ = 2−4 , which is not an
integer. The proof is complete.
n
n
Theorem 3.7. Let R = F1m1 ⊕ · · · ⊕ Fsms and R′ = F ′ 1 1 ⊕ · · · ⊕ F ′ t t be two finite
semi-simple rings.
(1) If mi , nj > 1 for i = 1, . . . , s and j = 1, . . . , t, and either the number of
fields of order 2 in R and R′ is not equal to 4 or the number of fields of
order 3 and 4 in R and R′ is not equal to 3, then R ∼
= R′ if and only if
′
).
Aut(JR ) ∼
Aut(J
=
R
(2) If all the fields Fi and Fj have the same parity of order different from 3
and 4, and either the number of fields of order 2 in R and R′ is not equal
to 4 or the number of fields of order 8 and 9 in R and R′ is greater than 1,
then R ∼
= R′ if and only if Aut(JR ) ∼
= Aut(JR′ ).
Proof. We know, by Krull, Remak and Schmidt Theorem [12, Theorem 3.3.8], that
every finite group can be expressed, up to isomorphism, as a direct product of some
indecomposable groups. Now, the result follows by using Corollary 3.4 and Lemmas
3.5 and 3.6.
Acknowledgment. The authors would like to thank Prof. Peter M. Neumann
for pointing out that the groups in Lemma 3.6 have different orders except for few
possibilities.
References
[1] S. Akbari and A. Mohammadian, On the zero-divisor graph of a commutative ring, J. Algebra
274(2) (2004), 847–855.
[2] D. F. Anderson, A. Frazier, A. Lauve and P. S Livingston, The zero-divisor graph of a commutative ring, II, Lecture Notes in Pure and Appl. Math 220 (2001), 61–72.
[3] D. F. Anderson and P. S. Livingston, The zero-divisor graph of a commutative ring, J. Algebra
217 (1999), 434–447.
[4] A. Azimi, A. Erfanian and M. Farrokhi D. G., The Jacobson graph of commutative rings, To
appear in J. Algebra Appl.
[5] M. R. Dixon and T. A. Fournelle, The indecomposability of certain wreath products indexed
by partially ordered sets, Arch. Math. 43 (1984), 193–207.
[6] J. Han, The zero-divisor graph under group action in a noncommutative ring, J. Korean Math.
Soc. 45(6) (2008), 1647–1659.
[7] F. Harary, The automorphism group of a hypercube, J. UCS 6(1) (2000), 136–138.
[8] M. W. Liebeck, Graphs whose full automorphism group is a symmetric group, J. Austral.
Math. Soc. Ser. A 44(1) (1988), 46–63.
[9] B. R. Macdonald, Finite Rings with Identity, Marcel Dekker, Inc., New York, 1974.
[10] A. Mohammadian, On zero-divisor graphs of Boolean rings, Pacific J. Math. 251(2) (2011),
375–383.
[11] S. Park and J. Han, The group of graph automorphisms over a matrix ring, J. Korean Math.
Soc. 48(2) (2011), 301–309.
[12] D. J. S. Robinson, A Course in the Theory of Groups, Second Edition, Springer-Verlag, New
York, 1995.
Department of Pure Mathematics, Ferdowsi University of Mashhad, Mashhad, Iran
E-mail address: [email protected]
Department of Pure Mathematics, Ferdowsi University of Mashhad, Mashhad, Iran
E-mail address: [email protected]
10
A. AZIMI, A. ERFANIAN AND M. FARROKHI D. G.
Department of Pure Mathematics, Ferdowsi University of Mashhad, Mashhad, Iran
E-mail address: [email protected]
| 0math.AC
|
Towards ECDSA key derivation from deep embeddings for
novel Blockchain applications
Christian S. Perone
arXiv:1711.04069v1 [cs.CV] 11 Nov 2017
[email protected]
http://blog.christianperone.com
Abstract. In this work, we propose a straightforward method to derive Elliptic Curve Digital
Signature Algorithm (ECDSA) key pairs from embeddings created using Deep Learning and Metric
Learning approaches. We also show that these keys allows the derivation of cryptocurrencies (such
as Bitcoin) addresses that can be used to transfer and receive funds, allowing novel Blockchain-based
applications that can be used to transfer funds or data directly to domains such as image, text,
sound or any other domain where Deep Learning can extract high-quality embeddings; providing
thus a novel integration between the properties of the Blockchain-based technologies such as trust
minimization and decentralization together with the high-quality learned representations from Deep
Learning techniques.
1
Introduction
Decentralized cryptocurrencies such as Bitcoin [15], enabled a wide spectrum of novel applications. Based
on the Blockchain, the underlying distributed core of Bitcoin, applications ranging from smart contracts,
verifiable data, and transactions, audit systems, among others, made a significant impact on the way that
we can make transactions and build systems with trust minimization or no third-party involved at all.
Supported and advocated by a vibrant community, Blockchain is a technology that can be applied to
many contexts beyond cryptocurrency and financial assets [20].
In the recent years, we also saw the fast development of Deep Learning [11], especially in the Computer
Vision area, where we were able to see major milestones such as the well known AlexNet [10], VGGNet
[17] GoogleNet [21] and more recently the ResNets [8]. The representation of high-dimensional data, such
as natural images, by low-dimensional embeddings learned using deep neural networks with multiple
hierarchical features, allowed not only image classification, but also many different applications such as
visual search [9] [2], image hashing [13] [14] [4], image captioning [23] and many other applications.
Although the synergy between the Blockchain or Blockchain-based technology and Machine Learning/AI
was mentioned in the health context, such as in [20] for the application of secure large-scale data
management mechanism to coordinate the information of individuals, and more recently in [18] for
the application of a verifiable data audit system, there are very few published works or implemented
applications exploring this integration between learned representations and the Blockchain beyond the
trend of Machine Learning model auditing. In [20], the author also speculatively cites an application of
the Blockchain and AI systems, but for the context of communication between AI entities in order to
conduct certain transactions.
In this context, our work focuses on the exploration of this potential synergy between representation
learning and the Blockchain by means of novel practical application example and experimentation.
Our contributions are as follows: first, we focus on the usage of Convolutional Neural Networks
(CNNs) [12] and metric learning [5] to produce embeddings for the purpose of cryptography key-pair
derivation. Later, we develop a method for deriving Elliptic Curve Digital Signature Algorithm (ECDSA)
key-pairs from deep embeddings extracted from natural images, as well as an application example using the
Blockchain technology that will allow us to create transactions that can only be redeemed by individuals
owning the same similar images or the transaction owner. The ECDSA is used by the Bitcoin to promote
cryptography guarantees for the transactions, ensuring that funds can only be spent by their owners, thus
we use the same mechanisms in this work to promote cryptographic guarantees for the fund transactions.
Later, we discuss future extensions of this work and other potential applications of this integration between
representation learning and cryptocurrencies/Blockchain.
2
2
Towards ECDSA key derivation from deep embeddings for novel Blockchain applications
Related Work
As we mentioned in the previous section, there are very few published works related to AI/Machine
Learning and Blockchain/cryptocurrencies. In [20], the author explores some Blockchain applications in a
highly-speculative manner. One of the applications that the author cite is the usage of the Blockchain
by AI entities as a mean to provide a permanent transparent public record that can be reviewed and
inspected.
In [19], they cite an application called Monegraph, created to provide property ownership of assets
such as online graphics or digital media, where a user can pay a small network fee via namecoin and have
its Twitter account and URL being placed on the Blockchain. However, they do not account for similarity
issues. In [19], they also cite a digital art and copyright protection project called Ascribe, however, they
use Machine Learning for similarity search while crawling the web, so there is no evidence of a link between
digital art extracted features embedded on the Blockchain. Only the hash of the digital content and
metadata is added to Blockchain.
Some works such as Crypto-Nets [24] used homomorphic encryption to provide privacy-preserving
guarantees, but no keys were derived from the embeddings since the goal was to provide privacy.
A thorough search of the relevant literature yielded no relevant similar work. To the best of our
knowledge, this is the first study exploring the use of embeddings in the context of the Blockchain.
3
3.1
Background
ECDSA curves and the secp256k1
The Elliptic Curve Digital Signature Algorithm, also called ECDSA, is the technique used by the Bitcoin
cryptocurrency to guarantee that the funds can only be used by their rightful owners. It is beyond the
scope of this work to describe how ECDSA works, so we’ll just describe the important concepts that will be
mentioned in this work. In order to use Elliptic Curve Cryptography (ECC), all parties must agree on all
the elements defining the elliptic curve, the domain parameters. Many standard bodies published domain
parameters of elliptic curves for different field sizes. The standard used by Bitcoin is called secp256k1
and was proposed by [6], where it was constructed in a special way to allow efficient computation. The
secp256k1’s parameters were also selected in a predictable way, which significantly reduces the possibility
that the curve’s creator inserted any sort of back-door into the curve [1].
3.2
Bitcoin Addresses and Transactions
A pseudo-random number generator (PRNG) is typically used
in Bitcoin to generate the private key for the secp256k1, where
a 256-bit number is randomly generated and then used as the
secret exponent of the secp256k1 curve. It is important to note
that not all 256-bit numbers are valid ECDSA private keys and
this range of valid private keys is governed by the secp256k1
ECDSA standard, however, the upper bound of the valid exponent is very close to the largest 256-bit number. The main goal
of this work is to derive this private key using deterministic representations created by Metric Learning over features extracted
from deep Convolutional Neural Networks (CNNs) instead of
using a pseudo-random number generator.
Key pair and address generation. The process to create
a Bitcoin address starts with the generation of the non-zero
random integer d private key and then a public key Q = dG
is derived, where G is the base point parameter. After that,
the process to generate a Bitcoin address can be seen in the
Figure 1, where basically, there are two rounds of two different
hash functions (SHA-256 and RIPEMD-160) over the public key,
after that the SHA256D is used to create a checksum and then
the address is encoded into a Base58 encoding scheme. What
is important to note here, is that the final Bitcoin address isn’t Fig. 1: An overview of the Bitcoin adthe public key itself, but the hash of the public key.
dress generation.
Towards ECDSA key derivation from deep embeddings for novel Blockchain applications
3
Transactions. There are two main standard transaction types
that are used on the Bitcoin network nowadays. The first one is
called “Pay To Public Key Hash” or simply P2PKH. This transaction, as the name states, pays for the
owner of the public key that hashes to the specified address. Since the sender can’t provide the public key,
when redeeming the funds of the transaction, the owner of the public key must provide his public key that
hashes to the specified address and also the signature.
The other commonly used transaction type is called “Pay To Script Hash”, where the responsibility for
supplying the conditions to redeem a transaction is moved from the sender of the funds to the redeemer
of the funds. This transaction type allows flexibility for specifying larger and complex scripts without
letting the sender to worry about it or even know the script contents.
4
Methods and Materials
Learning a good representation that can be used to represent different natural images is an important
task. In our context, we not only want to learn a good representation with similarity properties but also
a representation that can be binarized and used as the deterministic seed for the ECDSA private key
derivation.
There has been considerable research on distance metric learning over the past few years [25] [3]. The
goal of metric learning is to adapt some pairwise real-valued metric function, using the information from
training examples [3]. Most methods learn the metric in a weakly-supervised way from pairs or triplets
[16]. Since the creation of an architecture to learn the embeddings isn’t the main goal of this work, we
used a simple siamese network [5] with contrastive loss to learn the representations such that the squared
L2 distances in the embedding space directly correspond to image similarity.
Since the indefinite contraction of the pairs was found to be detrimental to the training dynamics [13],
we employed the same proposed double-margin loss as seen in [13]:
L(zα , zβ ) =y max(kzα − zβ k22 − m1 , 0) + (1 − y) max(m2 − kzα − zβ k22 , 0)
(1)
Where the m1 and m2 are two different margins that we choose experimentally, zα , zβ are the training
pairs and y is equal to 1 when the pairs are a genuine pair or 0 when they are impostors.
As we mentioned before, we need a representation that when binarized will have an approximate
performance result of the original representation for retrieval. To accomplish such representation property,
we applied a triangular regularization prior on the activations of the last layer of the the network (the
layer we use to extract embeddings) described in the element-wise equation below:
freg (ψ) = λ ∗ (1 − |(σ(x) ∗ 2) − 1|)
(2)
Where σ(·) is the sigmoid activation function, x is the preActivations Distribution
activation and λ is a balancing term. We used a pre-trained
ResNet-50 [7] as the network architecture for feature extraction 3.5
3.0
and we added a fully-connected layer of 256 units as the last
2.5
layer. The network training process was done using random
2.0
positive and negative sample pairs from the ImageNet dataset. 1.5
After training the network with the proposed regularization (eq. 1.0
1), we can see that the activations lie mostly on the upper and 0.5
lower bounds of the sigmoid activation as seen in Figure 2.
0.0 0.0
0.2
0.4
0.6
0.8
1.0
In the Figure 3, we show the Euclidean distance between
(x)
the feature vector of a single random sample from the ImageNet
belonging to the class 129, which represents the Spoonbill bird. Fig. 2: Activation distribution after
As we can see, the most similar classes are also close to the training using the proposed regularizaSpoonbill class, that also represent birds such as Flamingos, tion.
hence the lower Euclidean distance.
In Figure 4, we show the t-SNE[22] visualization of the feature vectors before and after the binarization,
where we can see that similar clustering patterns are preserved after the binarization process.
After training the network, we use it to extract feature representations from images and then binarize
the feature vector v ∈ R256 into a binary vector v ∈ Z256
2 , where Z2 = [0, 1]. After that, we use the binary
feature vector as the secret for deriving the public key using the secp256k1 base point parameters and
then we follow the same Bitcoin address generation protocol where a series of hashed are applied over
4
Towards ECDSA key derivation from deep embeddings for novel Blockchain applications
ImageNet classes ranging from 0 to 499
0
20
40
0
100
500
600
200
300
400
700
800
900
ImageNet classes ranging from 500 to 1000
0
20
40
0
1
2
3
4
5
6
Euclidean Distance
7
8
Fig. 3: Euclidean distance similarity for one sample.
tSNE Embedding (binarized features)
tSNE Embedding (binarized features)
40
40
800
600
0
400
20
200
40
40
20
0
20
0
40
(a) tSNE of non-binarized features.
20
ImageNet Classes
ImageNet Classes
20
800
600
0
400
20
200
40
40
20
0
20
40
(b) tSNE of binarized features.
Fig. 4: tSNE embeddings.
0
Towards ECDSA key derivation from deep embeddings for novel Blockchain applications
5
the public key coordinates to derive the final Bitcoin address. In Figure 5 we show an overview of the
complete described flow.
Fig. 5: An overview of the training procedure and method to derive the ECDSA key pair using the
representations learned by metric learning using a CNN.
The algorithm for deriving the public key can be seen in the Algorithm 1, where we describe the steps
to derive the public key from the feature vectors.
Algorithm 1 ECDSA key-pair derivation algorithm
1: procedure DeriveKeyPairs
2:
network ← pre-trained siamese network
3:
image ← input image
4:
repr ← network(image)
5:
secret ← binarize(repr)
6:
opt passwd ← P BKDF 2(passwd)
7:
priv key ← secret k opt passwd
8:
pub key ← priv key ⊗ G
. Forward pass to get representation repr ∈ R128
. Binarize features, repr bin ∈ Z256
2 , where Z2 = [0, 1]
. Use a password (optional)
. Concatenate the secret with the password (optional)
. Derive the ECDSA public key using secp256k1
As we can see from the Algorithm 1, the procedure to derive the key is straightforward. First we feed
the desired image into the network to compute the forward pass, after that we use the extracted features
and binarize them to create the ECDSA secret. As optional stage, we can concatenate (taking bit-length
into consideration) the derived secret with a user password where a PBKDF2 function is applied to avoid
some attack vectors such as rainbow tables, and then finally we use the ECDSA private-key to derive the
public-key by doing the field multiplication with the secp256k1 parameters.
After creating the public-key, we can derive a Bitcoin address using the scheme shown in Figure 1. With
the Bitcoin address in hands, we can transfer funds to this addres in a way that it can be redeemed only
when a similar image on the learned low-dimensional manifold instead of other pixel space or traditional
image hashing is used to derive the private-key, or if using the optional password: only someone with the
password and similar image will be able to redeem the transfered funds.
5
Discussion
As we saw, through a very simple process, we can create an ECDSA key-pair that is derived not from a
random seed but from a binarized feature vector representing an image. This key-pair can be used to sign
or encrypt data and in our context, can be used to derive Bitcoin addresses that only owners of the same
or similar images (on the learned low-dimensional manifold) can redeem the contents of the transaction.
By using a P2PKH (Pay to Public Key Hash) transaction, one can transfer cryptocurrency funds to the
image.
The described technique is not limited to the Bitcoin cryptocurrency but can be used for any Blockchainbased technology that relies on the ECDSA. Although we provided a very simple siamese architecture
with a simple regularization mechanism to keep the Euclidean distance on the low dimensional manifold,
6
Towards ECDSA key derivation from deep embeddings for novel Blockchain applications
our technique isn’t limited to this particular architecture, therefore other ”deep hash” techniques such as
HashNets [4] can be employed to create the representations that will be later used to derive the ECDSA
key-pair. Our method is also not limited to the image domain, since the representations can be created
from text or even other different input domains. In the Bitcoin context, users can transfer funds to natural
images, texts, sounds or any other domain where Deep Learning can successfully extract high-quality
embeddings, providing a new door for many potential applications based on the Blockchain technology.
To conclude, we showed in this work an approach to derive ECDSA keys from feature vectors, allowing
many novel applications where users can transfer funds or store decentralized data on natural images,
text and sounds. To the best of our knowledge, this is the first study to propose a concrete method to
allow the integration between Deep Learning and the Blockchain technology. Further studies are certainly
required in order to assess the potential attack vectors to this approach, however we believe that this
integration would certainly open the door to novel applications that can provide trust minimization for
transfering data or funds to many different data domains.
6
Additional Information
Software: To develop this work, we used the Keras framework using the TensorFlow backend. To create
the ECDSA keys and derive Bitcoin addresses we used the Protocoin1 framework for creating the ECDSA
keys and Bitcoin addresses.
Disclaimer: This method is provided by the author and contributors ”as is” and any express or implied
warranties, including, but not limited to, the implied warranties of merchantability and fitness for a
particular purpose are disclaimed. In no event shall the copyright owner or contributors be liable for any
direct, indirect, incidental, special, exemplary, or consequential damages (including, but not limited to,
procurement of substitute goods or services; loss of use, data, or profits; or business interruption) however
caused and on any theory of liability, whether in contract, strict liability, or tort (including negligence
or otherwise) arising in any way out of the use of this method, even if advised of the possibility of such
damage. The author isn’t responsible for any loss of cryptocurrencies or any financial asset due to the use
of the method described in this article.
References
1. Secp256k1 - Bitcoin Wiki, https://en.bitcoin.it/wiki/Secp256k1
2. Bell, S.: Learning visual similarity for product design with convolutional neural networks. Siggraph 34(4), 1–9
(2015)
3. Bellet, A., Habrard, A., Sebban, M.: A Survey on Metric Learning for Feature Vectors and Structured Data.
Bmvc2015 p. 57 (2015)
4. Cao, Z., Long, M., Wang, J., Yu, P.S.: HashNet: Deep Learning to Hash by Continuation. arXiv (2017)
5. Chopra, S., Hadsell, R., Y., L.: Learning a similiarty metric discriminatively, with application to face verification.
Proceedings of IEEE Conference on Computer Vision and Pattern Recognition pp. 349–356 (2005)
6. Daniel R. L. Brown: Standards for Efficient Cryptography 2 (SEC 2) : Recommended Elliptic Curve Domain
Parameters. Standards for Efficient Cryptography p. 37 (2010)
7. He, K., Zhang, X., Ren, S., Sun, J.: Deep Residual Learning for Image Recognition. In: 2016 IEEE Conference
on Computer Vision and Pattern Recognition (CVPR). pp. 770–778 (2016)
8. He, K., Zhang, X., Ren, S., Sun, J.: Identity mappings in deep residual networks. Lecture Notes in Computer
Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics) 9908
LNCS, 630–645 (2016)
9. Jing, Y., Liu, D., Kislyuk, D., Zhai, A., Xu, J., Donahue, J., Tavel, S.: Visual Search at Pinterest. Proceedings
of the 21th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining pp. 1889–1898
(2015)
10. Krizhevsky, A., Sutskever, I., Hinton, G.E.: ImageNet Classification with Deep Convolutional Neural Networks.
Advances In Neural Information Processing Systems pp. 1–9 (2012)
11. LeCun, Y., Bengio, Y., Hinton, G., Y., L., Y., B., G., H.: Deep learning. Nature 521(7553), 436–444 (2015)
12. LeCun, Y., Huang, F.J.H.F.J., Bottou, L.: Learning Methods for Generic Object Recognition with Invariance
to Pose and Lighting. Computer Vision and Pattern Recognition, 2004. CVPR 2004. Proceedings of the 2004
IEEE Computer Society Conference on 2, II–97 – 104 (2004)
13. Lin, J., Morere, O., Chandrasekhar, V.: DeepHash: Getting Regularization, Depth and Fine-Tuning Right.
arXiv preprint arXiv: . . . p. 20 (2015)
1
http://protocoin.readthedocs.io/
Towards ECDSA key derivation from deep embeddings for novel Blockchain applications
7
14. Liong, V.E., Lu, J., Wang, G., Moulin, P., Zhou, J.: Deep hashing for compact binary codes learning. In:
Proceedings of the IEEE Computer Society Conference on Computer Vision and Pattern Recognition. vol.
07-12-June, pp. 2475–2483 (2015)
15. Nakamoto, S.: Bitcoin: A Peer-to-Peer Electronic Cash System. Www.Bitcoin.Org p. 9 (2008), https://
bitcoin.org/bitcoin.pdf
16. Schroff, F., Kalenichenko, D., Philbin, J.: FaceNet: A unified embedding for face recognition and clustering.
Proceedings of the IEEE Computer Society Conference on Computer Vision and Pattern Recognition 07-12June, 815–823 (2015)
17. Simonyan, K., Zisserman, A.: Very Deep Convolutional Networks for Large-Scale Image Recognition. ImageNet
Challenge pp. 1–10 (2014)
18. Suleyman, M., Laurie, B.: Trust, confidence and Verifiable Data Audit (2017), https://deepmind.com/blog/
trust-confidence-verifiable-data-audit/
19. Swan, M.: Blockchain: blueprint for a new economy. O’Reilly, first edit edn. (2015)
20. Swan, M.: Blockchain Thinking : the Brain as a Decentralized Autonomous Corporation [Commentary] (2015)
21. Szegedy, C., Liu, W., Jia, Y., Sermanet, P., Reed, S., Anguelov, D., Erhan, D., Vanhoucke, V., Rabinovich, A.,
Hill, C., Arbor, A.: Going Deeper with Convolutions (2014)
22. Van Der Maaten, L., Hinton, G.: Visualizing Data using t-SNE. Journal of Machine Learning Research 9,
2579–2605 (2008)
23. Vinyals, O., Toshev, A., Bengio, S., Erhan, D.: Show and Tell: Lessons learned from the 2015 MSCOCO Image
Captioning Challenge. TPAMI 99(PP), 1–1 (2016)
24. Xie, P., Bilenko, M., Finley, T., Gilad-Bachrach, R., Lauter, K., Naehrig, M.: Crypto-Nets: Neural Networks
over Encrypted Data (2014), https://arxiv.org/pdf/1412.6181.pdfhttp://arxiv.org/abs/1412.6181
25. Yang, L., Jin, R.: Distance metric learning: A comprehensive survey. Michigan State Universiy pp. 1–51 (2006)
| 1cs.CV
|
Medoids in almost linear time via multi-armed bandits
arXiv:1711.00817v3 [stat.ML] 7 Nov 2017
Vivek Bagaria∗
Stanford University
[email protected]
Govinda M. Kamath∗
Stanford University
[email protected]
Martin J. Zhang∗
Stanford University
[email protected]
Vasilis Ntranos∗
UC, Berkeley
[email protected]
David Tse
Stanford University
[email protected]
Abstract
Computing the medoid of a large number of points in high-dimensional space is
an increasingly common operation in many data science problems. We present
an algorithm Med-dit which uses O(n log n) distance evaluations to compute
the medoid with high probability. Med-dit is based on a connection with the
multi-armed bandit problem. We evaluate the performance of Med-dit empirically
on the Netflix-prize and the single-cell RNA-Seq datasets, containing hundreds of
thousands of points living in tens of thousands of dimensions, and observe a 5-10x
improvement in performance over the current state of the art. Med-dit is available
at https://github.com/bagavi/Meddit.
1
INTRODUCTION
An important building block in many modern data analysis problems such as clustering is the efficient
computation of a representative point for a large set of points in high dimension. A commonly used
representative point is the centroid of the set, the point which has the smallest average distance from
all other points in the set. For example, k-means clustering [28, 23, 21] computes the centroid of
each cluster under the squared Euclidean distance. In this case, the centroid is the arithmetic mean of
the points in the cluster, which can be computed very efficiently. Efficient computation of centroid is
central to the success of k-means, as this computation has to be repeated many times in the clustering
process.
While commonly used, centroid suffers from several drawbacks. First, the centroid is in general
not a point in the dataset and thus may not be interpretable in many applications. This is especially
true when the data is structured like in images, or when it is sparse like in recommendation systems
[19]. Second, the centroid is sensitive to outliers: several far away points will significantly affect
the location of the centroid. Third, while the centroid can be efficiently computed under squared
Euclidean distance, there are many applications where using this distance measure is not suitable.
Some examples would be applications where the data is categorical like medical records [29]; or
situations where the data points have different support sizes such as in recommendation systems
[19]; or cases where the data points are on a high-dimensional probability simplex like in single cell
RNA-Seq analysis [25]; or cases where the data lives in a space with no well known Euclidean space
like while clustering on graphs from social networks.
An alternative to the centroid is the medoid; this is the point in the set that minimizes the average
distance to all the other points. It is used for example in k-medoids clustering [15]. On the real line,
the medoid is the median of the set of points. The medoid overcomes the first two drawbacks of the
centroid: the medoid is by definition one of the points in the dataset, and it is less sensitive to outliers
than the centroid. In addition, centroid algorithms are usually specific to the distance used to define
the centroid. On the other hand, medoid algorithms usually work for arbitrary distances.
∗
Contributed equally and listed alphabetically.
Figure 1: We plot the probability that RAND and Med-dit do not return the true medoid as a function
of the number of distance evaluations per point over 1000 monte-carlo trials. We note that Med-dit
needs much fewer distance evaluations than RAND for comparable performance. The computation of
the true medoid here is also computationally prohibitive and is discussed in Section 4.
The naive method to compute the medoid would require computing all pairwise distances
between
points in the set. For a set with n points, this would require the computation of n2 distances, which
would be computationally prohibitive when there are hundreds of thousands of points and each point
lives in a space with dimensions in tens of thousands.
In the one-dimensional case, the medoid problem reduces to the problem of finding the median,
which can be solved in linear time through Quick-select [12]. However, in higher dimensions,
no linear-time algorithm is known. RAND [8] is an algorithm that estimates the average distance
of each
pointto all the other points by sampling a random subset of other points. It takes a total
n
of O n log
distance computations to approximate the medoid within a factor of (1 + ∆) with
high probability, where ∆ is the maximum distance between two points in the dataset. We remark
that this is an approximation algorithm, and moreover ∆ may not be known apriori. RAND was
leveraged by TOPRANK [26] which uses the estimates obtained by RAND to focus on a small subset
of candidate points, evaluates the average distance of these points exactly, and picks the minimum
4
5
of those. TOPRANK needs O(n 3 log 3 n) distance computations to find the exact medoid with high
probability under a distributional assumption on the average distances. trimed [24] presents a clever
3
algorithm to find the medoid with O(n 2 2Θ(d) ) distance evaluations,
which works very well when the
√
points live in a space of dimensions less than 20 (i.e., 2d n) under a distributional assumption on
the points (in particular on the distribution of points around the medoid). However, the exponential
dependence on dimension makes it impractical when d ≥ 50, which is the case we consider here.
Note that their result requires the distance measure to satisfy the triangle inequality.
In this paper, we present Med(oid)-(Ban)dit, a sampling-based algorithm that finds the exact
medoid with high probability. It takes O(n log n) distance evaluations under natural assumptions on
the distances, which are justified by the characteristics observed in real data (Section 2). In contrast to
trimed, the number of distance computations is independent of the dimension d; moreover, there is
no specific requirement for the distance measure, e.g. the triangle inequality or the symmetry. Thus,
Med-dit is particularly suitable for high-dimensional data and general distance measures.
In Figure 1, we showcase the performance of RAND and Med-dit on the largest cluster of the single
cell RNA-Seq gene expression dataset of [1]. This consists of 27, 998 gene-expressions of each of
109, 140 cells, i.e 109, 140 points in 27, 998 dimensions. We note that Med-dit evaluates around 10
2
times fewer distances than RAND to achieve similar performance. At this scale running TOPRANK and
trimed are computationally prohibitive.
The main idea behind Med-dit is noting that the problem of computing the medoid can be posed
as that of computing the best arm in a multi-armed bandit (MAB) setting [9, 13, 17]. One views
each point in the medoid problem as an arm whose unknown parameter is its average distance to all
the other points. Pulling an arm corresponds to evaluating the distance of that point to a randomly
chosen point, which provides an estimate of the arm’s unknown parameter. We leverage the extensive
literature on the multi-armed bandit problem to propose Med-dit, which is a variant of the Upper
Confidence Bound (UCB) Algorithm [17].
Like the other sampling based algorithms RAND and TOPRANK, Med-dit also aims to estimate the
average distance of every point to the other points by evaluating its distance to a random subset of
points. However, unlike these algorithms where each point receives a fixed amount of sampling
decided apriori, the sampling in Med-dit is done adaptively. More specifically, Med-dit maintains
confidence intervals for the average distances of all points and adaptively evaluates distances of
only those points which could potentially be the medoid. Loosely speaking, points whose lower
confidence bound on the average distance are small form the set of points that could potentially be the
medoid. As the algorithm proceeds, additional distance evaluations narrow the confidence intervals.
This rapidly shrinks the set of points that could potentially be the medoid, enabling the algorithm to
declare a medoid while evaluating a few distances.
In Section 2 we delve more into the problem formulation and assumptions. In Section 3 we present
Med-dit and analyze its performance. We extensively validate the performance of Med-dit empirically on two large-scale datasets: a single cell RNA-Seq gene expression dataset of [1], and the
Netflix-prize dataset of [3] in Section 4.
2
PROBLEM FORMULATION
Consider n points x1 , x2 , · · · , xn lying in some space U equipped with the distance function d :
U × U 7→ R+ . Note that we do not assume the triangle inequality or the symmetry for the distance
function d. Therefore the analysis here encompasses directed graphs, or distances like Bregman
divergence and squared Euclidean distance. We use the standard notation of [n] to refer to the set
{1, 2, · · · , n}. Let di,j , d(xi , xj ) and let the average distance of a point xi be
X
1
µi ,
di,j .
n−1
j∈[n]−{i}
The medoid problem can be defined as follows.
Definition 1. (The medoid problem) For a set of points X = {x1 , x2 , · · · , xn }, the medoid is the
point in X that has the smallest average distance to other points. Let xi∗ be the medoid. The index i∗
and the average distance µ∗ of the emdoid are given by
i∗ , argmin µi ,
µ∗ , min µi .
i∈[n]
i∈[n]
The problem of finding the medoid of X is called its medoid problem.
In the worst case over distances and points, we would need to evaluate O(n2 ) distances to compute
the medoid. Consider the following example.
Example 1. Consider a setting where there are n points. We pick a point i uniformly at random
from [n] and set all distances to point i to 0. For every other point j ∈ [n] − {i}, we pick a point k
uniformly at random from [n] − {i, j} and set dj,k = dk,j = n. As a result, each point apart from i
has a distance of n to at least one point. Thus picking the point with distance 0 would require at least
O(n2 ) distance evaluations. We note that this distance does not satisfy the triangle inequality, and
thus is not a metric.
The issue with Example 1 is that, the empirical distributions of distances of any point (apart from the
true medoid) have heavy tails. However, such worse-case instances rarely arise in practice. Consider
the following real-world example.
Example 2. We consider the `1 distance between each pair of points in a cluster – 109, 140 points in
27, 998 dimensional probability simplex – taken from a single cell RNA-Seq expression dataset from
3
Figure 2: Histogram of `1 distance of 4 randomly chosen points (from 20, 000 points). We note that
the standard deviations are around 0.05.
[1]. Empirically the `1 distance for a given point follows a Gaussian-like distribution with a small
variance, as shown in Figure 2. Modelling distances among high-dimensional points by a Gaussian
distribution is also discussed in [20]. Clearly the distances do not follow a heavy-tailed distribution
like Example 1.
2.1
The Distance Sampling Model
Example 2 suggests that for any point xi , we can estimate its average distance µi by the empirical
mean of its distance from a subset of randomly sampled points. Let us formally state this. For any
point xi , let Di = {di,j , j 6= i} be the set of distances associated with xi . Define the random distance
samples Di,(1) , Di,(2) , · · · , Di,(k) to be i.i.d. sampled with replacement from Di . We note that
E[Di,(j) ] = µi , ∀i ∈ [n], j ∈ [k]
Hence the average distance µi can be estimated as
µ̂i =
k
1X
Di,(k) .
k i=1
To capture the concentration property illustrated in Example 2, we assume the random variables
Di,(j) , i.e. sampling with replacement from Di , are σ-sub-Gaussian 2 for all i.
2.2
Connection with Best-Arm Problem
For points whose average distance is close to that of the medoid, we need an accurate estimate of
their average distance, and for points whose average distance is far away, coarse estimates suffice.
2
X is a σ-sub-Gaussian if P (X > t) ≤ 2e
2
− t2
σ
.
4
Algorithm 1 Med-dit
1: Evaluate distances of each point to a randomly chosen point and build a (1 − δ)-confidence
interval for the mean distance of each point i: [µ̂i (1) − Ci (1), µ̂i (1) + Ci (1)].
2: while true do
3:
At iteration t, pick point At that minimises µ̂i (t − 1) − Ci (t − 1).
4:
if distances of point At are evaluated less than n − 1 times then
evaluate the distance of At to a randomly picked point; update the confidence interval of At .
5:
6:
else
7:
Set µ̂i (t) to be the empirical mean of distances of point At by computing its distance to all
8:
9:
(n − 1) other points and set CAt (t) = 0.
if there exists a point xi∗ such that ∀i 6= i∗ , µ̂i∗ (t) + Ci∗ (t) < µ̂i (t) − Ci (t) then
return xi∗ .
Thus the problem reduces to adaptively choosing the number of distance evaluations for each point to
ensure that on one hand, we have a good enough estimate of the average distance to compute the true
medoid, while on the other hand, we minimise the total number of distance evaluations.
This problem has been addressed as the best-arm identification problem in the multi-armed bandit
(MAB) literature (see the review article of [13] for instance). In a typical setting, we have n arms. At
the time t = 0, 1, · · · , we decide to pull an arm At ∈ [n], and receive a reward Rt with E[Rt ] = µAt .
The goal is to identify the arm with the largest expected reward with high probability while pulling as
few arms as possible.
The medoid problem can be formulated as a best-arm problem as follows: let each point in the
ensemble be an arm and let the average distance µi be the loss of the i-th arm. Medoid corresponds
to the arm with the smallest loss, i.e the best arm. At each iteration t, pulling arm i is equivalent to
sampling (with replacement) from Di with expected value µi . As stated previosuly, the goal is to find
the best arm (medoid) with as few pulls (distance computations) as possible.
3
ALGORITHM
P
1
We have n points x1 , · · · , xn . Recall that the average distance for arm i is µi = n−1
j6=i di,j . At
iteration t of the algorithm, we evaluate a distance Dt to point At ∈ [n]. Let Ti (t) be the number of
distances of point i evaluated upto time t. We use a variant of the Upper Confidence Bound (UCB)
[17] algorithm to sample distance of a point i with another point chosen uniformly at random.
We compute the empirical mean and use it as an estimate of µi at time t, in the initial stages of the
algorithm. More concretely when Ti (t) < n,
X
1
µ̂i (t) ,
Dτ .
Ti (t)
1≤τ ≤t,Aτ =i
Next we recall from Section 2 that the sampled distances are independent and σ-sub-Gaussian (as
points are sampled with replacement). Thus for all i, for all δ ∈ (0, 1), with probability at least 1 − δ,
µi ∈ [µ̂i (t) − Ci (t), µ̂i (t) + Ci (t)], where
s
2σ 2 log 2δ
Ci (t) =
.
(1)
Ti (t)
When Ti (t) ≥ n, we compute the exact average distance of point i and set the confidence interval to
zero. Algorithm 1 describes the Med-dit algorithm.
Theorem 1. For i ∈ [n], let ∆i = µi − µ∗ . If we pick δ = n23 in Algorithm 1, then with probability
1 − o(1), it returns the true medoid with the number of distance evaluations M such that,
X 24σ 2
M≤
log n ∧ 2n .
(2)
∆2i
i∈[n]
If ∆σi is Θ(1), then Theorem 1 gives that Med-dit takes O(n log n) distance evaluations. In addition,
if one assumed a Gaussian prior on the average distance of points (as one would expect from Figure
5
Figure 3: The distribution of mean distances of all the points in the 20, 000 cell RNA-Seq dataset.
3), then Med-dit would also take O(n log n) distance evaluations in expectation over the prior. See
Appendix 1.
If we use the same assumption as TOP-RANK, i.e. the µi are i.i.d. Uniform [c0 , c1 ], then the number of
1
4
3
5
distance evaluations needed by Med-dit is O(n 2 log 2 n), compared to O(n 3 log 3 n) for TOP-RANK.
Under these assumptions RAND takes O(n2 ) distance evaluations to return the true medoid.
We assume that σ is known in the proof below, whereas we estimate them in the empirical evaluations.
See Section 4 for details.
Remark 1. With a small modification to Med-dit, the sub-Gaussian assumption in Section 2 can be
further relaxed to a finite variance assumption [4]. One the other hand, one could theoretically adapt
Med-dit to other best-arm algorithms to find the medoid with O(n log log n) distance computations
at the sacrifice of a large constant [14]. Refer to Appendix 2 for details.
Proof. Let M be the total number of distance evaluations when the algorithm stops. As defined
above, let Ti (t) be the number of distance evaluations of point i upto time t.
We first assume that [µ̂i (t) − Ci (t), µ̂i (t) − Ci (t)] are true 1 − n23 -confidence interval (Recall that
δ = n23 ) and show the result. Then we prove this statement.
Let i∗ be the true medoid. We note that if we choose to update arm i 6= i∗ at time t, then we have
µ̂i (t) − Ci (t) ≤ µ̂i∗ (t) − Ci∗ (t).
For this to occur, at least one of the following three events must occur:
E1 = {µ̂i∗ (t) ≥ µi∗ (t) + Ci∗ (t)} ,
E2 = {µ̂i (t) ≤ µi (t) − Ci (t)} ,
E3 = {∆i = µi − µi∗ ≤ 2Ci (t)} .
To see this, note that if none of E1 , E2 , E3 occur, we have
(a)
(b)
(c)
µ̂i (t) − Ci (t) > µi − 2Ci (t) > µ1 > µ̂1 − C1 (t),
where (a), (b), and (c) follow because E2 , E3 , and E1 do not hold respectively.
We note that as we compute 1 − n23 −confidence
intervals at most n times for each point. Thus we
2
2
have at most n computations of 1 − n3 −confidence intervals in total.
6
Thus E1 and E2 do not occur during any iteration with probability 1 − n2 , because
2
) : |µi − µ̂i (t)| ≤ Ci (t), ∀ i ∈ [n], ∀ t.
(3)
n
This also implies that with probability 1 − Θ n1 the algorithm does not stop unless the event E3 , a
deterministic condition stops occurring.
w.p. (1 −
Let ζi be the iteration of the algorithm when it evaluates a distance to point i for the last time. From
the previous discussion, we have that the algorithm stops evaluating distances to points i when the
following holds.
s
2σ 2 log n3
∆i
∆i
Ci (ζi ) ≤
=⇒
≥
or Ci (ζi ) = 0,
2
2
Ti (ζi )
=⇒ Ti (ζi ) ≥
24σ 2
log n or Ti (ζi ) ≥ 2n.
∆2i
Thus with probability (1 − o(1)), the algorithm returns i∗ as the medoid with at most M distance
evaluations, where
X
X 24σ 2
log
n
∧
2n
.
M≤
Ti (ζi ) ≤
∆2i
i∈[n]
i∈[n]
To complete the proof, we next show that [µ̂i (t) − Ci (t), µ̂i (t) − Ci (t)] are true 1 − n23 -confidence
intervals of µi . We observe that if point i is picked by Med-dit less than n times at time t, then
Ti (t) is equal to the number of times the point is picked. Further Ci (t) is the true (1 − δ)-confidence
interval from Eq (1).
However, if point i is picked for the n-th time at iteration t ( line 7 of Algorithm 1) then the empirical
mean is computed by evaluating all n − 1 distances (many distances again). Hence Ti (t) = 2(n − 1).
As we know the mean distance of point i exactly, Ci (t) = 0 is still the true confidence interval.
We remark that if a point i is picked to be updated the (n + 1)-th time, as Ci (t) = 0, we have that
∀j 6= i,
µ̂i (t) + Ci (t) = µ̂i (t) − Ci (t) < µ̂j (t) − Cj (t).
This gives us that the stopping criterion is satisfied and point i is declared as the medoid.
4
EMPIRICAL RESULTS
We empirically evaluate the performance of Med-dit on two real world large-scale high-dimensional
datasets: the Netflix-prize dataset by [3], and 10x single cell RNA-Seq dataset [1].
We picked these large datasets because they have sparse high-dimensional feature vectors, and are at
the throes of active efforts to cluster using non-Euclidean distances. In both these datasets, we use
1000 randomly chosen points to estimate the sub-Gaussianity parameter by cross-validation. We set
δ = 1e-3.
Running trimed and TOPRANK was computationally prohibitive due to the large size and high
dimensionality (∼ 20, 000) of the above two datasets. Therefore, we compare Med-dit only to
RAND. We also ran Med-dit on three real world datasets tested in [24] to compare it to trimed and
TOPRANK.
4.1
Single Cell RNA-Seq
Rapid advances in sequencing technology has enabled us to sequence DNA/RNA at a cell-by-cell
basis and the single cell RNA-Seq datasets can contain up to a million cells [1, 16, 22, 30].
This technology involves sequencing an ensemble of single cells from a tissue, and obtaining the gene
expressions corresponding to each cell. These gene expressions are subsequently used to cluster these
7
Figure 4: We computed the true medoid of the data-set by brute force. The y-axis shows the
probability that the point with the smallest estimated mean distance is not the true medoid as a
function of the number of pulls per arm. We note that Med-dit has a stopping condition while RAND
does not. However we ignore the stopping condition for Med-dit here. Med-dit stops after 80
distance evaluations per point without failing in any of the 1000 trials, while RAND takes around 650
distance evaluations to reach a 2% probability of error.
cells in order to discover subclasses of cells which would have been hard to physically isolate and
study separately. This technology therefore allows biologists to infer the diversity in a given tissue.
We note that tens of thousands gene expressions are measured in each cell, and million of cells are
sequenced in each dataset. Therefore single cell RNA-Seq has presented us with a very important
large-scale clustering problem with high dimensional data [30]. Moreover, the distance metric of
interest here would be `1 distance as the features of each point are probability distribution, Euclidean
(or squared euclidean) distances do not capture the subtleties as discussed in [2, 25]. Further, very
few genes are expressed by most cells and the data is thus sparse.
4.1.1
10xGenomics Mouse Dataset
We test the performance of Med-dit on the single cell RNA-Seq dataset of [1]. This dataset from
10xGenomics consists of 27, 998 genes-expression of 1.3 million neurons cells from the cortex,
hippocampus, and subventricular zone of a mice brain. These are clustered into 60 clusters. Around
93% of the gene expressions in this dataset are zero.
We test Med-dit on subsets of this dataset of two sizes:
• small dataset - 20, 000 cells (randomly chosen): We use this as we can compute the true
medoid on this dataset by brute force and thus can compare the performance of Med-dit
and RAND.
• large dataset - 109, 140 cells (cluster 1) is the largest cluster in this dataset. We use the most
commonly returned point as the true medoid for comparision. We note that this point is the
same for both Med-dit and RAND, and has the smallest distance among the top 100 points
returned in 1000 trials of both.
Performance Evaluation : We compare the performance of RAND and Med-dit in Figures 1 and 4
on the large and the small dataset respectively. We note that in both cases, RAND needs 7-10 times
more distance evaluations to achieve 2% error rate than what Med-dit stops at (without an error in
1000 trials).
8
Figure 5: Left: the 20, 000 user Netflix-prize dataset. We computed the true medoid of the dataset by
brute force. The y-axis shows the probability that the point with the smallest estimated mean distance
is not the true medoid as a function of the number of pulls per arm. We note that Med-dit has a
stopping condition while RAND does not. However we ignore the stopping condition for Med-dit
here. Med-dit stops after 500 distance evaluations per point without failing in any of the 1000
trials, while RAND takes around 4500 distance evaluations to reach a 2% probability of error. Right:
the 100, 000 user Netflix-prize dataset. The y-axis shows the probability that the point with the
smallest estimated mean distance is not the true medoid as a function of the number of pulls per
arm. We note that Med-dit has a stopping condition while RAND does not. However we ignore the
stopping condition for Med-dit here. Med-dit stops after 500 distance evaluations per point with
1% probability of error, while RAND takes around 500 distance evaluations to reach a 2% probability
of error. The computation of the true medoid here is computationally prohibitive and for comparision
we use the most commonly returned point as the true medoid (which was same for both Med-dit
and RAND)
On the 20, 000 cell dataset, we note that Med-dit stopped within 140 distance evaluations per arm
in each of the 1000 trials (with 80 distance evaluations per arm on average) and never returned the
wrong answer. RAND needs around 700 distance evaluations per point to obtain 2% error rate.
On the 109, 140 cell dataset, we note that Med-dit stopped within 140 distance evaluations per arm
in each of the 1000 trials (with 120 distance evaluations per arm on average) and never returned the
wrong answer. RAND needs around 700 distance evaluations per point to obtain 2% error rate.
4.2
Recommendation Systems
With the explosion of e-commerce, recommending products to users has been an important avenue.
Research into this took off with Netflix releasing the Netflix-prize dataset [3].
Usually in the datasets involved here, we have either rating given by users to different products
(movies, books, etc), or the items bought by different users. The task at hand is to recommend
products to a user based on behaviour of similar users. Thus a common approach is to cluster similar
users and to use trends in a cluster to recommend products.
We note that in such datasets, we typically have the behaviour of millions of users and tens of
thousands of items. Thus recommendation systems present us with an important large-scale clustering
problem in high dimensions [7]. Further, most users buy (or rate) very few items on the inventory and
the data is thus sparse. Moreover, the number of items bought (or rated) by users vary significantly
and hence distance metrics that take this into account, like cosine distance and Jaccard distance, are
of interest while clustering as discussed in [19, Chapter 9].
9
Dataset
n, d
TOPRANK trimed med-dit
Europe
160k, 2
176k
2862
3514
Gnutella 6.3k, NA
7043
6328
83
MNIST 6.7k, 784
7472
6514
91
Table 1: This table shows the performance on real-world data-sets picked from [24]. We note that
performance of Med-dit on datasets, where the features are in low dimensions is not as good as that
of trimed, in high dimensions Med-dit performs much better
.
4.2.1
Netflix-prize Dataset
We test the performance of Med-dit on the Netflix-prize dataset of [3]. This dataset from Netflix
consists of ratings of 17, 769 movies by 480, 000 Netflix users. Only 0.21% of the entries in the
matrix are non-zero. As discussed in [19, Chapter 9], cosine distance is a popular metric considered
here.
As in Section 4.1.1, we use a small and a large subset of the dataset of size 20, 000 and a 100, 000
respectively picked at random. For the former we compute the true medoid by brute force, while for
the latter we use the most commonly returned point as the true medoid for comparision. We note that
this point is the same for both Med-dit and RAND, and has the smallest distance among the top 100
points returned in 1000 trials of both.
Performance Evaluation:
On the 20, 000 user dataset (Figure 5 left), we note that Med-dit stopped within 600 distance
evaluations per arm in each of the 1000 trials (with 500 distance evaluations per arm on average) and
never returned the wrong answer. RAND needs around 4500 distance evaluations per point to obtain
2% error rate.
On the 100, 000 users on Netflix-Prize dataset (Figure 5 right), we note that Med-dit stopped within
150 distance evaluations per arm in each of the 1000 trials (with 120 distance evaluations per arm on
average) and returned the wrong answer only once. RAND needs around 500 distance evaluations
per point to obtain 2% error rate.
4.3
Performance on Previous Datasets
We show the performance of Med-dit on data-sets — Europe by [11] , Gnutella by [27] , MNIST
by [18]— used in prior works in Table 1. The number of distance evaluations for TOPRANK and
trimed are taken from [24]. We note that the performance of Med-dit on real world datasets which
have low-dimensional feature vectors (like Europe) is worse than that of trimed, while it is better on
datasets (like MNIST) with high-dimensional feature vectors.
We conjecture that this is due to the fact the distances in high dimensions are average of simple
functions (over a each dimension) and hence have Gaussian-like behaviour due to the Central Limit
Theorem (and the delta method [6]).
4.4
Inner workings of Med-dit
Empirical running time: We computed the true mean distance µi , and variance σ, for all the points
in the small dataset by brute force. Figure 3 shows the histrogram of µi ’s. Using δ = 1e-3 in Theorem
1, the theoretical average number of distance evaluations is 266. Empirically over 1000 experiments,
the average number of distance computations is 80. This suggests that the theoretic analysis captures
the essence of the algorithm.
Progress of Med-dit : We see that at any iteration of Med-dit, only the points whose lower
confidence interval are smaller than the smallest upper confidence interval among all points have their
distances evaluated. At any iteration, we say that such points are under consideration. We note that a
point that goes out of consideration can be under consideration at a later iteration (this happens if the
smallest upper confidence interval increases). We illustrate the fraction of points under consideration
at different iterations (indexed by the number of distance evaluations per point) in the left panel of
Figure 6. We pick 10 snapshots of the algorithm and illustrate the true distance distributions of all
10
1
1
2
3
2
4
5
3
6
7
4
8
5
9
6,7,8
9
10
10
Figure 6: In the left panel, we show the number of points under consideration at different iterations
in Med-dit. We note that though the number of points under consideration show a decreasing trend,
they do not decrease monotonically. In the right panel, we show the distribution of the true distances
among the arms that are under consideration at various snapshots of the algorithm. We note that both
the mean and the variance of the distributions keeps decreasing. The last snapshot shows only the
distance of the point declared a medoid.
Figure 7: 99.99%-Confidence Intervals of the 150 points with smallest estimated mean distance in
the 20, 000 cell dataset at the termination of Med-dit. The true means are shown by the orange line,
while the estimated means are shown by the pink line. We note that the distances of 9 points are
computed to all other points. We also note that mean of the medoid is less than the 99.99%-lower
confidence bound of all the other points.
points under consideration at that epoch in the right panel of Figure 6. We note that both the mean
and the standard deviation of these distributions decrease as the algorithm progresses.3
Confidence intervals at stopping time : Med-dit seems to compute the mean distance of a few
points exactly. These are usually points with the mean distance close to the that of the medoid. The
points whose mean distances are farther away from that of the medoid have fewer distances evaluated
to have the confidence intervals rule them out as the medoid. The 99.99% confidence interval of the
top 150 points in a run of the experiment are shown in Figure 7 for the 20, 000 cell RNA-Seq dataset.
This allows the algorithm to save on distance computations while returning the correct answer.
Speed-accuracy tradeoff Setting the value of δ between 0.01 to 0.1 will result in smaller confidence
intervals around the estimates µ̂i , which will ergo reduce the number of distance evaluations. One
could also run Med-dit for a fixed number of iterations and return the point with the smallest mean
estimate. Both of these methods improve the running time at the cost of accuracy. Practically, the
first method has a better trade-off between accuracy and running time whereas the second method
has a deterministic stopping time.
3
When the number of points under consideration is more than 2000, we draw the distribution using the true
distances of 200 random samples.
11
References
[1] 10xGenomics. 1.3 Million Brain Cells from E18 Mice. 10x Genomics, 2017. available
at https://support.10xgenomics.com/single-cell-gene-expression/datasets/
1M_neurons.
[2] Tugkan Batu, Lance Fortnow, Ronitt Rubinfeld, Warren D Smith, and Patrick White. Testing
that distributions are close. In Foundations of Computer Science, 2000. Proceedings. 41st
Annual Symposium on, pages 259–269. IEEE, 2000.
[3] James Bennett, Stan Lanning, et al. The netflix prize. In Proceedings of KDD cup and workshop,
volume 2007, page 35. New York, NY, USA, 2007.
[4] Sébastien Bubeck, Nicolo Cesa-Bianchi, and Gábor Lugosi. Bandits with heavy tail. IEEE
Transactions on Information Theory, 59(11):7711–7717, 2013.
[5] Olivier Catoni et al. Challenging the empirical mean and empirical variance: a deviation
study. In Annales de l’Institut Henri Poincaré, Probabilités et Statistiques, volume 48, pages
1148–1185. Institut Henri Poincaré, 2012.
[6] Aad W Chapter 3 Van der Vaart. Asymptotic statistics, volume 3. Cambridge university press,
1998.
[7] Srivatsava Daruru, Nena M Marin, Matt Walker, and Joydeep Ghosh. Pervasive parallelism in
data mining: dataflow solution to co-clustering large and sparse netflix data. In Proceedings
of the 15th ACM SIGKDD international conference on Knowledge discovery and data mining,
pages 1115–1124. ACM, 2009.
[8] David Eppstein and Joseph Wang. Fast approximation of centrality. Graph Algorithms and
Applications 5, 5:39, 2006.
[9] Eyal Even-Dar, Shie Mannor, and Yishay Mansour. Pac bounds for multi-armed bandit and
markov decision processes. In International Conference on Computational Learning Theory,
pages 255–270. Springer, 2002.
[10] Eyal Even-Dar, Shie Mannor, and Yishay Mansour. Action elimination and stopping conditions
for the multi-armed bandit and reinforcement learning problems. Journal of machine learning
research, 7(Jun):1079–1105, 2006.
[11] P. Fränti, M. Rezaei, and Q. Zhao. Centroid index: cluster level similarity measure. Pattern
Recognition, 47(9):3034–3045, 2014.
[12] Charles AR Hoare. Algorithm 65: find. Communications of the ACM, 4(7):321–322, 1961.
[13] Kevin Jamieson and Robert Nowak. Best-arm identification algorithms for multi-armed bandits
in the fixed confidence setting. In Information Sciences and Systems (CISS), 2014 48th Annual
Conference on, pages 1–6. IEEE, 2014.
[14] Zohar Karnin, Tomer Koren, and Oren Somekh. Almost optimal exploration in multi-armed
bandits. In Proceedings of the 30th International Conference on Machine Learning (ICML-13),
pages 1238–1246, 2013.
[15] Leonard Kaufman and Peter Rousseeuw. Clustering by means of medoids. North-Holland,
1987.
[16] Allon M Klein, Linas Mazutis, Ilke Akartuna, Naren Tallapragada, Adrian Veres, Victor Li,
Leonid Peshkin, David A Weitz, and Marc W Kirschner. Droplet barcoding for single-cell
transcriptomics applied to embryonic stem cells. Cell, 161(5):1187–1201, 2015.
[17] Tze Leung Lai and Herbert Robbins. Asymptotically efficient adaptive allocation rules. Advances in applied mathematics, 6(1):4–22, 1985.
[18] Yann LeCun. The mnist database of handwritten digits. http://yann. lecun. com/exdb/mnist/,
1998.
[19] Jure Leskovec, Anand Rajaraman, and Jeffrey David Ullman. Mining of massive datasets.
Cambridge university press, 2014.
[20] George C Linderman and Stefan Steinerberger. Clustering with t-sne, provably. arXiv preprint
arXiv:1706.02582, 2017.
[21] Stuart Lloyd. Least squares quantization in pcm. IEEE transactions on information theory,
28(2):129–137, 1982.
[22] Evan Z Macosko, Anindita Basu, Rahul Satija, James Nemesh, Karthik Shekhar, Melissa
Goldman, Itay Tirosh, Allison R Bialas, Nolan Kamitaki, Emily M Martersteck, et al. Highly
parallel genome-wide expression profiling of individual cells using nanoliter droplets. Cell,
161(5):1202–1214, 2015.
12
[23] James MacQueen et al. Some methods for classification and analysis of multivariate observations.
In Proceedings of the fifth Berkeley symposium on mathematical statistics and probability,
volume 1, pages 281–297. Oakland, CA, USA., 1967.
[24] James Newling and François Fleuret. A sub-quadratic exact medoid algorithm. Proceedings of
the 20th International Conference on Artificial Intelligence and Statistics, 2017.
[25] Vasilis Ntranos, Govinda M Kamath, Jesse M Zhang, Lior Pachter, and David N Tse. Fast and
accurate single-cell rna-seq analysis by clustering of transcript-compatibility counts. Genome
biology, 17(1):112, 2016.
[26] Kazuya Okamoto, Wei Chen, and Xiang-Yang Li. Ranking of closeness centrality for largescale social networks. In International Workshop on Frontiers in Algorithmics, pages 186–195.
Springer, 2008.
[27] Matei Ripeanu, Ian Foster, and Adriana Iamnitchi. Mapping the gnutella network: Properties of
large-scale peer-to-peer systems and implications for system design. arXiv preprint cs/0209028,
2002.
[28] Hugo Steinhaus. Sur la division des corp materiels en parties. Bull. Acad. Polon. Sci, 1(804):801,
1956.
[29] Cathie Sudlow, John Gallacher, Naomi Allen, Valerie Beral, Paul Burton, John Danesh, Paul
Downey, Paul Elliott, Jane Green, Martin Landray, et al. Uk biobank: an open access resource
for identifying the causes of a wide range of complex diseases of middle and old age. PLoS
medicine, 12(3):e1001779, 2015.
[30] Grace XY Zheng, Jessica M Terry, Phillip Belgrader, Paul Ryvkin, Zachary W Bent, Ryan
Wilson, Solongo B Ziraldo, Tobias D Wheeler, Geoff P McDermott, Junjie Zhu, et al. Massively
parallel digital transcriptional profiling of single cells. Nature communications, 8:14049, 2017.
13
Appendices
1
The O(n log n) Distance Evaluations Under Gaussian Prior
We assume that the mean distances of each point µi are i.i.d. samples of N (γ, 1). We note that this
implies that ∆i , 1 ≤ i ≤ n are n iid random variables. Let ∆ be a random variable with the same
law as ∆i .
From the concentration of the minimum of n gaussians, we have that
p
p
min µi + 2 log n → γ.
i
This gives us that
∆−
p
d
2 log n → N (0, 1).
We note that by Eq (2), we have that the expected number of distance evaluations M is of the order of
log n
E[M ] ≤ nE
∧
n
,
∆2
where the expection is taken with respect to the randomness of ∆.
To show this, it is enough to show that,
log n
∧ n ≤ C log n,
E
∆2
for some constant C.
To compute that for this prior, we divide the real line into three intervals, namely
q
log n
• −∞,
,
n
•
q
log n √
n , c log n
,
√
• c log n, ∞ ,
and compute the expectation on these three ranges. We note that for if ∆ ∈ (−∞,
q
n
∆ ∈ ( logn n , ∞], log
∆2 ≤ n. Thus we have that,
I
q
log n
n ],
while for
II
z"
}|r
!#{ z
"
!#{
r }|
√
log n
log n
log n
log n
log n
∧ n ≤ E nI ∆ ≤
+ lim E
I
E
≤∆≤c
δ,→0
∆2
n
∆2
nδ
n(1−)
p
log n
+E
log
n
,
I
∆
≥
c
∆2
{z
}
|
III
where we use the Bounded Convergence Theorem to establish II.
We next show that all three terms in the above equation are of O(log n). We first show the easy cases
of I and III and then proceed to III.
√
• To establish that I is O(log n), we start by defining, q1 = P [∆ <
√
log n
E nI ∆ ≤
= nq1 .
n
14
log n
n ]
Further note that,
q1 ≤ exp −
1
2
r
p
2 log n −
log n
n
!2
,
(1− √1 )2
2n
1
=
.
n
Thus
n
q1 n
≤
log n
1−(1− √12n )2
,
log n
r
2
≤ exp((
(1 + o(1)) log n − log log n),
n
= o(1).
• To establish that III is O(log n), we note that,
p
p
1
log n
I
∆
≥
c
log
n
≤ 2 P (∆ ≥ c log n),
E
2
∆
c
1
≤ 2,
c
= Θ(1).
• Finally to establish that II is O(log n), we note that,
"
!#
r
√
√
log n
log n
log n
log n
2−
lim E
I
≤∆≤c
≤ lim n P ∆ ≤ c
,
δ,→0
δ,→0
∆2
nδ
nδ
n(1−)
p
1 p
c 2
≤ lim n1− exp −
2 log n − log n δ
,
δ,→0
2
n
(1− √ c δ )2
2n
1
1−
= lim n
,
δ,→0
n
= lim n
1−−(1− √ c
2nδ
)2
δ,→0
= lim n−+
δ,→0
√
2
2c
− c 2δ
nδ
2n
,
.
Letting δ to go to 0 faster than , we see that,
"
!#
r
√
log n
log n
log n
lim E
I
≤∆≤c
≤ O(log n).
δ,→0
∆2
nδ
n(1−)
This gives us that under this model, we have that under this model,
E[M ] ≤ O(n log n).
15
2
Extensions to the Theoretical Results
In this section we discuss two extensions to the theoretical results presented in the main article:
1. relax the sub-Gaussian assumption used for the analysis of Med-dit. We see that assuming
that the distances random variables have finite variances is enough for a variant of Med-dit
to need essentially the same number of distance evaluations as Theorem 1. The variant
would use a different estimator (called Catoni’s M-estimator [5]) for the mean distance of
each point instead of empirical average leveraging the work of [4]. This is based on the
work of [4].
2. note that there exists an algorithm which can compute the medoid with O(n log log n) distance evaluations. The algorithm called exponential-gap algorithm [14] and is discussed
below.
2.1
Weakening the Sub-Gaussian Assumption
We note that in order to have the O(n log n) sample complexity, Med-dit relies on a concentration
bound where the tail probability decays exponentially. This is achieved by assuming that for each
point xi , the random variable of sampling with replacement from Di is σ-sub-Gaussian in Theorem 1.
As a result, we can have the sub-Gaussian tail bound that for any point xi at time t, with probability
at least 1 − δ, the empirical mean µ̂i satisfies
s
2σ 2 log 2δ
.
|µi − µ̂i | ≤
Ti (t)
In fact, as pointed out by [4], to achieve the O(n log n) sample complexity, all we need is a performance guarantee like the one shown above for the empirical mean. To be more precise, we need the
following property:
Assumption 1. [4] Let ∈ (0, 1] be a positive parameter and let c, v be positive constants. Let
X1 , · · · , XT be i.i.d. random variables with finite mean µ. Suppose that for all δ ∈ (0, 1), there
exists an estimator µ̂ = µ̂(T, δ) such that, with probability at least 1 − δ,
c log 2δ 1+
1
.
|µ − µ̂| ≤ v 1+
T
Remark 2. If the distribution of Xj satisfies σ-sub-Gaussian condition, then Assumption 1 is satisfied
for = 1, c = 2, and variance factor v = σ 2 .
However, Assumption 1 can be satisfied with conditions much weaker than the sub-Gaussian condition.
One way is by substituing the empirical mean estimator by some refined mean estimator that gives
the exponential tail bound. Specifically, as suggested by [4], we can use Catoni’s M-estimator [5].
Catoni’s M-estimator is defined as follows: let ψ : R → R be a continuous strictly increasing function
satisfying
− log(1 − x +
x2
x2
) ≤ ψ(x) ≤ log(1 + x + ).
2
2
Let δ ∈ (0, 1) be such that T > 2 log( 1δ ) and introduce
v
u
2 log 1δ
u
αδ = t
.
2σ 2 log 1
T (σ 2 + T −2 log δ1 )
δ
If X1 , · · · , XT are i.i.d. random variables, the Catoni’s estimator is defined as the unique value
µ̂C = µ̂C (T, δ) such that
n
X
ψ(αδ (Xi − µ̂C )) = 0.
i=1
16
Catoni [5] proved that if T ≥ 4 log
probability at least 1 − δ,
1
δ
and the Xj have mean µ and variance at most σ 2 , then with
s
|µ̂C − µ| ≤ 2
σ 2 log 2δ
.
T
(4)
The corresponding modification to Med-dit is as follows.
1. For the initialization step, sample each point 4 log 1δ times to meet the condition for the
concentration bound of the Catoni’s M-estimator.
2. For each arm i, if Ti (t) < n, maintain the 1 − δ confidence interval [µ̂C,i − Ci (t), µ̂C,i +
Ci (t)], where µ̂C,i is the Catoni’s estimator of µi , and
s
σ 2 log 2δ
Ci (t) = 2
.
Ti (t)
Proposition 1. For i ∈ [n], let ∆i = µi − µ∗ . If we pick δ = n13 in the above algorithm, then with
probability 1 − o(1), it returns the true medoid with the with number of distance evaluations M such
that,
X 48σ 2
M ≤ 12n log n +
log n ∧ 2n
∆2i
i∈[n]
Proof. Let δ = n23 . The initialization step takes an extra 12n log n distance computations. Following
the same proof as Theorem 1, we can show that the modified algorithm returns the true medoid
with probability at least 1 − Θ( n1 ), and apart from the initialization, the total number of distance
computations can be upper bounded by
X 48σ 2
log
n
∧
2n
.
∆2i
i∈[n]
So the total number of distance computations can be upper bounded by
X 48σ 2
M ≤ 12n log n +
log n ∧ 2n .
∆2i
i∈[n]
Remark 3. By using the Catoni’s estimator, instead of the empirical mean, we need a much weaker
assumption that the distance evaluations have finite variance to get the same results as Theorem 1.
2.2
On the O(n log log n) Algorithm
The best-arm algrithm
exponential-gap [14] can be directly applied on the medoid problem,
P
−2
which takes O( i6=i∗ ∆−2
i log log ∆i ) distance evaluations, essentially O(n log log n) if ∆i are
constants. It is an variation of the familiy of action elimination algorithm for the best-arm problem.
A typical action elimination algorithm proceeds as follows: Maintaining a set Ωk for k = 1, 2, · · · ,
initialized as Ω1 = [n]. Then it proceeds in epoches by sampling the arms in Ωk a predetermined
number of times rk , and maintains arms according to the rule:
Ωk+1 = {i ∈ Ωk : µ̂a + Ca (t) < µ̂i − Ci (t)},
where a ∈ Ωk is a reference arm, e.g. the arm with the smallest µ̂i + Ci (t). Then the algorithm
terminates when Ωk contains only one element.
The above vanilla version of the action elimination algorithm takes O(n log n) distance evaluations, same as Med-dit. The improvement by exponential-gap is by observing that the suboptimal log n factor is due to the large deviations of |µ̂a − µa | with a = arg mini∈Ωk µ̂i . Instead,
exponential-gap use a subroutine median elimination [10] to determine an alternative reference arm a with smaller deviations and allows for the removal of the log n term, where median
17
elimination takes O( n2 log 1δ ) distance evaluations to return a -optimal arm. However, this will
introduce a prohibitively large constant due to the use of median elimination. Regarding the
technical details, we note both paper [14, 10] assume the boundedness of the random variables for
their proof, which is only used to have the hoefflding concentration bound. Therefore, with our
sub-Gaussian assumption, the proof will follow symbol by symbol, line by line.
18
| 7cs.IT
|
A novel normalized sign algorithm for system
identification under impulsive noise interference
Lu Lu • Haiquan Zhao • Kan Li • Badong Chen
Abstract. To overcome the performance degradation of adaptive filtering algorithms in the presence of impulsive noise, a novel normalized sign algorithm
(NSA) based on a convex combination strategy, called NSA-NSA, is proposed
in this paper. The proposed algorithm is capable of solving the conflicting
requirement of fast convergence rate and low steady-state error for an
individual NSA filter. To further improve the robustness to impulsive noises, a
mixing parameter updating formula based on a sign cost function is derived.
Moreover, a tracking weight transfer scheme of coefficients from a fast NSA
filter to a slow NSA filter is proposed to speed up the convergence rate. The
convergence behavior and performance of the new algorithm are verified by
theoretical analysis and simulation studies.
Keywords. Adaptive filtering • Convex combination • Normalized sign
algorithm • System identification • Impulsive noise
1
Introduction
In general, the performance of an adaptive filtering algorithm degrades when signals
are contaminated by impulsive or heavy-tailed noise. To overcome this limitation,
many algorithms were proposed, such as the sign algorithm (SA) [26], the signed
regressor algorithm (SRA) [7] and the sign-sign algorithm (SSA) [9]. Although the
SA has been successfully applied to system identification under impulsive noise, its
convergence rate is slow [26]. As a variant of SA, the convergence behavior of the
SRA is heavily dependent on the inputs, and there may exist some inputs for which
the SRA is unstable while the least mean square (LMS) algorithm is stable [7].
Among the family of SA algorithms, the SSA has the lowest computational complexiL. Lu • H. Zhao()
Key Laboratory of Magnetic Suspension Technology and Maglev Vehicle, Ministry of Education, and
School of Electrical Engineering, Southwest Jiaotong University, Chengdu, China
e-mail: [email protected], [email protected]
Kan Li
Computational Neuro-Engineering Laboratory, University of Florida, Gainesville, FL 32611, USA
email: likan@ufl.edu.
B. Chen
School of Electronic and Information Engineering, Xi’an Jiaotong University, Xi’an, China.
e-mail: [email protected]
ty and the most similar characteristic to SA [9]. In addition, the degradations of two
algorithms depend significantly on the initial weights. Similar to the normalized least
mean square (NLMS), the normalized versions of these sign algorithms can be easily
derived, including the normalized SA (NSA) [10], the normalized SRA (NSRA) [11]
and the normalized SSA (NSSA) [12]. The NSA can improve the robustness of the
filter against impulsive noises. However, its convergence performance is still not
good in general. Several variants have been proposed aiming at improving the convergence [5-6,8,13-14,16,27,30,32,]. Particularly, in [14], a dual SA (DSA) with a
variable step-size (VSS) was proposed, but it has a local divergence problem especially when a large disparity occurs between two successive step sizes. In [8], attempt
was made to obtain better stability and convergence performance by inserting another
step-size. Note that the above-mentioned efforts have all been made for a single adaptive filtering architecture.
On the other hand, to cope with impulsive noise, the family of mixed-norm algorithms were developed to combine the benefits of stochastic gradient adaptive filter
algorithms [3-4,17,24,31]. Chambers et al. introduced a robust mixed-norm (RMN)
algorithm, where the cost function is a combination of the error norms that underlie
the LMS and SA [4]. Later, Papoulis et al. [17,23] proposed a novel VSS RMN
(NRMN) algorithm, which circumvents the drawback of slow convergence for RMN
to some extent, by using time-varying learning rate..
The convex combination approach is another way to effectively balance the convergence rate and steady-state error. An adaptive approach using combination LMS
(CLMS) was proposed in [1], utilizing two LMS filters with different step sizes to
obtain fast convergence and small misadjustment. Nevertheless, when the signals are
corrupted by impulsive noise, the algorithms in [1] and [15] usually fail to converge.
To improve performance, an NLMS-NSA algorithm was developed where a combination scheme was used to switch between the NLMS and NSA algorithms [2]. Regrettably, in the initial stage of adaptation, the NLMS algorithm may cause large misadjustment especially when the noise becomes severe. Moreover, the adaptation rule of
the mixing parameter of NLMS-NSA is unsuitable for impulsive noise, such that the
algorithm fails to perform at a desirable level.
In this work, to address the above-mentioned problems, a NSA-NSA algorithm is
proposed by using the convex combination approach. This novel algorithm achieves
robust performance in impulsive noise environments by leveraging two independent
NSA filters with a large and a small step-sizes, respectively. To further enhance the
robustness against impulsive noise, the mixing parameter is adjusted using a sign cost
function. In addition, a tracking weight transfer of coefficients is proposed in order to
obtain fast convergence speed during a transition period. Our main contributions are
listed as follows: 1) propose a NSA-NSA that is well-suited for system identification
problems under impulsive noise; 2) modify an existing update scheme of the mixing
parameter, and analyze its behavior; 3) propose a novel weight transfer scheme that is
computationally simple yet can significantly improve the convergence rate.
The rest of this paper is organized as follows. In Section 2, we propose the NSANSA and develop a novel weight transfer scheme. In Section 3, simulation results in
different impulsive noise environments are presented. Lastly, Section 4 concludes the
paper.
2
Adaptive combination of NSA algorithms
2.1
The proposed algorithm
v (n)
w0
unknown system
e1 (n)
d (n)
w1 ( n) y ( n )
1
x(n)
w 2 ( n)
e( n )
( n)
y2 ( n )
e2 (n)
y (n)
1 ( n)
Fig. 1. Diagram of the proposed algorithm.
The diagram of adaptive combination scheme of two NSA filters is illustrated in Fig.
1, where x(n) and y (n) are the filter input and output signals respectively, d (n) is
the desired signal, y1 (n) and y2 (n) are symbols of the two component filters defined
by weight vectors w1 and w 2 , respectively, v(n) is the impulsive noise, and w 0 is
the weight vector of the unknown system. The overall error of the combined filter is
given by e(n) d (n) y (n) . To improve performance, both filters are combined with
a scalar mixing parameter (n) :
(1)
y (n) (n) y1 (n) [1 (n)] y2 (n)
(2)
e(n) (n)e1 (n) [1 (n)]e2 (n)
where (n) [0,1] is defined by a sigmoidal activation function with auxiliary parameter a (n)
(n) 1 (1 e a ( n ) ) .
(3)
A gradient descent adaptation of a (n) is given as
a (n 1) a(n)
a e2 (n)
2 a(n)
a(n) a e(n)[ y1 (n) y2 (n)] (n)[1 (n)].
(4)
Note that a is the step-size of the auxiliary parameter a (n) . This adaptation rule is
derived by the cost function J (n) e(n) 2 [1]. To improve the robustness against
impulsive noise, the new cost function is defined as J s (n) | e(n) | based on the classical sign-error LMS algorithm [26]. Therefore, the updated scheme of a (n) is derived by minimizing the cost function J s (n) as follows:
a (n 1) a(n)
a J s (n)
2 a (n)
(5)
where a is the step-size.
Using the chain rule, the gradient adaptation of J s (n) can be calculated as follows:
a J s (n) (n)
2 (n) a (n)
a(n) a sign{e(n)}[ y1 (n) y2 (n)] (n)[1 (n)]
a (n 1) a(n)
(6)
where a is a positive constant, and the sign function sign() can be expressed as
if x 0
1,
x
sign( x)
0, if x 0 .
(7)
|| x ||2
1, if x 0
At each iteration cycle, the weight update of NSA-NSA takes the form [10]
x(n) sign{ei (n)}
w i (n 1) w i (n) i
(i 1, 2)
(8)
i || x(n) ||22
where w i (n) is the weight vectors with length M, i is the constant step-size, i 0
is a regularization constant close to zero, and || ||2 represents the Euclidian-norm. As
a result, the combined filter is obtained by using the following convex combination
scheme
(9)
w (n) (n)w1 (n) [1 (n)]w 2 (n) .
2.2
Proposed weight transfer scheme
Inspired by the instantaneous transfer scheme from [22], a tracking weight transfer
scheme is proposed, as shown in Table 1. By using a sliding window approach, the
proposed scheme involves few parameters and retains the robustness against impulsive noise with low-cost. Like the instantaneous transfer scheme in [22], the parameter of proposed weight transfer scheme is not sensitive to the choice. This scheme can
speed up the convergence property of the overall filter, especially during the period of
convergence transition. Define N 0 as the window length. If n−1 mod N 0 is equal to
zero, then implement the following operations. It is well known that the standard convex combination scheme needs to check if a(n 1) a , so the only additional operation is the n mod N 0 operation. The judgment condition a (n 1) a represents the
condition when the fast filter (filter with large step-size) switches to the slow filter
(filter with small step size) at the transient stage. The operations (n 1) 0 and
(n 1) 1 are the limitations for a (n 1) a and a(n 1) a , respectively. The
operation w 2 (n 1) w1 (n 1) denotes the transfer of coefficients, which is only
applied in the transient stage. By applying the weight transfer, the adaptation of
w 2 (n 1) is similar to that of the fast filter, which speeds up the convergence rate of
2 NSA filter. Moreover, the cost of the proposed weight transfer scheme is smaller
than that of the original combination, because only one filter is adapted.
15
Fast filter
Slow filter(weight transfer)
Slow filter(no transfer)
Comb.,no transfer
Comb.,weight transfer
10
5
EMSE(dB)
0
-5
-10
-15
-20
-25
-30
-35
0
0.2
0.4
0.6
0.8
1
1.2
1.4
1.6
1.8
2
4
n
x 10
Fig. 2. Comparison of EMSE of NSA-NSA for Gaussian input in example 1.
15
Fast filter
Slow filter(weight transfer)
Slow filter(no transfer)
Comb.,no transfer
Comb.,weight transfer
10
EMSE(dB)
5
0
-5
-10
-15
-20
-25
-30
0
0.2
0.4
0.6
0.8
1
n
1.2
1.4
1.6
1.8
2
4
x 10
Fig. 3. Comparison of EMSE of NSA-NSA for Gaussian input in example 2.
Figs. 2 and 3 display the comparison of excess means-square error (EMSE) obtained from NSA-NSA with the tracking weight transfer scheme and no transfer
scheme (the mixing parameter is adjusted according to (6)). The same step-size is
chosen for this comparison. As can be seen, the overall performance of the filter bank
is improved by the transfer scheme. It shows from these figures that the proposed
weight transfer scheme exhibit faster convergence than no transfer scheme. The proposed algorithm is summarized in Table 1.
Table 1. The Proposed algorithm
Initialize N 0 , a , a , i , i , w i (0) 0, a (0) 0, (0) 0.5
Loop n=1→end do
yi (n) wTi (n)x(n) (i 1,2)
ei (n) d (n) yi (n)
y (n) (n) y1 (n) [1 (n)] y2 (n)
w1 ( n 1) w1 ( n) 1
sign{e1 ( n)}x( n)
1 || x( n) ||22
w 2 ( n 1) w 2 ( n) 2
sign{e2 ( n)}x( n)
2 || x( n) ||22
a( n 1) a (n) a sign{e(n)}[( y1 ( n) y2 (n)] (n)[1 (n)]
(n 1) 1 / (1 e a ( n 1) )
% Tracking weight transfer scheme (the proposed method)
if (mod( n 1, N 0 )equal to zero)
if a (n 1) a
a (n 1) a
(n 1) 0
endif
if a (n 1) a
a(n 1) a
(n 1) 1
w 2 (n 1) w1 (n 1)
endif
endif
Let n=n+1
end
2.3
Computational complexity
The computational complexity of the basic CLMS [1], NLMS-NSA [2] and NSANSA algorithms is listed in Table 2. Since the basic CLMS combines two LMS algorithms, it requires 4M+2 multiplications for the adaptation of the component filters.
The NLMS-NSA algorithm provides additional insensitivity to the input signal level
by combining the NLMS and NSA, it requires 6M+1 multiplications for the adaptation of the component filters. In contrast with the CLMS, the proposed algorithm uses
NSA as the fast filter to replace the NLMS filter, which reduces the computational
burden and the negative effect of impulsive noise. From (1) and (4), the basic-CLMS
and NLMS-NSA algorithms need 6 multiplications to compute the filter output and to
update a (n) . However, the proposed algorithm requires 5 multiplications to update
a (n) (see (1) and (6), respectively). According to (9), all the algorithms demand 2M
multiplications to calculate the explicit weight vector. Moreover, due to using the
slide window of tracking weight transfer scheme, the NSA-NSA can further reduce
the computation operations. Consequently, these would lead to significant computational efficiency.
Table 2. Summary of the computational complexity.
Algorithms
Basic-CLMS
[1]
NLMS-NSA
[2]
NSA-NSA
2.4
Component filter
adaptation
4M+2
Basic combination
6
Explicit weight
calculation
2M
Weight
transfer
2M
6M+1
6
2M
M+3
2M
No
6M
5(using (6))
The analysis of the mixing parameter
In this section, the convergence behavior of the mixing parameter is analyzed, and the
range of a will be discussed. When the error term e(n) is expanded with a Taylor
series [18-20], we have
e(n)
1 2 e( n ) 2
e(n 1) e(n)
a (n)
a (n) h.o.t
(10)
a(n)
2 2 a ( n)
where h.o.t. represents the higher order terms of the remainder of the Taylor series
e(n)
expansion. According to e(n) d (n) y (n) and (3),
can be obtained as fola(n)
lows:
e(n)
(n)[1 (n)][ y2 (n) y1 (n)] .
(11)
a(n)
The mixing parameter correction a(n) can be calculated from (6)
a(n) a sign(e(n))[( y1 (n) y2 (n))] (n) [1 (n)] .
Combining (10), (11) and (12), we can express (10) as
( y (n) y2 (n))2
e(n 1) e(n)[1 a 1
(n)2 (1 (n))2 ] .
| e( n ) |
The NSA-NSA can converge if
(12)
(13)
| e(n 1) || e(n) ||1
a ( y1 (n) y2 (n))2
| e( n ) |
(n)2 (1 (n))2 | . (14)
Hence
|1
a ( y1 (n) y2 (n)) 2
| e( n ) |
Solving the inequality with respect to a gives
0 a
2.5
(n) 2 (1 (n))2 | 1 .
2 | e(n) |
.
( y1 (n) y2 (n))2 (n)2 (1 (n))]2
(15)
(16)
Steady-state performance of the proposed algorithm
To measure the steady-state performance, the EMSEs of the filters are expressed as
[1]
(17)
J ex ,i ( ) lim E{ea2,i ( n )}, i 1, 2
n
J ex ( ) lim E{ea2 ( n )}
n
J ex,12 () lim E{ea ,1 (n)ea ,2 (n)}
n
(18)
(19)
denotes the expectation, J ex ,i () represents the individual EMSE of the
where E{}
ith filter, J ex () is the cross-EMSE of the combined filters, J ex ,12 ( ) is the steadystate correlation between the a priori errors of the elements of the combination,
ea ,i (n) and ea ( n) are a priori error, respectively, defined by
ea ,i (n) [w 0 wi (n)]T x(n) iT (n)x(n)
T
T
ea (n) [w 0 (n) w (n)] x(n) (n)x(n)
(20)
(21)
where i ( n) is the weight error vector of the individual filter, and (n) is the weight
error vector of the overall filter.
Additionally, for the modified combination (2), J ex,u () is defined as
J ex ,u ( ) lim E{u2 ( n )ea2,1 ( n ) [1 u ( n )]2 ea2,2 ( n ) 2u ( n )[1 u ( n)]ea ,1 ( n )ea ,2 ( n )}
n
(22)
where
1
u (n) (n)
0
a ( n) a
a a ( n) a
(23)
a ( n) a
and is a small positive constant.
Taking expectations of both sides of (6) and using y1 (n) y2 (n) ea ,2 (n) ea ,1 (n)
yields：
E{a(n 1)} E{a(n)} a E{sign(e(n))[ea ,2 (n) ea ,1 (n)] (n)[1 (n)]} .
According to the Price theorem [21,25], we have
(24)
E{sign[e( n)] ( n)}
2 1
e, n
E{e( n) ( n)}
(25)
where e , n is the standard deviation of the error e(n) , i.e., e2, n E{e2 (n)} , and
(n) can be defined as (n) [ea ,2 (n) ea ,1 (n)] (n)[1 (n)] . Therefore, (24) becomes
E{a (n 1)} E{a(n)} a E{e(n)[ea ,2 (n) ea ,1 (n)] (n)[1 (n)])
where a a
2 1
e, n
(26)
. Then, (26) can be rewritten as：
E{a (n 1)} [ E{a (n)} a E{[ea2,2 (n) ea ,1 (n)ea ,2 (n)] (n)[1 (n)]2 )
a E{[ea ,1 ( n)ea ,2 (n) ea2,1 (n)] 2 (n)[1 (n)])]a a
.
(27)
Assume (n) is independent of a prior error ea ,i (n) in the steady state, under this
assumption, E{a(n 1)} is governed by
E{a (n 1)} [ E{a(n)} a E{ (n)[1 (n)]2 J 2 )
a E{ 2 (n)[1 (n)]J1 ]a a
(28)
where J i J ex ,i () J ex,12 (), i 1, 2 . Suppose the NSA-NSA convergences, the
optimal mean combination weights under convex constraint are given by [1], which is
discussed in the three situations as follows:
1) If J ex,1 () J ex ,12 () J ex ,2 () , we have J1 0 and J 2 0 . Since a (n) and
(n) are limited in the effective range, an assumption can be expressed as
E{a(n 1)} [ E{a (n)} C ]a a as n
(29)
where C (1- )(J 2 -J1 ) is a positive constant. In this case, we can conclude
that
J ex ( ) J ex ,1 ( )
.
(30)
J ex ,u () J ex ,1 ()
Therefore, (30) shows that the NSA-NSA algorithm performs as well as the component filters.
2) If J ex,1 () J ex ,12 () J ex ,2 () , we have J1 0 and J 2 0 . Then, (28) can
+
+ 2
be rewritten as
E{a(n 1)} [ E{a (n)} C ]a a as n
(31)
for a positive constant C (1- )(J1 -J 2 ) and
+
+ 2
J ex ( ) J ex ,2 ( )
.
(32)
J ex ,u () J ex ,2 ()
From (32), the overall filter performs approximately equal to the better component
filter.
3) If J ex,12 () J ex ,i ( ), i 1, 2 , we have J1 0 and J 2 0 .
Assume (n) 0 when n , we obtain
[1 ()]J 2 ()J1
(33)
where () is given by
() [
J 2
] .
J1 J 2 1
(34)
Consequently, it can be concluded from (34) that: if J ex,1 () J ex ,2 () , then
() 0.5 ; if J ex,1 () J ex ,2 () , so 0.5 () 1 .
Consider the following formulas
J ex () 2 () J ex ,1 () [1 ()]2 J ex ,2 () 2 ()[1 ()]J ex ,12 ()
2
(35)
2
J ex,u () () J ex ,1 () [1 ()] J ex ,2 () 2 ()[1 ()]J ex ,12 () (36)
and rearranging (35), we have
J ex () (){J ex ,1 () [1 ()]J ex ,12 ()}
[1 ()]{[1 ()]J ex ,2 () () J ex ,12 ()}
(37)
(){J ex ,12 () ()[ J ex ,1 () J ex ,12 ()]}
[1 ()]{J ex ,12 () [1 ()][ J ex ,2 () J ex ,12 ()]}.
Then, we can rewrite (37) using (34) as
J ex () ()[ J ex,12 () ()J1 ]
[1 ()]{J ex ,12 () [1 ()]J 2 }
.
(38)
Since () J 2 / (J1 J 2 ) and 1 () J1 / (J1 J 2 ) , yielding
J ex () ()[ J ex ,12 ()
J1J 2
J1J 2
] [1 ()][ J ex,12 ()
] . (39)
J1 J 2
J1 J 2
Hence, we obtain
J ex () J ex ,u () J ex ,12 ( )
J1J 2
.
J1 J 2
(40)
According to () (1 , ) , the following bounds hold:
J ex () J ex ,u () J ex ,12 () ()J1 J ex ,1 ()
(41)
J ex () J ex ,u () J ex ,12 () ()J 2 J ex ,2 () .
(42)
That is
J ex () min{ J ex ,1 (), J ex ,2 ()}
.
(43)
J ex ,u ( ) min{ J ex ,1 ( ), J ex ,2 ()}
From the above three situations, it is clear that the proposed NSA-NSA filter performs equally or outperforms the best component filter.
3
Simulation results
To evaluate the performance of the proposed algorithm, three examples of system
(channel) identification were carried out. The results presented here were obtained
from 200 independent Monte Carlo trials. The software of Matlab 8.1 version (2013a)
was used to simulate the proposed algorithm under the computer environment of
AMD (R) A-10 CPU 2.10 GHz and 8Gb memory. To measure the performance of the
algorithms, EMSE using logarithmic scale(dB) was used, defined as:
(44)
EMSE 10 log10 {| ea2 (n) |} .
The unknown system was a ten-tap FIR filter given by random. White Gaussian
noise (WGN) with zero mean and unit variance was used as input. The system was
corrupted by additive WGN and an impulsive noise sequence. The impulsive noise
v(n) was generated from the Bernoulli-Gaussian (BG) distribution [4,17,23,29]
v (n) A(n) I (n)
(45)
where A(n) is a binary independent and identically distributed (i.i.d.) Bernoulli
process with p{ A(n) 1} c and p{ A(n) 0} 1 c , and c is the probability of
occurrence for the impulsive interference I(n). The mean value of v(n) is zero, and its
variance is given by
(46)
var{v (n)} c I2
where I2 var{I (n)} , and the parameters c is set as c 0.01 [4,17,23].
3.1
Example 1
For the first example, the parameter I2 in (46) was fixed at I2 10 4 / 12 , and the
10dB SNR WGN [4,17,23]. The unknown system changes abruptly at n=10000.
Figs. 4 and 5 show the performances of the proposed algorithm with different sets
of N 0 and a . The filter values of the NSA were 1 0.05 , 2 0.005 (which satisfies the stability condition), 1 2 0.0001 , and a 4 . Consider the stability of
evolution of the mixing parameter and the convergence rate, the best choice is
N 0 2 . In addition, we can observe from Fig. 5 that the best choice is a 10 .
Figs. 6 and 7 display the evolution of the mixing parameters (n) and a (n) in
NSA-NSA. Run 1 used the no transfer scheme [1], Run 2 and Run 3 represent the
mixing parameters based on the tracking weight transfer scheme, according to (4) and
(6), respectively. Results demonstrate that the proposed transfer scheme achieves
faster convergence rate and improve the filter robustness in the presence of impulsive
noise. Moreover, Figs. 6 and 7 shows that adjusting the mixing parameter a (n) using
(6) (Run3) results in better stability than other methods.
15
N0=6
10
EMSE(dB)
N0=5
5
N0=4
0
N0=3
-5
N0=2
-10
-15
-20
-25
-30
0
0.2
0.4
0.6
0.8
1
1.2
1.4
1.6
1.8
2
4
n
x 10
Fig. 4. The choice of parameter N 0 in example 1.
4
ρα=5
3
ρα=8
ρα=10
Magnitude
2
ρα=15
1
0
-1
-2
-3
-4
0
0.2
0.4
0.6
0.8
1
n
1.2
1.4
1.6
1.8
2
4
x 10
Fig. 5. The choice of parameter a in example 1.(the mixing parameter a (n) ).
To further show the performance advantage of the proposed method, Fig. 8 depicts
the learning curves of the NLMS-NSA and the NSA-NSA algorithms. This figure
verifies that the performance of the proposed algorithm is at least as good as the better
component in the combination. Both algorithms have the same misadjustment, since
the step size of the slow filters are the same. However, the fast filter of the NLMSNSA is the NLMS, which results in large misadjustment in high background noise
environments. Consequently, the NLMS-NSA suffers from higher misadjustment in
the initial convergence stage. Fig. 9 plots a comparison of NRMN [23], NSA [10],
VSS-NSA [27], VSS-APSA [29], and the proposed algorithm. Clearly, the NSA has a
tradeoff between fast convergence rate and low EMSE, while the proposed algorithm
shows a good balance between the steady-state error and convergence rate.
1
Run1
Run2
Run3
0.9
0.8
Magnitude
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
0
0.2
0.4
0.6
0.8
1
1.2
1.4
1.6
1.8
2
4
n
x 10
Fig. 6. Evolution of the mixing parameter λ(n) of NSA-NSA.
8
Run1
Run2
Run3
6
Magnitude
4
2
0
-2
-4
-6
0
0.2
0.4
0.6
0.8
1
n
1.2
1.4
1.6
1.8
2
4
x 10
Fig. 7. Evolution of the mixing parameter a (n) of NSA-NSA.
15
5
Slow filter
Fast filter
NLMS-NSA(μ 1=0.05, μ 2=0.005)
0
NSA-NSA(μ 1 =0.05,μ 2=0.005)
EMSE(dB)
10
-5
-10
-15
-20
-25
-30
-35
0
0.2
0.4
0.6
0.8
1
1.2
1.4
1.6
1.8
2
4
n
x 10
Fig. 8. Comparison of EMSE of NLMS-NSA algorithm and NSA-NSA for Gaussian input
when 1% impulsive noises are added.
15
VSS-NSA(λ=0.976)
VSS-APSA(K =8,α=0.9)
NRMN(A=0.0005, K=2)
NSA(μ =0.007)
NSA-NSA(μ 1 =0.05,μ 2 =0.005)
10
5
EMSE(dB)
0
-5
-10
-15
-20
-25
-30
-35
0
0.2
0.4
0.6
0.8
1
n
1.2
1.4
1.6
1.8
2
x 10
4
Fig. 9. Comparison of EMSE of NRMN, NSA, VSS-NSA, VSS-APSA algorithms and NSANSA for Gaussian input when 1% impulsive noises are added.
3.2
Example 2
Next, we consider the case of I2 10 4 / 20 and SNR=5dB, which corresponds to
case with the slightly impulsive case and highly Gaussian noises. The abrupt change
appeared in the system at the n=10000.
In this example, the step size of NSA-NSA filter was selected as 1 0.05 ,
2 0.008 , and 1 2 0.0001 . This selection of the parameters ensures good
performance of the algorithm in terms of the convergence rate and steady-state misad-
justment. Fig. 10 displays the choice of N 0 in example 2. We can see that the proposed method is not sensitive to this selection, with the optimal value at N 0 2 . Fig.
11 shows the EMSE of NSA-NSA for different a . The mixing parameter a 10
for the proposed algorithm was selected to guarantee the stability.
Figs. 12 and 13 show the time evolution of the mixing coefficients, where Run 1
represents the no weight transfer scheme [1], and Run 2 and Run 3 represent the mixing parameters based on the tracking weight transfer scheme given by (4) and (6),
respectively. Clearly, it can be observed from these figures that the best selection is
Run 3. The robust performance in the presence of impulsive noise is also improved by
using (9).
15
N0=6
10
EMSE(dB)
N0=5
5
N0=4
0
N0=3
-5
N0=2
-10
-15
-20
-25
-30
0
0.2
0.4
0.6
0.8
1
1.2
1.4
1.6
1.8
2
4
n
x 10
Fig. 10. The choice of parameter N 0 in example 2.
4
ρα=5
3
ρα=8
ρα=10
Magnitude
2
ρα=15
1
0
-1
-2
-3
-4
0
0.2
0.4
0.6
0.8
1
n
1.2
1.4
1.6
1.8
2
4
x 10
Fig. 11. The choice of parameter a in example 2. (the mixing parameter a (n) ).
1
Run1
Run2
Run3
0.9
0.8
Magnitude
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
0
0.2
0.4
0.6
0.8
1
1.2
1.4
1.6
1.8
2
4
n
x 10
Fig. 12. Evolution of the mixing parameter λ(n) of NSA-NSA.
12
Run1
Run2
Run3
10
8
Magnitude
6
4
2
0
-2
-4
-6
-8
0
0.2
0.4
0.6
0.8
1
n
1.2
1.4
1.6
1.8
2
4
x 10
Fig. 13. Evolution of the mixing parameter a (n) of NSA-NSA.
15
Slow filter
Fast filter
NLMS-NSA( μ1 =0.05, μ2=0.008)
10
5
EMSE(dB)
NSA-NSA( μ1=0.05, μ2 =0.008)
0
-5
-10
-15
-20
-25
-30
0
0.2
0.4
0.6
0.8
1
1.2
1.4
1.6
1.8
2
4
n
x 10
Fig. 14. Comparison of EMSE of NLMS-NSA algorithm and NSA-NSA for Gaussian input
when 1% impulsive noises are added.
15
VSS-NSA(λ =0.976)
VSS-APSA(K =8, α=0.9)
NRMN(A=0.0005,K =2)
NSA( μ=0.007)
NSA-NSA( μ 1 =0.05, μ 2 =0.008)
10
5
EMSE(dB)
0
-5
-10
-15
-20
-25
-30
0
0.2
0.4
0.6
0.8
1
n
1.2
1.4
1.6
1.8
2
4
x 10
Fig. 15. Comparison of EMSE of NRMN, NSA, VSS-NSA, VSS-APSA algorithms and NSANSA for Gaussian input when 1% impulsive noises are added.
Fig. 14 plots a comparison of NLMS-NSA and the proposed algorithms. Again, we
see that the EMSE of NSA-NSA is consistent with the theoretical analysis. Both algorithms achieve quite similar steady-state error, but the proposed algorithm has the
smaller misadjustment in the initial stage of convergence. This is due to the fact that
the NLMS algorithm is not well-suited for impulsive noise environment. Fig. 15
shows a comparison of the learning curves from NRMN [23], NSA [10], VSS-NSA
[27], VSS-APSA [29] and NSA-NSA for high Gaussian noise and low impulsive
noise environments. It is observed that the proposed algorithm achieves an improved
performance in the presence of impulsive noise.
3.3
Intersymbol Interference (ISI) channel identification under impulsive
noise environment
300
impulsive noise
200
Magnitude
100
0
-100
-200
-300
-400
0
0.2
0.4
0.6
0.8
1
1.2
1.4
1.6
1.8
2
4
n
x 10
Fig. 16. Impulsive noise in ISI channel.
In this section, we consider a real intersymbol interference (ISI) channel corrupted by
impulsive noise, which occurs quite often in communication systems. Here, we model
the ISI channel as
T
(47)
w 0 [0.04,
0.05, 0.07, 0.21, 0.5,
0.72, 0.36,0, 0.21, 0.03, 0.07]
.
eleven coefficients
In practice, the channel information is unknown. To deal with such problem, the
length of our filter was set to M=13. Quadrature phase shift keyin (QPSK) was used
as the input signal. A segment of 10000 samples was used as the training data and
another 10000 as the test data. The ISI channel was corrupted by impulsive noise, as
shown in Fig. 16. The performance of the proposed NSA-NSA1 is demonstrated, in
comparison with the NLMS-NSA2.
Fig. 17 shows the learning curves of the two algorithms in impulsive noise. Clearly, with impulsive noise, the performance of NSA-NSA is barely affected by large
disturbances, while the performance of NLMS-NSA deteriorates significantly due to
NLMS’s sensitivity to outliers.
1
With
QPSK
input,
the
adaptation
of
a(n)
of
NSA-NSA
is
given
as
a(n 1) a(n) a conj{sign{e(n)}}[ y1 (n) y2 (n)] (n)[1 (n)] , where conj{} denotes conjugate
2
operation.
The derivation of VSS-NSA, VSS-APSA, and NRMN are different from the original literatures, when input signal is the complex number. For paper length optimization, and in order
to focus on the simplicity of the proposed approach, we have decided to only compare to
NLMS-NSA algorithm.
25
15
Slow filter
Fast filter
NLMS-NSA(μ 1 =0.1,μ 2=0.03)
10
NSA-NSA(μ 1=0.1,μ 2 =0.03)
EMSE(dB)
20
5
0
-5
-10
-15
-20
-25
0
1000
2000
3000
4000
5000
6000
7000
8000
9000 10000
n
Fig. 17. Learning curves of NLMS-NSA and NSA-NSA in ISI channel identification (Testing
stage).
4
Conclusions
A novel NSA-NSA was proposed to improve the performance of NSA for system
identification under impulsive noise. The proposed adaptive convex scheme which
combines a fast and a slow NSA filter to achieve both fast convergence speed and low
steady-state error. Moreover, a sign cost function scheme to adjust the mixing parameter was introduced to improve the robustness of the algorithm under impulsive
noise. To further accelerate the initial convergence rate, a tracking weight transfer
scheme was applied in the NSA-NSA. Simulation results demonstrated that the proposed algorithm has better performance than the existing algorithms in terms of convergence rate and steady-state error.
Acknowledgments
The authors want to express their deep thanks to the anonymous reviewers for many
valuable comments which greatly helped to improve the quality of this work.
This work was supported in part by National Natural Science Foundation of China
(Grants: 61271340, 61571374, 61134002, 61433011, U1234203), the Sichuan Provincial Youth Science and Technology Fund (Grant: 2012JQ0046), and the Fundamental Research Funds for the Central Universities (Grant: SWJTU12CX026).
Reference
1. J. Arenas-García, A. R. Figueiras-Vidal, A. H. Sayed, Mean-square performance of a convex combination of two adaptive filters. IEEE Trans. Signal Process. 54(3), 1078–1090
(2006). doi: 10.1109/TSP.2005.863126
2. J. Arenas-Garcia, A. R. Figueiras-Vidal, Adaptive combination of normalised filters for
robust system identification. Electronics Lett. 41(15), 874–875 (2005). doi:
10.1049/el:20051936
3. J. A. Chambers, O. Tanrikulu, A. G. Constantinides, Least mean mixed-norm adaptive filtering. Electronics Lett. 30(19), 1574–1575 (1994). doi: 10.1049/el:19941060
4. J. Chambers, A. Avlonitis, A robust mixed-norm adaptive filter algorithm. IEEE Signal
Process. Lett. 4(2), 46–48 (1997). doi: 10.1109/97.554469
5. S. C. Douglas, A family of normalized LMS algorithms. IEEE Signal Process. Lett. 1(3),
49–51 (1994). doi: 10.1109/97.295321
6. S. C. Douglas, Analysis and implementation of the max-NLMS adaptive filter. In PRoc.
29th Asilomar Conf. on Signals, Systems, and Computers, 659–663 (1995).
7. E. Eweda, Analysis and design of signed regressor LMS algorithm for stationary and nonstationary adaptive filtering with correlated Gaussian data. IEEE Trans. Circuits Syst.
37(11), 1367–1374 (1990). doi: 10.1109/31.62411
8. S. B. Jebara, H. Besbes, Variable step size filtered sign algorithm for acoustic echo cancellation. Electronics Lett. 39(12), (2003). 936–938. doi: 10.1049/el:20030583
9. B. E. Jun, D. J. Park, Y. W. Kim, Convergence analysis of sign-sign LMS algorithm for
adaptive filters with correlated Gaussian data. In: IEEE International Conference on
Acoustics, Speech, and Signal Processing, 1380–1383 (1995).
10. S. Koike, Variable step size normalized sign algorithm for fast convergent adaptive filters
with robustness against impulsive noise. NEC research & development 41(3), 278–288
(2000).
11. S. Koike, Analysis of adaptive filters using normalized sign regressor LMS algorithm.
IEEE Trans. Signal Process. 47(10), 2710–2723 (1999). doi: 10.1109/78.790653
12. S. Koike, Convergence analysis of adaptive filters using normalized sign-sign algorithm,
IEICE Trans. Fundamentals of Electronics Communications and Computer Sciences, E88A(11), 3218–3224 (2006).
13. R. H. Kwong, E. W. Johnston, A variable step size LMS algorithm. IEEE Trans. Signal
Process. 40(7), 1633–1642(1992). doi: 10.1109/78.143435
14. C. P. Kwong, Dual sign algorithm for adaptive filtering. IEEE Trans. Commun. 34(12),
1272–1275 (1986). doi: 10.1109/TCOM.1986.1096490
15. L. Lu, H. Zhao, A novel convex combination of LMS adaptive filter for system identification. 2014 12th International Conference on Signal Processing (ICSP), Hangzhou 225–229
(2014).
16. V. J. Mathews, Z. Xie, A stochastic gradient adaptive filter with gradient adaptive step
size. IEEE Trans. Signal Process. 41(6), 2075–2087 (1993). doi: 10.1109/78.218137
17. D. P. Mandic, E. V. Papoulis, C. G. Boukis, A normalized mixed-norm adaptive filtering
algorithm robust under impulsive noise interference. In: IEEE International Conference on
Acoustics, Speech, and Signal Processing, 333–336 (2003).
18. D. P. Mandic, NNGD algorithm for neural adaptive filters. Electronics Lett. 36(9), 845–
846 (2000). doi: 10.1049/el:20000631
19. D. P. Mandic, J. A. Chambers, Toward the optimal learning rate for backpropagation.
Neural Processing Letters, 11(1), 1–5 (2000). doi: 10.1023/A:1009686825582
20. D. P. Mandic, A. I. Hanna, M. Razaz, A normalized gradient descent algorithm for nonlinear adaptive filters using a gradient adaptive step size. IEEE Signal Process. Lett. 8(11),
295–297 (2001). doi: 10.1109/97.969448
21. V. J. Mathews, S. H. Cho, Improved convergence analysis of stochastic gradient adaptive
filters using the sign algorithm. IEEE Trans. Acoust., Speech, Signal Process. 35(4), 450–
454 (1987). doi: 10.1109/TASSP.1987.1165167
22. V. H. Nascimento, R. C. de Lamare, A low-complexity strategy for speeding up the convergence of convex combinations of adaptive filters. In: IEEE International Conference on
Acoustics, Speech and Signal Processing, 3553–3556 (2012).
23. E. V. Papoulis, T. Stathaki, A normalized robust mixed-norm adaptive algorithm for system identification. IEEE Signal Process. Lett. 11(1), 56–59 (2004). doi:
10.1109/LSP.2003.819353
24. D. I. Pazaitis, A. G. Constantinides, LMS+F algorithm. Electronics Lett. 31(17), 1423–
1424 (1995). doi: 10.1049/el:19951026
25. R. Price, A useful theorem for nonlinear devices having Gaussian inputs. IRE Transactions
on Information Theory, 4(2), 69–72 (1958). doi: 10.1109/TIT.1958.1057444
26. A. H. Sayed, Fundamentals of Adaptive Filtering. (Wiley IEEE Press, 2003).
27. T. Shao, Y. R. Zheng, J. Benesty, A variable step-size normalized sign algorithm for
acoustic echo cancelation. In: IEEE International Conference on Acoustics, Speech, and
Signal Processing, 333–336 (2010).
28. T. Shao, Y. R. Zheng, J. Benesty, An affine projection sign algorithm robust against impulsive interferences. IEEE Signal Process. Lett. 17(4), 327–330 (2010). doi:
10.1109/LSP.2010.2040203
29. J. Shin, J. Yoo, P. Park, Variable step-size affine projection sign algorithm. Electronics
Lett. 48(9), 483–485 (2012). doi: 10.1049/el.2012.0751
30. J. Soo, K. K. Pang, A multi step size (MSS) frequency domain adaptive filter. IEEE Trans.
Signal Process. 39(1), 115–121 (1991). doi: 10.1109/78.80770
31. O. Tanrikulu, J. A. Chambers, Convergence and steady-state properties of the least-mean
mixed-norm (LMMN) adaptive algorithm. IEE Proceedings-Vision, Image and Signal
Processing, 143, 137–142 (1996).
32. P. Yuvapoositanon, J. Chambers, An adaptive step-size code-constrained minimum output
energy receiver for nonstationary CDMA channels. In: IEEE International Conference on
Acoustics, Speech, and Signal Processing, 465–468 (2003).
| 3cs.SY
|
arXiv:1708.09262v2 [math.NT] 3 Sep 2017
BOUNDED GENERATION OF SL2 OVER RINGS OF S-INTEGERS WITH
INFINITELY MANY UNITS
ALEKSANDER V. MORGAN, ANDREI S. RAPINCHUK, AND BALASUBRAMANIAN SURY
Abstract. Let O be the ring of S-integers in a number field k. We prove that if the group of units
O× is infinite then every matrix in Γ = SL2 (O) is a product of at most 9 elementary matrices. This
completes a long line of research in this direction. As a consequence, we obtain that Γ is boundedly
generated as an abstract group.
To Alex Lubotzky on his 60th birthday
1. Introduction
Let k be a number field. Given a finite subset S of the set V k of valuations of k containing the
k of archimedian valuations, we let O
set V∞
k,S denote the ring of S-integers in k, i.e.
Ok,S = {a ∈ k× | v(a) ≥ 0 for all v ∈ V k \ S} ∪ {0}.
As usual, for any commutative ring R, we let SL2 (R) denote the group of unimodular 2 × 2-matrices
over R and refer to the matrices
1 a
1 0
E12 (a) =
and E21 (b) =
∈ SL2 (R) (a, b ∈ R)
0 1
b 1
as elementary (over R).
It was established in [Va] (see also [L1]) that if the ring of S-integers O = Ok,S has infinitely
many units, the group Γ = SL2 (O) is generated by elementary matrices. The goal of this paper is
to prove that in this case Γ is actually boundedly generated by elementaries.
Theorem 1.1. Let O = Ok,S be the ring of S-integers in a number field k, and assume that the
group of units O× is infinite. Then every matrix in SL2 (O) is a product of at most 9 elementary
matrices.
The quest to validate this property has a considerable history. First, G. Cooke and P. J. Weinberger [CW] established it (with the same bound as in Theorem 1.1) assuming the truth of a
suitable form of the Generalized Riemann Hypothesis, which still remains unproven. Later, it was
shown in [LM] (see also [M]) by analytic tools that the argument can be made unconditional if
|S| ≥ max(5, 2[k : Q] − 3). On the other hand, B. Liehl [L2] proved the result by algebraic methods
for some special fields k. The first unconditional proof in full generality was given by D. Carter,
G. Keller and E. Paige in an unpublished preprint; their argument was streamlined and made
available to the public by D. W. Morris [MCKP]. This argument is based on model theory and
provides no explicit bound on the number of elementaries required; besides, it uses difficult results
from additive number theory.
In [Vs], M. Vsemirnov proved Theorem 1.1 for O = Z[1/p] using the results of D. R. Heath-Brown
[HB] on Artin’s Primitive Root Conjecture (thus, in a broad sense, this proof develops the initial
1
2
A. MORGAN, A. RAPINCHUK, AND B. SURY
approach of Cooke and Weinberger [CW]); his bound on the number of elementaries required is
≤ 5. Subsequently, the third-named author re-worked the argument from [Vs] to avoid the use of
[HB] in an unpublished note. These notes were the beginning of the work of the first two authors
that has eventually led to a proof of Theorem 1.1 in the general case. It should be noted that our
proof uses only standard results from number theory, and is relatively short and constructive with
an explicit bound which is independent of the field k and the set S. This, in particular, implies
that Theorem 1.1 remains valid for any infinite S.
The problem of bounded generation (particularly by elementaries) has been considered for Sarithmetic subgroups of algebraic groups other than SL2 . A few years after [CW], Carter and
Keller [CK1] showed that SLn (O) for n ≥ 3 is boundedly generated by elementaries for any ring
O of algebraic integers (see [T] for other Chevalley groups of rank > 1, and [ER] for isotropic, but
nonsplit (or quasi-split), orthogonal groups). The upper bound on the number of factors required to
write every matrix in SLn (O) as a product of elementaries given in [CK1] is 12 (3n2 − n) + 68∆ − 1,
where ∆ is the number of prime divisors of the discriminant of k; in particular, this estimate
depends on the field k. Using our Theorem 1.1, one shows in all cases where the group of units
O× is infinite, this estimate can be improved to 12 (3n2 − n) + 4, hence made independent of k –
see Corollary 4.6. The situation not covered by this result are when O is either Z or the ring of
integers in an imaginary quadratic field – see below. The former case was treated in [CK2] with
an estimate 12 (3n2 − n) + 36, so only in the case of imaginary quadratic fields the question of the
existence of a bound on the number of elementaries independent of the k remains open.
From a more general perspective, Theorem 1.1 should be viewed as a contribution to the sustained
effort aimed at proving that all higher rank lattices are boundedly generated as abstract groups. We
recall that a group Γ is said to have bounded generation (BG) if there exist elements γ1 , . . . , γd ∈ Γ
such that
Γ = hγ1 i · · · hγd i,
where hγi i denotes the cyclic subgroup generated by γi . The interest in this property stems from
the fact that while being purely combinatorial in nature, it is known to have a number of farreaching consequences for the structure and representations of a group, particularly if the latter is
S-arithmetic. For example, under one additional (necessary) technical assumption, (BG) implies
the rigidity of completely reducible complex representations of Γ (known as SS-rigidity) – see [R],
[PR2, Appendix A]. Furthermore, if Γ is an S-arithmetic subgroup of an absolutely simple simply
connected algebraic group G over a number field k, then assuming the truth of the MargulisPlatonov conjecture for the group G(k) of k-rational points (cf. [PR2, §9.1]), (BG) implies the
congruence subgroup property (i.e. the finiteness of the corresponding congruence kernel – see [Lu],
[PR1]). For applications of (BG) to the Margulis-Zimmer conjecture, see [SW]. Given these and
other implications of (BG), we would like to point out the following consequence of Theorem 1.1.
Corollary 1.2. Let O = Ok,S be the ring of S-integers, in a number field k. If the group of units
O× is infinite, then the group Γ = SL2 (O) has bounded generation.
We note that combining this fact with the results of [Lu], [PR1], one obtains an alternative
proof of the centrality of the congruence kernel for SL2 (O) (provided that O× is infinite), originally
established by J.-P. Serre [S1]. We also note that (BG) of SL2 (O) is needed to prove (BG) for some
other groups – cf. [T] and [ER].
Next, it should be pointed out that the assumption that the unit group O× is infinite is necessary
for the bounded generation of SL2 (O), hence cannot be omitted. Indeed, it follows from Dirichlet’s
BOUNDED GENERATION OF SL2 (OS )
3
Unit Theorem [CF, §2.18] that O× is finite only when |S| = 1 which happens precisely when S is
the set of archimedian valuations in the following two cases:
1) k = Q and O = Z. In this case, the group SL2 (Z) is generated by the elementaries, but has a
nonabelian free
√ subgroup of finite index, which prevents it from having bounded generation.
2) k = Q( −d) for some square-free integer d ≥ 1, and Od is the ring of algebraic integers
integers in k. According to [GS], the group Γ = SL2 (Od ) has a finite index subgroup that admits an
epimorphism onto a nonabelian free group, hence again cannot possibly be boundedly generated.
Moreover, P. M. Cohn [Co] shows that if d ∈
/ {1, 2, 3, 7, 11} then Γ is not even generated by
elementary matrices.
The structure of the paper is the following. In §2 we prove an algebraic result about abelian
subextensions of radical extensions of general field – see Proposition 2.1. This statement, which
may be of independent interest, is used in the paper to prove Theorem 3.7. This theorem is one
of the number-theoretic results needed in the proof of Theorem 1.1, and it is established in §3
along with some other facts from algebraic number theory. One of the key notions in the paper
is that of Q-split prime: we say that a prime p of a number field k is Q-split if it is non-dyadic
and its local degree over the corresponding rational prime is 1. In §3, we establish some relevant
for us properties of such primes (see §3.1) and prove for them in §3.2 the following refinement of
Dirichlet’s Theorem from [BMS].
Theorem 3.3. Let O be the ring of S-integers in a number field k for some finite S ⊂ V k containing
k . If nonzero a, b ∈ O are relatively prime (i.e., aO + bO = O) then there exist infinitely many
V∞
principal Q-split prime ideals p of O with a generator π such that π ≡ a (mod bO) and π > 0 in all
real completions of k.
Subsection 3.3 is devoted to the statement and proof of the key Theorem 3.7, which is another
key number-theoretic result needed in the proof of Theorem 1.1. In §4, we prove Theorem 1.1 and
Corollary 1.2. Finally, in §5 we correct the faulty example from [Vs] of a matrix in SL2 (Z[1/p]),
where p is a prime ≡ 1(mod 29), that is not a product of four elementary matrices – see Proposition
5.1, confirming thereby that the bound of 5 in [Vs] is optimal.
Notations and conventions. For a field k, we let kab denote the maximal abelian extension of k.
Furthermore, µ(k) will denote the group of all roots of unity in k; if µ(k) is finite, we let µ denote
its order. For n ≥ 1 prime to char k, we let ζn denote a primitive n-th root of unity.
In this paper, with the exception of §2, the field k will be a field of algebraic numbers (i.e., a
finite extension of Q), in which case µ(k) is automatically finite. We let Ok denote the ring of
algebraic integers in k. Furthermore, we let V k denote the set of (the equivalence classes of) nonk and V k denote the subsets of archimedean and nonarchimedean
trivial valuations of k, and let V∞
f
valuations, respectively. For any v ∈ V k , we let kv denote the corresponding completion; if v ∈ Vfk
then Ov will denote the valuation ring in kv with the valuation ideal p̂v and the group of units
Uv = O×
v.
k , and O = O
Throughout the paper, S will denote a fixed finite subset of V k containing V∞
k,S the
corresponding ring of S-integers (see above). Then the nonzero prime ideals of O are in a natural
bijective correspondence with the valuations in V k \ S. So, for a nonzero prime ideal p ⊂ O we
let vp ∈ V k \ S denote the corresponding valuation, and conversely, for a valuation v ∈ V k \ S we
let pv ⊂ O denote the corresponding prime ideal (note that pv = O ∩ p̂v ). Generalizing Euler’s
4
A. MORGAN, A. RAPINCHUK, AND B. SURY
ϕ-function, for a nonzero ideal a of O, we set
φ(a) = |(O/a)× |.
For simplicity of notation, for an element a ∈ O, φ(a) will always mean φ(aO). Finally, for a ∈ k× ,
we let V (a) = { v ∈ Vfk | v(a) 6= 0 }.
Given a prime number p, one can write any integer n in the form n = pe ·m, for some non-negative
integer e, where p /| m. We then call pe the p-primary component of n.
2. Abelian subextensions of radical extensions.
In this section, k is an arbitrary field. For a prime p 6= char k, we let µ(k)p denote the subgroup
d
of µ(k), consisting of elements satisfying xp = 1 for some d ≥ 0. If this subgroup is finite, we set
λ(k)p to be the non-negative integer satisfying |µ(k)p | = pλ(k)p ; otherwise, set λ(k)p = ∞. Clearly
Q
√
if µ(k) is finite, then µ = p pλ(k)p . For a ∈ k× , we write n a to denote an arbitrary root of the
polynomial xn − a.
The goal of this section is to prove the following.
Proposition 2.1. Let n ≥ 1 be an integer prime to char k, and let u ∈ k × be such
that u ∈
/ µ(k)p k×
√
n
n
for all p | n. Then the polynomial x − u is irreducible over k, and for t = u we have
n
,
k(t) ∩ kab = k(tm ) where m = Y
gcd(n, pλ(k)p )
p
p|n
with the convention that
gcd(n, p∞ )
is simply the p-primary component of n.
We first treat the case n = pd where p is a prime.
Proposition 2.2. Let p be a prime number 6= char k, and let u ∈ k × \ µ(k)p (k× )p . Fix an integer
√
d
d ≥ 1, set t = p u. Then
γ
k(t) ∩ kab = k(tp ) where γ = max(0, d − λ(k)p ).
We begin with the following lemma.
√
Lemma 2.3. Let p be a prime number 6= char k, and let u ∈ k × \ µ(k)p (k× )p . Set k1 = k( p u).
Then
(i) [k1 : k] = p;
(ii) µ(k1 )p = µ(k)p
√
(iii) None of the p u are in µ(k1 )p (k1× )p .
Proof. (i) follows from [La, Ch. VI, §9], as u ∈
/ (k× )p .
(ii): If λ(k)p = ∞, then there is nothing to prove. Otherwise, we need to show that for λ = λ(k)p ,
/ k1 . Assume the contrary. Then, first, λ > 0. Indeed, we have a tower of inclusions
we have ζpλ+1 ∈
k ⊆ k(ζp ) ⊆ k1 . Since [k1 : k] = p by (i), and [k(ζp ) : k] ≤ p − 1, we conclude that [k(ζp ) : k] = 1,
i.e. ζp ∈ k.
/ k, we have
Now, since ζpλ+1 ∈
q
(1)
k1 = k(ζpλ+1 ) = k( p ζpλ ).
BOUNDED GENERATION OF SL2 (OS )
5
√
√
But according to Kummer’s theory (which applies because ζp ∈ k), the fact that k( p a) = k( p b)
for a, b ∈ k× implies that the images of a and b in k× /(k× )p generate the same subgroup. So,
it follows from (1) that uζpi ∈ (k× )p for some i, and therefore u ∈ µ(k)p (k× )p , contradicting our
choice of u.
√
√
(iii): Assume the contrary, i.e. some p-th root p u can be written in the form p u = ζap for some
a ∈ k1× and ζ ∈ µ(k1 )p . Let N = Nk1 /k : k1× → k× be the norm map. Then
√
N ( p u) = N (ζ)N (a)p .
√
√
Clearly, N (ζ) ∈ µ(k)p , so N ( p u) ∈ µ(k)p (k× )p . On the other hand, N ( p u) = u for p odd, and
−u for p = 2. In all cases, we obtain that u ∈ µ(k)p (k× )p . A contradiction.
A simple induction now yields the following:
Corollary 2.4. Let p be a prime number 6= char k, and let u ∈ k × \ µ(k)p (k× )p . For a fixed integer
√
d
d ≥ 1, set kd = k( p u). Then:
(i) [kd : k] = pd ;
(ii) µ(kd )p = µ(k)p , hence λ(kd )p = λ(k)p .
Of course, assertion (i) is well-known and follows, for example, from [La, Ch. VI, §9].
Lemma 2.5. Let p be a prime number 6= char k, and let u ∈ k × \ µ(k)p (k× )p . Fix an integer
√
d
d ≥ 1, and set t = p u and kd = k(t). Furthermore, for an integer j between 0 and d define
√
d−j
j
ℓj = k(tp ) ≃ k( p u). Then any intermediate subfield k ⊆ ℓ ⊆ kd is of the form ℓ = ℓj for some
j ∈ {0, . . . , d}.
Proof. Given such an ℓ, it follows from Corollary 2.4(i) that [kd : ℓ] = pj for some 0 ≤ j ≤ d.
d
Since any conjugate of t is of the form ζ · t where ζ p = 1, we see that the norm Nkd /ℓ (t) is of the
d
form ζ0 tp , where again ζ0p = 1. Then ζ0 ∈ µ(kd )p , and using Corollary 2.4(ii), we conclude that
j
ζ0 ∈ k ⊆ ℓ. So, tp ∈ ℓ, implying the inclusion ℓd−j ⊆ ℓ. Now, the fact that [kd : ℓd−j ] = pj implies
that ℓ = ℓd−j , yielding our claim.
j
√
d
Proof of Proposition 2.2. Set λ = λ(k)p . Then for any d ≤ λ the extension k( p u)/k is abelian,
and our assertion is trivial. So, we may assume that λ < ∞ and d > λ. It follows from Lemma 2.5
j
that ℓ := k(t) ∩ kab is of the form ℓd−j = k(tp ) for some j ∈ {0, . . . , d}. On the other hand, ℓd−j /k
d−j
d−j
of tp , implying
is a Galois extension of degree pd−j , so must contain the conjugate ζpd−j tp
√
d−j
that ζpd−j ∈ ℓd−j . Since ℓd−j ≃ k( p
u), we conclude from Corollary 2.4(ii) that d − j ≤ λ,
γ
i.e. j ≥ d − λ. This proves the inclusion ℓ ⊆ k(tp ); the opposite inclusion is obvious.
Proof of Proposition 2.1.√Let n = pα1 1 · · · pαs s be the prime factorization of n, and for i = 1, . . . , s
set ni = n/pαi i . Let t = n u and ti = tni (so, ti is a pαi i -th root of u). Using again [La, Ch. VI, §9]
we conclude that [k(t) : k] = n, which implies that
(2)
[k(t) : k(ti )] = ni for all i = 1, . . . , r.
6
A. MORGAN, A. RAPINCHUK, AND B. SURY
Since for K := k(t) ∩ kab the degree [K : k] divides n, we can write K = K1 · · · Ks where Ki is
an abelian extension of k of degree pβi i for some βi ≤ αi . Then the degree [Ki (ti ) : k(ti )] must be
a power of pi . Comparing with (2), we conclude that Ki ⊆ k(ti ). Applying Proposition 2.2 with
d = αi , we obtain the inclusion
(3)
p
γi
γi
Ki ⊆ k(ti i ) = k(tni pi ) where γi = max(0 , αi − λ(k)pi ).
It is easy to see that the g.c.d. of the numbers ni pγi i for i = 1, . . . , s is
m= Y
n
gcd(n, pλ(k)p )
.
p|n
γ1
γs
Furthermore, the subgroup of k(t)× generated by tn1 p1 , . . . , tns ps coincides with the cyclic subgroup with generator tm . Then (3) yields the following inclusion
K = K1 · · · Ks ⊆ k(tm ).
Since the opposite inclusion is obvious, our claim follows.
Corollary 2.6. Assume that µ = |µ(k)| < ∞. Let P be a finite set of rational primes, and define
Y
µ′ = µ ·
p.
p∈P
Given u ∈ k× such that
u∈
/ µ(k)p (k× )p for all p ∈ P,
for any abelian extension F of k the intersection
E := F ∩ k
is contained in k
√
µ
u, ζµ′ .
√
µ′
u, ζµ′
Proof. Without loss of generality we may assume that ζµ′ ∈ F , and then we have the following
tower of field extensions
√
√
√
′
k µ u , ζµ′ ⊂ E µ u ⊂ k µ u , ζµ′ .
Q
√
√
′
We note that the degree k µ u , ζµ′ : k µ u , ζµ′ divides p∈P p. So, if we assume that the
assertion
the lemma
is false,
the
of √
then we should be able to find to find a prime
p∈
√
√P that divides
√
µ
µ
µ′
µ
′
′
degree E ( u) : k
u , ζµ , and therefore does not divide the degree k
u , ζµ : E ( u) .
√
√
√
The latter
implies that pµ u ∈ E ( µ u). But this contradicts Proposition 2.1 since E ( µ u) =
√
E · k ( µ u) is an abelian extension of k.
BOUNDED GENERATION OF SL2 (OS )
7
3. Results from Algebraic Number Theory
1. Q-split primes. Our proof of Theorem 1.1 heavily relies on properties of so-called Q-split
primes in O.
Definition. Let p be a nonzero prime ideal of O, and let p be the corresponding rational prime.
We say that p is Q-split if p > 2, and for the valuation v = vp we have kv = Qp .
For the convenience of further references, we list some simple properties of Q-split primes.
Lemma 3.1. Let p be a Q-split prime in O, and for n ≥ 1 let ρn : O → O/pn be the corresponding
quotient map. Then:
(a) the group of invertible elements (O/pn )× is cyclic for any n;
(b) if c ∈ O is such that ρ2 (c) generates (O/p2 )× then ρn (c) generates (O/pn )× for any n ≥ 2.
Proof. Let p > 2 be the rational prime corresponding to p, and v = vp be the associated valuation
of k. By definition, kv = Qp , hence Ov = Zp . So, for any n ≥ 1 we will have canonical ring
isomorphisms
(4)
O/pn ≃ Ov /p̂nv ≃ Zp /pn Zp ≃ Z/pn Z.
Then (a) follows from the well-known fact that the group (Z/pn Z)× is cyclic. Furthermore, the
isomorphisms in (4) are compatible for different n’s. Since the kernel of the group homomorphism
(Z/pn Z)× → (Z/p2 Z)× is contained in the Frattini subgroup of (Z/pn Z)× for n ≥ 2, the same is
true for the homomorphism (O/pn )× → (O/p2 )× . This easily implies (b).
Let p be a Q-split prime, let v = vp be the corresponding valuation. We will now define the level
ℓp (u) of an element u ∈ O×
v and establish some properties of this notion that we will need later.
Let p > 2 be the corresponding rational prime. The group of p-adic units Up = Z×
p has the
natural filtration by the congruence subgroups
It is well-known that
i
U(i)
p = 1 + p Zp for i ∈ N.
Up = C × U(1)
p
where C is the cyclic group of order (p − 1) consisting of all roots of unity in Qp . Furthermore,
(i)
the logarithmic map yields a continuous isomorphism Up → pi Zp , which implies that for any
u ∈ Up \ C, the closure of the cyclic group generated by u has a decomposition of the form
hui = C ′ × U(ℓ)
p
for some subgroup C ′ ⊂ C and some integer ℓ = ℓp (u) ≥ 1 which we will refer to as the p-level of
u. We also set ℓp (u) = ∞ for u ∈ C.
Returning now to a Q-split prime p of k and keeping the above notations, we define the p-level
ℓp (u) of u ∈ O×
v as the the p-level of the element in Up that corresponds to u under the natural
identification Ov = Zp . We will need the following.
Lemma 3.2. Let p be a Q-split prime in O, let p be the corresponding rational prime, and v = vp
the corresponding valuation. Suppose we are given an integer d ≥ 1 not divisible by p, a unit
u ∈ O×
v of infinite order having p-level s = ℓp (u), and integer ns , and an element c ∈ Ov such
that uns ≡ c (mod ps ). Then for any t ≥ s there exists an integer nt ≡ ns (mod d) for which
unt ≡ c (mod pt ).
8
A. MORGAN, A. RAPINCHUK, AND B. SURY
Proof. In view of the identification Ov = Zp , it is enough to prove the corresponding statement
for Zp . More precisely, we need to show the following: Let u ∈ Up be a unit of infinite order and
p-level s = ℓp (u). If c ∈ Up and ns ∈ Z are such that uns ≡ c (mod ps ), then for any t ≥ s there
(s)
exists nt ≡ ns (mod d) such that unt ≡ c (mod pt ). Thus, we have that uns ∈ cUp , and we wish to
show that
\
uns · hud i cUp(t) 6= ∅.
(t)
Since cUp is open, it is enough to show that
uns · hud i
(5)
\
cU(t)
p 6= ∅.
(s)
But since ℓp (u) = s and d is prime to p, we have the inclusion hud i ⊃ Up , and (5) is obvious.
2. Dirichlet’s Theorem for Q-split primes. We will now establish the existence of Q-split
primes in arithmetic progressions.
Theorem 3.3. Let O be the ring of S-integers in a number field k for some finite S ⊂ V k containing
k . If nonzero a, b ∈ O are relatively prime (i.e., aO + bO = O) then there exist infinitely many
V∞
principal Q-split prime ideals p of O with a generator π such that π ≡ a (mod bO) and π > 0 in all
real completions of k.
The proof follows the same general strategy as the proof of Dirichlet’s Theorem in [BMS] see Theorem A.10 in the Appendix on Number Theory. First, we will quickly review some basic
facts from global class field theory (cf., for example, [CF], Ch. VII) and fix some notations. Let Jk
denote the group of ideles of k with the natural topology; as usual, we identify k× with the (discrete)
subgroup of principal ideles in Jk . Then for every open subgroup U ⊂ Jk of finite index containing
k× there exists a finite abelian Galois extension L/k and a continuous surjective homomorphism
αL/k : Jk → Gal(L/k) (known as the norm residue map) such that
• U = Ker αL/k = NL/k (JL )k× ;
• for every nonarchimedean v ∈ V k which is unramified in L we let FrL/k (v) denote the Frobenius automorphism of L/k at v (i.e., the Frobenius automorphism FrL/k (w|v) associated to
some (equivalently, any) extension w|v) and let i(v) ∈ Jk be an idele with the components
1 , v ′ 6= v
,
i(v)v′ =
πv , v ′ = v
where πv ∈ kv is a uniformizer; then αL/k (i(v)) = FrL/k (v).
k , we define the following open subgroup of J :
For our fixed finite subset S ⊂ V k containing V∞
k
Y
Y
×
US :=
kv ×
Uv .
v∈S
v∈V k \S
Then the abelian extension of k corresponding to the subgroup US := US k× will be called the
Hilbert S-class field of k and denoted K throughout the rest of the paper.
Next, we will introduce the idelic S-analogs of ray groups. Let b be a nonzero ideal of O = Ok,S
with the prime factorization
(6)
b = pn1 1 · · · pnt t ,
BOUNDED GENERATION OF SL2 (OS )
9
let vi = vpi be the valuation in V k \ S associated with pi , and let V (b) = {v1 , . . . , vt }. We then
define an open subgroup
Y
RS (b) =
Rv
v∈V k
where the open subgroups Rv ⊆ kv× are defined as follows. For v real, we let Rv be the subgroup
of positive elements, letting Rv = kv× for all other v ∈ S, and setting Rv = Uv for all v ∈
/ S ∪ V (b).
It remains to define Rv for v = vi ∈ V (b), in which case we set it to be the congruence subgroup
(n )
Uvi i of Uvi modulo p̂nvii . We then let K(b) denote the abelian extension of k corresponding to
RS (b) := RS (b)k× (“ray class field”). (Obviously, K(b) contains K for any nonzero ideal b of O.)
Furthermore, given c ∈ k × , we let jb (c) denote an idele with the following components:
c , v ∈ V (b),
jb (c)v =
1 , v∈
/ V (b).
Then θb : k× → Gal(K(b)/k) defined by c 7→ αK(b)/k (jb (c))−1 is a group homomorphism.
The following lemma summarizes some simple properties of these definitions.
Lemma 3.4. Let b ⊂ O be a nonzero ideal.
(a) If a nonzero c ∈ O is relatively prime to b (i.e. cO + b = O) then θb (c) restricts to the Hilbert
S-class field K trivially.
(b) If nonzero c1 , c2 ∈ O are both relatively prime to b then c1 ≡ c2 (mod b) is equivalent to
(7)
prb (jb (c1 )RS (b)) = prb (jb (c2 )RS (b))
where prb : Jk →
Y
kv× is the natural projection.
v∈V (b)
Proof. (a): Since c is relatively prime to b, we have jb (c) ∈ US . So, using the functoriality properties
of the norm residue map, we obtain
θb (c)|K = αK(b)/k (jb (c))−1 |K = αK/k (jb (c))−1 = idK
because jb (c) ∈ US ⊂ US = Ker αK/k , as required.
(b): As above, let (6) be the prime factorization of b, let vi = vpi ∈ V k \ S be the valuation
associated with pi . Then for any c1 , c2 ∈ O, the congruence c1 ≡ c2 (mod b) is equivalent to
c1 ≡ c2 (mod p̂nvii ) for all i = 1, . . . , t.
(8)
On the other hand, for any v ∈ Vfk and any u1 , u2 ∈ Uv , the congruence u1 ≡ u2 (mod p̂nv ) for n ≥ 1
is equivalent to
u1 Uv(n) = u2 Uv(n) ,
(n)
where Uv is the congruence subgroup of Uv modulo p̂nv . Thus, for (nonzero) c1 , c2 ∈ O prime to
b, the conditions (7) and (8) are equivalent, and our assertion follows.
We will now establish a result needed for the proof of Theorem 3.3 and its refinements.
Proposition 3.5. Let b be a nonzero ideal of O, let a ∈ O be relatively prime to b, and let F be a
finite Galois extension of Q that contains K(b). Assume that a rational prime p is unramified in F
10
A. MORGAN, A. RAPINCHUK, AND B. SURY
and there exists an extension w of the p-adic valuation vp to F such that FrF/Q (w|vp )|K(b) = θb (a).
If the restriction v of w to k does not belong to S ∪ V (b) then:
(a) kv = Qp ;
(b) the prime ideal p = pv of O corresponding to v is principal with a generator π satisfying
π ≡ a (mod b) and π > 0 in every real completion of k.
(We note since v is unramified in F which contains K(b), we in fact automatically have that
v∈
/ V (b).)
Proof. (a): Since the Frobenius Fr(w|vp ) generates Gal(Fw /Qp ), our claim immediately follows
from the fact that it acts trivially on k.
(b): According to (a), the local degree [kv : Qp ] is 1, hence the residual degree f (v|vp ) is also 1,
and therefore
Fr(w|v) = Fr(w|vp )f (v|vp ) = Fr(w|vp ).
Thus,
αK(b)/k (i(v)) = Fr(w|v)|K(b) = θb (a) = αK(b)/k (jb (a))−1 ,
and therefore
i(v)jb (a) ∈ Ker αK(b)/K = RS (b) = RS (b)k× .
So, we can write
(9)
Then
i(v)jb (a) = rπ with r ∈ RS (b), π ∈ k× .
π = i(v)(jb (a)r−1 ).
Since a is prime to b, the idele jb (a) ∈ US , and then jb (a)r−1 ∈ US . For any v ′ ∈ V k \ (S ∪ {v}), the
v ′ -component of i(v) is trivial, so we obtain that π ∈ Uv′ . On the other hand, the v-component of
i(v) is a uniformizer πv of kv implying that π is also a uniformizer. Thus, p = πO is precisely the
prime ideal associated with v. For any real v ′ , the v ′ -components of i(v) and jb (a) are trivial, so π
equals the inverse of the v ′ -component of r, hence positive in kv′ . Finally, it follows from (9) that
prb (jb (a)) = prb (jb (π)r),
so π ≡ a (mod b) by Lemma 3.4(b), as required.
P roof of T heorem 3.3. Set b = bO and σ = θb (a) ∈ Gal(K(b)/k). Let F be the Galois closure
of K(b) over Q, and let τ ∈ Gal(F/Q) be such that τ |K(b) = σ. Applying Chebotarev’s Density
Theorem (see [CF, Ch. VII, 2.4] or [BMS, A.6]), we find infinitely many rational primes p > 2 for
which the p-adic valuation vp is unramified in F , does not lie below any valuations in S ∪ V (b),
and has an extension w to F such that FrF/Q (w|vp ) = τ . Let v = w|k, and let p = pv be the
corresponding prime ideal of O. Since p > 2, part (a) of Proposition 3.5 implies that p is Q-split.
Furthermore, part (b) of it asserts that p has a generator π such that π ≡ a (mod b) and π > 0 in
every real completion of k, as required.
Remark. Dong Quan Ngoc Nguyen pointed out to us that Theorem 3.3, hence the essential part
of Dirichlet’s Theorem from [BMS] (in particular, (A.11)), was known already to Hasse [H, Satz
13]. In the current paper, however, we use the approach described in [BMS] to establish the key
Theorem 3.7; the outline of the constructions from [BMS] as well as the technical Lemma 3.4 and
Proposition 3.5 are included for this purpose. We note that in contrast to the argument in [BMS],
BOUNDED GENERATION OF SL2 (OS )
11
our proofs of Theorems 3.3 and 3.7 involve the application of Chebotarev’s Density Theorem to
noncommutative Galois extension.
We will now prove a statement from Galois theory that we will need in the next subsection.
Lemma 3.6. Let F/Q be a finite Galois extension, and let κ be an integer for which F ∩ Qab ⊆
Q(ζκ ). Then F (ζκ ) ∩ Qab = Q(ζκ ).
Proof. We need to show that
(10)
[F (ζκ ) : F (ζκ ) ∩ Qab ] = [F (ζκ ) : Q(ζκ )].
Let
G = Gal(F (ζκ )/Q) and H = Gal(F/Q).
Then the left-hand side of (10) is equal to the order of the commutator subgroup [G, G], while the
right-hand side equals
[F : F ∩ Q(ζκ )] = [F : F ∩ Qab ] = |[H, H]|.
Now, the restriction gives an injective group homomorphism
ψ : G → H × Gal(Q(ζκ )/Q).
Since the restriction G → H is surjective, we obtain that ψ implements an isomorphism between
[G, G] and [H, H] × {1}. Thus, [G, G] and [H, H] have the same order, and (10) follows.
3. Key statement. In this subsection we will establish another number-theoretic statement
which plays a crucial role in the proof of Theorem 1.1. To formulate it, we need to introduce some
additional notations. As above, let µ = |µ(k)| be the number of roots of unity in k, let K be the
Hilbert S-class field of k, and let K̃ be the Galois closure of K over Q. Suppose we are given two
finite sets P and Q of rational primes. Let
Y
µ′ = µ ·
p,
p∈P
pick an integer λ ≥ 1 which is divisible by µ and for which K̃ ∩ Qab ⊆ Q(ζλ ), and set
Y
q.
λ′ = λ ·
q∈Q
Theorem 3.7. Let u ∈ O× be a unit of infinite order such that u ∈
/ µ(k)p (k× )p for every prime
p ∈ P , and let q be a Q-split prime of O which is relatively prime to λ′ . Then there exist infinitely
many principal Q-split primes p = πO of O with a generator π such that
(1) for each p ∈ P , the p-primary component of φ(p)/µ divides the p-primary component of
the order of u (mod p);
(2) π(mod q2 ) generates (O/q2 )× ;
(3) gcd(φ(p), λ′ ) = λ.
12
A. MORGAN, A. RAPINCHUK, AND B. SURY
Proof. As in the proof of Theorem 3.3, we will derive the required assertion by applying Chebotarev’s Density Theorem to a specific automorphism of an appropriate finite Galois extension.
Let K(q2 ) be the abelian extension K(b) of k introduced in subsection 3.2 for the ideal b = q2 .
Set
√
′
L1 = K(q2 )(ζλ′ ), L2 = k ζµ′ , µ u , L = L1 L2 and ℓ = L1 ∩ L2 .
Then
(11)
Gal(L/k) = { σ = (σ1 , σ2 ) ∈ Gal(L1 /k) × Gal(L2 /k) | σ1 |ℓ = σ2 |ℓ }.
So, to construct σ ∈ Gal(L/k) that we will need in the argument it is enough to construct appropriate σi ∈ Gal(Li /k) for i = 1, 2 that have the same restriction to ℓ.
Lemma 3.8. The restriction maps define the following isomorphisms:
(1) Gal(L1 /K) ≃ Gal(K(q2 )/K) × Gal(K(ζλ′ )/K);
Y
Gal(Q(ζqλ )/Q(ζλ )).
(2) Gal(K(ζλ′ )/K(ζλ )) ≃ Gal(Q(ζλ′ )/Q(ζλ )) ≃
q∈Q
Proof. (1): We need to show that K(q2 ) ∩ K(ζλ ) = K. But the Galois extensions K(q2 )/K and
K(ζλ )/K are respectively totally and unramified at the extensions of vq to K (since q is prime to
λ), so the required fact is immediate.
(2): Since K(ζλ′ ) = K(ζλ ) · Q(ζλ′ ), we only need to show that
(12)
K(ζλ ) ∩ Q(ζλ′ ) = Q(ζλ ).
We have
K(ζλ ) ∩ Q(ζλ′ ) ⊆ K̃(ζλ ) ∩ Qab = Q(ζλ )
by Lemma 3.6. This proves one inclusion in (12); the other inclusion is obvious.
Since q is Q-split, the group (O/q2 )× is cyclic (Lemma 3.1(a)), and we pick c ∈ O so that c
(mod q2 ) is a generator of this group. We then set
σ1′ = θq2 (c) ∈ Gal(K(q2 )/K)
in the notations of subsection 3.2 (cf. Lemma 3.4(a)). Next, for q ∈ Q, we let q e(q) be the q-primary
component of λ. Then using the isomorphism from Lemma 3.8(2), we can find σ1′′ ∈ Gal(K(ζλ′ )/K)
such that
(13)
σ1′′ (ζλ ) = ζλ but σ1′′ (ζqe(q)+1 ) 6= ζqe(q)+1 for all q ∈ Q.
We then define σ1 ∈ Gal(L1 /K) to the automorphism corresponding to the pair (σ1′ , σ1′′ ) in terms
of the isomorphism from Lemma 3.8(1) (in other words, the restrictions of σ1 to K(q2 ) and K(ζλ′ )
are σ1′ and σ1′′ , respectively).
µ′ /ν
√
√
√
−1
′
′
We fix a µ′ -th root µ u, and for ν|µ′ set ν u = µ u
(also denoted uν ). To construct
σ2 ∈ Gal(L2 /k), we need the following.
Lemma 3.9. Let σ0 ∈ Gal(ℓ/k). Then there exists σ2 ∈ Gal(L2 /k) such that
(1) σ2 |ℓ = σ0 ;
BOUNDED GENERATION OF SL2 (OS )
13
(2) for any p ∈ P , if pd(p) is the p-primary component of µ then
−(d(p)+1)
−(d(p)+1)
6= up
,
σ2 up
and consequently either σ2 (ζpd(p)+1 ) 6= ζpd(p)+1 or σ2 acts nontrivially on all pd(p)+1 -th roots
of u.
Proof. Since L1 /k is an abelian extension, we conclude from Corollary 2.6 that
√
(14)
ℓ ⊆ k µ u, ζµ′ ⊆ kab .
√
On the other hand, according to Proposition 2.1, none of the roots pµ u for p ∈ P lies in kab , and
the restriction maps yield an isomorphism
√
Y
√
√
√
′
Gal k pµ u, ζµ′ /k µ u, ζµ′ .
Gal k µ u, ζµ′ /k µ u, ζµ′ →
p∈P
√
√
′
such that
It follows that for each p ∈ P we can find τp ∈ Gal k µ u, ζµ′ /k µ u, ζµ′
−(d(p)+1)
−(d(q)+1)
−(d(q)+1)
−(d(p)+1)
for all q ∈ P \ {p}.
= uq
and τp uq
= ζ p · up
τp up
Now, let σ̃0 be any extension of σ0 to L2 . For p ∈ P , define
1 , σ̃0 up−(d(p)+1) = up−(d(p)+1)
χ(p) =
0 , σ̃0 up−(d(p)+1) 6= up−(d(p)+1)
Set
σ2 = σ̃0 ·
Y
τpχ(p) .
p∈P
In view of (14), all τp ’s act trivially on ℓ, so σ2 |ℓ = σ̃0 |ℓ = σ0 and (1) holds. Furthermore, the
choice of the τp ’s and the χ(p)’s implies that (2) also holds.
Continuing the proof of Theorem 3.7, we now use σ1 ∈ Gal(L1 /k) constructed above, set σ0 =
σ1 |ℓ, and using Lemma 3.9 construct σ2 ∈ Gal(L2 /k) with the properties described therein. In
particular, part (1) of this lemma in conjunction with (11) implies that the pair (σ1 , σ2 ) corresponds
to an automorphism σ ∈ Gal(L/k). As in the proof of Theorem 3.3, we let F denote the Galois
closure of L over Q, and let σ̃ ∈ Gal(F/Q) be such that σ̃|L = σ. By Chebotarev’s Density Theorem,
there exist infinitely many rational primes p > 2 that are relatively prime to λ′ · µ′ and for which
the p-adic valuation vp is unramified in F , does not lie below any valuation in S ∪ {vq }, and has
an extension w to F such that FrF/Q (w|vp ) = σ̃. Let v = w|k, and let p = pv be the corresponding
prime ideal of O. As in the proof of Theorem 3.3, we see that p is Q-split. Furthermore, since
σ|K(q2 ) = θq2 (c), we conclude that p has a generator π such that π ≡ c (mod q2 ) (cf. Proposition
3.5(b)). Then by construction π (mod q2 ) generates (O/q2 )× , verifying condition (2) of Theorem
3.7.
To verify condition (1), we fix p ∈ P and consider two cases. First, suppose σ(ζpd(p)+1 ) 6= ζpd(p)+1 .
Since p is prime to p, this means that the residue field O/p does not contain an element of order
pd(p)+1 (although, since µ is prime to p, it does contain an element of order µ, hence of order
pd(p) ). So, in this case φ(p)/µ is prime to p, and there is nothing to prove. Now, suppose that
14
A. MORGAN, A. RAPINCHUK, AND B. SURY
σ(ζpd(p)+1 ) = ζpd(p)+1 . Then by construction σ acts nontrivially on every pd(p)+1 -th root of u, and
d(p)+1
− u has no roots in kvp . Again, since p is prime to p, we see
therefore the polynomial X p
from Hensel’s lemma that u (mod p) is not a pd(p)+1 -th power in the residue field. It follows that
the p-primary component of the order of u (mod p) is not less than the p-primary component of
φ(p)/pd(p) , and (1) follows.
Finally, by construction σ acts trivially on ζλ but nontrivially on ζqλ for any q ∈ Q. Since p is
prime to λ′ , we see that the residue field O/p contains an element of order λ, but does not contain
an element of order qλ for any q ∈ Q. This means that λ | φ(p) but φ(p)/λ is relatively prime to
each q ∈ Q, which is equivalent to condition (3) of Theorem 3.7.
4. Proof of Theorem 1.1
First, we will introduce some additional notations needed to convert the task of factoring a given
matrix A ∈ SL2 (O) as a product of elementary matrices into the task of reducing the first row of
A to (1, 0). Let
R(O) = {(a, b) ∈ O2 | aO + bO = O}
(note that R(O) is precisely the set of all first rows of matrices A ∈ SL2 (O)). For λ ∈ O, one defines
two permutations, e+ (λ) and e− (λ), of R(O) given respectively by
(a, b) 7→ (a, b + λa) and (a, b) 7→ (a + λb, b).
These permutations will be called elementary transformations of R(O). For (a, b), (c, d) ∈ R(O) we
n
will write (a, b) ⇒ (c, d) to indicate the fact that (c, d) can be obtained from (a, b) by a sequence
of n (equivalently, ≤ n) elementary transformations. For the convenience of further reference, we
will record some simple properties of this relation.
Lemma 4.1. Let (a, b) ∈ R(O).
n
n
(1a) If (c, d) ∈ R(O) and (a, b) ⇒ (c, d), then (c, d) ⇒ (a, b).
m
n
m+n
(1b) If (c, d), (e, f ) ∈ R(O) are such that (a, b) ⇒ (c, d) and (c, d) ⇒ (e, f ), then (a, b) ⇒ (e, f ).
1
(2a) If c ∈ O such that c ≡ a(mod bO), then (c, b) ∈ R(O), and (a, b) ⇒ (c, b).
1
(2b) If d ∈ O such that d ≡ b(mod aO), then (a, d) ∈ R(O), and (a, b) ⇒ (a, d).
n
(3a) If (a, b) ⇒ (1, 0) then any matrix A ∈ SL2 (O) with the first row (a, b) is a product of ≤ n + 1
elementary matrices.
n
(3b) If (a, b) ⇒ (0, 1) then any matrix A ∈ SL2 (O) with the second row (a, b) is a product of ≤ n + 1
elementary matrices.
2
(4a) If a ∈ O× then (a, b) ⇒ (0, 1).
2
(4b) If b ∈ O× then (a, b) ⇒ (1, 0).
Proof. For (1a), we observe that the inverse of an elementary transformation is again an elementary
transformation given by [e± (λ)]−1 = e± (−λ), so the required fact follows. Part (1b) is obvious.
n
(Note that (1) implies that the relation between (a, b) and (c, d) ∈ R(O) defined by (a, b) ⇒ (c, d)
for some n ∈ N is an equivalence relation.)
In (2a), we have c = a + λb with λ ∈ O. Then
cO + bO = aO + bO = O,
BOUNDED GENERATION OF SL2 (OS )
15
so (c, a) ∈ R(O), and e+ (λ) takes (a, b) to (c, b). The argument for (2b) is similar.
(3a) Suppose A ∈ SL2 (O) has the first row (a, b). Then for λ ∈ O, the first row of the product
AE12 (λ) is (a, b + λa) = e+ (λ)(a, b), and similarly the first row of AE21 (λ) is e− (λ)(a, b). So, the
n
fact that (a, b) ⇒ (1, 0) implies that there exists a matrix U ∈ SL2 (O) which is a product of n
elementary matrices and is such that AU has the first row (1, 0). This means that AU = E21 (z) for
some z ∈ O, and then A = E21 (z)U −1 is a product of ≤ n + 1 elementary matrices. The argument
for (3b) is similar.
Part (4a) follows since e− −a e+ a−1 (1 − b) (a, b) = (0, 1). The proof of (4b) is similar.
Remark. All assertions of Lemma 4.1 are valid over any commutative ring O.
Corollary 4.2. Let q be a principal Q-split prime ideal of O with generator q, and let z ∈ O be
such that z(mod q2 ) generates (O/q2 )× . Given an element of R(O) of the form (b, q n ) with n ≥ 2,
1
and an integer t0 , there exists an integer t ≥ t0 such that (b, q n ) ⇒ (z t , q n ).
Proof. By Lemma 3.1(b), the element z(mod qn ) generates (O/qn )× . Since b is prime to q, one can
find t ∈ Z such that b ≡ z t (mod qn ). Adding to t a suitable multiple of φ(qn ) if necessary, we can
assume that t ≥ t0 . Our assertion then follows from Lemma 4.1(2a).
Lemma 4.3. Suppose we are given (a, b) ∈ R(O), a finite subset T ⊆ Vfk , and an integer n 6= 0.
1
Then there exists α ∈ Ok and r ∈ O× such that V (α) ∩ T = ∅, and (a, b) ⇒ (αr n , b).
k we let m denote the maximal ideal of
Proof. Let hk be the class number of k. If for each v ∈ S \V∞
v
Ok corresponding to v, then the ideal (mv )hk is principal, and its generator πv satisfies v(πv ) = hk
k;
and w(πv ) = 0 for all w ∈ Vfk \ {v}. Let R be the subgroup of k× generated by πv for v ∈ S \ V∞
note that R ⊂ O× . We can pick r ∈ R so that a′ := ar −n ∈ Ok . We note that since a and b are
relatively prime in O, we have V (a′ ) ∩ V (b) ⊂ S.
Now, it follows from the strong approximation theorem that there exists γ ∈ Ok such that
k,
v(γb) ≥ 0 and v(γb) ≡ 0(mod nhk ) for all v ∈ S \ V∞
and v(γb) = 0 for all v ∈ V (a′ ) \ S.
k . Set
Then, in particular, we can find s ∈ R so that v(γbs−1 ) = 0 for all v ∈ S \ V∞
By construction,
(15)
γ ′ := γs−1 ∈ O and b′ := γ ′ b ∈ Ok .
k
v(b′ ) = 0 for all v ∈ V (a′ ) ∪ (S \ V∞
),
implying that V (a′ ) ∩ V (b′ ) = ∅, which means that a′ and b′ are relatively prime in Ok .
Again, by the strong approximation theorem we can find t ∈ Ok such that
v(t) = 0 for v ∈ T ∩ V (a′ ) and v(t) > 0 for v ∈ T \ V (a′ ).
Set α = a′ + tb′ ∈ Ok . Then for v ∈ T ∩ V (a′ ) we have v(a′ ) > 0 and v(tb′ ) = 0 (in view of (15)),
while for v ∈ T \ V (a′ ) we have v(a′ ) = 0 and v(tb′ ) > 0. In either case,
v(α) = v(a′ + tb′ ) = 0 for all v ∈ T,
i.e. V (α) ∩ T = ∅. On the other hand,
a + r n tγ ′ b = r n (a′ + tb′ ) = r n α,
16
A. MORGAN, A. RAPINCHUK, AND B. SURY
1
which means that (a, b) → (αr n , b), as required.
Recall that we let µ denote the number of roots of unity in k.
Lemma 4.4. Let (a, b) ∈ R(O) be such that a = α · r µ for some α ∈ Ok and r ∈ O× where V (α) is
disjoint from S ∪ V (µ). Then there exist a′ ∈ O and infinitely many Q-split prime principal ideals
3
q of O with a generator q such that for any m ≡ 1(mod φ(a′ O)) we have (a, b) ⇒ (a′ , q µm ).
Proof. The argument below is adapted from the proof of Lemma 3 in [CK1]. It relies on the
properties of the power residue symbol (in particular, the power reciprocity law) described in the
Appendix on Number Theory in [BMS]. We will work with all v ∈ V k (and not only v ∈ V k \ S),
so to each such v we associate a symbol (“modulus”) mv . For v ∈ Vfk we will identify mv with the
corresponding maximal ideal of Ok (obviously, pv = mv O for v ∈ V k \ S); the valuation ideal and
the group of units in the valuation ring Ov (or Omv ) in the completion kv will be denoted m̂v and
Uv respectively. For any divisor ν|µ, we let
∗, ∗
mv ν
be the (bi-multiplicative,
skew-symmetric)
power residue symbol of degree ν on kv× (cf. [BMS, p.
x, y
= 1 if one of the elements x, y is a ν-th power in kv× (in particular, if v
85]). We recall that
mv ν
is either complex, or is real and one of the elements x, y is positive
in kv ) or if v is nonarchimedean
x, y
= 1 for almost all v ∈ V k .
∈
/ V (ν) and x, y ∈ Uv . It follows that for any x, y ∈ k× , we have
mv ν
Furthermore, we have the reciprocity law:
Y x, y
(16)
= 1.
mv ν
k
v∈V
pe11
· · · penn
Now, let µ =
be a prime factorization of µ. For each i = 1, . . . , n, pick vi ∈ V (pi ).
According to [BMS, (A.17)], the values
x, y
for x, y ∈ Uvi
mvi pei
i
cover all pei i -th roots of unity. Thus, we can pick units ui , u′i ∈ Uvi for i = 1, . . . , n so that
ui , u′i
= ζpei ,
i
mvi pei
i
a primitive pei i -th root of unity. Then
(17)
ζµ :=
n
Y
ui , u′
i
i=1
mvi
e
pi i
is a primitive µ-th root of unity. Furthermore, it follows from the Inverse Function Theorem or
Hensel’s Lemma that we can find an integer N > 0 such that
(18)
×
1 + m̂N
v ⊂ kv
µ
for all v ∈ V (µ).
BOUNDED GENERATION OF SL2 (OS )
17
We now write b = βtµ with β ∈ Ok and t ∈ O× . Since a, b are relatively prime in O, so are α, β,
hence V (α) ∩ V (β) ⊂ S. On the other hand, by our assumption V (α) is disjoint from S ∪ V (µ),
so we conclude that V (α) is disjoint from V (β) ∪ V (µ). Applying Theorem 3.3 to the ring Ok we
obtain that there exists β ′ ∈ Ok having the following properties:
(1)1
(2)1
(3)1
(4)1
b := β ′ Ok is a prime ideal of Ok and the corresponding valuation vb ∈
/ S ∪ V (µ);
′
β > 0 in every real completion of k;
β ′ ≡ β(mod αOk );
for each i = 1, . . . , n, we have
β ′ ≡ u′i (mod m̂vi ), and
β ′ ≡ 1(mod m̂v ) for all v ∈ V (pi ) \ {vi }.
Set b′ = β ′ tµ . It is a consequence of (3)1 that b ≡ b′ (mod aO), so by Lemma 4.1(2) we have
1
µ
(a, b) ⇒ (a, b′ ). Furthermore, it follows from (4)1 and (18) that β ′ /u′i ∈ kv×i , so
ui , β ′
ui , u′i
=
= ζ ei .
mvi pei
mvi pei pi
i
i
Since ζµ defined by (17) is a primitive µ-th root of unity, we can find an integer d > 0 such that
Y
n d
ui , β ′
α, β ′
α, β ′
d
·ζ =
·
.
(19)
1=
b µ µ
b µ
mvi pei
i=1
i
By construction, vb ∈
/ V (α) ∪ V (µ), so applying Theorem 3.3 one more time, we find α′ ∈ Ok
such that
(1)2 a := α′ Ok is a prime ideal of Ok and the corresponding valuation va ∈
/ S ∪ V (µ);
′
(2)2 α ≡ α(mod b);
(3)2 α′ ≡ udi (mod m̂N
vi ) for i = 1, . . . , n.
1
Set a′ = α′ r µ . Then a′ O= α′ Ois a prime ideal of O and a′ ≡ a(mod b′ O), so (a, b′ ) ⇒ (a′ , b′ ).
α′ , β ′
k (since β ′ > 0 in all real completions of
= 1 if either v ∈ V∞
Now, we note that
mv µ
k) or v ∈ Vfk \ (V (α′ ) ∪ V (β ′ ) ∪ V (µ)). Since the ideals a = α′ Ok and b = β ′ Ok are prime by
construction, we have V (α′ ) = {va } and V (β ′ ) = {vb }. Besides,
follows from (18) and (4)1 that
′ it
′
α
,
β
µ
for v ∈ V (pi ) \ {vi } we have β ′ ∈ kv× , and therefore again
= 1. Thus, the reciprocity law
mv µ
(16) for α′ , β ′ reduces to the relation
′ ′ ′ ′ Y
n ′
α , β′
α ,β
α ,β
= 1.
·
·
(20)
a µ
b µ
mvi µ
i=1
It follows from (2)2 and (3)2 that
′ ′
′ ′ d ′
α ,β
ui , β
α, β ′
α ,β
=
=
and
for all i = 1, . . . , n.
b µ
b µ
mvi µ
b µ
18
A. MORGAN, A. RAPINCHUK, AND B. SURY
Comparing now (19) with (20), we find that
′ ′ ′ ′ −1
β ,α
α ,β
= 1.
=
a µ
a µ
This implies (cf. [BMS, (A.16)]) that β ′ is a µ-th power modulo a, i.e. β ′ ≡ γ µ (mod a) for some
γ ∈ Ok . Clearly, the elements a′ = α′ r µ and γt are relatively prime in O, so applying Theorem
3.3 to this ring, we find infinitely many Q-split principal prime ideals q of O having a generator
q ≡ γt(mod a′ O). Then for any m ≡ 1(mod φ(a′ O)) we have
1
q µm ≡ q µ ≡ β ′ tµ ≡ b′ (mod a′ O),
3
so (a′ , b′ ) ⇒ (a′ , q µm ). Then by Lemma 4.1(1b), we have (a, b) ⇒ (a′ , q µm ), as required.
The final ingredient that we need for the proof of Theorem 1.1 is the following lemma which uses
the notion of the level ℓp (u) of a unit u of infinite order with respect to a Q-split ideal p introduced
in §3.1.
Lemma 4.5. Let p be a principal Q-split ideal of O with a generator π, and let u ∈ O× be a unit of
infinite order. Set s = ℓp (u), and let λ and m be integers satisfying λ|φ(p) and m ≡ 0(modφ(ps )/λ).
Given an integer δ > 0 dividing λ and b ∈ O prime to π such that b is a δ-th power (mod p) while
ν := λ/δ divides the order of u(mod p), for any integer t ≥ s there exists an integer nt for which
1
(π t , bm ) ⇒ (π t , unt ).
Proof. Let p be the rational prime corresponding to p. Being a divisor of λ, the integer δ is relatively
prime to p. So, the fact that b is a δ-th power mod p implies that it is also a δ-th power mod ps . On
the other hand, it follows from our assumptions that λm = δνm is divisible by φ(ps ), and therefore
(bm )ν ≡ 1(mod ps ). But since ν is prime to p, the subgroup of elements in (O/ps )× of order dividing
ν is isomorphic to a subgroup of (O/p)× , hence cyclic. So, the fact that the order of u(mod p), and
consequently the order u(mod ps ), is divisible by ν implies that every element in (O/ps )× whose
order divides ν lies in the subgroup generated by u(mod ps ). Thus, bm ≡ uns (mod ps ) for some
integer ns . Since p is Q-split, we can apply Lemma 3.2 to conclude that for any t ≥ s there exists
1
an integer nt such that bm ≡ unt (mod pt ). Then (π t , bm ) ⇒ (π t , unt ) by Lemma 4.1(2).
We will call a unit u ∈ O× fundamental if it has infinite order and the cyclic group hui is a direct
factor of O× . Since the group O× is finitely generated (Dirichlet’s Unit Theorem, cf. [CF, §2.18])
it always contains a fundamental unit once it is infinite. We note that any fundamental unit has
the following property:
u∈
/ µ(k)p (k× )p for any prime p.
We are now in a position to give
Proof of Theorem 1.1. We return to the notations of §3.3: we let K denote the Hilbert S-class
field of k, let K̃ be its normal closure over Q, and pick an integer λ ≥ 1 which is divisible by µ
and for which K̃ ∩ Qab ⊂ Q(ζλ ). Furthermore, since O× is infinite by assumption, we can find a
fundamental unit u ∈ O× . By Lemma 4.1(3), it suffices to show that for any (a, b) ∈ R(O), we have
(21)
8
(a, b) ⇒ (1, 0).
BOUNDED GENERATION OF SL2 (OS )
19
k ) ∪ V (µ) and n = µ, we see that there exist α ∈ O
First, applying Lemma 4.3 with T = (S \ V∞
k
×
and r ∈ O such that
1
V (α) ∩ (S ∪ V (µ)) = ∅ and (a, b) ⇒ (αr µ , b).
Next, applying Lemma 4.4 to the last pair, we find a′ ∈ O and a Q-split principal prime ideal q
3
such that vq ∈
/ S ∪ V (λ) ∪ V (φ(a′ O)) and (αr µ , b) ⇒ (a′ , q µm ) for any m ≡ 1(mod φ(a′ O). Then
(22)
4
(a, b) ⇒ (a′ , q µm ) for any m ≡ 1(mod φ(a′ O)).
To proceed with the argument we will now specify m. We let P and Q denote the sets of prime
divisors of λ/µ and φ(a′ O), respectively, and define λ′ and µ′ as in §3.3; we note that by construction
q is relatively prime to λ′ . So, we can apply Theorem 3.7 which yields a Q-split principal prime
ideal p = πO so that vp ∈
/ V (φ(a′ O)) and conditions (1) - (3) are satisfied. Let s = ℓp (u) be the
p-level of u. Condition (3) implies that
gcd(φ(p)/λ, λ′ /λ) = 1 = gcd(φ(p)/λ, φ(a′ O))
since λ′ /λ is the product of all prime divisors of φ(a′ O). It follows that the numbers φ(ps )/λ and
φ(a′ O) are relatively prime, and therefore one can pick a positive integer m so that
m ≡ 0(mod φ(ps )/λ) and m ≡ 1(mod φ(a′ O)).
Fix this m for the rest of the proof.
Condition (2) of Theorem 3.7 enables us to apply Corollary 4.2 with z = π and t0 = s to find
1
t ≥ s so that (a′ , q µm ) ⇒ (π t , q µm ). Since P consists of all prime divisors of λ/µ, condition (1) of
Theorem 3.7 implies that λ/µ divides the order of u(mod p). Now, applying Lemma 4.5 with δ = µ
1
and b = q µ , we see that (π t , q µm ) ⇒ (π t , unt ) for some integer nt . Finally, since u is a unit, we
2
have (π t , unt ) ⇒ (1, 0). Combining these computations with (22), we obtain (21), completing the
proof.
Corollary 4.6. Assume that the group O× is infinite. Then for n ≥ 3, any matrix A ∈ SLn (O) is
a product of ≤ 21 (3n2 − n) + 4 elementary matrices.
Proof. Indeed, since the ring O is Dedekind, it is well-known and easy to show that any A ∈ SLn (O)
can be to a matrix in SL2 (O) by at most 21 (3n2 − n) − 5 (cf. [CK1, p. 683]). Now, our result
immediately follows from Theorem 1.1.
Proof of Corollary 1.2. Let
e+ : α 7→
1 α
0 1
and e− : α 7→
1 0
α 1
be the standard 1-parameter subgroups. Set U ± = e± (O). In view of Theorem 1.1, it is enough
to show that each of the subgroups U + and U − is contained in a product of finitely many cyclic
subgroups of SL2 (O). Let hk be the class number of k. Then there exists t ∈ O× such that v(t) = hk
k
/ S. Then O = Ok [1/t]. So, letting U0± = e± (Ok ) and
for all
v ∈ S \V∞ and v(t) = 0 for all v ∈
t 0
, we will have the inclusion
h=
0 t−1
U ± ⊂ hhi U0± hhi.
20
A. MORGAN, A. RAPINCHUK, AND B. SURY
On the other hand, if w1 , . . . , wn (where n = [k : Q]) is a Z-basis of Ok then U0± = he± (w1 )i · · · he± (wn )i,
hence
(23)
as required.
U ± ⊂ hhihe± (w1 )i · · · he± (wn )ihhi,
Remark. 1. Quantitatively, it follows from the proof of Theorem 1.1 that SL2 (O) = U − U + · · · U −
(nine factors), so since the right-hand side of (23) involves n + 2 cyclic subgroups, with hhi at both
ends, we obtain that SL2 (O) is a product of 9[k : Q] + 10 cyclic subgroups. Also, it follows from
[Vs] that SL2 (Z[1/p]) is a product of 11 cyclic subgroups.
k , then the proof of Corollary 1.2 yields a factorization of SL (O) as a finite
2. If S = V∞
2
product hγ1 i · · · hγd i of cyclic subgroups where all generators γi are elementary matrices, hence
k , the factorization we produce involves some diagonal
unipotent. On the contrary, when S 6= V∞
(semisimple) matrices. So, it is worth pointing out in the latter case there is no factorization with
k and let γ ∈ SL (O) be unipotent. Then there exists
all γi unipotent. Indeed, let v ∈ S \ V∞
2
N = N (γ) such that for any a = (aij ) ∈ hγi we have v(aij ) ≤ N (γ) for all i, j ∈ {1, 2}. It follows
that if SL2 (O) = hγ1 i · · · hγd i where all γi are unipotent, then there exists N0 such that for any
a = (aij ) ∈ SL2 (O) we have v(aij ) ≤ N0 for i, j ∈ {1, 2}, which is absurd.
5. Example
For a ring of S-integers O in a number field k such that the group of units O× is infinite, we
let ν(O) denote the smallest positive integer with the property that every matrix in SL2 (O) is a
product of ≤ ν(O) elementary matrices. So, the result of [Vs] implies that ν(Z[1/p]) ≤ 5 for any
prime p, and our Theorem 1.1 yields that ν(O) ≤ 9 for any O as above. It may be of some interest
to determine the exact value of ν(O) in some situations. In Example 2.1 on p. 289, Vsemirnov
claims that the matrix
5 12
M=
12 29
is not a product of four elementary matrices in SL2 (Z[1/p]) for any p ≡ 1(mod 29), and therefore
ν(Z[1/p]) = 5 in this case. However this example is faulty because for any prime p, in SL2 (Z[1/p])
we have
2
5 12
1 0
1 2
M=
=
·
12 29
2 1
0 1
However, it turns out that the assertion that ν(Z[1/p]) = 5 is valid not only for p ≡ 1(mod 29) but
in fact for all p > 7. More precisely, we have the following.
Proposition 5.1. Let O = Z[1/p], where p is prime > 7. Then not every matrix in SL2 (O) is a
product of four elementary matrices.
In the remainder of this section, unless stated otherwise, we will work with congruences over the
ring O rather than Z, so the notation a ≡ b(mod n) means that elements a, b ∈ O are congruent
modulo the ideal nO. We begin the proof of the proposition with the following lemma.
Lemma 5.2. Let O = Z[1/p], where p is any prime, and let r be a positive integer satisfying
p ≡ 1(mod r). Then any matrix A ∈ SL2 (O) of the form
1 − pα
∗
, α, β ∈ Z
(24)
A=
∗
1 − pβ
BOUNDED GENERATION OF SL2 (OS )
21
which is a product of four elementary matrices, satisfies the congruence
0 1
A≡±
(mod r).
−1 0
Proof. We note right away that the required congruence is obvious for the diagonal entries, so we
only need to establish it for the off-diagonal ones. Since A is a product of four elementary matrices,
it admits one of the following presentations:
(25)
A = E12 (a)E21 (b)E12 (c)E21 (d),
or
(26)
A = E21 (a)E12 (b)E21 (c)E12 (d),
with a, b, c, d ∈ O.
First, suppose we have (25). Then reducing it modulo bO, we obtain the following congruence:
1 + d(a + c) a + c
A ≡ E12 (a + c)E21 (d) =
(mod b).
d
1
Looking back at (24) and comparing the (2, 2) entries, we obtain that 1 − pβ ≡ 1(mod b). It follows
that b ∈ O× , i.e. b = ±pγ for some integer γ. Similarly, one shows that c = ±pδ for some integer δ.
Next, we observe that the signs involved in b and c must be different. Indeed, otherwise we
would have
∗
∗
γ
δ
,
A = E12 (a)E21 (±p )E12 (±p )E21 (d) =
∗ 1 + pγ+δ
which is inconsistent with (24). Thus, A looks as follows:
∗
a(1 − pγ+δ ) ∓ pδ
γ
δ
A = E12 (a)E21 (±p )E12 (∓p )E21 (d) =
.
d(1 − pγ+δ ) ± pδ
∗
Consequently, the required congruences for the off-diagonal entries immediately follow from the
fact that p ≡ 1(mod r), proving the lemma in this case.
Now, suppose we have (26). Then
A−1 = E12 (−d)E21 (−c)E12 (−b)E21 (−a),
which means that A−1 has a presentation of the form (25). Since the required congruence in this
case has already been established, we conclude that
0 −1
−1
A ≡±
(mod r).
1 0
But then we have
A≡±
as required.
0 1
−1 0
(mod r),
To prove the proposition, we will consider two cases.
Case 1. p − 2 is composite. Write p − 2 = r1 · r2 , where r1 , r2 are positive integers > 1, and set
r = p − 1. Then
(27)
ri 6≡ ±1(mod r) for i = 1, 2.
22
A. MORGAN, A. RAPINCHUK, AND B. SURY
√
Indeed, we can assume that r2 ≤ p − 2. If r2 ≡ ±1(modr)
√ then r2 ∓ 1 would be a nonzero integral
multiple of r, and therefore r ≤ r2 + 1, hence p − 2 ≤ p − 2. But this is impossible since p > 3.
Thus, r2 6≡ ±1(mod r). Since r1 · r2 ≡ −1(mod r), condition (27) follows.
Now, consider the matrix
1 − p r1 · p
A=
r2
1−p
One immediately checks that A ∈ SL2 (O). At the same time, A is of the form (24). Then Lemma
5.2 in conjunction with (27) implies that A is not a product of four elementary matrices.
Case 2. p and p − 2 are both primes. In this paragraph we will use congruences in Z. Clearly,
a prime > 3 can only be congruent to ±1(mod 6). Since p > 5 and p − 2 is also prime, in our
situation we must have p ≡ 1(mod 6). Furthermore, since p > 7, the congruence p ≡ 0 or 2(mod 5)
is impossible. Thus, in the case at hand we have
p ≡ 1, 13, or 19(mod 30).
p3
If p ≡ 13(mod 30) in Z, then
≡ 7 (mod 30), and therefore p3 − 2 is an integral multiple of 5.
Set r = p − 1 and s = (p3 − 2)/5, and consider the matrix
1 − p3
5p3
A=
s
1 − p3
Then A is a matrix in SL2 (O) having form (24). Note that 5p3 ≡ 5(mod r), which is different from
±1(mod r) since r > 6. Now, it follows from Lemma 5.2 that A is not a product of four elementary
matrices.
It remains to consider the case where p ≡ 1 or 19(mod 30). Consider the following matrix:
900 53 · 899
A=
,
17
900
and note that A ∈ SL2 (Z) and
−1
A
=
900 −53 · 899
−17
900
.
It suffices to show that neither A nor A−1 can be written in the form
∗
c + a(1 + bc)
(28)
E12 (a)E21 (b)E12 (c)E21 (d) =
, with a, b, c, d ∈ O
b + d(1 + bc)
(1 + bc)
Set
t = b + d(1 + bc) and u = c + a(1 + bc).
Assume that either A or A−1 is written in the form (28). Then 1 + bc = 900, so
b, c ∈ {±pn , ±29pn , ±31pn , ±899pn | n ∈ N}.
On the other hand, we have the following congruences
t ≡ b(mod 30) and u ≡ c(mod 30).
Analyzing the above list of possibilities for b and c, we conclude that each of t and u is ≡
±pn (mod 30) for some integer n. Thus, if p ≡ 1(mod 30) then t, u ≡ ±1(mod 30), and if p ≡
19(mod 30) then t, u ≡ ±1, ±19(mod 30). Since 17 6≡ ±1, ±19(mod 30), we obtain a contradiction
BOUNDED GENERATION OF SL2 (OS )
23
in either case. (We observe that the argument in this last case is inspired by Vsemirnov’s argument
in his Example 2.1.)
Acknowledgements. The second author was supported by the Simons Foundation. The third author
gratefully acknowledges fruitful conversations with Maxim Vsemirnov. We thank Dong Quan Ngoc Nguyen
for his comments regarding Theorem 3.3.
References
[AT] E. Artin, J. Tate, Class Field Theory, AMS, 1967.
[CK1] D. Carter, G. Keller, Bounded Elementary Generation of SLn (O), Amer. J. Math. 105(1983), No. 3, 673-687.
[CK2] D. Carter, G. Keller, Elementary expressions for unimodular matrices, Commun. Alg. 12(1984), No. 4, 379-389.
[BMS] H. Bass, J. Milnor, J.-P. Serre, Solution of the congruence subgroup problem for SLn (n ≥ 3) and Sp2n (n ≥ 2),
Publ. math. I.H.É.S. 33(1967), 59-137.
[CF] J. W. S. Cassels, A. Fröhlich, Algebraic Number Theory, Thompson Book Company Inc., 1967.
[Co] P. M. Cohn, On the structure of the GL2 of a ring, Publ. math. I.H.É.S. 30(1966), 5-53.
[CW] G. Cooke, P.J. Weinberger, On the construction of division chains in algebraic number rings, with applications
to SL2 , Commun. Alg. 3(1975), No. 6, 481-524.
[ER] I. V. Erovenko, A. S. Rapinchuk, Bounded generation of S-arithmetic subgroups of isotropic orthogonal groups
over number fields, J. Number Theory 119(2006), No. 1, 28-48.
[GS] F.J. Grunewald, J. Schwermer, Free Non-abelian Quotients of SL2 Over Orders of Imaginary Quadratic Numberfields, J. Algebra 69(1981), 298-304.
[H] H. Hasse, Bericht über neuere Untersuchungen und Probleme aus der Theorie der algebraischen Zahlkörper I,
Jber. Deutsch. Math.-Verein. 35(1926), 1-55.
[HB] D.R. Heath-Brown, Artin’s conjecture for primitive roots, Quart. J. Math. Oxford (2), 37 (1986), 27-38.
[La] S. Lang, Algebra, Springer, 2002.
[LM] D. Loukanidis, V.K. Murty, Bounded generation for SLn (n ≥ 2) and Sp2n (n ≥ 1), preprint (1994).
[L1] B. Liehl, On the group SL2 over orders of arithmetic type., J. Reine Angew. Math. 323(1981), 153-171.
[L2] B. Liehl, Beschränkte Wortelänge in SL2 , Math. Z. 186(1984), 509-524.
[Lu] A. Lubotzky, Subgroup growth and congruence subgroups, Invent. Math. 119(1995), 267-295.
[MCKP] D.W. Morris, Bounded generation of SL(n, A) (after D. Carter, G. Keller and E. Paige), New York
J. Math. 13(2007), 383-421.
[M] V.K. Murty, Bounded and finite generation of arithmetic groups. (English summary), Number theory (Halifax,
NS, 1994), CMS Conf. Proc. 15, Amer. Math. Soc., Providence, RI, (1995), 249-261.
[MP] M.R. Murty, K.L. Petersen, The generalized Artin conjecture and arithmetic orbifolds, Groups and symmetries,
CRM Proceeding & Lecture Notes 47, Amer. Math. Soc., Providence, RI, (2009), 259-263.
[PR1] V.P. Platonov, A.S. Rapinchuk, Abstract properties of S-arithmetic groups and the congruence subgroup problem, Izv. Ross. Akad. Nauk. Ser. Mat. 56(1992), 483-508.
[PR2] V.P. Platonov, A.S. Rapinchuk, Algebraic Groups and Number Theory, Academy of Sciences, 1991.
[R] A.S. Rapinchuk, Representations of groups of finite width, Soviet Math. Dokl. 42(1991), 816-820.
[S1] J.-P. Serre, Le problème des groupes de congruence pour SL2 , Ann. Math., Second Series, 92(1970), No. 3,
489-527.
[S2] J.-P. Serre, Local Fields, Springer-Verlag, 1979.
[SW] Y. Shalom, G. Willis, Commensurated subgroups of arithmetic groups, totally disconnected groups and adelic
rigidity, Geom. Funct. Anal., 23(2013), No. 5, 1631-1683.
[T] O.I. Tavgen, Bounded generation of Chevalley groups over rings of algebraic S-integers, Izv. Akad. Nauk SSSR
Ser. Mat. 54(1990), No. 1, 97-122.
[Va] L.N. Vaserštein, The group SL2 over Dedekind rings of arithmetic type, Math. USSR- Sb. 18(1972), No. 2,
321-332.
[Vs] M. Vsemirnov, Short unitriangular factorizations of SL2 (Z[1/p]), Quart. J. Math., 65 (2014), 279-290.
24
A. MORGAN, A. RAPINCHUK, AND B. SURY
Department of Mathematics, University of Virginia, Charlottesville, VA 22904-4137, USA
E-mail address: [email protected]
E-mail address: [email protected]
Stat-Math Unit, Indian Statistical Institute, 8th Mile Mysore Road, Bangalore 560059, India
E-mail address: [email protected]
| 4math.GR
|
arXiv:1512.08555v1 [cs.DS] 28 Dec
1
Maximium Priority Matchings
Jonathan Turner
wucse-2015-06
Abstract
Let G = (V, E) be an undirected graph with n vertices and m edges,
in which each vertex u is assigned an integer priority in [1, n], with 1
being the “highest” priority. Let M be a matching of G. We define
the priority score of M to be an n-ary integer in which the i-th mostsignificant digit is the number of vertices with priority i that are incident
to an edge in M . We describe a variation of the augmenting path
method (Edmonds’ algorithm) that finds a matching with maximum
priority score in O(mn) time.
Introduction
A matching in an undirected graph is a subset of its edges, no two of which
share a common endpoint. In the maximum size matching problem, the
objective is to find a matching with the largest possible number of edges.
Edmonds showed how to solve the problem for general graphs [1] and more
efficient implementations of his method are described in [2] and [5]. Hopcroft
and Karp described a simpler algorithm for the case of bipartite graphs [4].
A matching is said to match a vertex u, if one of its edges is incident
to u. Given a graph in which vertices are assigned integer priorities ρ(u), a
maximum priority matching is one that maximizes the number of matched
vertices in the highest priority class, then maximizes the number of matched
vertices in the next priority class (without reducing the number matched in
the highest priority class), and so forth. We define a matching’s priority
1
e
2
5 c
a
3
b 2
4 f
d
g 4
1
h
6
i
1 j
8
priority score: 2111000100
Figure 1: Example showing the priority score for a matching
score to be the n-ary number in which the i-th most-significant digit is the
number of matched vertices with priority i. Figure 1 shows an example of
a graph with a matching whose priority score is 2111000100. Adding the
edge ac yields a matching with a score of 2211100100. A maximum priority
matching is a matching that has a priority score with maximum value. This
version of the matching problem arises as a subproblem in an approximation
algorithm for an np-complete scheduling problem for crossbar switches used
in internet routers [7].
In this paper, we show how Edmonds’ algorithm for the ordinary matching problem can be extended to solve priority matching problems. In section
2, we focus on the special case of two priorities (ρ(u) ≤ 2 for all u). This
problem is of independent interest and provides a useful introduction to the
methods used in the more general case, which is discussed in section 3.
2
Two priority case
The two priority case can be phrased more simply by defining a set S consisting of all vertices with priority 1. Our objective is then to find a maximum
2
size matching that also matches the largest possible number of vertices in S.
In Edmonds’ algorithm for maximum size matching, augmenting paths
are used to convert a matching to a larger one. Given a matching in a graph,
an augmenting path is a simple path in which the edges alternate between
matching edges and non-matching edges, with both endpoints unmatched.
By reversing the status of the matching and non-matching edges on such a
path, we can obtain a new matching with one more edge than the original.
Thus, so long as we can find an augmenting path, we can expand a given
matching. Edmonds showed that if a larger matching exists, the graph must
contain an augmenting path.
For the priority matching problem, we must adjust the definition of an
augmenting path. Let M be a matching of a graph G that does not match
all the vertices in the given set S. An augmenting path in G with respect
to M is a path p = u0 , . . . , ut in which every other edge belongs to M , u0 is
unmatched, u0 ∈ S and if ut is matched, ut 6∈ S.
Figure 2 shows examples of two such paths. Observe that in both cases,
if one replaces the path edges in M with the path edges not in M , we get a
new matching that matches at least one more vertex in S. Also, note that
when we expand a matching in this way, all previously matched vertices in
S remain matched.
Figure 2: Augmenting paths for two priority case (shaded vertices are in S)
Define the count of a two-priority matching to be the number of vertices in
S that are incident to edges in the matching. Now consider two matchings M
and M 0 , where M has a smaller count than M 0 . Let N be the graph defined
by the edges that are in M or in M 0 , but not both. Note that N consists
of a collection of disjoint paths and cycles with alternate edges in M and
M 0 . Since M 0 has a larger count than M , N must contain at least one path
p in which M 0 matches more elements of S than does M . Since all interior
vertices of a path are matched by both M and M 0 , the path endpoints must
3
account for the difference in the matchings’ counts. That means that at
least one endpoint must be in S and unmatched by M . If the other endpoint
is matched, it cannot also be in S. Hence, p satisfies the condtion for an
augmenting path with respect to M . Thus, for any matching that does not
have the largest possible count, there exists an augmenting path that can be
used to obtain a new matching with a larger count.
Observe that given a matching with the maximum possible count, but
less than maximum size, the original augmenting path method can be used to
obtain a matching with the same count, but one more more matching edge.
Hence one can easily convert a maximum count matching to one that has
both maximum count and size. Such a matching also maximizes the number
of matched vertices that are not in S, hence it satisfies the definition of a
maximum priority matching for the 2 priority case.
To complete the description of the augmenting path method, we still need
an algorithm to find an augmenting path. Our presentation is an adaptation
of that given in [6] for the maximum size matching problem. We start with
the special case of bipartite graphs.
The algorithm finds an augmenting path by building a collection of trees
rooted at unmatched vertices in S. Vertices that have not yet been added
to a tree are called unreached, while a vertex u in a tree is called odd or
even depending on the length of the tree path from u to the root of the
tree. Initially, the unmatched vertices in S are the only tree vertices. The
algorithm also maintains a list of eligible edges that initially contains all
edges incident to tree roots. It then repeats the following step until it either
finds an augmenting path or runs out of eligible edges.
Select an eligible edge e = {u, v} for which u is even, remove it from
the eligible list, then apply the applicable case from those listed below.
• If v is unreached and matched, let {v, w} be the matching edge
incident to v. Extend the tree containing u by making v a child
of u and w a child of v. If w is not in S, then the path from w
to the root of its tree is an augmenting path; otherwise, add all
non-matching edges incident to w to the eligible list.
4
b
h
c
g +
a +
g +
a +
b -
c -
e +
f
h -
d
d
i
e
eligible: ab, ac, gh
i
j
f
k
m
n
p
j +
+
eligible: bf, fi, jk, jm
k
m
n
p
Figure 3: Example of bipartite algorithm at start of augmenting path search
and after three steps (even vertices labelled +, odd vertices − and arrows
point to parents in trees)
• If v is unreached and unmatched then the path consisting of e
plus the tree path from u to its tree root is an augmenting path.
• If v is even then the path formed by combining e with the tree
path from u to the root of its tree and the tree path from v to the
root of its tree is an augmenting path. (Note that u and v are in
different trees, since the graph is bipartite.)
• If v is odd, ignore e and proceed to the next eligible edge.
Figure 3 illustrates the operation of the algorithm. Note that if edge bf is
selected next, the algorithm will ignore it, if fi is selected next, it will find
the odd-length augmenting path acfi, and if edge jk is selected next, it will
find the even-length augmenting path ghjkn.
To establish the correctness of the algorithm, we show that if it halts
without finding an augmenting path, then the graph must not contain one.
We start by noting a few properties of the algorithm.
1. Every tree root is unmatched and in S and each tree has exactly one
unmatched vertex.
5
2. Every matching edge either has both endpoints unreached, or one odd
and one even. In the latter case, the even endpoint is the child of the
odd endpoint in some tree.
3. Every edge joining two vertices in the same tree joins an odd vertex to
an even vertex.
4. If the algorithm fails to find an augmenting path, then when it fails,
every edge with an endpoint in some tree has at least one odd endpoint.
5. If the algorithm fails to find an augmenting path, then when it fails,
every even vertex is in S.
Properties 3 and 4 imply that if the algorithm fails to find an augmenting,
any path u0 , u1 , . . . , ut with u1 a child of u0 that alternates between matching
and non-matching edges, must also alternate between even and odd vertices.
So in particular, if the algorithm fails to find an augmenting path, but the
graph contains an augmenting path p = u0 , . . . , ut with u0 unmatched and
in S, all vertices in p must alternate between even and odd. If t is odd,
this implies that ut is odd and unmatched, but this contradicts the fact that
every odd vertex is matched. If t is even, then ut is even and matched, but
this contradicts property 5, since the matched endpoint of an augmenting
path cannot be in S.
Before proceeding to the case of general graphs, we show that for the
special case where S consists of all vertices of maximum degree, the algorithm
finds a matching that covers all vertices in S. Assume, to the contrary, that
the algorithm fails to find an augmenting path when there is some unmatched
vertex in S. Consider the collection of trees at the time the algorithm halts
and recall that by property 5, all the even vertices must be in S. Since every
tree has one more even vertex than it has odd vertices, some even vertex must
have an edge that connects it to an unreached vertex, but this contradicts
property 4. Hence, the algorithm matches all vertices in S.
Next, we show how to find augmenting paths in general graphs. As
with ordinary matchings, the key issue is handling odd-length cycles, known
as blossoms. Edmonds showed how to extend the ordinary augmenting path
search to recognize blossoms and shrink each blossom down to a single vertex,
6
a +
a +
b -
c +
- d
f -
e
+
g
+
h +
b -
i -
c +
- d
j +
k
-
m
+
+
e
h +
B1
i B
j + 2
f -
g
+
k
-
m+
Figure 4: Example of general graph with blossoms
producing a new graph which has an augmenting path if and only if the
original graph does. This is illustrated in Figure 4, which shows a blossom
B1 containing vertices c, d, e, f and g and a blossom B2 containing vertices j, k
and m. In the “shrunken graph” on the right, we have two augmenting paths
abB1 h and abB1 B2 ih. The corresponding paths in the original graph are
abcdegfh and abcfgkmjih. There is a straightforward procedure to obtain an
augmenting path in the underlying unshrunken graph, given an augmenting
path in the current shrunken graph.
We say than a vertex is internal if it is contained in some blossom, otherwise it is external. The base of a blossom is the unique vertex in the blossom
that has no incident matching edge with the other endpoint in the blossom.
So, in Figure 4, the base of B1 is c and the base of B2 is j.
Our algorithm modifies the blossom-shrinking procedure to accommodate
our modified augmenting paths. As in the bipartite case, the algorithm finds
an augmenting path by building a collection of trees rooted at unmatched
vertices in S. For convenience, we let β(u) denote the largest blossom containing a vertex u (if u is external, β(u) = u). Initially, all edges incident
7
to tree roots are on the list of eligible edges. The algorithm then repeats
the following step until it either finds an augmenting path in the current
shrunken graph or runs out of eligible edges.
Select an eligible edge e = {u, v} for which β(u) is even and remove it
from the list of eligible edges.
• If v is unreached and matched, let {v, w} be the matching edge
incident to v; extend the tree by making v a child of u and w a
child of v. If w is not in S, then the path from w to the root
of its tree in the current shrunken graph is an augmenting path;
otherwise, add all non-matching edges incident to w to the eligible
list.
• If v is unreached and unmatched, then the path consisting of e
plus the tree path from β(u) to the root of its tree in the current
shrunken graph is an augmenting path.
• If β(v) is even and in a different tree than β(u), then the path
formed by combining e with the tree path from β(u) to the root
of its tree in the current shrunken graph and the tree path from
β(v) to the root of its tree is an augmenting path.
• If β(v) is even and in the same tree as β(u), then the cycle formed
by combining e with the tree paths in the current shrunken graph
from β(u) and β(v) to their nearest common ancestor forms a
blossom. If some odd vertex x in the cycle is not in S, then there
is an augmenting path in the current shrunken graph from x to
the root of its tree that starts with the matching edge incident to
x, proceeds around the new blossom to its base and then up the
tree to the root. If there is no such vertex, add all non-tree edges
incident to odd vertices in the blossom cycle to the eligible list
and shrink the blossom.
• If β(v) is odd, ignore e and proceed to the next iteration
Figure 4 can be used to illustrate the operation of the algorithm (note that
the shaded vertices are in S). When the algorithm reaches the state shown
8
on the left side of the figure, the eligible edges are hf, dg, gk, eg and jm.
The algorithm ignores edges hf, dg and gk and forms one blossom when
processing edge eg and the other when processing jm. At this point, edges hf
and gk are again eligible. Processing either edge leads to the discovery of an
augmenting path. Also, observe that if d were not a member of S, then when
edge eg was processed, the algorithm would have found the augmenting path
abcfged. We defer the correctness proof of the algorithm to the next section.
3
Unlimited priorities
We can find a maximum priority matching for the general case using a generalization of the algorithm for the two priority case. We first maximize the
number of priority 1 vertices that are matched, then the number of priority 2 vertices and so forth. At each step, we find an augmenting path that
increases the number of matched priority i vertices without decreasing the
number of priority j vertices, for all j < i.
Define the i-score of a matching M as the n-ary integer with i digits
in which the j-th most significant digit is the number of priority j vertices
that are matched by M . For example, in Figure 1, the 2-score is 21 and the
5-score is 21110. Given a matching M with a maximum (i − 1)-score, an
i-augmenting path is a path p = u0 , . . . , ut in which edges alternate between
matching edges and non-matching edges, u0 is unmatched, ρ(u0 ) = i, and if
ut is matched, ρ(ut ) > i. Observe that because M has a maximum (i − 1)score, if ut is unmatched, it cannot have priority less than i. Consequently,
if we exchange the non-matching edges and matching edges in p, we obtain
a new matching with the same (i − 1)-score and a larger i-score than M .
To justify the use of augmenting paths, we must show that if a matching
M has a maximum (i − 1)-score but not a maximum i-score, then there must
be an i-augmenting path for M . Let M be such a matching and let M 0 be a
matching with a larger i-score. Let N be the graph consisting of edges that
are in M or M 0 but not both. N consists of a collection of disjoint paths
and even-length cycles. Since M 0 has a larger i-count than M , there must
be some path p in N , in which M 0 has a larger i-count than M . If p is an
odd-length path, its endpoints must be unmatched by M and at least one of
9
its endpoints must have priority i, making p an augmenting path for M . If
p is an even length path, the endpoint that is unmatched by M must have
priority i and the endpoint matched by M must have priority > i. Hence, p
is an augmenting path in this case, as well. We summarize this argument in
the following theorem.
Theorem 1 Let G = (V, E) be an undirected graph with priorities ρ(u)
and let M be a matching with a maximum (i − 1)-score. G contains an iaugmenting path with respect to M if and only if M does not have a maximum
i-score.
To find a matching with maximum overall priority score, we initialize i = 1,
then repeat the following step until i > n.
Search for an i-augmenting path; if one is found, augment the matching
by reversing the status of the path edges, otherwise increment i.
The heart of the method is the algorithm used to find an i-augmenting path.
At the start of each path search, all unmatched priority i vertices are tree
roots, and all edges incident to these vertices are in the eligible list. The
algorithm then searches for an i-augmenting path by repeating the following
step until it either finds a path or runs out of eligible edges.
Select an eligible edge e = {u, v} for which β(u) is even and remove it
from the list of eligible edges.
• If v is unreached and matched, let {v, w} be the matching edge
incident to v; extend the tree by making v a child of u and w a
child of v. If ρ(w) > i then the path in the current shrunken graph
from w to the root of its tree is an i-augmenting path; otherwise,
add all non-matching edges incident to w to the eligible list and
continue.
• If v is unreached and unmatched, then the path consisting of e
plus the tree path in the current shrunken graph from β(u) to the
root of its tree is an i-augmenting path.
10
• If β(v) is even and in a different tree than β(u), then the path
formed by combining e with the tree path from β(u) to the root
of its tree in the current shrunken graph and the tree path from
β(v) to the root of its tree is an i-augmenting path.
• If β(v) is even and in the same tree as β(u), then the cycle formed
by combining e with the tree paths in the current shrunken graph
from β(u) and β(v) to their nearest common ancestor forms a
blossom. If some odd vertex x on the blossom cycle has ρ(x) > i,
then there is an i-augmenting path in the current shrunken graph
from x to the root of the tree that starts with the matching edge
incident to x, continues around the blossom cycle to its base and
then up the tree to the root. If there is no such vertex, add all
non-tree edges incident to odd vertices on the blossom cycle to
the eligible list and shrink the blossom.
• If β(v) is odd, ignore e and proceed to the next iteration.
Once again, to establish the correctness of the algorithm, we need to show
that if it halts without finding an i-augmenting path, then the graph must
not contain one. Note the following properties of the algorithm.
1. Every tree root is unmatched and has priority i and each tree has
exactly one unmatched vertex.
2. Every matching edge either has both endpoints unreached, or one odd
and one even. In the latter case, the even endpoint is the child of the
odd endpoint in some tree.
3. For every internal vertex x, β(x) is even, and if x is unmatched, then
β(x) is unmatched.
4. If the algorithm fails to find an augmenting path, then when it fails,
any edge that has endpoints that are both even or internal is contained
within some blossom.
5. If the algorithm fails to find an augmenting path, then when it fails,
every vertex x that is even or internal has ρ(x) ≤ i.
11
Now, suppose that the original graph contains an augmenting path p =
u0 , . . . , ut , but the algorithm halts without finding a path. If both endpoints
of p are unmatched, then the endpoints cannot be in the same tree (by
property 1). In this case, let {uk , uk+1 } be an edge in p with uk and uk+1 in
different trees. By property 5, at least one of uk and uk+1 must be odd and
external. Assume, without loss of generality, that uk is odd and external.
Since {uk−1 , uk } is a matching edge, k is even. If just one endpoint (u0 ) of
p is unmatched, then t is even, ρ(ut ) > i and hence ut odd and external (by
property 4). Thus, in both cases, p contains an odd external vertex uk with
k even.
Let j be the smallest even integer for which uj is odd and external.
Since uj is external, there must some vertex in {u0 , . . . , uj−1 } that is odd
and external. (If not, all vertices in {u0 , . . . , uj−1 } must be contained in
a common blossom (property 5), and since u0 is unmatched, the blossom
must be also (property 3), hence must include the matching edge {uj−1 , uj },
contradicting the fact that uj is external.) Let i be the largest integer <
j for which ui is odd and external and note that i must be odd. This
implies that all vertices in {ui+1 , . . . , uj−1 } are in a common blossom and
that blossom is incident to two matching edges {ui , ui+1 } and {uj−1 , uj }.
This contradiction implies the correctness of the algorithm. (This argument
was adapted from [6].)
Each augmenting path search can be implemented to run in O(m log n)
time using the method described in [2] to represent the current shrunken
graph. This can be reduced to O(m) time using the data structure described
in [3].
Theorem 2 The augmenting path algorithm for priority matching computes
a matching with maximum priority score. It can be implemented to run in
O(mn) time.
We close by noting that the maximum priority matching is also a maximum size matching. If it were not, we could find an ordinary augmenting
path in the graph that would match two more vertices, giving it higher priority score.
12
4
Closing remarks
The maximum size matching problem can be solved in O(mn1/2 ) time using
the algorithms described in [4] for the bipartite case and [5] for the general
case. It is possible that one or both of these algorithms could be adapted to
handle maximum priority matching.
It might also be interesting to consider a weighted version of the problem.
In the maximum weight matching problem, we seek a matching that maximizes the sum of the edge weights of the edges in the matching. While one
cannot simultaneously maximize the weight and priority score of a matching,
one could conceivably maximize the weight of a matching with a specified
minimum priority score, or maximize the priority score of a matching with a
specified minimum weight. Alternatively, one might associate weights with
vertices and find matchings that maximize the weight of matched vertices.
References
[1] Edmonds, Jack. “Paths, trees and flowers,” Canadian Journal of Mathematics, 1965, pp. 449–467.
[2] Gabow, Harold N. “An efficient implementation of Edmonds’ algorithm
for maximum matching on graphs,” Journal of the Association for Computing Machinery, 1976, pp. 221–234.
[3] Gabow, Harold N and Robert E. Tarjan. “A linear time algorithm for a
special case of disjoint set union,” ACM Symposium on the Theory of
Computing (STOC), 1983, pp. 246-251.
[4] Hopcroft, John E. and Richard M Karp. “An O(n5/2 algorithm for maximum matching in bipartite graphs,” SIAM Journal on Computing, 1973,
pp 225–231.
p
[5] Micali, Silvio. and V. V. Vazirani. “An O( |V | · |E|) algorithm for
finding maximum matchings in general graphs,” IEEE Symposium on
the Foundations of Computer Science (FOCS), 1980, pp. 17-27.
13
[6] Tarjan, Robert E. Data structures and network algorithms. Society for
Industrial and Applied Mathematics, 1983.
[7] Turner, Jonathan S. “The bounded edge coloring problem and offline
crossbar scheduling,” Washington University Computer Science and Engineering Department technical report, wucs-2015-07, 2015.
14
| 8cs.DS
|
arXiv:1705.07755v1 [cs.NE] 19 May 2017
Improving classification accuracy of feedforward neural networks for spiking
neuromorphic chips
Antonio Jimeno Yepes, Jianbin Tang, Benjamin Scott Mashford
[email protected], [email protected], [email protected]
IBM Research Australia, Carlton, 3053, VIC, Australia
Abstract
Deep Neural Networks (DNN) achieve human level
performance in many image analytics tasks but
DNNs are mostly deployed to GPU platforms that
consume a considerable amount of power. New
hardware platforms using lower precision arithmetic achieve drastic reductions in power consumption. More recently, brain-inspired spiking neuromorphic chips have achieved even lower
power consumption, on the order of milliwatts,
while still offering real-time processing.
However, for deploying DNNs to energy efficient neuromorphic chips the incompatibility between continuous neurons and synaptic weights
of traditional DNNs, discrete spiking neurons and
synapses of neuromorphic chips need to be overcome. Previous work has achieved this by training
a network to learn continuous probabilities, before
it is deployed to a neuromorphic architecture, such
as IBM TrueNorth Neurosynaptic System, by random sampling these probabilities.
The main contribution of this paper is a new learning algorithm that learns a TrueNorth configuration
ready for deployment. We achieve this by training
directly a binary hardware crossbar that accommodates the TrueNorth axon configuration constrains
and we propose a different neuron model.
Results of our approach trained on electroencephalogram (EEG) data show a significant improvement with previous work (76% vs 86% accuracy) while maintaining state of the art performance
on the MNIST handwritten data set.
1 Introduction
Neural networks achieve human level performance in many
image analytics tasks but neural networks are mostly
deployed to GPU platforms that consume a considerable amount of power.
New hardware platforms using lower numerical precision achieve drastic reductions
in power consumption. Brain-inspired spiking neuromorphic chips such as the IBM TrueNorth Neurosynaptic System [Merolla et al., 2014] have achieved power consumption
on the order of milliwatts while still offering high accuracy
and real-time processing. Recent work [Esser et al., 2015;
Esser et al., 2016] has used these ideas to implement constrained neural networks to enable high performance image
analytics on TrueNorth [Merolla et al., 2014]. With recent
advances in hardware platforms already available, there is an
urge to provide software that fully exploits the potential of
this new type of hardware.
Recent approaches in machine learning have
explored training models constrained to binary
weights [Courbariaux et al., 2015] or low precision
arithmetic [Courbariaux et al., 2014; Cheng et al., 2015;
Hwang and Sung, 2014] and spiking deep belief neural
networks [Stromatias et al., 2015], achieving state of the
art performance in image analytics tasks. Lower power
consumption is achieved by using only additions (avoiding
expensive multiplication modules), which can benefit from
implementing these models in configurable architectures
such as FPGA or brain-inspired systems.
Recent methods for training adaptations of fully connected style layers to the TrueNorth core-to-core connectivity [Esser et al., 2015] applied to EEG data have shown that
results using neuromorphic platforms exhibit performance
levels that are below current work using unconstrained hardware [Nurse et al., 2016]. In this work, we revise the training
of deep neural networks for low energy neuromorphic architectures and demonstrate high performance levels comparable
to non-constrained neural networks on MNIST and EEG data
sets.
The selection of datasets in this study is guided by emerging areas of application for neuromorphic computing platforms. It has been proposed that the development of highperformance classifiers that run within a constrained power
consumption limit will enable a new generation of mobile sensing applications [Lane et al., 2015]. Early examples of such applications include healthcare-related wearables that involve machine vision and audio-signal processing [Poggi and Mattoccia, 2016], as well as real-time
analysis of imagery for visual monitoring and surveillance
[Antoniou and Angelov, 2016].
This paper is organized as follows. First feedforward neural networks and their training are introduced, followed by an
overview of the TrueNorth chip architecture. Then a revised
constrained model of the feedforward neural network train-
ing and deployment for TrueNorth are introduced. Finally
results on EEG data and the MNIST handwritten data set are
presented with discussion and future work.
2 Feedforward Neural networks
In feedforward neural networks, inputs xi are integrated in
units Ij using a set of weights wij and a bias term bj (which
can be shared among units) as shown in equation 1. During
training, the weights wij and bias bj are learnt.
X
xi wij + bj
(1)
Ij =
hardware [Benjamin et al., 2014; Painkras et al., 2013;
Pfeil et al., 2012].
In its current implementation it is
composed of 4,096 neurosynaptic cores.
Figure 2 shows the layout of one of these cores. Each core
has highly configurable 256 input axons and 256 output neurons. Input spikes enter the core through axons and the information is passed via the configurable binary synaptic crossbar
to the leaky integrate-and-fire neurons, which act as computation units. The neuron assigns weights to the input spikes,
integrates and updates its membrane potential, compares it to
a threshold and may spike and reset, based on the selected
threshold, reset, and stochastic modes [Cassidy et al., 2013].
i
♥✄☎✆✝✞s♥❛☛✡✠☞ ☞✝✆✄
Neurons output are a combination of the integration function Ij followed by a non-linearity, which can be a sigmoid
function. The neurons can be grouped into layers and these
layers can be stacked as shown in figure 1.
❞✄♥❞✆✠✡✄✞
✞s♥❛☛✡✠☞
☞✆✝✞✞❝❛✆
❇ ✁✁❢✂
❇ ✁✁❢✂
❇ ✁✁❢✂
❛✟✝♥✞
❇ ✁✁❢✂
♥✄☎✆✝♥✞
Figure 2: TrueNorth neurosynaptic-core
Figure 1: Neural network
Prediction is done using the output layer outcome. In classification, typically each neuron is associated to one prediction category. Information is integrated in the output layer
neurons, which are then be turned into probabilities using,
for instance, softmax (eq. 2). pl is the prediction probability
for category cl from all the possible K categories.
ecl
pl = P K
k=1
eck
(2)
During training, weight and bias values are typically learnt
using backpropagation [Duda et al., 2012], which minimizes
a loss function that represents classification error of the network using gradient descent. An example of such a loss function is the logarithmic loss, which is used in this work (eq. 3),
even though other functions could be explored.
E=−
X
[yk log(pk ) + (1 − yk )log(1 − pk )]
(3)
k
The next sections describe the neuromorphic hardware and
an explanation about how the neural networks can be constrained, learnt and deployed.
2.1
Deployment hardware
IBM TrueNorth [Merolla et al., 2014], a low power
and highly parallelized brain-inspired chip is used
here as an example deployment system, though our
method could be generalized to other neuromorphic
Programming for TrueNorth means writing core configurations and connections, which in a way could be seen as building a neural network. A programming language has been developed to abstract the chip configuration [Amir et al., 2013].
In our case, the learned neural networks need to be translated
into a valid TrueNorth configuration.
In previous work [Esser et al., 2015], a constrained network is trained, which learns the crossbar connection probability and biases. To deploy the model into TrueNorth these
probabilities are sampled into discrete synaptic crossbar connections.
3 Constrained training model
Neuromorphic chips such as TrueNorth provide a reduction
of power consumption, which is achieved by a constrained
hardware architecture. There are several constraints compared to more traditional hardware platforms that are explored in this section to define the training of neuromorphic
chips. These constraints are: use of spiking neurons, weights
constrained to a limited set of integer values and reduced connectivity. In the following sections, we describe these constraints in more detail and provide a way to do learning based
on backpropagation. We also describe how to use output layer
predictions integrated into backpropagation.
3.1
Spiking neurons
Spiking neurons have two possible outputs (0, no output, or 1,
spike) as shown in the revised neuron output shown in eq. 4.
1, if Ij > 0
nj =
(4)
0, otherwise
Spiking neurons behavior could be seen as a Heaviside step
function, which does not have a gradient for backpropagation
since it is not differentiable. This step function can have many
different smooth approximations using, for instance, the logistic function that is used to represent the neuron output n̂j
as shown in eq. 5. The parameter k is the steepness of the
function, which in our work has been set empirically to 1/2.
Other values of k could be explored or, may be, approximate
its value to a logistic distribution using large data sets.
1
(5)
1 + e−2kIj
The following gradient formulation can then be used in
backpropagation.
n̂j ≈
∂ n̂j
1
1
=
∗ (1 −
)
∂Ij
1 + e−2kIj
1 + e−2kIj
(6)
In this approach, each neuron is stateless, so its status is
reset after each time cycle, or tick. This is implemented in
TrueNorth by setting parameters as specified and each neurons leak equal to the rounded bias term, with its threshold to
0, its membrane potential floor to 0 [Esser et al., 2015].
3.2
Constrained weights
In TrueNorth, effective weights wij are divided in two components: the axon weight sij and the crossbar connection cij ,
so wij = cij sij . Taking this into account, the integration
function is defined as:
X
xi cij sij + bj
(7)
Iˆj =
i
Instead of learning the weights wij , the axon weights
are proportionally distributed as explained below the binary
crossbar connection cij is learnt during training.
The hardware accepts two possible values for the crossbar
connections (0,1). The effective weights are dependent on the
configuration of the axon types and crossbar connections. In
this setup, the total number of effective weights that can be
used is a maximum of five, i.e. four axon type weights and
no connection (which means zero weight). The distribution
of weights in sij depends on the axon type configuration in a
TrueNorth core (we follow [Esser et al., 2015]).
The backpropagation gradients are specified as below and
a similar formulation is used for bias values, which are round
to the closer integer value during deployment. In the next
section, we explain how the binary crossbar connections are
learnt.
∂ n̂j
∂ n̂j ∂ Iˆj ∂ Iˆj
=
;
= xi sij
∂cij
∂ Iˆj ∂cij ∂cij
3.3
(8)
This is in contrast to previous work in which connection
probabilities are learnt [Esser et al., 2015] that need to be
sampled prior to hardware deployment. As well, in contrast to
previous work [Esser et al., 2016; Courbariaux et al., 2015],
the proposed approach allows learning a larger number of
weights that are defined by the hardware configuration.
During training, in the forward propagation step the high
precision csh matrix from the shadow network is used to
calculate the c binary connection values (eq. 9), which are
used to calculate the effective network weights using (wij =
cij sij ). Effective network weights are used to calculate the
performance of the trained model (c.f. figure 3).
1, if csh
ij > 0.5
cij =
(9)
0, otherwise
During backpropagation, gradients are estimated using c
connection values that are used to update the csh high precision connection probabilities.
Algorithm 1 Forward and backward propagation
//Forward propagation
for l = 1 : layers do
for c = 1 : layer(l).cores do
Calculate core binary crossbar connections cij (c.f.
eq. 9) and effective weights wij
Calculate core neuron outputs Iˆj using cij
end for
end for
Estimate log loss in the last layer
//Backpropagation
for l = 1 : layers do
for c = 1 : layer(l).cores do
Calculate gradients using cij
Update csh
ij using estimated gradients
end for
end for
Figure 3: Crossbar connections during training and deployment
Binary crossbar connection learning
In the previous section, we have shown that binary crossbar
connections need to be learnt. We propose adapting previous work [Esser et al., 2016; Courbariaux et al., 2015] to
learn binary crossbar connections that are used to obtain the
effective network weights as shown in algorithm 1.
3.4
Neural network connections
The described method allows defining fully connected layer
networks, but TrueNorth cores have 256 inputs (axons) and
256 outputs (neurons), which constrains the size of fully connected layers. When designing a network, the first layer,
which processes the input data, is defined by blocks that tile
to cover the input data. Block size and stride are defined by
the number of pixels in the image classification case. Since
each of these blocks maps to a TrueNorth core, the number of
input axons required to map the input data per block is calculated as (blockSize2 ∗numberChannels) and the maximum
value is 256, which is the number of axons per core. In the
upper layers, block size and stride are defined by the number
of cores (instead of image size as in the input layer). Output
neurons of the previous layer are mapped to the input axons
of the cores in the upper layer.
3.5
Output layer
The last layer of the of the network is used for prediction.
Spiking neurons in the last layer can be assigned randomly
to one category [Esser et al., 2015]. The category with the
highest number of neurons firing will be the predicted one.
These counts need to be used in backpropagation. In our
work, we have used the number of spiking neurons per category csk divided by the number of neurons per category cnk
and used softmax as shown in equation 10 to estimate the
per category probability that is then used with the logarithmic loss (eq. 3).
csl
e cnl
p̂l = P
K
k=1
csk
(10)
e cnk
We have described neural networks, the constraints added
by neuromorphic chips and how to adapt the training of neural
networks to these constraints. In the following sections, we
present the results obtained using the MNIST data set and
EEG data.
4 Results
We have used two data sets in our experiments. An EEG data
set, described in more detail below, and the MNIST handwritten data set.
EEG data1 was recorded from a participants wearing electrodes placed on the scalp while they performed a self-paced
hand squeeze task. Full experimental details are described
elsewhere [Nurse et al., 2015]. Briefly, the participant wore
surface electromyography (EMG) electrodes attached to the
forearms that were used to detect the onset of each hand
squeeze. Data was sampled at 1 kHz in four minute sessions,
with a total recording time of approximately 30 minutes. Detection of hand squeeze onset was achieved by extracting the
root mean squared (RMS) power from the EMG signal, and
the corresponding EEG data was labeled with either left-hand
squeeze or right-hand squeeze (2 classes). Each EEG instance
has been translated to 46x46 images with one 8-bit channel.
In this study, EEG signals from participant D are used, which
allows comparing the results to previous work. This data set
contains 480/468 examples per class in the training set and
66/95 examples per class in the testing set. The data set consisted of a randomly allocated unseen test set (20% of the
1
https://github.com/EwanNurse/A-Generalizable-BCI-usingMachine-Learning-for-Feature-Discovery
data) and training set (80% of the data), allowing a direct
comparison with previous results [Nurse et al., 2015].
MNIST data set [LeCun et al., 1998] contains handwritten digits from 0 to 9 with 60,000 examples for training and
10,000 samples for testing. Images are 28x28 pixels, with
one 8-bit gray scale channel. Gray scale values have been
normalized dividing them by 255.
During training EEG and MNIST data are represented using floating point values. During deployment, the data values
a single input axon over n ticks, which means that floating
point values need to be represented as a set of spikes. Rate
code [Jimeno Yepes et al., 2016] has been used to encode the
data. In rate code, the number of spikes in the time window
represents the encoded value, k. Spikes are evenly distributed
within the time window.
Experiments were run using a MATLAB implementation
of the proposed method, this includes training and evaluation
of the trained models in a simulated MATLAB environment.
We have used two configurations for the axon type weights in
the neuron configuration s=[-1,1], used in previous work, and
s=[-2,-1,1,2].
Table 1 lists the networks used in the experiments.
Network configurations are derived from previous
work [Esser et al., 2015]. There is a small network with 3
layers that requires 9 TrueNorth cores and a larger network
with 4 layers that requires 30 TrueNorth cores.
Name
Small network
Large network
Layer
1
2
3
1
2
3
4
Definition
Block size 16 Stride 12
Block size 1 Stride 1
Block size 2 Stride 1
Block size 16 Stride 4
Block size 2 Stride 1
Block size 2 Stride 1
Block size 2 Stride 1
Table 1: Networks used in the experiments. MNIST small
has 3 layers and requires 9 TrueNorth cores. MNIST large
has 4 layers and requires 30 TrueNorth cores.
Performance of the deployed trained models has been measured using classification accuracy, which is calculated by
dividing the number of correctly predicted instances by all
instances in the test set. McNemar’s test has been used to
look if the outcome of the systems is statistically significant [Haralick, 1993]. During test, data is converted into
spikes using a rate code scheme. Since rate code is used and
there is no need to sample from the trained network crossbar
probabilities, results are stable (e.g. repetitions of the experiments provide the same result) compared to previous methods.
We have found that the initial optimal learning rate lr was
provided by equation 11, which takes into account the number
of neurons assigned to each category and the batch size.
lr = (neurons per category ∗ batch size)−1
(11)
4.1
EEG results
For EEG experiments, the initial learning rate lr was multiplied by 0.1. We have updated the learning rate every 100
iterations using a rate of 0.1. Batch size is 25 and we have
run a total of 500 epochs.
Ranges of values vary a lot for EEG data across features.
We have normalized each individual feature xi but instead
of using values between -1 and 1 as it is typically done with
unconstrained neural networks [Ioffe and Szegedy, 2015], the
values have been normalized to be between 0 and 1 as shown
in the formula below. A small ǫ value has been inserted to
avoid zero values.
train
x̂i =
i
√xi −E[xtrain
var(xi
]
)+ǫ
+ 0.5
(12)
2
Results are shown in Table 2 training the network with
two different axon type weights and the combination of both
trained models. Previously reported results using TrueNorth
showed a 76% accuracy for participant D on this 2-class
task [Nurse et al., 2016]. With the proposed method we obtain over 86%, which is the same accuracy as previously
reported for the same data set using an unconstrained neural network [Nurse et al., 2015]. Better results are obtained
when a more restricted set of axon type weights (s = [−1, 1])
is used. Normalization of the set at the feature level improves the performance of the trained models, which indicates that feature preprocessing can benefit predictive performance. Both results are statistically significant (p < 0.005).
Using more than 32 ticks does not show a statistically significant improvement.
T
1
2
4
8
16
32
64
s=[-2,-1,1,2]
Original
Normalized
61.43±3.72 76.58±2.17
64.66±4.82 78.01±1.58
68.94±2.40 80.93±1.78
71.12±2.22 81.30±2.46
73.73±3.21 82.55±1.48
76.27±3.14 83.17±0.99
76.52±2.23 83.60±0.52
s=[-1,1]
Original
Normalized
62.36±5.27 79.19±2.45
66.96±4.22 81.80±2.43
70.06±3.98 84.22±1.50
72.36±2.92 84.66±1.58
75.96±2.17 84.07±1.40
80.00±2.34 85.90±0.66
80.62±1.89 86.09±0.73
Table 2: EEG accuracy results using the large network and
two axon weights s configuration and with (Normalized) and
without normalization (Original). T stands for the number of
ticks used to encode test instances with rate code.
4.2
iterations were required for the training set using augmentation to converge compared to the number of iterations used
without augmentation.
Tables 3 and 4 show the results for several configurations. The large network shows better performance compared to the smaller one. The configuration with axon types
(s = [−1, 1]) provide better performance. Augmentation increases the performance, which is similar to previous published work [Esser et al., 2015]. Improvements in performance by these three configurations are statistically significant (p < 0.005). As with the EEG results, using more than
32 ticks does not show a statistically significant improvement.
MNIST results
Training has been carried out using batches of 100 images
for each iteration and with and without data augmentation.
To perform data augmentation, random changes to the images have been done in each epoch. We used a maximum
rotation of 7.5 degrees, a maximum shift of 2.5 and a maximum rescale of 7.5%. Augmentation changes are randomly
performed. For experiments without augmentation, the initial learning rate lr is multiplied by 0.1 and 500 epochs were
used. For augmentation experiments, the initial learning rate
lr is multiplied by 0.1, which is further multiplied by 0.1 after 2,000 epochs. 3,000 epochs were used for training. More
T
1
2
4
8
16
32
64
s=[-2,-1,1,2]
No aug
Aug
95.05±0.14 95.99±0.13
96.27±0.09 97.46±0.13
97.00±0.09 98.20±0.08
97.30±0.09 98.44±0.03
97.44±0.06 98.56±0.07
97.56±0.07 98.64±0.05
97.65±0.04 98.66±0.03
s=[-1,1]
No aug
Aug
93.73±0.21 95.20±0.16
96.07±0.20 97.17±0.11
97.28±0.11 98.22±0.11
97.75±0.06 98.58±0.09
97.95±0.08 98.76±0.05
98.03±0.05 98.84±0.02
98.08±0.04 98.86±0.03
Table 3: MNIST small network results using data augmentation (Aug) and not using it (No aug). T stands for the number
of ticks used to encode test instances with rate code.
T
1
2
4
8
16
32
64
s=[-2,-1,1,2]
No aug
Aug
95.45±0.13 96.97±0.13
96.50±0.11 98.07±0.11
97.16±0.06 98.60±0.07
97.53±0.09 98.78±0.04
97.68±0.05 98.91±0.04
97.80±0.05 98.97±0.04
97.82±0.05 98.99±0.03
s=[-1,1]
No aug
Aug
95.36±0.17 97.26±0.13
96.45±0.12 98.32±0.10
97.38±0.10 98.88±0.07
97.70±0.04 99.16±0.05
97.88±0.05 99.29±0.04
97.94±0.04 99.29±0.03
97.99±0.03 99.32±0.02
Table 4: MNIST large network results using data augmentation (Aug) and not using it (No aug). T stands for the number
of ticks used to encode test instances with rate code.
5 Discussion
Results on EEG data show that the performance of the trained
network increased from 76% [Nurse et al., 2016] to 86%,
which is close to state-of-the-art performance obtained using unconstrained fully connected layers [Nurse et al., 2015].
Results on MNIST show that the performance of the proposed
method is similar compared to previously reported work depending on the configuration.
The use of several ticks for encoding input data shows that
a plateau is reached after 8 ticks for MNIST, additional ticks
to encode data does not significantly increase performance.
We have examined the effective weights and bias values.
Figures 4 and 5 show the values for MNIST and figure 6
shows the values for the EEG data set.
In all combinations, approximately half of the weights have
a value of zero for all the layers, which implies that half of the
crossbar connections are disconnected. The other weights are
EEG neuron weight configuration s=[-1,1]
0.6
0.6
0.4
1
L1
L2
L3
L4
0.4
0.2
0.2
0
-2 -1 0
1
Effective Weight
0
Bias
-1
0
0.6
Effective Weight
0.2
0.6
0.4
0.4
0.2
0.2
0
-1
0
1
0.6
0.4
0.2
0
-5 -4 -3 -2 -1 0 1 2 3 4 5
Effective Weight
L1
L2
L3
0.8
0
Bias
-1
0
0
-5 -4 -3 -2 -1 0 1 2 3 4 5
1
Effective Weight
Bias
No augmentation
Aug
Neuron weight configuration s=[-2,-1,1,2]
0.6
0.6
0.2
Probability
0.4
L1
L2
L3
L4
0.8
0.6
0.4
Probability
1
L1
L2
L3
L4
1
L1
L2
L3
L4
0.4
0.2
0.2
0.6
0.4
0.4
0.2
-2 -1 0
1
2
0.6
0
Bias
-2 -1 0
1
0
-5 -4 -3 -2 -1 0 1 2 3 4 5
2
Effective Weight
No augmentation
Bias
Aug
Figure 4: Effective weights and bias values for the small network (3 layers, 9 cores) for MNIST data set. Two weight
configurations have been used s=[-1,1] and s=[-2,-1,1,2]. The
networks have been trained without augmentation (No augmentation) and with augmentation (Aug).
Neuron weight configuration s=[-1,1]
0.6
0.2
0.6
0.4
1
L1
L2
L3
L4
0.4
0.2
0.2
0
-1
0
Effective Weight
0.6
0.4
0.2
0
-5 -4 -3 -2 -1 0 1 2 3 4 5
1
L1
L2
L3
L4
0.8
Probability
0.4
L1
L2
L3
L4
0.8
Probability
1
L1
L2
L3
L4
Probability
Probability
0.6
0
Bias
-1
0
0
-5 -4 -3 -2 -1 0 1 2 3 4 5
1
Effective Weight
Bias
No augmentation
Aug
Neuron weight configuration s=[-2,-1,1,2]
0.6
0.2
0.6
0.4
1
L1
L2
L3
L4
0.4
0.2
0.2
0
-2 -1 0
1
2
Effective Weight
L1
L2
L3
L4
0.8
Probability
0.4
L1
L2
L3
L4
0.8
Probability
1
L1
L2
L3
L4
Probability
Probability
0.6
0.6
Bias
No augmentation
0
2
0.2
0
-5 -4 -3 -2 -1 0 1 2 3 4 5
Effective Weight
0.4
Bias
Original
0
-2 -1 0
1
2
0
-5 -4 -3 -2 -1 0 1 2 3 4 5
Effective Weight
Bias
Normalized
Figure 6: Effective weights and bias values for the large network (4 layers, 30 cores) for EEG data set. Two weight configurations have been used s=[-1,1] and s=[-2,-1,1,2]. The
networks have been trained using the original set or the normalized one.
-2 -1 0
1
6 Conclusions and future work
2
Effective Weight
We have shown that it is possible to train constrained fully
connected layers on neuromorphic chips for two sets of problems (image classification with MNIST and EEG data analytics), which provides better or similar performance compared
to previous work depending on the data set.
Results on EEG data sets which were previously analysed
through the original method yield over 86% classification accuracy using the proposed approach, which is comparable
to unconstrained neural networks. The trained models use a
small portion of the TrueNorth chip (30 cores vs. 4096 available in the current version of the chip), thus requiring a much
less than 70mW to work, which makes these models suitable
for portable autonomous devices with large autonomy.
The outcome of this research provides insights into the possibility of learning network parameters that translate into deployment networks for brain-inspired chips. We will explore
feasibility of expanding the introduced new methodology to
train other kinds of existing deep learning algorithms including convolutional neural networks and recurrent neural networks. Furthermore, analysis of the learnt parameters provide insights that might complement hardware design, thus
providing a more efficient deployment of the trained models.
7 Acknowledgment
0.4
0.2
0
-5 -4 -3 -2 -1 0 1 2 3 4 5
1
0.6
0.4
0.2
0
-5 -4 -3 -2 -1 0 1 2 3 4 5
Effective Weight
L1
L2
L3
0.8
0.2
0
-2 -1 0
L1
L2
L3
L4
0.8
Probability
0.6
1
L1
L2
L3
Probability
0.4
L1
L2
L3
0.8
Probability
1
L1
L2
L3
Probability
Probability
0.6
Bias
Original
Normalized
EEG neuron weight configuration s=[-2,-1,1,2]
0
1
L1
L2
L3
Probability
0.4
L1
L2
L3
0.8
Probability
1
L1
L2
L3
Probability
Probability
0.6
0.4
0
-5 -4 -3 -2 -1 0 1 2 3 4 5
1
0.2
Neuron weight configuration s=[-1,1]
0.6
0.2
0
-5 -4 -3 -2 -1 0 1 2 3 4 5
2
L1
L2
L3
L4
0.8
Probability
0.2
L1
L2
L3
L4
0.8
Probability
1
L1
L2
L3
L4
0.4
Probability
Probability
0.6
Probability
evenly distributed. Furthermore, bias values tend to be around
zero for all layers in many cases. On the other hand, when using augmentation in MNIST or in the EEG experiments, bias
values tend to be distributed mostly on the negative values for
the first layers. Negative values might indicate that there is a
tendency to prevent neurons from spiking.
The values for weights and biases imply that in a deployment scenario, hardware configuration could be further constrained to a limited set of bias values and further research
is required to understand implications in hardware design to
improve crossbar usage.
0
-5 -4 -3 -2 -1 0 1 2 3 4 5
Bias
The authors would like to thank the SyNAPSE team at the
IBM Almaden Research Center for their support.
Aug
References
Figure 5: Effective weights and bias values for the large network (4 layers, 30 cores) for MNIST data set. Two weight
configurations have been used s=[-1,1] and s=[-2,-1,1,2]. The
networks have been trained without augmentation (No augmentation) and with augmentation (Aug).
[Amir et al., 2013] Arnon Amir, Pallab Datta, William P
Risk, Andrew S Cassidy, Jeffrey Kusnitz, Steven K Esser,
Alexander Andreopoulos, Theodore M Wong, Myron
Flickner, Rodrigo Alvarez-Icaza, et al. Cognitive computing programming paradigm: a corelet language for
composing networks of neurosynaptic cores. In The
2013 International Joint Conference on Neural Networks
(IJCNN), pages 1–10. IEEE, 2013.
[Antoniou and Angelov, 2016] Antreas Antoniou and Plamen Angelov. A general purpose intelligent surveillance
system for mobile devices using deep learning. In Neural
Networks (IJCNN), 2016 International Joint Conference
on, pages 2879–2886. IEEE, 2016.
[Benjamin et al., 2014] Ben Varkey Benjamin, Peiran Gao,
Emmett McQuinn, Swadesh Choudhary, Anand R Chandrasekaran, Jean-Marie Bussat, Rodrigo Alvarez-Icaza,
John V Arthur, Paul A Merolla, and Kwabena Boahen.
Neurogrid: A mixed-analog-digital multichip system for
large-scale neural simulations. Proceedings of the IEEE,
102(5):699–716, 2014.
[Cassidy et al., 2013] Andrew Cassidy, Paul Merolla, John
Arthur, Steve Esser, Bryan Jackson, Rodrigo AlvarezIcaza, Pallab Datta, Jun Sawada, Theodore Wong, Vitaly
Feldman, Arnon Amir, Daniel Ben-Dayan Rubin, et al.
Cognitive computing building block: A versatile and efficient digital neuron model for neurosynaptic cores. In The
2013 International Joint Conference on Neural Networks
(IJCNN), pages 1–10. IEEE, 2013.
[Cheng et al., 2015] Zhiyong Cheng, Daniel Soudry, Zexi
Mao, and Zhenzhong Lan. Training binary multilayer
neural networks for image classification using expectation
backpropagation. arXiv preprint arXiv:1503.03562, 2015.
[Courbariaux et al., 2014] Matthieu Courbariaux, Yoshua
Bengio, and Jean-Pierre David. Low precision arithmetic
for deep learning. arXiv preprint arXiv:1412.7024, 2014.
[Courbariaux et al., 2015] Matthieu Courbariaux, Yoshua
Bengio, and Jean-Pierre David. Binaryconnect: Training
deep neural networks with binary weights during propagations. In Advances in Neural Information Processing
Systems, pages 3105–3113, 2015.
[Duda et al., 2012] Richard O Duda, Peter E Hart, and
David G Stork. Pattern classification. John Wiley & Sons,
2012.
[Esser et al., 2015] Steve K Esser, Rathinakumar Appuswamy, Paul Merolla, John V Arthur, and Dharmendra S Modha.
Backpropagation for energy-efficient
neuromorphic computing.
In Advances in Neural
Information Processing Systems, pages 1117–1125, 2015.
[Esser et al., 2016] Steven K Esser, Paul A Merolla, John V
Arthur, Andrew S Cassidy, Rathinakumar Appuswamy,
Alexander Andreopoulos, David J Berg, Jeffrey L McKinstry, Timothy Melano, Davis R Barch, et al. Convolutional networks for fast, energy-efficient neuromorphic
computing. arXiv preprint arXiv:1603.08270, 2016.
[Haralick, 1993] Robert M Haralick. Performance characterization in computer vision. In International Conference
on Computer Analysis of Images and Patterns, pages 1–9.
Springer, 1993.
[Hwang and Sung, 2014] Kyuyeon Hwang and Wonyong
Sung. Fixed-point feedforward deep neural network design using weights+ 1, 0, and- 1. In Signal Processing Sys-
tems (SiPS), 2014 IEEE Workshop on, pages 1–6. IEEE,
2014.
[Ioffe and Szegedy, 2015] Sergey Ioffe and Christian
Szegedy. Batch normalization: Accelerating deep network training by reducing internal covariate shift. arXiv
preprint arXiv:1502.03167, 2015.
[Jimeno Yepes et al., 2016] Antonio Jimeno Yepes, Jianbin
Tang, Shreya Saxena, Tobias Brosch, and Arnon Amir.
Weighted population code for low power neuromorphic
image classification. In Neural Networks (IJCNN), 2016
International Joint Conference on, pages 4294–4301.
IEEE, 2016.
[Lane et al., 2015] Nicholas D Lane, Sourav Bhattacharya,
Petko Georgiev, Claudio Forlivesi, and Fahim Kawsar. An
early resource characterization of deep learning on wearables, smartphones and internet-of-things devices. In Proceedings of the 2015 International Workshop on Internet
of Things towards Applications, pages 7–12. ACM, 2015.
[LeCun et al., 1998] Yann LeCun, Léon Bottou, Yoshua
Bengio, and Patrick Haffner. Gradient-based learning applied to document recognition. Proceedings of the IEEE,
86(11):2278–2324, 1998.
[Merolla et al., 2014] 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.
[Nurse et al., 2015] Ewan S Nurse, Philippa J Karoly,
David B Grayden, and Dean R Freestone. A generalizable
brain-computer interface (bci) using machine learning for
feature discovery. PloS one, 10(6):e0131328, 2015.
[Nurse et al., 2016] Ewan Nurse, Benjamin S Mashford, Antonio Jimeno Yepes, Isabell Kiral-Kornek, Stefan Harrer,
and Dean R Freestone. Decoding eeg and lfp signals using
deep learning: heading truenorth. In Proceedings of the
ACM International Conference on Computing Frontiers,
pages 259–266. ACM, 2016.
[Painkras et al., 2013] Eustace Painkras, Luis A Plana, Jim
Garside, Steve Temple, Francesco Galluppi, Cameron Patterson, David R Lester, Andrew D Brown, and Steve B
Furber. Spinnaker: A 1-w 18-core system-on-chip for
massively-parallel neural network simulation. IEEE Journal of Solid-State Circuits, 48(8):1943–1953, 2013.
[Pfeil et al., 2012] Thomas Pfeil, Andreas Grübl, Sebastian
Jeltsch, Eric Müller, Paul Müller, Mihai A Petrovici,
Michael Schmuker, Daniel Brüderle, Johannes Schemmel, and Karlheinz Meier. Six networks on a universal neuromorphic computing substrate. arXiv preprint
arXiv:1210.7083, 2012.
[Poggi and Mattoccia, 2016] Matteo Poggi and Stefano Mattoccia. A wearable mobility aid for the visually impaired
based on embedded 3d vision and deep learning. In Computers and Communication (ISCC), 2016 IEEE Symposium on, pages 208–213. IEEE, 2016.
[Stromatias et al., 2015] E Stromatias, D Neil, F Galluppi,
M Pfeiffer, SC Liu, and S Furber. Scalable energyefficient, low-latency implementations of spiking deep
belief networks on spinnaker. In 2015 International
Joint Conference on Neural Networks (IJCNN)(Killarney),
2015.
| 9cs.NE
|
Learning General Latent-Variable Graphical Models with
Predictive Belief Propagation and Hilbert Space Embeddings
arXiv:1712.02046v1 [cs.LG] 6 Dec 2017
Borui Wang
Stanford University
Abstract
In this paper, we propose a new algorithm for
learning general latent-variable probabilistic graphical models using the techniques of
predictive state representation, instrumental
variable regression, and reproducing-kernel
Hilbert space embeddings of distributions.
Under this new learning framework, we first
convert latent-variable graphical models into
corresponding latent-variable junction trees,
and then reduce the hard parameter learning
problem into a pipeline of supervised learning problems, whose results will then be used
to perform predictive belief propagation over
the latent junction tree during the actual inference procedure. We then give proofs of our
algorithm’s correctness, and demonstrate its
good performance in experiments on one synthetic dataset and two real-world tasks from
computational biology and computer vision
— classifying DNA splice junctions and recognizing human actions in videos.
1
Introduction
Many important problems in modern machine learning
and artificial intelligence can be gracefully captured by
probabilistic graphical models with latent variables.
However, learning such general latent-variable probabilistic graphical models has remained very difficult,
largely due to the intractability introduced by the
latent variables and the possibly complicated loopy
structures of the graphical models.
Previously, most learning algorithms for latentvariable graphical models would resort to local search
heuristics such as the Expectation Maximization algorithm (Dempster et al., 1977) that turn the parameter
learning problem into a maximization problem over a
non-convex objective function. However, such methods based on iterative local search can lead to bad local
optima and can be very slow to converge.
Geoffrey J. Gordon
Carnegie Mellon University
In order to overcome these shortcomings, recently researchers have been proposing and studying a new
family of non-parametric learning algorithms called
spectral learning that are based on the idea of method
of moments (Hsu et al., 2009; Anandkumar et al.,
2014). These spectral learning algorithms seek to learn
alternative parameterizations of the latent-variable
graphical models that are purely based on the observable variables and can be efficiently recovered from
their low-order observable moments through tensor decompositions. Therefore these algorithms enjoy the
advantages of being local-minimum-free and provably
consistent, and also can often be orders of magnitude
faster than the search-based methods such as the EM
algorithm.
Although the spectral learning algorithms have recently been successfully applied to tackle the learning
problems of more and more types of latent-variable
graphical models, there are still several severe limitations and problems to these spectral algorithms that
need to be solved. First of all, most of the spectral
learning algorithms that have been proposed so far
are only restricted to learning one or several specific
types of latent-variable graphical models (the majority of them are tree-structured), thus making themselves hard to generalize to other types of latentvariable graphical model structures (possibly beyond
trees). Second, most of the current spectral learning
algorithms can only deal with discrete-valued random
variables with moderate cardinalities due to their parameterization assumptions, thus cannot be easily applied to scenarios where the variables in the graphical models are continuous-valued or have large cardinalities, which are quite often encountered in many
important real-world applications. Third, the current spectral algorithms are generally idiosyncratic to
the specific graphical model structures that they are
targeted to learn, and thus cannot provide a flexible
learning framework or template for users to easily incorporate different prior knowledge and probabilistic
assumptions when facing different learning problems.
In order to solve these problems and to overcome these
Learning General Latent-Variable Graphical Models with Predictive Belief Propagation and HSE
2
Predictive State Representations
and Predictive Belief Propagation
In its original formulation, a predictive state representation (PSR) (Singh et al., 2004) is a compact
way to model the state of a controlled dynamical system using a vector consisting of predictions over a set
of observable future experiments (or tests) that can
be performed on the system. Under this representation, the state of the dynamical system can be compactly summarized as a K-dimensional vector of the
predictions over a finite set of K tests, which is called
the “core set of tests”, and the success probabilities of
all other future tests are simply determined by linear
functions of this state vector.
A key insight is that the idea of predictive state representations can actually be carried over to characterize the information flows of message passing and
belief propagation in probabilistic graphical models.
In essence, the state of a system is a time bottleneck that compactly summarizes everything we need
to know about the past in order to predict the future,
as illustrated in Figure 1(a) below.
When we perform belief propagation during inference, at each separator set, the classical approach is
to propagate a message containing the partial sumproduct results from one part of the junction tree to
Intuition: bottleneck
Past
Future
Belief Propagation
predict
separator set
data about past
(many samples)
data about future
(many samples)
state
limitations, in this paper we will develop a new algorithmic framework for learning general latent-variable
probabilistic graphical models based on the ideas of
predictive belief propagation and reproducing-kernel
Hilbert space embeddings of distributions. Our new
learning framework is built on the latent junction tree
graphical representation of latent-variable probabilistic graphical models, and reduces the hard parameter learning problem down to a pipeline of supervised
learning problems named two-stage regression, where
we can flexibly insert any supervised learning algorithms into the procedure as we see fit. When dealing with continuous variables, we will embed their distributions as elements in reproducing-kernel Hilbert
spaces, and then use the kernel trick to propagate and
integrate these embeddings over the latent junction
tree via tensor operations in a non-parametric fashion,
thus seamlessly extending our algorithm from the discrete domain to the continuous domain under the same
general framework. As a result, our learning framework is general enough to apply to all different structures of latent-variable graphical models (both trees
and loopy graphs), can handle both discrete variables
and continuous variables, and can allow users to easily incorporate prior information and probabilistic assumptions during learning.
compress
“Outside Tree”
expand
“Inside Tree”
bottleneck
Can find best rank-k bottleneck via matrix
factorization spectral method
(a)
Geoff Gordon—ICML tutorial—July, 2012
(b)
3
for regularization, don’t just want a state: want a low-d (hopefully minimal) one
so, select k (linear combinations of) predictions which explain as much variance in future as possible
Figure 1: (a) The classical interpretation of state as
a compact time bottleneck in dynamical systems. (b)
The analogy of predictive state in belief propagation
over latent-variable graphical models (and their corresponding latent junction trees). Blue nodes indicate
observable variables and white nodes indicate latent
variables.
the other across the separator set in order to compute
the posterior distribution of downstream variables using the evidence information from the upstream variables. Therefore, when the belief propagation agent
visits each separator set, the message that it is sending over can be essentially viewed as a representation
of the current “state” of the junction tree inference
system at that paticular separator set. This important analogy is illustrated in Figure 1(b) below.
Once this analogy of state is built for junctions trees,
we can further generalize the idea of PSR onto junction
trees. Now instead of using the partial result of sumproduct calculations over the “past” part (i.e., outside tree) of the junction tree as the state as well as
the message to send across the current separator set
(as were done by the classical Shafer-Shenoy algorithm
(Shenoy & Shafer, 1990) and Hugin algorithm (Lauritzen & Spiegelhalter, 1988; Anderson et al., 1989)),
we can use the predictions of the success probabilities
of a core set of future tests consisting of observations
in the “future” part (i.e., inside tree1 ) of the junction
tree to represent the state of the junction tree at the
current separator set, and send this prediction vector as the message into the inside tree during belief
propagation (see Figure 1(b)). More concretely, at a
separator set in a junction tree, among all the observable variables in the inside tree, there always exists a
minimal subset of observable variables whose posterior
joint probabilistic distributions given the evidence information passed in from the outside tree would completely determine the posterior joint distribution of all
the observable variables in the inside tree. This minimal subset of observable variables in the inside tree
would be designated as the core set of variables
1
The inside tree T (S) of a separator set S is defined
to be the subtree rooted at S (including S) in the juntion
tree; the outside tree is defined to be rest of the junction
tree excluding the inside tree T (S).
Borui Wang, Geoffrey J. Gordon
for this particular separator set, and thus all the possible joint value realizations of these variables can be
defined as the core set of tests for a PSR over the latent junction tree. Therefore, the predictive state at
a separator set can be defined as the posterior joint
probability distribution of its core set of variables (or
equivalently a sufficient statistic for this joint distribution).
Now during the inference, instead of passing partial
sum-product results around as the messages, we can
pass the predictive state vectors around as the new
form of predictive messages to calculate the desired
posterior. And during the learning process, we can use
two-stage instrumental regression to learn the operators at each separator set that process incoming predictive messages and distribute outgoing predictive messages, purely based on the observed quantities in the
training dataset. We name this novel message passing
framework predictive belief propagation.
…
…
Cp
β(S)
…
S
outside tree T\T(S)
C
α(SK)
α(S2)
…
k=1
cannot be directly observed in the training data. However, this estimation of W S that we would obtain from
K
can overlap with the variables in ∪ α(Sk ) and thus
k=1
K
the noise term on ⊗ θ
α(S)
T(S2)
inside tree T(S)
Figure 2: Illustration of a typical non-leaf separator
set S in a latent junction tree
3
K
⊗ E[θSk (α(Sk )) | E] and E[θS (α(S)) | E] themselves
θS (α(S)) may be biased, because the variables in α(S)
T(SK)
…
T(S1)
K
pirical samples of ⊗ θSk (α(Sk )) and θS (α(S)), since
k=1
…
…
Therefore, the goal of our new learning algorithm under this PSR representation is mainly to learn these
linear operators W S for each of the non-leaf separator sets S in the latent junction tree from our training data. A first naive attempt for this would be
to perform a linear regression directly using the em-
K
S2
α(S1)
for any outside tree evidence E.
linear regression over samples of ⊗ θSk (α(Sk )) and
SK
S1
K
⊗ E[θSk (α(Sk )) | E] = W S ×S E[θS (α(S)) | E]
k=1
k=1
Root
…
dence information that we observed in the outside tree
T \ T (S) of S, then we have that the outer product of E[θS1 (α(S1 )) | E], E[θS2 (α(S2 )) | E], ... ,
E[θSK (α(SK )) | E] is completely determined by a linear function of E[θS (α(S)) | E], for any E, because
α(S1 ), α(S2 ), ... , α(SK ) are all sets of variables that
are contained in the inside tree below S. That is to
say, there exists a linear operator W S such that2 :
Two-Stage Regression
After determining the form of messages that we are
passing around during the inference process, we now
need to figure out how different messages relate to each
other and how we could learn these inter-message relationships from the training data. Without loss of
generality, let’s consider a non-leaf separator set S in
a latent junction tree T , where S is connected with K
“childen” separator sets {S1 , S2 , ..., SK } below it (see
Figure 2). Let α(S) indicate the core set of observable variables that is associated with a separator set
S. Then according to the definition of PSR on latent junction trees, knowing the posterior joint probability distribution over α(S) would completely determine the posterior joint probability distribution over
all the observable variables in the inside tree T (S)
below S. Let θS [α(S)] denote a sufficient statistics
feature vector for the posterior joint distribution over
α(S) of a separator set S, and let E denote the evi-
k=1
Sk
(α(Sk )) and θS (α(S)) can
be correlated with each other. One powerful method
to overcome this problem of correlated noise in modern statistics and econometrics is instrumental variable regression (Stock & Watson, 2011; Hefny et al.,
2015). In our current setting, a valid instrumental variable must satisfy the criteria that it is correlated with
the input variable θS (α(S)), and that it is not correK
lated with the noise term on ⊗ θSk (α(Sk )). Thus an
k=1
ideal choice of instrumental variable that satisfies the
above criteria would be some features of the observable variables that are contained in the outside tree
T \T (S), because these observable variables in the outside tree are strongly correlated with θS (α(S)) but also
K
do not overlap with the variables in ∪ α(Sk ). Now let
k=1
β(S) indicate the set of observable variables in the outside tree T \ T (S) whose features η S (β(S)) we decide
to use as our instrumental variable. Then we can efficiently obtain an unbiased estimation of the linear operator W S in three straightforward steps: (1) regress
K
θS (α(S)) upon η S (β(S)); (2) regress ⊗ θSk (α(Sk ))
k=1
2
The mathematical notation ×S indicates the modespecific tensor multiplication along mode S.
Learning General Latent-Variable Graphical Models with Predictive Belief Propagation and HSE
upon η S (β(S)); (3) run a final linear regression from
the predictions obtained in step (1) to the predictions
obtained in step (2) to recover an unbiased estimation
of the value of W S . These three steps of supervised
learning constitute the two-stage regression procedure in our new learning algorithm.
Root
A
A
G
A
H
B
J
D
B
Main Algorithm
In this section, we present our main algorithm3 for
learning general latent-variable graphical models in
two consecutive parts: the learning algorithm followed
by the inference algorithm. Here we will describe the
basic version of the algorithm for the case where all the
variables are discrete-valued, and the extension to handle models with continuous-valued variables is given in
section 5. We give the full proof of the consistency of
our algorithm in the appendix.
At a high level, the desired input-output behavior of
our algorithm can be formulated as:
Input: the topology of the latent-variable probabilistic graphical model G, a training set of N i.i.d.
samples {xd1 , ..., xd|O| }N
d=1 , a set of observed evidence
{Xi = xi }i∈E (where E denotes the set of index for
the set of observable variables that are observed as
evidence), and the index Q of the query node XQ .
Output: Estimated posterior probability distribution
of the query node conditioned on the observations of
b Q | {Xi = xi }i∈E ].
the set of evidence nodes: P[X
4.1
The Learning Algorithm:
Model Construction: Convert the latent-variable
graphical model G into an appropriate corresponding
latent-variable junction tree T such that each observable variable in the junction tree T can be associated
with one leaf clique in T . (See Figure 3(a) and 3(b)
for an example.)
Model Specification: For each separator set S in
the junction tree T , among all the observable variables
that are associated with its inside tree T (S), determine a minimal subset α(S) of observable variables4
3
Some mathematical notations that we will use here are:
O denotes the set of observed variables {X1 , ..., X|O| }; H
denotes the set of hidden variables {X|O|+1 , ..., X|O|+|H | };
N( ) denotes the number of elements in a set; N ( ) denotes the number of possible realization values of a discrete
random variable; l( ) denotes the length of a vector; I(a)
denotes the order index of a certain value a among all the
possible realization values of a discrete random variable A;
γ(C) denotes the set of all separator sets that are connected
with a clique node C in a junction tree.
4
For each separator set Sl that is right above a leaf
clique node Cl , we require that α(Sl ) must be the set of
observable variables that are associated with Cl .
B
D
D
B
E
A
D
F
D
G
A
H
C
D
4
D
C
A
B
F
B
B
B
E
D
F
F
J
(a)
(b)
Figure 3: (a) An example latent-variable graphical
model. Blue nodes indicate observable variables and
white nodes indicate latent variables. (b) A corresponding latent junction tree that the graphical model
in (a) is converted to. Pink squares indicate clique
nodes and yellow squares indicate separator sets. Variable nodes with red circles around them are associated
with their current leaf clique nodes.
whose posterior joint probabilistic distributions given
the evidence information passed in from the outside
tree T \ T (S) would completely determine the posterior joint distribution of all the observable variables in
the T (S). This is the core set of variables for S, and
will be denoted as α(S) = {A1 , A2 , ..., AN[α(S)] }. Then
among all the observable variables that are associated
with the outside tree T \ T (S), select a subset of variables β(S) = {B1 , B2 , ..., BN[β(S)] }. Now pick features
for the inside tree φS = θS [α(S)] and the outside tree
ψ S = η S [β(S)], where we require that φS must be a
sufficient statistic for P[α(S) | β(S)]. (See Figure 4 for
an example.)
Stage 1A Regression (S1A): At each non-leaf separator set S in the latent junction tree T , learn a (possibly non-linear) regression model to estimate φ¯S =
E[φS | ψ S ]. The training data for this regression model
n
oN
are
θS (α(S)d ), η S (β(S)d )
across all N i.i.d.
d=1
training samples.
Stage 1B Regression (S1B): At each non-leaf separator set S in the latent junction tree T , where
S is connected with K “childen” separator sets
{S1 , S2 , ..., SK } below it (see Figure 3), learn a (possiK ¯
bly non-linear) regression model to estimate ⊗ φSk =
k=1
K
E[ ⊗ φ
Sk
S
| ψ ]. The training data for this regression
n K
oN
model are
⊗ θSk (α(Sk )d ), η S (β(S)d )
across
k=1
k=1
d=1
all N i.i.d. training samples.
Stage 2 Regression (S2): At each non-leaf separator set S in the latent junction tree T , use the feature
expectations estimated in S1A and S1B to train a lin-
Borui Wang, Geoffrey J. Gordon
K ¯
ear regression model to predict ⊗ φSk = W S ×S φ¯S ,
k=1
where W S is a linear operator associated with S.
The training data for this model are estimates of
K ¯
⊗ φSk , φ¯S for all the training samples that we obk=1
tained from S1A and S1B regressions.
(
~1, if X is not observed
ζ(X) =
ex , if X is observed to have value x
Then the upward message that we send from C to its
parent clique node Cp can be calculated as:
†
mC→Cp = ΦS ×{δ(C)}
⊗ ζ(X)
X∈δ(C)
Root Covariance Estimation: At the root clique
node Cr of the latent junction tree T , estimate the
expectation of the outer product of the inside tree feature vectors of all adjecent separator sets that are connected with Cr by taking average across all the N i.i.d.
training samples:
N
P
⊗ θS (α(S)d )
S∈γ(C
r)
d=1
b
.
T Cr = E[
⊗ θS (α(S))] =
N
S∈γ(Cr )
Message Passing:
From leaf to root (the upward phase):
For each non-root parent clique node C in T where
it is separated by a separator set S from its parent
node Cp , once it has received the messages from all of
its children nodes C1 , C2 , ..., CK , which are separated
from it by the separator sets S1 , S2 , ..., SK repectively,
compute
the upward message:
m
C→Cp
Root
A
S
A
Outside Tree
C
= WS
AB
A
Inside Tree
B
B
B
B
D
B
E
D
D
A
(S AB )
B
D
D
F
D
G
A
H
F
F
J
↵(S AB )
Figure 4: Illustration of the inside tree, outside tree,
α set and β set for an example separator set S AB .
4.2
×Sj
mCj →C
j∈{1,2,...,K}
and then send this message to its parent node Cp .
B
D
= W S ×S1 mC1 →C ×S2 mC2 →C ×S3 ... ×SK mCK →C
The Inference Algorithm:
Design Tensor Construction: For each leaf separator set Sl of the latent junction tree T , construct an
inside feature design tensor with modes φSl , A1 , A2 ,
..., AN[α(Sl )] , and dimensions l(φSl )×N (A1 )×N (A2 )×
... × N (AN[α(Sl )] ). Each fiber of ΦSl along mode φSl
would be the realization vector of the inside feature
function θSl at the corresponding values for variables
A1 , A2 , ..., AN[α(Sl )] , i.e.,
ΦSl :, I(a1 ), I(a2 ), ..., I(aN[α(Sl )] )
= θSl (a1 , a2 , ..., aN[α(Sl )] )
Initial Leaf Message Generation: At each leaf
clique node C of the latent junction tree T , let δ(C)
indicate the set of observable variables that are contained in C, and let S indicate the separator set right
above C. We define a function ζ(X) of an observable
variable X that evaluates to an all-one vector if X is
not observed in the evidence and evaluates to a onehot-encoding value indicator vector ex if X is observed
to have value x in the evidence. That is to say:
At the Root:
At the root clique node Cr in T where it is surrounded
by all of its children node C1 , ..., CK (each separated
by the separator sets S1 , ..., SK respectively), for each
k ∈ {1, 2, ..., K}, once Cr has received all the K − 1
upward messages from all of its other K − 1 children
cliques except for Ck , compute the downward message:
mCr →Ck
= T Cr ×S1 mC1 →Cr ×S2 ... ×Sk−1 mCk−1 →Cr
×Sk+1 mCk+1 →Cr ×Sk+2 ... ×SK mCK →Cr
= T Cr
×Sj
mCj →Cr
j∈{1,2,...,K}\k
and then send this message to its childen node Ck .
From root to leaf (the downward phase):
For each non-root parent clique node C in T where it
is separated from its parent node Cp by a separator set
S and separated from its children nodes C1 , C2 , ..., CK
by the separator sets S1 , S2 , ..., SK respectively, once
it has received the downward message mCp →C from
its parent node Cp , for each k ∈ {1, 2, ..., K}, compute
the downward message:
mC→Ck
= W S ×S mCp →C ×S1 mC1 →C ×S2 ... ×Sk−1 mCk−1 →C
×Sk+1 mCk+1 →C ×Sk+2 ... ×SK mCK →C
= W S ×S mCp →C
×Sj
mCj →C
j∈{1,2,...,K}\k
and then send this message to its children node Ck .
Learning General Latent-Variable Graphical Models with Predictive Belief Propagation and HSE
Computing Query Result: For the query node XQ ,
locate the leaf clique node that XQ is associated with
and call it CQ . Call CQ ’s parent node Cp and the separator set between them SQ . We first use the design ten†
sor ΦSQ and its Moore-Penrose psuedoinverse ΦSQ to
transform the downward incoming message mCp →CQ
and the upward outgoing message mCQ →Cp , respectively, and then compute the Hadamard product of
these transformed versions of the two messages to obtain an estimate of the unnormalized conditional probability of δ(CQ ) given all the evidence {Xi = xi }i∈E :
b
P[δ(C
Q ) | {Xi = xi }i∈E ] ∝
†
mCp →CQ ×SQ ΦSQ ◦ ΦSQ ×SQ mCQ →Cp
We can then marginalize out the variables in
δ(CQ )\XQ and renormalize to obtain the final query
result - the estimate of the conditional probability distribution of the query variable XQ given all the evib Q | {Xi = xi }i∈E ].
dence: P[X
[Additional Note]: Another important type of queries
that is very commonly encountered in practice and
that we can also easily compute here is the joint probb
ability of all the observed evidence: P[{X
i = xi }i∈E ].
It turns out that in the previous step, before marginalization and renormalization, the Hadamard product is
actually equal to:
†
mCp →CQ ×SQ ΦSQ ◦ ΦSQ ×SQ mCQ →Cp
b
= P[δ(C
Q ), {Xi = xi }i∈E ]
(See the appendix for the proof.) Now if we marginalize out all the variables in δ(CQ ) from the above probb
ability table, we can easily obtain P[{X
i = xi }i∈E ].
4.3
Proof of Consistency
The learning and inference algorithm that we presented above can be proved to be a consistent estimator of the inquired ground truth conditional probability distribution P[XQ | {Xi = xi }i∈E ]. Here we list
the four key results in the derivation of the proof, and
we present the actual full proof in the appendix.
Theorem 1. At each non-leaf separator set S, the
linear operator W S that is obtained from the Stage 2
Regression converges in probability to:
K
plim W S = P[ ∪ α(Si ) | α(S)] ×{α(S1 )} ΦS1 ×{α(S2 )}
And then the next two theorems directly follows
through mathematical induction:
Theorem 3. During the upward phase of the inference
procudure, all the upward messages that each non-leaf
non-root clique node C sends to its parent node Cp
(which is separated from C by the separator set S) converges in probability to5 :
plim mC→Cp
n→∞
= P[{Xi = xi }Xi ∈ρ(C) | α(S)] ×{α(S)} ΦS
†
Theorem 4. During the downward phase of the inference procudure, all the downward messages that each
non-leaf clique node C sends to any of its children node
Ci (which is separated from C by the separator set Si )
converges in probability to6 :
plim mC→Ci
n→∞
= P[{Xj = xj }Xj ∈µ(Ci ) , α(Si )] ×{α(Si )} ΦSi
5
Extending to Continuous Domain
through Hilbert Space Embeddings
of Distributions
One of the biggest advantages of our new learning
framework is that it can be seamlessly entended from
the discrete domain to the continuous domain in a
very general non-parametric fashion. Previously all
the existing algorithms for learning graphical models
with continuous variables would have to make certain
parametric assumptions (such as multivariate Gaussian) about the function forms that they are using
to model or approximate the continuous distributions
of variables. However, in many real-world data, the
underlying continuous distributions are highly complex and irregular, and such parametric approach
would severely limit the modeling expressiveness of
the graphical models and can often deviate from the
true underlying distributions. In contrast, in our algorithm, we can simply use the reproducing-kernel
Hilbert space (RKHS) embeddings of distributions as
sufficient statistic features and express all the (alternative) parameter learning and belief propagation operations as tensor algebra in the infinite-dimensional
Hilbert space, and then employ the kernel trick to
i=1
n→∞
ΦS2 ×{α(S3 )} ... ×{α(SK )} ΦSK ×{α(S)} ΦS
†
Theorem 2. At the root clique of the latent junction tree, the tensor T Cr that is obtained from the root
covariance estimation step converges in probability to:
plim T Cr = P[ ∪ α(S)] ×{α(S)} ΦS
n→∞
S∈γ(Cr )
S∈γ(Cr )
5
Here ρ(C) denotes the set of all observable variables
that are associated with the leaf clique nodes of the inside
tree T (C) that are actually being observed in the evidence.
6
Here µ(C) denotes the set of all observable variables
that are associated with the leaf clique nodes of the outside tree T \ T (C) that are actually being observed in the
evidence.
Borui Wang, Geoffrey J. Gordon
transform these operations back into tractable finitedimensional linear algebra calculations over Gram matrices. We now explain how to formulate our learning
algorithm over continuous domain with RKHS embeddings7 .
5.1
RKHS Embeddings of Continuous
Distributions
First, for each observable variables X in the latent
graphical model, specify a characteristic kernel:
def
KX (x, x0 ) = hφX (x), φX (x0 )iF
and pick the features for the inside tree to be φS =
φα(S) (α(S)), and the features for the outside tree to
be ψ S = φβ(S) (β(S)). And for each separator set Sl
that is right above a leaf clique node, we require that
φSl = φα(Sl ) (α(Sl )) =
⊗
φXj (xj ).
Xj ∈α(Sl )
n
oN
K
Now given the N i.i.d. α(S)d , β(S)d , ∪ α(Si )d
d=1
tuples in our training dataset, we define the following
linear operators:
Υα(S) = φα(S) (α(S)1 ), ... , φα(S) (α(S)N )
Υβ(S) = φβ(S) (β(S)1 ), ... , φβ(S) (β(S)N )
K
K
Υext(S) = ⊗ φα(Sk ) (α(Sk )1 )... ⊗ φα(Sk ) (α(Sk )N )
k=1
k=1
Let Υ∗ denote the adjoint (conjugate transpose) of Υ,
we also compute the corresponding Gram matrices:
α(S)∗
−1
= Cbα(S),β(S) × Cbβ(S),β(S)
× φβ(S) (b)
1
1
−1
∗
∗
=
Υα(S) × Υβ(S) ×
Υβ(S) × Υβ(S) + λI
N
N
× φβ(S) (b)
−1
∗
= Υα(S) × Gβ(S),β(S) + λN I
× Υβ(S) × φβ(S) (b)
{z
}
|
ωb
α(S)
=Υ
ωb , for Stage 1A;
and very similarly E[ ⊗ φα(Sk ) (α(Sk )) | β(S) = b] =
k=1
Υext(S) ωb for Stage 1B, where λ is the regularization
parameter.
We now collect all the estimated features for the inside
tree and the extended inside trees that we calculated
from the above S1A and S1B regressions into two operators:
Υµα(S) = Υα(S) ωβ(S)1 , ... , Υα(S) ωβ(S)N
Υµext(S) = Υext(S) ωβ(S)1 , ... , Υext(S) ωβ(S)N
And then for the stage 2 regression, we obtain the
linear operator W S as:
1
∗
WS =
Υµext(S) × Υµα(S)
N
1
−1
∗
×
Υµα(S) × Υµα(S) + λI
N
−1
∗
µext(S)
=Υ
× Gµα(S) ,µα(S) + λN I
× Υµα(S) ,
where the last equality follows from the matrix inversion lemma.
5.3
α(S)
Gα(S),α(S) = Υ
×Υ
∗
Gβ(S),β(S) = Υβ(S) × Υβ(S)
5.2
E[φα(S) (α(S)) | β(S) = b] = µ
bα(S)|β(S)=b
K
Then for each non-leaf separator set S in the latent
junction tree T , specify characteristic kernels for α(S)
and β(S):
def
Kα(S) (α(S), α(S)0 ) = hφα(S) (α(S)), φα(S) (α(S)0 )iG
def
Kβ(S) (β(S), β(S)0 ) = hφβ(S) (β(S)), φβ(S) (β(S)0 )iH
i=1
the Stage 1A and Stage 1B regressions, we would have:
Two-Stage Regression in the Hilbert
Space
After representing the continuous distributions of variables as points embedded in the corresponding RKHS,
we can easily perform the two-stage regression through
linear algebra operations over the Hilbert Space. For
example, if we are using kernel ridge regression to run
7
Here we will follow the notations of (Boots et al., 2013)
in our derivation.
Message Passing Inference in Hilbert
Space
Now we first compute the estimated root covariance
tensor:
N
P
⊗ φα(S) (α(S)d )
d=1 S∈γ(Cr )
Cr
,
T
=
N
and then the initial leaf messages that we send from
each leaf clique node C to its parent clique node Cp
would be:
mC→Cp = ⊗ ξ(X)
X∈δ(C)
where the function ξ(X) is defined to be8 :
φRX (x),
if X is observed to be x
v
ξ(X) =
φ
(x)dx
u X
, if X is not observed.
v−u
8
Let [u,v] denote the range of value that X can take.
Learning General Latent-Variable Graphical Models with Predictive Belief Propagation and HSE
In the actual calculations, we can use the empirical
average of φX (x)Rover finite number of samples in [u, v]
v
φX (x)dx
.
to approximate u
v−u
The remaining steps of the message-passing inference
procedure follow exactly the same form as in the discrete case, and eventually all the computation involved
will be nicely carried out over the finite-dimensional
Gram matrices. To our knowledge, this is the first algorithm that has the ability to learn continuous-valued
latent-variable graphical models in a completely nonparametric fashion without making any assumptions
about the function forms of the variables’ distributions.
6
Experiments
We designed three sets of experiments to evaluate the
performance of our proposed algorithm, including both
synthetic data and real data and ranging from both
discrete domain and continuous domain.
6.1
gorithm and average across all possible joint realizations of the variables (g, h, e). We report the results
in Figure 6 below. From the plot, we see that the
average KL divergence between our algorithm’s results and the ground truth posterior quickly decreases
and approaches 0 as the size of the training data increases. This result demonstrates that our algorithm
learns quickly to perform accurate inference over latent graphical models.
Synthetic Dataset
Figure 6: Our algorithm’s learning performance on the
synthetic dataset. We test our algorithm on 6 different
models. Each model has a different set of randomly
initialized parameters.
In this experiment, we test the performance of our algorithm on the task of learning and running inference
on the discrete-valued latent-variable graphical model
depicted in Figure 5 using artifically generated synthetic data and compare it with the EM algorithm.
G
J
H
D
A
Figure 7: Comparison between our algorithm and the
EM algorithm on the synthetic dataset.
?
F
B
C
E
Figure 5: A toy directed latent-variable graphical
model. Green nodes indicate observable variables and
red nodes indicate latent variables.
We also run the EM algorithm to learn the same model
with the same synthetic training dataset, and compare its performance and training time with our algorithm. The results are plotted in Figure 7. Our
algorithm achieves comparable learning performance,
but is much faster to train than EM.
6.2
We randomly set all the local conditional probability
tables in this directed graphical model as the ground
truth model parameters, and then sample a dataset
of 30000 sets of joint observations for the observable
variables. Next we apply our new algorithm to learn
this model, and evaluate its performance on the task of
inferring the posterior distribution of variable D given
the observed values at variables G, H, and E. We
compute the Kullback-Leibler divergence between our
algorithm’s inferred posterior and the ground truth
posterior calculated using the exact Shafer–Shenoy al-
DNA Splice Dataset
Here we consider the computational biology task of
classifying splice junctions in DNA sequences, and test
our learning algorithm using the DNA splice dataset
on the UCI machine learning repository (Asuncion &
Newman, 2007). This dataset contains a total of 3190
length-60 DNA sequences. Each sequence is labeled as
one of the three different categories: Intron/Exon site,
Exon/Intron site or Neither. Our goal is to learn to
classify unseen DNA instances. We adopt a generative
approach and use second-order nonhomogeneous Hid-
Borui Wang, Geoffrey J. Gordon
Figure 8: The performance of our learning algorithm
on the DNA splice dataset.
cal models with loopy structures, since the feature descriptors are continuous-valued and have complex distributions. But as we discussed in section 5, our new
algorithm can smoothly handle the learning problem
of such models through RKHS embeddings. In our
experiment, using Gaussian radial basis function kernels with bandwidth parameter σ = 3, regularization
parameter λ = 0.01 and model length K = 30, we
are able to achieve an overall recognition accuracy of
75.69%. Figure 10 plots the recognition accuracies and
the normalized confusion matrix of our results.
dern Markov Models to model the DNA splice junction
sequences. For each of the three categories, we use our
algorithm to learn a different second-order nonhomogeneous HMM. At test time, we compute the probabilities that a test instance is generated from each of
the three second-order HMMs, choosing the one with
the highest probability as the predicted category. We
are able to achieve an overall classification accuracy of
87.97%, and the detailed results are reported in Figure
8.
6.3
Human Action Recognition in Videos
In this experiment we consider the computer vision
problem of recognizing human actions from videos, using the classic KTH human action dataset (Schuldt
et al., 2004). The KTH dataset contains a total of 2391
video sequences of 6 different categories of actions:
boxing, handclapping, handwaving, jogging, running
and walking. Here we choose to use the second-order
non-homogeneous state space model to model the generative process of human action videos, as depicted
in Figure 9. For each video episode in the dataset,
we take K key frames evenly spaced across the whole
time frame, and then extract a 2800-dimensional histogram of oriented gradients (HOG) feature vector and
a 180-dimensional histogram of optical flow (HOF)
feature vector for each of the key frames. During
training, we concatenate the HOG and HOF feature
vectors into a 2980-dimensional vector that summarizes each key frame, and then use dimensionality
reduction to reduce it to 5 dimensions. These 5dimensional feature vectors are the observable variables Xt in our latent graphical model. For each of
the 6 action categories, we use our algorithm to learn
a different second-order nonhomogeneous state space
model. At test time, we compute the probabilities
P (X1 = x1 , X2 = x2 , ..., XK = xK | class label) that
a test video sequence is generated from each of the
6 second-order state space models, choosing the one
with the highest probability as the predicted action
category. This formulation poses a difficult problem
of learning continuous-valued latent-variable graphi-
Figure 9: Illustration of our second-order nonhomogeneous state space model for human action
recognition from videos.
Figure 10: The performance of our learning algorithm
on the KTH human action recognition dataset.
7
Conclusion
We have developed a new algorithm for learning general latent-variable graphical models using predictive
belief propagation, two-stage instrumental regression
and RKHS embeddings of distributions. We proved
that the algorithm gives a consistent estimator of the
inference results. We evaluate the algorithm’s learning performance on both synthetic and real datasets,
showing that it learns different types of latent graphical model efficiently and achieves good inference results in both discrete and continuous domains. We
believe that our algorithm provides a powerful and
flexible new learning framework for latent graphical
models.
Learning General Latent-Variable Graphical Models with Predictive Belief Propagation and HSE
References
Anandkumar, A., Ge, R., Hsu, D., Kakade, S., and
Telgarsky, M. Tensor decompositions for learning
latent variable models. Journal of Machine Learning
Research, 15:2773–2832, 2014.
Anderson, S., Olesen, K., Jensen, F., and Jensen, F.
HUGIN - a shell for building Bayesian belief universes for expert systems. In Proceedings of the 18th
International Joint Conference on Artificial Intelligence (IJCAI), pp. 1080–1085, 1989.
Asuncion, A. and Newman, D. J. UCI machine
learning repository.
https://archive.ics.
uci.edu/ml/datasets/Molecular+Biology+
(Splice-junction+Gene+Sequences), 2007.
Boots, B., Siddiqi, S., and Gordon, G. Closing the
learning-planning loop with predictive state representations. The International Journal of Robotics
Research, 30(7):954–966, 2011.
Boots, B., Gretton, A., and Gordon, G. Hilbert space
embeddings of predictive state representations. In
Proceedings of the 29th Conference on Uncertainty
in Artificial Intelligence (UAI 2013), pp. 92–101,
2013.
Dempster, A., Laird, N., and Rubin, D. Maximum
likelihood from incomplete data via the EM algorithm. Journal of the Royal Statistical Society, Series B, 39(1):1–38, 1977.
Hefny, A., Downey, C., and Gordon, G. Supervised
learning for dynamical system learning. In Advances
in Neural Information Processing Systems 28 (NIPS
2015), 2015.
Hsu, D., Kakade, S., and Zhang, T. A spectral algorithm for learning hidden Markov models. In Proceedings of the 22nd Annual Conference on Learning
Theory, 2009.
Lauritzen, S. and Spiegelhalter, D. Local computations with probabilities on graphical structures and
their application to expert systems. Journal of the
Royal Statistical Society, Series B, 50(2):157–224,
1988.
Littman, M., Sutton, R., and Singh, S. Predictive representations of state. In Advances in Neural Information Processing Systems 14 (NIPS 2001), 2001.
Parikh, A., Song, L., Ishteva, M., Teodoru, G., and
Xing, E. A spectral algorithm for latent junction
trees. In Proceedings of the 28th Conference on Uncertainty in Artificial Intelligence (UAI 2012), pp.
675–684, 2012.
Schuldt, C., Laptev, I., and Caputo, B. Recognizing
human actions: A local svm approach. In Proceedings of the 17th International Conference on Pattern
Recognition (ICPR 2004), 2004.
Shenoy, P. and Shafer, G. Axioms for probability and
belief-function propagation. In Proceedings of the
6th Conference on Uncertainty in Artificial Intelligence (UAI 1990), pp. 169–198, 1990.
Singh, S., James, M., and Rudary, M. Predictive state
representations: A new theory for modeling dynamical systems. In Proceedings of the 20th Conference
on Uncertainty in Artificial Intelligence (UAI 2004),
pp. 512–519, 2004.
Stock, J. H. and Watson, M. W. Introduction to
Econometrics. Addison-Wesley, 3rd edition, 2011.
| 2cs.AI
|
Repeated Games with Vector Losses: A Set-valued Dynamic
Programming Approach
arXiv:1603.04981v4 [cs.GT] 7 Jan 2018
Vijay Kamble∗
Patrick Loiseau†
Jean Walrand‡
January 9, 2018
Abstract
We consider infinitely repeated games with vector losses discounted over time. We characterize the set of minimal upper bounds on expected losses that a player can simultaneously
guarantee across the different dimensions. Specifically, we show that this set is the fixed point
of a set-valued dynamic programming operator. This approach also characterizes the strategies
that achieve these bounds. These optimal strategies are shown to be independent of the player’s
own past actions and stationary relative to a compact state space obtained by parameterizing
the set of the minimal bounds. We also present a computational procedure to approximate this
set and the optimal strategies.
We discuss two applications of our results: 1) characterization of the optimal strategy of
the uninformed player in zero-sum discounted repeated games with incomplete information on
one side; 2) characterization of the minmax optimal regret and the regret-optimal strategy in
repeated games with discounted losses. Our approximation procedure can be used to compute
approximately optimal strategies in both these applications. We illustrate this procedure by
computing approximately regret-optimal strategies for the problem of prediction using expert
advice from two and three experts under {0, 1}−losses. Our numerical evaluations demonstrate
improved performance over existing algorithms for this problem.
1
Introduction.
In several decision-making scenarios in uncertain and potentially adversarial environments, a decisionmaker cares about multiple objectives at the same time. For example, in certain defense operations,
an agent might be interested in simultaneously defending multiple targets against an enemy. In a
repeated game against an unknown opponent, a player wants to perform well relative to each strategy in some given class, in hindsight. In dynamic asymmetric information games where a player
lacks some information that other players have, a natural approach for the player is to choose a
strategy that gives appropriate worst-case guarantees simultaneously across the different underlying
possibilities. One can model many of these scenarios as a vector-valued sequential game between
the agent and an adversary.
∗
Dept. of Information and Decision Sciences, University of Illinois at Chicago ([email protected])
Univ. Grenoble Alpes, CNRS, INRIA, Grenoble INP, LIG, France, and, Max-Planck Institute for Software
Systems (MPI-SWS), Germany ([email protected])
‡
Dept. of Electrical Engineering and Computer Science, University of California, Berkeley ([email protected])
†
1
In this paper we focus on two-player repeated games with vector-valued losses, which is one of the
simplest classses of such sequential games. We give a systematic set-valued dynamic programming
approach for characterizing the lower Pareto frontier of the set of upper bounds on expected losses
that a player can simultaneously guarantee across the different dimensions. For the case of infinitely
repeated games with discounted losses, we characterize this Pareto frontier as the fixed point of a setvalued dynamic programming operator. This operator also characterizes the strategies that achieve
the different points on the frontier, which gives us hitherto unknown insights into the structure of
these optimal strategies. Finally, we leverage this characterization to propose an approximate
dynamic programming based scheme to approximate the optimal frontier and to compute simple
approximately optimal strategies.
Blackwell [10] pioneered the study of two-player vector-valued repeated games, resulting in the
“Blackwell Approachability” framework. This framework gives sufficient conditions for any convex
set of loss vectors to be approachable by a player, which means that there exists a strategy for the
player that ensures that the long-run average loss approaches this set with probability 1 regardless
of the adversary’s actions. Blackwell also defined an adaptive randomized strategy that ensures
this. Further, he proved the following remarkable “minmax” theorem: any convex set is either
approachable by the player or is excludable by the adversary, which means that the adversary has
a strategy that ensures that the long-run average loss remains outside this set with probability 1.
Approachability theory has been subsequently extended in various ways [39, 28, 33, 20]. However, the lack of both a “classical” dynamic programming based development of the ideas and the
corresponding characterizations of achievable sets for the case of discounted losses has remained
a significant theoretical gap. This paper takes a major step towards bridging this gap between
Bellman’s and Blackwell’s paradigms. A minmax theorem à la Blackwell isn’t expected to hold
for discounted losses. For instance, it doesn’t hold for finite scalar games, as we show in Appendix E. Nevertheless, our theory characterizes the minimal lower corner sets, i.e., sets of the
form {x ∈ RK : xk ≤ ak }, that can be guaranteed by a single player in the discounted repeated
game. That is, given any lower corner set, our characterization yields the answer to the query: can
the player guarantee that the expected loss vector remains within this set irrespective of the actions
of the adversary? Analogous questions for general convex polytopes of the form {x ∈ RK : Ax ≤ b}
can be answered by considering games with appropriate linear transformations of the vector losses.
Consequently, our approach bridges corresponding important gaps in other decision-making
problems where vector-valued repeated games and Blackwell’s approachability theory have found
applications. We focus on two particularly significant such applications. The first application is to
the theory of zero-sum discounted repeated games with lack of information on one side [6], where
our approach yields previously unknown insights into the structure of the optimal strategy for the
uninformed player. The second application is to the paradigm of regret minimization in repeated
games, where our theory results in the first exact characterization of the optimal regret and the
regret-optimal strategy for the case of discounted losses. We discuss both these applications in
detail.
Finally, we mention that set-valued dynamic programs have been used in other contexts in the
theory of dynamic games, for instance to compute the set of equilibrium payoffs in non-zero sum
repeated games with imperfect monitoring [3, 4]. For the use of dynamic programming in zero-sum
dynamic games one can refer to the classic paper by Shapley [36] on stochastic games. For a general
theory of dynamic programming in control problems and Markov decision processes (MDPs), see
[8, 9, 34].
The paper is organized as follows. In Section 2, we formally introduce the model of repeated
2
games with vector valued losses and define the objective of characterizing the set of losses that
can be guaranteed by a player. In Section 3, we introduce the set-valued dynamic programming
approach for characterizing these guarantees and the corresponding optimal strategies. In Section 4,
we present an approximation procedure to compute an -optimal strategy. Section 5 presents the
application to zero-sum repeated games with incomplete information on one side. Section 6 discusses
the application of our approach to regret minimization in repeated games. Section 7 concludes the
paper. The proofs of all our claims are presented in the appendix.
2
Model.
Consider a two-player vector-valued game G defined by an action set A = {1, · · · , m} for player
1, who is the decision-maker and whom we will call Alice, and the action set B = {1, · · · , n} for
player 2 who is the adversary and whom we will call Bob. For each pair of actions a ∈ A and b ∈ B,
Alice incurs a vector-valued loss r(a, b) ∈ RK .
The game G is played repeatedly in stages t = 1, 2, 3, · · · , T . Let GT denote this T -stage repeated
game. In each stage t, both Alice and Bob simultaneously pick their actions at and bt respectively,
and Alice bears the vector of losses r(at , bt ). Fix a discount factor β ∈ [0, 1). Then the vector of
total discounted losses is defined as:
X
T
T
X
t−1
t−1
β r(at , bt ) =
β rk (at , bt ); k = 1, · · · , K .
(1)
t=1
t=1
An adaptive randomized strategy πA for Alice specifies for each stage t, a mapping from the set
of observations till stage t, i.e., Ht = (a1 , b1 , · · · , at−1 , bt−1 ), to a probability distribution on the
action set A, denoted by ∆(A). Let ΠA be the set of all such strategies of Alice. Similarly, let ΠB
be the set of all adaptive randomized strategies for Bob. For a pair of strategies πA and πB , the
expected discounted loss on component k in the repeated game is given by:
RkT (πA , πB )
= EπA ,πB
X
T
β
t−1
rk (at , bt ) ,
(2)
t=1
where the expectation is over the randomness in the strategies πA and πB . Alice would like to
minimize her loss in every component k. However, reducing the loss in one dimension typically
implies increasing the loss in another dimension. For instance, consider the situation of protecting
two different targets against attacks: devoting more resources to protect one target makes the other
more vulnerable. Accordingly, it is important to characterize the set of best possible tradeoffs
between the different dimensions of the loss.
Now consider a fixed strategy πA ∈ ΠA . If Alice plays this strategy, then irrespective of the
strategy chosen by Bob, Alice guarantees that the long term expected vector loss is no larger than
T
k
max Rk (πA , πB ); k = 1, · · · , K
k ∈Π
πB
B
along each dimension. Let the set of all such simultaneous upper bounds that correspond to all the
strategies πA ∈ ΠA be defined as:
T
T
k
W ,
max Rk (πA , πB ); k = 1, · · · , K : πA ∈ ΠA .
(3)
k ∈Π
πB
B
3
Then characterizing the best possible tradeoffs across the different dimensions amounts to finding
the minimal points in the set W, i.e., its lower Pareto frontier, which is the set
VT , Λ(WT ) , {x ∈ WT : ∀ x0 ∈ WT \ {x}, ∃ k s.t. xk < x0k },
(4)
since all other points are strictly sub-optimal. Our goal in this paper is to characterize and compute
the set V∞ that can be achieved in the infinite horizon game G∞ and compute strategies for Alice
in ΠA that achieve different points in it. Our main technical contributions are as follows:
1. We show that the set V∞ is the fixed point of a set-valued dynamic programming operator
defined on the space of lower Pareto frontiers of closed convex sets with an appropriately
defined metric. We then show that the optimal strategies that guarantee different points in
this set are of the following form. V∞ can be parametrized so that each point corresponds
to a parameter value, which can be thought of as an “information state” belonging to a
compact space. Each state is associated with an immediate optimal randomized action and a
transition rule that depends on the observed action of the adversary (but not on the Alice’s
own action). In order to attain a point in V∞ , Alice starts with the corresponding state, plays
the associated randomized action, transitions into another state depending on Bob’s observed
action as dictated by the rule, plays the randomized action associated with the new state and
so on. In particular, the strategy does not depend on Alice’s own actions and it depends on
Bob’s actions only through this state that Alice keeps track of.
2. We also give a procedure to approximate V∞ and to compute an approximately optimal
strategy that only uses a coarse finite quantization of the parameter space. This strategy
can be simply implemented by a randomized finite-state automaton. Any desired diminishing
approximation error can be attained by choosing the appropriate quantization granularity
and number of iterations.
3
Set-valued dynamic programming.
We first present an informal description of our approach. Let V0 = {0}, i.e., the singleton set
containing only the zero vector in RK . We can show that one can obtain the set VT +1 from the set
VT , by decomposing Alice’s strategy in GT +1 into a strategy for the 1st stage, and a continuation
strategy for the remainder of the game from stage 2 onwards as a function of the action chosen by
both the players in the 1st stage. The inductive argument results from the fact that the minimal
guarantees that she can guarantee from stage 2 onwards are exactly the set VT . Suppose that at
the start of GT +1 , Alice fixes the following plan for the entire game: she will play a mixed strategy
α ∈ ∆(A) in stage 1. Then depending on her realized action a and Bob’s action b, from stage
2 onwards she will play a continuation strategy that achieves the upper-bound R(a, b) ∈ VT (she
will choose one such point R(a, b) for every a ∈ A and b ∈ B). Note that it is strictly sub-optimal
for Alice to choose any points outside VT from stage 2 onwards. Now this plan for the entire
game GT +1 gives Alice the following simultaneous upper bounds on the expected losses on the K
dimensions:
max
b∈B
X
a∈A
αa rk (a, b) + βRk (a, b) ; k = 1, · · · , K .
4
By varying the choice of α and the map R(a, b) we can obtain the set of all the simultaneous upper
bounds that Alice can achieve in the (T + 1)-stage game. The Lower Pareto frontier of this set is
exactly VT +1 . Thus there is an operator Φ, such that
VT +1 = Φ(VT )
for any T ≥ 0. In what follows, we will show that this operator is a contraction in the space of lower
Pareto frontiers of compact and convex sets, with an appropriately defined metric. This space is
shown to be complete, and thus the sequence VT converges in the metric to a set V∗ , which is the
unique fixed point of this operator Φ. As one would guess, this V∗ is indeed the set V∞ of minimal
simultaneous upper bounds that Alice can achieve in the infinitely repeated game G∞ .
The rest of this section formalizes these arguments. We will begin the formal presentation of
our results by first defining the space of Pareto frontiers that we will be working with.
3.1
A space of Pareto frontiers in [0, 1]K
For the remainder of the paper, 1 and 0 denote the the vector of ones and zeros, respectively, in
RK .
Definition 3.1. (a) Let u, v ∈ RK . We say that u v if uk ≤ vk for all k. Also, we say that
u ≺ v if u v and u 6= v. For some ≥ 0 if u v + 1, we say that v -dominates u. If = 0,
we simply say that v dominates u.
(b) A Pareto frontier in [0, 1]K is a subset V of [0, 1]K such that no v ∈ V is dominated by
another element of V.
(c) For two Pareto Frontiers U and V, we say that V -dominates U if for every point v ∈ V,
there is a point u ∈ U that it -dominates. If = 0, then we simply say that V dominates U.
(d) The lower Pareto frontier (or simply Pareto frontier) of S ⊂ [0, 1]K , denoted by Λ(S), is the
set of elements of S that do not dominate any another element of S.
The Pareto frontier of a set may be empty, as is certainly the case when the set is open. But
one can show that the Pareto frontier of a compact set is always non-empty. The proof is presented
in the Appendix A.
Lemma 3.1. Suppose that S is a compact subset of RK . Then Λ(S) is non-empty.
Since compactness is equivalent to a set being closed and bounded in Euclidean spaces, any closed
subset of [0, 1]K has a non-empty Pareto frontier. We next define the upset of a set, illustrated in
Figure 1.
Definition 3.2. Let A be a subset of B ⊆ RK . The upset of A in B is defined as up(A) = {x ∈
B | xk ≥ yk for all k, for some y ∈ A}, i.e., up(A) is the set of all points in B that dominate some
point in A. Equivalently, up(A) = {x ∈ B | x = y + v, for some y ∈ A and v 0}.
For a subset of [0, 1]K , we will refer to its upset in [0, 1]K as simply its upset. It is immediate
that the upset of a closed and convex subset of [0, 1]K is closed and convex. We define the following
space of Pareto frontiers:
5
Definition 3.3. F is the space of Pareto frontiers in [0, 1]K whose upset is closed and convex.
It is easy to show that F can be equivalently defined as the space of lower Pareto frontiers of
closed and convex subsets of [0, 1]K .1 We will now define a metric on this space. We first recall the
definition of Hausdorff distance induced by the L∞ norm.
Definition 3.4. Let A and B be two subsets of RK . The Hausdorff distance h(A, B) between the
two sets is defined as
h(A, B) = max{sup inf ||x − y||∞ , sup inf ||x − y||∞ }.
x∈A y∈B
y∈B x∈A
The Hausdorff distance defines a metric on the space of non-empty closed subsets of [0, 1]K , and
further, this space is compact and complete in this metric [24]. On the other hand, it only defines a
pseudometric on space of all non-empty subsets of [0, 1]K . Now, a possible straightforward metric
on the space F could be the one defined by the Hausdorff distance. But, as we show in Appendix B,
if K > 2, then a Pareto frontier in F may not be closed, and hence the Hausdorff distance at best
defines a pseudometric on F. Moreover, even this pseudometric is not appropriate for our purposes
as demonstrated by the following example.
Example: Consider a sequence of Pareto frontiers (Vn )n∈N where Vn is the union of the line
segment joining (0, 1) and (1/n, 1/n), and the segment joining (1/n, 1/n) and (1, 0), as depicted in
Figure 2. Then we would like this sequence of frontiers to converge to the Pareto frontier defined
by the singleton set {(0, 0)}, but the Hausdorff distance between Vn and {(0, 0)} doesn’t vanish
as n → ∞. Under the Hausdorff metric, the sequence (Vn )n∈N converges to the union of the line
segment joining (0, 1) and (0, 0), and the segment joining (0, 0) and (1, 0), which is not in F.
(𝟎, 𝟏)
up(V)
Vn
(𝟏/𝒏, 𝟏/𝒏)
(𝟎, 𝟎)
Figure 1: A Pareto frontier V and its
upset up(V) in [0, 1]2 .
(𝟏, 𝟎)
Figure 2: Approximations of (1 − β)V∗ for different β values with corresponding errors
It is thus clear that we need to define a different metric on F. We now proceed to define one with
the desired properties. We define the distance between two Pareto frontiers in F as the Hausdorff
distance between their upsets.2
One direction is clear since a Pareto frontier in F is the lower Pareto frontier of its upset, which is closed and
convex. The other direction follows from the observation that the upset of the lower Pareto frontier of a set is the
upset of the set itself and the upset of a closed and convex set is closed and convex.
2
Since the upsets of the sets in F are compact, the sup and the inf in the definition of the Hausdorff distance can
be replaced by min and max respectively.
1
6
Definition 3.5. For two Pareto frontiers U and V in F, we define the distance d(U, V) between
them as d(U, V) , h(up(U), up(V)).
We can then show that d is a metric on F, and F is both complete and compact in the metric d.
The latter essentially follows from the completeness and compactness of the space of closed subsets
of [0, 1]K in the Hausdorff metric. The proof is presented in Appendix A.
Proposition 3.1. a) d is a metric on F.
b) Let Vn n∈N be a sequence in F. Suppose that supm,k>n d(Vm , Vk ) → 0. Then there exists a
unique V ∈ F such that d(Vn , V) → 0.
c) Let Vn n∈N be a sequence in F. Then there is a subsequence Vnk k∈N and a V ∈ F such that
d(Vnk , V) → 0.
In the proof, it becomes clear that d induces these properties not just on F, but also on the more
general space of Pareto frontiers in [0, 1]K whose upset is closed (though not necessarily convex).
Finally, we end this section by presenting another way of defining the same metric d on F.
Definition 3.6. For two Pareto frontiers V and U in F, define
e(U, V) , inf{ ≥ 0 : ∀ u ∈ U, ∃ v ∈ V s.t. v u + 1}.
(5)
In other words, e(U, V) is the smallest ≥ 0 such that U -dominates V (note that e is not a
symmetric distance).3 We can then show the following.
Proposition 3.2. For any two Pareto frontiers V and U in F,
d(U, V) = max(e(U, V), e(V, U))
This means that the distance d between two frontiers V and U is less than or equal to if both U
and V -dominate each other. This way of defining d is attractive since it does not require defining
upsets of the Pareto frontiers as we do in Definition 3.5. The proof of this equivalence is presented
in Appendix A.
3.2
Dynamic programming operator and the existence of a fixed point.
By scaling and shifting the losses, we assume without loss of generality that rk (a, b) ∈ [0, 1 − β] for
all (a, b, k). Accordingly, the total discounted rewards of the game take values in [0, 1] irrespective
of the time horizon. Now, for any set S ⊆ [0, 1]K , define the following operator Ψ that maps S to a
subset of RK :
Ψ(S) =
max
b∈B
X
a∈A
αa rk (a, b)+βRk (a, b) ; k = 1, · · · , K
: α ∈ ∆(A), R(a, b) ∈ S ∀ a ∈ A, b ∈ B .
(6)
7
(𝟎, 𝟐)
(𝟒, 𝟐)
(𝟐, 𝟐)
𝒖 𝜶
𝟏
𝟏
𝟐
(𝟎, 𝟐)
(𝟐, 𝟎)
(𝟐 𝟏 − 𝜶 , 𝟐𝜶)
(𝟑, 𝟏)
(𝟐𝜶 + 𝟒 𝟏 − 𝜶 , 𝟐 𝟏 − 𝜶 )
𝟐
(𝟐, 𝟎)
(𝟒, 𝟐)
(𝟐, 𝟎)
Figure 4: Construction of the set Ψ(S) where S =
{(0, 0)}, for the game shown in Figure 3.
Figure 3: A game with vector losses.
This operator can be interpreted as follows. Assuming that S is the set of vectors of simultaneous
upper bounds on expected losses that Alice can ensure in GT , Ψ(S) is the set of vectors of simultaneous upper bounds on expected losses that she can ensure in GT +1 . If S is convex then Ψ(S) is
not necessarily convex as we demonstrate in the following example.
Example: Consider the game depicted in Figure 3. Suppose that S = {(0, 0)}, which is convex.
Then for any discount factor β and any (α, 1 − α) where α ∈ [0, 1], one obtains the guarantee:
u(α) = (max(2(1 − α), 2α + 4(1 − α)), max(2α, 2(1 − α))).
This is depicted in Figure 4. Thus, by varying α, we find that the set Ψ(S) is the union of the line
segment joining points (2, 2) and (3, 1) and the segment joining the points (3, 1) and (4, 2). Clearly,
this set is not convex.
Despite the fact that the operator Ψ doesn’t preserve convexity, we can nevertheless show that
if V is a Pareto frontier in F (which means that it is the Pareto frontier of a convex and closed
set), then the Pareto frontier of Ψ(V) is also in F (observe that in the example above, the Pareto
frontier of Ψ({(0, 0)}) is the line segment joining points (2, 2) and (3, 1)). Further, we can also show
that if V ∈ F is the set of vectors of simultaneous upper bounds on expected losses that Alice can
ensure in GT , then in any optimal plan for Alice in GT +1 , the continuation strategy from stage 2
onwards need not depend on her own action in stage 1 (but it could depend on Bob’s action in the
first stage). The proof of this result is presented in Appendix A.
Lemma 3.2. Let V ∈ F. Then:
1. Λ(Ψ(V)) ∈ F.
2. Any point u in Λ(Ψ(V)) is of the form:
u=
max
b∈B
X
a∈A
αa rk (a, b) + βQk (b) ; k = 1, · · · , K .
where Q(b) ∈ V for all b ∈ B.
3
The inf in the definition can be replaced by a min since e(U, V) can be equivalently defined as inf{ ≥ 0 : ∀ u ∈
up(U), ∃ v ∈ up(V) s.t. v u + 1}, and up(U) and up(V) are compact sets for any U, V ∈ F.
8
We next define the following dynamic programming operator Φ on F.
Definition 3.7. (Dynamic programming operator) For V ∈ F, we define Φ(V) = Λ(Ψ(V)).
From Lemma 3.2, we know that Φ(V) ∈ F whenever V ∈ F. Next, we claim that Φ is a
contraction in the metric d.
Lemma 3.3. e(Φ(U), Φ(V)) ≤ βe(U, V), and hence d(Φ(U), Φ(V)) ≤ βd(U, V).
Finally, we show that the dynamic programming operator has a unique fixed point and starting
from a Pareto frontier in F, the sequence of frontiers obtained by a repeated application of this
operator converges to this point.
Theorem 3.1. Let V ∈ F. Then the sequence (An = Φn (V))n∈N converges in the metric d to a
Pareto frontier V∗ ∈ F, which is the unique fixed point of the operator Φ, i.e., the unique solution
of Φ(V) = V.
We can then show that V∗ is indeed the optimal set V∞ that we are looking for.
Theorem 3.2. V∞ = V∗ .
3.3
Optimal strategies: existence and structure.
For a Pareto frontier V ∈ F, one can define a one-to-one function from some compact parameter set
P to V. Such a function parameterizes the Pareto frontier. We present one such parameterization
that will be used later in our approximation procedure. Define the set
P , ∪K
k=1 {(p1 , · · · , pk−1 , 0, pk , · · · , pK−1 ); pr ∈ [0, 1] for all r = 1, · · · , K − 1}.
P is thus the union of K, K − 1 dimensional faces of the hypercube [0, 1]K , where each face
is obtained by pinning the value along one dimension to 0. For instance for K = 2, we have
P = [0, 1] × {0} ∪ {0} × [0, 1], i.e., the union of the line segment joining (0, 0) and (0, 1), and the
segment joining (0, 0) and (1, 0). Now consider the function F : P × F → RK , where one defines,
F(p, V) = arg min t
(7)
x
s.t. x = t1 + p, t ∈ R,
x u, u ∈ V.
F(p, V) is essentially the component-wise smallest point of intersection of the line x = t1 + p (for
a fixed p) with the upset of V in [0, 2]K . In R2 , this is simply the family of lines y = x + p0 where
p0 = p2 − p1 , for p0 ∈ [−1, 1] (see Figure 5).
9
y = x + p0
(𝟎, 𝟐)
(𝟎, 𝟏)
V
F (p0 , V)
(𝟏, 𝟎)
(𝟎, 𝟎)
(𝟐, 𝟎)
Figure 5: Parameterization of V.
Then for a given V, the function F(., V) : P → up(V) defines a map such that for every point
u on V, there is a unique p ∈ P that maps only to that point. This is the p such that the line
x = t1 + p intersects V at u (if the line intersects V at two or more points, then one of those points
is dominated by the other(s), which is a contradiction). Note that for some values of p, the line
x = t1 + p may not intersect V, but it will definitely intersect the upset of V in [0, 2]K , which is
why in (7), we optimize over x that dominate u ∈ V, rather than directly optimizing over u ∈ V.
We can now express the dynamic programming operator in the form of such a parametrization.
Assume that V∗ is such that V∗ = Φ(V∗ ). Then for p ∈ P, one can choose α(p) ∈ ∆(A) and
q(b, p) ∈ P for each b ∈ B such that for k ∈ {1, · · · , K},
Fk (p, V∗ ) = max{
b∈B
X
αa (p)rk (a, b) + βFk (q(b, p), V∗ )}.
(8)
a∈A
Then we have the following result.
Theorem 3.3. For any p1 ∈ P, the upper bound F(p1 , V∗ ) ∈ V∗ on losses is guaranteed by Alice
in the infinite horizon game by first choosing action a1 ∈ A with probability αa1 (p1 ). Then if Bob
chooses an action b1 ∈ B, the optimal guarantees to choose from the second step onwards are then
βF(p2 , V∗ ) in βV∗ , where p2 = q(b1 , p1 ), which can be guaranteed by Alice by choosing action
a2 ∈ A with probability αa2 (p2 ), and so on.
This implies that P can be thought of as a compact state space for the strategy. Each state is
associated with an immediate optimal randomized action and a transition rule that depends on the
observed action of Bob. In order to attain a point in V∗ , Alice starts with the corresponding state,
plays the associated randomized action, transitions into another state depending on Bob’s observed
action as dictated by the rule, plays the randomized action associated with the new state and so on.
In particular, the strategy does not depend on the past actions of Alice and it depends on the past
actions of Bob only through this information state that Alice keeps track of. It is interesting to note
that unlike in stochastic games or Markov decision processes (MDPs) [34], the state transitions
are not exogenously defined but they are endogenously specified by the dynamic programming
operator.
10
4
Approximation.
In general, except for simple examples (we will see such an example in Section 6.3.2), it is difficult
to analytically compute V∗ and {(α(p), q(b, p)) : p ∈ P} that satisfies (8) by simply using the
fixed point relation. Hence, we now propose an approximate dynamic programming procedure to
approximate the optimal Pareto frontier and devise approximately optimal strategies. In order
to do so, we need to first define an approximation of any Pareto frontier. Consider the following
approximation scheme for a Pareto frontier V ∈ F. For a fixed positive integer N , define the set
1 2
N −1
, ,··· ,
, 1} ∀ r = 1, · · · , K − 1}.
N N
N
(9)
PN is thus obtained by approximating each of the K, K − 1 dimensional faces in P by a grid of
points. The number of points in this set is
PN = ∪K
k=1 {(p1 , · · · , pk−1 , 0, pk , · · · , pK−1 ); pr ∈ {0,
H(K, N ) , K(N + 1)K−1 − (K − 1).
Define the approximation operator to be
ΓN (V) = Λ ch F(p, V) : p ∈ PN
,
(10)
(11)
where F(p, V) was defined in (7). Here ch denotes the closed convex hull of a set. Thus ΓN (V) ∈ F
is the Lower Pareto frontier of a convex polytope, and it has at most H(K, N ) vertices, where each
vertex is the point of intersection of the line x = t1 + p with the upset of V for some p ∈ PN . We
can then show the following approximation guarantee.
Proposition 4.1. Consider a V ∈ F. Then
e(ΓN (V), V) = 0 and e(V, ΓN (V)) ≤
and hence
d(V, ΓN (V)) ≤
1
,
N
1
.
N
Next, we can express the compound operator ΓN ◦ Φ via a set of explicit optimization problems
as in (7), that only take V as input:
F(p, Φ(V)) = arg min t
(12)
x
s.t. x = t1 + p, t ∈ R,
X
x
αa r(a, b) + βQ(b) ∀ b ∈ B,
a∈A
α ∈ ∆(A), Q(b) ∈ V ∀ b ∈ B.
If V ∈ F is the lower Pareto frontier of a convex polytope, then this is a linear program, and
further ΓN ◦ Φ(V) is also the lower Pareto frontier of a convex polytope. We then we have the
following result.
11
Proposition 4.2. Let G0 = {0} and let Gn = (ΓN ◦ Φ)n (G0 ). Then
1
e(Gn , V ) ≤ β and e(V , Gn ) ≤
N
∗
∗
n
And thus
1
d(V , Gn ) ≤
N
∗
1 − βn
1−β
1 − βn
1−β
+ βn.
(13)
+ βn.
Hence for any , there is a pair (N, n) such that d(V∗ , Gn ) ≤ . This result implies an iterative
procedure for approximating V∗ by successively applying the compound operator ΓN ◦ Φ to G0 ,
by solving the above linear program at each step. Since Gn is a lower Pareto frontier of a convex
polytope, with at most H(K, N ) vertices for each n, the size of the linear program remains the
same throughout. More details on solving this linear program can be found in Appendix C.
The fact that e(Gn , V∗ ) ≤ β n implies that Gn β n -dominates V∗ for all n, and thus the optimal
upper bounds in V∗ cannot be larger than in Gn + β n 1. Thus as n gets larger, the set Gn + β n 1
approaches V∗ “from above”, and in the limit, ends up within a 1/(N (1 − β)) distance of V∗ .
4.1
Extracting an approximately optimal strategy.
From Gn , one can also extract an approximately optimal strategy πn in the infinite horizon game.
Suppose α∗ (p) and Q∗ (b, p) for b ∈ B are the optimal values that solve the program (12) to
compute F(p, Φ(Gn )) for different p ∈ PN . Then these define an approximately optimal strategy
in the following class:
Definition 4.1. A H(K, N )−mode stationary strategy π is a mapping from each p ∈ PN to the
pair
1. α(p) ∈ ∆(A), and
2. q1 (b, p), · · · , qK (b, p), z(b, p) , where for all b ∈ B, qk (b, p) ∈ PN for all k = 1, · · · , K
and z(b, p) ∈ ∆K .
Here, ∆K is the unit simplex in RK . The interpretation is as follows. One starts with some
initial mode, i.e., a value of p ∈ PN . Then at any step, if the current mode is p, then Alice first
chooses action a ∈ A with probability αa (p). Then if Bob plays action b ∈ B, Alice samples the
new mode to be qk (b, p) with probability zk (b, p) for each k, and after having sampled a new mode,
plays accordingly thereafter.
∗
Now, α (p) defines α(p) in πn , and q1 (b, p), · · · , qK (b, p), z(b, p) are defined such that they
satisfy
Q∗ (b, p) =
K
X
zk0 (b, p)F(qk0 (b, p), Gn ).
(14)
k0 =1
The interpretation is as follows. If V is the lower Pareto frontier of a convex polytope with each
vertex lying on the line x = t1 + p for some p ∈ PN , Q∗ (b, p) for each b ∈ B that results from
12
solving (12) will lie on one of the faces of this Pareto frontier. Thus Q∗ (b, p) can be expressed as
a convex combination of (at most K) extreme points of the face as expressed in (14).
Let Vπn be the corresponding Pareto frontier that is attained by the strategy πn (each point on
this frontier is guaranteed by choosing different possible initial randomizations over the H(K, N )
modes). Simply from the definition of V∗ as the optimal frontier, we know that Vπn dominates V∗ ,
i.e., e(Vπn , V∗ ) = 0. But we can also show the following.
Proposition 4.3.
d(Vπn , V∗ ) ≤
1
N
1 − βn
1−β
+ 2β n +
1
N
2 − β n − β n+1
.
(1 − β)2
(15)
Thus an approximately optimal strategy can be obtained by choosing an appropriate (N, n).
Remark: The procedure to approximate the frontier and extract an approximately optimal
strategy illustrates that our characterization of the minmax optimal strategy via the fixed point of
a dynamic programming operator opens up the possibility of using several approximate dynamic
programming procedures; for instance, see Chapter 6 in [9]. Here we have not tried to determine a
computation procedure that achieves the optimal error-complexity tradeoff. For a fixed (N, n), in
order to approximate the optimal frontier, the procedure needs to solve nH(K, N ) linear programs
to give the corresponding error bound in Proposition 4.2. In our implementation described in
Appendix C, each linear program is composed of mH(K, N ) + l + 1 variables and Km + K + 1
constraints (recall that H(K, N ) is polynomial in N for a fixed K).
One can focus on two terms in the approximation error separately: the first term is the quan1
tization error which is bounded by N (1−β)
and second is the iteration error which is bounded by
n
β . The second term is relatively benign since it decays exponentially in n. The first term requires
1
N = (1−β)
to achieve an error of , which grows rapidly as β gets closer to 1. For finding an
1
approximately optimal strategy, the scaling is like (1−β)
2 , which grows even faster. But for fixed
values of β not too close to 1, the computation isn’t prohibitive and this is precisely the case where
other no-regret algorithms fare poorly. Further, this computation can be done offline, and needs to
be performed only once. The resulting approximately optimal strategy requires a bounded memory
1
(O(log(H(K, (1−β)
2 )) bits) and is very simple to implement via a look-up table.
5
Application 1: Zero-sum repeated games with lack of information on one side.
One of the most celebrated and well studied models of dynamic games with incomplete information
is a model introduced by Aumann and Maschler [6] of zero-sum repeated games with incomplete
information on one side. It is described as follows.
There are K two person zero-sum games G1 , · · · , GK , each with m actions, A = {1, · · · , m}
for player 1, who is the minimizer (Alice), and n actions B = {1, · · · , n} for player 2, who is the
maximizer (Bob). Let the loss corresponding to actions a and b of Alice and Bob respectively be
denoted by rk (a, b) in game Gk .
∞
We define the
P game G as follows. One of the K games Gk is chosen by nature with probability
pk , such that k pk = 1. This distribution is known to both the players but the actual choice of
13
the game is informed to Bob and not to Alice. Let the chosen (random) game be denoted by G.
Then this game G is played infinitely often in stages t = 1, · · · , ∞. At each stage t, Alice and Bob
play their actions simultaneously. The payoff that is incurred by the players is not observed by
Alice at any stage (if she did, the game could be identified immediately), but she observes Bob’s
actions. Intuitively, Bob’s goal is thus to try to maximize Alice’s loss while utilizing his information
efficiently, since such utilization leads to information leakage, while Alice’s goal is to minimize her
loss while utilizing any leaked information.
An adaptive randomized strategy (also called a behavioral strategy) πA for Alice specifies for
each time t, a mapping from her set of observations till time t, i.e. HtA = (a1 , b1 , · · · , at−1 , bt−1 ),
to ∆(A). A behavioral strategy πB for Bob specifies for each time t a mapping from his set of
observations till time t and the chosen game G, i.e., HtB = (G, a1 , b1 , · · · , at−1 , bt−1 ), to ∆(B). We
k ; k = 1, · · · , K), where π k is his strategy
will express the behavioral strategy πB of Bob as πB = (πB
B
conditioned on the event {G = Gk }.
We now formally specify the objectives of the two players in G∞ . For a discount factor β ∈ (0, 1)
and for a choice of strategies πA and πB of the two players, the ex-ante expected loss is given by
R(πA , πB ) = EπA ,πB ,G
X
∞
β
t−1
rG (at , bt ) =
t=1
K
X
pk EπA ,πk
X
∞
B
β
t−1
rk (at , bt ) .
(16)
t=1
k=1
Alice’s objective is to minimize this payoff while Bob’s objective is to maximize it. The minmax or
the upper value of the game is given by
V = min
πA
K
X
pk max EπA ,πk
k=1
X
∞
B
k
πB
β
t−1
rk (at , bt ) .
(17)
t=1
The minimizing strategy in the outer minimization problem is the minmax optimal strategy for
Alice; we will simply call it her optimal strategy. Similarly, the maxmin or the lower value of the
game is given by
V=
max
min
k ;k=1,··· ,K) πA
(πB
X
K
pk EπA ,πk
X
∞
B
β
t−1
rk (at , bt ) .
(18)
t=1
k=1
The optimal strategy for Bob is similarly defined as his maxmin strategy, i.e. the maximizing
strategy in the outer maximization problem. In general, we have that V ≥ V, but in this case one
can show that a minmax theorem holds and V = V [6, 37, 41, 31].
Characterizing the maxmin optimal strategy for Bob: To characterize and compute
the maxmin strategy for Bob, one can use a dynamic programming approach that exploits the
structural relationship between the original game and the game after one stage has elapsed [37, 41].
Suppose V (p) is a function that assigns to every prior probability distribution p over the game
G, the maxmin value of the associated infinitely repeated game. Then it is easy to show [37, 41]
that the maxmin value as a function of the prior p is the fixed point of the following contractive
dynamic programming operator defined on the function V : ∆K → R (here ∆K is the unit simplex
14
in RK ) :
T(V )(p) =
max
min
k ∈∆B;k=1,··· ,K) qA ∈∆A
(qB
K
X
pk Eqk ,qA [rk (a, b)]
B
k=1
K
k (b)
X
p k qB
k
; k = 1, · · · , K .
+
β(
pk qB (b))V
PK
k (b)
p
q
k
k=1
B
b∈B
k=1
X
This operator can be understood as follows. Notice that in the first stage, the probability distributions over Bob’s actions chosen by him for each of the K games as a part of his strategy πB , makes
his realized action a (potentially) informative signal of the true game chosen by nature. Since
Alice can be assumed to know this strategy in the computation of the maxmin, she can perform
a Bayesian update of her belief about the chosen game after having observed Bob’s action. Thus
once the randomization of Bob in the first stage is fixed, there is a one-stage expected loss that is
minimized by Alice, and then every realized action of Bob results in a new game with an associated
maxmin value, that is identical in structure to the original game, except that the original prior is
replaced with the posterior distribution conditional on that action (note that the state transitions
in the space of posteriors are independent of Alice’s actions). Bob thus chooses a randomization
for each of the K games that maximizes the sum of these two values. Consistency then requires
that the function V (p) has to be the fixed point of this resulting operator. It also follows that the
optimal strategy for Bob is a stationary strategy that depends only on the posterior pt at stage t.
Characterizing the minmax optimal strategy for Alice: Now it turns out that a similar
approach cannot be used to compute and characterize the minmax optimal strategy for Alice, the
uninformed player. The problem is that in order to perform the Bayesian update as a part of
her strategy πA , Alice needs to know Bob’s strategy πB , which means that πA presupposes the
knowledge of πB , which contradicts the fact that the maxmin strategy is ‘universal’: it guarantees
that her loss is no more than V irrespective of the strategy chosen by Bob. Even if Bob’s optimal
strategy is unique, the best response strategy of Alice that computes the posterior updates at each
stage and plays optimally accordingly is vulnerable to bluffing by Bob. Thus the optimal strategy
of Alice cannot rely on the computation of these posterior distributions and must instead depend in
some form on Bob’s actions and the corresponding losses incurred in the different possible choices
of games.
To the best of our knowledge, the exact computation or even approximation of Alice’s optimal
strategy has been an open problem. Structurally, all that is known (see Corollary 3.25 in [37]) about
the optimal strategy is that Alice’s decision at stage t doesn’t depend on her own past actions, but
can depend on potentially all of Bob’s actions till time t. This also suggests the possibility that
any dynamic programming based procedure that may be developed to compute this strategy will
have a non-compact state space and thus suffer from the curse of dimensionality, i.e., the state may
include the entire history of actions.
We propose the following key step that resolves this problem. Instead of computing the upper
value V corresponding to the prior distribution p, suppose that one computes the following set:
W=
max E
k
πB
X
∞
t=1
β t−1 rk (at , bt ) ; k = 1, · · · , K : πA ∈ ΠA .
(19)
This is the set of upper bounds on losses that Alice can simultaneously achieve on the K components
of the vector of the long term discounted losses, by playing all the possible strategies in ΠA . In fact,
15
one need not compute the entire set W, but just its lower Pareto frontier Λ(W). If we determine
this set, then one can simply choose a point r(p) ∈ Λ(W) such that
r(p) = arg min
K
X
pk rk .
(20)
r∈Λ(W) k=1
The corresponding strategy of Alice that results in the simultaneous guarantee r is then the optimal
strategy in the original game. Thus we are interested in characterizing the set V∗ = Λ(W). But
this is exactly the set we characterized using the set-valued dynamic programming approach.
Using our approach, first, we immediately deduce the known fact that the optimal strategy of
Alice doesn’t depend on her own past actions (Theorem 3.3). Second, we also obtain a previously
unknown insight into its structure: Alice’s optimal strategy is stationary relative to a compact
state space as described in Theorem 3.3. This state space can be compared to the compact state
space of Bob’s optimal strategy, which is the space of probability distributions on the K games.
It is interesting to note that the state transitions for Bob’s strategy are exogenously determined
by the Bayesian updating of the posterior, whereas the state transitions for Alice’s strategy are
endogenously determined by the dynamic programming operator (as we had remarked earlier in
Section 3.3). Third, the compactness of the state space opens up the possibility to approximate
Alice’s optimal strategy, as we have demonstrated in Section 4. To the best of our knowledge, this
is the first known approximation procedure for the this problem.
Note that we solve a harder problem than the one we set out to solve, since instead of computing the minmax value corresponding to one prior p, in effect we are trying to simultaneously
compute the minmax values corresponding to all the possible priors. But it turns out that this
harder objective makes this problem suddenly become amenable to a dynamic programming based
approach. This should not be too surprising, since as we have seen for the case of the informed
player, in order to solve for the lower value corresponding to a prior p and to compute the optimal
strategy, one needs to simultaneously solve for games starting from all possible priors p ∈ ∆K .
6
Application 2: Regret minimization in repeated games.
The regret minimization paradigm models a repeated decision-making problem in an uncertain
environment as a repeated game between the decision-making agent and an adversary. The key
notion is that of “regret”, defined as the difference between the loss incurred by the decision maker’s
choice of actions, and the loss incurred by the best fixed action that could have been chosen in
hindsight against the sequence of actions chosen by the environment. The decision maker’s goal in
this game is to design an adaptive and possibly randomized strategy that minimizes the expected
regret in the worst case over all the possible sequences of actions of the environment. Since its
inception by Hannan in the 1950s [23], this formulation of an online decision-making problem under
uncertainty has been widely popular both in theory and practice. In this setting, several no-regret
√
algorithms exist, which ensure that if the decision-making horizon is N , the regret scales like O( N )
with high probability irrespective of the sequence of actions chosen by the adversary. Hence the
time-averaged regret (i.e., the cesaro sum) asymptotically vanishes in an infinitely repeated game,
not just in expectation, but with probability 1.
Blackwell [11] had noticed early on that his approachability framework can be used to obtain
no-regret strategies as formulated by Hannan by transforming the original repeated game into a
repeated game with vector-valued losses. In this new game, the number of vector components is
16
the number of actions available to the decision-maker, where each component keeps track of the
additional loss incurred by the strategy relative to the loss incurred if the corresponding action
was always chosen in the past. The goal of regret minimization in the original game now translates to the goal of minimizing the maximum worst-case expected loss across all the components
in this vector-valued game. Approachability theory immediately shows that the negative orthant
is approachable in this game, thus resulting in yet another no-regret algorithm. Stronger connections of the approachability framework with regret minimization and other learning problems like
calibration have subsequently been shown [1, 33].
In discounted repeated games, the loss criterion is the weighted sum of per-stage losses, with
the loss at stage t ≥ 1 weighted by β t−1 , where β ∈ (0, 1) is called the discount factor. Such
discounting is common in literature and is natural in practice, where minimizing current losses
is more important than √
the ones in the future. Several known no-regret algorithms guarantee an
expected regret of O(1/ 1 − β) as β → 1 in an infinitely repeated game, but for a given discount
factor β, their performance can be far from optimal. Overall, except for specific instances, exact
regret optimality has been difficult to achieve, and there is no known systematic procedure to
compute even approximately optimal algorithms.
In this scenario, we can couple Blackwell’s transformation with our dynamic programming based
approach to characterize the Pareto frontier of minimal regret relative to the different actions. Apart
from providing new insights into the structure of the regret optimal strategy, our approximation
scheme described in Section 4 allows us to compute an approximately regret optimal strategy.
We will formally present the regret minimization problem and its transformation into a vector
valued game, to which our approach can then be readily applied. We will then demonstrate the
applicability of our approximation scheme to the problem of prediction using expert advice with
binary losses and two experts. Since this is an extensively studied area, we begin by discussing
some of the relevant literature so as to provide some context for our contribution.
6.1
Related literature.
The first study of regret minimization in repeated games dates back to the pioneering work of
Hannan [23], who introduced the notion of regret optimality in repeated games and proposed
the earliest known no-regret algorithm. Since then, numerous other such algorithms have been
proposed, particularly for the problem of prediction using expert advice, see [29, 40, 14, 21], one
particularly well-known class being the multiplicative weights update class of algorithms. Other
settings with limited feedback have been considered, most notably the multi-armed bandit setting
[5, 13]. Stronger notions of regret such as internal regret, have also been studied [19, 16, 12, 38].
Regret minimization with non-uniformly weighted losses, of which the discounted loss is a special
case, has also been considered before [15, 17, 33]. While the average regret goes to zero if the weights
satisfy a non-summability condition, lower bounds exist ([15], Thm 2.7) that show that the optimal
regret is bounded away from 0 if the weights are summable, which is the case
√ with discounting.
Natural extensions of no-regret algorithms incur a cumulative regret of O(1/ 1 − β) (and hence
√
√
an average regret of O(1/ 1 − β)(1 − β) = O( 1 − β)) in this case; for instance see Thm 2.8 in
[15] and Prop. 2.22 in [33]. [17] derives better bounds for the case where future losses are given a
higher weight that current ones.
The results on exact regret minimization are few. In an early work, Cover [18] gave the optimal
algorithm for the problem of prediction using expert advice over any finite horizon T , for the
case of 2 experts, and where the losses are {0, 1}. [22] recently extended the result to the case
17
of 3 experts for both the finite horizon and geometrically distributed random horizon problems.
Although a geometric time horizon model seems to be related to the infinite horizon model with
discounted losses, the two problem formulations define regret differently, and thus lead to different
optimal regrets. We discuss this in Section 6.3.1. [2] considered a related problem, where a gambler
places bets from a finite budget repeatedly on a fixed menu of events, the outcomes of which
are adversarially chosen from {0, 1} (you win or you lose), and characterized the minmax optimal
strategies for the gambler and the adversary. [30] considered a similar repeated decision-making
problem where an adversary is restricted to pick loss vectors (i.e., a loss for each action of the
decision-maker in a stage) from a set of basis vectors, and characterized the minmax optimal
strategy for the decision-maker under both, a fixed and an unknown horizon. Most of the approaches
in these works are specific to their settings, and exploit the assumptions on the structure of the
loss vectors. But if the loss vectors are arbitrary, these approaches are difficult to generalize and
indeed it is recognized that characterizing the optimal regret and algorithm is difficult; cf. [30].4
6.2
The regret minimization problem and its transformation.
Let G be a two player game with m actions A = {1, . . . , m} for player 1, who is assumed to be
the minimizer and who we will call Alice (the decision-maker), and n actions B = {1, . . . , n} for
player 2, who is the adversary and who we will call Bob, in keeping with the previous notation.
For each pair of actions a ∈ A and b ∈ B, the corresponding loss for Alice is l(a, b) ∈ R. The losses
for different pairs of actions are known to Alice. The game G is played repeatedly for T stages
t = 1, 2, · · · , T . In each stage, both Alice and Bob simultaneously pick their actions at ∈ A and
bt ∈ B and Alice incurs the corresponding loss l(at , bt ). The loss of the repeated game is defined
P
to be the total discounted loss given by Tt=1 β t−1 l(at , bt ), where β ∈ (0, 1). We define the total
discounted regret of Alice as:
T
X
t=1
β
t−1
l(at , bt ) − min
a∈A
T
X
β t−1 l(a, bt ),
(21)
t=1
which is the difference between her actual discounted loss, and the loss corresponding to the single
best action that could have been chosen against the sequence of actions chosen by Bob in hindsight.
An adaptive randomized strategy πA for Alice specifies for each stage t, a mapping from the set
of observations till stage t, i.e., Ht = (a1 , b1 , · · · , at−1 , bt−1 ), to a probability distribution on the
action set A, denoted by ∆(A). Let ΠA be the set of all such strategies of Alice. The adversary
Bob is assumed to choose a deterministic oblivious strategy, i.e., his choice is simply a sequence of
actions πB = (b1 , b2 , b3 , · · · , bT ) chosen before the start of the game. Let ΠB be the set of all such
sequences.5 We would like to compute the worst case or minmax expected discounted regret which
4
All of the examples we discuss are of games with finite action spaces, which is the setting that we are concerned
with. But there are many works that consider exact minmax optimality in repeated games with general action sets,
with specific types of loss functions; see [26, 7, 27] and references therein.
5
Having an oblivious adversary is a standard assumption in regret-minimization literature [5, 13, 15]. This makes
sense when “nature” is modeled as an adversary in an application, which would be the case for instance in weather
forecasting. Also see Chapter 3 in [13] for a discussion of different adversary models and their implications on different
definitions of regret.
18
is defined as:
min
max EπA ,πB
X
T
πA ∈ΠA πB ∈ΠB
β
t−1
= min
max EπA ,πB max
πA ∈ΠA πB ∈ΠB
l(at , bt ) − min
a∈A
t=1
X
T
a∈A
β
t−1
t=1
T
X
β
t−1
l(a, bt )
t=1
l(at , bt ) −
T
X
β
t−1
l(a, bt ) ,
(22)
t=1
and the strategy for Alice that guarantees this value. Since Bob’s strategy πB is deterministic
and oblivious, the expectation in (22) is effectively only over the randomness in Alice’s strategy,
and moreover the second term does not depend on Alice’s strategy. Thus we can exchange the
expectation and the inner maximization and equivalently write (22) as:
min
max max EπA ,πB
X
T
πA ∈ΠA πB ∈ΠB a∈A
β
t−1
t=1
(l(at , bt ) − l(a, bt )) .
(23)
This quantity is called the pseudo-regret in regret minimization literature (see Chapter 3 in [13]).
The expected regret is equivalent to the pseudo-regret only when the adversary’s strategy is deterministic and oblivious; in general if the adversary’s strategy is randomized and adaptive, then the
pseudo-regret is smaller than the expected regret.
In order to address this objective, it is convenient to define a vector-valued game G, in which,
for a pair of actions a ∈ A and b ∈ B, the vector of losses is r(a, b) with m components (recall that
|A| = m), where
rk (a, b) = l(a, b) − l(k, b)
(24)
for k = 1, · · · , m. rk (a, b) is the single-stage additional loss that Alice bears by choosing action a
instead of action k, when Bob chooses b: the so called “single-stage regret” with respect to action k.
For a choice of strategies πA ∈ ΠA and πB ∈ ΠB of the two players, the expected loss on component
k in this vector-valued repeated game over horizon T is given by
RkT (πA , πB )
= E πA
X
T
β
t−1
rk (at , bt ) ,
(25)
t=1
where the expectation is over the randomness in Alice’s strategy. Now observe that by playing a
fixed strategy πA ∈ ΠA , irrespective of the strategy chosen by Bob, Alice guarantees that the total
k ). Suppose that we determine
expected loss on component k is no more than maxπk ∈ΠB RkT (πA , πB
B
the set of all simultaneous upper bounds that correspond to all the strategies πA ∈ ΠA , defined as:
T
T
k
W ,
max Rk (πA , πB )
: πA ∈ ΠA .
(26)
k ∈Π
πB
B
k=1,··· ,m
Then it is clear that the minmax optimal regret can be written as:
min
max max EπA
πA ∈ΠA πB ∈ΠB a∈A
X
T
t=1
β
t−1
(l(at , bt ) − l(a, bt )) = min max xk .
x∈WT
k
In fact, it suffices to characterize the lower Pareto frontier of the set WT , i.e., Λ(WT ), and the
strategies that achieve this frontier, since all other points are strictly sub-optimal. In particular, we
19
are interested in Λ(W∞ ). Here we remark that it may seem curious that instead of simply focusing
on computing the minmax optimal regret, we have expanded our objective to characterizing the
entire Pareto frontier of minimal regret relative to all actions. But this expanded objective is
precisely what opens up this problem to a dynamic programming based approach. One way to
see this is to realize that in order to achieve the minimal optimal regret, it is necessary to know
the optimal strategies that focus on reducing regret relative to certain actions more than others,
since these strategies will form the optimal response plans for the decision-maker as the profile
of accumulated regret relative to the different actions changes. We mention here that the idea
of characterizing the Pareto Frontier of all achievable regrets with respect to different actions has
been explored by [25], but this was in the specific context of prediction with expert advice with 2
experts and {0, 1}−losses, and for the finite time horizon problem.
Now we can show that the set Λ(W∞ ) (and Λ(WT ), in general) is exactly what we are able to
characterize and approximate using our set-valued dynamic programming approach. To see this,
notice that the only difference in assumptions is the set of strategies the adversary can use in the
two formulations. Recall that when we defined the dynamic programming operator for our vectorvalued repeated game, we assumed that Bob is allowed to choose any adaptive strategy (that could
in particular depend on the actions chosen by Alice). Thus our results are directly applicable to the
problem of minimizing pseudo-regret when the adversary is allowed to use a general strategy. But
we need to argue that they are also applicable to the problem of minimizing pseudo-regret/expected
regret in the case where Bob is restricted to using an oblivious deterministic strategy.
We do so by arguing that even if we restrict Bob to an oblivious deterministic strategy in
our original vector-valued repeated game, the set of minimal upper bounds on losses that Alice
can guarantee is still V∗ , the fixed point of our operator. First, recall that Bob’s best response
to Alice’s optimal strategy that achieves different points on the frontier V∗ is deterministic and
oblivious (offline). This is because Alice does not take into account her own actions to determine
the information state transitions (Theorem 3.3). In fact, if Alice is restricted to use strategies that
do not depend on her own actions chosen in the past, then the best response to such strategy
is always an offline deterministic strategy, and hence the minimal achievable frontier if Bob is
restricted to use offline deterministic strategies is V∗ . So all that we need to verify is that Alice
does not gain by using strategies that depend on her own past actions, when Bob is restricted to
using only offline deterministic strategies.
To see this is indeed the case, suppose that VT is the set of minimal guarantees that Alice can
achieve by using general randomized adaptive strategies in GT , assuming that Bob is restricted to
using deterministic offline strategies. Then in GT +1 , Alice’s strategy is a choice of a distribution over
her actions, α, and a mapping from the realized actions (a, b) to some continuation (randomized,
adaptive) strategy π(a, b) ∈ ΠA . But since Bob’s responses that maximize the losses on the different
components cannot depend on the realization of Alice’s action a, and can only depend on α, his best
responses from time 2 onwards would effectively be against the strategy π 0 (b) of Alice that chooses
the strategy π(a, b) with probability αa for each action a. Thus Alice could as well have chosen the
continuation strategy independently of her chosen action a. And this strategy ought to guarantee
a point in VT , since all other strategies that guarantee points outside VT are strictly-suboptimal.
Thus the guarantees that Alice can achieve in GT +1 is given by the set:
X
T +1
V =Λ
max
αa rk (a, b) + βQk (b) ; k = 1, · · · , K : α ∈ ∆(A), Q(b) ∈ V
∀b ∈ B
.
T
b∈B
a∈A
20
But this is exactly the dynamic programming operator in Definition 3.7. Hence we can conclude
that V∗ is indeed the set of minimal guarantees, even if Bob is restricted to using deterministic
offline strategies.
6.3
Example: Combining expert advice.
Consider the following model of combining expert advice. There are two experts who give Alice
recommendation for a decision-making task: say predicting which route will be the fastest to go
to work the next day. Each day Alice decides to act on the recommendation made by one of the
experts. The experts’ recommendations may be correct or wrong, and if Alice acts on an incorrect
recommendation, she bears a loss of 1. Otherwise she does not incur any loss. Each day, exactly
one expert is correct and the other is wrong.6 This model can be represented by the matrix shown
in Figure 6. The rows correspond to the choice made by Alice and the columns correspond to the
two different possibilities for the correct expect on each day. The matrix of single-stage regrets is
shown in Figure 7.
𝟏
𝟐
𝟏
Expert 1
𝟏
𝟎
Expert 1
Expert 2
𝟎
𝟏
Expert 2
Figure 6: Possible loss scenarios
𝟐
(𝟎, 𝟏) (𝟎, −𝟏)
(−𝟏, 𝟎) (𝟏, 𝟎)
Figure 7: Single-stage regret w.r.t. Expert 1
& 2.
1.0
1.3
(N, n)
1.2
= 0.5
(40, 8)
1.1
= 0.6
(50, 11)
1.0
0.8
0.6
= 0.7
0.4
(67, 17)
= 0.8
= 0.9
= 0.95
0.2
0.9
(101, 28)
0.8
(201, 66)
0.7
(400, 149)
0.6
0.0
0.0
0.2
0.4
0.6
0.8
0.5
1.0
0.5
0.6
0.7
0.8
0.9
0.95
β
Figure 8: Approximations of the optimal
frontier (1 − β)V∗ (β) for different β values
and the associated (N, n). (N, n) values in
each case are chosen so that the approximation error (Proposition 4.2) is less than 0.06.
n is chosen so that β n /(1 − β) ≤ 0.01.
Figure 9: Upper bounds on the optimal expected regret plotted as a function of the discount factor β.
6
It is not expected to be optimal for the adversary trying to maximize Alice’s regret to have both the experts be
correct or both be wrong. Hence these actions are omitted.
21
In Figure 8, the computed approximately optimal Pareto frontiers of regret for a range of values
of β are shown. (N, n) is chosen in each case so that the error in the approximation of (1−β)V∗ (β) is
at most 0.06. Further, n is chosen in each case so that β n /(1 − β) ≤ 0.01, and hence Proposition 4.2
guarantees that 0.01 added to the approximations of optimal regret for each β result in upper bounds
on the optimal regret. These upper bounds are plotted in Figure 9.
6.3.1
Comparison with other algorithms.
In this section, we compare the performance of approximately optimal algorithms derived from
our approach to two known algorithms for the problem of prediction with expert advice. We will
consider discount factors β = 0.8 and 0.9, and K = 2 and 3 experts.7
The first algorithm we consider is the well known the exponentially weighted average forecaster,
also known as “Hedge”. In this algorithm, if Lt (i) is the cumulative loss of expert i till time t, then
the probability of choosing expert i at time t + 1 is
pi (t + 1) ∝ exp(−ηLt (i)),
p
where η is a parameter. In the undiscounted problem, choosing η = 8 log K/T when the time
p
horizon T is known attains an upper bound of T log K/2 on the expected cumulative regret (Thm.
2.2, [15]). In a certain sense, this is shown to be asymptotically optimal in K and T for general loss
function taking values in [0, 1] (Thm. 3.7, [15]). In our implementation, we use discounted cumulap
tive losses in this algorithm, and choose η = 8 log K(1 − β 2 ). This resulting algorithm achieves an
p
p
upper bound of log K/2(1 − β 2 ) = log K/(2(1 − β)(1 + β)) on the expected discounted regret
in the infinitely repeated game (see proof of Thm. 2.2, and Thm. 2.8 in [15]).
Next, we consider the optimal algorithms given by Gravin, Peres and Sivan [22] (which we
will refer to as GPS) for the experts problem with a geometrically distributed time horizon with
K = 2 and 3 experts. In this model, at each stage, the game ends with a probability 1 − β
and continues with probability β. This is essentially the same as our model of discounted losses,
where the discount factor is interpreted as the probability of continuation at each stage. But the
difference between that formulation and our formulation is in the definition of regret. In their
formulation, the loss of the decision-maker is compared to the expected loss of the best expert
in the realized time horizon (where the expectation is over the randomness in the time horizon),
P
i.e., to ET [mini=1,2 Tt=1 lt (i)], where lt (i) is the loss of expert i at time t. On the other hand, in
our formulation, the loss of the decision-maker is compared to the expert with the lowest expected
P
P
t−1 l (i). Naturally, the optimal regret in their
loss, i.e., to mini=1,2 ET [ Tt=1 lt (i)] = mini=1,2 ∞
t
t=1 β
formulation is at least as high as the optimal regret in our formulation, and in fact it turns out to
be strictly higher. For example, for K = 2 and β = 0.9, the optimal regret in their formulation
is ≈ 1.147,8 while in our formulation, the optimal regret in this case is at most ≈ 0.9338 (see
Figure 9). Further, it is clear that their optimal strategy gives the same guarantee for our definition of regret as the optimal regret in their definition, i.e., for instance, for β = 0.9, their strategy
will guarantee an expected regret of at most ≈ 1.147 according to our definition. Note that in the
7
The code for all our numerical experiments can be found at https://github.com/vijaykamble/
vector-repeated-games.
8
The optimal regret for [22]’s formulation is √ 1 2 . The expression in [22] is off by a factor of β since they
2
1−β
discount the first period by β (i.e., the game could end before the first stage begins), whereas we discount it by 1.
22
context where discounting captures the temporal change in the present value of money, our formulation of regret is the natural one. The GPS algorithms for K = 2 and 3 are described in Appendix D.
Results for K=2.
We first consider K = 2 experts. We compare the performance of
our approximately optimal strategy with the performance of GPS and Hedge against 3 different
adversaries:
1. Adversary A: This adversary gives a loss of 1 to one expert (and 0 to the other) uniformly at
random. In the geometrically distributed time horizon model with regret as defined in [22],
this has been shown to be the maxmin optimal adversary.
√
2. Adversary B: This adversary chooses a Bernoulli(1/2 − 1 − β) loss for expert 1 and a
Bernoulli(1/2) loss for expert 2.9
3. Adversary C: In this adversary’s strategy, the probability that expert 1 incurs a loss (and
expert 2 doesn’t) at time t is 0.91/t if t is odd, and 0.9t if t is even.
Figures 11 and 10 compare the expected regret incurred by Hedge and GPS to the expected
regret incurred by our approximately optimal strategies for β = 0.8 and 0.9. For β = 0.8 we
computed a 203−mode (H(2, 101)) strategy and a 21−mode (H(2, 10)) strategy (n = 28 in both
cases). For β = 0.9 we computed a 403−mode (H(2, 201)) strategy and a 41−mode (H(2, 20))
strategy (n = 66 in both cases). The expected regret is estimated in each case by averaging over
10000 runs, where each run is a game with time horizon T = 100. The associated error bars
(±1.96 × standard error) are shown in the graph.
In both cases, our strategies significantly outperform Hedge and GPS against Adversaries B
and C, while they result in comparable performance against Adversary A. Observe that against all
the three adversaries, the regret incurred by both our strategies does not significantly exceed the
upper bounds of ≈ 0.6886 for β = 0.8 and ≈ 0.9338 for β = 0.9. The regret of GPS significantly
exceeds these values in both cases by a considerable margin against Adversary C. This in particular
eliminates the possibility of GPS being exactly optimal for our problem with high probability. As
expected, the average regret incurred by the GPS algorithm does not significantly exceed the values
of ≈ 0.8333 and ≈ 1.147 respectively, which are the optimal values of regret as defined by GPS.
Similarly, considering the performance against Adversary C, Hedge appears to be sub-optimal
by a large margin for β = 0.8. But its performance comes close to the upper bound of ≈ 0.9338
for β = 0.9. This seems aligned with the possibility of Hedge being asymptotically optimal for our
problem as β → 1.
Finally, it is also interesting to note that there is no significant difference in the performance
of the 203−mode and 21−mode strategies for β = 0.8, or in the performance of the 403−mode
and 41−mode strategies for β = 0.9. This suggests the possibility of obtaining better performance
bounds for less complex strategies, potentially using some other approximation schemes.
Results for K=3. The possible losses for the experts problem with 3 experts is shown in
Figure 12.10 The resulting matrix of single stage regret vectors are shown in Figure 13.
9
Adversary B could give a loss of 1 or a loss of 0 to both experts in some stage. In this case, our algorithm stays
23
Adversary A
0.70
0.69
0.68
0.67
GPS
Adversary B
0.62
0.60
0.58
0.56
0.54
0.52
Hedge 203-mode 21-mode GPS
Adversary C
0.80
0.75
0.70
Hedge 203-mode 21-mode
GPS
Hedge 203-mode 21-mode
Figure 10: Estimates of the expected regret of different algorithms for K = 2 and β = 0.8
against the different adversaries, along with associated error bars (±1.96 × standard
error). Upper bound on optimal regret is ≈ 0.6886.
Adversary A
Adversary B
Adversary C
1.15
0.94
1.10
0.80
1.05
0.92
0.75
1.00
0.90
GPS
0.70
Hedge 403-mode 41-mode GPS
0.95
Hedge 403-mode 41-mode
GPS
Hedge 403-mode 41-mode
Figure 11: Estimates of the expected regret of different algorithms for K = 2 and β = 0.9
against the different adversaries, along with associated error bars (±1.96 × standard
error). Upper bound on optimal regret is ≈ 0.9338.
For β = 0.8 and 0.9, we approximated the frontiers using (N, n) = (20, 20), resulting in a
H(3, 20) = 1321−mode strategy in each case. The approximate value of optimal regret was ≈ 0.9067
and ≈ 1.244 respectively. The values of β n /(1 − β) in the two cases (n = 20) are ≈ 0.057 and
≈ 1.215, resulting in upper bounds of ≈ 0.9637 and ≈ 2.459 respectively. The theoretical error is
quite high in this case and hence both these upper bounds and the strategies are expected to be
crude approximations. We compared the performance of these strategies with the performance of
GPS and Hedge against the following three adversaries.
1. Adversary D: This adversary chooses one of the six possible actions uniformly at random
at each stage.
2. Adversary E: This is an optimal adversary in the GPS model for 3 experts.11 For i ≤ j,
let dij be the difference in the cumulative losses of the expert with the ith lowest loss and the
expert with the j th lowest loss. Then if d12 = d13 = 0, then with equal probability of 1/3, the
algorithm gives a loss of 0 to one expert and a loss of 1 to each of the remaining two experts.
Otherwise, with probability (1/2), the algorithm gives no loss to the leading expert and a loss
of 1 to each of the others, and with probability (1/2) it gives a loss of 1 only to the leading
expert, and no loss to others.
in the same mode at the end of the stage.
10
The optimal adversary isn’t expected to give a loss of 0 or a loss of 1 to all experts, and hence these two actions
are omitted.
11
The optimal adversary in this case is not unique; we choose one of the many mentioned in [22].
24
𝟏
𝟐
Expert 1
(𝟎, 𝟏, 𝟏)
(𝟎, −𝟏, 𝟎)
𝟎
Expert 2
(−𝟏, 𝟎, 𝟎)
(𝟏, 𝟎, 𝟏)
𝟏
Expert 3
(−𝟏, 𝟎, 𝟎) (𝟎, −𝟏, 𝟎)
𝟏
𝟐
𝟑
𝟒
𝟓
𝟔
Expert 1
𝟏
𝟎
𝟎
𝟎
𝟏
𝟏
Expert 2
𝟎
𝟏
𝟎
𝟏
𝟏
Expert 3
𝟎
𝟎
𝟏
𝟏
𝟎
Figure 12: Possible loss scenarios with 3 experts.
𝟑
𝟓
𝟔
(𝟎, 𝟎, 𝟏)
(𝟎, 𝟏, 𝟎)
𝟒
(𝟎, 𝟎, −𝟏) (𝟎, −𝟏, −𝟏)
(𝟎, 𝟎, −𝟏)
(𝟏, 𝟎, 𝟎)
(𝟏, 𝟏, 𝟎)
(𝟏, 𝟎, 𝟎)
(𝟎, 𝟎, 𝟏) (−𝟏, 𝟎 − 𝟏)
(−𝟏, −𝟏, 𝟎)
(𝟎, 𝟏, 𝟎)
Figure 13: Single-stage regret w.r.t. Experts
1, 2 and 3.
√
3. Adversary F: This adversary gives a loss of Bernoulli(1/2 − 1 − β) to expert 1 and a loss
of Bernoulli(1/2) to each of the two remaining experts at each stage.12
The results (again, over 10000 runs with horizon T = 100) are shown in Figures 14 and 15 for
β = 0.8 and 0.9 respectively.
Adversary D
Adversary E
0.86
0.85
0.88
0.84
0.87
0.82
0.86
0.80
0.85
0.78
0.83
0.82
GPS
Hedge
1321-mode
Adversary F
0.84
GPS
Hedge
0.76
1321-mode GPS
Hedge
1321-mode
Figure 14: Estimates of the expected regret of different algorithms for K = 3 and β = 0.8
against the different adversaries, along with associated error bars (±1.96 × standard
error). Upper bound on optimal regret is ≈ 0.9637.
Adversary D
1.14
1.12
1.10
Adversary E
1.16
1.13
1.14
1.12
1.12
1.11
1.10
1.10
GPS
Hedge
Adversary F
1.14
1321-mode
GPS
Hedge
1.08
1321-mode GPS
Hedge
1321-mode
Figure 15: Estimates of the expected regret of different algorithms for K = 3 and β = 0.9
against the different adversaries, along with associated error bars (±1.96 × standard
error). Upper bound on optimal regret is ≈ 2.459.
Observe that the 1321−mode strategies, despite being crude approximations, significantly outperform both GPS and Hedge against Adversary F, and their performance is comparable to these
12
If Adversary C gives a loss of 1 or a loss of 0 to all the experts in some stage, our strategy doesn’t change its
mode at the end of that stage.
25
algorithms in the remaining cases. This again suggests the possibility of obtaining tighter error
bounds for less complex strategies through other approximation techniques (although this is far
from being a comprehensive test). Unlike for the case of K = 2, we were not able to design an
adversary in this case that would rule out the optimality of GPS or Hedge with high probability.
Note that for K = 3, the optimal regret in the GPS model is ≈ 1.1111 for β = 0.8 and ≈ 1.5294
for β = 0.9.13
6.3.2
Exact characterization of V∗ for K = 2 and β = 0.5.
In some simple examples, we can exactly determine the set V∗ by “solving” the fixed point relation
given by the dynamic programming operator. We demonstrate this by determining the optimal
Pareto frontier in the game of combining expert advice (Figure 7) from 2 experts for β = 0.5.
Note that the points (0, 1/(1 − β)) = (0, 2) and (1/(1 − β), 0) = (2, 0) lie on V∗ (achieved by
choosing Expert 1 always or Expert 2 always, respectively). We can thus represent V∗ by a convex
and decreasing function f (x) defined on x ∈ [0, 2] such that f (0) = 2 and f (2) = 0, so that
V∗ = {(x, f (x)) : x ∈ [0, 2]}. βV∗ for β = 0.5 is thus the set {(βx, βf (x)) : x ∈ [0, 2]} =
{(x, 0.5f (x/0.5) : x ∈ [0, 1]}, which thus can be represented by the convex, decreasing function
f¯(x) = f (2x)/2 defined on x ∈ [0, 1], where f¯(0) = 1 and f¯(1) = 0.
( x, 1
x) + 0.5V ⇤
( 0.5, 0.5) + 0.5V ⇤
(𝟎, 𝟐)
𝑥
(𝟎, 𝟐)
𝟎. 𝟓
up(U (x))
(𝟎, 𝟏)
up(U (0.5))
(𝟎, 𝟏)
(𝟎. 𝟓, 𝟎. 𝟓)
1−𝑥
(−𝟏, 𝟎)
(𝟎, 𝟎)
𝟎. 𝟓
V⇤
𝑥
(𝟐, 𝟎)
(𝟏, 𝟎)
(x, x
(𝟎, 𝟎)
(−𝟏, 𝟎)
1) + 0.5V ⇤
V⇤
𝟎. 𝟓
(𝟏, 𝟎)
(𝟐, 𝟎)
(0.5, 0.5) + 0.5V ⇤
(𝟎, −𝟏)
(𝟎, −𝟏)
Figure 16: Construction of up(U(x)).
Figure 17: Construction of up(U(0.5)).
Now for a fixed randomization over Alice’s actions, (1 − x, x), by choosing different points in V∗
from the next stage onwards, one obtains the set of guarantees
U(x) =
max(−x + 0.5 Q1 (1), x + 0.5 Q1 (2)),
∗
max(1 − x + 0.5 Q2 (1), x − 1 + 0.5 Q2 (2)) : Q(1), Q(2) ∈ V .
(27)
If we denote the set (−x, 1 − x) + 0.5V∗ (which is obtained by mapping each element u of V∗ to
(−x, 1 − x) + 0.5u)), by U1 (x), and the set (x, x − 1) + 0.5V∗ by U2 (x), it is straightforward to see
that
up(U(x)) = up(U1 (x)) ∩ up(U2 (x)),
13
The optimal regret in the GPS model for K = 3 experts is √ 2
3
26
1−β 2
.
where up(.) is the upset of the set in [0, 2]2 . This is depicted in Figure 16. The fixed point relation
says that
∗
V = Λ ∪x∈[0,1] up(U(x)) .
From the figure, one can see that V∗ is the curve traced by the lower left corner point of up(U(x))
as x varies between 0 and 1. Since we already know that the two extreme points on V∗ are (0, 2)
and (2, 0), for x = 0.5, we know that the lower left corner point of up(U(0.5)) is (0.5, 0.5), and
hence is contained in V∗ , as shown in Figure 17. Since we know that V∗ is symmetric around the
line x = y, we know that f (x) = f −1 (x), and thus it is sufficient to determine f (x) in the range
x ∈ [0, 0.5]. In this range, f satisfies the following fixed point relation (again, see Figure 16):
f (x) = f¯(2x) + 1 − x
= f (4x)/2 + 1 − x.
(28)
Taking the derivative twice on both sides, we obtain:
f 00 (x) = 8f 00 (4x).
√
3
This gives us f 00 (x) = ax− 2 for any a ∈ R. Integrating, we obtain f (x) = a x + x + 2. Since we
√
√
want f (0.5) = 0.5, we obtain a = −2 2. Thus we have f (x) = −2 2x + x + 2. Note that f (2) = 0,
and it turns out that f (x) restricted to the domain x ∈ [0, 2] is such that f (x) = f −1 (x). Thus
√
f (x) is the function we are looking for and V∗ = {(x, −2 2x + x + 2) : x ∈ [0, 2]}.
We can compare this exact characterization with the approximate frontier that we computed
using our approximation procedure for β = 0.5 (approximation error less than 0.06). Both these
frontiers are plotted in Figure 18. As we can observe, the two frontiers are close to identical.
1.0
Exact 0.5 V ∗ (0.5)
Approx. 0.5 V ∗ (0.5)
0.8
0.6
0.4
0.2
0.0
0.0
0.2
0.4
0.6
0.8
1.0
Figure 18: Comparison of the approximation of 0.5V∗ (0.5) and the exact characterization
√
0.5V∗ (0.5) = {(x, −2 x + x + 1) : x ∈ [0, 1]}.
Optimal Policy: To attain the point (x, f (x)) for x ∈ [0, 0.5], the optimal strategy of Alice
chooses a randomization (1−x, x); then if the adversary chooses action 1, the next point she chooses
to attain is (4x, f (4x)), whereas if he chooses action 2 then the next point she chooses to attain
is (0, f (0)). To attain the point (f (x), x) for x ∈ [0, 0.5], the optimal strategy of Alice chooses
a randomization (x, 1 − x); then if the adversary chooses action 2, the next point she chooses to
attain is (f (4x), 4x), whereas if he chooses action 1, then the next point she chooses to attain is
(f (0), 0).
27
7
Conclusion and future directions.
We presented a novel set-valued dynamic programming approach to characterize and approximate
the set of minimal guarantees that a player can achieve in a discounted repeated game with vector
losses. We showed that this optimal set is the fixed point of a contractive dynamic programming
operator, and it is the Pareto frontier of some convex and closed set. We also established the
structure of the optimal strategies that achieve the different points on this set. Finally we gave
a value-iteration based procedure to approximately compute this set and also find approximately
optimal strategies. This set-valued dynamic programming approach can also be used to characterize
and approximate optimal strategies for vector repeated games with a finite time horizon. In this
case, it follows that the optimal strategy will depend on an information state belonging to a compact
set, and also the stage t.
We showed that this approach gives us a way to compute approximately regret-minimizing
strategies for playing repeated games with discounted losses, and to approximate the optimal
strategy of the uninformed player in Aumann and Maschler’s well known model of zero-sum repeated
games with incomplete information on one side
The extension of this approach to the case of long-run average losses in infinitely repeated games
is much less straightforward, despite the fact that average cost dynamic programming for standard
dynamic optimization problems like MDPs is quite well understood. Such an extension could
potentially improve our understanding of the construction of no-regret algorithms and would fill a
significant portion of the remaining gap in viewing the classical dynamic programming paradigm
as a methodical approach to designing robust decision-making algorithms.
References
[1] Jacob Abernethy, Peter L Bartlett, and Elad Hazan, Blackwell approachability and no-regret
learning are equivalent, Proceedings of The 24th Annual Conference on Learning Theory, 2011.
[2] Jacob Abernethy, Manfred K Warmuth, and Joel Yellin, Optimal strategies from random walks,
Proceedings of The 21st Annual Conference on Learning Theory, 2008.
[3] Dilip Abreu, David Pearce, and Ennio Stacchetti, Optimal cartel equilibria with imperfect
monitoring, Journal of Economic Theory 39 (1986), no. 1, 251 – 269.
[4] Dilip Abreu, David Pearce, and Ennio Stacchetti, Toward a theory of discounted repeated games
with imperfect monitoring, Econometrica 58 (1990), no. 5, pp. 1041–1063.
[5] Peter Auer, Nicolò Cesa-Bianchi, Yoav Freund, and Robert E. Schapire, The nonstochastic
multiarmed bandit problem, SIAM Journal on Computing 32 (2002), no. 1, 48–77.
[6] Robert J. Aumann and Michael Maschler, Repeated games with incomplete information, MIT
Press, 1995.
[7] Peter L Bartlett, Wouter M Koolen, Alan Malek, Eiji Takimoto, and Manfred K Warmuth,
Minimax fixed-design linear regression, Proceedings of The 28th Annual Conference on Learning Theory, 2015.
[8] Dimitri P. Bertsekas, Dynamic programming and optimal control, vol. 1, Athena Scientific,
2005.
28
[9]
, Dynamic programming and optimal control, vol. 2, Athena Scientific, 2012.
[10] David Blackwell, An analog of the minimax theorem for vector payoffs., Pacific J. Math. 6
(1956), no. 1, 1–8.
[11]
, Controlled random walks, Proceedings of the International Congress of Mathematicians 1954 (J De Groot and J.C.H Gerretsen, eds.), vol. 3, 1956, pp. 336–338.
[12] Avrim Blum and Yishay Mansour, From external to internal regret, Journal of Machine Learning Research 8 (2007), no. Jun, 1307–1324.
[13] Sébastien Bubeck and Nicolò Cesa-Bianchi, Regret analysis of stochastic and nonstochastic
multi-armed bandit problems, Foundations and Trends in Machine Learning 5 (2012), no. 1,
1–122.
[14] Nicolò Cesa-Bianchi, Yoav Freund, David Haussler, David P. Helmbold, Robert E. Schapire,
and Manfred K. Warmuth, How to use expert advice, J. ACM 44 (1997), no. 3, 427–485.
[15] Nicolò Cesa-Bianchi and Gabor Lugosi, Prediction, learning, and games, Cambridge University
Press, 2006.
[16] Nicolò Cesa-Bianchi and Gabor Lugosi, Potential-based algorithms in on-line prediction and
game theory, Machine Learning 51 (2003), no. 3, 239–261.
[17] Alexey Chernov and Fedor Zhdanov, Prediction with expert advice under discounted loss, Algorithmic Learning Theory, Springer, 2010, pp. 255–269.
[18] Thomas M Cover, Behavior of sequential predictors of binary sequences, Tech. report, DTIC
Document, 1966.
[19] Dean P. Foster and Rakesh V. Vohra, Calibrated learning and correlated equilibrium, Games
and Economic Behavior 21 (1997), no. 1–2, 40–55.
[20] Gaëtan Fournier, Eden Kuperwasser, Orin Munk, Eilon Solan, and Avishay Weinbaum, Approachability with constraints, arXiv preprint arXiv:1712.00781 (2017).
[21] Yoav Freund and Robert E. Schapire, Adaptive game playing using multiplicative weights,
Games and Economic Behavior 29 (1999), no. 1–2, 79–103.
[22] Nick Gravin, Yuval Peres, and Balasubramanian Sivan, Towards optimal algorithms for prediction with expert advice, Proceedings of the 27th Annual ACM-SIAM Symposium on Discrete
Algorithms, 2016.
[23] James Hannan, Approximation to Bayes risk in repeated plays, Contributions to the Theory
of Games (M. Dresher, A. W. Tucker, and P. Wolfe, eds.), vol. 3, Princeton University Press,
1957, pp. 97–139.
[24] Jeff Henrikson, Completeness and total boundedness of the Hausdorff metric, MIT Undergraduate Journal of Mathematics, Citeseer, 1999.
[25] Wouter M Koolen, The Pareto regret frontier, Advances in Neural Information Processing
Systems, 2013.
29
[26] Wouter M Koolen, Alan Malek, and Peter L Bartlett, Efficient minimax strategies for square
loss games, Advances in Neural Information Processing Systems, 2014.
[27] Wouter M Koolen, Alan Malek, Peter L Bartlett, and Yasin Abbasi, Minimax time series
prediction, Advances in Neural Information Processing Systems, 2015.
[28] Ehud Lehrer, Approachability in infinite dimensional spaces, International Journal of Game
Theory 31 (2003), no. 2, 253–268.
[29] N. Littlestone and M.K. Warmuth, The weighted majority algorithm, Information and Computation 108 (1994), no. 2, 212–261.
[30] Haipeng Luo and Robert Schapire, Towards minimax online learning with unknown time horizon, Proceedings of the 31st International Conference on Machine Learning, 2014.
[31] Jean-François Mertens, Sylvain Sorin, and Shmuel Zamir, Repeated games, vol. 55, Cambridge
University Press, 2015.
[32] James R Munkres, Topology: a first course, 1975.
[33] Vianney Perchet, Approachability, regret and calibration: Implications and equivalences, Journal of Dynamics and Games 1 (2014), no. 2, 181–254.
[34] Martin L Puterman, Markov decision processes: Discrete stochastic dynamic programming,
John Wiley & Sons, 2014.
[35] Walter Rudin, Real and complex analysis, 3rd ed., McGraw-Hill, 1986.
[36] Lloyd S Shapley, Stochastic games, Proceedings of the National Academy of Sciences of the
United States of America 39 (1953), no. 10, 1095.
[37] Sylvain Sorin, A first course on zero sum repeated games, Springer, 2002.
[38] Gilles Stoltz and Gabor Lugosi, Internal regret in on-line portfolio selection, Machine Learning
59 (2005), no. 1–2, 125–159.
[39] Nicolas Vieille, Weak approachability, Mathematics of Operations Research 17 (1992), no. 4,
pp. 781–791.
[40] Volodimir G Vovk, Aggregating strategies, Proc. of Computational Learning Theory, 1990
(1990).
[41] Shmuel Zamir, Chapter 5: Repeated games of incomplete information: Zero-sum, Handbook
of Game Theory with Economic Applications (Robert Aumann and Sergiu Hart, eds.), vol. 1,
Elsevier, 1992, pp. 109–154.
A
A.1
Proof of all results
Proof of Lemma 3.1.
Consider the minimization problem:
min f (x) =
x∈S
K
X
k=1
30
xk .
Since f (x) is a continuous function defined on a compact set, it achieves this minimum value at
some point x∗ ∈ S. Hence there cannot be any point x0 ∈ S such that x0 ≺ x∗ , which means that
x∗ is on the Pareto frontier of S.
A.2
Proof of Proposition 3.1.
In order to prove the result, we need the following set of results about the Hausdorff distance:
Lemma A.1. a) h is a metric on the space of closed subsets of RK .
b) Assume that (An )n∈N is a Cauchy sequence of closed subsets of [0, 1]K . Then there is a unique
closed subset A of [0, 1]K such that h(An , A) → 0. This set A is defined as follows:
A = {x ∈ [0, 1]k : ∃xn ∈ An ∀ n s.t. xn → x}.
c) Assume that (An )n∈N is a sequence of closed subsets of [0, 1]K . Then there is a subsequence
(Ank )k∈N that converges to some closet subset A of [0, 1]K .
d) If the sets (An )n∈N in b) or c) are convex, then A is convex.
e) h(up(A), up(B)) ≤ h(A, B).
Proof. a)-c) This is the well-known completeness and compactness property of the Hausdorff metric;
see [24, 32].
d) Say that x, y ∈ A. Then x = limn xn and y = limn yn for xn ∈ An and yn ∈ An . By
convexity of each An , zn := λxn + (1 − λ)yn ∈ An . But then, zn → z := λx + (1 − λ)y. It follows
that z ∈ A, so that A is convex.
e) Let := h(A, B). Pick x ∈ up(A). Then x = y + v for some y ∈ A and v 0. There is some
y0 ∈ B with ky − y0 k∞ ≤ . Then x0 = min{y0 + v, 1} ∈ up(B), where 1 is the vector of ones in
RK , i.e., (1; k = 1, · · · , K), and the minimization is component-wise. We claim that kx0 − xk∞ ≤ .
If y0 + v ∈ [0, 1]K , this is clear. Assume yk0 + vk > 1. Then,
x0k = 1 < yk0 + vk and xk = yk + vk ≤ 1.
Thus,
0 ≤ x0k − xk < yk0 + vk − yk − vk = yk0 − yk .
Hence, |x0k − xk | ≤ |yk0 − yk | for any k. Thus, one has kx0 − xk∞ ≤ ky0 − yk∞ ≤ .
Now we can prove the proposition. First, to show that d is a metric on F, we just need to show
that if h(up(V), up(U)) = 0 then V = U. The other properties (e.g., triangle inequality etc.) follow
from the corresponding properties for the Hausdorff metric. Note that if h(up(V), up(U)) = 0, then
up(V) = up(U). Suppose that there is some u ∈ U such that u ∈
/ V. But since u ∈ up(V), we have
u = v + y for some v ∈ V and some y 0. But since h(up(V), up(U)) = 0, by the definition of the
Hausdorff distance, for each > 0, there is a point u in up(U) such that u v + 1. Consider a
sequence (n )n∈N such that n → 0, and consider the corresponding sequence (un )n∈N . Now since
up(U) is compact, (un )n∈N has a convergent subsequence that converges to some u∗ ∈ up(U) such
that u∗ v v + y = u, which contradicts the fact that u ∈ U. Thus U = V.
Next, we prove statement (b). Under the given assumptions, (up(Vn ))n∈N is Cauchy in the Hausdorff metric, so that, by Lemma A.1, there is a unique closed convex set such that h(up(Vn ), A) → 0.
31
But since h(up(Vn ), up(A)) ≤ h(up(Vn ), A) (from Lemma A.1), we have that
h(up(Vn ), up(A)) → 0 and hence up(A) = A. Thus the Pareto frontier V of A is then such that
d(Vn , V) → 0. To show uniqueness of V, assume that there is some U ∈ F such that d(Vn , U) → 0.
Then, the closed convex set up(U) is such that h(up(Vn ), up(U)) → 0. By Lemma A.1, this implies
that up(U) = up(V), so that U = V (from part (a) of the proposition) .
Finally, we prove statement (c). From statement (c) and (d) of Lemma A.1, the subsequence
(up(Vnk ))k∈N converges to some convex set A. But since h(up(Vnk ), up(A)) ≤ h(up(Vnk ), A), we
have up(A) = A. And thus the subsequence (Vnk )k∈N converges to the Pareto frontier of A, which
is in F.
Observe that it becomes clear from the above arguments that d induces these properties not just
on F, but also on the more general space of Pareto frontiers in [0, 1]K whose upset is closed.
A.3
Proof of Proposition 3.2.
Suppose that max(e(U, V), e(V, U)) ≤ . Consider a point x ∈ up(U) such that x = y + v where y ∈
U and v 0. Suppose that there is no x0 ∈ up(V) such that kx − x0 k∞ ≤ , i.e., for any x0 ∈ up(V),
kx − x0 k∞ > . This means that up(V) is a subset of the region {x0 : x0k > xk + for some k} (this
is the region S shown in the Figure 19). But since y = x − v, we have y x (y is in region S0 shown
in the Figure 19). But then for any w ∈ S0 , ky − wk∞ > . This contradicts the fact that for y
there is some y0 ∈ V, such that y +1 y0 . Thus d(U, V) ≤ . Now suppose that d(U, V) ≤ . Then
1
up(B)
S
B
𝝐
𝒙
𝒚
S0
S
A
1
0
Figure 19: Construction for the proof of Proposition 3.2.
for any x ∈ U, there is a x0 ∈ up(V) such that kx − x0 k∞ ≤ where x0 = y + v for y ∈ V and v 0.
Thus x + 1 x0 = y + v. The roles of U and V can be reversed. Thus max(e(U, V), e(V, U)) ≤ .
Observe that this proof uses the fact that the sup and inf in the definitions of h and e can be
replaced by max and min respectively, which is valid for the space F as discussed in footnotes 2
and 3.
A.4
Proof of Lemma 3.2:
In order to prove this lemma, we need a few intermediate results. We define the following notion
of convexity of Pareto frontiers.
32
Definition A.1. A Pareto frontier V is p-convex if for any v, u ∈ V and for each λ ∈ [0, 1], there
exists a point r ∈ V such that r λv + (1 − λ)u.
We then show the following equivalence.
Lemma A.2. For a Pareto frontier V ⊂ [0, 1]K , the following statements are equivalent:
1. V is in F.
2. V ∈ [0, 1]K is p-convex and up(V) is closed.
Proof. To show that 1 implies 2, we just need to show that V is p-convex. To see this, suppose that
u and v are two points in V. Since they also belong to up(V), which is convex, for each λ ∈ [0, 1],
λu + (1 − λ)v ∈ up(V) and thus there is some r ∈ V such that r λu + (1 − λ)v. Thus V is
p-convex.
To show that 2 implies 1, we just need to show that up(V) is convex if V is p-convex. To see
this, suppose that u + x and v + y are two points in up(V) where u, v ∈ V and x, y 0. By
p-convexity of V, for each λ ∈ [0, 1], there is a r ∈ V such that r λu + (1 − λ)v and thus
r λ(u + x) + (1 − λ)(v + y). Thus up(V) is convex.
We can now prove Lemma 3.2. Recall that,
X
αa rk (a, b) + βRk (a, b) ; k = 1, · · · , K
Ψ(V) =
max
b∈B
a∈A
: α ∈ ∆(A), R(a, b) ∈ V ∀ a ∈ A, b ∈ B .
First, note that Λ(Ψ(V)) = Λ(Ψ(up(V))). Now one can see that Ψ(up(V)) is the image of a
continuous function from the product space up(V)m×n ×∆(A) to a point in RK , which is a Hausdorff
space. Since up(V) is closed and bounded, it is compact. Also the simplex ∆(A) is compact. Thus
by Tychonoff’s theorem [35], the product space up(V)m×n × ∆(A) is compact. Hence by the closed
map lemma, f is a closed map and hence Ψ(up(V)) is closed. Hence up(Λ(Ψ(V))) is closed.
Next, recall that any point u in Λ(Ψ(V)) is of the form:
u=
max
X
b∈B
a∈A
αa rk (a, b) + βRk (a, b) ; k = 1, · · · , K
for some α ∈ ∆(A) and R(a, b) ∈ V. But since
P V is p-convex from Lemma A.2, for each b ∈ B,
there exists some Q(b) ∈ V such that Q(b) m
a=1 αa R(a, b). Hence statement 2 follows.
Now let
u=
and
max
b∈B
v=
max
b∈B
X
a∈A
X
a∈A
αa rk (a, b) + βQk (b) ; k = 1, · · · , K
ηa rk (a, b) + βRk (b) ; k = 1, · · · , K
33
be two points in Λ(Ψ(V)), where α, η ∈ ∆(A) and Q(b), R(b) ∈ V for all b ∈ B. For a fixed
λ ∈ [0, 1], let z = αλ + η(1 − λ). Then
λu + (1 − λ)v
X
X
= λ max
αa rk (a, b) + βQk (b) + (1 − λ) max
ηa rk (a, b) + βRk (b) ; k = 1, · · · , K
b∈B
max
b∈B
max
b∈B
b∈B
a∈A
X
a∈A
X
a∈A
a∈A
za rk (a, b) + β[λQk (b) + (1 − λ)Rk (b)] ; k = 1, · · · , K
za rk (a, b) + βLk (b) ; k = 1, · · · , K .
The first inequality holds since max is a convex function and the second follows since V is p-convex,
and hence L(b) ∈ V that satisfy the given relation exist. Thus Λ(Ψ(V)) is p-convex. Combined
with the fact that up(Λ(Ψ(V))) is closed, this implies that Λ(Ψ(V)) ∈ F using Lemma A.2.
A.5
Proof of Lemma 3.3.
Suppose e(U, V) = . Let
max
X
b∈B
a∈A
αa rk (a, b) + βRk (b) ; k = 1, · · · , K
be some point in Φ(V), where α ∈ ∆(A). Then for each b, we can choose Q(b) ∈ U such that
Q(b) R(b) + 1. We then have
max
X
b∈B
a∈A
X
αa rk (a, b) + βRk (b) + β(Qk (b) − Rk (b))
αa rk (a, b) + βQk (b)
= max
b∈B
≤ max
b∈B
= max
b∈B
a∈A
X
αa rk (a, b) + βRk (b) + β
a∈A
X
αa rk (a, b) + βRk (b) + β.
a∈A
Thus
max
b∈B
X
a∈A
But since
max
b∈B
max
αa rk (a, b) + βQk (b) ; k = 1, · · · , K
X
b∈B
X
a∈A
a∈A
αa rk (a, b) + βRk (b) ; k = 1, · · · , K + β1.
αa rk (a, b) + βQk (b) ; k = 1, · · · , K ∈ Ψ(U),
and since Φ(U) = Λ(Ψ(U)), there exists some L ∈ Φ(U) such that
34
L
Thus
max
b∈B
L
X
max
b∈B
a∈A
X
a∈A
αa rk (a, b) + βQk (b) ; k = 1, · · · , K .
αa rk (a, b) + βRk (b) ; k = 1, · · · , K
+ β1.
Thus
e(Φ(U), Φ(V)) ≤ β = βe(U, V).
A.6
(29)
Proof of Theorem 3.1.
Since Φ is a contraction in the metric d, the sequence {An } is Cauchy in F. Hence by Lemma A.1,
{An } converges to a Pareto frontier V∗ ∈ F. The continuity of the operator further implies that
V∗ = Φ(V∗ ). To show uniqueness, observe that if there are two fixed points U and V, then we have
d(U, V) = d(Φ(U, Φ(V)) ≤ βd(U, V), which implies that d(U, V) = 0 and hence U = V.
A.7
Proof of Theorem 3.2.
In G∞ , fix T ≥ 1 and consider a truncated game where Alice can guarantee the cumulative losses
in β T +1 V∗ after time T + 1. Then the minimal losses that she can guarantee after time T is the
set:
X
T +1
∗
T
αa rk (a, b) + β
Qk (b); k = 1, · · · , K | α ∈ ∆(A), Q(b) ∈ V ∀ b ∈ B
.
Λ
max β
b∈B
a∈A
This set is β T V∗ . By induction, this implies that the set of minimal losses that she can guarantee
after time 0 is V∗ .
The losses of the truncated game and of the original game differ only after time T + 1. Since
the losses at each step are bounded by (1 − β), the cumulative losses after time T + 1 are bounded
by
β T +1 (1−β)
1−β
= β T +1 . Consequently, the minimal losses of the original game must be in the set
u ∈ [0, 1]K : uk ∈ [xk − β T +1 , xk + β T +1 ] for all k, x ∈ V∗ .
Since T ≥ 1 is arbitrary, the minimal losses that Alice can guarantee in the original game must be
in V∗ .
A.8
Proof of Theorem 3.3.
Assume that Alice can guarantee every pair β T +1 u of cumulative losses with u ∈ V∗ after time
T +1 by choosing some continuation strategy in ΠA . Let x = F(p, V∗ ). We claim that after time T ,
Alice can guarantee a loss of no more than β T x on each component by first choosing aT = a with
35
probability αa (p) and then if Bob chooses b ∈ B, choosing a continuation strategy that guarantees
her F(p0 , V∗ ), where p0 = q(b, p). Indeed by following this strategy, her expected loss on component
k after time T is then
X
αa (p)rk (a, b) + β T +1 Fk (q(b, p), V∗ )} ≤ β T Fk (p, V∗ ) = β T xk .
{β T
a
Thus, this strategy for Alice guarantees that her loss after time T is no more than β T V∗ . Hence by
induction, following the indicated strategy (in the statement of the theorem) for the first T steps
and then using the continuation strategy from time T + 1 onwards, guarantees that her loss is not
more than F(p1 , V∗ ) after time 0. Now, even if Alice plays arbitrarily after time T +1 after following
the indicated strategy for the first T steps, she still guarantees that her loss is (componentwise) no
more than F(p1 , V∗ ) + β T +1 (1; k = 1, · · · , K)T . Since this is true for arbitrarily large values of T ,
playing the given strategy indefinitely guarantees that her loss is no more than F(p1 , V∗ ).
A.9
Proof of Proposition 4.1.
PM
P
Any point e in ΓN (V) is of the form M
k=1 λk = 1, and M ≤ K.
k=1 λk vk where vk ∈ up(V) and
0
But then by the definition of an upset, we have vk ∈ V for each k such that vk0 vk and hence
PM
PM
PM
0
0
k=1 λk vk ,
k=1 λk vk . By the p-convexity of V, there is some r ∈ V, such that r
k=1 λk vk
and hence r e. Thus e(ΓN (V), V) = 0.
Next, we will show that for any u ∈ V, there exists e ∈ ΓN (V) such that e u + (1/N )1.
For the rest of the proof, all the distances refer to distances in the L∞ norm. Consider a line
x = t1 +p, and suppose that the shortest distance between u and any point on this line is a > 0,
i.e., min ||u − x||∞ : x = t1 + p = a. Let x∗ = t∗ 1 + p be the point on the line that is closest
to u. If a+ , max{(x∗k − uk )+ : k = 1, · · · , K} and a− , max{−(x∗k − uk )− : k = 1, · · · , K}, then
a = max{a+ , a− }. Consider any point v that is the smallest point of intersection of x = t1 + p
and the set up(V). Then this point must lie in the set {t1 + p : t ∈ [t∗ − a+ , t∗ + a− ]}, because a)
if v = t0 1 + p for some t0 < t∗ − a+ , then it means that u dominates v which contradicts the fact
that v ∈ up(V), and b) if v = t0 1 + p for some t0 > t∗ + a− then v will strictly dominate u on each
dimension, but then the point v0 = (t∗ + a− )1 + p is strictly smaller than v and lies in up(V) and
on the line v = t0 1 + p, which contradicts the definition of v. Thus ||u − v||∞ ≤ a+ + a− ≤ 2a.
Thus we have shown that if the shortest distance between u and some line x = t1 + p is a, then
the distance between u and the smallest point of intersection of x = t1 + p and the set up(V) is no
more than 2a.
Now we will show that the for any u ∈ V, there is always a line x = t1 + p such that the
shortest distance between u and the line is no more than 1/(2N ). Let umin = mink {uk }. Then
u = umin 1 + (u − umin 1). Now the vector (u − umin 1) has value 0 on one dimension, and on
every other dimension it has value in [0, 1] (since u ∈ [0, 1]K ), and so it can be approximated by
some pk ∈ {0, 1/N, · · · , (N − 1)/N, 1} where the approximation error on any dimension is at most
1/(2N ). Thus there is a point e0 = umin 1 + p where p ∈ PN such that ku − e0 k ≤ 1/(2N ). Thus
there us always a line x = t1 + p such that the shortest distance between u and the line is no more
than 1/2N .
Together, we finally have that for any u ∈ V there is some point e00 , which is the smallest point
of intersection of some line x = t1 + p and the set up(V), such that ku − e00 k∞ ≤ 2 × (1/2N ) = 1/N ,
and thus e00 u + (1/N )1. Since there is always some point e ∈ ΓN (V) such that e e00 (recall
36
the definition (11) of ΓN (V) as the Pareto frontier of the set ch
e u + (1/N )1. Thus e(V, ΓN (V)) ≤ 1/N .
A.10
F(p, V) : p ∈ PN
), we have
Proof of Proposition 4.2.
We have Gn = ΓN ◦ Φ(Gn−1 )). Consider another sequence of Pareto frontiers
n
An = Φ (G0 )
.
(30)
n∈N
Then we have
d(An , Gn )
=
d(Φ(An−1 ), ΓN (Φ(Gn−1 )))
(a)
≤
d(Φ(An−1 ), Φ(Gn−1 )) + d(Φ(Gn−1 ), ΓN (Φ(Gn−1 )))
≤
βd(An−1 , Gn−1 ) +
(b)
1
,
N
(31)
where inequality (a) is the triangle inequality and (b) follows from (29) and Lemma 4.1. Coupled
with the fact that d(A0 , G0 ) = 0, we have that
1
2
n−1
d(An , Gn ) ≤
1 + β + β + ···β
N
1 1 − βn
=
.
(32)
N 1−β
Since Φ is a contraction, the sequence {An } converges to some Pareto frontier V∗ . Suppose that
we stop the generation of the sequences {An } and {Gn } at some n. Now since A0 = G0 = {0}, and
since the stage losses rk (a, b) ∈ [0, 1 − β], we have that d(A1 , A0 ) ≤ 1 − β. From the contraction
property, this implies that d(An+1 , An ) ≤ β n (1 − β). Thus d(V∗ , An ) ≤
triangle inequality we have
1 1 − βn
∗
d(V , Gn ) ≤
+ βn.
N 1−β
β n (1−β)
1−β
= β n , and thus by
(33)
Finally, to show that e(Gn , V∗ ) ≤ β n , observe that
e(Gn , An )
=
e(ΓN (Φ(Gn−1 )), Φ(An−1 ))
(a)
≤
e(ΓN (Φ(Gn−1 )), Φ(Gn−1 )) + e(Φ(Gn−1 ), Φ(An−1 ))
≤
0 + βe(Gn−1 , An−1 ).
(b)
(34)
Since A0 = G0 = {0}, this implies that e(Gn , An ) = 0 for all n. Here, (a) holds since if for three
frontiers U, V and L, U 1 -dominates V and V 2 -dominates Z, then U (1 + 2 )-dominates Z. (b)
follows from the contraction property of Φ under e. Further, e(An , V∗ ) ≤ d(An , V∗ ) ≤ β n from
above. Thus we have e(Gn , V∗ ) ≤ e(Gn , An ) + e(An , V∗ ) ≤ β n .
37
A.11
Proof of Proposition 4.3.
In order to prove this result, we need a few intermediate definitions and results. First, we need to
characterize the losses guaranteed by any H(K, N )−mode stationary strategy. Such a strategy π
defines the following operator on any function F : PN → RK (PN is defined in (9)):
∆πN (F)(p)
=
max
b∈B
X
αa (p)rk (a, b) +
K
X
k0 =1
a∈A
zk0 (b, p)βFk (qk0 (b, p)) ; k = 1, · · · , K ,
(35)
where qk0 (b, p) ∈ PN for all k 0 . Now for a function F : PN → RK , define the following norm:
kFk = max kF(p)k∞ .
p∈PN
It is easy to show that ∆πN is a contraction in the norm. We omit the proof for the sake of brevity.
Lemma A.3.
k∆πN (F) − ∆πN (G)k ≤ βkF − Gk.
(36)
We can then show the following result.
Lemma A.4. Consider a H(K, N )-mode strategy π. Then there is a unique function
Fπ : PN → RK
such that ∆πN (Fπ ) = Fπ . Further, The strategy π initiated at mode p where p ∈ PN guarantees the
vector of losses Fπ (p).
The first part of the result follows from the fact that the operator is a contraction and the
completeness of the space of vector-valued functions with a finite domain for the given norm. The
second part follows from arguments similar to those in the proof of Theorem 3.3. The arguments
are not repeated here for the sake of brevity. Now let
πn
πn
V , Λ ch({F (p) : p ∈ PN ) ,
where Fπn is the fixed point of the operator ∆πNn .
Define a sequence of functions Fn : PN → RK where Fn (p) = F(p, Φ(Gn−1 )) = F(p, Gn ). We
then have that
d(Vπn , V∗ ) ≤ d(Vπn , Gn ) + d(Gn , V∗ )
1 1 − βn
πn
≤ d(V , Gn ) +
+ βn.
N 1−β
(37)
(38)
From the definition of d, it is clear that d(Vπn , Gn ) ≤ kFπn − Fn k. Next we have
kFπn − Fn k
≤
kFπn − ∆πNn (Fn )k + k∆πNn (Fn ) − Fn k
(39)
=
k∆πNn (Fπn ) − ∆πNn (Fn )k + kFn+1 − Fn k
(40)
βkFπn − Fn k + kFn+1 − Fn k.
(41)
(a)
(b)
≤
38
(42)
Here (a) holds because ∆πNn (Fn ) = Fn+1 by the definition of the strategy πn , and because Fπn is a
fixed point of the operator ∆πNn . (b) holds because ∆πNn is a contraction. Thus we have
d(Vπn , Gn ) ≤ kFπn − Fn k ≤
kFn+1 − Fn k
.
1−β
(43)
And finally we have:
1
d(V , V ) ≤
N
πn
∗
1 − βn
1−β
+ βn +
kFn+1 − Fn k
.
1−β
(44)
To finish up, we need the following result:
Lemma A.5.
kFn+1 − Fn k ≤ d(Gn+1 , Gn ).
Proof. Let u = Fn+1 (p) and v = Fn (p) for some p. Now u is the point of intersection of Gn+1
and the line x = t1 + p. v is the point of intersection of the frontier Gn and the line x = t1 + p,.
Now suppose that ku − vk∞ > d(Gn+1 , Gn ). Then either for u, there is no r ∈ Gn such that
r u + 1d(Gn+1 , Gn ) or for v, there is no r ∈ Gn+1 such that r v + 1d(Gn+1 , Gn ). Either of the
two cases contradict the definition of d(Gn+1 , Gn ). Thus ku − vk∞ ≤ d(Gn+1 , Gn ).
Finally, by the triangle inequality we have
d(Gn+1 , Gn ) ≤ d(An+1 , An ) + d(Gn+1 , An+1 ) + d(Gn , An )
1 1 − βn
1 1 − β n+1
n
+
.
≤ (1 − β)β +
N
1−β
N 1−β
(45)
(46)
Combining with (44) we have the result.
B
Some remarks on Pareto frontiers of closed and convex sets.
The Pareto frontier of a closed set is not necessarily closed. Figure 1 is one example – the upset is
closed, but its Pareto frontier is open. But we can show that the Pareto frontier of a closed and
convex set in R2 is closed.14
Proposition B.1. Let V be the lower Pareto frontier of a closed and convex set S in R2 . Then V
is closed.
Proof. Suppose that {vn } is a sequence of points in V that converge to some point v. Then since
S is closed, v ∈ S. We will show that v ∈ V. Suppose not. Then there is some u ∈ V such that
u v. Suppose first that u1 < v1 and u2 < v2 . Then let = (min(v1 − u1 , v2 − u2 )/2 and consider
the L2 ball of radius around v, i.e.
Bv () = {y ∈ R2 : ky − vk2 ≤ }.
14
Of course, the Pareto frontier of a closed set may be empty, e.g., {(x, y) ∈ R2 : x = y}, in which case it is trivially
closed.
39
Then for any point y in Bv (), we have that u y. But since {vn } converges to v, there exists some
point in the sequence that is in Bv (), and u is dominated by this point, which is a contradiction.
Hence either u1 = v1 or u2 = v2 . Suppose w.l.o.g. that u1 < v1 and u2 = v2 . See Figure 20. Let
δ = (v1 − u1 )/2 and consider the ball of radius δ centered at v, i.e. Bv (δ). Let vn be a point in the
sequence such that vn ∈ Bv (δ). Now vn,1 > u1 and hence it must be that vn,2 < u2 . Now for some
λ ∈ (0, 1), consider a point r = λu + (1 − λ)vn such that r1 = u1 + δ. It is possible to pick such a
point since a) v1 = u1 + 2δ and b) |vn,1 − v1 | ≤ δ, which together imply that vn,1 ≥ u1 + δ (please
see the figure). Now r ∈ S since S is convex. Next r1 = v1 − δ < v1 , and also r2 < u2 = v2 since
𝜹
𝜹
𝒖
𝒗
𝜹%
𝒓
𝒗𝒏
Figure 20: Construction in the proof of Proposition B.1.
λ > 0 and vn,2 < u2 . Let δ 0 = v2 − r2 . Then consider the ball Bv (δ 0 ) centered at v. Clearly r y
for any y ∈ Bv (δ 0 ). But since {vn } converges to v, there exists some point in the sequence that is
in Bv (δ 0 ), and r is dominated by this point, which is again a contradiction. Thus v ∈ V.
Interestingly, this result doesn’t hold for closed and convex sets in RK for K > 2, as we show
in the example depicted in Figure 21. The closed and convex set S is a solid 3-dimensional cone
with apex (0, 0, 1) and base being the semicircular disc defined by the set {(x, y, z) ∈ R3 : x =
5, (y − 1)2 + (z − 1)2 ≤ 1, z ≤ 1}. The sequence (vn )n∈N lies on the Pareto frontier of this set as
shown in the figure, but it converges to the point (5, 0, 1), which dominates the point (0, 0, 1).
40
𝒛
(𝟎, 𝟎, 𝟏)
S
𝒚
𝒗∗ = (𝟓, 𝟎, 𝟏)
𝒗𝒏
𝒙
Figure 21: An example of a compact and convex set in R3 whose Pareto frontier is not closed. The
sequence (vn ) on the Pareto frontier converges to the point (5, 0, 1), which dominates (0, 0, 1).
C
Solving Linear Program (12).
When V is the lower Pareto frontier of a convex polytope, (12) is a linear program. In this program,
x is a dependent vector and can be eliminated. The only question remains is that of addressing
the constraint Q(b) ∈ V. The vertices of V are a subset of {F(p, Φ(V)) : p ∈ PN }. Thus Q(b)
can be chosen as a convex combination of points in {F(p, Φ(V)) : p ∈ PN }. This introduces
H(K, N ) variables for each b ∈ B along with the constraint that these variables sum to 1, thus
contributing mH(K, N ) variables and m constraints. Along with the variables α and t, this makes
mH(K, N ) + l + 1 variables inPtotal. And along with the K domination constraints for each
b ∈ B and the constraint that a αa = 1, this makes Km + m + 1 constraints in total (ignoring
non-negativity constraints on all variables except t).
Of the H(K, N ) variables associated with each b ∈ B that determine the point Q(b), we know
that at most K will be non-zero. This sparsity constraint can potentially be utilized to speed up
the computation, although we didn’t attempt to do so in our computations.
D
GPS algorithms.
The GPS algorithms are defined as follows [22]. Let ξ = (1 −
p
1 − β 2 )/β.
1. K=2: Let d be the difference in the cumulative (undiscounted) losses of the leading expert
(the one with the lower cumulative loss) and the lagging expert (the one with the higher loss).
Then at every stage, the algorithm chooses the leading expert with probability 1 − (1/2)ξ d
and the lagging expert with probability (1/2)ξ d .
41
2. K=3: For i ≤ j, let dij be the difference in the cumulative losses of the expert with the
ith lowest loss and the expert with the j th lowest loss. Then the algorithm chooses the
leading expert with probability 1 − ξ d12 /2 − ξ d13 +d23 /6, the second expert with probability
ξ d12 /2 − ξ d13 +d23 /6 and the lagging expert with probability ξ d13 +d23 /3.
E
A Blackwell-type minmax theorem doesn’t hold for finite games.
Here we give an example showing that a Blackwell-type minmax result doesn’t hold for expected
losses in finite games, even if the losses are scalar. Consider the matrix game shown in Figure 22.
And consider the convex set consisting of a single point 0.4 on the real line. Then observe that for
any mixed strategy (α, 1 − α) of player 1, the expected loss if player 2 chooses action 1 is (1 − α)
and if she chooses action 2, it is 0.25α + 0.5(1 − α) = 0.5 − .25α. Thus no strategy of player 1 can
guarantee that the expected loss is exactly 0.4 irrespective of the actions of player 2. But neither
does player 2 have any strategy that guarantees that the expected loss in the game is not 0.4. To
see this, observe that for any strategy (γ, 1 − γ) of player 2, player 1 can choose a randomization
over her two actions that will result in an expected loss of exactly 0.4.
𝟏
𝟐
𝟏
𝟎
𝟎. 𝟐𝟓
𝟐
𝟏
𝟎. 𝟓
Figure 22: A game with scalar losses.
42
| 8cs.DS
|
1
Energy Management for Energy Harvesting
Wireless Sensors with Adaptive Retransmission
Animesh Yadav, Member, IEEE, Mathew Goonewardena, Student Member, IEEE, Wessam
arXiv:1710.09034v1 [cs.IT] 25 Oct 2017
Ajib, Senior Member, IEEE, Octavia A. Dobre, Senior Member, IEEE, and Halima
Elbiaze, Member, IEEE
Abstract
This paper analyzes the communication between two energy harvesting wireless sensor nodes. The
nodes use automatic repeat request and forward error correction mechanism for the error control. The
random nature of available energy and arrivals of harvested energy may induces interruption to the signal
sampling and decoding operations. We propose a selective sampling scheme where the length of the
transmitted packet to be sampled depends on the available energy at the receiver. The receiver performs
the decoding when complete samples of the packet are available. The selective sampling information
bits are piggybacked on the automatic repeat request messages for the transmitter use. This way, the
receiver node manages more efficiently its energy use. Besides, we present the partially observable
Markov decision process formulation, which minimizes the long-term average pairwise error probability
and optimizes the transmit power. Optimal and suboptimal power assignment strategies are introduced
for retransmissions, which are adapted to the selective sampling and channel state information. With
finite battery size and fixed power assignment policy, an analytical expression for the average packet
drop probability is derived. Numerical simulations show the performance gain of the proposed scheme
with power assignment strategy over the conventional scheme.
A part of the paper is published in the Proceeding of the IEEE International Conference on Communications (IEEE ICC
2015), London, UK, 8-12 June 2015.
A. Yadav and O. A. Dobre are with the Faculty of Engineering and Applied Science, Memorial University, St. John’s, NL,
Canada, (e-mail: {animeshy, odobre}@mun.ca), W. Ajib and H. Elbiaze are with the Department of Computer Science, Université
du Québec à Montréal (UQAM), Montreal, QC, Canada (e-mail: {ajib.wessam, elbiaze.halima}@uqam.ca) and M. Goonewardena
is École de Technologie Supérieure (ÉTS), Montréal, QC, Canada. (e-mail:[email protected]).
2
Index Terms
Wireless sensors networks, energy harvesting, packet drop probability, partially observable Markov
decision processes.
I. I NTRODUCTION
The use of energy harvesting (EH) sources to power wireless communication systems has recently
received considerable attention [1]–[7]. The EH devices offer green communication and can operate
autonomously over long periods of time. Because of these benefits, the EH devices are also increasingly
considered in wireless sensor networks (WSNs) to power the sensor nodes [8]–[15].Sensors nodes are
low cost distributed devices which operate on minimal energy. They are very prevalent in applications
related to monitoring and controlling the environments, especially the remote and dangerous ones [16].
Usually, sensor nodes are operated by small capacity non-renewable batteries, thus, suffering from finite
lifespan of operation. Sensor nodes with EH capabilities can be an alternative to increase the lifespan
and lower the maintenance cost. Energy can be harvested from the environment using for instance solar,
vibration or thermoelectric effects. Unlike EH, another practical alternative to increase the lifespan of the
nodes is to use massive antenna arrays at the receiver nodes to mitigates severe energy constraints given
by the inexpensive transmitter nodes [17], [18].
Typically, the energy arrival amount at the EH devices is random. Thus, for such nodes, the challenging
objective is the adequate management of the collected energy to enable reliable and continuous operation.
Recently, a considerable amount of works on wireless networks solely powered by harvested energy have
emanated [4]–[11] to address this objective. Although both transmitter and receiver nodes can harvest
energy, the research is primarily focused either on the transmitter [8]–[11] or receiver [5]–[7], [19]. There
are many practicals scenarios where the transmitter and receiver nodes can harvest energy to increase their
lifespan, such as the scenario of transmitter and multiple intermediate nodes in a multi-hop WSN, and
multiple transmitter nodes communicating with a single sink node. These scenarios are more challenging
due to the presence of many random sources of energy.
3
Fewer works have considered the EH capability at both transmitter and receiver nodes simultaneously
[4], [12]–[15]. In [4], the authors considered a static additive white Gaussian noise (AWGN) channel and
used a rate-based utility as a function of both transmitter and receiver powers. They proposed directional
water-filling based power allocation policy in an offline setting. The problem of online power control for
a wireless link with automatic repeat request (ARQ) scheme is studied in [12]. The authors investigated
three fixed policies under various assumptions, such as knowledge of the receiver battery availability
at the transmitter node, finite and infinite battery storage. In [13], the authors analyzed a wireless link
which employs type-II hybrid automatic repeat request (HARQ) scheme. They derived the packet drop
probability (PDP) for predetermined transmit energy levels. In [14], the authors obtained a lower bound
on maximum achievable throughput and proposed a common threshold policy. In [15], we introduced an
energy-aware adaptive retransmission scheme, where the receiver node performs the selective sampling
(SS) and decoding operations based on the energy availability.
The receiver node spends the energy dominantly in sampling and decoding operations, if a forward error
correction (FEC) coding is employed [20]. Moreover, for small distances, the transmit energy is often
smaller than the energy needed in the decoding operation [21]. Nonetheless, because of the randomness in
the amount of energy arrivals, the receiver operations may be suspended, which leads to energy wastage.
Thus, the receiver might favor to sample a fraction of the full packet depending on the available energy
[5], which we refer to as SS. On the other hand, the transmitter with exact SS information (SSI) can
retransmit only a portion of the packet, which is not sampled by the receiver.
Furthermore, the time-varying characteristic of the wireless channel and harvested energy might contribute to a higher packet error probability (PEP). Hence, the transmitter must adequately adapt the transmit
power level to the channel state information (CSI), while meeting the constraint of energy causality, to
ensure a lower PEP. The causality constraint affirms that the cumulative used energy cannot surpass the
cumulative harvested energy by nodes at any given time. Furthermore, based on the SSI knowledge, the
transmitter adapts the packet size to ensure an efficient utilization of the receiver energy. In pursuance of
providing the SSI to the transmitter, we resort to the ARQ protocol’s acknowledgement (ACK)/negative
acknowledgement (NAK) feedback messages. Consequently, the retransmission scheme, which we denote
4
by ACK/NAKx, needs to have some additional feedback messages.
In this paper, we consider a generic communication between two EH wireless nodes with the aforementioned retransmission protocol. A decision-theoretic approach is used to find the optimal transmit power
strategy. Firstly, the problem is formulated as a partially observable Markov decision process (POMDP),
which is a suitable approach for formulating problems that require sequential decision making in a
stochastic setting, when some of the system states are unknown [22]. We solve the POMDP problem
using the value iteration method by computing the value function for the belief of the unknown state.
Since the memory and computational complexity requirements are limited for sensor nodes, we propose
a suboptimal and a computationally lower greedy power assignment method.
The outline of this paper is as follows. The system model is presented in Section II. The adaptive
retransmission scheme is detailed in Section III. Section IV introduces the optimal and suboptimal methods
aiming to allocate the power over the time slots. An analytical upper bound on the PDP is derived
in Section V. Simulated numerical results and discussions are presented in Section VI, followed by
conclusions in Section VII. A list of symbols with their descriptions used in this paper is given in Table I
ACK
NAK
NAKx
ACK
NAKx
ACK
h
ETx
h
ETx
X
h
ETx
h
ETx
h
ETx
ETx
k=1
Ts
t−5
ETx
ETx
k=2
t−4
ETx
k=3
k=1
k=2
time index
Tf
t−3
t−2
t−1
t
t+1
Fig. 1: Time-slotted packet transmission time line at the transmitter node. ’99K’ and ’−→’ denote the
ARQ message and EH arrival events, respectively. Shaded areas in the slot denote the amount of energy
used for transmission.
5
II. S YSTEM M ODEL
A. Transmission Model
We consider a point-to-point communication between two EH wireless sensor nodes. Sensor nodes have
limited capacity rechargeable batteries, which are charged by renewable energy sources. In the considered
model, when a transmitted packet is erroneously decoded, the receiver requests its retransmission. A
maximum of K ∈ Z retransmission requests are permitted. A packet consists of c information bits, taken
from the data buffer, encoded with an (m, c) FEC code (e.g., convolutional code), and then modulated
through an M -ary quadrature amplitude modulation, where M denotes the cardinality of the constellation.
This forms the packet of length ⌈m/ log2 M ⌉ symbols, where ⌈·⌉ is the ceiling operator.
The CSI and SSI are known at the transmitter through the ARQ feedback messages. Each sensor node
is aware of its own battery state information (BSI), but not of the BSI of the other node. However, the
transmitter estimates the one slot delayed BSI of the receiver via SSI.
A discrete time-slotted model is considered as depicted in Fig. 1. Each time slot is of Ts seconds
duration and indexed as t∈{1, 2,. . .}. The packet transmission and corresponding ARQ message reception
are completed within a slot, i.e., the round-trip time is Ts . Several slots constitute a frame of duration
Tf . A frame has variable duration depending upon the number of slots being used in packet transmission,
including retransmissions. Thus, the minimum and maximum values of frame duration are Ts and KTs ,
respectively. After a maximum of K unsuccessful attempts, the transmitter drops the packet and chooses
a new one to transmit.
Without loss of generality, the system model considered here can be extended to generic short-range
communication systems involving different modulation formats, sophisticated channel coding methods,
and transmission strategies relying on multiple antennas and sub-carriers techniques.
B. Energy Consumption Model
The transmitter and receiver sensor nodes spend energy to transmit and retrieve the information bits,
respectively. For short-range commmunication, the energy consumption in a wireless link can be broken
down into two dominant factors [20]: the energy consumed at the power amplifiers PPA at the transmitter,
6
and the circuit blocks at both transmitter and receiver. The circuit blocks of the transmitter consist of
a digital-to-analog convertor, mixers, active filters, and frequency synthesizers, while mainly of a low
noise amplifier, intermediate frequency amplifier, active filters, analog-to-digital convertor, and frequency
synthesizer at the receiver. Further, for coded systems, the energy expended in the decoding operation
needs to be included [20], [21] at the receiver. Thus, for a coded system, the total energy expenditure at
time slot t, at both transmitter and receiver nodes is, respectively, given as
PTx = (1 + α)Pout +PC,Tx ,
{z
}
|
(1)
PPA
PRx = Pdec + PC,Rx + Pfb ,
(2)
where Pout is the transmit power, α = (ξ/η) − 1, with η as the drain efficiency and ξ as the peak-toaverage power ratio. PC,Tx and PC,Rx are the total power spent in the circuit blocks of the transmitter and
receiver, respectively. The power consumed in transmitting the ARQ messages is denoted by Pfb . Pdec
denotes the power used in the decoding operation and is ignored for the uncoded system. Typical values
of Pdec are around 70-80% of the power dissipated in the circuit blocks [21]. The index t is dropped in
(1) and (2) to simplify the presentation.
C. Energy Harvesting Model
The transmitter and receiver nodes are connected to two separate but similar renewable EH sources. In
particular, two independent and identically distributed (i.i.d.), Bernoulli random processes are considered
to model the energy arrivals, similar to [23]. The Bernoulli model is tractable and captures the intermittent and irregular behavior of the energy arrival. It is worth mentioning that this work is, in essence,
independent of the energy arrival process; this will be shown later in the simulation results, where the
h Joule (J) with
compound Poisson arrival model [24] is used as well. At the start of every slot, ETx
probability (w.p.) ρTx and zero J w.p. 1 − ρTx is harvested at the transmitter. The receiver node follows
h .
a similar energy arrival process with probability ρRx and amount ERx
When the two nodes are in close vicinity and have the same type of harvesting source, then the two
EH processes are spatially correlated. In this case, the harvested energy pairs at time slot t are given as
7
[12]
t
t
(ETx
, ERx
)=
(0, 0)
h )
(0, ERx
w.p. p00 ,
w.p. p01 ,
(3)
h , 0)
(ETx
w.p. p10 ,
(E h , E h ) w.p. p11 ,
Tx
Rx
with the condition that p00 +p01 +p10 +p11 = 1. For example, when p00 = p01 = p00 = p10 = p11 = 0.25
then both nodes harvest energies independent from each other. When p01 = p10 = 0 and p00 = p11 = 0.5,
the harvested energies are highly correlated.
t and B t denote the energy levels of the battery at the start of the tth time slot, and E t =
Let BTx
Rx
Tx
t
PTx Ts and ERx
= PRx Ts denote the energy consumed in transmitting a packet, as well as sampling
and decoding operations at the transmitter and receiver, respectively. The battery level at the transmitter
follows the Markovian evolution:
t+1
=
BTx
t + E h − E t , B max },
min{BTx
Tx
Tx
Tx
t − Et ,
BTx
Tx
w.p. ρTx
(4)
w.p. 1 − ρTx ,
max denotes the transmitter node’s battery capacity. Replacing the subscript Tx in (4) with Rx
where BTx
gives the receiver side battery evolution. For presentation simplicity, the energies are normalized by a
min and E min that are spent in transmitting and receiving a packet of
minimum possible energy, i.e., ETx
Rx
min and
smallest size, respectively. Consequently, the transmitter energy level is an integer multiple of ETx
h /E min . Similarly, the battery
the change in the battery state whenever harvesting takes place is LTx , ETx
Tx
min and the EH amount is L
h
min
energy level at the receiver side is an integer multiple of ERx
Rx , ERx /ERx .
D. Channel Model
The wireless channel from the transmitter to the receiver is assumed to be Rayleigh faded and modeled
as a finite state Markov chain (FSMC) [25], [26]. This model captures the main features of fading channels,
and approximates the fading as a discrete-time Markov process. Essentially, all possible fading gains are
8
modelled as a set of finite and discrete channel states. The FSMC channel is described as follows: discrete
states of the channel G = {g1, g2 , . . . , g|G| }, state transition probabilities Ω = {p(gj |gi ) : g1 < gi, gj <
g|G| }, and steady state probabilities represented by ωo (gi ), i = 1, 2, . . . , |G|.
Based on the FSMC channel model, the entire range of channel gains are partitioned into |G| + 1
|G|
non-overlapping intervals with boundary values denoted as {γi }i=0 , with increasing order of their values
from γ0 = 0 to γ|G| = ∞. The fading gain interval [γi−1 , γi ) represents the gi channel state, which is
considered fixed during the time slot t and changes to gj in time slot t + 1 with probability p(gj |gi ).
The channel state is considered fixed during time slot t and changes to another in time slot t + 1 with
probability p(gj |gi ) .
III. A DAPTIVE R ETRANSMISSION S CHEME
In low harvesting rate, the receiver might not perform the sampling and decoding operations together
or perform only the sampling operation in one time slot. In such time slots, the receiver performs SS
and stores the samples, which can be later combined with the remaining parts of the packet for building
a full packet. Besides, the receiver sends back the SSI via ARQ messages. The transmitter then adapts
the packet length for the next transmission. In the adaptive retransmission scheme, ARQ messages carry
1 + log2 β bits, where β ∈ {1, 2, 4, 8, . . . , ⌈m/ log2 M ⌉}, and have a total of x + 2 messages where
{0, 1, . . . , β}, β > 1
x∈
(5)
∅,
otherwise.
For β = 1, the scheme becomes the conventional one. The additional x messages are essentially the
SSI which are carried back to the transmitter as ARQ messages; henceforth, we refer to the adaptive
retransmission scheme by ACK/NAKx. Details of each message are as follows:
ACK: The packet decoding at the receiver is successful. In reply, the transmitter chooses a new packet
for transmission in the subsequent time slot.
NAK: The packet decoding is erroneous. In reply, the transmitter chooses the same packet for
transmission in the subsequent time slot.
9
TABLE I: List of symbols
Symbol
K and k
Description
Maximum number of retransmissions
c and m
and its index
Number of information and coded bits
M and Rc
Modulation order and code rate
Symbol
S, |S| and St
BTx , BRx , |BTx |, and |BRx |
USt , |USt |, at , BSt
Description
System state space, its cardinality, and
state at time slot t
Transmitter and receiver battery states
space and their cardinality
Action states space, its cardinality, action, and maximum value of an action at
Ts and Tf
Slot duration and frame duration
Z, |Z| and Zt
time slot t
Observation state space, its cardinality,
and observation state at time slot t
ARQ feedback messages
t
t
ETx
and ERx
Transmitter and receiver energy expen-
ACK, NAK, NAKx
Pdec and Pfb
diture at time slot t
Decoding and ARQ message transmit
Pe (g, at )
Packet error probability with energy
Pout and PPA
power expenditure at receiver
Transmit power and power amplifier out-
P2 (d, g, at )
min
aETx
via channel state g
Modulation dependent bit error probabil-
PC,Tx and PC,Rx
put power
Transmitter and receiver circuit block
Perr
ity
Approximate packet error probability af-
Ad and dfree
ter NAK or decoding failure
Weight spectral coefficient and free dis-
r(s, at )
tance of convolution code
Cost function of state St = s after taking
ρTx and ρRx
action a
Transmitter and receiver nodes probabil-
Minimum energy to transmit and receive
π and Jπ (S0 )
ities of EH
Transmitter policy and total expected
a packet of minimum size
h
min
h
min
ETx
/ETx
and ERx
/ERx
Total number of discrete channel states,
̟(Gt )
Pdrop and P̄drop
cost with given start state S0
Belief of channel state Gt at time slot t
PDP and average PDP
η and ξ
max
max
BTx
and BRx
h
h
ETx
and ERx
min
min
ETx
and ERx
LTx and LRx
G, |G|, γi , and Gt
power
Amplifier drain efficiency and PAPR
Maximum battery size of transmitter and
receiver nodes
Transmitter and receiver EH amounts
its cardinality, ith interval fading gain,
gi and ωo (gi )
and channel state at time slot t
Channel state of interval [γi−1 , γi ) and
ψ(i, j)
its steady state probability
β and x
Number of division of a packet and
number of additional ARQ messages
Stationary
probability
distribution
with transmitter and receiver energies
Ξq,r,w,y
i,j,z,k
min
min
(iETx
, jERx
)
Transition probability of going from
state (i, j, z, k) to state (q, r, w, y)
NAKx: ⌈xm/(β log2 M )⌉ symbols of the transmitted packet are sampled and the rest is discarded
due to the lack of energy. In reply, in the subsequent time slot, the transmitter sends a packet with the
remaining ⌈m(β − x)/(β log2 M )⌉ symbols.
10
Note that the message NAK is different from NAK0, which corresponds to the case when the receiver
does not have enough energy to sample the smallest fraction of the transmitted packet. However, in
both cases, the transmitter retransmits the full packet. As for the conventional scheme, the ACK/NAKx
messages help in estimating the CSI to the transmitter node. Furthermore, it is now evident that the
SS is a function of the available energy at the receiver. Based on the chosen value of β , the receiver
selects ⌈xm/(β log2 M )⌉ symbols for sampling in any given time slot, where x is related to the available
energy at the receiver. The transmitter sends the ⌈m(β − x)/(β log2 M )⌉ symbols after receiving NAKx
message from the receiver. The length of the transmitted packet depends on the energy available at the
transmitter. For example, if β = 4, the variable x takes values from {0, 1, 2, 3, 4}. For x = 3 and if
BRx ≥ xEC,Rx /β = 3EC,Rx /4, the receiver samples ⌈xm/(β log2 M )⌉ = ⌈0.75m/ log 2 M ⌉ symbols
min
then a part of the
and selects NAK3 to feedback to the transmitter. At the transmitter, if BTx ≥ ETx
packet of size ⌈m(β − x)/(β log2 M ⌉) = ⌈0.25m/ log 2 M ⌉ symbols is sent out; otherwise, there is
no transmission. In another example, if BRx ≥ (EC,Rx + Edec ), the receiver samples and decodes the
packet and selects the NAK or ACK message depending on the decoding outcome. In the conventional
retransmission scheme, if the receiver lacks energy to sample the full packet, it samples the packet till the
energy lasts and does not store the samples, and requests the packet retransmission, which requires full
amount of energy. In conclusion, the conventional retransmission scheme wastes energy when compared
with the proposed adaptive one.
IV. P OWER A SSIGNMENT S TRATEGY
Considering the adaptive retransmission scheme, we formulate in this section the power assignment
as a sequential decision problem and then discuss two methods for managing efficiently the harvested
energy at transmitter. At each time slot, the transmitter chooses the energy levels that minimize the
average PEP. The decision is based on the retransmission index, BSI, sequences of past observations and
power assignments at the transmitter. After each transmission, the transmitter receives a feedback, referred
to as observation {ACK, NAK, NAKx}, from the receiver. Furthermore, based on the observation, the
transmitter can also adapt the modulation and coding scheme. However, for sake of tractability, we only
consider the transmit packet size and power adaptations. The problem is considered in infinite horizon.
11
A. Problem Formulation
We formulate the problem by defining the following components: a set of time slots T = {1, 2, . . .}
over which decisions are made, and a set of system states S , a set of transmitter BSI BTx , a set of
FSMC channel states G , a set of retransmission indices K = {1, 2, . . . , K}, a set of actions U , a set
of transition probabilities P , a set of observations Z, and a cost corresponding to every decision. Let
S = BTx × G × K = {(b1 , g1 , k1 ), (b2 , g1 , k1 ), . . . , (b|BTx | , g|G| , k|K| )} denote the complete discrete state
space of the system with a total of |BTx |×|G|×|K| states, where b, g, and k represent the transmitter battery
state, channel state, and retransmission index state, respectively. The state of the system, channel, and
observation at time slot t are represented as St ∈ S , Gt ∈ G , and Zt ∈ Z , respectively. The retransmission
index k tracks the system state within each frame and is reset to one when its maximum value K is
reached or when ACK is received, whichever comes first. Due to EH, the cardinality of the actions set
varies in every time slot. Hence, the set of actions at time slot t is denoted by Ust , {0, 1, 2, . . . , Bst },
where Bst ∈ BTx represents the current state battery level. This is a set consisting of feasible choices
of energy levels corresponding to the transmission of a full packet. An action at ∈ Ust represents the
min = P
energy level at ETx
out Ts in time slot t. For each action taken, the system receives an observation
belonging to the set Z . Note that a set of receiver BSI can be included in the system state space. Since
the exact receiver BSI state is unkown to the transmitter, its value can be estimated using the ARQ
messages likewise the channel state. However, including more unknown states to the system state space
increases the complexity in solving the problem.
We first define the PEP used for the proposed adaptive retransmission scheme. Note that the packet
can be transmitted in parts, as presented in Section-III, and the receiver can decode the packet only
when all the samples are available. Since the different parts of the packet have passed through different
channel states, the part which has passed through the worse channel state leads to the decoding failure. To
simplify, we assume that the full packet is transmitted through the worse channel state, and the receiver
decodes it. Thus, the approximate PEP is the one of the worse part. Consequently, we use the following
approximate PEP expression in the rest of the paper.
Definition 1: For β > 1, the PEP after the transmitter receives Zt = NAK in current time slot t and
12
retransmission index Kt = k is approximated as
Perr (Gt , at ) ≈ max{Pe (Gt−l , at−k )},
l
1 ≤ l ≤ k,
(6)
min via
where Pe (Gt , at ) is the probability that a full packet transmitted in time slot t, with energy at ETx
channel state Gt = g, received in error. at−k corresponds to the energy level used by the transmitter
when sending a new packet at the retransmission k = 1. Through numerical simulations, we have verified
that the approximated PEP approaches the simulated PEP; hence, the approximation (6) is reliable.
Furthermore, for β = 1 the above expression becomes same as Pe (Gt , at ).
Furthermore, PEP is a function of the modulation type and FEC coding used. With the convolution
code, for example, PEP is calculated as [27]:
m
m
X
Ad P2 (d, g, a) ,
Pe g, a ≤ 1 − 1 −
(7)
d=dfree
where dfree is the free distance and Ad is the weight spectra coefficients of the convolutional code.
P2 (d, g, a) is the modulation dependent bit error probability. For example, the bit error probability of
p
binary phase-shift-keying can be approximated by P2 (d, g, a) ≈ 0.5erfc( (dγ̃(g)Pout )/σn2 ), where σn2
is the noise power, erfc(·) is the complementary error function, and γ̃(g) is the average power gain
R γi
R γi
p(γ)dγ , where p(γ) is the
γp(γ)dγ)/ γi−1
in channel state g, which can be found as γ̃(g) = ( γi−1
probability density function of γ , which is distributed exponentially. Here, we assume that the error
detection code is able to find all remaining errors.
t = b, G = g, K = k). Let the ARQ message
Now, we consider a system state at time t as St = (BTx
t
t
at time t be denoted by Zt , where Zt = NAK for decoding failure, NAKx for incomplete transmission
or Zt = ACK for a decoding success. After an action at is taken at time slot t, the current system state
goes to a new state with the transition probability p(St+1 = s′ |St = s, at ) and is associated with a cost.
Let s = (b, g, k) be the current system state, then r(St = s, at ) is the cost defined as:
Perr (g, at ) at ≤ b, Zt = NAK,
r(s, at ) =
0
otherwise.
The cost function is independent of the receiver available energy as it is unknown at the transmitter.
(8)
13
At time slot t, the probability of transition from state s = (b, g, k) to state s′ = (b′ , g′ , k′ ) after taking
an action at , similar to [11], is given as
p(s′ |s, at ) = δ(k′ , k+ )p(g′ |g)ζ((b′ , at , b, k, g),
(9)
where k+ , (k mod K)1Zt+1 6=ACK + 1, with the indicator function 1A equal to 1 if the event A is true,
and to zero otherwise. δ(·, ·) is the Kronecker delta function. δ(k′ , k+ ) ensures that the transmission index
increases by one at each state transition and is reset to one when the maximum retransmission times is
reached or when Zt+1 = ACK is received, whichever occurs first. ζ(b′ , at , b, k, g) is the probability that
the transmitter with current channel state and retransmission state (g, k) moves from battery state b to
another state b′ after taking an action at . For k ≥ 1,
ρRx Perr (g, at )
ρRx (1 − Perr (g, at ))
′
′
ζ(b , at , b, k, g) = η(b , at , b) ×
1 − ρRx
0
Zt+1 = NAK,Zt = NAK, NAKx
Zt+1 = ACK,Zt = NAK, NAKx
(10)
Zt+1 = NAKx, Zt = NAK, NAKx
otherwise,
where η(b′ , at , b) , ρTx δ(b′ , b + LTx − at ) + (1 − ρTx )δ(b′ , b − at ).
At time slot t, the transmitter uses the history of both observation sequence, i.e., zt , [Z1 , . . . , Zt ]
with Z1 = ACK, and previously selected transmit power at−1 , [a1 , . . . , at−1 ] to choose the transmit
power at from the set USt of the admissible power level. The transmit power is selected to minimize the
total expected cost for the current and remaining packets:
a⋆t
∞
X
⋆
, arg min E r(St , at ) +
r(Sk , ak ) zt ,at−1
at ∈USt
k=t+1
for
t = 1, 2, . . . ,
(11)
where a⋆t and a⋆k denote the optimal transmit power assignments for the current time slots t and for future
time slot k = t + 1, respectively. E{·} is the expectation operator.
Let the policy π : S → U specifies the rule for the selection of an action by the transmitter in a given
time slot. Hence, a policy is basically a mapping between what happened in the past and what has to be
14
done at the current state. To find a policy π that minimizes the total expected cost, we cast (11) as an
infinite-horizon Markov decision process (MDP) as
X
T
1
r(St , at ) S0 , zt , at−1 ,
Jπ (S0 ) = lim E
T →∞ T
(12)
t=1
where S0 is a known start state. The optimal policy π ⋆ minimizes the expected long-term average cost
given by (12). The optimal policies obtained in infinite-horizon MDP problems are often stationary, and
hence, simpler to implement compared to what is obtained in a finite-horizon MDP problem that varies
in each time slot. Furthermore, since the total system state space is countable and discrete, and USt is
finite for each St ∈ S , there exists an optimal stationary deterministic policy π ⋆ that minimized the total
expected cost.
B. Solution Methods
This section discusses the solution methods for solving the MDP considered in this work. The following
Bellman equation [28] is used to solve (12),
λ⋆ + h⋆ (s) =
min
at ∈Us ,at ≤Bs
h
i
X
r(s, at ) +
p(s′ |s, at )h⋆ (s′ ) ,
(13)
s′ ∈S
where λ⋆ is the optimal cost and h⋆ (s) is an optimal differential cost or relative value function for each
state s ∈ S . The Bellman equation is a well estabilished and commonly used method for solving a
sequential decision making problem. Interested readers may refer to [22], [28] for further insights. Let
π ⋆ (s) denote the solution of the MDP solved via the value iteration algorithm.
In the formulation described above, the exact CSI is unknown at the transmitter while making the
decisions. Since one of the system state variable is partially known, the problem at hand is commonly
referred to as POMDP [22]. Consequently, based on the observation history, a belief channel state space
of the system is formed. It represents a sufficient statistic for the history of the previous actions and
observations, and adequate actions can be chosen depending upon the belief state. The belief channel
state ̟(Gt ) = p(Gt |zt , at−1 ) is defined as a probability distribution over all possible states conditioned
on the history of previous actions and observations. The belief state at time slot t can be obtained by
expanding the inferred CSI distribution via the Bayes rule
15
̟(Gt ) =
G
X
p(Gt |Gt−1 = gj , zt , at−1 )p(Gt−1 = gj |zt , at−1 )
j=1
=
G
X
(14)
p(Gt |Gt−1 = gj )p(Gt−1 = gj |zt , at−1 ),
j=1
where we use the Markov CSI variation assumption to write (14). Further, with some simple mathematical
manipulations, (14) can be written as:
p(Gt−1 |zt , at−1 ) = PG
p(Zt−1 |at−1 , Gt−1 )p(Gt−1 |zt−1 , at−2 )
l=1 p(Zt−1 |at−1 , Gt−1
= gl )p(Gt−1 = gl |zt−1 , at−2 )
,
(15)
where for Gt = g
p(Zt |at , g) =
ρRx Perr (g, at )
ρRx (1 − Perr (g, at ))
1 − ρRx
p(Zt−1 |at , g)
Zt = NAK, at > 0,
Zt = ACK, at > 0,
(16)
Zt = NAKx, at > 0,
at = 0.
Note that when at = 0, the transmitter is in energy outage; thus, no transmission takes place. Consequently, the acknowledged state also remains the same as the previously received one and so is the
probability.
The POMDP can be solved using dynamic programming to find the optimal policy. However, solving
the POMDP optimally is computationally infeasible for systems with a total number of states higher than
15 [29]. In our case, the total number of system states is large, and hence, the optimal solution is not
presented. POMDPs are PSPACE-complete, i.e., they have high computational complexity and require
large memory that grows exponentially with the horizon [30]. Furthermore, PSPACE-complete problems
are even harder than NP-complete problems. However, many heuristics exist to find suboptimal policies,
e.g., maximum-likelihood policy heuristic (MLPH) [31].
16
C. Proposed Solutions
1) MLPH Power Assignment: We solve the problem in (13) using the MLPH method. In this approach,
we first determine the state that the channel is most likely in, i.e.,
(17)
gML = arg max ̟(Gt ).
Gt ∈G
With γML as the belief channel state at the tth time slot, the corresponding ML state is denoted as
sML = (b, g, k). Then, the transmit power policy is set as
at , π ⋆ (sML ).
(18)
Furthermore, MLPH finds the most probable state of the system from the belief state. When two or
more states are equally likely, MLPH chooses one arbitrarily.
2) Greedy Power Assignment: For low-power wireless sensors, the optimal solution should be avoided
due to the computational complexity constraint. Thus, we turn to a suboptimal greedy power assignment
scheme by modifying (11) as:
āt , arg min E{r(s, at )|zt , at−1 }
at ∈USt
.
(19)
for t = 1, 2, . . . .
The main idea of greedy power assignment is to avoid computation of future dependent expected
cost values. This incurs performance loss, however, at the expense of lower computational and storage
requirement. The greedy power assignment scheme can be rewritten as:
āt = arg min
at ∈USt
G
X
r(s, at )p(Gt = gi |zt , at−1 ).
(20)
i=1
In order to estimate the greedy power assignment, (20) has to be implemented recursively. Using (14)
and (15), we can write the following recursive implementation for the greedy power assignment.
1) Measure Zt , compute p(Zt |at−1 , Gt−1 ) as a function of Gt−1 , and calculate p(Gt |zt , at−1 ) using
(15).
2) Calculate p(Gt |zt , at−1 ) using the Markov prediction step (14).
3) Calculate at via (19).
17
For the initial packets indices t ∈ {1, 2}, we use the initial steady state distribution of states ωo instead
of p(Gt−1 |zt−1 , at−2 ).
3) Implementation Issues: Here, we compare the implementation complexity issues of MLPH and
greedy heuristics. The computation required to solve MLPH is too high to cater by low power wireless
nodes. Thus, similar to [32], we use the memory resource of the sensor nodes rather than the computational
complexity. A look-up table T, which has been pre-computed and stored in the nodes memory, is used
to find the adequate transmit power. It contains the actions for different probabilities of EH, transmitter
side battery, channel, acknowledgement and retransmit index states. The node, at every time slot, updates
the channel belief state ̟(Gt ) and looks up the transmit power at , corresponding to this value.
The memory requirement for storing the look-up table T depends on the total number of the system
space states |S| and the number of actions |U |. The look-up table is stored for different values of
the probabilities of EH. If each EH probability value is divided into κ levels, then the total memory
requirement is κ × |U | × |S|2 bits. Additionally, 10|S| bits of memory are required to store the belief
vector of size |S|, and each element is quantized into 10 levels. On the other hand, the greedy algorithm
requires neither computation nor memory resource of the sensors. It only computes the immediate cost
as a function of the current state of the system including the belief state of the channel. This computation
has very low complexity when compared to computing the expected future costs.
V. PACKET D ROP P ROBABILITY A NALYSIS
In this section, the queuing process induced by the adaptive retransmission scheme is analyzed for
the link between two sensors nodes. In particular, PDP is derived by leveraging tools from the queuing
theory. The PDP is the probability that the transmitted packet has been dropped due to repeatedly decoding
failure or not decoded due to the lack of energy at the receiver over K retransmission attempts. In this
section, we consider that the channel state remains constant for the duration of one frame transmission
and changes to a new state with some transition probability at the start of the new frame. The EH and
consumption models are defined in Section II.
In order to make the PDP analysis tractable, we consider equal and fixed power policy, where the
min . For example, for the case when β = 4,
energy required to transmit a full packet is fixed to βETx
18
if the transmitter battery has energy sufficient to transmit only 1/2 portion of the packet, then the
min . On the other hand, if the transmitter battery has less than the E min , the
transmit energy is 2ETx
Tx
transmit energy is zero. Hence, we approximate the system by discrete-time FSMC, which has the
state space S = BTx × BRx × G × Z × K = {s1 , s2 , . . . , s|S| }. The state at time t is denoted by
t
t
St = (BTx
= i, BRx
= j, Gt = g, Zt = z, Kt = k) where i, j , g, z , and k are the state values of
the battery at the transmitter and receiver nodes, channel, acknowledgement and retransmission index,
respectively.
Let P̄drop denote the average PDP. The packet drop event in a finite battery system is due to either
decoding failure or unavailability of energy at the transmitter or at the receiver during K retransmissions.
As depicted in Fig. 1, a frame consists of minimum 1 to maximum K slots. The acknowledgement
state is z = ACK and the retransmission index state is k = 1 at the start of the frame. At the end of
each slot, k is incremented by 1 if NAKx is received; otherwise, k = 1 if ACK is received. Moreover,
after K retransmission attempts, the value of k is reset to 1. If a NAK is received in the K th attempt,
then the acknowledge state is reset to ACK to indicate the start of a next packet transmission.
The PDP as a function of K ≥ 1 can be written as
P̄drop (K) =
X
i,j
ψ(i, j)Eg Pdrop (K|i, j, g, z = ACK, k = 1) ,
(21)
min and
where ψ(i, j) is the stationary probability that the transmitter and receiver nodes have energy iETx
min , respectively, at the start of the frame. P
jERx
drop (K|i, j, g, z, k) is the PDP conditioned on the channel
min , receiver BSI jE min , acknowledgement state z and
gain being in state Gt = g, transmitter BSI iETx
Rx
retransmission index k at the beginning of the frame. It is given by
Pdrop (K|i, j, g, ACK, 1) = 1 − Psuc ,
(22)
where Psuc is the probability that the packet is successfully decoded within K attempts. Thus, Psuc
is the sum of all possible events contributing to successful packet transmission. It is given by Psuc =
PK
k=1 Psuc,k , where Psuc,k is the probability of success at the k th retransmission index. Accounting for
19
EH events at the transmitter and receiver EHNs, Psuc,k can be upper bounded as:
(
β
X
ϕRx,x
Psuc,k ≤ 1 − ρTx ρRx Perr (ak ) + (1 − ρTx )ρRx ϕTx Perr (ak ) + ρTx (1 − ρRx ) Perr (ak )(ϕRx + ϕdec ) +
x=0
β
X
ϕRx,x
+(1 − ρTx )(1 − ρRx ) ϕTx Perr (ak )(ϕRx + ϕdec ) +
x=0
)
(23)
× (1 − Psuc,k−1 ),
min
min
min
where ϕTx = 1(ak ETx
min
≥ETx ) , ϕRx = 1(jERx
≥ERx ) , ϕdec = 1(jERx
≥Edec ) and ϕRx,x = 1( x+1 EC,Rx >jERx
≥ x EC,Rx ) .
β
β
Perr (ak ) = Perr (g, ak ), where ak denotes the value of action taken at retransmission time index k.
Hereafter, the dependency of g is removed from Perr (g, ak ) since the channel state is assumed fixed
during the packet transmission.
max , B max )] can be comThe stationary probability distribution ψ = [ψ(0, 0), · · · , ψ(i, j), · · · , ψ(BTx
Rx
puted by solving ψ = ψΨg , where Ψg is the transition probability matrix whose elements are given as
P
t+1
t+1
t = i, B t = j, g) , under constraint
= r|BTx
= q, BRx
Eg Pr(BTx
Rx
(i,j) ψg (i, j) = 1. Moreover,
t+1
t+1
t
t
= r|BTx
= i, BRx
= j, g) =
= q, BRx
Eg Pr(BTx
G
X
t+1
t+1
t
t
= r|BTx
= i, BRx
= j),
= q, BRx
ωo (gl )Pr(BTx
(24)
l=1
where the left hand side term is the expected probability that BSI of the transmitter and receiver is q and
r conditioned on previous BSI of i and j , respectively. Furthermore, the right hand side term of (24) can
be given as
Tx
Pr(Bt+1
=
Rx
q, Bt+1
=
r|BtTx
=
i, BtRx
= j) =
|Z| K
X
X
Pr(q, r, w, y|i, j, z = ACK, k = 1).
(25)
w y=1
t+1
t+1
t = i, B t = j, g).
= r|BTx
= q, BRx
We use the transition probability matrix Ξ to evaluate Pr(BTx
Rx
The elements of matrix Ξ represent the transition probability of going from state (i, j, z, k) to another
state (q, r, w, y), which is denoted by Ξq,r,w,y
i,j,z,k with fixed g . We have identified the following four cases
to calculate these elements:
Case i) For z ∈ {ACK/NAKx}, k = 1, . . . , K, and both transmitter and receiver are harvesting
energy, then Ξq,r,w,y
i,j,z,k = ρTx ρRx w11 , where w11 is described in Table IIa. In this case, the receiver does
not feedback NAKx messages and the transmitter resends the full packet since both nodes are harvesting.
Case ii) For z ∈ {ACK/NAKx}, and k = 1, . . . , K , and the transmitter is harvesting energy, while
the receiver is not, then Ξq,r,w,y
i,j,z,k = ρTx (1 − ρRx )w10 , where w10 is described in Table IIb. In this case,
20
the receiver can feedback NAKx messages whenever it does SS. In response, the transmitter can send
the appropriate fraction of the packet.
Case iii) For z ∈ {ACK/NAKx} and k = 1, . . . , K , and the transmitter is not harvesting energy, while
the receiver is harvesting, then Ξq,r,w,y
i,j,z,k = (1 − ρTx )ρRx w01 , where w01 is described in Table IIc. In this
case, the receiver never transmits NAKx messages as it is harvesting the entire time slot. The transmitter
node can decide to transmit or not depending upon the availability of minimum energy. However, if the
current acknowledgement state value is NAKx and the transmitter decides not to transmit, then the next
acknowledgement state remains NAKx.
Case iv) Both transmitter and receiver are not harvesting energy, and k = 1, . . . , K , then Ξq,r,w,y
i,j,z,k =
(1 − ρTx )(1 − ρRx )w00 , where w00 is described in Table IId. In this case, assuming β = 4, if the current
system state is St = (i, j, NAK2, k) such that ϕTx = 1 and ϕRx,1 = 1, then the system moves to state
St+1 = (q, r, NAK1, k + 1) with probability (1 − ρTx )(1 − ρRx ). In another example, if the current system
state is St = (i, j, NAK, k) such that ϕTx = 1 and ϕRx,1 = 1, the system moves to the new state
St+1 = (q, r, NAK1, k + 1) with probability (1 − ρTx )(1 − ρRx ). The energy levels at the receiver nodes
min , a
are defined as aRx = ERx /ERx
dec =
Edec
min ,
ERx
and aRx,x =
xEC,Rx
min .
βERx
VI. N UMERICAL R ESULTS
In this section, we evaluate the performance of the adaptive ACK/NAKx scheme and power assignment
strategy by numerical simulations. Results are compared with the conventional scheme in order to
demonstrate the benefits. The conventional retransmission scheme is denoted by ACK/NAK.
Three metrics are used to evaluate the performance: the average packet transmission time Tp (t), which
is the average time taken per packet to be successfully delivered; PDP, i.e., Pdrop (t), the probability
of dropping a packet after K retransmission attempts; the spectral efficiency, which is the ratio of the
number of successfully transmitted packets to the total number of packets selected from the data buffer
to transmit within a fixed transmission time.
The parameters summarized in Table III are used in numerical simulations unless otherwise mentioned.
The probabilities of EH for both nodes are assumed to be the same, i.e., ρTx = ρRx = ρ. Note that the
slot duration Ts = 1 second, and thus, the power and energy values can be used interchangeably. The
21
e . Furthermore, for low values of β such as
equal power assignment in each time slot is denoted by Pout
4, Pfb is assumed negligible in the simulations.
TABLE III: Simulation Parameters
Parameters
G
c
M and Rc
K
β
Ts and T
ξ, η and α
σn2
PC,Tx and PC,Rx
ρTx and ρRx
Pout and PPA
min
min
ETx
and ERx
max
max
BTx
and BRx
Pdec
h
h
ETx
and ERx
Value
3
128 bits
2 and 1/2
4
4
1 s√and 150 s
√
3( M − 1)/( M + 1), 0.25 and 1
5 mW
0.1 W
[0, 1]
{5, 15} mW and {10, 30} mW
t
t
ETx
/β and ERx
/β
6PTx and 3PTx
7PC,Rx
3PTx Ts and 1.5PRx Ts
10 0
Greedy ACK/NAKx (K = 2)
MLPH ACK/NAKx (K = 2)
GreedyNAKx (K = 3)
MLPH ACK/NAKx (K = 3)
Average PDP
10 -1
K=2
K=3
10 -2
10 -3
0.4
0.5
0.6
0.7
Probability of EH (ρ)
0.8
0.9
Fig. 2: Average packet drop probability (PDP) Pdrop (t) for K = {2, 3}.
1
22
We first compare the performance of the MLPH with the greedy transmit power assignments. Fig. 2
plots the average PDP versus the probability of EH. In order to reduce the overall system states, we
h = 1.2P T . The MLPH and greedy algorithms have similar
set the values of K to 2 and 3, and ERx
Rx s
performance in lower EH rate regime, whereas the MLPH algorithm shows higher gains in higher EH
rate regime. As expected, the performances improve with higher number of retransmission attempts, i.e.,
K = 3. However, when the state size increases, the MLPH becomes impractical and the greedy power
assignment strategy becomes a natural choice. Thus, the following numerical examples only consider the
greedy approach.
8
ACK/NAK, P eout = 5 mW
ACK/NAKx, P e = 5 mW
Average packet transmission time (s)
7
out
ACK/NAK, P eout = 15 mW
e
ACK/NAKx, P out = 15 mW
6
Greedy ACK/NAK
Greedy ACK/NAKx
5
Peout = 5 mW
4
Peout = 15 mW
3
2
Greedy
1
0.2
0.3
0.4
0.5
0.6
0.7
Probability of EH (ρ)
0.8
0.9
1
Fig. 3: Average packet transmission time for K = 4.
In Fig. 3, the average packet transmission time Tp (t) is shown versus the probability of EH, and
the performance of the ACK/NAKx and ACK/NAK in the greedy and equal power assignment settings,
e
= 5 and 15 mW for equal power assignments. We can observe that
respectively, is compared. We set Pout
the ACK/NAKx scheme has the lowest average packet transmission time compared to that of ACK/NAK
23
in both greedy and equal power assignment settings. Furthermore, the performance of the ACK/NAKx
scheme, which employs the equal power assignment of 15 mW is better than the conventional scheme,
which employs the greedy power assignment in the low EH regime. As expected, for low harvesting rates,
all schemes have longer transmission time. Additionally, the ACK/NAK scheme exhibits equal average
e
transmission times for both greedy assignment and equal power assignment with Pout
= 15 mW. This
result means that a higher transmit power helps the equal power assignment algorithm in overcoming the
channel states that are in deep fade. However, it does not help in using the receiver energy efficiently.
Moreover, the performance of ACK/NAKx over the ACK/NAK scheme is significant in lower EH rate
regime. This is because the receiver node, in the latter scheme, processes the received packet by sampling
followed by decoding it. Due to the lack of energy, the signal processing operation stops, which results
in packet drop and loss of energy. Therefore, the receiver has to wait longer to get enough energy to
sample and decode the packet in a single time slot.
10 0
Peout = 5 mW
Average PDP
10 -1
Peout = 15 mW
10 -2
ACK/NAK, P eout = 5mW
ACK/NAKx, P eout = 5 mW
ACK/NAK, P eout = 15 mW
10 -3
0.2
Greedy
ACK/NAKx, P eout = 15 mW
Greedy ACK/NAK
Greedy ACK/NAKx
0.3
0.4
0.5
0.6
0.7
Probability of EH (ρ)
0.8
0.9
Fig. 4: Average packet drop probability (PDP) Pdrop (t) for K = 4.
1
24
In Fig. 4, the PDP of the ACK/NAKx scheme is compared to that of the ACK/NAK scheme, when
employing the greedy and equal power assignments strategies, respectively. The simulation parameters
are the same as those used for Fig. 3. One can see that all retransmission schemes experience high PDP
in low harvesting rate regime. However, the proposed scheme exhibits performance gain particularly in
low EH regime. The greedy ACK/NAKx shows even better gains in all the harvesting rate regimes.
0.9
Spectral efficiency (bits/channel use)
0.8
Greedy
0.7
Peout = 15 mW
0.6
0.5
0.4
ACK/NAK, P eout = 5 mW
Peout = 5 mW
ACK/NAKx, P eout = 5 mW
ACK/NAK, P e = 15 mW
0.3
out
ACK/NAKx, P eout = 15 mW
Greedy ACK/NAK
Greedy ACK/NAKx
0.1
0.2
0.3
0.4
0.5
0.6
Probability of EH (ρ)
0.7
0.8
0.9
1
Fig. 5: Spectral efficiency for a fixed transmission time for K = 4.
To get further insight into the performance gain, Fig. 5 compares the spectral efficiency of the
ACK/NAKx and the ACK/NAK schemes for the fixed transmission time T = 150 s. The ACK/NAKx
scheme has better performance over the ACK/NAK scheme under equal power assignment. Moreover,
the greedy ACK/NAKx has better performance over equal power ACK/NAKx and ACK/NAK. Again, as
expected, the gains are significant in the low EH regime.
In Fig. 6, the simulated and analytical PDP calculated using (21) are compared. In order to reduce the
h = 2P T , B max = 3P T , E h = 1.2P T ,
number of states of discrete-time FSMC, we set new ETx
Tx s
Tx s
Rx s
Tx
Rx
25
10 0
Average PDP
Peout = 5 mW
10 -1
Peout = 15 mW
Analy. ACK/NAKx, P eout = 5 mW
Sim. ACK/NAKx, Peout = 5 mW
Analy. ACK/NAKx, P eout = 15 mW
Sim. ACK/NAKx, Peout = 15 mW
10 -2
0.2
0.3
0.4
0.5
0.6
0.7
Probability of EH (ρ)
0.8
0.9
1
Fig. 6: Comparison of analytical (21) and simulated packet drop probability (PDP).
max = 2P T , and P
BRx
Rx s
dec = 5PC,Rx . It can be seen that the analytical and simulated curves for
e ∈ {5, 15} mW have a similar behavior; however, loose. This is because the exact tight bound for
Pout
the PDP of any modulation with convolution coding is not known for low signal-to-noise ratio (SNR).
Consequently, we use the upper bound on the PEP (7), which is loose in low and tight in high SNR
regime, respectively, in deriving the PDP expression (21).
For further exposition, Fig. 7 compares the performance of the ACK/NAKx and ACK/NAK under
a stochastic EH setup modeled by the compound Poisson process [1], [24]. The compound Poisson
process closely models the EH due to the solar power [33], [34]. In this model, the energy arrivals
follow a Poisson distribution with intensity λ, i.e., inter-arrival time is exponentially distributed with
h
mean 1/λ. The energy amount in each arrival, i.e., E{Tx,Rx}
is i.i.d., with mean Ē{Tx,Rx} . The number
of arrivals in one time slot follows a Poisson distribution with mean λTs . The simulated PDP of the
26
e = 5 mW assignments
ACK/NAKx and ACK/NAK are compared for the greedy and equal power Pout
strategy. The simulation parameters are the same as the ones in Fig. 3, except the harvesting model with
ĒTx = 3PTxTs /2 and ĒRx = 1.5PRxTs /2. The behavior of the schemes is the same as in the case with
the Bernoulli EH arrival process.
10 0
Peout = 5 mW
Average PDP
10 -1
Greedy
10 -2
10 -3
ACK/NAK, P eout = 5 mW
ACK/NAKx, P e = 5 mW
out
Greedy ACK/NAK
Greedy ACK/NAKx
10 -4
0.5
1
1.5
2
EH intensity (λ)
2.5
3
Fig. 7: Average packet drop probability (PDP) Pdrop (t) for K = 4.
It is worth mentioning that the overall gain of the NAK/NAKx with low complexity greedy power
assignment method is approximately 10−15% when compared to ACK/NAK. While this gain may appear
low, it is significant in WSN where multiple sensors are in operation.
VII. C ONCLUSIONS
AND
F UTURE D IRECTIONS
An ARQ based adaptive retransmission scheme between a pair of EH wireless sensor nodes is investigated. In a conventional scheme, the receiver may suspend the sampling and decoding operations
due to insufficient energy, and hence, suffer a loss of both data and harvested energy. To overcome this
27
problem, a selective sampling scheme was introduced, where the receiver selectively samples the received
data and stores it. The selection depends on the amount of energy available. The receiver performs the
decoding operation when both complete samples of the packet and enough energy are available. Selective
sampling information is fed back to the transmitter by resorting to the conventional ARQ scheme. The
transmitter uses this information to re-size the packet length. A POMDP formulation was setup to further
optimize the transmit power. A suboptimal greedy power assignment method was developed, which is
well suited for low power wireless nodes from the implementation perspective. An analytical upper bound
on the PDP is derived for the proposed adaptive retransmission scheme. Simulation results agree with
the analytical solution when the fixed power assignment policy is used. Numerical results demonstrated
that the adaptive retransmission scheme and power assignment strategy provide better performance over
the conventional scheme.
The proposed adaptive retransmission framework paves the way to several other interesting research
avenues. The proposed scheme can be investigated with more sophisticated retransmission scheme, i.e.,
type-II HARQ. To conceptualize the performance of the proposed scheme in a practical scenario, one
can consider a system setup with multiple transmitters and one receiver, all with EH. Problems to be
studied under this setting are more challenging due to the involvement of an increased number of random
variables associated to each node.
R EFERENCES
[1] O. Ozel , K. Tutuncuoglu, J. Yang, S. Ulukus, and A. Yener, “Transmission with energy harvesting nodes in fading wireless
channels: Optimal policies,” IEEE J. Select. Areas Commun., vol. 29, no. 8, pp. 1732–1743, Sep. 2011.
[2] M. Gorlatova, A. Wallwater, and G. Zussman, “Networking low-power energy harvesting devices: Measurements and
algorithms,” in Proc. IEEE Int. Conf. Comput. Commun., Shanghai, China, Apr. 10–15 2011, pp. 1602–1610.
[3] J. Yang and S. Ulukus, “Optimal packet scheduling in an energy harvesting communication system,” IEEE Trans. Commun.,
vol. 60, no. 1, pp. 220–230, Jan. 2012.
[4] K. Tutuncuoglu and A. Yener, “Communicating with energy harvesting transmitter and receivers,” in Proc. Inform. Theory
and Applications Workshop, San Diego, CA, USA, Feb. 05–10 2012, pp. 240–245.
[5] H. Mahdavi-Doost and R. D. Yates, “Energy-harvesting receivers: Finite battery capacity,” in Proc. IEEE Int. Symp. Inform.
Theory, Istanbul, Turkey, Mar. 19–21 2013, pp. 1799–1803.
28
[6] R. D. Yates and H. Mahdavi-Doost, “Energy-harvesting receivers:Otimal sampling and decoding policies,” in Proc. IEEE
Global Signal and Inform. Processing, Austin, TX, USA, Dec. 03–05 2013, pp. 367–370.
[7] H. Mahdavi-Doost and R. D. Yates, “Fading channels in energy-harvesting receivers,” in Proc. Conf. Inform. Sciences Syst.
(CISS), Princeton, USA, Mar. 19–21 2014, pp. 1–6.
[8] A. Kansal, J. Hsu, S. Zahedi, and M. B. Srivastava, “Power management in energy harvesting sensor netwroks,” ACM
Trans. Embedded Comput. Syst., vol. 6, no. 4, pp. 1–38, Sep. 2007.
[9] C. R. Murthy, “Power management and data rate maximization in wireless energy harvesting sensors,” Int. J. Wireless Inf.
Netw., vol. 16, no. 3, pp. 102–117, Jul. 2009.
[10] Z. Shenqiu, A. Seyedi, and B. Sikdar, “An analytical approach to the design of energy harvesting wireless sensor nodes,”
IEEE Trans. Wireless Commun., vol. 12, no. 8, pp. 4010–4024, Aug. 2013.
[11] A. Aprem, C. R. Murthy, and N. B. Mehta, “Transmit power control policies for energy harvesting sensors with
retransmissions,” IEEE J. Select. Topics Signal Processing, vol. 7, no. 5, pp. 895–906, Oct. 2013.
[12] S. Zhou, T. Chen, W. Chen, and Z. Niu, “Outage minimization for a fading wireless link with energy harvesting transmitter
and receiver,” IEEE J. Select. Areas Commun., vol. 33, no. 3, pp. 496–511, Mar. 2015.
[13] M. K. Sharma and C. R. Murthy, “Packet drop probability analysis of ARQ and HARQ-CC with energy harvesting
transmitters and receivers,” in Proc. IEEE Global Signal and Inform. Processing, Atlanta,Georgia. USA, Dec. 3–5 2014,
pp. 148–152.
[14] J. Doshi and R. Vaze, “Long term throughput and approximate capacity of transmitter-receiver energy harvesting channel
with fading,” in Proc. IEEE Int. Conf. Commun. Syst., Macau, Nov.19–21 2014, pp. 46–50.
[15] A. Yadav, M. Gonnewardhena, W. Ajib, and H. Elbiaze, “Novel retransmission scheme for energy harvesting transmitter
and receiver,” in Proc. IEEE Int. Conf. Commun., London, UK, Jun.8–12 2015, pp. 4810–4815.
[16] J. A. Stankovic, T. F. Abdelzaher, C. Lu, L. Sha, and J. C. Hou, “Real-time communication and coordination in embedded
sensor networks,” Proc. IEEE, vol. 91, no. 7, pp. 1002–1022, Jul. 2003.
[17] D. Ciuonzo, P. Salvo Rossi, and S. Dey, “Massive MIMO channel-aware decision fusion,” IEEE Trans. Signal Processing,
vol. 63, no. 3, pp. 604–619, Feb. 2015.
[18] A. Shirazinia, S. Dey, D. Ciuonzo, and P. Salvo Rossi, “Massive MIMO for decentralized estimation of a correlated source,”
IEEE Trans. Signal Processing, vol. 64, no. 10, pp. 2499–2512, May 2016.
[19] Q. Bai, A. Mezghani, and J. A. Nossek, “Throughput maximization for energy harvesting receivers,” in Proc. Int. Works.
on Smart Antennas, Stuttgart, Germany, Mar. 13–14 2013, pp. 1–8.
[20] S. Cui, A. Goldsmith, and A. Bahai, “Energy-efficiency of MIMO and cooperative MIMO techniques in sensor networks,”
IEEE J. Select. Areas Commun., vol. 22, no. 6, pp. 1089–1098, Aug. 2004.
[21] P. Grover, K. Woyach, and A. Sahai, “Towards a communication-theoretic understanding of system-level power consumption,” IEEE J. Select. Areas Commun., vol. 29, no. 8, pp. 1744–1755, Sep. 2011.
[22] M. Putterman, Markov Decision Processes: Discrete Stochastic Dynamic Programming.
New York, USA: Wiley-
29
Interscience, 1994.
[23] J. A. Paradiso and M. Feldmeier, A Compact, Wireless, Self-Powered Pushbutton Controller.
Ubicomp 2001, Springer
Berlin Heidelberg, 2001, pp. 299–304.
[24] J. Xu and R. Zhang, “Throughput optimal policies for energy harvesting wireless transmitters with non-ideal circuit power,”
IEEE J. Select. Areas Commun., vol. 32, no. 2, pp. 322–332, Feb. 2014.
[25] H. S. Wang and N. Moayeri, “Finite-state Markov channel-A useful model for radio communication channels,” IEEE Trans.
Veh. Technol., vol. 44, no. 1, pp. 163–171, Feb. 1995.
[26] Q. Zhang and S. A. Kassam, “Finite-state Markov model for Rayleigh fading channels,” IEEE Trans. Commun., vol. 47,
no. 11, pp. 1688–1692, Nov. 1999.
[27] M. B. Pursley and D. J. Taipale, “Error probabilities for spread-spectrum packet radio with convolutional codes and Viterbi
decoding,” IEEE Trans. Commun., vol. 35, no. 1, pp. 1–12, Jan. 1987.
[28] D. P. Bertsekas, Dynamic Programming and Optimal Control, 2nd ed.
Athena Scientific, 2000.
[29] M. L. Littman, A. R. Cassandra, and L. P. Kaelbling, “Learning policies for partially observable environments: Scaling
up,” in Proc. Int. Conf. Mach. Learning, Tahoe City, CA, USA, Jul.9–15 1995, pp. 362–370.
[30] C. H. Papadimitrious and J. N. Tsitsiklis, “The complexity of Markov decision processes,” Mathematics of Operations
Research, vol. 12, no. 3, pp. 441–450, Aug. 1987.
[31] I. Nourbakhsh, R. Powers, and S. Birchfield, “DERVISH: An office-navigating robot,” Artificial Intell. Magazine, vol. 16,
no. 2, pp. 53–60, Summer, 1995.
[32] R. Srivastava and C. E. Koksal, “Energy optimal transmission scheduling in wireless sensor networks,” IEEE Trans. Wireless
Commun., vol. 9, no. 5, pp. 1550–1560, May 2010.
[33] Q. Bai and J. A. Nossek, “Modulation optimization for energy harvesting transmitters with compound Poisson energy
arrivals,” in Proc. IEEE Works. on Sign. Proc. Adv. in Wirel. Comms., Darmstadt, Germany, Jun. 16–19 2013, pp. 764–768.
[34] P. Lee, Z. A. Eu, M. Han, and H. Tan, “Empirical modeling of a solar-powered energy harvesting wireless sensor node
for time-slotted operation,” in Proc. IEEE Wireless Commun. and Netw. Conf., Quintana, Mexico, Mar.28–31 2011, pp.
179–184.
30
(a) Case-I
w11
Perr (ak )
1 − Perr (ak )
0
Conditions
w = NAK, y = mod (k, K) + 1
max
max
q = min {i + LTx − ak , BTx
}, r = min {j + LRx − aRx , BRx
}
w = ACK, y = 1
max
max
q = min {i + LTx − ak , BTx
}, r = min {j + LRx − aRx , BRx
}
otherwise
(b) Case-II
Conditions
w = NAK, y = mod (k, K) + 1
max
q = min{i + LTx − ak , BTx
}, r = j − aRx
for z = NAK, ϕRx
max
q = min{i + LTx − ak , BTx
}, r = j − adec − aRx,x for z = NAKx, ϕdec = 1, ϕRx,x
w = ACK, y = 1
max
q = min{i + LTx − ak , BTx
}, r = j − aRx
for z = NAK, ϕRx
max
q = min{i + LTx − ak , BTx
}, r = j − adec − aRx,x for z = NAKx, ϕdec = 1, ϕRx,x
w = NAKx, y = mod (k, K) + 1
max
q = min{i + LTx − ak , BTx
}, r = j − aRx,x for z = NAK, NAKx, ϕRx,x = 1
otherwise
w10
Perr (ak )
1 − Perr (ak )
1
0
(c) Case-III
w01
Perr (ak )
1 − Perr (ak )
1
0
Conditions
w = NAK, y = mod(k, K) + 1
max
q = i − ak , r = min{j + LRx − aRx , BRx
}
for z = NAK, ϕTx = 1
max
q = i − ak , r = min{j + LRx − aRx , BRx
} for z = NAKx, ϕTx = 1
w = ACK, y = 1
max
q = min{i + LTx − ak , BTx
}, r = j − aRx
for z = NAK, ϕTx = 1
max
q = min{i + LTx − ak , BTx
}, r = j − adec − aRx,x for z = NAKx, ϕTx = 1
w = NAK, NAKx, y = mod (k, K) + 1
max
q = i, r = min{j + LRx , BRx
}, ϕTx = 0
otherwise
(d) Case-IV
w00
Perr (ak )
1 − Perr (ak )
1
0
Conditions
w = NAK, y = mod(k, K) + 1
q = i − ak , r = j − aRx
for z = NAK, ϕTx = ϕRx = 1
q = i − ak , r = j − aRx
for z = NAKx, ϕTx = ϕRx = ϕdec = 1
w = ACK, y = 1
q = i − ak , r = j − aRx
for z = NAK, ϕTx = ϕTx = 1
q = i − ak , r = j − adec for z = NAKx, ϕTx = ϕRx = ϕdec = 1
w = NAKx, y = mod (k, K) + 1
q = i − ak , r = j − aRx,x ϕRx,x ,
for z = NAK, ϕTx = ϕRx,x = 1
q = i − ak , r = j − aRx,x ϕRx,x ,
for z = NAKx, ϕTx = ϕRx,x = 1
q = i, r = j,
for z = NAKx, ϕTx = 0
otherwise
TABLE II: Values of w11 , w10 , w01 , and w00 .
=1
=1
=1
=1
| 7cs.IT
|
Group Synchronization on Grids
arXiv:1706.08561v1 [math.ST] 26 Jun 2017
Emmanuel Abbe∗, Laurent Massoulie†,
Andrea Montanari‡,
Allan Sly§,
Nikhil Srivastava¶
June 28, 2017
Abstract
Group synchronization requires to estimate unknown elements (θ v )v∈V of a compact group
G associated to the vertices of a graph G = (V, E), using noisy observations of the group
differences associated to the edges. This model is relevant to a variety of applications ranging
from structure from motion in computer vision to graph localization and positioning, to certain
families of community detection problems.
We focus on the case in which the graph G is the d-dimensional grid. Since the unknowns θv
are only determined up to a global action of the group, we consider the following weak recovery
question. Can we determine the group difference θ−1
u θ v between far apart vertices u, v better
than by random guessing? We prove that weak recovery is possible (provided the noise is small
enough) for d ≥ 3 and, for certain finite groups, for d ≥ 2. Viceversa, for some continuous
groups, we prove that weak recovery is impossible for d = 2. Finally, for strong enough noise,
weak recovery is always impossible.
1
Introduction
In the group synchronization problem, we are given a (countable) graph G = (V, E), a group G
and, for each edge (u, v) ∈ E, a noisy observation Y u,v . The objective is to estimate group elements
(θ v )v∈V associated to the vertices v ∈ V , under the assumption that the Y u,v are noisy observations
of the group difference between the adjacent vertices. Roughly speaking (see below for a precise
definition):
Y uv = θ −1
u θ v + noise .
(1.1)
In order for the above to be unambiguous, we will assume that an orientation (u, v) is fixed arbitrarily for each edge.
It is useful to introduce two concrete examples.
Example 1.1. The simplest example is G = Z2 = {(+1, −1), · }, the group with elements (+1, −1)
and operation given by ordinary multiplication (equivalently, the group of integers modulo 2). For
∗
Program in Applied and Computational Mathematics, and EE Department, Princeton University
Inria, MSR-Inria Joint Centre
‡
Department of Electrical Engineering and Department of Sttistics, Stanford University
§
Department of Mathematics, Princeton University
¶
Department of Mathematics, University of California, Berkeley
†
1
each edge (u, v) ∈ E we are given Yuv a noisy observation of θ u θ v = θ −1
u θ v . For instance we can
assume that, for some p ∈ [0, 1/2),
(
θu θv
with probability 1 − p,
(1.2)
Y uv =
−θ u θ v
with probability p.
with the (Y uv )(u,v)∈E conditionally independent given (θ v )v∈V . In other words Y uv is the output
of a binary symmetric channel with flip probability p and input θ u θ v .
We will refer to this case as Z2 synchronization.
Example 1.2. Consider G = O(m): the group of m × m orthogonal matrices, with the following
noise model. Let (Z uv )(u,v)∈E be an i.i.d. collection of matrices with i.i.d. standard normal entries,
and define
Y uv = PO(m) (θ −1
u θ v + σZ uv ) .
(1.3)
Here PO(m) is the projector for the Frobenius norm k · kF onto the orthogonal group, namely for a
matrix M with singular value decomposition M = U ΣV T , we set PO(m) (M ) = U V T .
Group synchronization plays an important role in a variety of applications.
Structure from motion is a central problem in computer vision: given multiple images of an object
taken from different points of view (and in presence of noise or occlusions) we want to reconstruct the
3-dimensional structure of the object [MMM13, CMG13, OS15, WBS16]. A possible intermediate
step towards this goal consists in estimating the relative orientation of the object with respect
to the camera in each image. This can be formulated as a group synchronization problem over
G = SO(3), whereby θ u describes the orientation of image u, and pairwise image registration is
used to construct the relative rotations Y uv .
Graph localization and positioning. Consider a set of nodes with positions x1 , . . . , xn ∈ Rd . We
want to reconstruct the nodes positions from noisy measurements of the pairwise distances kxu −
xv k2 . This question arises in sensor network positioning [HB01, OMK10], imaging [CSC12, SS11],
manifold learning [TDSL00], to name only a few applications. It is often the case that measurements
are only available for pairs u, v ∈ [n] that are close enough, e.g. only if kxu − xv k2 ≤ ρ for ρ a
certain communication range [Sin08, JM13].
Graph localization can be interpreted as a group synchronization problem in multiple ways.
First, we can interpret the unknown position xv as a translation and hence view it as a synchronization problem over the group of translations in d dimensions. Alternatively we can adopt a
divide-and-conquer approach following [CSC12]. First, we consider cliques in the graph and find
their relative positions. Then we reconstruct the relative orientations of various cliques, which can
be formulated as an SO(d) synchronization problem.
Community detection and the symmetric stochastic block model. The k-groups symmetric stochastic
block model is a random graph over n vertices generated as follows [Moo17, Abb17]. First, partition
the vertex set into k subsets of size n/k, uniformly at random. Then connect vertices independently,
conditional on the partition. Two vertices are connected with probability p if they belong to the
same subset, and with a smaller probability q < p otherwise. Given a realization of this graph, we
would like to identify the partition. This problem is in fact closely related to synchronizations over
Zk (the group of integers modulo k). Extensions of the stochastic block model where edges are
2
endowed with labels have also been considered [LMX15]. In particular the so-called censored block
model considered in [SLKZ15] corresponds precisely to Example 1.1 on an Erdős-Rényi graph.
The literature on group synchronization is fairly recent and rapidly growing. The articles
[Sin11, WS13] discuss it in a variety of applications and propose several synchronization algorithms,
mostly based on spectral methods or semidefinite programming (SDP) relaxations. Theoretical
analysis — mostly in the case of random (or complete) graphs G — is developed in [ABBS14,
BSAB14, JMRT16, PWBM16]. Most of these studies use perturbation theoretic arguments which
crucially rely on the fact that the Laplacian (or connection Laplacian, [BSS13]) of the underlying
graph has a spectral gap. This paper shows that nontrivial recovery is possible even in the absence of
a spectral gap, as in the case of grids with d ≥ 3, whose Laplacian pseudoinverses have appropriately
bounded trace rather than norm.
In the present paper we are interested in G being the d-dimensional grid1 , d ≥ 2. Namely,
V = Zd , and –to be definite– we orient edges in the positive direction:
E ≡ (x, y) : y − x ∈ {e1 , . . . , ed } ,
(1.4)
where ei = (0, · · · , 0, 1, 0, . . . , 0) is the i-th element of the canonical basis in Rd . We expect other
d-dimensional graphs (e.g. random geometric graphs) to present a qualitatively similar behavior.
By construction, we can hope to determine the unknowns (θ x )x∈Zd only up to a global action
by a group element. In other words, we cannot distinguish between (θ x )x∈Zd and (gθ x )x∈Zd for
some g ∈ G. We thus ask the following weak recovery question:
Is it possible to estimate θ −1
x θ y better than random guessing, as kx − yk2 → ∞?
Note that, in absence of noise (i.e. if Y uv = θ −1
u θ v exactly), the answer is always positive: we can
multiply the observations Y uv ’s along any path connecting x to y to reconstruct exactly θ −1
x θy .
However for any arbitrarily small noise level, errors add up along the path and this simple procedure
is equivalent to random guessing for kx − yk2 → ∞. The weak recovery question hence amounts to
asking whether we can avoid error propagation.
Focusing on the case of compact matrix groups, we will present the following main results:
Low noise, d ≥ 3. For sufficiently low noise, we prove that weak recovery is possible for d ≥ 3
and any group.
High noise. Vice-versa, weak recovery is impossible in any dimension at sufficiently high noise
(or for d = 1 at any positive noise).
Discrete groups. For the special case of Z2 -synchronization, we prove that weak recovery is
possible (at low enough noise) for all d ≥ 2. We expect the same to hold generally for other
discrete groups.
Continuous groups, d = 2. Vice-versa, for the simplest continuous group, SO(2), we prove that
weak recovery is impossible for d = 2.
The above pattern is completely analogous to the one of phase transitions in spin models within
statistical physics [Geo11]. We refer to Section 3 for a discussion of the connection with statistical
physics.
1
The case d = 1 is somewhat trivial.
3
The rest of the paper is organized as follows. Section 2 presents formal definitions and statements
of our main results. In order to achieve optimal synchronization, it is natural to consider the
Bayes posterior of the unknowns (θ v )v∈V , cf. Section 3. While this does not lead directly to
efficient algorithms, it clarifies the connection with statistical physics. Some useful intuition can
be developed by considering the case2 in which θv ∈ R and Yuv = θv − θu + Zuv with (Zuv )(u,v)∈E
i.i.d. noise. This can be treated by elementary methods, cf. Section 4. Finally, Section 5 and 6
prove our positive results (reconstruction is possible) with other proofs deferred to the appendices.
Notations. Throughout the paper we use boldface symbols (e.g. θ x , Y xy ) to denote elements of
the group G, and normal symbols for other quantities (including vectors and matrices).
2
Main results
As mentioned above, G = (V, E) will be the infinite d-dimensional grid, and G a compact matrix
group. Without loss of generality, we will assume G ⊆ O(m) (the group of m × m orthogonal
matrices). We attach to each vertex x ∈ V an element θ x ∈ G which may be deterministic or
random chosen independently from some distribution.
We are given observations Y = (Y xy )(x,y)∈E , Y xy ∈ G, that are conditionally independent
given θ. We assume that observations are unbiased in the following sense:
E{Y xy |θ} = λ θ −1
x θy ,
(2.1)
where the parameter λ ∈ [0, 1] is a natural measure of the signal-to-noise ratio. In particular, λ = 1
corresponds to noiseless observations. The two examples given in the introduction fit this general
definition:
• For Z2 synchronization (cf. Example 1.1) we have E{Y xy |θ} = (1 − 2p) θ −1
x θ y , and therefore
λ = (1 − 2p).
2
• For O(m) synchronization (cf. Example 1.2) we have E{Y xy |θ} = λ(σ 2 ) θ −1
x θ y where σ 7→
λ(σ 2 ) is a continuous function on [0, ∞) with λ(σ 2 ) → 1 as σ 2 → 0 and λ(σ 2 ) → 1 as σ 2 → ∞
(see Appendix A).
A simple mechanism to produce the noisy observations Y xy consists in introducing a probability
kernel Q on G and stipulate that, for each edge (x, y),
−1
P(Y x,y ∈ · |θ) = P(Y x,y ∈ · |θ −1
x θ y ) = Q( · |θ x θ y ) .
(2.2)
In other words, all observations are obtained by passing θ −1
x θ y through the same noisy channel.
While our results do not necessarily assume this structure, both of the examples given above are
of this type.
An estimator is a collection of measurable functions Tuv : Y 7→ Tu,v (Y ) ∈ G indexed by all
vertex pairs u, v ∈ V (here Y = (Y xy )(x,y)∈E denotes the set of all observations).
2
Strictly speaking, this is not a special case of the problem studied in the rest of the paper, because G = R is not
a compact group.
4
Definition 2.1. We say that the weak recovery problem is solvable for the probability distribution
P over (θ, Y ) defined above if there exists an estimator T , and ε > 0, such that
≥ ε > 0.
(2.3)
lim inf P θ x Txy (Y )θ −1
y ∈ · − PHaar ·
kx−yk→∞
TV
Our first result establishes that the problem is solvable if noise is small enough in d ≥ 3
dimensions.
Theorem 1. If d ≥ 3, then there exists λUB ∈ (0, 1) such that, if λ > λUB then the weak recovery
problem is solvable.
If noise is strong enough, the problem becomes unsolvable.
Theorem 2. Assume that:
1. P(Y x,y ∈ · |θ) = P(Y x,y ∈ · |θ −1
x θ y ).
2. P(Y x,y ∈ · |θ −1
x θ y ) has density q(y|θ 0 ), θ 0 ∈ G with respect to the Haar probability measure.
Let pc (d) ∈ (0, 1] the critical threshold for percolation on the d-dimensional grid. If
inf q(y|θ 0 ) > 1 − pc (d) ,
y,θ0
(2.4)
then the weak recovery problem is not solvable.
In particular, for d = 1, the recovery is not solvable as soon as the noise is strictly positive
inf y,θ q(y|θ) > 0 (since pc (d = 1) = 1).
In d = 2 the situation is more complicated. For certain discrete groups the problem is solvable
at low enough noise: we consider here the case G = Z2 , but we expect the same conclusion to
hold more generally. A result related to the next one was established in [HM82] using a Peierls
argument (Section 3 outlines the connection with the statistical physics formulation). We present
here an independent proof that also provides an efficient recovery algorithm.
Theorem 3. Consider d = 2, and G = Z2 , with uniform flip probability p. Then there exists
p∗ ∈ (0, 1) such that, if p ≤ p∗ then the weak recovery problem is solvable.
On the contrary, we expect that weak recovery is not possible in d = 2 dimensions, for continuous
groups even for very weak noise. This is analogous to the celebrated Mermin-Wagner theorem in
statistical mechanics [MW66, Mer67]. For the sake of simplicity, we focus on the case of G =
SO(2) which is isomorphic to U(1), the group of complex variables of unit modulus, with ordinary
multiplication. Let Z a U(1)-valued random variable with density g satisfying
g ∈ C 2,
inf
s∈[0,2π]
g(eis ) > 0 .
(2.5)
We consider observation on the edges corrupted by multiplicative noise
Y xy = θ −1
x θ y Zxy ,
(2.6)
where (Zxy )(x,y)∈E ∼iid g.
Theorem 4. If d = 2 and G = SO(2) with noise model satisfying (2.5) and (2.6), then the weak
recovery problem is not solvable.
5
3
Bayesian posterior and connection to statistical physics
In this section it is convenient to assume a more general model in which the observations Y xy ∈
Rm×m are not necessarily elements of the matrix group G. We assume that the conditional distribution of the observations Y xy given the unknowns θ x is absolutely continuous with respect to
a reference measure P# (independent of θ). In practice, we will take P# to be either the Haar
measure on G, or the Lebesgue measure on Rm×m . We denote the corresponding density by
1
dP
(Y x,y |θ) =
exp − u(θ −1
x θ y ; Y xy ) .
dP#
Z0
(3.1)
where u : G × Rm×m → R ∪ {+∞} is a measurable function bounded below. Applying Bayes
formula, we can write the posterior µY (B) = P(θ ∈ B|Y ) as
n
o
X
1
exp −
u(θ −1
(3.2)
µY (dθ) =
x θ y ; Y xy ) µ0 (dθ) ,
Z(Y )
(x,y)∈E
where µ0 (dθ) = µ0 (dθ1 ) · · · µ0 (dθ n ) is the product Haar measure over the unknowns and Z(Y ) is
a normalization constant. The joint distribution (3.2) takes the form of a Gibbs measure on the
graph G.
Remark 3.1. For Eq. (3.2) to make sense, the graph G needs to be finite. However, the Bayesian
interpretation implies immediately that quantities of interest have a well defined limit over increasing sequences of graphs. In particular, we can take G to be the finite grid with vertex set
V = {−L, . . . , L}d , and edges E = {(x, y) ∈ V × V : y − x ∈ {e1 , . . . , ed }}. Then the quantity
sup
Txy ( · )
P θ x Txy (Y )θ −1
y ∈ · − PHaar
·
(3.3)
TV
is obviously non-decreasing in L (because larger L corresponds to a larger class of estimators) and
hence admits a well defined limit. We will refer succinctly to this L → ∞ limit as the model on
‘the d-dimensional grid’.
In the rest of this section, it will be useful to distinguish between the arguments of the posterior
density (that we will keep denoting by (θ x )x∈V ), and the true unknowns that we will denote by
(θ 0,x )x∈V . We further assume that the function u satisfies
u(θτ ; Y ) = u(τ ; θ −1 Y ) = u(θ; Y τ −1 ).
(3.4)
for any θ, τ ∈ G and any Y ∈ Rm×m . This condition is verified by all of our examples. Thanks
to this symmetry, for any {τ x }x∈V and any Y , the distribution µY ( · ) of θ in (3.2) coincides with
−1
that of {θ x τ −1
x }x∈V where θ is distributed according to µỸ (·), and Ỹ xy = τ x Y xy τ y . By taking
τ x = θ 0,x for all x, we can assume that θ 0,x = I m for all x, which then leads to the (Y xy )(xy)∈E
being i.i.d. with common distribution
Y xy ∼
1
exp − u(I m ; Y xy ) P# (dY xy ) .
Z1
(3.5)
In the jargon of statistical physics, Gibbs measures of the form (3.2) with associated parameters
distribution (3.5) are known as spin-glasses on the ‘Nishimori line.’ These were first introduced
6
for the case θ x ∈ {+1, −1} [Nis81] and subsequently generalized to other groups in [GHLDB85].
Several results about spin glasses on the Nishimori line were derived in [ON93, Nis01] and the
connection with Bayesian statistics was emphasized in [Iba99, Mon08]. The weak recovery phase
transition corresponds to a paramagnetic-ferromagnetic phase transition in physics language.
Example 3.2. The simplest example is the so-called random bond Ising model which is obtained
by taking θ x ∈ {+1, −1} and
o
n X
1
(3.6)
exp β
Y xy θ x θ y ,
µY (θ) =
Z(Y )
(x,y)∈E
where Y xy = +1 with probability 1 − p and Y xy = −1 with probability p. The Nishimori line is
given by the condition β = (1/2) log((1 − p)/p). It is easy to see that this is equivalent to the Bayes
posterior for the Z2 synchronization model of Example 1.1, if we take θ 0,x = +1.
This model has attracted considerable interest within statistical physics. In particular, highprecision numerical estimates of the phase transition location yield pc ≈ 0.1092 (in d = 2) and
pc ≈ 0.233 (in d = 3) [PTPV09, HTPV07].
Example 3.3. Take G = O(m) (the group of orthogonal matrices), and assume
Y xy = θ −1
0,x θ 0,y + σ Z xy
(3.7)
where Z xy is a noise matrix with i.i.d. entries (Z xy )ij ∼ N(0, 1). This model is analogous to the
one of Example 1.2, although we do not project observations onto the orthogonal group.
After a simple calculation, the Gibbs measure (3.2) takes the form
n X
o
1
T
exp β
Tr θ x Y xy θ y
µ0 (dθ) ,
(3.8)
µY (dθ) =
Z(Y )
(x,y)∈E
where β = 1/σ 2 . By the symmetry under O(m) rotations, for the purpose of analysis we can assume
Y xy = I m + σ Z xy which is the usual setting in physics.
Example 3.4. In the case G = SO(2) we can identify θ x with an angle in [0, 2π), and let
Y xy = θ 0,y − θ 0,x + Z xy ,
mod 2π ,
(3.9)
where Z xy is noise with density proportional to exp(−u(z)) for u(z) a periodic function bounded
below.
The Gibbs measure (3.2) takes the form
n
X
o
1
(3.10)
µY (dθ) =
exp −
u Y xy − θ y + θ x µ0 (dθ) .
Z(Y )
(x,y)∈E
For the purpose of analysis we can assume Y xy = Z xy . This is known as the ‘XY model’ in physics.
Our results have direct implications on these models that we summarize in the following statement.
Corollary 3.1. Consider the Gibbs measure (3.2) on the d-dimensional grid, with parameters
Y xy ∈ G distributed according to Eq. (3.5) and satisfying Eq. (2.1). Then, the following hold:
7
1. For d ≥ 3, and G ⊆ O(m) is any compact matrix group, then there exists λUB < 1 such that
the model is in a ferromagnetic phase for any λ > λUB .
2. For the case of Example 3.2 (i.e. G = Z2 ) and d ≥ 2, there exists p∗ ∈ (0, 1) such that the
model is in a ferromagnetic phase for any p ≤ p∗ .
3. For the case of Example 3.4 (i.e. G = SO(2)) and d = 2 the model is not in a ferromagnetic
phase provided z 7→ u(z) is bounded.
4. For any group G, d ≥ 2, there exists a constant c(d) such that, if kuk∞ ≤ c(d) then the model
is not in a ferromagnetic phase.
Furthermore point 1 applies to Example 3.3 as well.
Proof. These statements are merely a translation of Theorems 1, 2, 3, 4 for the case in which channel
observations take values in G. For the case in Example 3.3, note that we can always project Y xy
onto the group O(m), hence recovering the setting of Example 1.2. Since weak recovery is possible
in the latter, it is also possible in the former.
As already pointed out in Section 2, the existence of a ferromagnetic phase for Example 3.2
(i.e. G = Z2 ) was already obtain in [HM82]. Note however that [HM82] estabilish existence
of a spontaneous magnetization, while here we prove the existence of long range point-to-point
correlation, which is equivalent to weak recovery.
4
A toy example
It is instructive to consider a simple example in which G = R is the group of translations on the
real line. This case does not fit the framework of the rest of this paper, but presents the same
dichotomy between d = 2 and d ≥ 3 and can be solved by elementary methods.
Throughout this section, we adopt additive notation, and hence the observation on edge (x, y)
takes the form
Yx,y = θy − θx + Zx,y .
(4.1)
where {Zxy }(x,y)∈E are i.i.d. random variables with mean 0 and variance σ 2 .
To simplify our treatment, we assume the graph to be the discrete torus, with vertex set
V = {1, 2, . . . , L}d and edges E = {(x, x + ej ) : x ∈ V, j ∈ {1, . . . , d}} (where we identify L + 1
with 1). Denoting by D the difference operator on G, the observation can be written as
Y = Dθ + Z .
(4.2)
As usual, θ can be determined only up to a global shift. To resolve this ambiguity, it is convenient
to assume that θ is centered: hθ, 1i = 0. Consider the least square estimator θ̂(Y ) = D † Y = DL† Y
where † denotes the pseudoinverse. A standard calculation [Was13, Theorem 13.13] yields the
following formula for the mean square error
1
E{kθ(Y ) − θk22 }
Ld
σ2
σ2
= d Tr0 (D T D)† = d Tr0 L† .
L
L
MSE(L, σ 2 ) ≡
8
(4.3)
(4.4)
Here we denoted by L = D T D the Laplacian of G and by Tr0 the trace on the subspace orthogonal
to the all-ones vector. The eigenpairs of the Laplacian are [Moh97]:
v(p)x =
Hence
1
eihp,xi ,
λ(p) =
d
X
[2 − 2 cos(p)] ,
Ld/2
i=1
o
n 2π
(n1 , . . . , nd ) : ni ∈ {0, . . . , L − 1} .
p ∈ BL ≡
L
MSE(L, σ 2 ) =
σ2
Ld
X
p∈BL \{0}
1
.
λ(p)
(4.5)
(4.6)
(4.7)
For large L, the sum can be estimated by approximating it via Riemann integrals to yield the
following fact.
Fact 4.1. The mean square error of least-square estimation within the translation synchronization
model of Eq. (4.1) is
L
for d = 1,
12 + OL (1)
1
2
1
MSE(L, σ ) = 2π log L + OL (1) for d = 2,
(4.8)
σ2
C(d) + oL (1)
for d ≥ 3.
where C(d) is a dimension dependent constant.
We observe that this qualitative behavior is the same that we obtain for continuous compact
groups, cf. Theorem 1 and Theorem 4: the weak recovery problem is solvable only for d ≥ 3.
5
Proof of Theorem 1
Throughout this section we assume a probability distribution P over θ, Y satisfying the unbiasedness condition Eq. (2.1). For most of our analysis, we consider general estimators Tuv : Y 7→
Tuv (Y ) ∈ Rm×m whose output is not necessarily in G, and let T uv = Tuv (Y ) (as projecting them
into G at the end can only increase their accuracy). Also, we set u(n) = (n, . . . , n) ∈ Zd , and
denote by P+ the set of infinite increasing paths in the grid, that start at 0.
Throughout the proof, we will use repeatedly the following two elementary facts. First, for any
two matrices A, B,kABkF ≤ kAkF kBkF . Second, if B is an orthogonal matrix, then kABkF =
kAkF .
We start by defining the estimator Tx,y (Y ) for x = 0, y = u(n), and will then generalize it to
other pairs x, y.
Lemma 5.1. Consider d ≥ 3. Then there exists an estimator T = (Tu,v ), and absolute constants
λ0 < 1, C0 such that, for all λ > λ0 and all even n,
E θ 0 T 0,u(n) θ −1
(5.1)
u(n) = I m ,
n
o
2
E θ 0 T 0,u(n) θ −1
(5.2)
u(n) − I m F ≤ C0 m(1 − λ) .
9
Proof. Benjamini, Pemantle and Peres [BPP+ 98] construct a probability measure µ over paths in
P+ satisfying the so called exponential intersection property (EIT). Namely, there exist absolute
constants β∗ < 1, C∗ such that
(µ × µ){(γ1 , γ2 ) ∈ P+ × P+ : |γ1 ∩ γ2 | ≥ k} ≤ C∗ β∗k .
(5.3)
Let P+ (v) be the set of increasing paths starting at 0 and ending at v ∈ Zd . For n even, we construct
a probability measure µn over P+ (u(n)) as follows. Define the hyperplane H(n) = {(x1 , x2 , x3 ) ∈
R3 : x1 + x2 + x3 = 3n/2}, and let Rn (x1 , x2 , x3 ) = (n − x1 , n − x2 , n − x3 ) denote the reflection
with respect to this hyperplane. For γ ∼ µ, let γ (1,n) denote the path obtained by stopping γ when
it hits H(n), and denote by γ (n) its extension obtained by reflecting the with respect to H(n).
We let µn be the probability distribution of γ (n) (note that γ (n) ends at u(n) by construction). It
follows immediately from Eq. (5.3) that µn satisfies the EIT for some new absolute constants C, β,
that are independent of n:
µn × µn {(γ1 , γ2 ) ∈ P+ (u(n)) × P+ (u(n)) : |γ1 ∩ γ2 | ≥ k} ≤ Cβ k .
(5.4)
For a path γ ∈ P+ (u(n)), denote the ordered sequence of directed edges in γ by I1 (γ), . . . , I3n (γ),
where Ij (γ) ∈ E, j ∈ [3n], and define
Y γ := Y I1 (γ) Y I2 (γ) · · · Y I3n (γ) ,
1
T 0,u := 3n Eγ (Y γ ) ,
λ
(5.5)
(5.6)
where Eγ denotes expectation with respect to µn . Note that by the assumption (2.1) we have
EY γ = λ3n θ −1
0 θ u(n) for any γ ∈ P+ (u(n)) and therefore
ET 0,u(n) = θ −1
0 θ u(n) .
(5.7)
Observe that if two paths γ1 , γ2 in P+ (u(n)) intersect in an edge e then they must intersect in the
same position since the paths are increasing, i.e. we must have e = Ik (γ1 ) = Ik (γ2 ) for some k.
Writing for simplicity u = u(n), and denoting by Eγ1 ,γ2 expectation with respect to γ1 , γ2 ∼iid µn
1
T
E T 0,u T T
0,u = 6n Eγ1 ,γ2 EY γ1 (Y γ2 )
λ
1
T
T
= 6n Eγ1 ,γ2 EY I1 (γ1 ) . . . (EY I3n (γ1 ) Y T
I3n (γ2 ) )Y I3n−1 (γ2 ) . . . Y I1 (γ2 )
λ
(a) 1
= 6n Eγ1 ,γ2 λ|γ1 |+|γ2 |−2|γ1 ∩γ2 | I m
λ
= Eγ1 ,γ2 λ−2|γ1 ∩γ2 | I m .
where (a) follows by repeatedly applying the identity Y e Y T
e = Im for any edge e, each time an
intersection appears, and taking expectation
with
respect
to
Y e1 , Y e2 fornot repeated edges. By
this last expression, the trace τ of E T 0,u T T
reads
τ
=
m E λ−2X where X is a random
0,u
variable counting the number of intersections in two paths γ1 , γ2 independently drawn from µn .
10
Thus for λ2 > β,
m−1 τ =
X
λ−2x [P(X ≥ x) − P(X ≥ x + 1)]
(5.8)
X
(5.9)
x≥0
=1+
x>0
P(X ≥ x) λ−2x − λ−2x+2
≤ 1 + (1 − λ2 )
X
C β/λ2
x>0
2
= 1 + (1 − λ )
Cβ
,
λ2 − β
x
where the inequality follows from Eq. (5.4). Thus
o
n
o
n
2
−1
T
E θ 0 T 0,u θ −1
+m
= Tr E θ 0 T 0,u T T
u − Im F
0,u θ 0 − 2Tr E θ 0 T 0,u θ u
=τ −m
≤ (1 − λ2 )m
Cβ
,
λ2 − β
(5.10)
(5.11)
(5.12)
(5.13)
(5.14)
where we used Eq. (5.7) together with our previous bound on τ . The second statement of the
Lemma follows.
Lemma 5.2. Consider any d ≥ 3 and fix ε > 0. For n ∈ N, j ∈ {1, . . . , d}, let v(j, n) ≡ n ej Then
there exists an estimator T = (Tu,v )u,v∈V , and a constant λ(ε) < 1, such that, for all λ > λ(ǫ) and
all n,
n
o
P θ 0 T 0,v(j,n) θ −1
−
I
≥
ε
≤ ε.
(5.15)
m
v(j,n)
F
Proof. Without loss of generality, assume j = 1, and set for simplicity v(n) = v(j, n). Consider
(∗)
first the case of n even and let w(n) ≡ (n/2, n/2, n/2, 0, . . . , 0). Let (T x,y )x,y∈V be the estimator
of Lemma 5.1 (where we use only the observations on the subgraph induced by the hyperplane
{x ∈ Zd : x4 = · · · = xd = 0}). Define
(∗)
(∗)
T 0,v(n) = T 0,w(n) T w(n),v(n) .
(5.16)
1 + kX 1 X 2 − IkF ≤ (1 + kX 1 − IkF )(1 + kX 2 − IkF ),
(5.17)
From the inequality
we get
(∗)
(∗)
−1
−1
−
Ik
θ
1
+
kθ
T
−
Ik
θ
−
Ik
≤
1
+
kθ
T
kθ 0 T 0,v(n) θ −1
F − 1.
F
F
0 0,w(n) w(n)
w(n) w(n),v(n) v(n)
v(n)
By Lemma 5.1 and Markov’s inequality, the probability that one of the Frobenius norms in the righthand side exceeds δ > 0 is at most C0 m(1−λ)/δ2 . Thus with probability at least 1−2C0 m(1−λ)/δ2 ,
one has
2
kθ 0 T 0,v(n) θ −1
v(n) − IkF ≤ (1 + δ) − 1.
The right-hand side is at most 3δ for δ ≤ 1. The announced result follows for the choice λ(ε) =
1 − ε3 /(18mC0 ).
11
We can now prove our main result, that is a strengthening of Theorem 1.
Theorem 5. Consider any d ≥ 3 and fix ε > 0. Then there exists an estimator T = (Tu,v )u,v∈V ,
and a constant λd (ε) < 1, such that, for all λ > λd and all n,
n
o
P θ x T x,y θ −1
−
I
≥
ε
≤ ε.
(5.18)
m F
y
Proof. Without loss of generality, assume x = 0. Further, for j ∈ {0, . . . , d} define w(j) ≡
(#)
(y1 , . . . , yj , 0, . . . , 0). In particular, w(0) = 0 and w(d) = y. Let (T x,y ) be the estimator of
Lemma 5.2, and define
(#)
(#)
(#)
T 0,y = T w(0),w(1) T w(1),w(2) · · · T w(d−1),w(d) .
By Lemma 5.2, for all λ > λ(ε0 ) we have
(#)
P max θ w(j−1) T w(j−1),w(j)θ −1
w(j) − I m
1≤j≤d
F
≥ ε0 ≤ d ε0 .
(5.19)
(5.20)
By repeated application of Inequality (5.17), on the complement of the event in the right-hand side,
one has
kθ 0 T 0,y θ −1
y
d
Y
(#)
−
I
k
+
1
− 1 ≤ ε0 2d .
kθ w(j−1) T w(j−1),w(j)θ −1
− I mk ≤
m
w(j)
j=1
The claim follows by taking ε0 = ε/(2d ) and by setting λd (ε) = λ(ε/2d ).
6
Proof of Theorem 3
We give a multi-scale scheme to reconstruct the unknowns θ = (θ x )x∈Z2 . Without loss of generality
we will consider pairs of vertices u, v in the positive quadrant. For k ≥ 0 let ℓk = 210k(k+1) . We
partition the lattice Z2 into blocks of side-length ℓk as follows,
Bu(k) = {(x1 , x2 ) ∈ Z2 : ui = ⌈xi /ℓk ⌉}
(6.1)
Let B (k) be the set of blocks at level k and let Du,k denote the unique block in B (k) containing u.
(k)
For each block B ∈ B (k) we will define synchronization random variables WB ∈ {−1, 1} that are
Q
(k)
(k)
measurable with respect to {Y xy }x,y∈B . Our estimate for θ u θ −1
v is
k≥0 WDu,k WDv,k . For some
(k)
(k)
large enough k⋆ we have that Du,k⋆ = Dv,k⋆ and so WDu,k WDv,k = 1 for all k ≥ k⋆ . The product of
synchronization variables at u up to level k will be denoted as
W̃u(k) =
k
Y
(ℓ)
WDu,ℓ .
(6.2)
ℓ=1
We say that two blocks B, B ′ ∈ B (k) are adjacent (denoted B ∼ B ′ ) if there exist x ∈ B, x ∈ B ′
such that (x, x′ ) ∈ E. In this case there are exactly ℓk such pairs. We say that B ∼ B ′ is an honest
edge if the following event holds
X
9
(k)
′
ℓk .
(6.3)
Y xx′ θ x θ x′ ≥
A (B, B ) =
10
′
′
x∈B,x ∈B
12
This condition will mean that edges between vertices along the cut will be informative as we try to
synchronize them.
Next we recursively define the set of good level k blocks G (k) . A block B ∈ B (k) is good if
• There is at most one bad (k − 1)-level sub-block of B, that is
o
n
(k−1)
(k−1)
≤ 1.
Bi ∈ B
: Bi ⊂ B, Bi 6∈ G
(6.4)
• All level k − 1 sub-block edges are honest,
\
(6.5)
A(k−1) (B1 , B2 ) .
B1 ,B2 ∈B(k−1)
B1 ,B2 ⊂B, B1 ∼B2
Claim 6.1. There exists p⋆ > 0 such that, if 0 < p < p⋆ then for all B ∈ B (k)
P(B ∈ G (k) ) ≥ 1 − 2−200k−200 .
(6.6)
Proof. We will establish (6.6) inductively. Note that blocks at level 0 are good. First we estimate
1
the probability that the honest edge condition holds. Assuming that p⋆ ≤ 40
,
9
P A(k−1) (B1 , B2 ) = P Bin(ℓk−1 , 1 − p) ≥ ℓk−1
10
9
39
≥ P Bin(ℓk−1 , ) ≥ ℓk−1 ≥ 1 − exp −κ210k(k−1)
40
10
for some κ > 0. Thus
P A(k−1) (B1 , B2 ) ≥ 1 − 2−400k−800
(6.7)
for all sufficiently large k. By taking p⋆ small enough equation (6.7) holds for small k as well and
thus for all k. Hence, since there are 240k level k − 1 sub-blocks in each level k block we have that,
P
\
∈B(k−1)
B1 ,B2
B1 ,B2 ⊂B
40k+1
· 2−400(k−1)−800 ≥ 1 − 2−200k−201 .
A(k) (B1 , B2 )
≥1−2
(6.8)
Since there are no bad sub-blocks at level 0 this implies (6.6) for k = 1. For some k ≥ 2, assume
inductively that equation (6.6) holds up to k − 1. Then, since the event that blocks are good are
independent, for B ∈ B (k) ,
o
n
P
B ′ ∈ B (k−1) : B ′ ⊂ B, B ′ ∈
/ G (k−1) ≥ 2 = P Bin(240k , 2−200(k−1)−200 ) ≥ 2
40k
2
(2−200k )2 ≤ 2−320k ≤ 2−200k−240 .
≤
2
Combining with equation (6.8) we have that
P(B ∈ G (k) ) ≥ 1 − 2−200k−200 ,
as required.
13
(k)
Next we describe how to inductively construct the synchronization variables WB
block B ∗ . For B1 ∼ B2 k-level sub-blocks of B ∗ we let
X
Y B1 ,B2 = sign
W̃x(k) W̃y(k) Y xy
in a k + 1
B1 ∋x∼y∈B2
(k)
We assign the WB as follows:
1. A quartet is a collection of 4 sub-blocks B1 ∼ B2 ∼ B3 ∼ B4 ∼ B1 that form a square of side
Q
(k)
length 2ℓk . A quartet is incoherent if 4i=1 Y Bi ,Bi+1 = −1 where we take B5 = B1 . Let IB ∗
(k)
be the set of sub-blocks of B ∗ that appear in no incoherent quartets. It is possible for IB ∗ to
(k)
be disconnected, in that case take IB ∗ to be the largest component.
(k)
(k)
(k)
2. If possible, assign WB for all B ∈ IB ∗ such that for all adjacent sub-blocks B1 , B2 ∈ IB ∗ we
have that
(k)
(k)
WB1 WB2 = Y B1 ,B2
(6.9)
(k+1)
Denote the event that such an assignment is possible as HB ∗ . If such an assignment is not
(k)
(k)
(k)
possible set all the WB = 1. Set WB = 1 for all B ∈ (IB ∗ )c .
(k)
In the following we will write I = I (k) = IB ∗ omitting arguments when clear from the context. Note
(k)
(k+1)
that on the event HB ∗ , the WB can be found efficiently by assigning the variables iteratively to
satisfy equation (6.9).
Claim 6.2. For k ≥ 1, if B ∈ G (k) is good then the following hold:
(k)
1. HB holds.
(k)
2. There exists a random variable SB ∈ {−1, 1} such that if x ∈ B and on the event
k−1
\n
j=0
o
{Dx,j ∈ G (j) } ∩ {Dx,j ∈ I (j) }
we have that
(k)
θ x = SB W̃x(k) .
3. Furthermore, for any B ′ ∈ G (k) with B ′ ∼ B,
X
(k)
SB W̃x(k) θ x ≥ (1 − 2−8 + 2−10k )ℓk .
(6.10)
(6.11)
(6.12)
x∈B∩∂B ′
(Here ∂B ′ ≡ {x ∈ Z2 : dist(x, B ′ ) = 1}.)
(k)
Note that we do not (and cannot) construct SB and observe that it is used in the analysis
but not the construction. It accounts for the fact that we can only hope to recover the θ u up to a
global multiplicative shift.
14
Proof of Claim 6.2. We proceed inductively. In the base case when k = 0 for x = B ∈ G (0) we may
(0)
(0)
set Sx = θ x . With the convention that an empty product is 1 we have that W̃x = 1 and so
θ x = Sx(0) W̃x(0) .
Now we assume the claim holds for all k′ < k and consider a good block B ∈ G (k) .
1. For any good (k − 1)-level sub-blocks, B1 ∼ B2 in B
X
Y B1 ,B2 = sign
W̃x(k−1) W̃y(k−1) Y xy
B1 ∋x∼y∈B2
(k−1) (k−1)
SB2
= sign SB1
(k−1)
X
SB1
(k−1)
W̃x(k−1) SB2
B1 ∋x∼y∈B2
W̃y(k−1) Y xy
(6.13)
Our inductive hypothesis implies that there are at most 2−8 ℓk−1 vertices x in this sum with
(k−1)
(k−1)
SB
W̃x
6= θ x , thus
X
X
(k−1)
(k−1)
SB1 W̃x(k−1) SB2 W̃y(k−1) Y xy ≥
θ x θ y Y xy − 4 · 2−8 ℓk−1 ,
(6.14)
B1 ∋x∼y∈B2
B1 ∋x∼y∈B2
and so since A(k−1) (B1 , B2 ) holds,
X
(k−1)
(k−1)
SB1 W̃x(k−1) SB2 W̃y(k−1) Y xy
≥
B1 ∋x∼y∈B2
9
−8
−4·2
ℓk−1 > 0 .
10
(6.15)
Combining with equation (6.13) we have that
(k−1) (k−1)
.
Y B1 ,B2 = sign SB1 SB2
(6.16)
It follows that every quartet of good sub-blocks is coherent. If all of the (k − 1)-level quartets
(k−1)
(related by a
of sub-blocks of B are coherent then there are exactly two assignments of WBi
(k−1)
(k−1)
multiplicative factor of −1) satisfying WB1 WB2
= Y B1 ,B2 . If there is one or more incoherent
quartet, this must include the single bad sub-block. The sub-blocks in I are good and there exist
(k)
(k)
two assignments satisfying WB1 WB2 = Y B1 ,B2 for all B1 , B2 ∈ I which are,
(k−1)
WBi
(k−1)
≡ SBi
or
(k−1)
WBi
(k−1)
≡ −SBi
.
(6.17)
(k)
(k)
(k)
In either case the procedure will construct WBi satisfying (6.17) on I and HB holds. We set SB
so that
(k−1)
(k−1)
(k)
≡ SBi .
SB WBi
2. To verify condition (6.11) we see that for x ∈ Bi ,
(k)
(k)
(k)
(k−1)
SB W̃x(k) = SB WBi W̃x(k−1) = SBi
where the last equality used the inductive hypothesis.
15
W̃x(k−1) = θ x ,
3. It remains to check the condition on the boundary of B adjacent to some good block B ′ . Since
any sub-block in I c must be in a quartet with a bad sub-block, there are at most 3 on any side of
B. Thus, summing over sub-blocks Bi of B we have that
X
X
X
(k)
(k)
SB W̃x(k) θ x =
SB W̃x(k) θ x
x∈B∩∂B ′
Bi :Bi ∼B ′ x∈Bi ∩∂B ′
≥
X
(k)
X
SB W̃x(k) θ x − 3ℓk−1
Bi ∈I:Bi ∼B ′ Bi ∋x∼y∈B ′
≥ (1 − 2−8 + 2−10(k−1) )ℓk−1 (220k − 3) − 3ℓk−1
≥ (1 − 2−8 + 2−10k )ℓk
which establishes (6.12).
By the proceeding claim, if u and v are in the same k-level block on the event
(k)
Juv
=
k−1
\n
j=0
o
{Du,j , Dv,j ∈ G (k) } ∩ {Du,j , Dv,j ∈ I}
we have that
(k)
(k)
W̃u(k−1) W̃v(k−1) = θ u SB θ v SB = θ u θ v .
(k−1)
(6.18)
(k−1)
so W̃u
W̃v
correctly recovers θ u θ v . A sufficient condition for Du,k ∈ G (k) ∩ I is that Du,k
and the 8 sub-blocks surrounding it are all good. Thus
(k)
)≥1−
P(Juv
X
18 P(Du,k ∈ G (k) ) ≥ 1 − 18
k ′ ≥1
X
2−200k−200 ≥
k ′ ≥1
Thus
P(W̃u(k−1) W̃v(k−1) = θ u θ v ) ≥
and so the success probability of recovery is at least
and v which completes the proof of Theorem 3.
8
10
>
1
2
9
.
10
8
10
independent of the distance between u
Acknowledgements
This work was partially supported by the following grants: NSF CAREERAward CCF-1552131,
NSF CSOI CCF-0939370 (E.A.); NSF CCF-1319979, NSF DMS-1613091 (A.M); NSF CCF-1553751
and a Sloan Research Fellowship (N.S). We thank the American Institute of Mathematics (San Jose,
CA) where part of this work was carried out.
A
Proof of Eq. (2.1) for O(m) synchronization
Here we prove the remark that –under the model of Example 1.2– E{Y xy |θ} = λ(σ 2 ) θ −1
x θ y . Fixing
for simplicity x = 1, y = 2 and dropping the indices x, y unless necessary, we have Y = Ũ Ṽ
T
X̃ = θ −1
1 θ 2 + σZ has singular value decomposition X̃ = Ũ ΣṼ .
16
T
where
T
T
2
Let X = θ 1 X̃θ −1
2 = U ΣV . Our claim is equivalent to E{U V |θ} = λ(σ )I. By rotational
invariance of the Gaussian distribution, we have X = I + σ G̃ for (Gij )1≤i,j≤m ∼iid N(0, 1) or
–equivalently– X = QT (I + σ G)Q for any Q in O(m). Using the last representation, E =
E{U V T |θ} = E{QT U V T Q|θ} for I + σ G = U ΣV T . This implies that QT EQ = E for any
orthogonal matrix Q, which can hold only if E = λI for some scalar λ.
Continuity and the limit values of λ(σ 2 ) are straightforward.
B
Proof of Theorem 2
Let p ≡ 1 − inf y,θ0 q(y|θ 0 ). We can write the conditional probability density q(y|θ 0 ) as
q(y|θ 0 ) = (1 − p) + p q∗ (y|θ) .
(B.1)
Hence observations (Y xy )(x,y)∈E can be generated as follows. First draw independent random
variables (Uxy )(x,y)∈E ∼iid Bernoulli(p). Then, for each (x, y) ∈ E such that Uxy = 1, draw
an independent observation Y xy ∼ q∗ ( · |θ −1
x θ y ). For (x, y) ∈ E such that Uxy = 0, draw Y xy
according to the Haar measure.
To upper bound the total variation distance in Eq (2.3) we consider the easier problem in which
instead of Y , we are given all the Bernoulli variables U = (Uxy )(x,y)∈E and, for each (x, y) ∈ E such
that Uxy = 1 we are given the group difference Dxy = θ −1
x θ y . Denoting by D = {D xy }(x,y)∈E,Uxy =1 ,
we then have
P θ x Txy (Y )θ −1
≤ sup P θ x Txy (U ; D)θ −1
. (B.2)
y ∈ · − PHaar ·
y ∈ · − PHaar ·
TV
TV
Texy
Consider the percolation process defined by the variables U (whereby edge (x, y) ∈ E is open if
Uxy = 1), and denote by x ∼U y the event that x and y are in the same percolation cluster. If x and
y are not in the same percolation cluster, then the conditional distribution of θ−1
x θ y conditional on
U ; D is uniformly on G. This implies that
≤ P(x ∼U y) .
(B.3)
P θ x Txy (Y )θ −1
y ∈ · − PHaar ·
TV
For p < pc (d), the right hand side goes to 0 as kx − yk → ∞ [Gri89], which yields the claim.
C
Proof of Theorem 4
For s ∈ R and Z ∼ g( · ), we define
ψ(s) = kP Zeis ∈ · − P Z ∈ · k2L2 (g) =
Z
=
Z
2π
0
2π
0
!2
g(ei(t−s) )
− 1 g(ei(t) ) dt
g(ei(t) )
!2
g(ei(t−s) )
g(ei(t) ) dt − 1 .
i(t)
g(e )
(C.1)
Note that ψ(s) is twice differentiable, nonnegative and that ψ(0) = 0 so ψ ′ (s) = 0 and for some
κ = κ(g) > 0,
|ψ(s)| ≤ κ|s|2 .
17
(C.2)
Let u, v ∈ Z2 with L = ku − vk2 , and define the function h : Z2 → R by
log 1 + min(kx − uk2 ; L)
.
h(x) = 1 −
log(L + 1)
(C.3)
2
Note that h(u) = 1, h(v) = 0 and h(x) = 0 for kx − uk2 ≥ L. Fix θ ∈ U(1)Z , s ∈ [0, 2π) and define
(s)
θ (s) by letting θ x = eis θ x . Denote by Pθ (Y ∈ · ) the conditional distribution of the observations
given hidden variables θ. We then have, for a constant C,
Pθ(s) Y ∈ · − Pθ Y ∈ ·
2
TV
2
≤ Pθ(s) Y ∈ · − Pθ Y ∈ ·
L2
Y
=
1 + Pνs Y xy ∈ · − Pν0 Y xy ∈ ·
(x,y)∈E
=
Y
(x,y)∈E
≤
Y
(x,y)∈E
≤
Y
(x,y)∈E
kx−uk≤L
(C.4)
2
L2
1 + ψ s(h(x) − h(y)) − 1
C
(1 + kx − uk2 ) log2 L
−1
(C.5)
(C.6)
1 + κ s2 |h(x) − h(y)|2 ) − 1
1+
(C.7)
−1
= O(1/ log L).
(C.8)
(C.9)
Taking expectation over s uniformly random in [0, 2π) (denoted by Es ), we have, for any measurable
set B,
(s),−1
Es Pθ(s) θ u Tuv (Y )θ v ∈ B = Es Pθ(s) θ (s)
∈ eis B ,
(C.10)
u Tuv (Y )θ v
and therefore
(s),−1
= O(log(1/L)) .
∈ eis B − Pθ θ u Tuv (Y )θ −1
Es Pθ(s) θ (s)
v ∈B
u Tuv (Y )θ v
(C.11)
We next take expectation with respect to (θ x )x∈Z2 i.i.d. uniform in U(1). Note that under this
distribution, also (θ x )x∈Z2 are i.i.d. uniform in U(1). Letting P( · ) = EPθ ( · ), we have
is
−1
= O(log(1/L)) .
Es P θ u Tuv (Y )θ −1
v ∈ e B − P θ u Tuv (Y )θ v ∈ B
(C.12)
Pθ θ u Tuv (Y )θ −1
v ∈ B − PHaar (B) = O(log(1/L)) .
(C.13)
For any fixed B, ξ, Ps (ξ ∈ eis B) = PHaar (B) and hence we get
This proves the impossibility of weak recovery.
References
[Abb17]
Emmanuel Abbe, Community detection and stochastic block models: recent developments, arXiv:1703.10146 (2017).
18
[ABBS14]
Emmanuel Abbe, Afonso S Bandeira, Annina Bracher, and Amit Singer, Decoding
binary node labels from censored edge measurements: Phase transition and efficient
recovery, IEEE Transactions on Network Science and Engineering 1 (2014), no. 1,
10–22.
[BPP+ 98]
Itai Benjamini, Robin Pemantle, Yuval Peres, et al., Unpredictable paths and percolation, The Annals of Probability 26 (1998), no. 3, 1198–1211.
[BSAB14]
Nicolas Boumal, Amit Singer, P-A Absil, and Vincent D Blondel, Cramér–rao bounds
for synchronization of rotations, Information and Inference 3 (2014), no. 1, 1–39.
[BSS13]
Afonso S Bandeira, Amit Singer, and Daniel A Spielman, A cheeger inequality for the
graph connection laplacian, SIAM Journal on Matrix Analysis and Applications 34
(2013), no. 4, 1611–1630.
[CMG13]
Avishek Chatterjee and Venu Madhav Govindu, Efficient and robust large-scale rotation averaging, Proceedings of the IEEE International Conference on Computer Vision,
2013, pp. 521–528.
[CSC12]
Mihai Cucuringu, Amit Singer, and David Cowburn, Eigenvector synchronization,
graph rigidity and the molecule problem, Information and Inference 1 (2012), no. 1,
21–67.
[Geo11]
Hans-Otto Georgii, Gibbs measures and phase transitions, vol. 9, Walter de Gruyter,
2011.
[GHLDB85] Antoine Georges, David Hansel, Pierre Le Doussal, and Jean-Philippe Bouchaud, Exact properties of spin glasses. ii. nishimori’s line: new results and physical implications, Journal de Physique 46 (1985), no. 11, 1827–1836.
[Gri89]
Geoffrey Grimmett, Percolation?, Springer, 1989.
[HB01]
Jeffrey Hightower and Gaetano Borriello, Location systems for ubiquitous computing,
Computer 34 (2001), no. 8, 57–66.
[HM82]
T Horiguchi and T Morita, Existence of the ferromagnetic phase in a random-bond
Ising model on the square lattice, Journal of Physics A: Mathematical and General 15
(1982), no. 2, L75.
[HTPV07]
Martin Hasenbusch, Francesco Parisen Toldin, Andrea Pelissetto, and Ettore Vicari, Critical behavior of the three-dimensional ±J Ising model at the paramagneticferromagnetic transition line, Physical Review B 76 (2007), no. 9, 094402.
[Iba99]
Yukito Iba, The nishimori line and bayesian statistics, Journal of Physics A: Mathematical and General 32 (1999), no. 21, 3875.
[JM13]
Adel Javanmard and Andrea Montanari, Localization from incomplete noisy distance
measurements, Foundations of Computational Mathematics 13 (2013), no. 3, 297–345.
19
[JMRT16]
Adel Javanmard, Andrea Montanari, and Federico Ricci-Tersenghi, Phase transitions
in semidefinite relaxations, Proceedings of the National Academy of Sciences 113
(2016), no. 16, E2218–E2223.
[LMX15]
Marc Lelarge, Laurent Massoulié, and Jiaming Xu, Reconstruction in the labelled
stochastic block model, IEEE Transactions on Network Science and Engineering 2
(2015), no. 4, 152163.
[Mer67]
N David Mermin, Absence of ordering in certain classical systems, Journal of Mathematical Physics 8 (1967), no. 5, 1061–1064.
[MMM13]
Pierre Moulon, Pascal Monasse, and Renaud Marlet, Global fusion of relative motions
for robust, accurate and scalable structure from motion, Proceedings of the IEEE
International Conference on Computer Vision, 2013, pp. 3248–3255.
[Moh97]
Bojan Mohar, Some applications of laplace eigenvalues of graphs, Graph symmetry,
Springer, 1997, pp. 225–275.
[Mon08]
Andrea Montanari, Estimating random variables from random sparse observations,
European Transactions on Telecommunications 19 (2008), no. 4, 385–403.
[Moo17]
Cristopher Moore, The computer science and physics of community detection: landscapes, phase transitions, and hardness, arXiv:1702.00467 (2017).
[MW66]
N David Mermin and Herbert Wagner, Absence of ferromagnetism or antiferromagnetism in one-or two-dimensional isotropic heisenberg models, Physical Review Letters
17 (1966), no. 22, 1133.
[Nis81]
Hidetoshi Nishimori, Internal energy, specific heat and correlation function of the bondrandom Ising model, Progress of Theoretical Physics 66 (1981), no. 4, 1169–1181.
[Nis01]
, Statistical physics of spin glasses and information processing: an introduction,
vol. 111, Clarendon Press, 2001.
[OMK10]
Sewoong Oh, Andrea Montanari, and Amin Karbasi, Sensor network localization from
local connectivity: Performance analysis for the mds-map algorithm, Information Theory (ITW 2010, Cairo), 2010 IEEE Information Theory Workshop on, IEEE, 2010,
pp. 1–5.
[ON93]
Y Ozeki and Hidetoshi Nishimori, Phase diagram of gauge glasses, Journal of Physics
A: Mathematical and General 26 (1993), no. 14, 3399.
[OS15]
Onur Ozyesil and Amit Singer, Robust camera location estimation by convex programming, Proceedings of the IEEE Conference on Computer Vision and Pattern
Recognition, 2015, pp. 2674–2683.
[PTPV09]
Francesco Parisen Toldin, Andrea Pelissetto, and Ettore Vicari, Strong-disorder
paramagnetic-ferromagnetic fixed point in the square-lattice ±J Ising model, Journal
of Statistical Physics 135 (2009), no. 5, 1039–1061.
20
[PWBM16] Amelia Perry, Alexander S Wein, Afonso S Bandeira, and Ankur Moitra, Messagepassing algorithms for synchronization problems over compact groups, arXiv:1610.04583
(2016).
[Sin08]
Amit Singer, A remark on global positioning from local distances, Proceedings of the
National Academy of Sciences 105 (2008), no. 28, 9507–9511.
[Sin11]
, Angular synchronization by eigenvectors and semidefinite programming, Applied and computational harmonic analysis 30 (2011), no. 1, 20–36.
[SLKZ15]
Alaa Saade, Marc Lelarge, Florent Krzakala, and Zdeborová, Spectral detection in the
censored block model, IEEE International Symposium on Information Theory (ISIT),
2015, p. 11841188.
[SS11]
Amit Singer and Yoel Shkolnisky, Three-dimensional structure determination from
common lines in cryo-em by eigenvectors and semidefinite programming, SIAM journal
on imaging sciences 4 (2011), no. 2, 543–572.
[TDSL00]
Joshua B Tenenbaum, Vin De Silva, and John C Langford, A global geometric framework for nonlinear dimensionality reduction, science 290 (2000), no. 5500, 2319–2323.
[Was13]
Larry Wasserman, All of statistics: a concise course in statistical inference, Springer
Science & Business Media, 2013.
[WBS16]
Kyle Wilson, David Bindel, and Noah Snavely, When is rotations averaging hard?,
European Conference on Computer Vision, Springer, 2016, pp. 255–270.
[WS13]
Lanhui Wang and Amit Singer, Exact and stable recovery of rotations for robust synchronization, Information and Inference (2013), iat005.
21
| 7cs.IT
|
Fused Trees: Simple BST balancing method by partial & scheduled rebuilds
Tae Woo Kim
Department of Computer Science
Yonsei University, Korea
[email protected]
Abstract
This paper proposes a simple method to balance a binary search tree, as an attempt to develop the simplest and
most explicit procedure. For a given constant 𝑘 in (0, 0.5), a rebuilt subtree of size 𝑛 rebuilds itself again after
max(1, ⌊𝑘𝑛⌋) insert/delete operations.
With a cost, this scheduled rebuild offers implementational advantages. This cost is frequent rebuilds and a single
integer value to be stored at all nodes. The return is short-coded and explicit extra procedures the only thing needed
to ensure logarithmic search costs and amortized logarithmic insert/delete costs.
1. Introduction
Seeking for a “just-working” solution means preferring the “simplest” solution. Performance deterioration,
neglecting the latest technology and all kinds of things unacceptable in theory can happen in practice, just for
simplicity. An example is the software engineering, or its education. In those fields, the philosophy is often to
produce “just-working” programs, to cite Arne Anderson [1]. As we can see, often, simplicity is the virtue.
This philosophy about “just-working” solutions is also applied to the balanced search tree. The key feature of a
balanced search tree would be the logarithmic costs, sometimes in a probabilistic or amortized sense, for search,
insert, and delete operations. Many examples of a balanced search tree exist, such as the AVL tree [2], or the binary
B-tree [3]. However, in Munro, Papadakis, and Sedgewick [4], they were rarely used in the actual (non-research)
computational practices. Evidently, their cumbersome implementation made their logarithmic efficiencies to be
meaningless, which just emphasizes the importance of a simple balanced search tree. The AA tree [1] is an
example of such simplified tree, which sacrifices some performance for simplicity but still provides logarithmic
costs. This sacrifice can be said to be reasonable since the only thing needed for a “just-working” balanced search
tree is the logarithmic costs.
The balanced search tree must be considered for the “average” programmer. It is a fundamental (and cumbersome)
data structure needed in numerous applications, but nobody welcomes a tiresome progress to be done every time.
Also, note that the notion of a programmer has been becoming broader in modern society. Anyone can become a
software engineer in the field, but not all of them are a computer science graduate or a highly experienced person
[5]. Data structures that require strong comprehension or long-coded procedures, which just fits the most balanced
search trees, get easily neglected in the field. This is because of its heavily time-consuming
implementation/maintenance, or because it is simply too-hard to try for some people.
Thus, we introduce a new method to balance binary search trees. It was done to develop the simplest and most
explicit kind by using the following observations.
-
The process of “detecting unbalance and then rebalancing” can be replaced by “scheduled rebalance”. To be
specific, a rebalance can be deterministically scheduled to occur at the very first moment where such
unbalance can happen. This makes no complicated tree rules, which are used in many balanced search trees
to detect an imbalance, needed to maintain a balance.
-
A single integer value, which will be stored at all nodes, is enough to simply and explicitly implement the
scheduled rebalances.
-
The rebuild procedures can be implemented more explicitly than tree rotation procedures. Tree rotations can
sometimes be done through short-coded procedures [1], but requires a consideration of various conditions.
These can include all possible cases of the tree, complicated tree attributes and restrictions, etc. On the other
hand, rebuild procedures can require less prerequisite considerations, and can be simply operated (See
Section 4.4 for details).
Following is our fused tree’s balancing method, explainable in just one sentence. For a given constant 𝑘 in (0,
0.5), a rebuilt subtree of size 𝑛 rebuilds itself again after max(1, ⌊𝑘𝑛⌋) insert/delete operations. In details, 𝑘 is
α
a given constant that has a similar role with the α in the scapegoat tree, except that 𝑘 is equivalent to 1 −
2
(See Section 3 for details). It shall be decided before the initialization of the tree. Also, the scheduled rebuilds will
be implemented by a single integer value, the “fuse”. The “fuse” refers to the remaining insert/delete operations
until rebuild, and is stored at all nodes. It is set to ⌊𝑘𝑛⌋ at rebuilds (or 0 at a node’s creation), reduced by 1 after
every insert/delete operation. When it is reduced to be no more than 0, it explicitly means that a rebuild is required
at this node. The integer value is named “fuse” since it acts like an explosive’s fuse, which is the string attached
to an explosive and represents the remaining time until an explosion.
This whole simple and explicit method can be easily implemented by adding small extra codes to a naïve binary
search tree (See Section 4 for details). In details, a rebuild procedure and some extra code for the insert/delete
operation is the only thing needed.
In addition, the fused tree is based on the scapegoat tree [6]. The scapegoat tree is already said to be simple [6,
7], but the fused tree made differences from it for even more simplicity. This leads to different performances and
implementations, explained in Section 6.
2. Notations and Notes
The following notations are for node attributes that are actually implemented. Note that they will be used mainly
during the discussions over the operations (Section 4).
For node 𝑁,
𝑁. 𝑘𝑒𝑦 is the key stored at 𝑁
𝑁. 𝑙𝑒𝑓𝑡 is the left child node of 𝑁
𝑁. 𝑟𝑖𝑔ℎ𝑡 is the right child node of 𝑁
𝑁. 𝑓𝑢𝑠𝑒 is the “fuse” value stored at 𝑁
The following notations are for subtree attributes, which are not actually implemented. They will be used mainly
during the discussions over the correctness of the tree (Section 5).
For subtree 𝑇,
𝑇. 𝑟𝑜𝑜𝑡 is the root of 𝑇
𝑇. 𝑙𝑒𝑓𝑡 is the subtree rooted by 𝑇. 𝑟𝑜𝑜𝑡. 𝑙𝑒𝑓𝑡. It refers to the left wing(subtree) of 𝑇
𝑇. 𝑟𝑖𝑔ℎ𝑡 is the subtree rooted by 𝑇. 𝑟𝑜𝑜𝑡. 𝑟𝑖𝑔ℎ𝑡. It refers to the right wing(subtree) of 𝑇
𝑇. 𝑓𝑢𝑠𝑒 is 𝑇. 𝑟𝑜𝑜𝑡. 𝑓𝑢𝑠𝑒
𝑇. 𝑠𝑖𝑧𝑒 is the size of 𝑇
𝑇. 𝑙𝑎𝑠𝑡_𝑟𝑒𝑏𝑢𝑖𝑙𝑑 is the size of 𝑇 after its last rebuild.
Note that in this paper, the term “subtree” is used only for trees with at least one node.
In addition, lowercase letters (such as 𝑛, 𝑘) were used to denote numerical values, and uppercase letters (such as
𝑁, 𝑇) were used to denote nodes or (sub)trees.
3. Preliminary Discussions (α-weight-balanced)
For an α in [0.5, 1), a node 𝑁 is α-weight-balanced if and only if
𝑇. 𝑙𝑒𝑓𝑡. 𝑠𝑖𝑧𝑒 ≤ α ∙ 𝑇. 𝑠𝑖𝑧𝑒 and,
𝑇. 𝑟𝑖𝑔ℎ𝑡. 𝑠𝑖𝑧𝑒 ≤ α ∙ 𝑇. 𝑠𝑖𝑧𝑒
where 𝑁 = 𝑇. 𝑟𝑜𝑜𝑡.
Also, a binary search tree is α-weight-balanced if and only if all its nodes are α-weight-balanced. Note that a
rebuilt binary search tree is 1/2-weight-balanced.
The height of an α-weight-balanced binary search tree is no more than ⌊log 1 𝑛⌋ [6].
𝛼
Thus, in scapegoat trees, this pre-fixed constant α is used to determine whether a rebuild is needed and to
consequentially maintain the tree with logarithmic height.
Similarly, the fused tree will use a pre-fixed constant 𝑘 , where 0 < 𝑘 <
consequentially maintain a
1
2−2𝑘
1
2
, to schedule a rebuild and to
-weight-balanced tree (See Section 4 for more details about how it maintains
balance, and Section 5 for the proof). Note that
1
2
<
1
2−2𝑘
< 1.
4. Operations
Operations will be explained based on its small differences with the naïve BST operations. The differences will
also be noted under the provided pseudocode.
Note that the only difference with naïve ones is about designating the location of rebuild (in the case of
insert/delete operations) and changing the “fuse” values.
4.1 Search
The search operations are performed in the same as in naïve binary search trees. No other procedures or rebuilding
methods are required.
4.2 Insert
Insert operations require a few more procedures compared to naïve binary search trees.
1. All the ancestor nodes of the successfully-inserted node get their “fuse” value decreased by 1.
2. During the process, if a node where its “fuse” value is no more than 0 is found, then that is the location where
a rebuild is needed. If several locations exist, the one with top height becomes the final location.
(Example 4.2.1) In Figure 4.2.1, the integer “1” (red node) gets successfully-inserted into the fused tree, as in a
naïve BST. Next, in Figure 4.2.2, all the ancestor nodes of the inserted node get their “fuse” value reduced by 1.
Several nodes where their “fuse” value was no more than 0 existed, and the one with top height (orange node)
was chosen as the final location of rebuild.
In an object-oriented style, the insert operation can be implemented through two main procedures. One for the
interface and one for the internal recursive function. Note the use of a dummyLeafNode in the latter procedure,
which literally refers to a meaningless leaf node.
Procedure 4.2.1 Insert(value)
Input : value as the value to be inserted.
1: (Root, rebuildLocation) = _Insert(Root, value)
2: if rebuildLocation is not null then
3:
Rebuild(rebuildLocation)
- In line 1, the procedure also receives a result about the location to be rebuilt.
- Also, in line 2 to 3, the procedure checks whether a rebuild is needed, and does a rebuild if necessary.
Procedure 4.2.2 _Insert(root, value)
Input : root as the root of the tree to insert the value, and value as the value to be inserted.
Output : Two results returned. First result is the updated root. Second result is the updated location of
rebuild.
1: if root is null then
2:
set root as a new node
3:
return (root, dummyLeafNode)
4: else if value < root.key then
5:
(root.left, rebuildLocation) = _Insert(root.left, value)
6: else if value > root.key then
7:
(root.right, rebuildLocation) = _Insert(root.right, value)
8: else then
9:
return (root, null)
10:
11: if rebuildLocation is not null then
12:
reduce root.fuse by 1
13:
if root.fuse ≤ 0 then
14:
set rebuildLocation as root
15: return (root, rebuildLocation)
- All the lines of the ‘return’ statements and where returned results are assigned get changed to also return/assign
the location of rebuild.
- Also, in line 11 to 14, the procedure checks if the insert operation was successful and if so, it reduces the “fuse”
value of the root. Then, in the case where the “fuse” value became no more than 0, it updates the location of
rebuild.
4.3 Delete
Same kinds of extra procedures are also required in the delete operation. Just like in insert operations, all “fuse”
values in the trail get decreased and a rebuild occurs at the required location (See section 4.2 for details). However,
note that only the ancestor nodes of the deleted node get their “fuse” values decreased.
Also, as in the insert operations, the implementation can be done through two main procedures. By an interface,
and an internal recursive function. The dummyLeafNode is also used in the delete procedures, too.
Procedure 4.3.1 Delete(value)
Input : value as the value to be deleted.
1: (Root, rebuildLocation) = _Delete(Root, value)
2: if rebuildLocation is not null then
3:
Rebuild(rebuildLocation)
- In line 1, the procedure also receives a result about the location to be rebuilt.
- Also, in line 2 to 3, the procedure checks whether a rebuild is needed, and does a rebuild if necessary.
Procedure 4.3.2 _Delete(root, value)
Input : root as the root of the tree to delete the value, and value as the value to be deleted.
Output : Two results returned. First result is the updated root. Second result is the updated location of
rebuild.
1: if root is null then
2:
return (null, null)
3:
4:
5:
6:
7:
8:
9:
10:
11:
12:
13:
14:
15:
16:
else if value < root.key then
(root.left, rebuildLocation) = _Delete(root.left, value)
else if value > root.key then
(root.right, rebuildLocation) = _Delete(root.right, value)
else then
if both root.left and root.right exist then
set root.key as getMax(root.left).key
(root.left, rebuildLocation) = _Delete(root.left, root.key)
else if root.left exists then
return (root.left, dummyLeafNode)
else if root.right exists then
return (root.right, dummyLeafNode)
else then
return (null, dummyLeafNode)
17:
18:
if rebuildLocation is not null then
19:
reduce root.fuse by 1
20:
if root.fuse ≤ 0 then
21:
22:
set rebuildLocation as root
return (root, rebuildLocation)
Same type of change occurs, compared to the insert operation.
- All the lines of ‘return’ statements and where returned results are assigned get changed to also return/assign the
location of rebuild.
- Also, in line 18 to 21, the procedure checks if the delete operation was successful and if so, it reduces the “fuse”
value of the root. Then, in the case where the “fuse” value became no more than 0, it updates the location of
rebuild.
4.4 Rebuild
Many tree-rebuild procedures exist, such as in [6, 8, 9].
Compared to most of the procedures, only a single process is additionally required in fused trees. That is, the “fuse”
value for all the nodes in the rebuilt tree must be reset. To be specific, for the root of subtree 𝑇, ⌊𝑘 ∙ 𝑇. 𝑠𝑖𝑧𝑒⌋ shall
become its new “fuse” value.
1
Note that for cases where 𝑇. 𝑠𝑖𝑧𝑒 < , the “fuse” value will be reset to 0, though it does not mean it actually
𝑘
requires a rebuild “now”.
(Example 4.4.1) The orange node in Figure 4.4.1 was chosen as the final location of rebuild. Figure 4.4.2 is the
resultant tree, rebuilt and “fuse” values reset.
The paper introducing the scapegoat tree also introduces a way to rebuild a tree [6]. First, it flattens the tree (as if
into a list), and second, recursively constructs a 1/2-weight-balanced tree from the flat tree. In our approach, to
reset all “fuse” values, only the latter part needs to be slightly modified.
Procedure 4.4.1 BuildTree(head, size)
Input : head as the first (most left) node of the flat tree, and size as the size of the flat tree.
Output : Returns the last (most right) node of the constructed 1/2-weight-balanced tree. Its ‘left’ attribute
refers to the root of the constructed tree.
1: if size is 0 then
2:
set head.left as null
3:
set head.fuse as 0
4:
return head
5: set r as BuildTree(head, ⌊(size − 1)/2⌋ )
6: set s as BuildTree(r.right, ⌊(size − 1)/2⌋ )
7: set r.fuse as ⌊k ∙ size⌋
8: set r.right as s.left
9: set s.left as r
10: return s
- In line 3 and 7, the “fuse” attribute gets assigned.
However, this procedure is not always used in actual implementations, possibly because of its complexity. A
frequently-used procedure of using a linear array (or another sequential data structure) exists, which is similar to
the procedure in [8]. The process is simple and explicit. First, it copies all the keys of the subtree into a linear
array in a sorted order, and second, recursively constructs a new 1/2-weight-balanced tree from the array. In our
approach, only the latter part needs a slight change to reset all the “fuse” values.
Procedure 4.4.2 ConvertArrayIntoTree(array, start, finish)1
Input : array as the linear array with all the data of the tree copied in a sorted order, start as the first
index of the array, and finish as the last index of the array
Output : The root of the 1/2-weight-balanced tree is returned.
1: if start > finish then
2:
return null
3: set root.key as array[ ⌊(start + finish)/2⌋ ]
4: set root.fuse as ⌊k ∙ (finish-start+1)⌋
5: set root.left as ConvertArrayIntoTree(array, start, ⌊(start + finish)/2⌋ − 1)
6: set root.right as ConvertArrayIntoTree(array, ⌊(start + finish)/2⌋ + 1, finish)
7: return root
- In line 4, the “fuse” attribute gets assigned.
Note that in the pseudocodes, though the floor function was used to convert fractions into integers, a (automatic)
truncation could be used instead since the values to be floored/truncated will always be positive.
Beyond these, several more applicable rebuild algorithms exist, such as in [8, 9].
Additionally, in actual implementation, rebuild procedures may ignore trees with size less than 3, since already
balanced in all possible forms.
5. Proof of Logarithmic Costs
5.1 α-weight-balanced
It is proven that the fused tree is always α-weight-balanced as follows. This will be done by separating the tree
into two cases mainly based on its size.
First, for small trees (size less than 2/𝒌), they are always rebuilt after an insert/delete operation. This will
be explained through Lemma 5.1.1 and Corollary 5.1.2.
Lemma 5.1.1 For any subtree 𝑇, 𝑇. 𝑓𝑢𝑠𝑒 ≤ ⌊𝑘 ∙ 𝑇. 𝑠𝑖𝑧𝑒⌋
Proof.
For subtree 𝑋 after 𝑖 insert/delete operations, define 𝑓(𝑋, 𝑖) = ⌊𝑘 ∙ 𝑋. 𝑠𝑖𝑧𝑒⌋ − 𝑋. 𝑓𝑢𝑠𝑒
i)
𝑖 = 0 or the 𝑖th operation included a rebuild
𝑓(𝑋, 𝑖) = 0 (∵ 𝑋. 𝑓𝑢𝑠𝑒 = ⌊𝑘 ∙ 𝑋. 𝑠𝑖𝑧𝑒⌋)
ii)
𝑖th operation was a rebuild-unincluded insert operation
⌊𝑘 ∙ 𝑋. 𝑠𝑖𝑧𝑒⌋ increases by 1 or stays the same
𝑋. 𝑓𝑢𝑠𝑒 always decreases by 1
∴ 𝑓(𝑋, 𝑖) > 𝑓(𝑋, 𝑖 − 1)
iii)
𝑖th operation was a rebuild-unincluded delete operation
⌊𝑘 ∙ 𝑋. 𝑠𝑖𝑧𝑒⌋ decreases by 1 or stays the same
𝑋. 𝑓𝑢𝑠𝑒 always decreases by 1
∴ 𝑓(𝑋, 𝑖) ≥ 𝑓(𝑋, 𝑖 − 1)
By i), ii), iii), 𝑓(𝑋, 𝑖) ≥ 0
∴ 𝑇. 𝑓𝑢𝑠𝑒 ≤ ⌊𝑘 ∙ 𝑇. 𝑠𝑖𝑧𝑒⌋
Corollary 5.1.2 If 𝑇. 𝑠𝑖𝑧𝑒 <
2
𝑘
for some subtree 𝑇 , then after an insert/delete operation to 𝑇 , a rebuild is
followed for 𝑇 or a subtree including it.
Proof.
2
𝑇. 𝑓𝑢𝑠𝑒 ≤ ⌊𝑘 ∙ 𝑇. 𝑠𝑖𝑧𝑒⌋ ≤ 1 (∵Lemma 5.1.1, 𝑇. 𝑠𝑖𝑧𝑒 < )
𝑘
Therefore, after an insert or delete operation, 𝑇. 𝑓𝑢𝑠𝑒 ≤ 0.
This makes only two cases possible for the location of rebuild.
i)
An ancestor node of 𝑇. 𝑟𝑜𝑜𝑡 was chosen as the location of rebuild.
ii)
𝑇. 𝑟𝑜𝑜𝑡 was chosen as the location of rebuild.
∴ After the insert/delete operation, a rebuild is followed for 𝑇 or a subtree including it.
Second, for rebuilt trees with size no less than 2/𝒌, they are guaranteed to be
𝟏
𝟐−𝟐𝒌
-weight-balanced even
after no more than ⌊𝒌 ∙ 𝑻. 𝒔𝒊𝒛𝒆⌋ − 𝟏 insert/delete operations. It is explained in the following Lemma 5.1.3,
and implies that such trees are
1
2−2𝑘
-weight-balanced even without a rebuild.
2
Lemma 5.1.3 For a rebuilt subtree 𝑇 where 𝑇. 𝑠𝑖𝑧𝑒 ≥ , after less than ⌊𝑘 ∙ 𝑇. 𝑠𝑖𝑧𝑒⌋ insert/delete operations,
𝑘
the resultant subtree is
1
2−2𝑘
-weight-balanced.
Proof.
Assume there were 𝑎 insertions and 𝑏 deletions.
0 ≤ 𝑎 + 𝑏 ≤ ⌊𝑘 ∙ 𝑇. 𝑠𝑖𝑧𝑒⌋ − 1
For the initial balanced subtree 𝑇, assume 𝑛0 = 𝑇. 𝑠𝑖𝑧𝑒 for convenience.
Also say the resultant subtree as 𝑇′, and 𝑛 = 𝑇 ′ . 𝑠𝑖𝑧𝑒
Note that 𝑛 = 𝑛0 + 𝑎 − 𝑏 .
For any subtree 𝑐ℎ𝑖𝑙𝑑 between 𝑇. 𝑙𝑒𝑓𝑡 and 𝑇. 𝑟𝑖𝑔ℎ𝑡,
𝑐ℎ𝑖𝑙𝑑. 𝑠𝑖𝑧𝑒 ≤
𝑛0
2
(∵ 𝑇 is 1/2-weight-balanced)
Thus, for any subtree 𝑐ℎ𝑖𝑙𝑑’ between 𝑇’. 𝑙𝑒𝑓𝑡 and 𝑇’. 𝑟𝑖𝑔ℎ𝑡,
𝑐ℎ𝑖𝑙𝑑′. 𝑠𝑖𝑧𝑒 ≤
𝑛0
2
+𝑎
Since 𝑎 + 𝑏 ≤ ⌊𝑘𝑛0 ⌋ − 1 ≤ 𝑘𝑛0 − 1,
𝑏 ≤ 𝑘𝑛0 − 𝑎 − 1.
Therefore,
𝑛 = 𝑛0 + 𝑎 − 𝑏 ≥ 𝑛0 − 𝑘𝑛0 + 2𝑎 + 1
1
1
⟺
𝑛≥
(𝑛 − 𝑘𝑛0 + 2𝑎 + 1)
2 − 2𝑘
2 − 2𝑘 0
𝑛0
2𝑎𝑘 + 1
=
+𝑎+
2
2 − 2𝑘
𝑛0
≥
+𝑎
2
≥ 𝑐ℎ𝑖𝑙𝑑′. 𝑠𝑖𝑧𝑒
∴ 𝑇′. 𝑙𝑒𝑓𝑡. 𝑠𝑖𝑧𝑒 ≤
∴ 𝑇′ is
1
2−2𝑘
1
2−2𝑘
𝑇′. 𝑠𝑖𝑧𝑒, and 𝑇′. 𝑟𝑖𝑔ℎ𝑡. 𝑠𝑖𝑧𝑒 ≤
1
2−2𝑘
𝑇′. 𝑠𝑖𝑧𝑒
-weight-balanced
By combining two of the cases, it can be explained that the fused tree is always
1
2−2𝑘
-weight-balanced. This can
be done by dividing the cases for a fused tree as the i)initial state, ii)rebuilt state, iii)transferred state, and iv)notrebuilt state.
Theorem 5.1.4 The fused tree is
1
2−2𝑘
-weight-balanced.
Proof.
For a subtree 𝑇 in the fused tree, four cases can be supposed.
i)
𝑇 had no insertion/deletion after its creation.
𝑇. 𝑙𝑒𝑓𝑡. 𝑠𝑖𝑧𝑒 = 𝑇. 𝑟𝑖𝑔ℎ𝑡. 𝑠𝑖𝑧𝑒 = 0
Thus,
1
2−2𝑘
-weight-balanced.
ii)
𝑇 was rebuilt after the last insertion/deletion to 𝑇.
Subtree 𝑇 is 1/2-weight-balanced.
Thus,
1
2−2𝑘
-weight-balanced.
iii)
A subtree including 𝑇 but not 𝑇 was rebuilt after the last insertion/deletion to 𝑇.
All nodes of 𝑇 were transferred into a 1/2-weight-balanced tree.
iv)
Neither 𝑇 nor a subtree including 𝑇 was rebuilt after the last insertion/deletion to 𝑇.
By the contraposition of Corollary 5.1.2, 𝑇. 𝑙𝑎𝑠𝑡_𝑟𝑒𝑏𝑢𝑖𝑙𝑑 ≥
Therefore, by Lemma 5.1.3, 𝑇 is
By i),ii),iii),iv), the fused tree is
1
2−2𝑘
1
2−2𝑘
-weight-balanced.
-weight-balanced.
2
𝑘
can be deduced.
To repeat, the height of an α-weight-balanced binary search tree is no more than ⌊log 1 𝑛⌋. Thus, search operations
𝛼
have logarithmic costs in the fused tree.
5.2 Amortized Cost of Insert/Delete Operations
The cost of a rebuild-included insert/delete operation can be amortized to have a logarithmic cost. This is since a
rebuild, which costs Ο(𝑇. 𝑙𝑎𝑠𝑡_𝑟𝑒𝑏𝑢𝑖𝑙𝑑), only happens after Θ(𝑇. 𝑙𝑎𝑠𝑡_𝑟𝑒𝑏𝑢𝑖𝑙𝑑) insert/delete operations. To be
more precise, a rebuild cost for max [𝑇. 𝑙𝑎𝑠𝑡_𝑟𝑒𝑏𝑢𝑖𝑙𝑑 + max(1, ⌊𝑘 ∙ 𝑇. 𝑙𝑎𝑠𝑡_𝑟𝑒𝑏𝑢𝑖𝑙𝑑⌋)] nodes gets amortized
into max(1, ⌊𝑘 ∙ 𝑇. 𝑙𝑎𝑠𝑡_𝑟𝑒𝑏𝑢𝑖𝑙𝑑⌋) (the number of operations). The cost gets amortized in an explicit sense.
Thus, by using the potential method of amortized analysis [10], we can show that the amortized cost of an
insert/delete operation is logarithmic. This can be done by (over)charging rebuild-unincluded insert/delete
operations
𝑘+1
𝑘
times its actual cost, and then using the accumulated “credit” to pay for the rebuild cost.
Note that in some cases, there is no “credit” accumulated to pay for the rebuild cost. Such cases are possible when
2
𝑇. 𝑠𝑖𝑧𝑒 < + 1 for a rebuilt subtree 𝑇, where its next insert/delete operation includes a (consecutive) rebuild.
𝑘
These cases can be handled exceptionally, since the rebuild cost in such cases are Ο(1).
Theorem 5.2.1 If a fused tree was built from a sequence of 𝑛 insert operations and 𝑚 search/delete operations,
the amortized cost of an insert/delete operation is Ο(log 𝑛).
Proof.
For a fused tree 𝑇, assume the cost of the rebuild-unincluded insert/delete operation is bound by
𝑎 ∙ log 1 𝑇. 𝑠𝑖𝑧𝑒 + C .
α
Also, assume the cost of rebuild is bounded by 𝑏 ∙ 𝑇. 𝑠𝑖𝑧𝑒 + C .
For after 𝑖 insert/delete operations, say the resultant tree as 𝑇𝑖 .
Also, define the potential function for node 𝑁 as
𝑘+1
Φ𝑁,𝑖 = 𝑏 ∙
(𝑖 − 𝐿𝑁 )
𝑘
where the creation of 𝑁 or the last rebuild for the subtree rooted at 𝑁 was after 𝐿𝑁 insert/delete operations.
Note that since 𝑖 ≥ 𝐿𝑁 , Φ𝑁,𝑖 ≥ 0 .
Several cases can be supposed for the subsequent events of an insertion/deletion.
i)
The 𝑖th insert/delete operation of node 𝑁 did not cause a rebuild.
ci ≤ 𝑎 ∙ log 1 𝑇𝑖 . 𝑠𝑖𝑧𝑒 + 𝐶
𝛼
ĉi = ci +
∑
(Φ𝑀,𝑖 − Φ𝑀,𝑖−1 )
𝑀 is an ancestor node of 𝑁
≤ 𝑎 ∙ log 1 𝑇𝑖 . 𝑠𝑖𝑧𝑒 + 𝑏 ∙
𝛼
𝑘+1
∙ log 1 𝑇𝑖 . 𝑠𝑖𝑧𝑒 + C
𝑘
𝛼
𝑘+1
) log 1 𝑇𝑖 . 𝑠𝑖𝑧𝑒 + C
𝑘
𝛼
∴ ĉi = Ο(log 𝑇𝑖 . 𝑠𝑖𝑧𝑒)
= (𝑎 + 𝑏 ∙
ii)
A rebuild for subtree 𝑋 followed the 𝑖th insertion/deletion of node 𝑁.
This implies 𝑖 = 𝐿𝑁 + max(1, ⌊𝑘 ∙ 𝑋. 𝑙𝑎𝑠𝑡_𝑟𝑒𝑏𝑢𝑖𝑙𝑑⌋), and leads to two cases.
⌊𝑘 ∙ 𝑋. 𝑙𝑎𝑠𝑡_𝑟𝑒𝑏𝑢𝑖𝑙𝑑⌋ ≤ 1
i)
This means Φ𝑋,𝑖 = Φ𝑋,𝑖−1 = 0 (The potential cannot “pay off” the credit).
Since 𝑋. 𝑙𝑎𝑠𝑡_𝑟𝑒𝑏𝑢𝑖𝑙𝑑 <
2
𝑘
and 𝑋. 𝑠𝑖𝑧𝑒 ≤ 𝑋. 𝑙𝑎𝑠𝑡_𝑟𝑒𝑏𝑢𝑖𝑙𝑑 + 1,
2
+1
𝑘
Note that 𝑋. 𝑠𝑖𝑧𝑒 = Ο(1)
𝑋. 𝑠𝑖𝑧𝑒 <
ci ≤ 𝑎 ∙ log 1 𝑇𝑖 . 𝑠𝑖𝑧𝑒 + 𝑏 ∙ 𝑋. 𝑠𝑖𝑧𝑒 + 2C
𝛼
ĉi = ci +
∑
(Φ𝑀,𝑖 − Φ𝑀,𝑖−1 )
𝑀 is an ancestor node of 𝑋.𝑟𝑜𝑜𝑡
𝑘+1
∙ log 1 𝑇𝑖 . 𝑠𝑖𝑧𝑒 + 2C
𝑘
𝛼
𝛼
2
𝑘+1
< 𝑎 ∙ log 1 𝑇𝑖 . 𝑠𝑖𝑧𝑒 + 𝑏 ∙ ( + 1) + 𝑏 ∙
∙ log 1 𝑇𝑖 . 𝑠𝑖𝑧𝑒 + 2C
𝑘
𝑘
𝛼
𝛼
𝑘+1
2
= (𝑎 + 𝑏 ∙
) log 1 𝑇𝑖 . 𝑠𝑖𝑧𝑒 + 𝑏 ∙ ( + 1) + 2C
𝑘
𝑘
α
∴ ĉi = Ο(log 𝑇𝑖 . 𝑠𝑖𝑧𝑒)
≤ 𝑎 ∙ log 1 𝑇𝑖 . 𝑠𝑖𝑧𝑒 + 𝑏 ∙ 𝑋. 𝑠𝑖𝑧𝑒 + 𝑏 ∙
⌊𝑘 ∙ 𝑋. 𝑙𝑎𝑠𝑡_𝑟𝑒𝑏𝑢𝑖𝑙𝑑⌋ > 1
ii)
Say 𝑛0 = 𝑋. 𝑙𝑎𝑠𝑡_𝑟𝑒𝑏𝑢𝑖𝑙𝑑
ci ≤ 𝑎 ∙ log 1 𝑇𝑖 . 𝑠𝑖𝑧𝑒 + 𝑏 ∙ 𝑋. 𝑠𝑖𝑧𝑒 + 2C
α
ĉi = ci +
∑
(Φ𝑀,𝑖 − Φ𝑀,𝑖−1 )
𝑀 is an ancestor node of 𝑋.𝑟𝑜𝑜𝑡
𝑘+1
∙ log 1 𝑇𝑖 . 𝑠𝑖𝑧𝑒 + Φ𝑋,𝑖 − Φ𝑋,𝑖−1 + 2C
𝑘
α
α
𝑘+1
𝑘+1
(𝑘𝑛0 − 1) + 2C
= 𝑎 ∙ log 1 𝑇𝑖 . 𝑠𝑖𝑧𝑒 + 𝑏 ∙ 𝑋. 𝑠𝑖𝑧𝑒 + 𝑏 ∙
∙ log 1 𝑇𝑖 . 𝑠𝑖𝑧𝑒 + 0 − 𝑏 ∙
𝑘
𝑘
α
α
𝑘+1
𝑘+1
(𝑘𝑛0 − 1) + 2C
≤ 𝑎 ∙ log 1 𝑇𝑖 . 𝑠𝑖𝑧𝑒 + 𝑏 ∙ (𝑛0 + 𝑘𝑛0 ) + 𝑏 ∙
∙ log 1 𝑇𝑖 . 𝑠𝑖𝑧𝑒 − 𝑏 ∙
𝑘
𝑘
α
α
𝑘+1
= 𝑎 ∙ log 1 𝑇𝑖 . 𝑠𝑖𝑧𝑒 + 𝑏 ∙
log 1 𝑇𝑖 . 𝑠𝑖𝑧𝑒 + 𝑏 + 2C
𝑘
α
α
𝑘+1
= (𝑎 + 𝑏 ∙
) log 1 𝑇𝑖 . 𝑠𝑖𝑧𝑒 + 𝑏 + 2C
𝑘
α
∴ ĉi = Ο(log 𝑇𝑖 . 𝑠𝑖𝑧𝑒)
≤ 𝑎 ∙ log 1 𝑇𝑖 . 𝑠𝑖𝑧𝑒 + 𝑏 ∙ 𝑋. 𝑠𝑖𝑧𝑒 + 𝑏 ∙
Since 𝑇𝑖 . 𝑠𝑖𝑧𝑒 ≤ 𝑛,
Ο(log 𝑇𝑖 . 𝑠𝑖𝑧𝑒) = Ο(log 𝑛).
Therefore, the amortized cost for an insert/delete operation is Ο(log 𝑛).
This completes the following Theorem 5.2.2.
Theorem 5.2.2 In a fused tree, for a sequence of 𝑛 insert operations and 𝑚 search/delete operations, the
amortized cost is Ο(log 𝑛) per insert/delete operation and Ο(log 𝑥) for search operations, where x is the size
of the tree the search is performed on.
6. Comparison with the Scapegoat Tree
Since the fused tree was based on the scapegoat tree, it shall be compared with it in the following.
Both use partial rebuilds to balance search trees, based on an α-weight-balance.
-
This offers the choice between lower height and rarer rebuilds.
-
Insert/delete operations have an amortized logarithmic cost.
-
Applicable for multidimensional trees. [7]
Fused trees have some implementational differences with the scapegoat tree.
-
Uses consistent methods for both insert and delete operations.
Deciding a rebuild is only done by constantly checking and updating “fuse” values. Not by separated
methods of 1)comparing the depth of the inserted node with its appropriate depth, or 2)comparing the
tree’s current size with the max-reached size of the tree, which needs constant update.
-
The process of deciding a rebuild was simplified.
No need to check whether the tree is actually unbalanced. Thus, also no need to (recursively) count the
size of the tree to find the non-α-weight-balanced node.
-
When several locations for rebuild exists, the one with max height is chosen.
-
A rebuild needs to reset all “fuse” values.
Fused trees have some performance differences with the scapegoat tree.
-
Strictly α-weight-balanced. May lead to smaller height compared to the scapegoat tree.
-
Possibly more frequent rebuilds. A rebuild can occur even for balanced subtrees.
-
Requires an extra integer value at all nodes.
7. Conclusion
With a cost, the scapegoat tree was simplified to be even simpler. This cost was more frequent rebuilds and an
extra integer value to be stored at all nodes. The return was a simple and explicit method of balancing trees, mainly
explainable in one sentence. In situations where this cost is not much concerned, the fused tree will fit perfect as
a simple and “working” balanced search tree.
References
[1] A. Andersson, "Balanced search trees made simple," in Proceedings of the 3rd Workshop on Algorithms and
Data Structures, 1993, pp. 60-71, doi:10.1007/3-540-57155-8 236.
[2] G. Adelson-Velsky and E. Landis, "An algorithm for the organization of information," Dokladi Akademia
Nauk SSSR, vol. 146, 1962, pp. 263-266.
[3] R. Bayer, "Binary B-trees for virtual memory," in Proceedings of the 1971 ACM SIGFIDET (now SIGMOD)
Workshop on Data Description, 1971, pp. 219-235, doi:10.1145/1734714.1734731.
[4] M. J. Ian, T. Papadakis and R. Sedgewick, "Deterministic skip lists," in Proceedings of the third annual
ACM-SIAM symposium on Discrete algorithms, 1992, pp. 367-375.
[5] "Stack Overflow Developer Survey 2017,"
https://insights.stackoverflow.com/survey/2017.
Stack
Overflow,
2017.
[Online].
Available:
[6] I. Galperin and R. L. Rivest, "Scapegoat trees," in Proceedings of the fourth annual ACM-SIAM Symposium
on Discrete algorithms, 1993, pp. 165-174.
[7] A. Andersson, "General Balanced Trees," Journal of Algorithms, vol. 30, 1999, pp. 1-18, doi:
10.1006/jagm.1998.0967.
[8] H. Chang and S. S. Iyangar, "Efficient algorithms to globally balance a binary search tree," Communications
of the ACM, vol. 27, 1984, pp. 695-702, doi:10.1145/358105.358191.
[9] W. A. Martin and D. N. Ness, "Optimizing binary trees grown with a sorting algorithm," Communications
of the ACM, vol. 15, 1972, pp. 88-93, doi:10.1145/361254.361259.
[10] T. H. Cormen, C. E. Leiserson, R. L. Rivest and C. Stein, "Amortized Analysis," in Introduction to
Algorithms, 3rd ed., MIT Press, 2009, pp. 459-463.
| 8cs.DS
|
Automated Verification and Synthesis of
Embedded Systems using Machine Learning
Lucas Cordeiro
Department of Computer Science, University of Oxford, UK
E-mail: [email protected]
arXiv:1702.07847v2 [cs.LO] 28 Feb 2017
ABSTRACT
The dependency on the correct functioning of embedded systems
is rapidly growing, mainly due to their wide range of applications, such as micro-grids, automotive device control, health care,
surveillance, mobile devices, and consumer electronics. Their
structures are becoming more and more complex and now require multi-core processors with scalable shared memory, in order
to meet increasing computational power demands. As a consequence, reliability of embedded (distributed) software becomes a
key issue during system development, which must be carefully addressed and assured. The present research discusses challenges,
problems, and recent advances to ensure correctness and timeliness regarding embedded systems. Reliability issues, in the development of micro-grids and cyber-physical systems, are then
considered, as a prominent verification and synthesis application.
In particular, machine learning techniques emerge as one of the
main approaches to learn reliable implementations of embedded
software for achieving a correct-by-construction design.
1.
INTRODUCTION
Generally, embedded computer systems perform dedicated functions with a high degree of reliability. They are used in a variety of sophisticated applications, which range from entertainment
software, such as games and graphics animation, to safety-critical
systems, including nuclear reactors and automotive controllers [1].
Embedded systems are ubiquitous in modern day information systems, and are also becoming increasingly important in our society,
especially in micro-grids, where reliability and carbon emission reduction are of paramount importance [2], and in cyber-physical
systems (CPS), which demand short development cycles and again
a high-level of reliability [3]. As a consequence, human life has
also become more and more dependent on the services provided
by this type of system and, in particular, their success is strictly
related to both service relevance and quality.
Figure 1 shows embedded systems examples, which typically consist of a human-machine interface (e.g., keyboard and LCD), a
processing unit (e.g., real-time computer system), and an instrumentation interface (e.g., sensor, network, and actuator) [1]. Indeed, many current embedded systems, such as unmanned aerial
vehicles (UAVs) [4] and medical monitoring systems [5], become
interesting solutions only if they can reliably perform their target
tasks. Besides, when physical interaction with the real world is
needed, which happens in CPS, additional care must be taken,
mainly when human action is directly replaced, as in vehicle driving. Regarding the latter, even human-in-the-loop feedback control can be employed, which raises deeper concerns w.r.t. reliability of human behavior modeling and system implementation.
Consequently, it is important to go beyond design correctness and
also address behavior correctness, which may be performed by incorporating system models. In particular, these models can be
Figure 1: An embedded system is part of a well-specified
larger system (intelligent product).
used for synthesizing a given system, ensuring that all needed
functions are correctly implemented and the correct behavior exhibited, i.e., the system is indeed correct by its method of construction [6]. Here, machine learning emerges as a powerful technique to automatically learn the correct behavior of the system,
which must provably satisfy a given correctness specification σ.
Specifically, synthesizers can use σ as starting point and then
incrementally produce a sequence of candidate solutions that satisfy σ, by integrating deductive methods with inductive inference
(learning from counterexamples) [7]. As a result, a given candidate solution can be iteratively refined to match the specification
σ based on a counterexample-guided learning approach.
2.
VERIFICATION AND SYNTHESIS
CHALLENGES FOR EMBEDDED SYSTEMS
State-of-the-art verification methodologies for embedded systems
generate test vectors (with constraints) and use assertion-based
verification and high-level processor models, during simulation [8],
as shown in Figure 2. Here, the main challenges regarding the verification of embedded systems lie on improving coverage, pruning
the state-space exploration during verification, and incorporating
system models, which allow specific checks regarding system behavior and not only code correctness. Additionally, embedded
system verification raises additional challenges, such as: (1) time
and energy constraints; (2) handling of concurrent software; (3)
platform restrictions; (4) legacy designs; (5) support to different
programming languages and interfaces; and (6) handling of nonlinear and non-convex optimization functions.
Indeed, the first two aspects are of extreme relevance in microgrids and cyber-physical systems, in order to ensure reliability,
which is a key issue for (smart) cities, industries, and consumers,
and the third one is essential in systems that implement device
models, such as digital filters and controllers, which present a behavior that is highly dependent on signal inputs and outputs and
whose deployment may be heavily affected by hardware restric-
(RP3) prove correctness and timeliness of embedded systems, by
taking into account stringent constraints imposed by hardware.
synthesis failed
(2a)
candidate
solution
Synthesize
Verify
INPUTS
counterexample
(1)
initial example
of a candidate
solution
(RP5) provide tools and approaches capable of addressing different programming languages and application interfaces, with the
goal of reducing the time needed to adapt current verification
techniques to new developments and technologies.
counterexample
(2c)
correct-byconstruction
code
verification
successful
(3)
(4)
assert
Embedded Software
(x>0)
Microprocessor
model
Specification
(RP4) incorporate knowledge about system purpose and associated features to detect system-level and behavior failures.
(2b)
data
[1..7]
Generate test
vectors with
constraints
(RP6) develop automated synthesis approaches that are algorithmically and numerically sound, in order to handle embedded
(control) software that is tightly coupled with the physical environment by considering uncertain models and FWL effects.
4.
Figure 2: Verification and synthesis methodologies for
embedded systems.
tions. The fourth aspect is inherent to a large number of embedded systems from telecommunications, control systems, and
medical devices. In particular, software developed for those systems has been extensively tested and verified, and also optimized
for efficiency over years. Therefore, when a new product is derived from a given platform, a lot of legacy code is usually reused
for reducing development time and improving code quality. The
fifth aspect is related to the evolution of development processes
and technologies, which may delay the application of suitable verification and synthesis approaches if verifiers and synthesizers do
not support different programming languages and interfaces. The
last one is related to the widespread use of embedded systems in
autonomous vehicle navigation systems, which demand optimization solving during their execution for a wide range of functions,
including non-linear and non-convex optimization functions.
Those challenges place difficulties for developing (reliable) synthesizers for embedded systems, especially for CPS and micro-grids,
where the controlled object (e.g., physical plant) typically exhibits
continuous behavior whereas the controller (usually implemented
by a real-time computer system) operates in discrete time and
over a quantized domain (cf. intelligent product in Figure 1). In
particular, synthesizers for those systems need to consider the effects of the quantizers (A/D and D/A converters), when a digital
equivalent of the controlled object is considered, i.e., a model of
their physical environment. Additionally, finite-precision arithmetic and their related rounding errors need to be considered
when correct-by-construction code is generated for embedded systems. The main challenge lies on exploiting effectively and efficiently counterexamples provided by verifiers to automatically
learn reliable embedded software implementations (cf. Figure 2).
3.
RESEARCH PROBLEM (RP)
This research statement tackles six major problems in computeraided verification and synthesis for embedded systems, which are
(partially) open in current published research.
(RP1) provide suitable encoding into SMT, which may extend
the background theories typically supported by SMT solvers, with
the goal of reasoning accurately and effectively about realistic
embedded (control) software.
(RP2) exploit SMT techniques to leverage bounded model checking of multi-threaded software, in order to mitigate the stateexplosion problem due to thread interleaving.
CURRENT ACHIEVEMENTS AND
FUTURE TRENDS
In order to support SMT encoding (RP1), Cordeiro, Fischer,
and Marques-Silva proposed the first SMT-based BMC for full
C programs, called Efficient SMT-Based Context-Bounded Model
Checker (ESBMC) [9], which was later extended to support C++,
CUDA, and Qt-based consumer electronics applications. This approach was also able to find undiscovered bugs related to arithmetic overflow, buffer overflow, and invalid pointer, in standard
benchmarks, which were later confirmed by the benchmarks’ creators (e.g., NOKIA, NEC, NXP, and VERISEC). Other SMTbased BMC approaches have also been proposed and implemented
in the literature [10, 11], but the coverage and verification time of
all existing ones are still limited to specific classes of programs, especially for those that contain intensive floating-point arithmetic
and dynamic memory allocation. One possible research direction is to bridge the gap between BMC tools and SMT solvers to
propose background theories and develop more efficient decision
procedures to handle specific classes of programs.
The SMT-based BMC approach proposed by Cordeiro, Fischer,
and Marques-Silva was further developed to verify correct lock acquisition ordering and the absence of deadlocks, data races, and
atomicity violations in multi-threaded software based on POSIX
and CUDA libraries (RP2) [12, 13], considering monotonic partialorder reduction and state-hashing techniques, in order to prune
the state-space exploration. Recent advances for verifying multithreaded C programs have been proposed to speed up the verification time, which significantly prune the state-space exploration; however, the class of concurrent programs (e.g., OpenCL
and MPI) that can be verified is still very limited. One possible research direction is to further extend BMC of multi-threaded
programs via Sequentialization [14] and also analyze interpolants
to prove non-interference of context switches [15].
Novel approaches to model check embedded software using k induction and invariants were proposed and evaluated in the literature (RP3), which demonstrate its effectiveness in some real-life
embedded-system applications [16]. However, the main challenge
still remains open, i.e., to compute and strengthen loop invariants
to prove program correctness and timeliness in a more efficient
and effective way, in order to be competitive with other modelchecking approaches [17]. In particular, invariant-generation algorithms have substantially evolved over the last years, with the goal
of discovering inductive invariants of programs or continuously refine them during verification [18]. Yet there is still a lack of studies for exploiting the combination of different invariant-generation
algorithms (e.g., interval analysis, linear inequalities, polynomial
equalities and inequalities) and how to strengthen them.
State-of-the-art SMT-based context-BMC approaches were ex-
tended to verify overflow, limit cycle, stability, and minimum
phase, in digital systems (RP4). Indeed, digital filters and controllers were tackled, in order to verify system-level properties of
those systems, specified as linear-time temporal logic [19, 20]. In
particular, a specific UAV application was tackled, with the goal
to verify its attitude controllers. In general, however, there is
still a lack of studies to verify system-level properties related to
embedded systems; emphasis should be given to micro-grids and
cyber-physical systems, which require high-dependability requirements for computation, control, and communication. Additionally, the application of automated fault detection, localization,
and correction techniques to digital systems represents an important research direction to make BMC tools useful for engineers.
Although ESBMC was extended to support C/C++ and some
variants (RP5), new application interfaces and programming languages are often developed, which require suitable verifiers. Indeed, it would be interesting if a new programming language
model could be loaded, which along with a BMC core could check
different programs. Some work towards that was already presented by [21], which employed operational models for checking
Qt-based programs from consumer electronics. In summary, the
BMC core is not changed, but instead an operational model, which
implements the behavior and features of Qt libraries, is used for
providing the new code structure to be checked. Such research
problem is closely related to the first one (RP1) and has the
potential to devise a new paradigm in software verification.
State-of-the-art synthesis approaches (RP6) for embedded (control) systems typically disregard the platform in which the embedded system software operates and restrict itself to generate code
that do not take into account FWL effects. However, the synthesized system must include the physical plant to avoid serious system’s malfunctioning (or even a catastrophe) due to the embedded
(control) software, e.g., the Mars Polar Lander did not account for
leg compressions prior to landing [22]. Research in this direction
has made some progress to design, implement, and evaluate an
automated approach for generating correct-by-construction digital controllers that is based on state-of-the-art inductive synthesis
techniques [6]. However, there is still little evidence whether that
approach can scale for larger systems modeled by other types of
representations (e.g., state-space). Another research direction for
synthesizers is to automatically produce UAV trajectory and mission planning code, by taking into account system’s dynamics and
nonholonomic constraints. As a result, verifiers and synthesizers
need to handle a wide range of functions, including non-linear and
non-convex optimization problems [23]. Machine learning techniques could be employed here to learn from counterexamples,
i.e., in the inductive step, synthesizers could learn the model from
raw data, and in the deductive step, the model could be applied
to predict the behaviour of new data [24].
5.
CONCLUSIONS
This research statement presented the main challenges related to
the verification of design correctness, in embedded systems, and
also raised some important side considerations about synthesis.
Given that software complexity has significantly increased in embedded products, there is still the need for stressing and exhaustively covering the entire system state space, in order to verify
low-level properties that have to meet the application’s deadline,
access memory regions, handle concurrency, and control hardware
registers. Besides, there is a trend towards incorporating knowledge about the system to be verified, which may take software
verification and synthesis one step further, where not only code
correctness will be addressed, but also full system reliability. Finally, it seems interesting to provide behavioral models when new
application interfaces or programming language features are used,
in order to extend the capabilities of current verification tools,
without changing the core BMC module.
As future perspective, the main goal of this research is to extend
BMC as a verification and synthesis tool for achieving correctby-construction embedded system implementations. Special attention will be given to CPS and modern micro-grids, considering
small-scale versions of a distributed system, so that reliability and
other system-level properties (e.g., carbon emission reduction in
smart cities) are amenable to automated verification and synthesis, probably through behavior models.
6.
REFERENCES
[1] H. Kopetz: Real-Time Systems - Design Principles for Distributed
Embedded Applications. Real-Time Systems Series, Springer, ISBN
978-1-4419-8236-0, pp. 1–376, 2011.
[2] Xua X., Jiaa H., Wanga D., Yub D., Chiangc H.: Hierarchical energy
management system for multi-source multi-product microgrids.
Renewable Energy, v. 78, pp. 621–630, 2015.
[3] Lee E.: The Past, Present and Future of Cyber-Physical Systems: A
Focus on Models. Sensors 15(3): pp. 4837–4869, 2015.
[4] Groza A., Letia I., Goron A., Zaporojan S.: A formal approach for
identifying assurance deficits in unmanned aerial vehicle software.
Progress in Systems Engineering, Springer, pp. 233–239, 2015.
[5] Cordeiro L., Fischer B., Chen H., Marques-Silva J.: Semiformal
Verification of Embedded Software in Medical Devices Considering
Stringent Hardware Constraints. ICESS, pp. 396–403, 2009.
[6] Abate A., Bessa I., Cattaruzza D., Cordeiro L., David C., Kesseli P.,
Kroening D.: Sound and Automated Synthesis of Digital Stabilizing
Controllers for Continuous Plants. HSCC, 2017 (to appear).
[7] Sanjit A. Seshia: Combining Induction, Deduction, and Structure for
Verification and Synthesis. Proc. of the IEEE 103(11): pp. 2036–2051,
2015.
[8] Behrend J., Lettnin D., Gruenhage A., Ruf J., Kropf T., Rosenstiel
W.: Scalable and Optimized Hybrid Verification of Embedded
Software. J. Electronic Testing 31(2): pp. 151–166, 2015.
[9] Cordeiro L., Fischer B., Marques-Silva J.: SMT-based Bounded Model
Checking for Embedded ANSI-C Software. IEEE Trans. Software Eng.
38(4), pp. 957–974, 2012.
[10] Armando A., Mantovani J., Platania L.: Bounded Model Checking of
Software Using SMT Solvers Instead of SAT Solvers. SPIN, LNCS
3925, pp. 146-162, 2006.
[11] Merz F., Falke S., Sinz C.: LLBMC: Bounded Model Checking of C
and C++ Programs using a Compiler IR. VSTTE, LNCS 7152, pp.
146–161, 2012.
[12] Cordeiro L., Fischer B.: Verifying Multi-threaded Software using
SMT-based Context-Bounded Model Checking. ICSE, pp. 331–340,
2011.
[13] Pereira P. Albuquerque H., Marques H., Silva I., Carvalho C., Santos
V., Ferreira R., Cordeiro L.: SMT-Based Context-Bounded Model
Checking for CUDA Programs. Concurrency and Computation: P&E.,
2017 (to appear).
[14] Inverso O., Tomasco E., Fischer B., La Torre S., Parlato G.: Bounded
Model Checking of Multi-threaded C Programs via Lazy
Sequentialization. CAV, LNCS 8559, pp. 585-602, 2014.
[15] K. McMillan: Widening and Interpolation. SAS, LCNS 6887, pp. 1,
2011.
[16] Gadelha M., Ismail H., Cordeiro L.: Handling Loops in Bounded
Model Checking of C Programs via k-Induction. STTT 19(1), pp.
97–114, 2017.
[17] Rocha, W., Rocha, H., Ismail H., Cordeiro, L., Fischer, B.: DepthK:
A k-Induction Verifier Based on Invariant Inference for C Programs
(Competition Contribution). TACAS, 2017 (to appear).
[18] Beyer D., Dangl M., Wendler P.: Boosting k-Induction with
Continuously-Refined Invariants. CAV, LNCS 9206, pp. 622–640,
2015.
[19] Morse J., Cordeiro L., Nicole D., Fischer B.: Model Checking LTL
Properties over ANSI-C Programs with Bounded Traces. Software
and System Modeling 14(1): pp. 65–81, 2015.
[20] Bessa I., Ismail H., Palhares, R., Cordeiro L., Chaves Filho J.:
Formal Non-Fragile Stability Verification of Digital Control Systems
with Uncertainty. IEEE Trans. Computers 66(3), pp. 545–552, 2017.
[21] Sousa F., Garcia M., Cordeiro, L., Lima Filho E.: Bounded Model
Checking of C++ Programs based on the Qt Cross-Platform
Framework. Softw. Test., Verif. Reliab., 2017 (to appear).
[22] Jackson D., Vaziri M.: Correct or usable? the limits of traditional
verification. ESEC/SIGSOFT FSE, pp. 11, 2016.
[23] Araujo, R., Albuquerque, H., Bessa, I., Cordeiro, L., Chaves Filho, J.:
Counterexample Guided Inductive Optmization. Sci. Comput.
Program., 2017 (under review).
[24] Alur R., Bodı́k R., Juniwal G., Martin M., Raghothaman M., Seshia
S., Singh R., Solar-Lezama A., Torlak E., Udupa A.: Syntax-guided
synthesis. FMCAD, pp. 1–8, 2013.
| 3cs.SY
|
Effective algorithms for homogeneous utility functions.
arXiv:1801.00777v1 [cs.DS] 29 Dec 2017
Alexander Shananin
∗
Sergey Tarasov
†
tweets: I am an economist so I can ignore
computational constraints.
I am a computer scientist, so I can ignore
gravity.
Lance Fortnow, a computer scientist
Abstract
Under the assumption of (positive) homogeinity (PH in the sequel) of the corresponding
utility functions, we construct polynomial time algorithms for the weak separability, the collective consumption behavior and some related problems. These problems are known to be at
least N P -hard if the homogeinity assumption is dropped.
Keywords: the utility function, the economic indices theory, the collective axiom of revealed
preference, the weak separability property, the class of the differential form of the demand.
1
Introduction
Globalization of the world economics results in structural changes that seems to fell out of the scope of
the traditional indicators such as, for instance, domestic product, and it seems that new instruments
like statistics panel should be involved in analysis. In this setting some already studied problems of
economics theory are becoming highly relevant. This paper is mainly motivated by the two following
questions.
1. Is it possible to describe structural changes within the framework of the traditional Pareto’s
theory of rational representative consumer (or its generalizations) for the case of several consumers?
2. How to analyze the segmentation of consumer’s market and how to compute statistics panel?
In economics literature this question is often referred to as separability problem.
Recall that Pareto’s theory assumes that the demand can be described within the model of
a rational representative consumer maximizing his utility function under the budget constraints
wherein the utility is postulated to be well-behaved, i. e. it is assumed to be concave, increasing, nonsatiated (nonnegative) maps from Rn+ to R+ continuous up to the boundary. Within the framework of
economic indices it is necessary to make another requirement and postulate that utility is PH. Indeed
(see, e. g. [10]), money balance of the consumer goods in the initial range and in the aggregated
∗
†
e-mail:
e-mail:
[email protected] Supported in part by RFBR grant 17-01-00507
[email protected] Supported in part by RFBR grant 17-07-00300
1
indicators is equivalent to Euler’s identity for the utility function. On the other hand, in economic
literature PH-assumption is criticized both from theoretical and empirical grounds. In theory, the
Engel curves for the PH-utility function should be rays, and this contradicts the intuition on demand
saturation with the growth of income of the household. Empirical studies of individual households
also show that their Engel curves are not rays.
Nevertheless, in some important settings PH can be justified. For instance, in the neoclassical
theory of consumer demand PH-assumption is verified in [11]. Namely, it is reasonable to assume
that PH holds not for an individual household but for the whole group of households with comparable
incomes (in other words, groups with similar living standards). Especially, the size of a particular
group varies as changes of income of an individual household result in modification of its consumer
behavior and possible transitions to different groups. This pattern leads to an essential nonlinearity
of Engel curves. And it turns out that the validity of PH-assumption for the cumulative consumer
behavior strongly depends on the redistribution of prices under changes of incomes of social groups,
and, in particular, whether these changes could be described by the Bergsonian social welfare function.
In this paper for the case of PH-utilities we construct effective algorithms bypassing exaustive
search for some important problems of the revealed preference theory that are NP-hard in the nonhomogeneous setting. In this respect the main message of the paper is that models allowing for
PH utility can be analyzed essentially faster, which may be essential for, say, monitoring or similar
purposes.
The organization of the paper is as folows. In the next section we recall some standard facts
of the revealed preference theory. The third and the fourth sections are devoted to the desciption
and justification of the algorithms for the weak separability and, respectively, for the collective
consumption behavior problems.
For convenience the details of some proofs are moved to the Appendix.
2
The revealed preference theory
Let us recall some standard facts from the revealed preference theory.
The rationality problem consists in the following: given the market statistics for a certain period
of time S = {{pt , q t }, pt = (pt1 , . . . , ptn ) ∈ Rn+ , q t = (q1t , . . . , qnt ) ∈ Rn+ , t = 1, . . . , T } (here pt and q t
are the bundles of prices and the corresponding quantities of goods purchased at time t, respectively),
to find whether the statistics can be rationalized, where rationalization means the existence of a
certain (concave) utility function f : Rn+ → R+ usually called the index of goods, which is consistent
with S. By definition, a utility function f (·) is consistent with market statistics S if
q t ∈ Argmax {f (q) | pt q ≤ pt q t , q ≥ 0},
t = 1, . . . , T.
Hereinafter we assume that utility functions are well-behaved, i. e. they are assumed to be
concave, increasing, non-satiated (nonnegative) maps from Rn+ to R+ continuous up to the boundary.
Besides that we assume that utilities satisfy an additional more restrictive reqiurement of positive
homogeneity (PH)1 . The results of this paper may provide a convincing argument in favour of this
choice, if the model involved allows for it: if we drop homogeneity assumption then all the problems
listed below are hard from the point of the complexity theory as they are known to be at least NP hard and may be even harder, say, belong to the complexity class ∃R-hard. Additionally, numerical
1
This property is also referred to as homotheticity in economics literature, but it seems that our choice is more
suitable, as not only level sets of the PH utilities ought to be homothetic, but, e. g., Euler identity should also hold.
2
results (see, for instance, [9]) show that the usage of PH utility functions are more adequate and
produce essentialy less noise.
Recall that by the homogeneous Afriat’s theorem (see, e.g. [1, 17]) the statistics S is rationalizable in the class of the well-behaved PH utilities iff the following system of linear inequalities is
consistent
λt pt q t ≤ λτ pτ q t , λt > 0, λτ > 0, t, τ = 1, . . . , T.
(1)
PT
Evidently, we can assume that the additional normalization condition holds i=1 λi = 1.
The particular utility function can be recovered from an arbitrary solution of (1) by the formula
f (x) = min λt pt x.
t=1,...,T
(2)
Now we remind two fundamental problems of the revealed preference theory and prove that for
PH ulilities each of these problems can be solved effectively by a polynomial time algorithm. On the
other hand, under the standard setting not assuming PH, these problems are proven to be at least
NP -hard, i. e. intractable for large enough inputs.
3
Weak separability
Assume that goods are partitioned into two (non intersecting) groups of, say, k q-goods and l y-goods.
Then the market statistics bundle can be written as S = {pt , xt ; q t , y t }t=1,...,T , here pt , xt are prices
and q t , y t are the corresponding volumes of purchases and thus {pt ; q t }t=1,...,T is a market statistics
bundle for q-goods and, respectively, {xt ; y t }t=1,...,T is a market statistics bundle for y-goods.
Assume that the overall statistics S is rationalizable by a well-behaved utility function u(q, y),
where q = {q t }, y = {y t}, t = 1, . . . , T .
We say that S is rationalizable by weak separability (in y-goods) if there exist a well-behaved
“micro” subutility function v(y) rationalizing statistics for y-goods and a well-behaved “macro (aggregated)” utility function ũ(q, v(y)), strictly increasing in v-variable, such that u(q, y) = ũ(q, v(y)).
This problem is studied in many publications, see, for instance, [2, 17, 14, 8, 5].
Necessary and sufficient conditions for (non homogeneous) weak separability are proven, e. g., in
[17, Theorem 3]. Furthemore, [17, Theorem 5] gives necessary and sufficient conditions for the case
of the so called “homothetic separability”, where the subutility function v(y) is additionally assumed
to be PH (homothetic). Unfortunately, both characterizations are expressed by a set of quadratic
inequalities and moreover are intractable from the point of view of complexity theory. It is proved
in [5, 6] that testing weak separability is NP -complete.
We prove that if we assume complete PH-separability, i. e. that not only subutility is PH
but macro utility is also PH, then testing weak separability can be reduced to convex program that
can be solved by an effective polynomial time algorithm.
The following necessary and sufficient conditions for complete PH-separability can be obtained
along the same lines as the corresponding conditions for homothetic separability in [17, Theorem 5].
For convenience the details of the proof are moved to the Appendix.
3
λt xt y t ≤ λτ xτ y t ,
λt > 0, λτ > 0
k
l
k
l
X
X
X
X
t t
t t
τ t
µ t λτ (
pi qi +
xi yi ) ≤ µτ (λτ
pi qi + λt
xti yit )
i=1
i=1
i=1
T
X
(3)
(4)
i=1
λi = 1
(5)
i=1
t, τ = 1, . . . , T.
(6)
Let us show how to put the system involved into convex form.
For convenience we use capital letters for all constant parameters except indices and denote the
def
logarithm of some [positive] parameter x by x̃. For instance, αf
λt = log α λt . We rewrite the system
using this notation.
λ̃t + log(X t Y t ) ≤ λ̃τ + log(X τ Y t )
µ̃t + λ̃τ + log(
+ log[exp(λ̃τ )
k
X
i=1
k
X
Pit
Piτ
Qti
+
l
X
(7)
i=1
Qti
+ exp(λ̃t )
i=1
T
X
Xit Yit ) ≤ µ̃τ +
l
X
Xit Yit ]
(8)
i=1
exp(λ̃i ) = 1
(9)
i=1
t, τ = 1, . . . , T.
Now the lhs of (8) is already convex. To fix the rhs we substitute for eachP
λi its expression from
the normalization condition (5). We obtain a function of the form log[const − r exp(Lr )], where Lr
are affine forms of λei and thus the function involved is concave. At last we need to fix the problem
with the normalization equality (5) and this is done by adding a new nonnegative slack variable γ.
Finally, we get the convex program
γ → min
λ̃t + log(X Y ) ≤ λ̃τ + log(X τ Y t )
t
µ̃τ + λ̃t + log(
k
X
i=1
Pit Qti +
l
X
(10)
(11)
t
Xit Yit ) ≤ µ̃τ + log[(1 −
T
X
exp(λ̃i ))
i=1, i6=τ
i=1
(1 −
T
X
i=1, i6=t
4
k
X
Piτ Qti +
(12)
i=1
exp(λ̃i ))
l
X
i=1
Xit Yit ]
(13)
T
X
exp(λ̃i ) + γ ≤ 1, γ ≥ 0
(14)
i=1
t, τ = 1, . . . , T.
And we conclude that the statistics S is weakly separable iff the optimum of the related convex
program is zero. Moreover, we can use modern effective polynomial-time routines to solve this
program. To be a bit more exact by “effective” we mean a procedure which solves the approximate
version of the convex program involved within the accuracy ε in time poly(max(L, log Rε )), where L
is the length of the input and R is the so called localization radius. By definition, if the program
is consistent then at least one solution should fall into the ball of radius R with the center at the
origin.
By construction we have proved the following theorem.
Theorem 1. There is an effective polynomial time algorithm to solve weak separability problem
Remark 1. PH-separability is equivalent to the so called indirect weak separability (IWS) (see,
e. g. [3, 5]) and as a by product IWS for PH utilities can be solved by an effective algorithm.
Let us define the generalized separability problem as a problem of checking whether there
exists some separable partition of the market statistics2 .
Corollary 1. The generalized PH-separability problem belongs to NP .
4
Collective rationality
Afriat’s theorem mentioned above is an effective tool to test whether the market statistics can be
rationalized. But how could violation of rationality be dealt, for instance, in the case of some
essential ongoing structural changes? In view of the uprecedented and unexpected changes of the
modern economics this question seems to acquire the status of a real challenge of crucial importance.
One of the approaches to deal with this problem is the collective revealed preference rationality
model introduced in [4]. It consists in partitioning of the initial (nonrationalizable) market statistics
bundle into a certain number of, say k, rationalizable bundles of the same cardinality, having the
same initial price subbundles but different goods subbundles that componentwise sum to the initial
goods subbundle. In other words, we want to represent the utility as a k-dimensional vector function
def
such that each component is a well-behaved
utility function for the statistics Sα = {pt , α q t }, t =
Pk PH
1, . . . , T , where for all i ∈ {1, . . . , n} α=1 α qit = qit . Using analogy with smooth setting3 we call
minimal such k the discrete class number of the demand form.
By the homogeneous Afriat’s theorem4 we can express collective rationality by the following
2
Evidently, to avoid trivial solutions we should assume that the required partition should be balanced, i. e.,
comparable to the size of the statistics.
3
In [11] this parameter is related to the E. Cartan’s class number of the differential form of the demand. Loosely
speaking, it is equal to the minimal number of the independent variables (functions) needed to express the form.
4
This problem becomes intractable [16] if we drop homogeneity assumption (and use the standard Afriat’s conditions).
5
system.
α
λt
n
X
pti α qit
≤
α
λτ
i=1
n
X
pτi α qit ,
α = 1, . . . k
i=1
k
X
α t
qj
= qjt ,
j = 1, . . . , n,
α=1
t, τ = 1, . . . , T.
As above, we use capital letters for all constant parameters except indices and denote the logarithm of some [positive] parameter x by x̃. After taking logarithms we obtain
αf
λt
+ log[
n
X
i=1
Pit
exp(αf
qit )]
k
X
α t
qj
≤
α
g
λτ
+ log[
= Qtj ,
n
X
Piτ α qit ],
α = 1, . . . k
(15)
i=1
j = 1, . . . , n,
(16)
α=1
t, τ = 1, . . . , T.
Now we can use almost the same trick as before to transform the resulting system into a convex
program.
Note that the lhs of (15) is already convex. To fix the rhs we substitute for each α qit its expression from the relevant balance condition from (16). We obtain a function of the form log[const −
P
αft
r exp(Lr )], where Lr are affine forms of qi and thus the function involved is concave. At last we
need to fix problem with balance equalities (16) and this is done by adding new nonnegative slack
variables γit , t = 1, . . . , T, i = 1, . . . , n. Finally, we get the convex program
T X
n
X
γit → min
t=1 i=1
αf
λt
+ log[
n
X
i=1
n
n
X
X
τ
t
t
ft ) exp(αf
α
g
exp(αf
qjt ))]
exp(P
q
)]
≤
λ
+
log[
P
(Q
−
τ
i
i
i
i
i=1
k
X
α=1
qit ) + γit ≤ Qti ,
exp(αf
γit ≥ 0
t, τ = 1, . . . , T,
j=1, j6=i
α = 1, . . . k,
i = 1, . . . , n.
By construction we proved the following theorem.
Theorem 2. The discrete class number of the demand form is less or equal to k iff the resulting
convex program has optimal solution zero.
As above we can use modern effective polynomial-time routines to solve the convex problem
involved for any k and thus not only can we identify the minimal k but we are able to retrieve some
required utility vector function as well.
6
References
[1] Afriat S. The construction of utility functions from expenditure data. //International Economic
Review. 8. N1. 66–67. 1967.
[2] Blackorby C., Primont D., and Russel R. Duality, Separability, and Functional Structure: Theory
and Economic Application. Amsterdam: North-Holland. 1979.
[3] Brown D.J., Shannon C. Uniqueness, stability, and comparative statics in rationalizable Walrasian markets. Econometrica 68, 1529–1540. 2000.
[4] Cherchye L., De Rock B., and Vermeulen F. The Collective model of household consumption: a
non parametric characterization. Econometrica. Vol. 75. No. 2. 553-574. 2007.
[5] Cherchye, L., Demuynk, T., De Rock, B., Hjertstrand P. Revealed preference tests for weak
separability: an integer programming approach. Journal of Econometrics. 186. 129–141. 2015.
[6] Echenique F. Testing for separability is hard. arXiv:1401.4499 [cs.GT]
[7] Heufer J., Hjertstrand P. Homothetic Efficiency: Theory and Applications, Journal of Business
& Economic Statistic. DOI: 10.1080/07350015.2017.1319372. 2017
[8] Fleissig A., Whitney G.A. A nonparametric test of weak separability and consumer preferences.
Journal of Econometrics. 147, 275281. 2008.
[9] Klemashev N., Shananin A. Inverse problems of demand analysis and their applications to
computation of positively-homogeneous KonüsDivisia indices and forecasting. Journal of Inverse
and Ill-Posed Problems. 2015.
[10] Kondrakov I., Pospelova L., and Shananin A. Generalized nonparametric method. Applications
to the analysis of commodity markets. In Russian. Proc. MIPT 2. 3245. 2010.
[11] Petrov A., Shananin A. Integrability conditions, Income distribution, and social structures.
Lecture notes in Economics and mathematical systems, 453. Springer. 1998.
[12] Shananin A., Tarasov S. Computing the class the differential demand form from discrete data.
Proc. MIPT conference. 2015.
[13] Smeulders, B., Cherchye, L., De Rock, B., Spieksma, F., Talla Nobibon, F. Complexity results for the weak axiom of revealed preference for collective consumption models. Journal of
Mathematical Economics. 58. 82–91. 2015.
[14] Swofford J.L., Whitney, G.A. A revealed preference test for weakly separable utility maximization with incomplete adjustment. Journal of Econometrics. 60, 235249. 1994.
[15] Talla Nobibon, F., Cherchye L., Crama, F, Demuynk, T., De Rock, B., and Spieksma, F.
Revealed preference tests of collectively rational consumption behavior: formulations and algorithms. Operations Research 64, 1197-1216. 2016.
[16] Talla Nobibon, F., Spieksma, F. On the complexity of testing the collective axiom of revealed
preference. Mathematical Social Sciences. 60. No. 2. 123–136. 2010.
7
[17] Varian H.R. Non-parametric tests of consumer behaviour. //The Review of Economic Studies.
1983, 50(1). 99–110. 1983.
5
Appendix
Proposition 1. Formulas (3)–(6) give necessary and sufficient conditions for complete PH-separability.
Proof. At first, recall that Euler’s homogeneous function theorem can be reformulated as follows
for the case of PH well-behaved functions f (·) : Rn+ → R+ . If x0 ∈ Rn+ and p ∈ ∂f (x0 ) then
px0 = f (x0 ).
Necessity. In the notation used above we need to prove the existence of the well-behaved PH
functions u0 (q, z) and u1 (y) such that
(q t , y t) ∈ Argmax[u0 (q, u1(y)) | pt q + xt y ≤ pt q t + xt y t , q ≥ 0, y ≥ 0], t = 1, . . . , T.
(17)
As PH well-behaved functions are monotonic then it follows from (17) that y t ∈ Argmax[u1 (y) |
xt y ≤ xt y t , y ≥ 0], t = 1′ . . . , T.
def
def
pq+sz
xy
Define the Young transforms: ν0 (p, s) =
inf
and ν1 (x) =
inf
.
u0 (q,z)
u1 (x)
{q≥0, z≥0 |, u0 (q,z)>0}
{y≥0 | u1 (x)>0}
By definition it holds
ν1 (xt )u1(y t )
ν1 (xτ )u1(y t )
ν0 (pt , ν1 (xt ))u0 (q t , u1 (y t))
ν0 (pτ , ν1 (xτ ))u0 (q t , u1 (y t))
t, τ
=
≤
=
≤
=
xt y t
xτ y t
pt q t + ν1 (xt )u1(y t )
pτ q t + ν1 (xτ )u1(y t )
1, . . . , T.
1
1
Now necessity follows as λt = ν1 (x
t ) , µt = ν (pt ,ν (xt )) t = 1, . . . , T are solutions of the system
0
1
(3)–(6).
def
t
t
Sufficiency. Set u1 (y) = min{λt xt | t = 1, .. . , T }. By
definition y ∈ Argmax[u1 (y) | x y ≤
def
xt y t , y ≥ 0], t = 1, . . . , T. Set u0 (q, z) = min{µt pt q +
1
z
λt
| t = 1, . . . , T }.
Let q ≥ 0, y ≥ 0, pt q + xt y ≤ pt q t + xt y t . Then it holds
1
1
u1 (y t) =
u0 (q t , u1 (y t )),
λt
µt
1
1
1
u0 (q, u1(y)).
xt y ≥ u1 (y), pt q + xt y ≥ pt q + u1 (y) ≥
λt
λt
µt
pt q t + xt y t = pt q t +
And it follows that u0 (q t , u1 (y t)) ≥ u0 (q, u1(y)) and thus (17) holds.
8
| 8cs.DS
|
arXiv:1704.05134v2 [cs.NE] 19 Apr 2017
Learning Linear Feature Space Transformations
in Symbolic Regression
Jan Žegklitz
Petr Pošı́k
Czech Technical University in Prague
Czech Institute of Informatics, Robotics and Cybernetics
Jugoslávských partyzánů 1580/3
Prague 6, Czech Republic 160 00
[email protected]
Czech Technical University in Prague
Faculty of Electrical Engineering
Technická 1902/2
Prague 6, Czech Republic 166 27
[email protected]
ABSTRACT
We propose a new type of leaf node for use in Symbolic Regression (SR) that performs linear combinations of feature variables
(LCF). These nodes can be handled in three different modes – an
unsynchronized mode, where all LCFs are free to change on their
own, a synchronized mode, where LCFs are sorted into groups in
which they are forced to be identical throughout the whole individual, and a globally synchronized mode, which is similar to the
previous mode but the grouping is done across the whole population. We also present two methods of evolving the weights of
the LCFs – a purely stochastic way via mutation and a gradientbased way based on the backpropagation algorithm known from
neural networks – and also a combination of both. We experimentally evaluate all configurations of LCFs in Multi-Gene Genetic Programming (MGGP), which was chosen as baseline, on a number of
benchmarks. According to the results, we identified two configurations which increase the performance of the algorithm.
KEYWORDS
genetic programming, symbolic regression
1
INTRODUCTION
Symbolic regression (SR) is an inductive learning task with the goal
to find a model in the form of a symbolic mathematical expression
that fits the available training data. SR is a landmark application
of Genetic Programming (GP) [7]. GP is an evolutionary optimization technique that is inspired by biological evolution to evolve
computer programs that perform well in a given task.
Recently, several methods emerged [1, 2, 8, 12] that explicitly
evolve models in a form of (possibly regularized) “top-level” linear
combinations of evolved complex features. Such models can be
learned much faster since the evolution does not have to deal with
the linear parts.
In some SR tasks, the underlying function could be modeled
more easily if we had access to a suitable rotation of the feature
space, or to suitable projections of the features. Such transformations can be achieved by linear combinations of the features of the
problem. Such linear combinations are already available in virtually any SR system that allows for numeric constants. These are
usually tuned by mutation and the linear combinations must be
constructed via structural manipulation operators. In this article
we explore the possibility of using explicit linear combinations of
features at the bottom of the evolved expression trees. These are
added to the original features and can then be non-linarly combined by evolution.
We have chosen Multi-Gene Genetic Programming (MGGP) [4,
12] as the base algorithm for the research as it is very close to
regular GP but uses top-level linear combinations to speed up the
search.
The rest of the article is organized as follows: in Section 2 we
briefly review MGGP as it is the foundation of our work. Section
3 we describe the linear combinations of features and how do they
work. Section 4 is dedicated to the description of the experimental evaluation of our proposals. Section 5 presents the results and
discusses the implications. Section 6 concludes the paper and provides suggestions for future work.
2 MULTI-GENE GENETIC PROGRAMMING
MGGP [4, 12, 13] is a tree-based genetic programming algorithm
utilizing multiple linear regression. The main idea behind MGGP is
that each individual is composed of multiple independent expression trees, called genes, which are put together by a linear combination to form a single final expression. The parameters of this
top-level linear combination are computed using multiple linear
regression where each gene acts as an independent feature.
In this article we base upon a particular implementation of MGGP
– GPTIPS2 [11]1 . This particular instance of MGGP uses two crossover
operators: (i) high-level crossover that, according to a probability which is a parameter of the algorithm, selects genes from two
parents and swaps them between those parents, throwing off the
genes that would exceed the upper limit on the number of genes;
(ii) low-level crossover which is a classical Koza-style subtree crossover.
Also, there are two mutation operators: (i) subtree mutation
which is a classical Koza-style subtree mutation; (ii) constant mutation which mutates the numerical values of leaves representing
constants by adding a normally distributed random number.
Both the crossover and mutation operators are chosen stochastically.
3 LINEAR COMBINATIONS OF FEATURES
We introduce a new type of leaf node – a Linear Combination of
Features or LCF for short. This type of node is similar to a leaf node
representing a variable, or feature. However, while an ordinary
feature-node evaluates simply to the value of that feature, a LCF
node evaluates to a linear combination of all the features present
1
We didn’t use that implementation but re-implemented it.
in the solved problem. Mathematically, the node implements the
following function
lc f (x) = a + bx
The second reason is especially important one because in the GP
environment there are (generally) no constraints on the inner structure of the expressions. Due to the finite precision of binary representation of real numbers in computers, this can lead to infinite
derivatives even if, mathematically, they are finite.
(1)
where x is the vector of feature values and a and the vector b are
constants.
The LCFs effectively perform affine transformations of the feature space and we argue that they can provide more effective tools
to deal with e.g. rotated functions and, in general, provide more
flexibility to the GP algorithm.
3.3 Operation modes of LCFs
When an LCF node is generated (during initialization or structural
mutation) an index is assigned to it. This index can be any integer
between (and including) 1 and the number of the features of the
solved problem. We define three operation modes that differ in
how the weights of LCFs are handled in relation to their index.
Terminology. In the rest of the paper we will refer to the additive
constant a and multiplicative coefficients b simply as to weights.
3.1
3.3.1 Unsynchronized mode. In the unsynchronized mode there
are no restrictions whatsoever on the weights. In other words, any
LCF is allowed to have any weights, regardless of its index.
Initialization of LCFs
LCF nodes have two or more weights (depending on the dimensionality of the problem) that have to be somehow initialized at
the start of the algorithm.
The initialization method is based on the idea that, at the start,
there is no feature space transformation happening. This means
that each LCF is initialized such that the additive weight and all
multiplicative weights except for one are set to zero. The only nonzero multiplicative weight is set to one2 .
3.2
3.3.2 Synchronized mode. In this mode, all LCFs with the same
index in a particular individual are forced to have the same weights.
This way, all the LCFs in a model form a single affine transformation of the feature space, effectively producing a simpler model.
From a technical point of view, all LCFs are still treated independently. In order to get the desired behaviour, a special handling is
required:
• After each backpropagation phase (if tuned by gradientbased approach) the values of partial derivatives for the
parameters are summed up inside the index-determined
groups. This way, all such nodes will be updated in the
same way.
• If it is detected that a model has, for some of the indexes,
two sets of weights (a result of a structural mutation or a
crossover with another model), both sets and their arithmetic mean are evaluated and the best performing setting
is used.
Tuning of LCF weights
In order to be of any use, the weights of LCF nodes must be modified during the evolution. We present two such methods.
The first approach, probably the simplest possible, is weights mutation, i.e. a dedicated mutation operator. We use an approach similar to one used for tuning constants in leaf nodes – a gaussian
mutation. When weights mutation occurs, a single random LCF
node is selected and each weight is offset by a random number
from gaussian distribution.
The second approach is gradient-based tuning. Weights mutation is simple but it is not informed as it relies just on the selection
strategy to promote good mutations. However, since the structure
of the expressions as well as the cost function is known, it is possible to compute the gradient of the expressions w.r.t. the weights.
We use an approach fundamentally identical to the one used in
neural networks – error backpropagation technique3 [10]. When
the individual partial derivatives are known, any first-order update
method can be used to modify the weights to produce more fit expression.
In our setup we use the iRprop− update mechanism [5]. There
are two reasons for using this particular method:
3.3.3 Globally synchronized mode. The globally synchronized mode
is similar to the synchronized mode but the index-based synchronization encompasses the whole population instead of single models. The motivation behind this mode is that should there truly be
a globally suitable transformation of the input space in the data,
the models can all work together to find this transformation.
Since there is only a single set of LCFs, using only mutation
to tune them makes no sense because there is no population of
LCF sets the selection could pick from and hence mutation would
be only a random walk. Therefore we always use gradient-based
approach, alone or accompanied by the mutation as a means to
help it escape from local optima.
(1) it is very simple to implement yet is very efficient [6]4 , and
(2) is numerically robust as it operates only with the signs of
the partial derivatives rather than their magnitudes.
4 EXPERIMENTAL EVALUATION
In this section we describe the experimental setup for the evaluation of our proposal.
2
We also tried random initialization – all weights sampled randomly – but preliminary experiments showed no benefit over the described init. method.
3 We strictly separate the task of determining the values of the partial derivatives (i.e.
the gradient) w.r.t. the parameters and the actual update of the parameters. When we
use the term “(error) backpropagation”, we mean only the procedure of determining
the values of the partial derivatives and not the update mechanism.
4 Although [6] shows that iRprop+ is superior to iRprop− , our preliminary experiments have shown that using iRprop+ causes more overfitting and not as good results,
hence we decided to use iRprop− .
4.1 Algorithm configurations
We proposed several ways of how LCFs can work in two aspects
of the algorithm: operation mode, and method of tuning the LCF
weights. In the following text we shall use “codenames” for each
of the algorithm configurations.
2
Each codename is composed of two letters. The first letter describes the LCF operation mode and second letter describes how
are the LCF weights tuned. The code letters are described in Table
1. The baseline, i.e. unmodified MGGP, has none of the two letters
and will be referred to as baseline, or (in tables) by two dashes, one
for each missing letter.
Toy problems. S2D, S5D, S10D, RS2D, RS5D, RS10D are the
toy problems in 2, 5 and 10 dimensions. The S* problems are a
simple sigmoid function applied to the first variable, independent
on the others. In the RS* problems, the sigmoid is rotated by π4 in
all pairs of axes, i.e. all variables are important. The problems are
uniformly randomly sampled from the interval [−10, 10]D . There
are 100 · D samples in the training set and 250 · D samples in the
testing set.
The goal of these benchmarks is to support our expectation that
the LCFs can provide means to find a suitable linear transformation
of the feature space.
Table 1: Description of codename letters.
aspect
code letter
description
operation mode
U
S
G
unsynchronized
synchronized
globally synchronized
weights tuning
M
B
C
mutation
backpropagation
all, i.e. both M and B
Realistic problems. K11C is similar to Keijzer11 in [9] but with
added numerical coefficients throughout the formula6 . The training set is 500 uniform samples from [−3, 3]2 , the testing set is a
grid in the same range with a spacing of 0.01 in each dimension
(361201 samples).
UB5D (Unwrapped Ball 5D) [15] is a 5D artificial benchmark.
The true relationship is f (x) = ÍN 10
2 where N = 5. The
5+
As a baseline, pure MGGP is used (see Section 2). Our proposed modifications then build upon the baseline by introducing
the LCFs5 and handling them in the way described by the codename.
Example. A configuration with codename UC is a configuration
where LCFs are unsynchronized and both backpropagation and
mutation are used for tuning the LCF weights. In one more example, configuration SM is such that LCFs are synchronized and
the weights are tuned using only mutation.
4.2
Parameters
The new features are, naturally, accompanied by a number of parameters that configure how exactly do those features behave. Table 2 presents all of the parameters, including the ones that come
from the baseline algorithm, as well as their values. The values are
based on the default values provided by the GPTIPS2 [11] package.
The values of parameters related to the LCFs were chosen by hand
based on preliminary examination of the configurations. Except
for configurations using backpropagation, we kept the MGGP related parameters for all other configurations too. For the case of
backpropagation, we decided to half the size of population because
more time is spent with each individual by tuning the LCF parameters so this enables the algorithm to do more structural exploration.
Function set. The functions available to the algorithm were: +,
1 , tanh, sin x , ln(1 + ex ), e−x 2 and 2nd to 6th
−, ×, sin, cos, exp, 1+e
x
x
powers.
4.3
i =1 (x i −3)
training set is 1024 uniform samples from [−0.25, 6.35]5 , the testing set is 5000 uniform samples from the same range.
ASN (Airforil Self-Noise), acquired from the UCI repository [3],
is a 5D dataset regarding the sound pressure levels of airfoils based
on measurements from a wind tunnel. Training/testing set comes
from a random 0.7/0.3 split (1503 samples in total).
CCS (Concrete Compressive Strength) [16], acquired from the
UCI repository [3], is an 8D dataset representing a highly nonlinear function of concrete age and ingredients. Training/testing
set comes from a random 0.7/0.3 split (1030 samples in total).
ENC and ENH (Energy Efficiency) [14], acquired from the UCI
repository [3], are 8D datasets regarding the energy efficiency of
cooling and heating of buildings. Training/testing set comes from
a random 0.7/0.3 split (768 samples in total per dataset).
SU7 and SU-I7 (swingup) are two datasets from a reinforcement learning domain. They are value functions of an inverted
pendulum swing-up problem computed by a numeric approximator. They have 2 dimensions (pendulum angle and angular velocity)
and the value is the value of the state w.r.t. the goal state which,
for the SU variant, is located at [−π, 0] and equivalently [π, 0] (due
to the circular nature of the problem). The SU-I variant represents
identical function but the angle coordinate is shifted by π2 . Training/testing set comes from a random 0.7/0.3 split (441 samples in
total).
MM7 (2-coil magnetic manipulation) is a dataset from a reinforcement learning domain. It is a value function of a linear magnetic manipulation problem with 2 coils computed by a numeric
approximator. It has 2 dimensions (the manipulated ball’s position
and velocity) and the value is the value of the state w.r.t. the goal
state. Training/testing set comes from a random 0.7/0.3 split (729
samples in total).
Testing benchmarks and environment
We have designed two sets of tests. First we test the LCF concept
on a simple toy problem. Second we test the overall performance
with 9 full benchmarks, both real-world and artificial ones.
Testing methodology. For each benchmark, each algorithm configuration was run 30 times, each time with a different seed and
different sampling (for the artificial benchmarks) or different training/testing split (for the real-world benchmarks). Each run has
5
7
6
However, the original, unmodified variables are still available to the algorithm.
3
f (x) = (27.22x 1 − 4.54)(−0.39x 2 ) + 11.46 sin ((0.21x 1 − 1) (x 2 + 16.6) + 1.97)
These datasets are depicted in the supplementary material.
Table 2: All parameters of the algorithm and the values we used for them. Numbers in parentheses for |P |, |T | and E are used
for high-level configurations with backpropagation. For the purposes of computing Pr Sm , PrW m is considered zero when no
weights mutation is present in the algorithm.
LCF params
MGGP params
parameter
description
value
Gmax
Nmax
Dmax
|P |
|T |
E
Pr x
Prm
Pr LLx
Pr H Lx
r H Lx
Pr Sm
PrCm
σCm
maximum number of genes
maximum number of nodes per gene
maximum depth of a gene
number of individuals in the population
the size of the tournament
number of top individuals copied to the next generation without any modification (elitism)
probability of a crossover event
probability of a mutation event
probability of a crossover being the low-level one
probability of a crossover being the high-level one
probability that a gene will be selected in high-level crossover
probability of a mutation being a subtree mutation
probability of a mutation being a mutation of constant leaf node(s)
variance of the gaussian distribution used in constant leaf node mutation
PrW m
σW m
Bpsteps
probability of a mutation being a mutation of weights in an LCF
variance of the gaussian distribution used in weights mutation
the number of backprop.-update steps as Bpsteps minus the number of nodes in all genes combined
minimum number of backprop.-update steps each generation per individual (overrides Bpsteps )
Bpmin
2
Discussion. The toy problems showed that the basic idea is supported – unrotated problems were easy for all configurations while
the rotated ones were easy for the configurations using LCFs and
not for baseline. We can see that the configurations that were statistically better than the baseline show high LCF ratio, i.e. they
really use the LCFs.
However, we can see that UM and SM, i.e. configurations with
LCF weights tuned only by mutation, were (statistically) neither
better, nor worse than the baseline (though the maximum achieved
values are better). We can also see that these configurations have
much smaller LCF ratio than the other LCF-enabled configurations.
This can be explained as the lack of the ability to tune the LCF
weights accurately enough to be beneficial for the expressions. This
can be seen as an indication that mutation is not very good approach to search for the linear combinations.
Note that the LCFs are also used in the non-rotated problem although they are of no benefit there. This is due to two facts: (i)
there is no penalty for using them, so they can be used similarly
Testing environment. Everything is written in Python 3 and uses
the NumPy library for vector and matrix calculations including the
linear regression. All experiments were carried out on the National
Grid Infrastructure MetaCentrum (see Acknowledgements) which
is a heterogeneous computational grid. We ensured that for each
dataset all runs of all configurations on that dataset were carried
out on machines of the same cluster which all have the same configuration.
RESULTS AND DISCUSSION
In this section we present the results of experiments we described
in the previous section.
5.1
0.05
3
25
worse than (denoted by ✗) or indifferent to (denoted by blank space)
the baseline. This result was established using Mann-Whitney ranksum
test on the testing R 2 values with the significance level α = 0.05
with Bonferonni correction for 11 comparisons8 resulting in an effective α ≈ 0.0045.
The column denoted “LCF” shows mean fraction of non-constant
leaf nodes that are LCFs9 .
a wall-clock time limit of 7 minutes and is terminated after this
amount of time passes (except up to a negligible amount of time
that passes between the time checks). Fitness is R 2 on the training
dataset and is maximized. After the algorithm finishes, the resulting model (the one with the best fitness found over the whole runtime of the algorithm) is evaluated on the testing set. In the next
section we report the results both on the training and testing sets
so that some judgement on overfitting can be made.
5
10
∞
11
100 (50)
10 (5)
15 (8)
0.84
0.14
1 − Pr H Lx
0.2
0.5
1 − PrCm − PrW m
0.05
0.1
Toy problems
Here we present the results on the toy problems S5D and RS5D.
The results can be seen in Tables 3 and 4. Results on (R)S2D and
(R)S10D are not presented because 2D case was too easy for all
algorithms, and the results on 10D case are very similar to the presented 5D case.
The column denoted “vb”, which stands for “versus baseline”,
shows whether the configuration was better than (denoted by ✓),
8 We present only 8 here because the randomized initialization (see footnote 2)
we dismissed because of no impact but three such configurations were part of this
comparison.
9 For example, expression containing 3 “pure” variables and 7 LCFs would have this
value equal to 0.7.
4
Table 3: Results on the S5D toy problem.
testing R2
mode
tuning
training R2
median
max
min
median
max
min
- -
1
1
UM
1
UB
1
UC
1
SM
1
SB
1
SC
1
GB
1
GC
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
vb
baseline. This comparison was established using identical test as
the “versus baseline” test in previous subsection.
mean
mean
LCF
depth
0
4.33
0.482
4.1
0.576
3.5
0.521
3.77
0.469
4.13
0.475
4.2
0.48
3.73
0.38
4.23
0.352
4.87
Table 5: Summary results for each algorithm configuration.
max
min
- -
0.995
UM
0.993
UB
1
0.997
0.942
1
0.93
1
1
1
1
1
0.158
1
1
1
1
1
0.855
1
0.872
UC
1
SM
0.995
SB
1
SC
1
GB
1
GC
0.974
median
max
min
0.991
0.996
0.912
1
0.907
1
1
1
1
0.999
−0.168
1
1
1
1
1
0.659
1
0.736
0.99
1
1
0.993
1
1
1
0.962
mean
mean
LCF
depth
0
10.9
0.524
10.6
✓
0.98
4.6
✓
0.974
4.63
0.579
10.9
✓
0.9
8.1
✓
0.954
7.77
✓
0.817
8.53
✗
0.651
6.7
vb
U
C
S
M
S
B
S
C
G
B
G
C
better than baseline
indifferent to baseline
worse than baseline
1
8
5
3
1
5
3
1
1
8
3
4
2
3
4
2
6
3
3
6
Negative R 2 . We can see that in some cases, some configurations
achieved a negative R 2 on the testing set11 . In all of those cases, it
happened in only a single run of the 30 runs for that configuration
and dataset.
Table 6: Performance on the K11C dataset.
to ordinary variables, and (ii) the genes containing the LCFs can
receive a negligible coefficient in the top-level linear combination
in case the true relationship is among the genes (that one will receive a coeffcient of 1) so they could be considered as effectively
not present.
Also interesting is the depth usage. On the rotated problem, the
baseline uses almost all the available depth while UB, for example,
uses less than half the depth10 . This is a clear indication of the
ability of LCFs to provide good data to the rest of the expressions.
5.2
U
B
training R2
mode
tuning
mode
tuning
median
testing R2
U
M
We will now present more detailed results of the configurations
for each dataset. We will list only results for configurations UB, UC,
SB and SC as these had the most positive results in Table 5. Full
results are available in the supplementary material to this article.
The results will be presented in the form of tables (see Tables 6
through 14) and box plots (see Figures 1 through 9). The tables are
identical to tables for toy problems, with exactly the same meaning
of columns, except we don’t present the mean depth here.
Table 4: Results on the RS5D toy problem.
training R2
mode
tuning
median
max
min
--
0.981
UB
0.998
UC
0.998
SB
0.991
SC
0.992
0.997
0.971
0.999
0.99
1
0.992
0.998
0.954
0.998
0.954
testing R2
mean
median
max
min
0.976
0.995
0.965
0.999
0.978
1
−3.24e +29
0.998
0.945
0.997
0.948
0.996
0.997
0.989
0.99
vb
LCF
0
✓
0.873
✓
0.874
✓
0.603
✓
0.622
1.00
0.99
0.98
R2
0.97
0.96
0.95
0.94
Realistic problems
baseline
UB
UC
SB
SC
Figure 1: Training (left) and testing (right) R 2 performance
on the K11C dataset. Note that for the testing UC plot, one
outlier at -3.24e29 is not shown (as well as the mean marker).
First we present a summary result of which configurations were
better than the baseline across individual datasets. The result can
be seen in Table 5. The table shows the number of datasets where
each configuration was better than, tied with, or worse than the
10
11
Negative R 2 means that the fit is worse than that of the constant model equal to
the mean of the target data.
Additionally, similar effect as in the previous paragraph can apply – some deeper
genes can have a small top-level linear coefficient than other gene(s).
5
Table 7: Performance on the UB5D dataset.
testing R2
median
max
min
median
max
min
--
0.885
0.866
UB
0.857
UC
0.858
SB
0.839
SC
0.839
0.976
0.808
0.887
0.828
0.932
0.824
0.972
0.802
0.93
0.816
0.968
0.796
0.856
0.58
0.892
0.807
0.967
0.796
0.908
0.795
0.828
0.826
0.816
0.818
training R2
max
min
median
max
min
--
0.869
0.844
0.823
UB
0.901
✗
0.802
UC
0.899
✗
0.553
SB
0.889
✗
0.601
SC
0.893
0.89
0.848
0.924
0.869
0.931
0.854
0.906
0.868
0.908
0.857
0.868
−8.68e +07
0.892
0.806
0.88
0.758
0.898
−4.74e +04
0.873
−291
LCF
0
✗
1.00
0.9
0.95
0.859
0.858
0.851
0.846
vb
LCF
0
✓
0.87
✓
0.885
0.676
0.707
0.90
R2
R2
0.8
0.85
0.7
0.80
baseline
UB
UC
SB
0.75
SC
Table 8: Performance on the ASN dataset.
testing R2
median
max
min
--
0.842
UB
0.849
UC
0.841
SB
0.804
SC
0.8
0.892
0.72
0.914
0.729
0.894
0.705
0.842
0.675
0.867
0.71
median
max
min
0.824
0.885
0.625
0.893
−0.719
0.88
0.623
0.829
0.624
0.861
0.653
0.818
0.818
0.77
0.76
vb
LCF
0
0.834
0.828
✗
0.651
✗
0.68
UB
UC
SB
SC
Table 10: Performance on the ENC dataset.
mean
mode
tuning
training R2
baseline
Figure 4: Training (left) and testing (right) R 2 performance
on the CCS dataset. Note that for baseline, SB and SC, testing
outliers (one configuration each) at -8.68e7, -4.74e4 and -291
respectively are not shown.
Figure 2: Training (left) and testing (right) R 2 performance
on the UB5D dataset.
mode
tuning
mean
median
vb
1.0
0.6
testing R2
mode
tuning
mean
mode
tuning
training R2
Table 9: Performance on the CCS dataset.
1.0
0.9
training R2
testing R2
median
max
min
--
0.974
UB
0.974
UC
0.975
SB
0.974
SC
0.973
0.981
0.969
0.988
0.97
0.986
0.972
0.979
0.969
0.98
0.969
mean
median
max
min
0.97
0.968
0.982
0.963
0.982
0.957
0.985
0.961
0.973
0.965
0.976
0.962
UC
SB
0.969
0.971
0.968
vb
LCF
0
0.751
0.772
0.609
0.609
1.00
R2
0.8
0.99
0.7
R2
0.98
0.6
baseline
UB
UC
SB
0.97
SC
0.96
Figure 3: Training (left) and testing (right) R 2 performance
on the ASN dataset. Note that for the testing UB plot, one
outlier at -0.719 is not shown.
baseline
UB
SC
Figure 5: Training (left) and testing (right) R 2 performance
on the ENC dataset.
Discussion. In the summary results (see Table 5) we can see that
the configurations using globally synchronized mode were either
worse or of similar performance as the baseline. We hypothesize
that this is caused by the fact that the shared linear combinations
are modified based on all individuals in the population of which
some can have a good structure but equally some can have a bad
structure. In the end, good and bad individuals fight against each
other which, in turn, makes the usefulness of the shared combinations doubtful.
We can also see that the (locally) synchronized mode does not
perform as well as the unsynchronized mode. We hypothesize
6
Table 11: Performance on the ENH dataset.
testing R2
median
max
min
median
max
min
--
0.998
0.997
UB
0.997
UC
0.998
SB
0.997
SC
0.997
0.998
0.996
0.998
0.993
0.998
0.995
0.998
0.993
0.998
0.99
0.998
0.995
0.998
0.991
0.998
0.994
0.998
0.993
0.998
0.988
0.997
0.997
0.997
0.997
training R2
testing R2
mean
mode
tuning
mean
mode
tuning
training R2
Table 13: Performance on the SU-I dataset.
median
max
min
median
max
min
--
0.931
0.885
0.73
UB
0.97
0.732
UC
0.976
0.592
SB
0.942
0.61
SC
0.952
0.979
0.841
0.993
0.938
0.991
0.915
0.988
0.884
0.989
0.836
0.931
0.97
0.175
0.987
0.886
0.988
0.865
0.992
0.769
0.99
0.788
UB
UC
SB
vb
LCF
0
1.000
1.0
0.998
0.9
0.996
0.955
0.962
0.928
vb
LCF
0
✓
0.895
✓
0.912
0.569
0.623
✓
0.8
R2
R2
0.994
0.7
0.992
0.990
0.6
baseline
UB
UC
SB
SC
baseline
Figure 6: Training (left) and testing (right) R 2 performance
on the ENH dataset.
SC
Figure 8: Training (left) and testing (right) R 2 performance
on the SU-I dataset. Note that for baseline, one testing outlier at 0.175 is not shown.
Table 12: Performance on the SU dataset.
Table 14: Performance on the MM dataset.
median
--
0.955
UB
0.985
UC
0.985
SB
0.977
SC
0.968
median
0.988
0.879
0.994
0.963
0.996
0.93
0.991
0.881
0.993
0.885
0.909
0.971
0.966
0.955
0.958
mean
max
min
0.978
−0.664
0.994
0.881
0.992
0.916
0.984
0.819
0.978
0.694
vb
LCF
0
training R2
median
0.987
0.954
0.997
0.973
0.996
0.969
0.991
0.967
0.997
0.947
0.894
--
0.966
✓
0.885
UB
0.988
✓
0.598
UC
0.988
0.633
testing R2
max
min
✓
SB
0.976
SC
0.974
1.00
0.99
0.98
0.97
0.96
0.95
0.94
0.93
0.92
mean
median
max
min
0.96
0.971
0.983
0.93
0.995
0.969
0.995
0.943
0.986
0.961
0.996
0.935
UC
SB
0.985
0.985
0.973
vb
LCF
0
✓
0.763
✓
0.797
✓
0.559
✓
0.563
R2
R2
1.00
0.95
0.90
0.85
0.80
0.75
0.70
0.65
testing
max
min
R2
mode
tuning
mode
tuning
training
R2
baseline
UB
UC
SB
SC
Figure 7: Training (left) and testing (right) R 2 performance
on the SU dataset. Note that for baseline, one testing outlier
at -0.664 is not shown.
baseline
UB
SC
Figure 9: Training (left) and testing (right) R 2 performance
on the MM dataset. Note that for baseline, one testing outlier
at 0.175 is not shown.
that this is caused by the need to resolve conflicts when structural
changes (i.e. mutation or crossover) happen (see Section 3.3.2).
However, after the synchronization the model can be much worse
than its parent(s) before the crossover (or mutation). Also, this
synchronization takes additional time and slows the process down
(though not very much).
Another result clearly visible from the summary results is that
modifying the LCFs only with mutation makes almost no difference to the baseline. This supports similar results on the toy problems. On the other hand, every configuration using backpropagation (except GB which we already discussed) is better than the
7
The presented work is just a first glance at the possibilities. The
proposed approach was tested only on problems of low dimensionality. A proper testing on high-dimensional problems is necessary.
Future research could also focus on tuning the parameters of the
algorithm to find e.g. how much backpropagation tuning is suitable, or which update method fits best into this setting. Another
idea worth probing is extending this concept to other nodes than
leaves, bringing similar tuning capabilities inside the trees.
baseline more than once. We see this as a clear indication that using the backpropagation technique to tune the LCF weights is a
viable approach.
On the K11C dataset (see Table 6 and Figure 1) we can see quite
a strong result in favor of UB and UC configurations. Again, this
result confirms the benefit of the LCFs since the true relationship itself contains several constants that multiply or offset the variables
– precisely what the LCFs are capable of.
On the UB5D dataset (see Table 7 and Figure 2), baseline was
by far the best. We hypothesize that this is caused by the fact that
the true relationship is one big fraction but there is no division
operator in the function set. Because of this, the LCF-based configurations are, in principle, not able to find a structure that would
enable them to find a good LCFs. Therefore they are wasting time,
compared to the baseline which does not attempt to do this and
aims for approximating the relationship by “brute force”. Similar
result, though not as strong, can be observed on the ASN dataset
(see Table 8 and Figure 3). Since it is a real-world dataset, we don’t
know the true relationship and we cannot draw the same conclusions as for UB5D.
Worth noting is the fact that the selected presented configurations, especially UB and UC, performed very well on the datasets
from reinforcement learning (RL) domain. The first possible explanation could be that these configurations are good on RL value
functions. However, there are many possible RL problems and
their value functions so such generalization cannot be made. Second possible explanation is that a common feature of these functions – they are all very smooth, without very sharp peaks, oscillations and noise – is well suited for LCFs. This is more viable explanation since the toy problems and the K11C dataset also share
this feature, but it is still only a hypothesis. The last possible explanation is that these datasets were easy because they are just twodimensional. The low dimensionality certainly plays an important
role, though it is difficult to asses the measure of importance. However, even if it was so, it would be an indication that LCFs are very
good for problems of low dimensionality.
6
ACKNOWLEDGMENTS
This work was supported by the Czech Science Foundation project
Nr. 15-22731S with a student support from the Grant Agency of
the Czech Technical University in Prague, grant No. SGS17/093/
OHK3/1T/13. Access to computing and storage facilities owned
by parties and projects contributing to the National Grid Infrastructure MetaCentrum provided under the programme ”Projects
of Large Research, Development, and Innovations Infrastructures”
(CESNET LM2015042), is greatly appreciated.
REFERENCES
[1] Ignacio Arnaldo, Krzysztof Krawiec, and Una-May O’Reilly. 2014. Multiple Regression Genetic Programming. In Proceedings of the 2014 Annual Conference on
Genetic and Evolutionary Computation (GECCO ’14). ACM, New York, NY, USA,
879–886. DOI:http://dx.doi.org/10.1145/2576768.2598291
[2] Ignacio Arnaldo, Una-May O’Reilly, and Kalyan Veeramachaneni. 2015. Building Predictive Models via Feature Synthesis. In Proceedings of the 2015 Annual
Conference on Genetic and Evolutionary Computation (GECCO ’15). ACM, New
York, NY, USA, 983–990. DOI:http://dx.doi.org/10.1145/2739480.2754693
[3] K. Bache and M. Lichman. 2013. UCI Machine Learning Repository. (2013).
http://archive.ics.uci.edu/ml.
[4] Mark Hinchliffe, Hugo Hiden, Ben McKay, Mark Willis, Ming Tham, and Geoffery Barton. 1996. Modelling Chemical Process Systems Using a Multi-Gene
Genetic Programming Algorithm. In Late Breaking Paper, GP’96. Stanford, USA,
56–65.
[5] Christian Igel and Michael Hüsken. 2000. Improving the Rprop learning algorithm. In Proceedings of the second international ICSC symposium on neural computation (NC 2000), Vol. 2000. Citeseer, 115–121.
[6] Christian Igel and Michael Hsken. 2003. Empirical evaluation of the improved Rprop learning algorithms. Neurocomputing 50 (2003), 105–123. DOI:
http://dx.doi.org/10.1016/S0925-2312(01)00700-7
[7] John R. Koza. 1992. Genetic Programming: On the Programming of Computers by Means of Natural Selection. MIT Press, Cambridge, MA, USA.
http://mitpress.mit.edu/books/genetic-programming
[8] Trent McConaghy. 2011. FFX: Fast, Scalable, Deterministic Symbolic Regression
Technology. In Genetic Programming Theory and Practice IX, Rick Riolo, Ekaterina Vladislavleva, and Jason H. Moore (Eds.). Springer New York, 235–260. DOI:
http://dx.doi.org/10.1007/978-1-4614-1770-5 13
[9] James McDermott, David R. White, Sean Luke, Luca Manzoni, Mauro Castelli,
Leonardo Vanneschi, Wojciech Jaskowski, Krzysztof Krawiec, Robin Harper,
Kenneth De Jong, and Una-May O’Reilly. 2012. Genetic Programming Needs
Better Benchmarks. In Proceedings of the 14th Annual Conference on Genetic and
Evolutionary Computation (GECCO ’12). ACM, New York, NY, USA, 791–798.
DOI:http://dx.doi.org/10.1145/2330163.2330273
[10] David E. Rumelhart, Geoffrey E. Hinton, and Ronald J. Williams. 1986. Learning
representations by back-propagating errors. Nature 323, 6088 (09 Oct 1986),
533–536. DOI:http://dx.doi.org/10.1038/323533a0
[11] Dominic P.
Searson.
2015.
GPTIPS
2.
(2015).
http://sites.google.com/site/gptips4matlab
[12] Dominic P. Searson. 2015. GPTIPS 2: An Open-Source Software Platform for Symbolic Data Mining. Springer International Publishing, Cham, 551–573. DOI:
http://dx.doi.org/10.1007/978-3-319-20883-1 22
[13] Dominic P Searson, David E Leahy, and Mark J Willis. 2010. GPTIPS: an open
source genetic programming toolbox for multigene symbolic regression. In Proceedings of the International MultiConference of Engineers and Computer Scientists, Vol. 1. 77–80.
[14] Athanasios Tsanas and Angeliki Xifara. 2012. Accurate quantitative estimation of energy performance of residential buildings using statistical machine learning tools.
Energy and Buildings 49 (2012), 560–567. DOI:
http://dx.doi.org/10.1016/j.enbuild.2012.03.003
CONCLUSIONS AND FUTURE WORK
In this article we presented a new type of leaf node for use in SR
– linear combination of feature variables – which we then used
in the baseline algorithm of MGGP. We have presented two approaches to tuning their weights, one based on mutation and the
other one based on error backpropagation technique. We also presented three operation modes, one very flexible with no constraints,
one enforcing creation of affine transformations of the feature space
and one enforcing such a transformation on the whole population.
All sensible configurations of the proposed algorithm were tested
on a set of benchmarks with the focus on showing the differences
from the baseline algorithm. The toy problems, designed specifically to test the ability of LCFs, showed that they are capable of
handling existing feature space transformation. The results on realistic problems have shown that configurations using globally synchronized mode are of no benefit or make the algorithm worse,
and we provided a possible explanation. On the other hand, two
configurations stood out as clear improvement over the baseline.
8
[15] E.J. Vladislavleva, G.F. Smits, and D. den Hertog. 2009. Order of Nonlinearity as a
Complexity Measure for Models Generated by Symbolic Regression via Pareto
Genetic Programming. Evolutionary Computation, IEEE Transactions on 13, 2
(April 2009), 333–349. DOI:http://dx.doi.org/10.1109/TEVC.2008.926486
[16] I.-C. Yeh. 1998. Modeling of strength of high-performance concrete using artificial neural networks. Cement and Concrete Research 28, 12 (1998), 1797–1808.
DOI:http://dx.doi.org/10.1016/S0008-8846(98)00165-3
9
| 9cs.NE
|
Prioritized Sweeping Neural DynaQ with
Multiple Predecessors, and Hippocampal
Replays
arXiv:1802.05594v1 [cs.AI] 15 Feb 2018
Lise Aubin, Mehdi Khamassi, and Benoı̂t Girard
Sorbonne Université, CNRS, Institut des Systèmes Intelligents
et de Robotique (ISIR), F-75005 Paris, France
[email protected]
Abstract. During sleep and awake rest, the hippocampus replays sequences of place cells that have been activated during prior experiences.
These have been interpreted as a memory consolidation process, but recent results suggest a possible interpretation in terms of reinforcement
learning. The Dyna reinforcement learning algorithms use off-line replays
to improve learning. Under limited replay budget, a prioritized sweeping
approach, which requires a model of the transitions to the predecessors,
can be used to improve performance. We investigate whether such algorithms can explain the experimentally observed replays. We propose
a neural network version of prioritized sweeping Q-learning, for which
we developed a growing multiple expert algorithm, able to cope with
multiple predecessors. The resulting architecture is able to improve the
learning of simulated agents confronted to a navigation task. We predict that, in animals, learning the world model should occur during rest
periods, and that the corresponding replays should be shuffled.
Keywords: Reinforcement Learning, Replays, DynaQ, Prioritized Sweeping, Neural Networks, Hippocampus, Navigation
1
Introduction
The hippocampus hosts a population of cells responsive for the current position
of the animal within the environment, the place cells (PCs), a key component of
the brain navigation system [1]. Since the seminal work of [2], it has been shown
that PCs are reactivated during sleep – obviously without any locomotion –
and that these reactivations are functionally linked with the improvement of the
learning performance of a navigation task [3]. Similar reactivations have been
observed in the awake state [4], while the animal is immobile, either consuming
food at a reward site, waiting at the departure site for the beginning of the
next trial or stopped at a decision point. These reactivations contain sequences
of PCs’ activations experienced in the awake state (forward reactivations) [5],
sequences played in the reverse order (backward reactivations) [4], and sometimes
never experienced sequences (resulting from the concatenation of experienced
sequences) [6]. These reactivations have been interpreted in the light of the
act
1
.6
.2
dist
0
T2
R1
T1
R2
.
.
.
0
0.2
0.6
1
φ= 0.6
0.2
0
.
.
1
0.5
Location
Memory
Fig. 1. Model of the rat experiment used in [6]. The maze is discretized into
32 positions (squares). The agent can use 4 discrete actions (N,E,S,W). The input
state φ is the concatenation of 32 location components and two reward memory components. The location part of φ represents the activation of 32 place cells co-located
with the maze discrete positions, their activity act depends on the Manhattan distance of the agent to the cell. All figures by Aubin & Girard, 2018; available at
https://doi.org/10.6084/m9.figshare.5822112.v2 under a CC-BY4.0 license.
memory consolidation theory [7]: they would have the role of copying volatile
hippocampal memories into the cortex [8] for reorganization and longer-term
storage [9]. Some recent results have however shown that these reactivations
also have a causal effect on reinforcement learning processes [3,10].
A number of reinforcement learning (RL) algorithms make use of reactivations of their inputs, reminiscent of hippocampal reactivations, that are thus
candidates to explain this phenomenon [11]. Among them, the Dyna family of
algorithms [12] is of special interest because it was specifically designed to make
the best possible use of alternation between on-line and off-line learning phases
(i.e. phases during which the agent acts in the real world or in simulation). We
concentrate here on the Q-learning version of Dyna (Dyna-Q). When operating
on-line, Dyna-Q is indistinguishable from the original model-free Q-learning algorithm: it computes reward prediction error signals, and uses them to update
the estimated values of the (state, action) couples, Q(s, a). In its original version
[12], when off-line, the Dyna algorithm reactivates randomly chosen quadruplets
composed of an initial state, a chosen action, and the predicted resulting state
and reward (produced by a learned world-model, this phase being thus modelbased ), in order to refine the on-line estimated values. However, when the number
of reactivations is under a strict budget constraint, it is more efficient to select
those that will provide more information: those that effectively generated a large
reward prediction error in the last on-line phase, and those that are predicted
to do so by the world model, a principle called prioritized sweeping [13,14].
We are here interested in mimicking the process by which the basal ganglia,
which is central for RL processes [15], can use the state representations of the
world that are provided by the hippocampus. The manipulated state descriptor
will thus be a population activity vector, and we will represent the Q-values and
the world model with neural network approximators [16].
In the following, we describe the rat experimental setup proposed in [6],
and how we simulated it. In this task, a state can have multiple predecessor
states resulting from the execution of a single action, we thus present a modified
Dyna-Q learning algorithm, with a special stress on the neural-network algorithm we designed to learn to approximate binary relations (not restricted to
functions) with a growing approach: GALMO for Growing Algorithm to Learn
Multiple Outputs. Our results successively illustrate three main points. First,
because of interferences between consecutively observed states during maze experience, themselves due to the use of a neural-network function approximator,
the world model had to be learned with shuffled states during off-line replay.
Second, GALMO allows to efficiently solve the multiple predecessor problem.
Third, the resulting system, when faced with a training schedule similar to [6],
generates a lot of disordered state replays, but also a non-negligible set of varied
backward and forward replay sequences, without explicitly storing and replaying
sequences.
2
2.1
Methods
Experimental task
We aim at modeling the navigation task used in [6]: two successive T-mazes (T1
and T2 on Fig. 1), with lateral return corridors. The left and right rewarding
sites deliver food pellets with different flavors. The training involves daily changing contingencies, forcing rats to adapt their choice to turn either left or right at
the final choice (T2) based on the recent history of reward. These contingencies
are: 1) always turn right, while the left side of the maze is blocked; 2) always
turn left, while the right side of the maze is blocked; 3) always turn right; 4)
always turn left; 5) alternate between left and right on a lap-by-lap basis.
Rats attempting to run backward on the maze were physically prevented to
do so by the experimenter. They had forty trials the first day to learn task 1, and
forty trials the second day to learn task 2. Then, depending on their individual
learning speed, rats had between seventeen and twenty days to learn task 3, 4
and 5 (a single condition being presented each day). Once they reached at least
80% success rate on all tasks, rats were implanted with electrodes; after recovery,
recording sessions during task performance lasted for six days.
During the six recording sessions, the reward contingency was changed approximately midway through the session and hippocampal replays were analyzed
when rats paused at reward locations. Original analyses of replayed sequences
[6] revealed that: during same-side replays (i.e., replays representing sequences
of previously visited locations on the same arm of the maze as the current rat
position) forward and backward replays started from the current position; during opposite-side replays (i.e., representing locations on the opposite arm of the
maze) forward replays occurred mainly on the segment leading up to reward
sites, and backward replays covered trajectories ending near reward sites. In
general, the replay content did not seem to only reflect recently experienced trajectories, since trajectories experienced 10 to 15 minutes before were replayed
as well. Indeed, there were more opposite-side replays during task 3 and 4 than
during the alternation task. Finally, among all replays, a few were shortcuts
never experienced before which crossed a straight path on the top or bottom of
the maze between the reward sites.
2.2
Simulation
We have reproduced the T-maze configuration with a discrete environment
composed of 32 squares (Fig. 1, left), each of them represents a 10 × 10 cm
area. States are represented by a vector φ, concatenating place cells activity and
a memory of past rewards (Fig. 1, right). The modeled place cells are centered
on the discrete positions, their activity (color-coded on Fig. 1) decreases with
the Manhattan distance between the simulated rat position to the position they
encode (top of Fig. 1). When a path is blocked (contingencies 1 and 2), the
activity field does not expand beyond walls and will thus shrink, as is the case
of real place cells [17]. To represent the temporal dimension, which is essential
during the alternation task, we have added two more components in the state’s
vector representation (Fig. 1, right): the left side reward memory (L) and the
right side reward memory (R). They take a value of 1 if the last reward was
obtained on that side, 0.5 if the penultimate reward was on that side, and 0 if
that side has not been rewarded during the last two reward events. Therefore,
after two successful laps, the task at hand can be identified by the agent based
on the value of this memory (Tab. 1). This ability to remember the side of the
rewards is supposed to be anchored both on the different position and flavor cues
that characterize each side. Since it has been shown that, beyond purely spatial
information, the hippocampus contains contextual information important for the
task at hand [18], we hypothesize that this memory is also encoded within the
hippocampus, along with the estimation of the agent’s current position.
The agent can choose between four actions: North, South, East and West.
As in the real experiment, the agent cannot run backward.
Table 1. State of the L and R memory components of φ and corresponding meaning
in terms of task at hand, after two successful laps.
L
1
0
0.5
1
R
0
1
1
0.5
Task identification (after 2 laps)
Always turn right (Tasks 1 & 3)
Always turn left (Tasks 2 & 4)
Alternation (Task 5), go left next time
Alternation (Task 5), go right next time
2.3
Neural DynaQ with a prioritized sweeping algorithm
Our algorithm is based on a Dyna architecture [12] which means that, as
in model-based architectures, we need to learn a world model composed of a
reward and a transition model [19]. In order to implement prioritized sweeping
[13,14], the transition model must be designed so as to allow the prediction of
the predecessors of a state s given an action a, because it will be needed to backpropagate the reward prediction computed in state s to its predecessors. Hence,
our architecture is composed of two distinct parts: one dedicated to learning the
world model, and the other one to learning the Q-values.
Algorithm 1 LearnWM: learn the world model
collect S // a set of (φt , φt−1 , a, r) quadruplets
for k ∈ {N, S, E, W } do
SPk ← {(φt , φt−1 ) : (φt , φt−1 , a, r) ∈ S and a = k}
k
← {(φt , r) : (φt , φt−1 , a, r) ∈ S and a = k}
SR
for f ∈ {P, R} do
// P,R: Predecessor and Reward types of networks
Nfk ← null // list of networks (outputs)
Gfk ← null // list of networks (gates)
k
k
create Nnew
; append Nnew
to Nfk
k
k
create Gnew ; append Gnew to Gfk
GALMO(Sfk , Nfk , Gfk ) // refer to Algo 2 for this specific training procedure
end for
end for
Learning the world model. Two sets of neural networks compose the world
model. Four reward networks NRa , one for each action a, learn the association
between (state, action) couples and rewards (NRa : s → r(s, a)). Four other
networks NPa learn the states for which a transition to a given state s is produced
after execution of action a, i.e., the predecessors of s (NPa : s → {s0 }).
Owing to the nature of the task (navigation constrained by corridors) and
the states’ representation, the data that must be learned are not independent.
Indeed, successive state vectors are very similar due to the overlap between
place-fields, and are always encountered in the same order during tasks execution (because the agent always performs the same stereotyped trajectories along
the different corridors). However, it is well known that the training of a neural
network is guaranteed to converge only if there is no correlation in the sequence
of samples submitted during learning, a condition that is often not respected
when performing on-line reinforcement learning [20]. We indeed observed that
in the task at hand, despite its simplicity, it was necessary to store the successive
observations and to train the world model off-line with a shuffled presentation of
the training samples (for the general scheme of the off-line training, see Algo. 1).
For that reason, we created a dataset S compiling all transitions, i.e (φt , φt−1 ,
a, r) quadruplets from all tasks. When there is no predecessor of φt by action
a (as can be the case when this action would require to come through a wall),
the transition is represented as (φt , 0, a, r): those ”null” transitions allow NPa
networks to represent the fact that the transition does not exist.
Correct alternation from left to right
T2
T2
T1
Memory part of Φ:
Followed by a wrong choice at T2
R
R
T2
T1
L=1
R=0.5
L=0.5
R=1
L=0.5
R=1
L=1
R=0.5
T1
same action
different
predecessors
L=1
R=0.5
L=1
R=0.5
R
Fig. 2. Example of multiple predecessors in the alternation task. The agent
first correctly goes to the right (left). It then goes to the left (middle) where, at the
reward site, its predecessor state has a (L = 0.5, R = 1) memory component. It then
makes a wrong decision and goes to the left again (right), but is not rewarded: at this
last position, the location component of the predecessor state (white mouse) is identical
but the memory component is different (L = 1, R = 0.5) from the previous lap. Violet
gradient: past trajectory; white mouse: previous position; gray mouse: current position;
white R: agent rewarded; black R: current position of the reward.
Despite its simplicity, the navigation task modeled here has some specificities:
during task 5 (alternation), some states have more than one predecessor for a
given action (see an example on Fig. 2), the algorithm must thus be capable
of producing more than one output for the same input. To do that, we have
created a growing type of algorithm inspired by mixture of expert algorithms
[21] (which we call here the GALMO algorithm, see Algo. 2), based on the
following principles:
– The algorithm should allow the creation of multiple Ni networks (if needed)
so that a single input can generate multiple outputs. Each of these network is
coupled with a gating network Gi , used after training to know if the output
of Ni has to be taken into account when a given sample is presented.
– When a sample is presented, the algorithm should only train the Ni network
that generates the minimal error (to enforce network specialization), and
remember this training event by training Gi to produce 1 and the other
Gk6=i to produce 0.
– The algorithm should track the statistics of the minimal training errors of
each sample during an epoch, so as to detect outliers (samples whose error
is much higher than the others’). GALMO assumes that these outliers are
caused by inputs who should predict multiple outputs and are stuck in predicting the barycenter of the expected outputs. A sample is considered an
outlier when its error is larger than a threshold θ, equal to the median of the
current error distribution, plus w times the amplitude of the third quartile
(Q3 − median). When such a detection occurs, a new network is created on
the fly, based on a copy of the network that produced the minimal error for
the sample. The new network is then trained once on the sample at hand.
Algorithm 2 GALMO: Growing algorithm to learn multiple outputs
INPUT: S, N , G
OUTPUT: N , G
// S = h(in0 , out0 ), ..., (inn , outn )i : list of samples
// N = hN0 i : lists of neural networks (outputs)
// G = hG0 i : lists of neural networks (gates)
θ ← +∞
for nbepoch ∈ {1, maxepoch} do
M ← null // M is a list of the minimal error per sample
for each (in,out)∈ S do
E ← null // E is a list of errors for a sample
for each N ∈ N do
append kN (in) − outkL1 to E
end for
if min(E) < θ then
backprop(Nargmin(E) , in, out)
backprop(Gargmin(E) , in, 1)
for each G ∈ G with G 6= Gargmin(E) do
backprop(G, in, 0)
end for
else
create Nnew ; append Nnew to N
Nnew ← copy(Nargmin(E) )
backprop(Nnew , input =in, target =out)
create Gnew ; append Gnew to G
backprop(Gnew , in, 1)
end if
end for
θ ← median(M) + w ∗ (Q3(M) − median(M))
end for
In principle, the algorithm could be modified to limit the maximal number
of created networks, or to remove the networks that are not used anymore, but
these additions were not necessary here.
Neural Dyna-Q. The second part of the algorithm works as a classical
neural network-based Dyna-Q [16] with prioritized sweeping [13,14]. As in [16],
a
the Q-values are represented by four 2-layer feedforward neural networks NQ
(one per action). During on-line phases, the agent makes decisions that drive its
movements within the maze, and stores the samples in a priority queue, their
priority is the absolute value of the reward prediction error, i.e., |δ|. Every time
the agent receives a reward, similarly to rats, it stops and replays are simulated
with a budget B (Algo. 3): the samples with the highest priority are replayed
Algorithm 3 Neural Dyna-Q with prioritized sweeping & multiple predecessors
INPUT: φt=0 , NP , GP , NR , GR
a∈{N,S,E,W }
OUTPUT: NQ
PQueue ← {} // PQueue: empty priority queue
nbTrials ← 0
repeat
a ← softmax(NQ (φt ))
take action a, receive r, φt+1
a
backprop(NQ
, input = φt , target = r + γmaxa (NQ (φt+1 ))
a
Put φt in PQueue with priority |NQ
(φt ) − (r + γmaxa (NQ (φt+1 )))|
if r> 0 then
nbReplays ← 0
Pr = hi // empty list of predecessors
repeat
φ ← pop(PQueue)
for each GP ∈ GP do
if GP (φ) > 0 then
k ← index(GP )
append NPk (φ) to Pr
end if
end for
for each p ∈ Pr s.t norm(p) > do
for each a ∈ {N, S, E, W } do
a
a
a
(φ)))
(p) + γmaxa (NQ
, input = p, target = NR
backprop(NQ
a
a
a
(φ)))|
(p) + γmaxa (NQ
(p) − (NR
Put p in PQueue with priority |NQ
nbReplays ← nbReplays + 1
end for
end for
until PQueue empty OR nbReplays ≥ B
end if
φt ← φt+1
nbTrials ← nbTrials +1
until nbTrials = maxNbTrials
first, their potential predecessors are then estimated and placed in the queue
with their respective priorities, and so on until the replay budget is exhausted.
The various parameters used in the simulations are summarized in Tab. 2.
3
3.1
Results
Learning the world model
Because of correlations in sample sequences, the world model is learned off-line:
the samples are presented in random order, so as to break temporal correlations.
We illustrate this necessity with the learning of the reward networks NR : when
trained on-line (Fig. 3, left), the reward networks make a lot of erroneous pre-
Table 2. Parameter values.
value
parameter
4000 maxepch: number of epoch replays to train the world model
3
w: gain of the outlier detector threshold in GALMO
20
B: replay budget per stop at reward sites
2
number of layers in NP , NR and NQ
10, 16, 26 size of the hidden layers in NQ , NR and NP (respectively)
±0.05, ±0.0045, ±0.1
weight initialization bound in NQ , NR and NP (resp.)
0.5, 0.1, 0.1
learning rate in NQ , NR and NP (resp.)
0.9, 1, 1 sigmoid slope in NP , NR and NQ (resp.) (hidden layer)
0.5, 0.4, 0.4 sigmoid slope in NP , NR and NQ (resp.) (output layer)
Fig. 3. Reward predictions are inaccurate when the model is trained on-line (Left
panel) and accurate when it is trained off-line (Right panel). L, R: memory configuration. Note the use of a logarithmic scale, so as to make visible errors of small amplitude.
dictions for each possible task, while when trained off-line with samples presented
in randomized order, the predictions are correct (Fig. 3, right).
With a single set of NP networks, the error of the whole set of states decreases
steadily with learning, except for four states which have multiple predecessors
(Fig. 4, top left). With the GALMO algorithm, when the error of these states
reaches the threshold θ (in red on Fig. 4, top right), networks are duplicated
and specialized for each of the possible predecessors. We repeated the experiment 10 times. It always converged, with a number of final networks comprised
between 2 and 5 (3 times 2 networks, 1 time 3, 5 times 4, and 1 time 5).
3.2
Reinforcement learning with multiple predecessors
We compare the efficiency of the Dyna-Q model we developed with the corresponding Q-learning (i.e. the same architecture without replays with the world
model). As expected, Q-learning is able to learn the task, measured here with
the proportion of erroneous choices at the decision point T2 (Fig. 4, bottom
left). On average it does not fully converge before 1000 epochs of training, the
Dyna-Q learns much faster, thanks to the replay mechanism (Fig. 4, bottom
right), converging on average after 200 trials.
Error evolution at state T2
Error evolution at state T2
1
0.9
0.8
0.8
0.7
0.7
0.6
0.6
Errors
Errors
1
0.9
0.5
0.5
0.4
0.4
0.3
0.3
0.2
0.2
0.1
0
0
0.1
200
400
600
800
1000
1200
1400
1600
1800
2000
0
0
200
400
600
800
#episodes
1000
1200
1400
1600
1800
2000
#episodes
Fig. 4. Top: Learning error dynamics without (left) and with (right)
GALMO. Errors of all samples (gray) during epochs of training. GALMO allows
for the creation of multiple prediction networks to handle the states where multiple
outputs have to be generated. Bottom: Learning without (left) and with (right)
replays. Evolution of the proportion of decision errors at point T2 during the alternation task. Blue: 10 run average, light blue: standard deviation.
3.3
Preliminary analysis of generated replays
Always Right
B
RND
Same
Opposite
Always Left
B
Alternate
B
RND
Same
Opposite
Same
Central
F
RND
Opposite
Central
O
Fig. 5. Type of replays. B: backward, F: forward, RND: random.
We analyze a posteriori the state reactivations caused by the prioritized
sweeping DynaQ algorithm in always turn right, always turn left and alternate
tasks. Prioritized sweeping does not rely on explicit replay of sequences, but it
however favors them. We considered sequences or replays implying three or more
consecutive steps; with 128 possible states, a 3-state sequence has a chance level
of 0.01% of being produced by a uniform random selection process. We observed
in all cases (Fig. 5) that a bit more than 80% of the state reactivations did not
correspond to actual sequences. Most of the sequences are backward, except for
the alternate task, which also generated 4.5% of forward ones. As in [6], we classified these sequences as being on the same side as the current agent location, on
the opposite side, or in the central part of the maze. There is no clear pattern
here, except that central reactivations were observed in the alternate task only.
4
Discussion
We proposed a new neural network architecture (GALMO) designed to associate
multiple outputs to a single input, based on the multiple expert principle [21]. We
implemented a neural version of the DynaQ algorithm [16], using the prioritized
sweeping principle [13,14], using GALMO to learn the world model. This was
necessary because the evaluation task, adapted from [6], contained some states
that have multiple predecessor.
We showed that this system is able to learn the multiple predecessors cases,
and to solve the task faster than the corresponding Q-learning system (i.e., without replays). This required learning the world-model off-line, with data presented
in shuffled order, so as to break the sequential correlations between them, which
prevented the convergence of the learning process. A neuroscience prediction derives from this result (independently from the use of GALMO): if the learning
principles of the rat brain are similar to those of the gradient descent for artificial
neural network, then the world model has to be learned off-line, which would
be compatible with non-sequential hippocampal replays. Besides, the part of the
DynaQ algorithm that uses the world model to update the Q-values predicts a
majority of non-sequential replays, but also 15 to 20% of sequential reactivations,
both backward and forward.
Concerning GALMO, it has been tested with a quite limited set of data,
and should thus be evaluated against larger sets in future work. In our specific
case, the reward networks NRa did not require the use of GALMO; a single network could learn the full (s, a) → r mapping as rewards were deterministic. But
should they be stochastic, GALMO could be used to learn the multiple possible
outcomes. Note that, while it has been developed in order to learn a predecessor
model in a DynaQ architecture, GALMO is much more general, and would in
principle be able to learn any one-to-many mapping. Finally, in the model-based
and dyna reinforcement learning contexts, having multiple predecessors or successors is not an exceptional situation, especially in a robotic paradigm. The
proposed approach is thus of interest beyond the task used here.
Acknowledgements
The authors would like to thank Olivier Sigaud for fruitful discussions. This
work has received funding from the European Unions Horizon 2020 research and
innovation programme under grant agreement No 640891 (DREAM Project).
This work was performed within the Labex SMART (ANR-11-LABX-65) supported by French state funds managed by the ANR within the Investissements
d’Avenir programme under reference ANR-11-IDEX-0004-02.
References
1. O’Keefe, J., Dostrovsky, J.: The hippocampus as a spatial map. preliminary evidence from unit activity in the freely-moving rat. Brain research 34(1) (1971)
171–175
2. Wilson, M.A., McNaughton, B.L., et al.: Reactivation of hippocampal ensemble
memories during sleep. Science 265(5172) (1994) 676–679
3. Girardeau, G., Benchenane, K., Wiener, S.I., Buzsáki, G., Zugaro, M.B.: Selective
suppression of hippocampal ripples impairs spatial memory. Nature neuroscience
12(10) (2009) 1222–1223
4. Foster, D.J., Wilson, M.a.: Reverse replay of behavioural sequences in hippocampal
place cells during the awake state. Nature 440(7084) (2006) 680–3
5. Lee, A.K., Wilson, M.A.: Memory of Sequential Experience in the Hippocampus
during Slow Wave Sleep. Neuron 36(6) (2002) 1183–1194
6. Gupta, A.S., van der Meer, M.A.A., Touretzky, D.S., Redish, A.D.: Hippocampal
Replay Is Not a Simple Function of Experience. Neuron 65(5) (2010) 695–705
7. Chen, Z., Wilson, M.A.: Deciphering neural codes of memory during sleep. Trends
in Neurosciences (2017)
8. Peyrache, A., Khamassi, M., Benchenane, K., Wiener, S.I., Battaglia, F.P.: Replay
of rule-learning related neural patterns in the prefrontal cortex during sleep. Nature
Neuroscience 12(7) (2009) 919–926
9. McClelland, J.L., McNaughton, B.L., O’reilly, R.C.: Why there are complementary
learning systems in the hippocampus and neocortex: insights from the successes
and failures of connectionist models of learning and memory. Psychological review
102(3) (1995) 419
10. De Lavilléon, G., Lacroix, M.M., Rondi-Reig, L., Benchenane, K.: Explicit memory
creation during sleep demonstrates a causal role of place cells in navigation. Nature
neuroscience 18(4) (2015) 493–495
11. Cazé, R., Khamassi, M., Aubin, L., Girard, B.: Hippocampal replays under the
scrutiny of reinforcement learning models. submitted (2018)
12. Sutton, R.S.: Integrated architectures for learning, planning, and reacting based on
approximating dynamic programming. In: Proceedings of the seventh international
conference on machine learning. (1990) 216–224
13. Moore, A.W., Atkeson, C.G.: Prioritized sweeping: Reinforcement learning with
less data and less time. Machine learning 13(1) (1993) 103–130
14. Peng, J., Williams, R.J.: Efficient learning and planning within the dyna framework. Adaptive Behavior 1(4) (1993) 437–454
15. Khamassi, M., Lacheze, L., Girard, B., Berthoz, A., Guillot, A.: Actor-critic models of reinforcement learning in the basal ganglia: from natural to arificial rats.
Adaptive Behavior 13 (2005) 131–148
16. Lin, L.H.: Self-improving reactive agents based on reinforcement learning, planning
and teaching. Machine learning 8(3/4) (1992) 69–97
17. Paz-Villagrán, V., Save, E., Poucet, B.: Independent coding of connected environments by place cells. European Journal of Neuroscience 20(5) (2004) 1379–1390
18. Eichenbaum, H.: Prefrontal–hippocampal interactions in episodic memory. Nature
Reviews Neuroscience 18(9) (2017) 547
19. Sutton, R., Barto, A.: Reinforcement Learning: An Introduction. Cambridge, MA:
MIT Press (1998)
20. Tsitsiklis, J.N., Van Roy, B.: Analysis of temporal-diffference learning with function
approximation. In: Advances in neural information processing systems. (1997)
1075–1081
21. Jacobs, R.A., Jordan, M.I., Nowlan, S.J., Hinton, G.E.: Adaptive mixtures of local
experts. Neural computation 3(1) (1991) 79–87
| 2cs.AI
|
Tame Decompositions and Collisions
arXiv:1402.5945v1 [math.AC] 24 Feb 2014
Konstantin Ziegler
B-IT, Universität Bonn
D-53113 Bonn, Germany
[email protected]
http://cosec.bit.uni-bonn.de/
April 10, 2018
Abstract
A univariate polynomial f over a field is decomposable if f =
g ◦ h = g(h) for nonlinear polynomials g and h. It is intuitively clear
that the decomposable polynomials form a small minority among all
polynomials over a finite field. The tame case, where the characteristic
p of Fq does not divide n = deg f , is fairly well-understood, and we
have reasonable bounds on the number of decomposables of degree
n. Nevertheless, no exact formula is known if n has more than two
prime factors. In order to count the decomposables, one wants to
know, under a suitable normalization, the number of collisions, where
essentially different (g, h) yield the same f . In the tame case, Ritt’s
Second Theorem classifies all 2-collisions.
We introduce a normal form for multi-collisions of decompositions
of arbitrary length with exact description of the (non)uniqueness of the
parameters. We obtain an efficiently computable formula for the exact
number of such collisions at degree n over a finite field of characteristic
coprime to p. This leads to an algorithm for the exact number of
decomposable polynomials at degree n over a finite field Fq in the tame
case.
1
Introduction
The composition of two univariate polynomials g, h ∈ F [x] over a field F
is denoted as f = g ◦ h = g(h), and then (g, h) is a decomposition of f ,
and f is decomposable if g and h have degree at least 2. In the 1920s, Ritt,
1
Fatou, and Julia studied structural properties of these decompositions over
C, using analytic methods. Particularly important are two theorems by Ritt
on the uniqueness, in a suitable sense, of decompositions, the first one for
(many) indecomposable components and the second one for two components,
as above. Engstrom (1941) and Levi (1942) proved them over arbitrary fields
of characteristic zero using algebraic methods.
The theory was extended to arbitrary characteristic by Fried & MacRae
(1969), Dorey & Whaples (1974), Schinzel (1982, 2000), Zannier (1993), and
others. Its use in a cryptographic context was suggested by Cade (1985).
In computer algebra, the decomposition method of Barton & Zippel (1985)
requires exponential time. A fundamental dichotomy is between the tame
case, where the characteristic p does not divide deg g, and the wild case,
where p divides deg g, see von zur Gathen (1990a,b). A breakthrough result
of Kozen & Landau (1989) was their polynomial-time algorithm to compute
tame decompositions; see also von zur Gathen, Kozen & Landau (1987);
Kozen, Landau & Zippel (1996); Gutierrez & Sevilla (2006), and the survey
articles of von zur Gathen (2002) and Gutierrez & Kozen (2003) with further
references.
Schur’s conjecture, as proven by Turnwald (1995), offers a natural connection between indecomposable polynomials with degree coprime to p and
certain absolutely irreducible bivariate polynomials. On a different, but related topic, Avanzi & Zannier (2003) study ambiguities in the decomposition
of rational functions over C.
It is intuitively clear that the univariate decomposable polynomials form
only a small minority among all univariate polynomials over a field. A
set of distinct decompositions of f is called a collision. The number of
decomposable polynomials of degree n is thus the number of all pairs (g, h)
with deg g · deg h = n reduced by the ambiguities introduced by collisions.
An important tool for estimating the number of collisions is Ritt’s Second
Theorem. Ritt worked with F = C and used analytic methods. Subsequently,
his approach was replaced by algebraic methods and Ritt’s Second Theorem
was also shown to hold in positive characteristic p. The original versions of
this required p > deg(g ◦ h). Zannier (1993) reduced this to the milder and
more natural requirement g 0 6= 0 for all g in the collision. His proof works
over an algebraic closed field, and Schinzel’s (2000) monograph adapts it to
finite fields.
The task of counting compositions over a finite field of characteristic p
was first considered in Giesbrecht (1988). Von zur Gathen (2014a) presents
general approximations to the number of decomposable polynomials. These
come with satisfactory (rapidly decreasing) relative error bounds except when
p divides n = deg f exactly twice. Blankertz, von zur Gathen & Ziegler (2013)
2
determine exactly the number of decomposable polynomials in one of these
difficult cases, namely when n = p2 .
Zannier (2008) studies a different but related question, namely compositions f = g ◦ h in C[x] with a sparse polynomial f , having t terms. The
degree is not bounded. He gives bounds, depending only on t, on the degree
of g and the number of terms in h. Furthermore, he gives a parametrization
of all such f , g, h in terms of varieties (for the coefficients) and lattices (for
the exponents). Bodin, Dèbes & Najib (2009) also deal with counting.
Zieve & Müller (2008) derive an efficient method for describing all complete
decompositions of a polynomial, where all components are indecomposable.
This turns Ritt’s First Theorem into an applicable form and Medvedev &
Scanlon (2014) combine this approach with results from model theory to
describe the subvarieties of the k-dimensional affine space that are preserved
by a coordinatewise polynomial map. Both works lead to slightly different
canonical forms for the complete decomposition of a given polynomial. Zieve
& Müller (2008) employ Ritt moves, where adjacent indecomposable g, h in a
complete decomposition are replaced by g ∗ , h∗ with the same composition,
but deg g = deg h∗ 6= deg h = deg g ∗ . Such collisions are the theme of Ritt’s
Second Theorem and von zur Gathen (2014b) presents a normal form with
an exact description of the (non)uniqueness of the parameters.
Our work combines the “normalizations” of Ritt’s theorems by Zieve &
Müller (2008) and von zur Gathen (2014b) to classify collisions of two or
more decompositions, not necessarily complete and of arbitrary length. We
make the following contributions.
• We obtain a normal form for collisions described by a set of degree
sequences for (possibly incomplete) decompositions. (Theorem 3.14 and
Theorem 3.15)
• The (non)uniqueness of the parameters leads to an exact formula for the
number of such collisions over a finite field with characteristic coprime
their degree. (Theorem 4.1)
• We conclude with an efficient algorithm for the number of decomposable
polynomials at degree n over a finite field of characteristic coprime n.
(Algorithm 4.2)
The latter extends the explicit formulae of von zur Gathen (2014a) for n a
semiprime or the cube of a prime.
We proceed in three steps. In Section 2, we introduce notation and
establish basic relations. In Section 3, we introduce the relation graph of a set
of collisions which captures the necessary order and possible Ritt moves for
3
any in decomposition. This leads to a complete classification of collisions by
Theorem 3.14 and Theorem 3.15. We conclude with the corresponding formula
for the number of such collisions over a finite field and the corresponding
procedure in Section 4.
2
Notation and Preliminaries
A nonzero polynomial f ∈ F [x] over a field F of characteristic p ≥ 0 is monic
if its leading coefficient lc(f ) equals 1. We call f original if its graph contains
the origin, that is, f (0) = 0. For g, h ∈ F [x],
f = g ◦ h = g(h) ∈ F [x]
(2.1)
is their composition. If deg g, deg h ≥ 2, then (g, h) is a decomposition of f .
A polynomial f ∈ F [x] is decomposable if there exist such g and h, otherwise
f is indecomposable. A decomposition (2.1) is tame if p - deg g, and f is tame
if p - deg f .
Multiplication by a unit or addition of a constant does not change decomposability, since
f = g ◦ h ⇐⇒ af + b = (ag + b) ◦ h
for all f , g, h as above and a, b ∈ F with a 6= 0. In other words, the set of
decomposable polynomials is invariant under this action of F × × F on F [x].
Furthermore, any decomposition (g, h) can be normalized by this action, by
taking a = lc(h)−1 ∈ F × , b = −a · h(0) ∈ F , g ∗ = g((x − b)a−1 ) ∈ F [x], and
h∗ = ah + b. Then g ◦ h = g ∗ ◦ h∗ and g ∗ and h∗ are monic original.
It is therefore sufficent to consider compositions f = g ◦ h where all three
polynomials are monic original. For n ≥ 1 and any positive divisor d of n, we
write
Pn (F ) = {f ∈ F [x] : f is monic original of degree n},
Dn (F ) = {f ∈ Pn : f is decomposable},
Dn,d (F ) = {f ∈ Pn : f = g ◦ h for some (g, h) ∈ Pd × Pn/d }.
We sometimes leave out F from the notation when it is clear from the context
and have over a finite field Fq with q elements,
#Pn = q n−1 .
(2.1a)
It is well known that in a tame decomposition, g and h are uniquely determined
and we have over Fq
#Dn,d = q n+n/d−2
(2.2)
4
if n is coprime to p.
The set Dn of all decomposable polynomials in Pn satisfies
Dn =
[
Dn,d .
(2.3)
d|n
1<d<n
In particular, Dn = ∅ if n is prime. Our collisions turn up in the resulting
inclusion-exclusion formula for #Dn if n is composite.
Let N = {1 < d < n : d | n} be the set of nontrivial divisors of n and
D ⊆ N a nonempty subset of size k. This defines a set
Dn,D =
\
Dn,d
d∈D
of k-collisions. We obtain from (2.3) the inclusion-exclusion formula
#Dn =
X
(−1)k+1
X
#Dn,D .
(2.4)
D⊆N
#D=k
k≥1
For #D = 1, the size of Dn,D is given in (2.2). For #D = 2, the central
tool for understanding is Ritt’s Second Theorem as presented in the next
subsection.
For f ∈ Pn (F ) and a ∈ F , the original shift of f by a is
f [a] = (x − f (a)) ◦ f ◦ (x + a) ∈ Pn (F ).
(2.5)
Original shifting defines a group action of the additive group of F on Pn (F ).
Shifting respects decompositions in the sense that for each decomposition
(g, h) of f we have a decomposition (g [h(a)] , h[a] ) of f [a] , and vice versa. We
denote (g [h(a)] , h[a] ) as (g, h)[a] . The stabilizer of a monic original polynomial
f under original shifting is F if f is linear and {0} otherwise.
2.1
Normal Form for Ritt’s Second Theorem
In the 1920s, Ritt, Fatou, and Julia investigated the composition f = g ◦ h =
g(h) of univariate polynomials over a field F for F = C. It emerged as an
important question to determine the collisions (or nonuniqueness) of such
decompositions, that is, different components (g, h) 6= (g ∗ , h∗ ) with equal
composition g ◦ h = g ∗ ◦ h∗ and equal sets of degrees: deg g = deg h∗ 6=
deg h = deg g ∗ .
Ritt (1922) presented two types of essential collisions:
xe ◦ xk w(xe ) = xke we (xe ) = xk we ◦ xe ,
5
(2.6)
∗
Td∗ (x, z e ) ◦ Te∗ (x, z) = Ted
(x, z) = Te∗ (x, z d ) ◦ Td∗ (x, z),
(2.7)
where w ∈ F [x], z ∈ F × = F \ {0}, and Td∗ is the dth Dickson polynomial
of the first kind. And then he proved that these are all possibilities up to
composition with linear polynomials. This involved four unspecified linear
functions, and it is not clear whether there is a relation between the first and
the second type of example. Without loss of generality, we use the originalized
dth Dickson polynomial Td (x, z) = Td∗ (x, z) − Td∗ (0, z) which also satisfy (2.7).
Von zur Gathen (2014b) presents a normal form for the decompositions
6 0 and the standard
in Ritt’s Theorem under Zannier’s assumption g 0 (g ∗ )0 =
assumption gcd(e, d) = 1, where d = k + e deg w in (2.6). This normal form
is unique unless p | m.
Theorem 2.8. (Ritt’s Second Theorem, Normal Form, tame case) Let d >
e ≥ 2 be coprime integers, and n = de coprime to the characteristic of F .
Furthermore, let f = g ◦ h = g ∗ ◦ h∗ be monic original polynomials with
deg g = deg h∗ = d, deg h = deg g ∗ = e.
Then either (i) or (ii) holds, and (iii) is also valid.
(i) (Exponential Case) There exists a monic polynomial w ∈ F [x] of degree
s and a ∈ F so that
f = (xke we (xe ))[a]
where d = se+k is the division with remainder of d by e, with 1 ≤ k < e.
Furthermore
(g ∗ , h∗ ) = (xe , xk w(xe ))[a] ,
(g, h) = (xk we , xe )[a] ,
and (w, a) is uniquely determined by f and d. Conversely, any (w, a)
as above yields a 2-collision via the above formulas.
(ii) (Trigonometric Case) There exist z, a ∈ F with z 6= 0 so that
f = Tn (x, z)[a] .
Furthermore we have
(g, h) = (Td (x, z e ), Te (x, z))[a] ,
(g ∗ , h∗ ) = (Te (x, z d ), Td (x, z))[a] ,
and (z, a) is uniquely determined by f . Conversely, any (z, a) as above
yields a 2-collision via the above formulas.
(iii) For e = 2, the Trigonometric Case is included in the Exponential Case.
For e ≥ 3, the Exponential and Trigonometric Cases are mutually
exclusive.
6
If p - n, then the case where gcd(d, e) 6= 1 is reduced to the previous one
by the following result about the left and right greatest common divisors
of decompositions. It was shown over algebraically closed fields by Tortrat
(1988, Proposition 1); a more concise proof using Galois theory is due to Zieve
& Müller (2008, Lemma 2.8). We use the version of von zur Gathen (2014b,
Fact 6.1(i)), adapted to monic original polynomials.
Proposition 2.9. Let d, e, d∗ , e∗ ≥ 2 be integers and de = d∗ e∗ coprime to p.
Furthermore, let g ◦ h = g ∗ ◦ h∗ be monic original polynomials with deg g = d,
deg h = e, deg g ∗ = d∗ , deg h∗ = e∗ , and ` = gcd(d, d∗ ), r = gcd(e, e∗ ).
Then there are unique monic original polynomials a and b of degree ` and r,
respectively, such that
g = a ◦ u,
g ∗ = a ◦ u∗ ,
h = v ◦ b,
h∗ = v ∗ ◦ b,
for unique monic original polynomials u, u∗ , v, v ∗ of degree d/`, d∗ /`, e/r,
and e∗ /r, respectively.
This determines Dn,{d,e} exactly if p - n = de.
For coprime integers d ≥ 2 and e ≥ 1, we define the sets
Ed,e =
P
d
k
for e = 1,
e
{x w ∈ Pd : d = s · e + k with 1 ≤ k < e
(2.10)
and w ∈ Fq [x] monic of degree s},
otherwise,
Td,e = {Td (x, z e ) ∈ Pd : z ∈ F×
q }
of exponential and trigonometric components, respectively. For d < e, we
have s = 0, k = d in (2.10), and therefore
Ed,e = {xd }.
(2.10a)
This allows the following reformulation of Theorem 2.8.
Corollary 2.11. Let f ∈ Dn,{d,e} . Then either (i) or (ii) holds and (iii) is
also valid.
(i) There is a unique monic original g ∈ Ed,e and a unique a ∈ F such that
f = (g ◦ xe )[a] .
7
(ii) There is a unique monic original g ∈ Tde,1 and a unique a ∈ F such
that
f = g [a] .
(iii) If e = 2, then case (ii) is included in case (i). If e ≥ 3, they are
mutually exclusive.
Conversely, we have
[F ]
Dn,{d,e} = (Ed,e ◦ Ee,d )[F ] ∪ Tde,1 ,
where the union is disjoint if and only if e ≥ 3, and
#Dn,{d,e} = q · (q bd/ec + (1 − δe,2 )(q − 1)).
With respect to the size under original shifting, we have the following
consequences.
Proposition 2.14. For F = Fq , coprime d ≥ 2 and e ≥ 1, both coprime to
p, we have
[F ]
#Td,eq
[F ]
#Ed,eq
(
=
=
q
q(q − 1)/ gcd(q − 1, e)
d−1
q
for d = 2,
otherwise,
for e = 1,
bd/2c+1
− q(q − 1)/2 for e = 2,
bd/ec+1
q
otherwise.
q
Proof. (i) For d = 2, we have T2,e = {x2 } independent from e. Since
p 6= 2, the original shifts (x2 )[a] = x2 + 2ax range over all monic original
[F ]
polynomials of degree 2 as a runs over all field elements. Thus T2,e = P2
and the size follows from (2.1a).
For d > 2, the coefficient of xd−2 in Td (x, z e ) ∈ Td,e is −dz e . Since there
are exactly (q − 1)/ gcd(q − 1, e) distinct eth powers z e for nonzero
elements z ∈ Fq , this shows
#Td,e = (q − 1)/ gcd(q − 1, e).
For the claimed formula it is sufficient to show that for T ∈ Td,e and
a ∈ F , we have
T [a] ∈ Td,e if and only if a = 0.
8
For d odd, we also have T an odd polynomial. Hence the coefficient of
xd−1 in T is 0 and the coefficient of xd−1 in T [a] is ad. This proves the
claim. For d even, the same argument applies with “odd” replaced by
“even”.
For a nonzero polynomial f ∈ F [x] and b in some algebraic closure K of F ,
let multb (f ) denote the root multiplicity of b in f , so that f = (x − b)multb (f ) u
with u ∈ K[x] and u(b) 6= 0.
[F ]
(ii) For e = 1, we have Ed,e = Pd = Pd
of size q d−1 by (2.1a).
For e > 1, we have #Ed,e = q bd/ec . It is sufficient to show that for
f ∈ Ed,e and a ∈ F , we have
f [a] ∈ Ed,e if and only if a = 0.
We have directly f [0] = f ∈ Ed,e . Conversely, let f [a] = f¯ = xk w̄e ∈ Ed,e ,
and compare the derivatives
0
f [a] = (x + a)k−1 w(x + a)e−1 (kw(x + a) + e(x + a)w0 (x + a)),
f¯0 = xk−1 w̄e−1 (k w̄ + exw̄0 ),
respectively. These are nonzero, since p - d = deg(f¯) = deg(f [a] ), and
we compute the root multiplicity of 0 as
mult0 (f
[a] 0
(
) = [a = 0] · (k − 1) + multa (w) · (e − 1) +
multa (w)
if p | multa (w),
multa (w) − [a 6= 0] otherwise.
= e multa (w) + [a = 0](k − 1) − [p - multa (w) and a 6= 0],
mult0 (f¯0 ) = k − 1 + (e − 1) · mult0 (w̄) + mult0 (w̄)
= e mult0 (w̄) + k − 1.
If f [a] = f¯, we find modulo e
k − 1 = [a = 0] · (k − 1) − [p - multa (w) and a 6= 0].
This holds if
• a = 0 or
• p | multa (w) and k = 1.
9
It remains to show that the latter case is included in the former. In other
words, that p | multa (w) and k = 1 imply a = 0. Let ∆ = w(x + a) − w̄
of degree less than s, since both are monic. Then
0
0 = f [a] − f¯0 = w(x + a)e − w̄e + e((x + a)w0 (x + a) − xw̄0 )
!
e 2 e−2
= e∆w̄
+
∆ w̄
+ ...
2
+ (e − 1)xw̄0 + ex∆0 + eaw̄0 + ea∆0
e−1
If ∆ = 0, we are done. Otherwise, deg ∆ ≥ 0 and the coefficient of
xdeg ∆+s(e−1) is
e lc(∆) + s(e − 1)[e = 2 and deg ∆ = 0].
If e > 2, this is nonzero, a contradiction.
We have w(x + a) = w̄. Feeding this and k = 1 back into the definition
of f [a] and f¯, we obtain for their difference
f [a] − f¯ = (x + a)w̄e − f (a) − xw̄e = aw̄e − f (a).
With deg(w) = s > 0 this implies a = 0.
3
Normal Form for Collisions
The results of the previous section suffice to describe 2-collisions of decompositions g ◦ h = g ∗ ◦ h∗ with length 2 each. This section describes the
structure of “many”-collisions of decompositions with arbitrary, possibly pairwise distinct, lengths. Let d = (d1 , d2 , . . . , d` ) be an ordered factorization of
n = d1 · d2 · . . . · d` with ` nontrivial divisors di ∈ N , 1 ≤ i ≤ `, and define
the set
Dn,d = {f ∈ Pn : f = g1 ◦ · · · ◦ g` with deg gi = di for all 1 ≤ i ≤ `}
of decomposable polynomials with decompositions of length ` and degree
sequence d. For a set D = {d(1) , d(2) , . . . , d(c) } of c ordered factorizations of n,
we define
Dn,D =
\
Dn,d
d∈D
10
(k)
(k)
(k)
= {f ∈ Pn : f = g1 ◦ · · · ◦ g`k with deg gi
(k)
= di
for all 1 ≤ k ≤ c and 1 ≤ i ≤ `k }.
For #D = 1, we have
Dn,D = Dn,d = Pd1 ◦ Pd2 ◦ · · · ◦ Pd` ,
#Dn,D = #Dn,d = q
P
1≤i≤`
di −`
,
where D = {d} and d = (d1 , d2 , . . . , d` ). The rest of this section deals with
#D > 1.
We determine the structure of Dn,D . First, we replace D by a refinement D∗ ,
where all elements are suitable permutations of the same ordered factorization
of n. Second, we define the relation graph of D∗ that captures the degree
sequences for polynomials in Dn,D . Finally, we classify the elements of Dn,D
as a composition of unique trigonometric or unique exponential components
as defined in (2.10).
3.1
A refinement of D
Let d = (d1 , d2 ) and e = (e1 , e2 ) be distinct ordered factorizations of n.
Let ` = gcd(d1 , e1 ), d∗1 = d1 /`, e∗1 = e1 /`, r = gcd(d2 , e2 ), d∗2 = d2 /r, and
e∗2 = e2 /r. Then Proposition 2.9 shows
Dn,{d,e} = Dn,{d∗ ,e∗ }
for d∗ = (`, d∗1 , d∗2 , r) and e∗ = (`, e∗1 , e∗2 , r) with gcd(d∗1 , e∗1 ) = 1 = gcd(d∗2 , e∗2 )
and therefore d∗1 = e∗2 and d∗2 = e∗1 . We generalize this procedure to two
ordered factorizations of arbitrary length. For squarefree n this is similar to
the computation of a coprime (also: gcd-free) basis for {d1 , d2 , e1 , e2 }, if we
keep duplicates and the order of factors; see Bach & Shallit (1997, Section 4.8).
For squareful n, the factors with gcd > 1 require additional attention.
Let d = (d1 , d2 , . . . , d` ) be an ordered factorization of n and call the underlying unordered multiset d = {d1 , d2 , . . . , d` } of divisors its basis. A refinement of
d is an ordered factorization d∗ = (d∗11 , . . . , d∗1m1 , d∗21 , . . . , d∗2m2 , . . . , d∗`1 , . . . , d∗`m` ),
Q
where di = 1≤k≤mi d∗ik for all 1 ≤ i ≤ `. We write d∗ | d and have directly
Dn,d∗ ⊆ Dn,d .
(3.1)
Every ordered factorization is a refinement of (n). A complete refinement
of d = (di )1≤i≤` is obtained by replacing every di by one of its ordered
factorization into primes.
11
Two ordered factorizations d = (d1 , . . . , d` ) and e = (e1 , . . . , e` ) of n with
the same basis, define a permutation σ = σ(d, e) on the indices 1, 2, . . . , `
through
di = eσ(i)
(3.2)
for 1 ≤ i ≤ `. We require
σ(i) < σ(j) for all i < j with di = dj
(3.3)
to make σ unique. In other words, σ has to preserve the order of repeated
divisors. If even stronger,
σ(i) < σ(j) for all i < j with gcd(di , dj ) > 1,
(3.4)
then we call d and e associated. Any two complete refinements d∗ of d and e∗
of e, respectively, are associated and we have from (3.1)
Dn,{d∗ ,e∗ } ⊆ Dn,{d,e} .
(3.5)
We ask for associated refinements d∗ and e∗ , respectively, that describe the
same set of collisions as d and e. Algorithm 3.6 solves this task and returns
“coarsest” associated refinements d∗ and e∗ that yield equality in (3.5). We
call the output d∗ of Algorithm 3.6 the refinement of d by e and denote it
by d∗ = d // e. Similarly, e∗ = e // d is the refinement of e by d and this is
well-defined, since interchanging the order of the input merely interchanges
the order of the output.
Lemma 3.7. For two ordered factorizations d and e of n, the following are
equivalent.
(i) len(d // e) = len(d) and len(e // d) = len(e).
(ii) d // e = d and e // d = e.
(iii) d and e are associated.
Proof. Let d = (d1 , . . . , d` ) and e = (e1 , . . . , e` ) be associated and σ = σ(d, e)
the unique permutation satisfying (3.2) and (3.4). Then we have in step 5 of
Algorithm 3.6
(
gcd(d∗i,m+1 , e∗j,`+1 )
=
di = ej
1
if j = σ(i),
otherwise,
for all 1 ≤ i ≤ ` and 1 ≤ j ≤ m. Thus Algorithm 3.6 returns d, e on input
d, e and (ii) and (i) follow.
12
Algorithm 3.6: Refine d and e
Input: two ordered factorizations d = (d1 , . . . , d` ) and e = (e1 , . . . , em )
of n
Output:
refinements d∗ | d and e∗ | e
two associated
1 . . . 1 d1
1 . . . 1 d2
∗
= (d∗i,j ) 1≤i≤`
1 d ←
..
1≤j≤m+1
.
1 . . . 1 d`
1 . . . 1 e1
1 . . . 1 e2
∗
e ←
= (e∗j,i ) 1≤j≤m
..
1≤m≤`+1
.
2
3
4
5
6
7
8
9
10
11
1 . . . 1 em
for i = 1, . . . , ` do
for j = 1, . . . , m do
c ← gcd(d∗i,m+1 , e∗j,`+1 )
d∗i,j ← c and e∗j,i ← c
d∗i,m+1 ← d∗i,m+1 /c and e∗j,`+1 ← e∗j,`+1 /c
end
end
remove last column (of all 1’s) from d∗ and e∗
d∗ ← (d∗i,j ) 1≤k≤`m /* rewrite row-by-row as a sequence
*/
e∗ ← (e∗j,i )
*/
k=(i−1)m+j
1≤i≤`,1≤j≤m
12
13
14
1≤k≤m`
k=(j−1)`+i
1≤j≤m,1≤i≤`
/* rewrite row-by-row as a sequence
remove all 1’s from d∗ and e∗
return d∗ , e∗
13
Conversely, let d = (d1 , . . . , d` ), e = (e1 , . . . , em ), and ` = len(d) ≥
len(e) = m. We assume that d and e are not associated and define i∗ as the
minimal index 1 ≤ i∗ ≤ ` such that there is no injective map τ : {1, . . . , i∗ } →
{1, . . . , m} with
di = eτ (i) for all 1 ≤ i ≤ i∗ ,
τ (i) < τ (j) for all 1 ≤ i < j ≤ i∗ with gcd(di , dj ) > 1
in analogy to (3.2) and (3.4), respectively.
We have two possible cases for the execution of the inner loop, steps 4-8,
for i = i∗ .
• If c = d∗i∗ ,m+1 in step 5 for some j, then e∗j,`+1 6= d∗i∗ ,m+1 (otherwise, we
could extend some injective τ for i∗ − 1 by i∗ 7→ j) and e∗j,`+1 splits into
at least two nontrivial factors in steps 6 and 7, thus len(e//d) ≥ len(e)+1
and e // d 6= e.
• Otherwise c 6= d∗i∗ ,m+1 in step 5 for all j, and d∗i∗ ,m+1 splits into at least
two nontrivial factors in steps 6 and 7, thus len(d // e) ≥ len(d) + 1 and
d // e 6= d.
Proposition 3.8. Let n be a positive integer and d, e two ordered factorizations of n with length ` and m, respectively. Then the following holds.
(i) Algorithm 3.6 works as specified and requires O(`m) gcd-computations
and O(`m) additional integer divisions.
(ii) We have Dn,{d,e} = Dn,{d//e,e} = Dn,{d//e,e//d} .
Proof. If we ignore the last column of d∗ and e∗ , respectively, we obtain
matrices that are each other’s transpose before and after every execution of
the inner loop in Algorithm 3.6. We use this property to define a sequence
of integer matrices M(k) ∈ Z(`+1)×(m+1) for 0 ≤ k ≤ `m to simultaneously
capture d∗ and e∗ after the inner loop has bee executed k times.
For k = 0, let
1
1
M(0) =
1
e1
1
1
...
...
..
.
1
1
1 ... 1
e2 . . . em
14
d1
d2
d`
1
(0)
= (mi,j ) 1≤i≤`+1
1≤j≤m+1
and for k = (i − 1)m + j > 0, 1 ≤ i ≤ `, 1 ≤ j ≤ m, we define M(k) as M(k−1)
(k−1)
(k−1)
(k−1)
(k−1)
with mi,j replaced by c, mi,m+1 replaced by mi,m+1 /c, and m`+1,j replaced
(k−1)
by m`+1,j /c, respectively. Thus, we have the following invariants. For every
0 ≤ k ≤ `m and every 1 ≤ i ≤ `, the ith row of M(k) is a factorization of di .
Analogously, for every 1 ≤ j ≤ m, the jth column of M(k) is a factorization
of ej .
We have d∗ as M(k) with the last row removed, and e∗ as M(k) with the
last column removed and then transposed. In particular, the output d // e is
the first ` rows of M`m read as a sequence with 1’s ignored. Analogously,
the output e // d is the first m columns of M`m read as a sequence with 1’s
ignored.
(i) By the invariants of M(k) mentioned above, the output d // e is a
refinement of the input d. Analogously, the output e // d is a refinement of the
input e. The outputs also have the same basis, namely the entries of M(`m)
different from 1.
A bijection σ on {1, . . . , `m} is given by k = (i − 1)m + j 7→ (j − 1)` + i
with 1 ≤ i ≤ `, 1 ≤ j ≤ m. And this satisfies (3.2) since dk = ci,j = eσ(k)
for all such i, j. We also show (3.3) for σ. Let 1 ≤ k < k 0 ≤ `m with
k = (i − 1)m + j and k 0 = (i0 − 1)m + j 0 . We have to prove, that if
σ(k) > σ(k 0 ), then gcd(dk , dk0 ) = 1. The condition is equivalent to i < i0 and
j > j 0.
Lemma 3.8a. Let 1 ≤ i ≤ m, 1 ≤ j ≤ `, (i − 1)m + j ≤ k ≤ `m and
Q
(k)
c(k) the state of Algorithm 3.6 after k executions. Let Ri,j = j<j 0 ≤`+1 ci,j 0
Q
(k)
and Bi,j = i<i0 ≤m+1 ci0 ,j . Then gcd(Ri,j , Bi,j ) = 1. In particular, after the
algorithm has terminated, we have gcd(ci,j 0 , ci0 ,j ) = 1 for all i0 > i, j 0 > j.
Proof of Lemma 3.8a. Concentrate on the element ci,j 0 . By construction
Q
Q
k>j 0 ci,k and
k>i ck,j 0 are coprime. In particular, their factors ci,j and
ci0 ,j 0 .
This shows that d∗ and e∗ are associated if we restrict σ to indices k with
dk > 1.
Finally, the only arithmetic costs are the gcd-computations in step 5 and
the integer divisions in step 7.
(ii) We begin with the first equality. The matrix d∗ corresponds to
an ordered factorization, when read row-by-row and 1’s ignored. Let d(k)
correspond to the state of the matrix d∗ after the inner loop has been executed
exactly k times for 0 ≤ k ≤ `m. Thus d(0) = d, d(`m) = d // e, and we show
inductively
Dn,{d(k) ,e} = Dn,{d(k+1) ,e}
(3.8b)
15
for all 0 ≤ k < `m.
Let k + 1 = (i − 1)m + j with 1 ≤ i ≤ `, 1 ≤ j ≤ m. If c = 1 in step 5,
then d(k+1) = d(k) and (3.8b) holds trivially. Otherwise c > 1, and d(k+1) is
the proper refinement of d(k) , where the entry d∗i,m+1 in d(k) is replaced by the
pair (c, d∗i,m+1 /c).
We have to show that if a polynomial has decomposition degree sequences
d(k−1) and e, then it also has decomposition degree sequence d(k) . This follows
from the following generalization of Proposition 2.9.
Lemma 3.8c. Let g1 ◦ g2 ◦ · · · ◦ g` = h1 ◦ h2 ◦ · · · ◦ hm be two decompositions
with degree sequence d and e, respectively. Let 1 ≤ i ≤ `, 1 ≤ j ≤ m,
c = gcd(di , ej ), and
gcd(d1 · . . . · di−1 · di , e1 · . . . · ej−1 ) = gcd(d1 · . . . · di−1 , e1 · . . . · ej−1 · ej ).
(3.8d)
Then there are unique monic original polynomials u and v of degree c and
di /c, respectively, such that
gi = u ◦ v.
(3.8e)
Therefore, if a monic original polynomial f has decomposition degree sequences
d and e, then it also has decomposition degree sequence d∗ = (d1 , . . . , di−1 , c, di /c, di+1 , . . . , d` ).
Proof of Lemma 3.8c. Let A = g1 ◦ · · · ◦ gi−1 , B = h1 ◦ · · · ◦ hj−1 , and
b = gcd(deg(A), deg(B)). Then (3.8d) reads gcd(deg(A ◦ gi ), deg(B)) =
gcd(deg(A), deg(B◦hj )). This implies gcd(deg(B)/ gcd(deg(A), deg(B)), deg(gi )) =
gcd(deg(A)/ gcd(deg(A), deg(B)), deg(hj )) and since the first arguments of
both outer gcd’s are coprime, this quantity is 1. This proves
gcd(deg(A ◦ gi ), deg(B ◦ hj )) = gcd(deg(A), deg(B)) · gcd(deg(g), deg(h))
deg(A)
deg(hj )
· gcd(
,
)
gcd(deg(A), deg(B)) gcd(deg(gi , hj ))
· gcd(
deg(B)
deg(gi )
,
)
gcd(deg(A), deg(B)) gcd(deg(gi , hj ))
= gcd(deg(A), deg(B)) · gcd(deg(gi ), deg(hj ))
= bc.
(3.8f)
Then Proposition 2.9 applied to left components of the bi-decompositions
A ◦ (gi ◦ · · · ◦ g` ) = B ◦ (hj ◦ · · · ◦ hm )
16
(3.8g)
guarantees the existence of unique, monic original C, A0 , B 0 with deg(C) = b
and gcd(deg(A0 ), deg(B 0 )) = 1, such that
A = C ◦ A0 and B = C ◦ B 0 .
(3.8h)
We substitute (3.8h) back into (3.8g), ignore the common left component C
due to the absence of equal-degree collisions, and write with the associativity
of composition
(A0 ◦ gi ) ◦ (gi+1 ◦ · · · ◦ g` ) = (B 0 ◦ hj ) ◦ (hj+1 ◦ · · · ◦ hm ).
From (3.8f), we have gcd(deg(A0 ◦ gi ), deg(B 0 ◦ hj )) = gcd(di , ej ) = c. With
Proposition 2.9, we obtain some monic original w and A00 of degree c and
deg(A0 ◦ gi )/c = deg(A0 ) · di /c, respectively, such that
A0 ◦ gi = w ◦ A00 .
(3.8i)
We have gcd(deg(gi ), deg(A00 )) = di /c and a final application of Proposition 2.9
to the right components of (3.8i) provides the decomposition for g, claimed
in (3.8e).
To apply this result with d = d(k) and e, we have to provide (3.8d). For
(k + 1) = (i − 1)m + j, we split D = d1 · . . . · di−1 and E = e1 · . . . · ej−1 into
their common (left upper subset) C and the remainders R and B, respectively.
By Lemma 3.8a, we have gcd(R, B) = 1 and therefore gcd(D, E) = C. The
same lemma shows gcd(di , B) = 1 = gcd(R, ej ) and we have
gcd(Ddi , E) = C gcd(di , B) = C gcd(R, ej ) = gcd(D, Eej ),
as required for (3.8b).
Finally, interchanging the rôles of d and e yields
Dn,{d,e} = Dn,{d//e,e} = Dn,{d,e//d} = Dn,{d//e,e//d,d,e} = Dn,{d//e,e//d} ,
since composition degree sequence d // e implies d and similarly e // d implies
e.
Example 3.9. Let n = 7! = 5040, d = (12, 420), and e = (14, 360). We have
as refinements
d // e = (2, 6, 7, 60) ,
e // d = (2, 7, 6, 60) ,
(3.10)
and any f ∈ Dn,{d,e} has a unique decomposition f = a ◦ g ◦ b with a ∈ P2 ,
g ∈ D42,{(6,7),(7,6)} , and b ∈ P60 by Proposition 2.9.
17
Given a set D with more than two ordered factorizations, we repeatedly
replace pairs d, e ∈ D by d//e and e//d, respectively, until we reach a refinement
D∗ invariant under this operation. This process terminates by Lemma 3.7.
The result depends on the order of the applied refinements, but any order
ensures the desired properties described by the following proposition.
Proposition 3.11. Let n be a positive integer and D a set of c ordered
factorizations of n. There is a set D∗ of at most c ordered factorizations of n
with the following properties.
(i) All ordered factorizations of D∗ are pairwise associated.
(ii) Dn,D∗ = Dn,D .
(iii) D∗ can be computed from D with at most O(c2 ) calls to Algorithm 3.6.
Proof. For c = 1, we have D = {d} and D∗ = {d} satisfies all claims.
For c = 2, we have D = {d, e} for ordered factorizations d 6= e, and
D∗ = {d // e, e // d} satisfies all claims by Proposition 3.8.
Let c > 2 and D = {d(1) , . . . , dc }. By induction assumption, we can
assume all d(i) for 1 ≤ i < c − 1 be pairwise associated. Let d(c) = f and
D∗ = {d(1) // f, d(2) // f, . . . , d(c−1) // f, f // d(1) }. Clearly Dn,D = Pn,D∗ and it
remains to show that all elements of D∗ are pairwise associated.
By construction, we have d(1) // f associated with f // d(1) and by transitivity
of associatedness the following lemma suffices.
Lemma 3.11a. If d∗ and e∗ are associated, then so are d∗ // f and e∗ // f for
any factorization f.
Proof. Let σ = σ(d∗ , e∗ ) and compare the matrices
M = M(d∗ , f) and N = M(e∗ , f).
The claimed bijection between the indices of M and N is given by mapping
row i to row σ(i) (followed by identity on the columns).
Assume for contradiction that i is the minimal row index such that
Mi,∗ 6= Nσ(i),∗ and j is the minimal column index such that Mi,j 6= Nσ(i),j .
Let Nσ(i),j = aMi,j with a > 1. Then there is a column j 0 > j, such that
a | Mi,j 0 , since the rows Mi,∗ and Nσ(i),∗ are both factorizations of di = eσ(i) .
Also there is a row i0 > i, such that a | Mi0 ,j , since the earlier occurrences of
a in that column are pairwise matched.
By Lemma 3.8a, this is a contradiction. And analogously, if Mi,j = aNσ(i),j
with a > 1.
18
Any D∗ satisfying Proposition 3.11(i)-(ii) is called a normalization of D.
For a normalized D = {d(k) : 1 ≤ k ≤ c, we have the same basis d(k) for all
1 ≤ k ≤ c and call this multiset the basis of D, denoted by D.
Example 3.12. We add the ordered factorization f = (20, 252) to D = {d, e}
of Example 3.9 and obtain from (3.10) through refinement with f
d∗ = (d // e) // f = (2, 2, 3, 7, 5, 12) ,
e∗ = (e // d) // f = (2, 7, 2, 3, 5, 12) ,
f ∗ = (f // d) // e = (2, 2, 5, 3, 7, 12) .
(3.12b)
Any f ∈ Pn,{d,e,f} = Pn,{d∗ ,e∗ ,f ∗ } has a unique decomposition f = a ◦ g ◦ b
with a ∈ P2 , g ∈ D210,{(2,3,7,5),(7,2,3,5),(2,5,3,7)} , and b ∈ P12 . The normalized set
{d∗ , e∗ , f ∗ } has basis {2, 2, 3, 5, 7, 12}.
3.2
The relation graph of D
An ordered factorization d = (d1 , d2 , . . . , d` ) defines a relation ≺d on its basis
d = {d1 , d2 , . . . , d` } by
di ≺d dj for 1 ≤ i < j ≤ `.
In other words, di ≺d dj if di appears before dj in the ordered factorization d,
where we distinguish between repeated factors in the multiset d. We define
the relation graph Gd as directed graph with
• vertices d = {d1 , d2 , . . . , d` } and
• directed edges (dj , di ) = di ← dj for di ≺d dj .
This graph is a transitive tournament, that is a complete graph with directed
edges, where a path d ← e ← f implies an edge d ← f for any vertices
d, e, f ∈ d.
Now, let D = {d(1) , d(2) , . . . , d(c) } be a normalized set of c ordered factorizations with common basis D = {d1 , d2 , . . . , d` }. The relation ≺D is the union
of the relations ≺d(k) for 1 ≤ k ≤ c and the relation graph GD is the union
of the relation graphs Gd(k) for 1 ≤ k ≤ c. The undirected graph underlying
GD is still complete, but may be intransitive. See Figure 1 for the relation
graphs of Example 3.9 and Example 3.12.
We can express the relation ≺D with the permutations (3.2). Let σk =
σ(d(1) , d(k) ) for 1 ≤ k ≤ c. Then σ1 is the identity on 1, 2, . . . , ` and we have
di ≺d(k) dj
19
2
6
2'
2
12
3
7
60
5
7
Figure 1: Relation graphs of (3.10) and (3.12b); in the latter, 20 denotes the
first 2 in each ordered factorization.
if and only if σk (i) < σk (j).
A Hamiltonian path e = e1 ← · · · ← e` in a graph G visits each vertex
exactly once. We call e transitive, if its transitive closure is a subgraph of G.
In other words, e is transitive if ei ← ej is an edge in G for all 1 ≤ i < j ≤ `.
Q
For a relation graph G with vertices d1 , d2 , . . . , d` and n = 1≤i≤` di , we define
DG = {f ∈ Pn : for every transitive Hamiltonian path e1 ← · · · ← e`
in G, there is a decomposition f = g1 ◦ g2 ◦ · · · ◦ g`
with deg gi = ei for 1 ≤ i ≤ `}.
If G = {d} is a singleton, we have DG = Pd .
Proposition 3.13. Let n be a positive integer, D a normalized set of ordered
factorizations of n, and G the relation graph of D. We have
Dn,D = DG .
Proof. Every transitive tournament Gd for d ∈ D, has d as its unique transitive
Hamiltonian path. Since G is the union of all such Gd , we have “⊇”.
For “⊆”, we have to show that every polynomial with decomposition
degree sequences D also has decomposition degree sequence d∗ for every
transitive Hamiltonian path d∗ in G. We proceed on two levels. First, we
derive all transitive Hamiltonian paths in G from “twisting” the paths given
by D. Second, we show that the corresponding “twisted” decomposition
degree sequences follow from the given ones.
20
di
di
di +2
di−1
di +2
di−1
di +1
di +1
Figure 2: A “swap” between two transitive Hamiltonian paths di−1 ← di ←
di+1 ← di+2 and di−1 ← di ← di+1 ← di+2 along the bidirectional edge
between di and di+1 .
Let d∗ be a transitive Hamiltonian path in G and d ∈ D arbitrary. We
use Bubble-Sort to transform d into d∗ and call the intermediate states
after k passes d(k) , 0 ≤ k ≤ c, such that d(0) = d and d(c) = d∗ .
Algorithm 3.14: Bubble-Sort d according to d∗
1 ` ← len(d)
(0)
2 k ← 0, d
←d
(k)
3 while d
6= d∗ do
4
k ← k + 1, d(k) ← d(k−1) /* copy previous state
*/
5
for i = 1, . . . , ` − 1 do
6
σ = σ(d(k) , d∗ )
7
if σ(i) > σ(i + 1) then
(k) (k)
(k)
(k)
8
(di , di+1 ) ← (di+1 , di ) /* swap
*/
9
end
10
end
11 end
12 c ← k
In other words, d(k) is obtained from d(k−1) by at most ` − 1 “swaps” of
(k)
(k)
adjacent vertices. Figure 2 visualizes a swap of di and di+1 as in step 8.
The fundamental properties of Bubble-Sort guarantee correctness and
c ≤ `(` − 1)/2, see Cormen, Leiserson, Rivest & Stein (2009, Problem 2.2).
Furthermore, the following holds.
(k)
(k)
(i) Every pair (di , di+1 ) of swapped vertices in step 8 is connected by a
bidirectional edge in G.
(ii) Every d(k) , 0 ≤ k ≤ c, is a transitive Hamiltonian path in G.
21
(k)
For (i), we have the edge di
(k)
(k)
← di+1 from d(k−1) and the edge d∗σ(i+1) =
(k)
di+1 ← di = d∗σ(i) from d∗ with σ as in step 6.
For k = 0, (ii) holds by definition. For k > 0 it follows inductively from
k − 1, since a swap merely replaces the 4-subpath di−1 ← di ← di+1 ← di+2
by di−1 ← di+1 ← di ← di+2 , where the outer edges are guaranteed in G by
transitivity of d(k−1) and the inner edge by (i). Thus, the swapped path is
also a transitive Hamiltonian path in G.
Now, we mirror the “swaps” of vertices by “Ritt moves” of components as
introduced by Zieve & Müller (2008).
Claim 3.13a (Ritt moves). Let g1 ◦· · ·◦g` = h1 ◦· · ·◦h` be decompositions with
degree sequence d and e, respectively. Let d and e be associated, σ = σ(d, e),
and 1 ≤ i < ` with σ(i) > σ(i + 1). Then
∗
gi ◦ gi+1 = gi∗ ◦ gi+1
∗
with deg(gi ) = deg(gi+1
) and deg(gi+1 ) = deg(gi∗ ). Therefore, if some monic
original polynomial f has decomposition degree sequences d and e, it also has
the decomposition degree sequence d∗ = (d1 , . . . , di−1 , di+1 , di , di+2 , . . . , d` ).
The claim is based on the following lemma.
Lemma 3.13b. Let d and e be associated ordered factorizations, σ = σ(d, e),
1 ≤ i ≤ len(d), and j = σ(i). Then
gcd(d1 · . . . · di−1 , e1 · . . . · ej−1 ) = gcd(d1 · . . . · di−1 · di , e1 · . . . · ej−1 )
= gcd(d1 · . . . · di−1 , e1 · . . . · ej−1 · ej ).
In particular, (3.8d) holds.
Proof of Lemma 3.13b. For any 1 ≤ k < j, with gcd(ek , ej ) = gcd(ek , di ) > 1,
we have σ −1 (k) < i due to (3.4). In other words, σ −1 maps all indices
1 ≤ k < j, where gcd(ek , di ) > 1, into the set {1, . . . , i − 1}. Therefore
e1 · . . . · ej−1
gcd(
, di ) = 1,
gcd(d1 · . . . · di−1 , e1 · . . . · ej−1 )
gcd(d1 · . . . · di−1 · di , e1 · . . . · ej−1 ) = gcd(gcd(d1 · . . . · di−1 , e1 · . . . · ej−1 )di , e1 · . . . · ej−1 )
= gcd(d1 · . . . · di−1 , e1 · . . . · ej−1 ).
Let j 0 = σ(i + 1) < σ(i) = j, A = g1 ◦ · · · ◦ gi−1 , C = gi+2 ◦ · · · ◦ g` ,
A = h1 ◦ · · · ◦ hj 0 −1 , B 0 = hj 0 +1 ◦ · · · ◦ hj−1 , and C 0 = hj+1 ◦ · · · ◦ h` , such that
0
A ◦ gi ◦ gi+1 ◦ C = A0 ◦ hj 0 ◦ B 0 ◦ hj ◦ C 0 .
22
Lemma 3.13b for i and i + 1 yields
gcd(deg(A ◦ gi ), deg(A0 ◦ hj 0 ◦ B 0 )) = gcd(deg(A), deg(A0 ◦ hj 0 ◦ B 0 )),
gcd(deg(A ◦ gi ◦ gi+1 ), deg(A0 )) = gcd(deg(A ◦ gi ), deg(A0 ◦ hj 0 )),
respectively. From the former, we derive
deg(A0 ◦ hj 0 ◦ B 0 )
)
gcd(deg(A), deg(A0 ◦ hj 0 ◦ B 0 ))
deg(A0 ◦ hj )
= gcd(gi ,
).
gcd(deg(A), deg(A0 ◦ hj 0 ))
1 = gcd(gi ,
And then continue the latter as
gcd(deg(A ◦ gi ◦ gi+1 ), deg(A0 ))
= gcd(deg(A ◦ gi ), deg(A0 ◦ hj 0 ))
= gcd(deg(A), deg(A0 ◦ hj 0 )) · gcd(gi ,
= gcd(deg(A), deg(A0 ◦ hj 0 )).
deg(A0 ◦ hj 0 )
)
gcd(deg(A), deg(A0 ◦ hj 0 ))
(3.13c)
Let G = gi ◦ gi+1 and H = hj . We have gcd(di , di+1 ) = 1 due the “twisting
condition” σ(i + 1) < σ(i) and therefore gcd(deg(G), deg(H)) = di+1 . We
apply Lemma 3.8c with gi = G, hj = H, and c = di+1 in the notation of that
claim, and find, since (3.13c) provides condition (3.8d),
∗
G = gi∗ ◦ gi+1
∗
with deg(gi∗ ) = di+1 and deg(gi+1
) = di as required.
Repeated application of Claim 3.13a shows that for every f ∈ Dn,D ,
d ∈ D, and every transitive Hamiltonian path d∗ in G, we have d(k) as in
Algorithm 3.14 as decomposition degree sequence. In particular, d(c) = d∗ .
3.3
The Decomposition of Dn,D
Every directed graph admits a decomposition into strictly connected components, where any two distinct vertices are connected by paths in either
direction. Since a relation graph G is the union of directed complete graphs,
its strictly connected components Gi , 1 ≤ i ≤ `, are again relation graphs
and form a chain G1 ← G2 ← · · · ← G` . Figure 3 shows the connected
components of the relation graphs Figure 1
23
6
60
2
7
3
2'
7
2
12
5
Figure 3: The three strongly connected components of each relation graph in
Figure 1, respectively.
24
Theorem 3.14. Let G be a relation graph with strongly connected components
G1 ← G2 ← · · · ← G` . We have
DG = DG1 ◦ DG2 ◦ · · · ◦ DG`
and for any f ∈ DG , we have uniquely determined gi ∈ DGi such that
f = g1 ◦ g2 ◦ · · · ◦ g` . Furthermore, over a finite field F = Fq with q elements,
we have
Y
#DG =
#DGi .
1≤i≤`
Proof. For f ∈ Pn , where n =
equivalent.
Q
v∈G
v, we show that the following are
(i) The polynomial f has decomposition degree sequence d for every transitive Hamiltonian path d in G.
(ii) The polynomial f has decomposition degree sequence d = d1 ← d2 ←
· · · ← d` for every concatenation of transitive Hamiltonian paths di in
Gi for 1 ≤ i ≤ `.
Assume (i) and let d = d1 ← d2 ← · · · ← d` be the concatenation of
transitive Hamiltonian paths di in Gi for 1 ≤ i ≤ ` as in (ii). Then di is
a Hamiltonian path in G. Since the underlying undirected graph of G is
complete, we have di ← dj in G for any vertices di ∈ Gi and dj ∈ Gj in
distinct strictly connected components with i < j. Thus d is also transitive
and f has decomposition degree sequence d by (i).
Conversely, assume (ii) and observe that the decomposition of G into
strictly connected components induces a decomposition of every transitive
Hamiltonian path d in G into Hamiltonian paths di in Gi . These are transitive,
since transitivity is a local condition and f has decomposition degree sequence
d by (ii).
Uniqueness and thus the counting formula follow from the absence of
equal-degree collisions in the tame case.
We split the edge set E of a strictly connected relation graph G with
→
−
vertices V into its uni-directional edges E = {(u, v) ∈ E : (v, u) ∈
/ E} and its
→
−
bi-directional edges (2-loops) E = {{u, v} ⊆ V : {(u, v), (v, u)} ∈ E} = E \ E .
We call the corresponding graphs on V the directed and the undirected
subgraph of G, respectively. The directed subgraph of G is a directed acyclic
graph since G is the union of transitive tournaments. The undirected subgraph
of G is connected. It is also the union of the permutation graphs of σk ,
1 ≤ k ≤ c.
25
3
7
2
5
Figure 4: The strongly connected component on 4 vertices of Figure 3
decomposed into its undirected subgraph (red) and its directed subgraph
(blue) with Max-Sink-sorting 7 ≺ 2 ≺ 5 ≺ 3.
26
→
−
The directed subgraph G captures the requirements on the position
of the degrees in a decomposition sequence. The undirected subgraph G
captures the admissible Ritt moves d’après Zieve & Müller (2008) and thus
the requirements on the shape of the components.
Every directed acyclic graph admits a topological sorting v1 , v2 , . . . , v` of
→
−
its vertices, where a directed edge vi ← vj in G implies i < j, see Cormen
et al. (2009, Section 22.4). A directed acyclic graph may have several distinct
topological sortings. Tarjan (1976) suggested to use Depth-First-Search
→
−
on G . The time step, when Depth-First-Search visits a vertex for the
last time, is called the finish time of the vertex and listing the vertices with
increasing finish time yields a topological sorting. The result is unique, if the
tie-break rule for expanding in Depth-First-Search is deterministic. We
use the following terminology.
Let U (v) denote the open G-neighborhood of a vertex v. It is always
nonempty. We call a vertex v locally maximal, if its value is greater or equal
than the value of every vertex in U (v). Since vertices with equal values are
never connected by an edge in G, a locally maximal v is always strictly greater
than all vertices in U (v). Furthermore, there is at least one locally maximal
vertex, namely a “globally” maximal one. There is a unique enumeration of
the locally maximal vertices d1 , d2 , . . . , dm such that
d1 ← d2 ← · · · ← dm
is a directed path in G. Furthermore, we define for 1 ≤ i ≤ m,
Vi = U (di ) \ U (di+1 ) and Wi = Vi ∪ {di },
V0 = W0 = {v ∈ G : no edge di ← v in G for any 1 ≤ i ≤ m},
Vm+1 = Wm+1 = {v ∈ G : no edge v ← di in G for any 1 ≤ i ≤ m}.
→
−
The Wi , 0 ≤ i ≤ m + 1, form a partition of all vertices of G and we formulate
the tie-break rule for Depth-First-Search as follows. Given vertices
u ∈ Wi and v ∈ Wj with i < j, the vertex u is preferred. Given vertices
u, v ∈ Wi , the vertex with the larger value is preferred. Since vertices with
→
−
equal value are always connected by a unidirectional edge in G due to (3.4),
the search has never to choose between to vertices with the same value and
Depth-First-Search with this tie-break rule yields a unique topological
→
−
sorting. We call it the Max-Sink topological sorting of G . Figure 4 shows
the largest strongly connected component of Figure 3 and its Max-Sink
topological sorting.
27
Theorem 3.15. Let G be a strongly connected relation graph with at least two
→
−
vertices, directed subgraph G , and undirected subgraph G. Let d1 , d2 , . . . , d`
→
−
be the Max-Sink topological sorting of G and let ei be the product of all
vertices in the open G-neighborhood of di . For every f ∈ DG either (i) or (ii)
holds, and (iii) is also valid.
(i) (Exponential Case) There are unique gi ∈ Edi ,ei for 1 ≤ i ≤ ` and a ∈ F
such that
f = (g1 ◦ g2 ◦ · · · ◦ g` )[a] .
(ii) (Trigonometric Case) There are unique z, a ∈ F with z 6= 0 such that
f = Td1 d2 ···d` (x, z)[a] .
(iii) If G contains no edge that connects two vertices both larger than 2, then
the Trigonometric Case is included in the Exponential Case. Otherwise,
they are mutually exclusive.
Conversely,
[F ]
Dn,G = Td1 d2 ···dm ,1 ∪ (Ed1 ,e1 ◦ Ed2 ,e2 ◦ · · · ◦ Edm ,em )[F ] .
(3.15a)
The ei are well-defined, since there are no empty neighborhoods in the
connected graph G with at least two vertices.
Proof. We begin with the proof of existence, then show uniqueness and
conclude with the “converse” (3.15a).
→
−
The Max-Sink topological sorting d1 , d2 , . . . , d` of G yields a transitive
Hamiltonian path
d = d1 ≺ d2 ≺ · · · ≺ d`
in G. For the rest of the proof, we identify the max-sink topological sorting
with the corresponding transitive Hamiltonian path.
We (re)label the locally maximally vertices d1 , d2 , . . . , dm and the elements
(1) (2)
(` )
of Vi as di , di , . . . , di i for 0 ≤ i ≤ m + 1 and `i = #Vi such that
(1)
(` )
(1)
(` )
(1)
d = (d0 , . . . , d0 0 , d1 , d1 , . . . , d1 1 , d2 , d2 , . . . ,
(`
)
(1)
(`
)
m−1
m+1
(`m )
dm−1
, dm , d(1)
m , . . . , dm , dm+1 , . . . , dm+1 )
= (V0 , d1 , V1 , d2 , . . . , dm , Vm , Vm+1 ),
28
(1)
(2)
(` )
where the Vi are read as tuple (di , di , . . . , di i ). Then f has a decomposition
(1)
(` )
(1)
(` )
(1)
f = G0 ◦ · · · ◦ G0 0 ◦ G1 ◦ G1 ◦ · · · ◦ G1 1 ◦ G2 ◦ G2 ◦ . . .
◦
(`m−1 )
Gm−1
(`m )
◦ Gm ◦ G(1)
◦
m ◦ · · · ◦ Gm
(1)
Gm+1
◦ ··· ◦
(3.15b)
(`m+1 )
Gm+1
(j)
with Gi ∈ Pd(j) for 0 ≤ i ≤ m + 1, 1 ≤ j ≤ `i , and Gi ∈ Pdi for 1 ≤ i ≤ m.
i
We assume for the moment that all edges in G contain a 2. Then Theorem 2.8 reduces to the exponential case, and we proceed as follows. First, we
(j)
[a ]
show that every Gi for 1 ≤ i ≤ m, 1 ≤ j ≤ `i , is of the form gi i for unique
gi ∈ Ed(j) ,e(j) and unique ai ∈ F . Then, we extend this to i = 0 and i = m + 1.
i
i
Finally, we show that the shifting parameters ai are “compatible” such that a
single shifting parameter a suffices.
For every 1 ≤ i ≤ m, we use Bubble-Sort Algorithm 3.14 with
Lemma 3.16 to obtain the decomposition degree sequence
(`i+1 )
(V0 , d1 , V1 , . . . , di , Vi , di
(`
(mi )
, . . . di
)
, di+1 , V̂i+1 , . . . , dm , V̂m , Vm+1 ),
(m )
where U (di ) = Vi ∪ {di i+1 , . . . di i } and the latter elements have been
Q
(j)
omitted from Vi+1 , . . . , Vm . We have ei = 1≤j≤mi di and this implies the
two decomposition degree sequences
(V0 , d1 , V1 , . . . , di , ei , di+1 , V̂i+1 , . . . , dm , V̂m , Vm+1 ), (V0 , d1 , V1 , . . . , ei , di , di+1 , V̂i+1 , . . . , dm , V̂m , Vm+1
Thus, there are unique gi ∈ Edi ,ei and ai ∈ F such that in (3.15b), we have
(1)
(`i )
Gi ◦ Gi ◦ · · · ◦ Gi
(1)
(`i )
= (gi ◦ xdi ◦ · · · ◦ x(di
) [ai ]
)
(j)
.
The same form applies to i = 0, since there is at least one di with 1 ≤ i ≤ m,
1 ≤ j ≤ `i that is in the G-neighborhood of some element of V0 due to the
strong connectedness of G. And since there is no locally maximal element
(j)
in V0 all components are of the form xd0 with possible some shift applied.
Every connection in G relates the corresponding shifting parameters and since
G has a Hamiltonian path, they are all determined by a single choice.
Now, for the general case, where some collisions may be trigonometric,
but not exponential. For any two locally maximal vertices di and dj there is
some vertex d ∈ U (di ) ∩ U (dj ). This shows, that either all blocks fall into the
exponential case or all blocks fall into the trigonometric case. The two cases
are disjoint if and only if there is some edge in G that connects two vertices
both with value greater than 2.
29
The stabilizer of original shifting is {0} for nonlinear monic original
polynomials and there are no equal-degree collisions. Hence the representation
is unique.
The converse (3.15a) is a direct computation.
Lemma 3.16. Let i < j and dj in the open G-neighborhood of di . Then for
every i < k < j, we have dk in the open G-neighborhood of di or dj or both.
Proof. The tournaments underlying G are acyclic. Therefore, if di ≺ dk ≺ dj
and dj ≺ dk , then at least one other edge is bidirectional, too.
4
Exact Counting of Decomposable Polynomials
The classification of Theorem 3.15 yields the exact number of decomposable
polynomials at degree n over a finite field Fq .
Theorem 4.1. Let G be a strongly connected relation graph with undirected
subgraph G. Let d1 , d2 , . . . , d` be the vertices of G and ei be the product of all
vertices in the (open) G-neighborhood of di . Let δG,2 be 1 if there is no edge
in G between two vertices both larger than 2 and let δG,2 be 0 otherwise. Then
#DG =
q d−1
Q
q · (
if G = {d},
di ∈G
q bdi /ei c + (1 − δG,2 ) · (q − 1)) otherwise.
Proof. For G = {d}, this follows from (2.1a). Otherwise from the (non)uniqueness
of the parameters in Theorem 3.15.
We are finally ready to employ the inclusion-exclusion formula (2.4) from
the beginning. For a nonempty set D of nontrivial divisors of n, it requires
#Dn,D = #Dn,D for D = {(d, n/d) : d ∈ D}. We compute the normalization
D∗ by repeated application of Algorithm 3.6 and derive the relation graph
of D∗ . Then #Dn,D = #PG and the latter follows from Theorem 3.14 and
Theorem 4.1.
This is easy to implement, see Algorithm 4.2, and yields the exact expressions for #Dn (Fq ) at lightning speed, see Table 1. Where no exact expression
was previously known, we compare this to the upper and lower bounds of
von zur Gathen (2014a).
30
Algorithm 4.2: Count Decomposables
Input: positive integer n
Output: #Dn (Fq ) as a polynomial in q for n coprime to q
1 if n = 1 or n is prime then
2
return 0
3 end
4 total ← 0
5 N ← {1 < d < n : d | n}
6 for ∅ =
6 D ⊆ N do
7
D ← {(d, n/d) : d ∈ D}
8
D∗ ← normalization of D
∗
9
G ← relation graph of D
10
collisions ← 1
11
for strongly connected components Gj of G do
12
Gj ← undirected subgraph of Gj
13
if Gj = {d} then
14
connected ← q d
15
else
16
{d1 , d2 , . . . , d` } ← Gj
17
for i = 1, . . . , ` do
18
U ← open neighborhood of di in Gj
Q
19
ei ← v∈U
20
end
Q
21
connected ← di ∈Gj q bdi /ei c
22
if some edge in Gj connects two vertices both larger than 2
then
23
connected ← connected + q − 1
24
end
25
connected ← connected · q
26
end
27
collisions ← collisions · connected
28
end
29
k ← #D
30
total ← total + (−1)k collisions
31 end
32 return total
31
Table 1: Exact values of #Dn (Fq ) in the tame case for composite n ≤ 50,
consistent with the upper and lower bounds (in the last column) or exact
values (no entry in the last column) of von zur Gathen (2014a, Theorem 5.2).
5
Conclusion
We presented a normal form for multi-collisions of decompositions of arbitrary
length with exact description of the (non)uniqueness of the parameters. This
lead to an efficiently computable formula for the exact number of such collisions
at degree n over a finite field of characteristic coprime to p. We concluded
with an algorithm to compute the exact number of decomposable polynomials
at degree n over a finite field Fq in the tame case.
We introduced the relation graph of a set of collisions which may be of
independent interest due to its connection to permutation graphs. It would
be interesting to characterize sets D of ordered factorizations that lead to
identical contributions #Dn,D and to quickly derive #Dn,D∪{e} form #Dn,D
or conversely. Finally, this work deals with polynomials only and the study
of rational functions with the same methods remains open.
6
Acknowledgements
Many thanks go to Joachim von zur Gathen for useful discussions and pointers
to the literature. This work was funded by the B-IT Foundation and the
Land Nordrhein- Westfalen.
References
Roberto M. Avanzi & Umberto M. Zannier (2003). The equation
f (X) = f (Y ) in rational functions X = X(t), Y = Y (t). Compositio Math.
139(3), 263–295. doi:10.1023/B:COMP.0000018136.23898.65.
Eric Bach & Jeffrey Shallit (1997). Algorithmic Number Theory, Vol.1:
Efficient Algorithms. MIT Press, Cambridge MA, second printing edition.
ISBN 0-262-02405-5.
David R. Barton & Richard Zippel (1985). Polynomial Decomposition
Algorithms. Journal of Symbolic Computation 1, 159–168.
Raoul Blankertz, Joachim von zur Gathen & Konstantin Ziegler
(2013). Compositions and collisions at degree p2 . Journal of Symbolic
32
Computation 59, 113–145. ISSN 0747-7171. URL http://dx.doi.org/
10.1016/j.jsc.2013.06.001. Also available at http://arxiv.org/abs/
1202.5810. Extended abstract in Proceedings of the 2012 International
Symposium on Symbolic and Algebraic Computation ISSAC ’12, Grenoble,
France (2012), 91–98.
Arnaud Bodin, Pierre Dèbes & Salah Najib (2009). Indecomposable
polynomials and their spectrum. Acta Arithmetica 139(1), 79–100.
John J. Cade (1985). A New Public-key Cipher Which Allows Signatures.
In Proceedings of the 2nd SIAM Conference on Applied Linear Algebra.
SIAM, Raleigh NC.
Thomas H. Cormen, Charles E. Leiserson, Ronald L. Rivest &
Clifford Stein (2009). Introduction to Algorithms. MIT Press, Cambridge MA, London UK, 3rd edition. ISBN 978-0-262-03384-8 (hardcover),
978-0-262-53305-8 (paperback), 1312 pages .
F. Dorey & G. Whaples (1974). Prime and Composite Polynomials. Journal of Algebra 28, 88–101. URL http://dx.doi.org/10.1016/
0021-8693(74)90023-4.
H. T. Engstrom (1941). Polynomial Substitutions. American Journal of Mathematics 63, 249–255. URL http://www.jstor.org/stable/
pdfplus/2371520.pdf.
Michael D. Fried & R. E. MacRae (1969). On the invariance of chains
of Fields. Illinois Journal of Mathematics 13, 165–171.
Joachim von zur Gathen (1990a). Functional Decomposition of Polynomials: the Tame Case. Journal of Symbolic Computation 9, 281–299. URL
http://dx.doi.org/10.1016/S0747-7171(08)80014-4.
Joachim von zur Gathen (1990b). Functional Decomposition of Polynomials: the Wild Case. Journal of Symbolic Computation 10, 437–452. URL
http://dx.doi.org/10.1016/S0747-7171(08)80054-5.
Joachim von zur Gathen (2002). Factorization and Decomposition of Polynomials. In The Concise Handbook of Algebra, edited by Alexander V.
Mikhalev & Günter F. Pilz, 159–161. Kluwer Academic Publishers.
ISBN 0-7923-7072-4.
Joachim von zur Gathen (2014a). Counting decomposable univariate
polynomials. To appear in Combinatorics, Probability and Computing,
33
Special Issue. Extended abstract in Proceedings of the 2009 International
Symposium on Symbolic and Algebraic Computation ISSAC ’09, Seoul,
Korea (2009). Preprint (2008) available at http://arxiv.org/abs/0901.
0054.
Joachim von zur Gathen (2014b). Normal form for Ritt’s Second Theorem.
Finite Fields and Their Applications 27, 41–71. ISSN 1071-5797. URL http:
//dx.doi.org/10.1016/j.ffa.2013.12.004. Also available at http://
arxiv.org/abs/1308.1135.
Joachim von zur Gathen, Dexter Kozen & Susan Landau (1987).
Functional Decomposition of Polynomials. In Proceedings of the 28th Annual
IEEE Symposium on Foundations of Computer Science, Los Angeles CA,
127–131. IEEE Computer Society Press, Washington DC. URL http:
//dx.doi.org/10.1109/SFCS.1987.29.
Mark William Giesbrecht (1988). Some Results on the Functional
Decomposition of Polynomials. Master’s thesis, Department of Computer
Science, University of Toronto. Technical Report 209/88. Available as
http://arxiv.org/abs/1004.5433.
Johannes Grabmeier, Erich Kaltofen & Volker Weispfenning
(editors) (2003). Computer Algebra Handbook – Foundations, Applications,
Systems. Springer-Verlag, Berlin, Heidelberg, New York. ISBN 3-540-654666. URL http://www.springer.com/978-3-540-65466-7.
Jaime Gutierrez & Dexter Kozen (2003). Polynomial Decomposition.
In Grabmeier, Kaltofen & Weispfenning (2003), section 2.2.4 (pages 26–28).
URL http://www.springer.com/978-3-540-65466-7.
Jaime Gutierrez & David Sevilla (2006). On Ritt’s decomposition
theorem in the case of finite fields. Finite Fields and Their Applications
12(3), 403–412. URL http://dx.doi.org/10.1016/j.ffa.2005.08.004.
Dexter Kozen & Susan Landau (1989). Polynomial Decomposition
Algorithms. Journal of Symbolic Computation 7, 445–456. URL http:
//dx.doi.org/10.1016/S0747-7171(89)80027-6. An earlier version was
published as Technical Report 86-773, Cornell University, Department of
Computer Science, Ithaca, New York, 1986.
Dexter Kozen, Susan Landau & Richard Zippel (1996). Decomposition
of Algebraic Functions. Journal of Symbolic Computation 22, 235–246.
34
H. Levi (1942). Composite Polynomials with coefficients in an arbitrary
Field of characteristic zero. American Journal of Mathematics 64, 389–400.
Alice Medvedev & Thomas Scanlon (2014). Invariant varieties for polynomial dynamical systems. Annals of Mathematics 179(1), 81–177. URL
http://dx.doi.org/10.4007/annals.2014.179.1.2. Also available at
http://arxiv.org/abs/0901.2352v3.
J. F. Ritt (1922). Prime and Composite Polynomials. Transactions of the
American Mathematical Society 23, 51–66. URL http://www.jstor.org/
stable/1988911.
Andrzej Schinzel (1982). Selected Topics on Polynomials. Ann Arbor;
The University of Michigan Press. ISBN 0-472-08026-1.
Andrzej Schinzel (2000). Polynomials with special regard to reducibility.
Cambridge University Press, Cambridge, UK. ISBN 0521662257.
Robert Endre Tarjan (1976). Edge-Disjoint Spanning Trees and DepthFirst Search. Acta Informatica 6, 171–185. URL http://dx.doi.org/10.
1007/BF00268499.
Pierre Tortrat (1988). Sur la composition des polynômes. Colloquium
Mathematicum 55(2), 329–353.
Gerhard Turnwald (1995). On Schur’s Conjecture. Journal of the
Australian Mathematical Society, Series A 58, 312–357. URL http://
anziamj.austms.org.au/JAMSA/V58/Part3/Turnwald.html.
U. Zannier (1993). Ritt’s Second Theorem in arbitrary characteristic.
Journal für die reine und angewandte Mathematik 445, 175–203.
Umberto Zannier (2008). On composite lacunary polynomials and the
proof of a conjecture of Schinzel. Inventiones mathematicae 174, 127–138.
ISSN 0020-9910 (Print) 1432-1297 (Online). URL http://dx.doi.org/10.
1007/s00222-008-0136-8.
Michael E. Zieve & Peter Müller (2008). On Ritt’s Polynomial
Decomposition Theorems. Submitted, URL http://arxiv.org/abs/0807.
3578.
35
| 0math.AC
|
arXiv:1509.02473v2 [math.ST] 3 May 2017
A new non-parametric detector of univariate outliers for
distributions with unbounded support
Jean-Marc Bardet
and
Solohaja-Faniaha Dimby∗
S.A.M.M., Université de Paris 1 Panthéon-Sorbonne
90, rue de Tolbiac, 75634, Paris, France
Abstract
The purpose of this paper is to construct a new non-parametric detector of univariate outliers
and to study its asymptotic properties. This detector is based on a Hill’s type statistic. It satisfies
a unique asymptotic behavior for a large set of probability distributions with positive unbounded
support (for instance: for the absolute value of Gaussian, Gamma, Weibull, Student or regular
variations distributions). We have illustrated our results by numerical simulations which show the
accuracy of this detector with respect to other usual univariate outlier detectors (Tukey, MAD or
Local Outlier Factor detectors). The detection of outliers in a database providing the prices of used
cars is also proposed as an application to real-life database.
Keywords: Outlier detection; order statistics; Hill estimator; non-parametric test.
∗
The authors gratefully acknowledge the enterprise Autobiz
1
Introduction
Let (X1 , · · · , Xn ) be a sample of positive, independent, identically distributed random variables with
unbounded distribution. This article aims to provide a non-parametric outlier detector among the
”large” values of (X1 , · · · , Xn ).
Remark 1. If we wish to detect outliers among the ”small” values of (X1 , · · · , Xn ), it would be possible
to consider max(X1 , · · · , Xn )−Xi instead of Xi , for i = 1, · · · , n. Moreover, if Xi , i = 1, · · · , n, are not
positive random variables, as in the case of quantile regression residuals, we can consider |Xi | instead
of Xi .
There are numerous outlier detectors in such a framework. Generally, such detectors consist of
statistics directly applied to each observation, deciding if this observation can be considered or not
as an outlier (see for instance the books of Hawkins, 1980, Barnett and Lewis, 1994, Rousseeuw and
Leroy, 2005, or the article of Beckman and Cook, 1983). The most frequently used, especially in the
case of regression residuals, is the Student-type detector (see a more precise definition in Section 4).
However, it is a parametric detector that is theoretically defined for a Gaussian distribution. Another
well-known detector is the robust Tukey detector (see for example Rousseeuw and Leroy, 2005). Its
confidence factor is computed from quartiles of the Gaussian distribution, though it is frequently used
for non-Gaussian distributions. Finally, we can also cite the M ADe detector using a confidence factor
computed from the median of absolute value of Gaussian distribution (see also Rousseeuw and Leroy,
2005).
Hence all the most commonly used outlier detectors are based on Gaussian distribution and they are
not really accurate for heavier distributions (for regression residuals, we can also cite the Grubbs-Type
detectors introduced in Grubbs, 1969, extended in Tietjen and Moore, 1972). Such a drawback could
be avoided by considering a non-parametric outlier detector. However, in the literature there are few
non-parametric outlier detectors. We could cite the Local Outlier Factor (LOF) introduced in Breunig
et al. (2000), which is also valid for multivariate outliers. For any integer k, the LOF algorithm compares the density of each point to the density of its k-closest neighbors. Unfortunately a theoretical or
numerical procedure for choosing the number k of cells and its associated threshold still does not exist.
There are also other detectors essentially based on a classification methodology (for instance: Knorr et
al., 2000) or robust statistics (for instance: Hubert and Vendervieren, 2008, for univariate skewed data
and Hubert and Van der Veeken, 2008, for multivariate ones).
An interesting starting point for defining a non-parametric detector of outlying observations is provided
by the order statistics (X(1) , . . . , X(n) ) of (X1 , · · · , Xn ). Thus, Tse and Balasooriya (1991) introduced a
detector based on increments of order statistics, but only for the exponential distribution. Recently, a
procedure based on the Hill estimator was also developed for detecting influential data points in Paretotype distributions (see Hubert et al., 2012). The Hill estimator (see Hill, 1975) has been defined from
the following property: the family of r.v. j log(X(n−j+1) ) − log(X(n−j) ) 1≤j≤k(n) is asymptotically
(when min k(n) , n − k(n) −→ ∞) a sample of independent r.v. following exponential distributions
n→∞
(Rényi exponential representation) for distributions in the max-domain of attraction of Gγ where Gγ
is the cumulative distribution function of the extreme value distribution (see Beirlant et al., 2004).
Here we will use an extension of this property for detecting a finite number of outliers among the
sample (X1 , · · · , Xn ). Indeed, an intuitive idea for detecting them is the following: the presence of
outliers generates a jump in the family of r.v. X(n−j+1) /X(n−j) j , therefore we also see a jump in
the family of the r.v. log(X(n−j+1) ) − log(X(n−j) ) j . Thus an outlying data detector can be obtained
when the maximum of this family exceeds a threshold (see details in (2.8) or (3.3)). In the sequel, see
some assumptions on probability distributions for applying this new test of outlier presence. This also
provides an estimator of the number of outliers. It is relevant to say that this test is not only valid
for Pareto-type distribution (for instance Pareto, Student or Burr probability distributions), but more
generally to a class of regular variations distributions. It can also be applied to numerous probability distributions with an exponential decreasing probability distribution function (such as Gaussian,
Gamma or Weibull distributions). So our new outlier detector is a non-parametric estimator defined
from an explicit threshold, which does not require any tuning parameter and can be applied to a very
large family of probability distributions.
Numerous Monte-Carlo experiments carried out in the case of several probability distributions attest to
the accuracy of this new detector. It is compared to other famous outlier detectors or extended versions
of these detectors and the simulation results obtained by this new detector are convincing especially
since it ignores false outliers. Moreover, an application to real-life data (price, mileage and age of used
cars) is done, allowing the detection of two different kinds of outliers.
We have drafted our paper along following lines. Section 2 contains the definitions and several probabilistic results while Section 3 describes how to use them to build a new outlier detector. Section 4
is devoted to Monte-Carlo experiments, Section 5 presents the results of the numerical application on
used car variables and the proofs of this paper are to be found in Section 6.
2
Definition and first probabilistic results
For (X1 , · · · , Xn ) a sample of positive i.i.d.r.v. with unbounded distribution, define:
G(x) = P(X1 > x)
for x ∈ R.
(2.1)
It is clear that G is a decreasing function and G(x) → 0 when x → ∞. Hence, define also the pseudoinverse function of G by
G−1 (y) = sup{x ∈ R, G(x) ≥ y}
y ≥ 0.
(2.2)
G−1 is also a decreasing function. Moreover, if the support of the probability distribution of X1 is
unbounded then G−1 (x) → ∞ when x → 0.
Now, we consider both the following spaces of functions:
n
f2 (α)
1
• A1 = f : [0, 1] → R, such as for any α > 0, f (αx) = f1 (x) 1 +
+O
when
log(x)
log2 (x)
o
x → 0 where f1 : [0, 1] → R satisfies limx→0 f1 (x) = ∞ and f2 is a C 1 ([0, ∞)) diffeomorphism .
n
• A2 = g : [0, 1] → R, there exist a > 0 and a function g1 : [0, 1] → R satisfying limx→0 g1 (x) = ∞,
o
1
and for all α > 0, g(αx) = α−a g1 (x) 1 + O log(x)
when x → 0 .
Example 2.1. We will show below that numerous famous ”smooth” probability distributions such as
absolute values of Gaussian, Gamma or Weibull distributions satisfy G−1 ∈ A1 . Moreover, numerous
heavy-tailed distributions such as Pareto, Student or Burr distributions have G−1 ∈ A2 .
Using the order statistics X(1) ≤ X(2) ≤ · · · ≤ X(n) , define the following ratios (τj ) by:
τj =
X(j+1)
X(j)
if X(j) > 0, and τj = 1 if not, for any j = 1, · · · , n − 1
τj0 = (τj − 1) log(n)
for any j = 1, · · · , n − 1
(2.3)
(2.4)
In the sequel, we are going to provide some probabilistic results on the maximum of these ratios.
Proposition 1. Assume G−1 ∈ A1 . Then, for any J ∈ N∗ , and with (Γi )i∈N∗ a sequence of r.v.
satisfying Γi = E1 + · · · + Ei for i ∈ N∗ where (Ei )i∈N∗ is a sequence of i.i.d.r.v. with exponential
distribution of parameter 1,
D
0
max {τn−j
} −→
j=1,··· ,J
max
n→∞ k=1,··· ,J
f2 (Γk+1 ) − f2 (Γk ) .
(2.5)
Here f2 corresponds to G−1 in the manner described in the definition of A1 .
Now, we consider a particular case of functions belonging to A1 . Let A01 be the following function space:
A01 = f ∈ A1 and there are C1 , C2 ∈ R satisfying f2 (α) = C1 + C2 log α for all α > 0 .
Example 2.2. Here there are some examples of classical probability distributions satisfying G−1 ∈ A01 :
• Exponential distribution E(λ): In this case, G−1 (x) = − λ1 log(x), and this implies G−1 ∈ A01
with f1 (x) = − λ1 log(x) and f2 (α) = log α (C1 = 0 and C2 = 1).
• Gamma distributions Γ(a) In this case, G(x) =
1
Γ(a)
R∞
x
ta−1 e−t dt for a ≥ 1 and we obtain,
using an asymptotic expansion of the incomplete gamma function (see Abramowitz and Stegun,
1964):
G−1 (x) =
1
− log x + (a − 1) log(− log x) + O(|(ln x)−1 |)
Γ(a)
x → 0.
As a consequence, we deduce G−1 ∈ A01 with
f1 (x) =
1
− log x + (a − 1) log(− log x)
Γ(a)
and
f2 (α) = log α (C1 = 0 and C2 = 1).
• Absolute value of standardized Gaussian distribution |N (0, 1)|: In this case, we can write
√
R∞
2
G(x) = √22π x e−t /2 dt = erfc(x/ 2), where erf c is the complementary Gauss error function.
But we know (see for instance Blair et al., 1976) that for x → 0, then erfc−1 (x) = √12 −
1/2
+ O(|(ln x)−1 |). As a consequence, for any α > 0,
log(πx2 ) − log(− log x)
log α
erfc−1 (α x) = erfc−1 (x) 1 +
+ O(|(ln x)−2 |)
2 log x
x → 0.
(2.6)
Consequently G−1 ∈ A01 with
f1 (x) =
p
−2 log x − log(− log x) − 2 log π
and
f2 (α) =
1
log α,
2
implying C1 = 0 and C2 = 12 .
k
• Weibull distributions: In this case, with a ≥ 0 and 0 < b ≤ 1, G(x) = e−(x/λ) with λ > 0 and
1/k
k ∈ N∗ , for x ≥ 0. Then it is obvious that G−1 (x) = λ − log x
and therefore G−1 ∈ A01 with
1/k
f1 (x) = λ − log x
and f2 (α) = k1 log α (implying C1 = 0 and C2 = 1/k).
When G−1 ∈ A01 , it is possible to specify the limit distribution of (2.5). Thus, we show the following
result:
Proposition 2. Assume that G−1 ∈ A01 . Then
J
Y
0
P max {τn−j } ≤ x −→
1 − e−jx/C2 .
j=1,··· ,J
(2.7)
n→∞ j=1
Such a result is interesting since it provides the asymptotic behavior of a vector of normalized and
centered ratios τi . Its asymptotic distribution is the distribution of a vector of independent exponentially
distributed r.v’s. However the parameters of these exponential distributions are different. Thus, if we
consider the statistic
0
Tb = max {τn−j
},
j=1,··· ,J
(2.8)
the computation of the cumulative distribution function of Tb requires consideration of the function
Q
y ∈ [0, ∞) 7→ R(y) = Jj=1 1 − e−jy . This function converges quickly to 1 when J increases. Hence
we numerically obtain that for J ≥ 3, R(3.042) ' 0.95. Then we deduce from (2.7) and (2.8) that for
J ≥ 3,
P Tb ≤ 3.042 × C2 ' 0.95.
This implies that for instance that for J ≥ 3 and n large enough,
• P Tb ≤ 3.042 ' 0.95
• P Tb ≤ 1.521 ' 0.95
when X follows a Gamma distribution
when |X| = |N (0, 1)|,
0
with the computation of C2 for each distribution. We remark that the ratio τn−1
is the main contributor
to the statistic Tb and it contains almost all the information. For giving equivalent weights to the other
ratios τk0 , k ≤ n − 1 and so as not to be troubled by the nuisance parameter C2 , it is necessary to modify
the statistic Tb. Then we consider:
Ten = max
j=1,··· ,J
j
0
τn−j
1
×
sJ
where
J
1 X 0
sJ =
j τn−j .
J
(2.9)
j=1
For (un )n and (vn )n two sequences of real numbers, denote un
∼
vn when un /vn −→ 1. The
n→∞
n→∞
following proposition can be established:
Proposition 3. Assume that G−1 ∈ A01 . Then, for a sequence (Jn )n satisfying Jn −→ ∞ and
n→∞
Jn / log n −→ 0,
n→∞
Pr Ten ≤ x
∼
1 − e−x
Jn
for x > 0.
(2.10)
n→∞
In the case where G−1 ∈ A2 , similar results can also be established, we demonstrated below.
Example 2.3. Here there are some examples of classical distributions such as G−1 ∈ A2 :
• Pareto distribution P(α): In this case, with c > 0 and C > 0, G−1 (x) = C x−c for x → 0, and
this implies G−1 ∈ A2 with a = c.
• Burr distributions B(α): In this case, G(x) = (1 + xc )−k for c and k positive real numbers.
Thus G−1 (x) = (x−1/k − 1)1/c for x ∈ [0, 1], implying G−1 ∈ A2 with a = (ck)−1 .
• Absolute value of Student distribution |t(ν)| with ν degrees of freedom: In the case of a
Student distribution with ν degrees of freedom, the cumulative distribution function is Ft(ν) (x) =
1
2 (1
+ I(y, ν/2, 1/2)) with y = ν(ν + x2 )−1 and therefore G|t(ν)| (x) = I(y, ν/2, 1/2), where I is
the normalized beta incomplete function. Using the handbook of Abramowitz and Stegun (1964),
we have the following expansion G|t(ν)| (x) =
2ν ν/2−1
B(ν/2,1/2)
x−ν + O(x−ν+1 ) for x → 0, where B is the
usual Beta function. Therefore,
G−1
|t(ν)| (x) =
B(ν/2, 1/2) −1/ν
x
+ O(x−1/ν−1 )
2ν ν/2−1
Consequently G−1
|t(ν)| ∈ A2 with a = 1/ν.
x → ∞.
Remark 2. The case of standardized log-normal distribution is singular. Indeed, the probability dis
√x
tribution of X is the same than the one of exp(Z) where Z ∼ N (0, 1). Therefore, G(x) = 21 erfc log
2
√
implying G−1 (x) = exp 2 erfc−1 (2x) . Using the previous expansion (2.6), we obtain for any α > 0:
√
2 erfc−1 (2x α)
√
log α
= exp
2 erfc−1 ( 2x) 1 +
+ O(|(ln x)−2 |)
2 log x
G−1 (α x) = exp
= G−1 (x) 1 + O(|(ln x)−1/2 |).
Therefore, the standardized log-normal distribution is such that G−1 ∈
/ A1 ∪ A2 .
For probability distributions such as G−1 ∈ A2 we obtain the following classical result (see also Embrechts et al., 1997):
Proposition 4. Assume that G−1 ∈ A2 . Then,
J
Y
P max {log(τn−j )} ≤ x −→
1 − e−jx/a .
j=1,··· ,J
(2.11)
n→∞ j=1
Hence the case of G−1 ∈ A2 also provides interesting asymptotic properties on the ratios. In the
forthcoming section devoted to the construction of an outlier detector from previous results, we are
going to consider a test statistic which could be as well applied to distributions with functions G−1
belonging to A01 and A2 .
3
A new non-parametric outlier detector
We are going to consider the following test problem:
H : there is no outlier in the sample
0
.
H : there is at least one outlier in the sample
1
(3.1)
However we have to specify which kind of outlier and therefore which kind of contamination we consider.
Our guide for this is typically the case of oversized regression residuals. Thus we would like to detect
from (X1 , . . . , Xn ) when there is a ”gap” between numerous Xi coming from a common probability
distribution and one or several (but not a lot!) Xj which are larger than the other one and generated
from another distributions. As a consequence the previous test problem can be specified as follows:
H : (X1 , . . . , Xn ) are i.i.d. r.v. with G−1 ∈ A01 ∪ A2
0
H1 : there exists K ∈ N∗ such as (X(1) , . . . , X(n−K) ) are i.i.d. r.v. with G−1 ∈ A01 ∪ A2 . (3.2)
and for i = 0, . . . , K − 1, the distribution of X
satisfy G−1
> G−1 .
(n−i)
X(n−i)
A consequence of this specification is the following: we would like to detect outliers which appear as
oversized data. Hence, under H1 we could expect that there exists a ”jump” between the smallest
outlier and the largest non contaminated data. As a consequence under H1 we could expect that the
ratio τ(n−J) is larger than it should be.
For doing such a job, we propose to consider the following outlier detector based on ratios and which
could be used as well when G−1 belongs to A01 or A2 . Hence, define:
b Jn = log 2
D
b Jn
L
max j log(τn−j )
j=1,··· ,Jn
where
b Jn = median j log(τn−j )
L
.
1≤j≤Jn
(3.3)
Then, we obtain the following theorem:
Theorem 3.1. Assume that G−1 ∈ A01 ∪ A2 . Then, for a sequence (Jn )n satisfying Jn −→ ∞ and
n→∞
Jn / log n −→ 0,
n→∞
b Jn ≤ x
Pr D
∼
1 − e−x
Jn
.
(3.4)
n→∞
b Jn we prefer an estimation of the parameter of the exponential
Remark 3. In the definition of D
distribution with a robust estimator (median) instead of the usual efficient estimator (empirical mean),
since several outliers could corrupt this estimation.
Therefore, Theorem 3.1 can be applied for distributions with G−1 belonging to A01 or A2 , i.e. as well as
for Gaussian, Gamma or Pareto distributions. Hence, for a type I error α ∈ (0, 1), the outlier detector
b Jn can be computed, and with t = − log 1 − (1 − α)1/Jn ,
D
b Jn ≤ t then we conclude that there is no outlier in the sample.
• If D
b Jn > t then the largest index b
b Jn ≥ t indicates that we decide that
• If D
k0 such as b
k0 log(τn−bk0 )/L
the observed data (X(n−bk0 +1) , X(n−bk0 +2) , . . . , X(n) ) can be considered as outliers, implying that
there are b
k0 detected outliers.
As a consequence this outlier detector allows a decision of the test problem (3.1) and also the identification of the exact outliers.
4
Monte-Carlo experiments
We are going to compare the new outlier detector defined in (3.3) with usual univariate outlier detectors.
b Jn , we present the results of Monte-Carlo
After giving some practical details of the application of D
experiments under several probability distributions.
Practical procedures of outlier detections
b Jn is simple, and in practice just requires the specification of 2 parameters:
The definition of D
b J when there are K
Table 1: Numerical choice of Jn : average frequencies for potential outliers with D
outliers from the seven considered probability distributions, and for several values of K, sample size n
and parameter J. Here α = 0.007, 20000 independent replications are done and multiplicative outliers
(see below) are generated.
J
12
14
16
18
20
22
24
26
28
0.595
0.588
0.583
0.573
0.568
0.565
0.561
0.557
0.554
0.694
0.708
0.702
0.696
0.694
0.689
0.686
0.683
0.679
0.727
0.719
0.712
0.705
0.700
0.691
0.687
0.685
0.680
K = 10
0.796
0.818
0.823
0.822
0.817
0.813
0.811
0.810
0.804
n = 1000, K = 5
0.760
0.752
0.744
0.738
0.732
0.727
0.722
0.716
0.711
K = 10
0.822
0.848
0.857
0.858
0.856
0.853
0.850
0.847
0.844
n = 5000, K = 5
0.802
0.793
0.786
0.779
0.773
0.769
0.763
0.759
0.754
K = 10
0.846
0.878
0.891
0.894
0.896
0.893
0.890
0.888
0.887
n = 100, K = 5
K = 10
n = 500, K = 5
• The type I error α is the risk to detect outliers in the sample while there is no outlier. Hence, a
natural choice could be the ”canonical” α = 0.05. However, we chose to be strict concerning the
risk of false detection, i.e. we chose α = 0.007 (as it was chosen by Tukey himself for building
boxplots) which implies that we prefer not to detect ”small” outliers and hence we avoid to detect
a large number of outliers while there is no outlier.
• The number Jn of considered ratios. On the one hand, it is clear that the smaller Jn , the smaller
the detection threshold, therefore more sensitive is the detector to the presence of outliers. On
the other hand, the larger Jn , the more precise is the estimation of the parameter of asymptotic
b Jn is √n) and larger is the possible number
exponential distribution (the convergence rate of L
of detected outliers. We carried out numerical simulations using 20000 independent replications,
for several probability distributions (the seven distributions presented below) for several values
of the number of outliers K, sample size n and parameter J. Results are reported in Table 1. A
first conclusion: the larger n and K the larger test power. Another conclusion, but this is not a
surprise, is the fact that the ”optimal” choice of J depends on K and n.
As a consequence, for at least detecting K = 10 outliers, we use Jn = 1 + [4 ∗ log3/4 (n)] that is
an arbitrary choice satisfying Jn = o(log(n)) and fitting well the results of these simulations, i.e.
for n = 100, Jn = 13, for n = 1000, Jn = 18 and for n = 5000, Jn = 20.
b Jn to five common and well-known univariate outlier detectors
We have compared the new detector D
computed from the sample (X1 , · · · , Xn ).
1. The Student’s detector (see for instance Rousseeuw and Leroy, 2005): an observation from the
sample (X1 , · · · , Xn ) will be consider as an outlier when P(Xk > X n + ss × σ n ) where X n and
σ 2n are respectively the usual empirical mean and variance computed from (X1 , · · · , Xn ), and ss
is a threshold. This threshold is usually computed from the assumption that (X1 , · · · , Xn ) is a
Gaussian sample and therefore ss = qt(n−1) (1 − α/2) , where qt(n−1) (p) denotes the quantile of
the Student distribution with (n − 1) degrees of freedom for a probability p.
2. The Tukey’s detector (see Tukey, 1977) which provides the famous and usual boxplots: Xk is
considered to be an outlier from (X1 , · · · , Xn ) if Xk > Q3 + 1.5 × IQR, where IQR = Q3 − Q1,
with Q3 and Q1 the third and first empirical quartiles of (X1 , · · · , Xn ). Note that the confidence
factor 1.5 was chosen by Tukey such that the probability of Gaussian random variable to be
decided as an outlier is close 0.7% which is good trade-off.
3. An adjusted Tukey’s detector as it was introduced and studied in Hubert and Vandervieren (2008):
Xk is considered to be an outlier from (X1 , · · · , Xn ) if Xk > Q3 + 1.5Q3 + 1.5 e3 M C × IQR, where
M C is the medcouple, defined by M C = medianXi ≤Q2≤Xj h(Xi , Xj ), where Q2 is the sample
median and the kernel function h is given by h(Xi , Xj ) =
(Xj −Q2)−(Q2−Xi )
.
Xj −Xi
This new outlier
detector improves considerably the accuracy of the usual Tukey’s detector for skewed distributions.
4. The M ADe detector (see for instance Rousseeuw and Leroy, 2005): Xk is considered as an outlier
from (X1 , · · · , Xn ) if |Xk − Q2| > 3 ∗ 1.483 ∗ median(|X1 − Q2|, · · · , |Xn − Q2|). The coefficient
1.483 is obtained from the Gaussian case based on the relation SD ' 1.483 × median(|X1 −
Q2|, · · · , |Xn − Q2|), while the confidence factor c = 3 is selected for building a conservative test.
5. The Local Outlier Factor (LOF), which is a non-parametric detector (see for instance Breunig et
al., 2000). This procedure is based on this principle: an outlier can be distinguished when its
normalized density (see its definition in Breunig et al., 2000) is larger than 1 or than a threshold
larger than 1. However, the computation of this density requires to fix the parameter k of the
used k-distance and a procedure or a theory for choosing a priori k does not still exist. The
authors recommend k > 10 and k < 50 (generally). We chose to fix k = Jn , where Jn is used for
b J . Then, using the same kind of simulations than those reported in Table
the computation of D
1, we tried to optimize the choice of a threshold sLOF defined by: if LOF (Xi ) > sLOF then the
observation Xi is considered to be an outlier. We remark that it is not really possible to choose
a priori k and sLOF with respect to α. Table 2 provides the results of simulations from 10000
independent replications and the seven probability distributions. We have chosen to optimize a
sum of empirical type I (case K = 0) and II (case K = 5 and K = 10) errors. This leads one to
choose sLOF = 8 for n = 100 as well as for n = 1000.
Table 2: Numerical choice of the LOF threshold sLOF : Average frequencies of potential outliers with
LOF detector when there are K outliers from the seven considered probability distributions, and for
several values of K, sample size n and threshold sLOF . Note that 10000 independent replications are
generated and multiplicative outliers (see below) are generated.
sLOF
K
n = 100
n = 1000
2
4
6
8
10
12
K=0
0.903
0.447
0.285
0.222
0.174
0.150
K=5
1.000
1.000
1.000
0.996
0.961
0.863
K = 10
1.000
0.860
0.422
0.232
0.141
0.106
K=0
0.955
0.533
0.328
0.237
0.196
0.179
K=5
1.000
1.000
1.000
1.000
1.000
0.995
K = 10
1.000
1.000
0.981
0.888
0.744
0.581
Student, Tukey and M ADe detectors are more or less based on Gaussian computations. We would
not be surprised if those methods failed to detect outliers when the distribution of X is ”far” from the
Gaussian distribution (but these usual detections of outliers, for instance the Student detection obtained
on Studentized residuals from a least squares regression, are done even if the Gaussian distribution is
not attested). Moreover, the computations of these detectors’ thresholds are based on an individual
detection of outlier, i.e. a test deciding if a fixed observation Xi0 is an outlier or not. Hence, if we apply
them to each observation of the sample, the probability to detect an outlier increases with n. This is
not exactly the same as deciding whether if there are no outliers in a sample, which is the object of test
problem (3.1).
b Jn , it is appropriate to change the thresholds of these detectors
Then, to compare these detectors to D
following test problem (3.1) and its precision (3.2). Hence we have to define a threshold s > 0 of this
test from the relation P(∃k = 1, · · · , n, Xk > s) = α. Therefore, from the independence property
1 − α = P(X1 < s)n implying that s has to satisfy P(X1 > s) = (1 − (1 − α)1/n ) ' α/n when n is large
and α close to 0 (typically α = 0.01). In the sequel we are going to compute the confidence factors of
previous famous detectors to the particular case of absolute values of Gaussian variables. Hence, when
for Z a N (0, 1) random variable and if n is large, then P(|Z| > sG ) '
α
n
with sG = qN (0,1) (1 −
α
2n ).
Then, we define:
1. The Student detector 2: we consider that Xk from (X1 , · · · , Xn ) is an outlier when Xk > X n +
p
p
√
p
α
cS × σ n , with 2/ 2π + cS (π − 2)/π = sG implying cS = qN (0,1) (1 − 2n
) − 2/π π/(π − 2)
(here we assume that n is a large number inducing that the Student distribution with (n − 1)
degrees of freedom could be approximated by the standard Gaussian distribution).
2. The Tukey detector 2: we consider that Xk from (X1 , · · · , Xn ) is an outlier when Xk > Q3 + cT ×
IQR. In the case of the absolute value of a standard Gaussian variable, Q1 ' 0.32 and Q3 ' 1.15
implying cT = (sG − 1.15)/0.83 ' 1.20 × sG − 1.38.
3. The M ADe detector 2: we consider that Xk from (X1 , · · · , Xn ) is an outlier when Xk − Q2 >
cM × median(|X1 − Q2|, · · · , |Xn − Q2|). In the case of the absolute value of a standard Gaussian
variable, Q2 ' 0.67 and median(|X1 − Q2|, · · · , |Xn − Q2|) ' 0.40. This induces cM ' 2.50 ×
sG − 1.69.
First results of Monte-Carlo experiments for samples without outlier: the size of the test
We apply the different detectors in different frames and for several probability distributions which are:
• The absolute value of Gaussian distribution with expectation 0 and variance 1, denoted N (0, 1)
(case A01 );
• The exponential distribution with parameter 1, denoted E(1) (case A01 );
• The Gamma distribution with parameter 3, denoted Γ(3) (case A01 );
• The Weibull distribution with parameters (3, 4), denoted W (3, 4) (case A01 );
• The absolute value of a Student distribution with 2 degrees of freedom, denoted |t(2)| (case A2 );
• The standard log-normal distribution, denoted log −N (0, 1) (not case A01 or A2 );
• The absolute value of a Cauchy distribution, denoted |C| (case A2 ).
In the sequel, we will consider samples (X1 , · · · , Xn ) following these probability distributions, for n =
100 and n = 1000, and for several numbers of outliers.
We begin by generating independent replications of samples without outlier, which corresponds to be
under H0 . Then we apply the outlier detectors. The results are reported in Table 3.
A first conclusion from this simulations is the following: as we already said, the original Student, Tukey
e Jn because their empirical sizes came out
and M ADe detectors can not be compared to the detector D
larger: they are not constructed to answer to our test problem. Therefore we are going now to consider
only their global versions Student 2, Tukey 2 and M ADe 2. Moreover, as they have been constructed
from the Gaussian case, these second versions of detectors provide generally poor results in case of
non-Gaussian distributions especially for Student, log-normal and Cauchy distributions (where outliers
are always detected while there are no generated outliers).
Table 3: The size of the test: average frequencies (Av. Freq.) of potential outliers with the different
outlier detectors, for the different probability distributions, n = 100 and n = 1000, while there is no
generated outlier in samples. Here α = 0.007 and 20000 independent replications are generated.
N (0, 1)
E(1)
Γ(3)
W (3, 4)
|t(2)|
log −N (0, 1)
|C|
e Jn
Av. Freq. D
0.007
0.008
0.008
0.008
0.010
0.010
0.018
Av. Freq. LOF
0.001
0.029
0.013
0
0.643
0.259
0.970
Av. Freq. Student
0.861
0.993
0.921
0.308
1
1
1
Av. Freq. Tukey
0.804
0.991
0.918
0.327
1
1
1
Av. Freq. Adj. Tukey
0.213
0.344
0.408
0.353
0.844
0.710
0.981
Av. Freq. M ADe
0.751
0.995
0.879
0.163
1
1
1
Av. Freq. Student 2
0.002
0.110
0.017
0
0.655
0.515
0.936
Av. Freq. Tukey 2
0.022
0.486
0.119
0
0.951
0.937
1
Av. Freq. M ADe 2
0.023
0.624
0.112
0
0.971
0.975
1
N (0, 1)
E(1)
Γ(3)
W (3, 4)
|t(2)|
log −N (0, 1)
|C|
e Jn
Av. Freq. D
0.009
0.009
0.009
0.009
0.014
0.011
0.016
Av. Freq. LOF
0.005
0.023
0.019
0.001
0.843
0.281
0.998
Av. Freq. Student
1
1
1
0.986
1
1
1
Av. Freq. Tukey
1
1
1
0.939
1
1
1
0.492
0.899
0.950
0.796
1
1
1
1
1
1
0.664
1
1
1
Av. Freq. Student 2
0.006
0.552
0.112
0
1
0.998
1
Av. Freq. Tukey 2
0.008
0.943
0.259
0
1
1
1
Av. Freq. M ADe 2
0.008
0.991
0.246
0
1
1
1
n = 100
n = 1000
Av. Freq. Adj Tukey
Av. Freq. M ADe
Second results of Monte-Carlo experiments for samples with outliers
Now, we consider the cases where there is a few number of outliers in the samples (X1 , · · · , Xn ).
Denote K the number of outliers, and ` > 0 a real number which represents a positive real number.
We generated 3 kinds of contaminations:
• A shift contamination: (X(1) , · · · , X(n−K) , X(n−K+1) +`, · · · , X(n) +`) instead of (X(1) , · · · , X(n) ).
We chose ` = 10. By the way the cluster of outliers is necessarily separated from the cluster of
non-outliers.
• A multiplicative contamination: (X(1) , · · · , X(n−K) , 3X(n−K+1) , · · · , 3X(n) ) instead of (X(1) , · · · , X(n) ).
By the way the cluster of outliers is necessarily separated from the cluster of non-outliers.
• A point contamination: (X(1) , · · · , X(n−K) , `, · · · , `) instead of (X(1) , · · · , X(n) ). We chose ` =
1000. By the way, the cluster of outliers is generally separated to the cluster of non-outliers (but
not necessary, especially for Student or Cauchy probability distributions).
First, we consider the second versions of Student, Tukey and M ADe detectors. But we also consider
parametric versions of these detectors, which are denoted Student-para, Tukey-para and M ADe -para:
the thresholds and confidence factors are computed and used with the knowledge of the probability
distribution of the sample (these thresholds change following the considered probability distributions).
e Jn or LOF are non-parametric detectors. We chose these
Hence, they are parametric detectors while D
parametric versions because they allow to obtain the same size of all the detectors and then a comparison of the test powers is more significant.
First results are reported in Tables 4 (n = 100) and 5 (n = 1000) for K shifted outliers (K = 5
or K = 10). From Tables 4 and 5, it appears:
• Student 2, as well as Student-para detectors are not really good choices for the detection of outliers
because they are not robust statistics (the empirical variance is totally modified by the values of
outliers).
• Tukey and M ADe detectors provide more and less similar results. But even their parametric
versions are not able to detect outliers for skewed distributions (Student, log-normal and Cauchy
e Jn clearly provides better results, even if they are not very
distributions). With the same size, D
accurate (especially for the Cauchy distribution).
• LOF detector does not provide accurate results (especially when K = 10). It is certainly a more
interesting alternative in case of multivariate data.
Table 4: Average frequencies (Av. Freq.) of potential outliers with the different outlier detectors and
average number (Av. Numb.) of detected outliers when outliers are detected, for n = 100, with K
shifted outliers. Here α = 0.007 and 20000 independent replications are generated.
n = 100
K
N (0, 1)
E(1)
Γ(3)
W (3, 4)
|t(2)|
log −N (0, 1)
|C|
e Jn
Av. Freq. D
0
0.007
0.008
0.008
0.008
0.010
0.010
0.018
5
0.998
0.562
0.574
1
0.848
0.813
0.066
10
1
0.973
0.959
1
1
1
0.076
5
5.12
5.23
5.25
5.14
5.09
5.07
9.17
10
10.60
10.64
10.58
10.62
10.05
10.05
10.01
0
0.001
0.111
0.018
0
0.655
0.515
0.934
5
0
0.013
0.004
0
0.538
0.411
0.687
10
0
0
0
0
0.241
0.126
0.628
5
0
1
1
0
1.04
1.02
1.67
10
0
0
0
0
1.01
1
1.66
0
0.022
0.483
0.118
0
0.951
0.937
1
5
1
1
1
1
1
1
1
10
1
1
0.999
1
1
1
1
5
5.01
5.35
5.05
5
5.10
5.09
10.32
10
10
10.14
9.67
10
10
10
13.19
0
0.023
0.619
0.109
0
0.971
0.975
1
5
1
1
1
1
1
1
1
10
1
1
0.999
1
1
1
1
5
5.01
5.67
5.06
5
5.24
5.37
12.66
10
10
10.38
9.86
10
10
12.16
16.43
0
0.001
0.029
0.013
0
0.643
0.259
0.970
5
0.998
0.521
0.123
1
0.309
0.129
1
10
0.046
0.011
0
0.200
0.03
0.003
0.548
5
5
3.97
2.86
5
2.71
2.03
3.39
10
9.11
1.10
-
9.47
1.56
1.40
2.27
0
0.002
0
0.001
0.005
-
0
-
5
0
0
0
1
-
0
-
10
0
0
0
0
-
0
-
5
0
0
0
5
-
0
-
10
0
0
0
0
-
0
-
0
0.022
0.013
0.014
0.025
0.008
0.009
0.007
5
1
0.991
0.960
1
0.009
0.026
0.005
10
1
0.903
0.782
1
0.007
0.026
0.004
5
5.01
4.87
4.40
5.01
1
1.03
1.94
10
10
7.27
4.50
10
1
1.01
1.88
0
0.023
0.014
0.013
0.025
0.007
0.009
0.007
5
1
0.995
0.969
1
0.009
0.024
0.006
10
1
0.966
0.867
1
0.008
0.025
0.005
5
5.01
4.92
4.47
5.01
1
1.02
1.94
10
10
8.51
5.46
10
1
1.02
1.89
Av. Numb.
Av. Freq. Student 2
Av. Numb.
Av. Freq. Tukey 2
Av. Numb.
Av. Freq. M ADe 2
Av. Numb.
Av. Freq. LOF
Av. Numb.
Av. Freq. Student-para
Av. Numb.
Av. Freq. Tukey-para
Av. Numb.
Av. Freq. M ADe -para
Av. Numb.
Table 5: Average frequencies (Av. Freq.) of potential outliers with the different outlier detectors and
average number (Av. Numb.) of detected outliers when outliers are detected, for n = 1000, with K
shifted outliers. Here α = 0.007 and 10000 independent replications are generated.
n = 1000
K
N (0, 1)
E(1)
Γ(3)
W (3, 4)
|t(2)|
log −N (0, 1)
|C|
e Jn
Av. Freq. D
0
0.009
0.009
0.009
0.009
0.014
0.011
0.016
5
1
0.892
1
1
0.241
0.517
0.072
10
1
0.997
0.991
1
0.987
1
0.088
5
5.23
5.27
5.32
5.27
5.34
5.14
8.85
10
10.37
10.82
10.75
10.41
10.05
10.06
7.83
0
0.006
0.565
0.111
0
1
0.998
1
5
1
1
1
1
1
1
1
10
1
1
1
1
1
1
1
5
5
5.02
5.01
5
4.69
4.98
4.74
10
10
10
9.97
10
7.73
9.75
4.59
0
0.008
0.949
0.256
0
1
1
1
5
1
1
1
1
1
1
1
10
1
1
1
1
1
1
1
5
5.01
7.90
5.27
5
22.34
20.09
64.41
10
10.01
12.64
10.23
10.01
22.49
20.06
67.76
0
0.008
0.992
0.237
0
1
1
1
5
1
1
1
1
1
1
1
10
1
1
1
1
1
1
1
5
5.01
9.86
5.26
5
26.94
28.52
83.04
10
10.01
14.55
10.23
10.01
27.10
28.46
86.59
0
0.001
0.029
0.013
0
0.843
0.281
0.970
5
1
0.802
0.473
1
0.291
0.113
0.65
10
0.961
0.014
0.004
1
0.125
0.011
0.571
5
5.01
4.27
3.37
5
2.21
1.21
3.07
10
9.93
3.36
3.50
10
1.96
1.63
2.92
0
0.006
0.004
0.005
0.007
-
0
-
5
1
0.652
0.979
1
-
0
-
10
1
0.017
0.275
1
-
0
-
5
5
1.21
2.48
5.01
-
0
-
10
10
1
1.07
10
-
0
-
0
0.008
0.008
0.008
0.010
0.007
0.007
0.007
5
1
1
1
1
0.007
0.013
0.007
10
1
1
1
1
0.007
0.013
0.007
5
5.01
5.01
5.01
1.02
1.95
1
2.00
10
10.01
10.01
9.93
10.01
1
1
2
0
0.008
0.007
0.008
0.010
0.007
0.007
0.007
5
1
1
1
1
0.007
0.013
0.037
10
1
1
1
1
0.008
0.013
0.007
5
5.01
5.01
5.01
5
1.01
1.01
2
10
10.01
10.01
9.94
10.01
1
1
2
Av. Numb.
Av. Freq. Student 2
Av. Numb.
Av. Freq. Tukey 2
Av. Numb.
Av. Freq. M ADe 2
Av. Numb.
Av. Freq. LOF
Av. Numb.
Av. Freq. Student-para
Av. Numb.
Av. Freq. Tukey-para
Av. Numb.
Av. Freq. M ADe -para
Av. Numb.
Those conclusions are confirmed when both the other contaminations (multiplicative and point) are
used for generating outliers (see Table 6 and 7).
Conclusions of simulations
b Jn provide a very good trade-off for the detection of outliers for a very large
The log-ratio detector D
choice of probability distributions compared with usual outliers detectors or their fitted versions. The
second non-parametric outlier detector, LOF, is not really as accurate (especially for K = 10), its size
is not controlled and results did not depend on the choice of α (a theoretical study should be done for
writing the threshold as a function of α).
5
Application to real data
We apply the theoretical results to real datasets of detailed data on individual transactions in the
used car market. The purpose of the experiment was to detect as many outliers as possible. The
original dataset contains information about n = 6079 transactions on the car Peugeot 207 1.4 HDI 70
Trendy Berline including year and month which is the manufacture date, the price, and the number
of kilometres driven. We then define 3 variables: Age (the age of the car, in months), Price (in euros)
and Mileage (in km). We chose these cars because they were advertised often enough to permit us to
create a relatively homogeneous sample. Figure 1 depicts the relationship between the price and some
variables: Price with Mileage, Price with Age. Such data were collected by Autobiz society, and can
be used for forecasting the price of a car following its age and mileage. Hence it is crucial to construct
a model for the price from a reliable data set including the smallest number of outliers.
We now apply our test procedure to identify eventual outlying observations or atypical combination
between variables. After preliminary studies, we chose two significant characteristics for each car of
the sample. The first one is the number of kilometres per month. The second one is the residual
obtained, after an application of the exponential function, from a robust quantile regression between
the logarithm of the price as the dependent variable and Age and Mileage as exogenous variables (an
alternative procedure for detecting outliers in robust regression has been developed in Gnanadesikan
and Kettenring, 1972). The assumption of independence is plausible for both these variables’ residuals.
Figure 2 exhibits the boxplots of the distributions of those two variables.
b Jn is carried out on those two variables with Jn = 20 (given by the empirical
The outlier test D
choice obtained in Section 4 with n = 6079). As the sample size is large, we can accept to eliminate
data detected as outliers while there are not really outliers and we chose α = 0.05. The results are
Table 6: Average frequencies (Av. Freq.) of potential outliers with the different outlier detectors and
average number (Av. Numb.) of detected outliers when outliers are detected for n = 100 and n = 1000,
with K multiplicative outliers. Here α = 0.007 and 20000 independent replications are generated.
n = 100
K
N (0, 1)
E(1)
Γ(3)
W (3, 4)
|t(2)|
log −N (0, 1)
|C|
e Jn
Av. Freq. D
0
0.007
0.008
0.008
0.008
0.010
0.010
0.018
5
1
0.620
1
1
0.743
0.797
0.221
10
1
1
1
1
1
1
0.910
5
5.03
5.04
5.03
5.03
5.05
5.03
5.17
10
10.01
10.01
10.01
10.01
10
10
10
0
0.002
0
0.001
0.005
-
0
-
5
0.022
0
0.021
1
-
0
-
10
0.019
0
0.022
0.190
-
0
-
5
1
0
1
4.08
-
0
-
10
1
0
1
1.04
-
0
-
0
0.022
0.013
0.014
0.025
0.008
0.009
0.007
5
1
0.983
1
1
0.065
0.304
0.024
10
1
0.980
1
1
0.063
0.305
0.022
5
5
3.83
5
5
1.03
1.21
1
10
9.99
4.45
9.84
10
1.03
1.19
1.01
0
0.023
0.014
0.013
0.025
0.007
0.009
0.007
5
1
0.981
1
1
0.065
0.303
0.023
10
1
0.977
1
1
0.063
0.303
0.022
5
5
3.81
5
5
1.03
1.21
1
10
9.99
4.48
9.77
10
1.03
1.20
1.01
n = 1000
K
N (0, 1)
E(1)
Γ(3)
W (3, 4)
|t(2)|
log −N (0, 1)
|C|
e Jn
Av. Freq. D
0
0.009
0.009
0.009
0.009
0.014
0.011
0.016
5
1
1
1
1
0.933
0.997
0.200
10
1
1
1
1
1
1
0.947
5
5.06
5.06
5.06
5.06
5.08
5.06
5.34
10
10.02
10.03
10.03
10.03
10.03
10.03
10.02
0
0.006
0.004
0.005
0.007
-
-
-
5
1
0.996
1
1
-
0
-
10
1
0.664
1
1
-
0
-
5
5
2.46
5
5
-
0
-
10
10
1.25
5.10
10
-
0
-
0
0.008
0.008
0.008
0.010
0.007
0.007
0.007
5
1
1
1
1
0.061
0.441
0.021
10
1
1
1
1
0.061
0.443
0.021
5
5
5
5
5
1.02
1.32
1.01
10
10
9.97
10
10
1.03
1.34
1.01
0
0.008
0.007
0.008
0.010
0.007
0.007
0.007
5
1
1
1
1
0.061
0.441
0.021
10
1
1
1
1
0.060
0.450
0.022
5
5
5
5
5
1.03
1.32
1.01
10
10
9.97
10
10
1.03
1.34
1.01
Av. Numb.
Av. Freq. Student-para
Av. Numb.
Av. Freq. Tukey-para
Av. Numb.
Av. Freq. M ADe -para
Av. Numb.
Av. Numb.
Av. Freq. Student-para
Av. Numb.
Av. Freq. Tukey-para
Av. Numb.
Av. Freq. M ADe -para
Av. Numb.
Table 7:
Average frequencies (Av. Freq.) of potential outliers with the different outlier detectors
and average number (Av. Numb.) of detected outliers when outliers are detected, for n = 100 and
n = 1000, with K point outliers. Here α = 0.007 and 20000 independent replications are generated.
n = 100
K
N (0, 1)
E(1)
Γ(3)
W (3, 4)
|t(2)|
log −N (0, 1)
|C|
e Jn
Av. Freq. D
0
0.007
0.008
0.008
0.008
0.010
0.010
0.018
5
1
1
1
1
1
1
0.719
10
1
1
1
1
1
1
0.979
5
5.11
5.14
5.19
5.15
5.43
5.27
5.83
10
10.63
10.67
10.75
10.68
11.10
10.98
11.25
0
0.002
0
0.001
0.005
-
0
-
5
0
0
0
1
-
0
-
10
0
0
0
0
-
0
-
5
0
0
0
5
-
0
-
10
0
0
0
0
-
0
-
0
0.022
0.013
0.014
0.025
0.008
0.009
0.007
5
1
1
1
1
1
1
0.005
10
1
1
1
1
1
1
0.004
5
5.01
5
5.01
5.01
5
5
1.01
10
10
10
10
10
10
10
1
0
0.023
0.014
0.013
0.025
0.007
0.009
0.007
5
1
1
1
1
1
1
0.006
10
1
1
1
1
1
1
0.006
5
5.01
5.01
5.01
5.01
5.01
5.01
1.03
10
10
10
10
10
10
10
1
n = 1000
K
N (0, 1)
E(1)
Γ(3)
W (3, 4)
|t(2)|
log −N (0, 1)
|C|
e Jn
Av. Freq. D
0
0.009
0.009
0.009
0.009
0.014
0.011
0.016
5
1
1
1
1
1
1
0.246
10
1
1
1
1
1
1
0.709
5
5.23
5.27
5.32
5.26
5.69
5.28
7.76
10
11.03
11.11
11.22
11.13
11.09
11.35
12.73
0
0.006
0.004
0.005
0.007
-
0
-
5
1
1
1
1
-
0
-
10
1
1
1
1
-
0
-
5
5
5
5
5
-
0
-
10
10
10
10
10
-
0
-
0
0.008
0.008
0.008
0.010
0.007
0.007
0.007
5
1
1
1
1
1
1
0.007
10
1
1
1
1
1
1
0.007
5
5.01
5.01
5.01
5.01
5.01
5.01
1
10
10.01
10.01
10.01
10.01
10.01
10.01
1.01
0
0.008
0.007
0.008
0.010
0.007
0.007
0.007
5
1
1
1
1
1
1
0.007
10
1
1
1
1
1
1
0.008
5
5.01
5.01
5.01
5.01
5.01
5.01
1
10
10.01
10.01
10.01
10
10.01
10.01
1.01
Av. Numb.
Av. Freq. Student-para
Av. Numb.
Av. Freq. Tukey-para
Av. Numb.
Av. Freq. M ADe -para
Av. Numb.
Av. Numb.
Av. Freq. Student-para
Av. Numb.
Av. Freq. Tukey-para
Av. Numb.
Av. Freq. M ADe -para
Av. Numb.
Figure 1: Relationship between the dependent variables and the regressors: Price with Mileage (left),
●
20000
●
●
●
5000
10000
15000
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
0
0
1000
2000
3000
4000
Price with Age (right).
Figure 2: Boxplots of kilometres per month (left) and of absolute values of quantile regression residuals
(right).
Table 8:
b Jn applied to 3 samples: the number of kilometres per month (km/m),
The outlier test D
max(km/m) − km/m and the residuals obtained from a quantile regression of the log-prices onto the
age and the mileage.
Sample
Jn
b Jn
D
t
Outliers
km/m (Sup)
20
6.7232
5.96721
n=6
km/m (Inf)
20
5.1200
5.96721
n=0
Res
20
6.3322
5.96721
n=2
presented in Tables 8, 9 and 10. Note that, concerning the study of kilometres per month (km/m),
we directly applied the test to this variable for detecting eventual ”too” large values, but also to
max(km/m) − (km/m) for detecting eventual ”too” small values.
Conclusions of the applications
We first remark that we did not get the same outliers from the different analysis. It could be expected
because the test on residuals worked as a multivariate test and identify atypical association between
the three variables Age, Mileage and Price while the tests done on kilometres/month identifies outlying
values in a bivariate case i.e. a typical association between the two variables Age and Mileage. From
a practitioner’s point of view it may be advisable to apply the test for the two cases together one
by one to be sure to detect the largest number of outliers. A second remark concerns the ”type” of
the detected outliers. We can state that concerning kilometres/month, outliers are simply the largest
values (the test did not identify outliers for ”too” small values). But for the regression residuals, the
detected outliers clearly correspond to typing errors on the prices (the prices have been replaced by the
mileages!). Thus, two kinds of outliers have been detected.
Table 9: Detailed analysis of the detected outliers obtained from the sample of kilometers per month
(large values).
Detected Outliers
Price
Mileage
outlier(1)
9590
70249
outlier(2)
11690
outlier(3)
Age
Kilometers per Month
Predicted Price
16
4391
9909
61484
14
4392
10286
10490
61655
14
4404
10280
outlier(4)
9390
61891
14
4421
10272
outlier(5)
11500
39826
9
4425
11285
outlier(6)
11900
65411
15
4361
10111
Table 10: Detailed analysis of outliers detected from the residual’s sample.
6
Detected Outliers
Price
Mileage
Age
Predicted Price
Outlier(1)
34158
34158
28
10626
Outlier(2)
29000
29000
11
11600
Proofs
Proof of Proposition 1. We begin by using the classical following result (see for example Embrechts et
al. 1997):
D
X(n−J) , X(n−J+1) , · · · , X(n) = G−1 ΓJ+1 /Γn+1 , G−1 ΓJ /Γn+1 , · · · , G−1 Γ1 /Γn+1 ,
(6.1)
where (Γi )i∈N∗ is a sequence of random variables such as Γi = E1 + · · · + Ei for i ∈ N∗ and (Ei )j∈N∗ is
a sequence of i.i.d.r.v. with distribution E(1). Consequently, we have
G−1 Γ /Γ
G−1 ΓJ−1 /Γn+1
G−1 Γ1 /Γn+1
n+1
J
D
,
,··· ,
.
τn−J , τn−J+1 , · · · , τn−1 =
G−1 ΓJ+1 /Γn+1
G−1 ΓJ /Γn+1
G−1 Γ2 /Γn+1
1
a.s.
But for j ∈ N∗ , G−1 Γj /Γn+1 = G−1
×Γj . From the strong law of large numbers, Γn+1 −→ ∞,
Γn+1
n→∞
−1
therefore since G ∈ A1 , we almost surely obtain:
G−1 Γj /Γn+1 = f1
1
Γn+1
× 1−
f2 (Γj )
1
+O
.
log(Γn+1 )
log2 (Γn+1 )
Using once again the strong law of large numbers, we have Γn+1 ∼ n −→ ∞ almost surely. Hence, we
n→∞
can write for all j = 1, · · · , J,
G−1 Γj /Γn+1
G−1 Γj+1 /Γn+1
1−
f2 (Γj )
log(Γn+1 )
+O
1
log2 (Γn+1 )
1−
f2 (Γj+1 )
log(Γn+1 )
+O
1
log2 (Γn+1 )
=
= 1+
f2 (Γj+1 ) − f2 (Γj )
1
+O
,
2
log(Γn+1 )
log (Γn+1 )
(6.2)
using a Taylor expansion. By considering now the family (τj0 )j defined by τj0 = (τj − 1) log(n) and the
limit of the previous expansion, we obtain
0
0
0
τn−J
, τn−J+1
, · · · , τn−1
D
−→
f2 (ΓJ+1 ) − f2 (ΓJ ) , f2 (ΓJ ) − f2 (ΓJ−1 ) , · · · , f2 (Γ2 ) − f2 (Γ1 ) . (6.3)
n→∞
The function (x1 , · · · , xJ ) ∈ RJ 7→ max(x1 , · · · , xJ ) ∈ R is a continuous function on RJ and therefore
we obtain (2.5).
Proof of Proposition 2. We use the asymptotic relation (2.5). Since G−1 ∈ A01 , for k = 1, · · · , J,
f2 (Γk+1 ) − f2 (Γk ) = C2 log Γk+1 /Γk = C2 log Γk+1 /ΓJ+1 − C2 log Γk /ΓJ+1 ,
But the random variable f2 (Γk+1 ) − f2 (Γk ) = f2 (Γk + Ek+1 ) − f2 (Γk ) is absolutely continuous with
respect to Lebesgue measure since Γk and Ek+1 are independent random variables. Using once again
the property (6.1), and since for an exponential distribution E(1), G−1 (x) = − log(x), then
D 0
0
0
− log ΓJ /ΓJ+1 , − log ΓJ−1 /ΓJ+1 , · · · , − log Γ1 /ΓJ+1 = E(1)
, E(2)
, · · · , E(J)
.
0 ≤ E0 ≤ · · · ≤ E0
where (Ej0 )j is a sequence of i.i.d.r.v. following a E(1) distribution and E(1)
(2)
(J) is the
order statistic from (E10 , · · · , EJ0 ). Consequently,
f2 (ΓJ+1 ) − f2 (ΓJ ), f2 (ΓJ ) − f2 (ΓJ−1 ) , · · · , f2 (Γ2 ) − f2 (Γ1 )
= C2 − log ΓJ /ΓJ+1 , log ΓJ /ΓJ+1 − log ΓJ−1 /ΓJ+1 , · · · , log Γ2 /ΓJ+1 − log Γ1 /ΓJ+1
D
0
0
0
0
0
= C2 E(1)
, E(2)
− E(1)
, · · · , E(J)
− E(J−1)
With y = x/C2 and using (2.5), this implies
P
max
{τj0 } ≤ x
j=n−J,··· ,n−1
−→
n→∞
−→
n→∞
0
0
0
0
0
P E(1)
≤ y, E(2)
≤ y + E(1)
, · · · , E(J)
≤ y + E(J−1)
0
0
J! P E10 ≤ y, E10 ≤ E20 ≤ y + E10 , · · · , EJ−1
≤ EJ0 ≤ y + EJ−1
.
The explicit computation of this probability is possible. Indeed:
0
0
P E10 ≤ y, E10 ≤ E20 ≤ y + E10 , · · · , EJ−1
≤ EJ0 ≤ y + EJ−1
Z y+eJ−2 Z y+eJ−1
Z y Z y+e1 Z y+e2
=
···
e−e1 e−e2 e−e3 · · · e−eJ−1 e−eJ de1 de2 de3 · · · deJ−1 deJ
0
e1
= 1−e
−y
Z
e2
y Z y+e1
eJ−2
y+e2
eJ−1
y+eJ−2
Z
Z
···
0
e1
e2
e−e1 e−e2 e−e3 · · · e−2eJ−1 de1 de2 de3 · · · deJ−1
eJ−2
Z Z
Z
Z y+eJ−3
y y+e1 y+e2
1
−y
−2y
1−e
···
e−e1 e−e2 e−e3 · · · e−3eJ−2 de1 de2 de3 · · · deJ−2
= 1−e
2
0
e1
e2
eJ−3
..
..
..
..
..
..
..
=
.
.
.
.
.
.
.
Z y Z y+e1
1
1 − e−y 1 − e−2y × · · · × 1 − e−(J−2)y
e−e1 e−(J−1)e2 de1 de2
=
(J − 2)!
Z0 e1
y −Je1
1
−y
−2y
−(J−1)y
=
e
de1
1−e
1−e
× ··· × 1 − e
(J − 1)!
0
1
=
1 − e−y 1 − e−2y × · · · × 1 − e−Jy .
J!
Then, we obtain (2.7).
Proof of Proposition 3. Such a result can be obtained by modifications of Propositions 1 and 2. Indeed,
we begin by extending Proposition 1 in the case where Jn −→ ∞ and Jn / log n −→ 0. This is possible
n→∞
n→∞
D
since Γn+1 /n = 1 + n−1/2 εn with εn −→ N (0, 1) from usual Central Limit Theorem. Using the Deltan→∞
D
method, we also obtain log(Γn+1 /n) = n−1/2 ε0n with ε0n −→ N (0, 1). Hence, for any j = 1, · · · , Jn ,
n→∞
from (6.3),
G−1 Γ /Γ
j
n+1
−1
= log(n)
G−1 Γj+1 /Γn+1
1
D
= f2 (Γj+1 ) − f2 (Γj ) + O
log(n)
1
D
= C2 log(Γj+1 /Γjn ) + O
.
log(n)
0
0
Denote Fn the cumulative distribution function of τn−J
,
·
·
·
,
τ
, and Fen the one of f2 (ΓJn +1 ) −
n−1
n
f2 (ΓJn ), · · · , f2 (Γ2 ) − f2 (Γ1 ) = C2 log(ΓJn +1 /ΓJn ), · · · , log(Γ2 /Γ1 ) . Then, using the second equation
0
τn−j
D
of the proof of Proposition 1, for all (x1 , · · · , xJn ) ∈ (0, ∞)Jn ,
Fn (x1 , · · · , xJn ) = Fen (x1 + u1n , · · · , xJn + uJnn ),
1
with uin = O log(n)
. But it is clear that the probability measure of f2 (ΓJn +1 ) − f2 (ΓJn ), · · · , f2 (Γ2 ) −
f2 (Γ1 ) is absolutely continuous with respect to the Lebesgue measure on RJn . Thus, the partial
derivatives of the function Fen exist. Then from a Taylor-Lagrange expansion,
Fen (x1 + u1n , · · · , xJn + uJnn ) = Fen (x1 , · · · , xJn ) +
Jn
X
ujn ×
j=1
j
j=1 un
PJn
where (x01 , · · · , x0Jn ) ∈ (0, ∞)Jn . Hence, we obtain
Jn
C 0 log
n
for some positive real numbers C and
C 0.
Fn (x1 , · · · , xJn )
×
∂
Fn (x01 , · · · , x0Jn ),
∂xj
∂
0
∂xj Fn (x1 , · · ·
, x0Jn ) ≤ C
j
j=1 un
PJn
≤
Consequently, we have:
Fen (x1 , · · · , xJn ).
∼
n→∞
Now, we are going back to the proof of Proposition 2 by computing Fen (x1 , · · · , xJn ). This leads to
compute the following integral:
Z yJ −1 +eJ −2 Z
Z y1 Z y2 +e1 Z y3 +e2
n
n
···
0
e1
e2
eJn −2
yJn +eJn −1
e−e1 e−e2 e−e3 · · · e−eJn −1 e−eJn de1 de2 de3 · · · deJn −1 deJn ,
eJn −1
with yi = xi /C2 , and with the same iteration than in the proof of Proposition 2, we obtain
Fen (x1 , · · · , xJn ) =
Jn
Y
1 − e−jxJn −j+1 /C2 .
j=1
Then, by considering the vector ((n − j)τj0 )n−Jn ≤j≤n−1 and the continuity of the function max, we have
for all x ≥ 0
Pr
max
{(n − j)τj0 } ≤ x
j=n−Jn ,··· ,n−1
∼
1 − e−x/C2
Jn
.
(6.4)
n→∞
To achieve the proof, we use the Slutsky’s Theorem. Indeed, we have sJn =
1
Jn
PJn
0
j=1 j τn−j
with
Jn → ∞. As a consequence,
s Jn
=
PJn
Jn
C2 X
j=1 j
j log Γj+1 /Γj + O
Jn
Jn log(n)
(6.5)
Jn
C2 X
Jn
j log Γj+1 /Γj + O
Jn
log(n)
(6.6)
Jn
1 X
Jn
εj + O
,
Jn
log(n)
(6.7)
j=1
=
j=1
=
j=1
with (ε1 , · · · , εJn ) a family of i.i.d.r.v. with exponential distribution of parameter 1/C2 . Using the Law
of Large Numbers and the condition Jn / log(n) −→ 0, we deduce that
n→∞
P
sJn −→ C2 .
n→∞
Then the proof can be concluded using (6.4), (6.8) and Slutsky’s Theorem.
(6.8)
Proof of Proposition 4. We begin by considering the proof of Proposition 1. Hence, since G−1 ∈ A2
and always with Γn+1
∼
n almost surely, we obtain for k = 1, · · · , J,
n→∞
G−1 Γ /Γ
Γ−a g (1/Γ
n+1
1
n+1 ) 1 + O(1/ log(Γn+1 ))
k
k
log
= log
G−1 Γk+1 /Γn+1
Γ−a
k+1 g1 (1/Γn+1 ) 1 + O(1/ log(Γn+1 ))
= −a log Γk /Γk+1 + O(1/ log(n)).
Then, we directly use the proof of Proposition 2.
Proof of Theorem 3.1. First consider the case G−1 ∈ A01 . Using the proof of Proposition 1, we have
G−1 Γ /Γ
f (Γ ) − f2 (Γj )
1
j
n+1
= 2 j+1
log
+O
.
2
−1
log(n)
G Γj+1 /Γn+1
log (n)
Consequently, using G−1 ∈ A01 and therefore the definition of f2 , we obtain:
log(τj ) =
C2
1
.
log Γj+1 /Γj + O
2
log(n)
log (n)
To prove (3.4), it is sufficient to use again the proof of Proposition 3, to normalize the numerator and
b Jn , which converges in probability to log 2/C2
denominator with log n and therefore to consider log n × L
(indeed, the median of the sample converges to log 2/λ which is the median of the distribution of the
E(λ) distribution).
When G−1 ∈ A2 , we can use the same argument that the ones of the proof of Proposition 3 with C2
replaced by a (the reminder 1/ log n obtained from the definition of A2 allows the achievement the proof
when Jn is negligible compared to log n).
Acknowledgements. The authors are extremely grateful to the referees for their impressive and
scrupulous reports with many relevant corrections, suggestions and comments which helped to improve
the contents of the paper. We also would like to thank Mia Hubert for having kindly provided R routines.
References
Abramowitz, M. and Stegun, I.A. (1964). Handbook of Mathematical Functions with Formulas, Graphs,
and Mathematical Tables. New York: Dover Publications.
Barnett,V. and Lewis, T. (1994). Outliers in Statistical Data. Wiley Series in Probability & Statistics,
Wiley.
Beckman, R.J. and Cook, R.D. (1983). Outlier....s, Technometrics, 25, 119-149.
Beirlant, J., Goegebeur, Y., Teugels, J. and Segers, J. (2004). Statistics of Extremes: Theory and
Applications. Wiley Series in Probability & Statistics, Wiley.
Beirlant, J., Vynckiera P. and Teugels, J. (1996). Tail Index Estimation, Pareto Quantile Plots Regression Diagnostics, Journal of the American Statistical Association, 91, 1659-1667.
Blair, J.M., Edwards C.A. and Johnson J.H. (1976). Rational Chebyshev approximations for the inverse
of the error function, Math. Comp. 30, 827-830.
Breunig, M.M., Kriegel, H.-P., Ng, R.T. and Sander, J. (2000). LOF: Identifying Density-based Local
Outliers. Proceedings of the 2000 ACM SIGMOD International Conference on Management of Data.
Embrechts, P., Kleppelberg, C. and Mikosch, T. (1997). Modelling Extreme Events for Insurance and
Finance. Springer.
Gnanadesikan, R. and Kettenring, J.R. (1972). Robust estimates, residuals, and outlier detection with
multiresponse data, Biometrics, 28, 81-124.
Grubbs, F.E. (1969). Procedures for Detecting Outlying Observations in Samples. Technometrics, 11,
1-21.
Hawkins, D.M. (1980). Identification of Outliers. Chapman and Hall
Hill, B. (1975). A simple general approach to inference about the tail of a distribution. The Annals of
Statistics, 3, 1163-1174.
Hubert, M., Dierckx, G. and Vanpaemel, D. (2013). Detecting influential data points for the Hill
estimator in Pareto-type distributions. Computational Statistics and Data Analysis, 65, 13-28.
Hubert, M. and Van der Veeken, S. (2008). Outlier detection for skewed data. Journal of Chemometrics,
22, 235-246.
Hubert, M. and Vandervieren, E. (2008). An adjusted boxplot for skewed distributions. Computational
Statistics and Data Analysis, 52, 5186-5201.
Knorr, E.M., Ng, R.T. and Tucakov, V. (2000). Distance-based outliers: algorithms and applications,
The VLDB Journal, 8, 237-253.
Rousseeuw, P.J. and Hubert, M. (2011). Robust statistics for outlier detection WIREs Data Min.
Knowl. Disc., 1, 73-79.
Rousseeuw, P.J. and Leroy, A.M. (2005). Robust Regression and Outlier Detection. Wiley Series in
Probability and Statistics, Wiley.
Tietjen, G.L. and Moore, R.H. (1972). Some Grubbs-Type Statistics for the Detection of Several
Outliers, Technometrics, 14, 583-597.
Tse, Y.K. and Balasooriya, U. (1991). Tests for Multiple Outliers in an Exponential Sample, The
Indian Journal of Statistics, Series B, 53, 56-63.
Tukey, J.W. (1977). Exploratory data analysis. Reading, MA, Addison-Wesley.
| 10math.ST
|
1
Private Information Retrieval from Transversal
Designs
Julien L AVAUZELLE
Laboratoire LIX, École Polytechnique, Inria & CNRS UMR 7161
Université Paris-Saclay
arXiv:1709.07952v1 [cs.IT] 22 Sep 2017
Abstract
Private information retrieval (PIR) protocols allow a user to retrieve entries of a database without revealing the
index of the desired item. Information-theoretical privacy can be achieved by the use of several servers and specific
retrieval algorithms. Most of known PIR protocols focus on decreasing the number of bits exchanged between the
client and the server(s) during the retrieval process. On another side, Fazeli et. al. introduced so-called PIR codes in
order to reduce the storage overhead on the servers. However, only a few works address the issue of the computation
complexity of the servers.
In this paper, we show that a specific encoding of the database provides PIR protocols with reasonable communication complexity, low storage overhead and optimal computational complexity for the servers. This encoding
is based on incidence matrices of transversal designs, from which a natural and efficient recovering algorithm is
derived. We also present instances of our construction, making use of finite geometries and orthogonal arrays, and
we finally give a generalisation of our main construction for resisting collusions of servers.
Index Terms
Private information retrieval, linear code, transversal design, orthogonal array, finite geometry
I. I NTRODUCTION
A. Private Information Retrieval
A private information retrieval (PIR) protocol aims at ensuring a user that he can retrieve some part Di of a remote
database D without revealing the index i to the server(s) which hold the database. For example, such protocols
can be applied in medical data storage where physicians would be able to access parts of the genome while hiding
the specific gene they analyse. The PIR paradigm was originally introduced by Chor, Goldreich, Kushilevitz and
Sudan [6, 7].
A naive solution to the problem would consist in downloading the entire database each time the user wants a single
symbol. But the communication complexity would then be overwhelming, so we look for PIR protocols exchanging
less bits. However, Chor et. al. proved that, when the k -bits database is stored on a single server, a PIR protocol
cannot be information-theoretically secure with less than Θ(k) bits of communication [7]. Two alternatives were
then considered: restricting the protocol to computational security (initiated by Chor and Gilboa [5]), or allowing
several servers to store the database. Our work focuses on the last one.
In many such PIR protocols the database is replicated on the servers. Informally, the idea is that each server is asked
to compute some partial information related to a random-like query sent by the user, the index of the desired item
being hidden in each query. Then the user collects all the servers’ answers and retrieves the desired symbol with
an appropriate algorithm. For instance, Chor et. al. [7] use XOR properties on log(k)-dimensional binary vectors
to mask the index of the desired item and to retrieve the associated symbol. Their protocol achieves decreasing
communication as a function of the number of servers: with ` servers, the communication is O(` log ` k 1/ log ` ) bits.
A few years later, Katz and Trevisan [13] showed that any smooth locally decodable code (LDC) gives rise to a
Submitted to editors on July 28, 2017. A shorter version of this paper was presented at the Tenth International Workshop on Coding and
Cryptography 2017, September 18-22, 2017, Saint-Petersburg, Russia.
2
PIR protocol whose number of servers and communication complexity is basically the same than the LDC locality
([20] is a good survey on LDCs and their applications in PIR protocols). Building on this idea,√many PIR schemes
(notably [3, 19, 10, 9]) successively decreased the communication complexity, achieving O(k log log k/ log k ) with
only two servers. However, only few of them tried to lighten the computational and storage cost on the server side.
By preprocessing the database, Beimel, Ishai and Malkin [4] were the first to address the minimization of the server
storage/computation. Then, initiated by Fazeli, Vardy and Yaakobi [11], recent works used the concept of PIR codes
to address the storage issue. The idea is to turn an `-server replication-based PIR protocol into a more-than-`-server
distributed PIR protocol with a smaller overall storage overhead. For this purpose, the user encodes the database and
distributes pieces of the associated codeword among the servers, such that they hold distinct parts of the database
(plus some redundancy). Through this transformation, both communication complexity and computational cost keep
the same order of magnitude, but the storage overhead corresponds to the PIR code’s one, which can be brought
arbitrarily close to 1 when sufficiently many servers are used. Several recent works also address the PIR issue on
previously coded databases [18], and/or aim at reaching the so-called capacity of the model [17]. However, while
the storage drawback seems to be solved, potentially huge computational cost can still represent a barrier to the
practicality of such PIR protocols.
B. Motivations and results
Indeed, as pointed out by Yekhanin [20], “the overwhelming computational complexity of PIR schemes (...) currently
presents the main bottleneck to their practical deployment”. Consider a public database which is frequently queried,
e.g. a database storing stock exchange prices where private queries could be very relevant. Fast retrieval is necessary
is this context, hence one cannot afford each run of the PIR protocol to be computationally inefficient, for instance
Ω(k), where k is the size of the database. Therefore, we aim at building PIR protocols with sublinear computational
complexity in the length of the database stored by each server.
Naively, the computational complexity of a PIR protocol could be drastically reduced by letting all the possible
answers to its queries to be precomputed. Of course, storing all these answers dramatically increases the needed
storage, and we prefer to focus on another construction due to Augot, Levy-dit-Vehel and Shikfa [2], anterior to
the PIR codes breakthrough [11], and that we shortly explain now.
Their construction uses a specific family of high-rate locally decodable codes called multiplicity codes introduced
by Kopparty, Saraf and Yekhanin [14]. But instead of replicating the database on the several servers, the authors
split an encoded version c of the database D into parts c(1) , . . . , c(`) , and share these parts on the servers. The main
difference with PIR codes [11] is that Augot et. al.’s construction does not purpose to emulate a lighter PIR protocol
with an existing one, but uses specific properties of the encoding as a way to split the database on several servers.
In short, the multiplicity codes they use feature both the privacy of the PIR protocol and the storage reduction for
the servers. We refer to Section VII for more details on the construction.
In this work, we consider this “codeword support splitting” idea in the context of incidence properties in block
designs. We thus build a large family of codes leading to PIR schemes with low storage and low computational
complexity on the server side.
C. Organisation
We start by giving two formal definitions of PIR protocols in Section II, depending on whether the database is
replicated or distributed on the servers. We also present the standard construction of replication-based PIR protocols
from smooth locally decodable codes. Section III is devoted to recall definitions of combinatorial structures and
their associated codes that are used to build our 1-private PIR protocols in Section IV. Section V lists instances
of our construction with efficient parameters. Finally, a generalisation of our construction is given in Section VI
in order to keep up with collusions of servers, and a comparison with the PIR protocols coming from multiplicity
codes is presented in Section VII.
3
II. D EFINITIONS AND RELATED CONSTRUCTIONS
We first recall that we are only concerned with information-theoretically secure PIR protocols. In this paper, we
denote by U the user (or client) of the PIR protocol. User U owns a database denoted by D = (Di )1≤i≤k ∈ Fkq ,
where Fq represents the finite field with q elements. Database D hence contains |D| = k log q bits. We also denote
by S1 , . . . , S` the ` servers involved in the PIR protocol.
Given A, B two sets, with |B| = n < ∞, we denote by AB the set of n-tuples a = (ab )b∈B of A-elements indexed
by B , which can also be seen as functions from B to A. For T ⊂ B , we also write a|T := (at )t∈T the restriction
of the tuple a to the coordinates of T .
A. Two definitions for PIR protocols
A vast majority of existing PIR schemes start by simply cloning the database D on all the servers S1 , . . . , S` . Then,
the role of each server Sj is to compute some combination of symbols from D, related to the query sent by U .
This computation has a non-trivial cost, so in a certain sense, the computational complexity of the privacy of the
PIR scheme is mainly devoted to the servers.
More formally, one can define replication-based PIR protocols as follows:
Definition II.1 (standard, or replication-based PIR protocol). Assume that every server Sj , 1 ≤ j ≤ `, stores a
copy of the database D. An `-server replication-based PIR protocol is a set of three algorithms (Q, A, R) running
the following steps on input i ∈ [1, k]:
1) Query generation: the randomized algorithm Q generates ` queries Q(i) := (q1 , . . . , q` ). Query qj is sent to
server Sj .
2) Servers’ answer: each server Sj computes an answer aj = A(qj , D) and sends it back to the user1 .
3) Reconstruction: denote by a = (a1 , . . . , a` ) and q = (q1 , . . . , q` ). User U computes and outputs r = R(i, a, q).
The PIR protocol is said:
– correct if r = Di when the servers follow the protocol;
– t-private if, for every (i, i0 ) ∈ [1, k]2 and every T ⊆ [1, `] such that |T | ≤ t, the distributions Q(i)|T and
Q(i0 )|T are the same. We also say that the PIR protocol resists t collusions of servers.
We call communication complexity the number of bits sent between the user and the servers, and server (resp. user)
computational complexity the maximal number of Fq -operations made by a server in order to compute an answer
aj (resp. made by R to reconstruct the desired item).
According to this definition, one sees that the servers must jointly carry the ` copies of the database, so the storage
overhead of the scheme is (` − 1)|D| bits. Moreover, since D is a raw database without specific structure, the
algorithm A has no reason to be trivial and can incur superlinear computations for the servers — which is verified
for most of current replication-based PIR protocols.
A way to reduce the computation cost of PIR protocols is to preprocess the database. Therefore we need to model
PIR protocols for which the database can be encoded and distributed over the servers. From now on, let c = (ci )i∈I
denote an encoding of the database D, i.e. the image of D by an injective map Fkq → FIq , with |I| = n ≥ k . Besides,
(i )
(j)
for convenience we assume that I = [1, s] × [1, `] and for readability we write c(i1 ,i2 ) = ci12 and c(j) = (cr )r∈[1,s] .
Definition II.2 (distributed PIR protocol). Assume that for 1 ≤ j ≤ `, server Sj holds the part c(j) of the encoded
database. An `–server distributed PIR protocol is a set of three algorithms (Q, A, R) running the following steps
on input i ∈ I :
1) Query generation: the randomized algorithm Q generates ` queries Q(i) := (q1 , . . . , q` ). Query qj is sent to
server Sj .
2) Servers’ answer: each server Sj computes an answer aj = A(qj , c(j) ) and sends it back to the user.
3) Reconstruction: denote by a = (a1 , . . . , a` ) and q = (q1 , . . . , q` ). User U computes and outputs r = R(i, a, q).
1
algorithm A := Aj may depend on j
4
Correctness and privacy properties are identical to those of replication-based PIR protocols. Similarly, one can
also define communication and computational complexities, and since the database D has been encoded, we finally
define the storage overhead as the number of redundancy bits stored by the servers, that is, (s` − k) log q .
In this paper, we focus on distributed PIR protocols with low computational complexity on the server side. More
precisely, we build PIR protocols where the answering algorithm A consists only in reading some symbols of the
database. Thus, our PIR protocols are computationally optimal on the server side, in a sense that, compared to the
non-private retrieval, they incur no extra computational burden for the each server taken individually.
B. PIR protocols from locally decodable codes
As pointed out in the introduction, Augot et. al. [2] used a family of locally decodable codes (LDC) to design a
distributed PIR scheme. LDCs are known to give rise to PIR protocols for a long time, but we emphasize that the
main idea from [2] is to benefit from the fact that the encoded database can be smartly partitioned with respect to
the queries of the local decoder.
Based on the seminal work of Katz and Trevisan [13], we briefly remind how to design a PIR protocol based on
a perfectly smooth locally decodable code. First, let us define (linear) locally decodable codes.
Definition II.3 (locally decodable code). Let Σ be a finite set, 2 ≤ ` ≤ k ≤ n be integers, and δ, ∈ [0, 1]. A code
C : Σk → Fnq is (`, δ, )–locally decodable if and only if there exists a randomized algorithm D such that, for every
input i ∈ [1, k] we have:
– for all m ∈ Σk and all y ∈ Fnq , if |{j ∈ [1, n], yj 6= C(m)j }| ≤ δn, then
P(D(y) (i) = mi ) ≥ 1 − ,
where the probability is taken over the internal randomness of D;
– D reads at most ` symbols yq1 , . . . , yq` of y .
Notation D(y) refers to the fact that D has oracle access to single symbols yqj of the word y . The parameter `
is called the locality of the code. Moreover, the code C is said perfectly smooth if on an arbitrary input i, each
individual query of the decoder D is uniformly distributed over the coordinates of the word y .
Now let us say a user wants to use a PIR protocol on a database D ∈ Σk , and assume there exists a perfectly
smooth locally decodable code C ⊂ Fnq of dimension k and locality `. Figure 1 presents a distributed PIR protocol
based on C .
1) Initialisation step. User U encodes D into a codeword c0 ∈ C . Each server S1 , . . . , , S` holds a copy
of c0 . In the formalism of Definition II.2, it means that c(j) := c0 , for j = 1, . . . , `.
2) Retrieving step for symbol Di . Denote by D a local decoding algorithm for C .
1) Queries generation: user U calls D to generate at random a query (q1 , . . . , q` ) for decoding the
symbol Di . Query qj is sent to server Sj .
2) Servers’ answer: each server Sj reads the encoded symbol aj := c0qj . Then Sj sends aj to U .
3) Reconstruction: user U collects the ` codeword symbols (c0qj )j∈[1,`] and feeds the local decoding
algorithm D in order to retrieve Di .
Fig. 1: A distributed PIR protocol based on a locally decodable code C .
The main drawback of these LDC-based PIR protocols is their storage overhead, since the ` servers must store
`n/k = `/R times more data than the raw database (R := k/n represents the information rate, or rate, of the code).
This issue becomes especially crucial as building LDCs with small locality and high rate is highly non-trivial.
The idea of Augot, Levy-dit-Vehel and Shikfa [2] for reducing the storage overhead is to benefit from a natural
partition of the support of multiplicity codes [14]. Assume that each codeword c ∈ C can be split into ` disjoint
5
parts c(1) , . . . , c(`) , such that each coordinate qj of any possible query (q1 , . . . , q` ) of the PIR protocol corresponds
to reading some symbols on c(j) . By sending the part c(j) to server Sj , the PIR protocol of Figure 1 can be
improved in order to save storage. We devote Section VII to more explanation on this construction, as well as to
a comparison with our schemes.
Finally, one can notice that the communication complexity of LDC-based PIR protocols depends on the locality of
the code, while its smoothness serves their privacy. We also point out two important remarks.
1) Assuming a noiseless transmission and honest-but-curious servers (i.e. they want to discover the index of
the desired symbol but never give wrong answers), one does not need a powerful local decoding algorithm.
Indeed, it should be possible to reconstruct the desired symbol Di by local decoding only one erasure on the
codeword. For instance, computing a single low-weight parity-check sum should be enough.
2) Smoothness is sufficient for 1-privacy, but we need more structure for preventing collusions of servers.
Coupled with the fact that we want to split the database over several servers, these remarks lead us to design other
kinds of encoding, which answer as close as possible the needs of private information retrieval protocols. Our
construction relies on combinatorial structures, namely transversal designs, that we recall in the upcoming section.
III. T RANSVERSAL DESIGNS AND CODES
Let us give here the definition of transversal designs and how to build linear codes upon them. We refer to [1], [16]
and [8] for complementary details.
Definition III.1 (block design). A block design is a pair D = (X, B) where X is a finite set of so-called points,
and B is a set of non-empty subsets of X called the blocks.
Definition III.2 (incidence matrix). Let D = (X, B) be a block design. An incidence matrix MD of D is a matrix
of size |B| × |X|, whose (i, j)−entry, for i ∈ B and j ∈ X , is:
1 if the block i contains the point j
0 otherwise
The q -rank of MD is the rank of MD over the field Fq . We also define the incidence vector 1B ∈ {0, 1}B of a
block B ∈ B as the row vector indexed by B in MD .
Of course, any design admits many incidence matrices, depending on the way points and blocks are ordered.
However, all these incidence matrices are equal up to some permutation of their rows and columns, and, in particular,
they all have the same q -rank. Hence, we call q -rank of a design the q -rank of any of its incidence matrices. Moreover,
from now on we consider incidence matrices of designs up to an ordering of points and blocks, and we abusively
refer to the incidence matrix MD of a design D.
Example III.3. Let A2 (F3 ) be the affine plane over the finite field F3 , and X be the set consisting of its 9 points:
X = {(0, 0), (0, 1), (0, 2), (1, 0), (1, 1), (1, 2), (2, 0), (2, 1), (2, 2)} .
We define the block set B as the set of the 12 affine lines of A2 (F3 ):
B = { {(0, 0), (0, 1), (0, 2)},
{(0, 0), (1, 1), (2, 2)},
{(0, 0), (2, 1), (1, 2)},
{(0, 0), (1, 0), (2, 0)},
{(1, 0), (1, 1), (1, 2)},
{(1, 0), (2, 1), (0, 2)},
{(1, 0), (0, 1), (2, 2)},
{(0, 1), (1, 1), (2, 1)},
{(2, 0), (2, 1), (2, 2)}
{(2, 0), (0, 1), (1, 2)}
{(2, 0), (1, 2), (0, 2)}
{(0, 2), (1, 2), (2, 2)} } .
6
The pair D = (X, B) is then a block design, and its associated (12 × 9)–incidence matrix is
1 1 1 0 0 0 0 0 0
1 0 0 1 0 0 1 0 0
1 0 0 0 1 0 0 0 1
1 0 0 0 0 1 0 1 0
0 1 0 1 0 0 0 0 1
0 1 0 0 1 0 0 1 0
.
MD =
0
1
0
0
0
1
1
0
0
0 0 1 1 0 0 0 1 0
0 0 1 0 1 0 1 0 0
0 0 1 0 0 1 0 0 1
0 0 0 1 1 1 0 0 0
0 0 0 0 0 0 1 1 1
A computation shows that over the field F2 , matrix MD is full-rank, while over F3 , it has only rank 6.
Definition III.4 (transversal design). Let s, ` ≥ 2 and λ ≥ 1 be integers. A transversal design, denoted TDλ (`, s),
is a block design (X, B) equipped with a partition G = {G1 , . . . , G` } of X called the set of groups, such that:
– |X| = `s;
– any group in G has size s and any block in B has size `;
– any unordered pair of elements from X is contained either in exactly one group or in exactly λ blocks.
When λ = 1, we use the simpler notation TD(`, s).
Remark III.5. A block cannot be secant to a group in more than one point, otherwise the third condition of the
definition would be disproved. Moreover, since the block size equals the number of groups, any block must meet
any group. Hence the following holds:
∀(B, G) ∈ B × G, |B ∩ G| = 1 .
The definition also implies there must lie exactly λs2 blocks in B .
Example III.6. Let D = (X, B) be the block design defined in Example III.3. Define G to be any set of 3 parallel
lines from B which partitions the point set X . For instance, one can consider
G = { {(0, 0), (0, 1), (0, 2)}, {(1, 0), (1, 1), (1, 2)}, {(2, 0), (2, 1), (2, 2)} } .
Then, T = (X, B \ G, G) is a transversal design TD(3, 3). Indeed, T is composed of `s = 9 points, ` = 3 groups
of size s = 3 and s2 = 9 blocks of size ` = 3 each. Moreover, in the affine plane every unordered pair of points
belongs simultaneously to a unique line, which is represented in T either by a group or by a block. More generally,
for any prime power q , a transversal design TD(q, q) can be built with the affine plane A2 (Fq ). A generalisation
of this construction will be given in Subsection V-A.
A simple way to build linear codes from block designs is to associate a parity-check equation of the code to each
incidence vector of a block of the design. We recall P
that the dual code C ⊥ of a code C ⊆ Fnq is the linear vector
space consisting of vectors h ∈ Fnq such that ∀c ∈ C, ni=1 ci hi = 0.
Definition III.7 (code of a design). Let Fq be a finite field, D = (X, B) be a block design and MD be its incidence
matrix. The code Codeq (D) is the Fq -linear code of length |X| whose dual code is spanned by the rows of MD .
Remark III.8. The code Codeq (D) is uniquely defined up to a chosen order of the points X . For different orders,
the arising codes remain permutation-equivalent. Also notice that the way blocks are ordered does not affect the
code.
For any design D, the dimension over Fq of Codeq (D) equals |X| − rankq (MD ). Since MD has coefficients in
{0, 1}, one must notice that rankq (MD ) = rankp (MD ), where p is the characteristic of the field Fq .
7
Remark III.9. Standard literature (e.g. [1]) sometimes defines Codeq (D) (and not Codeq (D)⊥ ) to be the vector
space generated by the incidence matrix of the design. We favor this convention because Codeq (D) will serve to
encode the database in our PIR scheme.
Example III.10. The design D from Example III.3 gives rise to C
9 and dimension 3. A full-rank generator matrix of C is given by:
1 1 1 1 1 1 1
G= 0 0 0 1 1 1 2
0 1 2 0 1 2 0
= Code3 (D), a linear code over F3 , of length
1 1
2 2 .
1 2
One may notice that this code is the generalized Reed-Muller code of degree 1 and order 2 over F3 , that is, the
evaluation code of bivariate polynomials of total degree at most 1 over the whole affine plane F23 .
Definition III.11 (systematic encoding). Let C ⊆ Fnq be a linear code of dimension k ≤ n. A systematic encoding
for C is a one-to-one map φ : Fkq → C , such that there exists an injective map σ : [1, k] → [1, n] satisfying:
∀m ∈ Fkq , ∀i ∈ [1, k], mi = φ(m)σ(i) .
The set σ([1, k]) ⊆ [1, n] is called an information set of C .
In other words, a systematic encoding allows to view the message m as a subword of its associated codeword
φ(m) ∈ C . For instance, it is useful for efficiently retrieving m from c when the codeword c has not been corrupted.
A systematic encoding exists for any code C , is not necessarily unique, and can be computed through a Gaussian
elimination over any generator matrix of the code. Also notice that this computation can be tedious for large codes.
IV. 1- PRIVATE PIR PROTOCOLS BASED ON TRANSVERSAL DESIGNS
In this section we present our construction of PIR protocols relying on transversal designs. The idea is that the
knowledge of one point of a block of a transversal design gives (almost) no information on the other points lying
on this block. The code associated to such a design then transfers this property to the coordinates of codewords.
Hence, we obtain a PIR protocol which can be proven 1-private, that is, which ensures perfect privacy for noncommunicating servers. Though this protocol cannot resist collusions, we will see in Section VI that a natural
generalisation leads to t-private PIR protocols with t > 1.
Notice that both Fazeli et. al.’s work [11] and ours make use of codes in order to save storage in PIR protocols.
Nevertheless, we emphasize that the constructions are very different, since Fazeli et. al. emulate a PIR protocol
from an existing one while we build our PIR protocols from scratch.
A. The transversal-design-based distributed PIR protocol
Let T be a transversal design TD(`, s) and n = |X| = `s. Denote by C = Codeq (T ) ⊆ Fnq the associated Fq -linear
code, and let k = dimFq C . Our PIR protocol is defined in Figure 2. We then summarize the steps of the construction
in Figure 3.
B. Analysis
We analyse our PIR scheme by proving the following:
Theorem IV.1. Let D be a database with k entries over Fq , and T = TD(`, s) be a transversal design, whose
incidence matrix has rank `s − k over Fq . Then, there exists a distributed `-server 1-private PIR protocol with:
–
–
–
–
only one Fq -symbol to read for each server,
` − 1 field operations over Fq for the user,
` log(sq) bits of communication (` log s are uploaded, ` log q are downloaded),
a (total) storage overhead of (`s − k) log q bits on the servers.
8
Parameters: T = (X, B, G) is a TDλ (`, s); C = Codeq (T ) has length n = `s and dimension k .
1) Initialisation step.
1) Encoding. User U computes a systematic encoding of the database D ∈ Fkq , resulting in the
codeword c ∈ C .
2) Distribution. Denote by c(j) = c|Gj the symbols of c whose support is the group Gj ∈ G . Each
server Sj receives c(j) , for 1 ≤ j ≤ `.
2) Retrieving step for symbol ci for i ∈ X . Denote by j ∗ ∈ [1, `] the index of the unique group Gj ∗
(j ∗ )
which contains i — that is, ci = cr for some r ∈ [1, s]. Also denote by B ∗ the subset of blocks
containing i. The three steps of the distributed PIR protocol are:
1) Queries generation. U picks uniformly at random a block B ∈ B ∗ . For j 6= j ∗ , user sends the
unique index qj ∈ B ∩ Gj to server Sj . Server Sj ∗ receives a random query qj ∗ uniformly picked
$
in Gj ∗ . To sum up (←
− stands for “picked uniformly at random in”) :
$
− Gj ∗ for j ∗ such that i ∈ Gj ∗
Q(i)j ∗ ←
$
B ←
− B∗
Q(i)j ← B ∩ Gj ,
j 6= j ∗
2) Servers’ answer. Each server Sj (including Sj ∗ ) reads aj := cqj and sends it back to the user. That
is,
A(qj , c(j) ) = cqj .
3) Reconstruction. Denote by a = {a1 , . . . , a` } and q = {q1 , . . . , q` }. User U computes
X
X
r = R(i, a, q) := −
aj = −
cqj
j6=j ∗
j6=j ∗
and outputs r.
Fig. 2: A 1-private distributed PIR protocol based on the Fq -linear code defined by a transversal design.
Transversal design
incidence
matrix
TD-based
/ linear code
database
encoding
/ Distributed PIR scheme
Fig. 3: Summary of steps the construction of a transversal-design-based PIR scheme.
Proof. Recall the PIR protocol we are dealing with is defined in Figure 2.
Correctness. By definition of the code C = Codeq (T ), the incidence vector 1B of anyPblock B ∈ B belongs to the
dual code C ⊥ . Hence, for c ∈ C , the inner product 1B · c vanishes, or said differently, x∈B cx = 0. We recall that
j ∗ represents the index of the group which contains i. Since P
the servers Sj , j 6=Pj ∗ , receive queries corresponding
to the points of a block B which contains i, we have ci = − x∈B\{i} cx = − j6=j ∗ cqj , and our PIR protocol is
correct as long as there is no error on the symbols aj := cqj returned by the servers.
Security (1-privacy). We need to prove that for all j ∈ [1, `], it holds that P(i | qj ) = P(i), where probabilities are
taken over the randomness of B ← B ∗ . The law of total probability implies
P(i | qj ) = P(i | qj and i ∈ Gj )P(i ∈ Gj ) + P(i | qj and i ∈
/ Gj )P(i ∈
/ Gj )
= P(i | i ∈ Gj )P(i ∈ Gj ) + P(i | i ∈
/ Gj )P(i ∈
/ Gj ) = P(i) ,
and the reasons why we eliminated the random variable qj in the conditional probabilities are:
– in the case i ∈ Gj (that is, j = j ∗ ), by definition of our PIR protocol we know that qj is uniformly random,
so qj and i are independent;
9
– in the case i ∈
/ Gj , by definition of a transversal design, there are as many blocks containing both qj and i as
there are blocks containing qj and any i0 in X \ Gj (the number of such blocks is always λ). So once again,
the value of the random variable qj is not related to i.
Communication complexity. Exactly one index in [1, s] and one symbol in Fq are exchanged between each server
and the user. So the overall communication complexity is ` × (log(s) + log(q)) = ` log(sq) bits.
Storage overhead. The number of bits stored on a server is s log q , giving a total storage overhead of (`s − k) log q ,
where k = dim C .
Computation complexity. Each server Sj only needs to read the symbol defined by query qj , hence our protocol
incurs no extra computational cost.
Theorem IV.1 shows that, if we want to optimize the practical parameters of our PIR scheme, we basically need
to look for small values of `, the number of groups. However, one observes that the dimension k of Codeq (T )
strongly depends on ` and n, and tiny values of ` can lead to trivial or very small codes. This issue should be
carefully taken into account, since instances with k < ` represent PIR protocols which are more communication
expensive to use than the trivial one which simply retrieves the whole database. Hence, it is very natural to raise
the main issue of our construction:
Problem IV.2. Find codes C = Codeq (T ) arising from transversal designs T = TD(`, s) with few groups (small
`) and large dimension k = dimFq C compared to their length n = `s.
The following section is devoted to the construction of such transversal designs.
V. E XPLICIT CONSTRUCTIONS OF 1- PRIVATE TD- BASED PIR PROTOCOLS
From now on, we denote by `(k) the number of servers involved in a given PIR protocol running on a database
with k entries, and by n(k) the actual number of symbols stored by all the servers. As it is proved in Theorem IV.1,
these two parameters are crucial for the practicality of our PIR schemes, and they respectively correspond to the
block size and the number of points of the transversal design used in the construction. In practice, we look for
small values of ` and n as explained in Problem IV.2.
In this section, we first give two classical instances of transversal designs derived from finite geometries (Subsections V-A and V-B), leading to good PIR parameters. We then show how orthogonal arrays produce transversal
designs, and we more deeply study a family of such arrays leading to high-rate codes. Subsection V-D is finally
devoted to another family of orthogonal arrays whose divisibility properties ensure to give an upper bound on the
storage overhead of related PIR protocols.
A. Transversal designs from affine geometries
Transversal designs can be built with incidence properties between subspaces of an affine space.
Construction V.1 (Affine transversal design). Let Am (Fq ) be the affine space of dimension m over Fq , and
H = {H1 , . . . , Hq } be q hyperplanes that partition Am (Fq ). We define a transversal design TA (m, q) as follows:
– the point set X consists in all the points in Am (Fq );
– the groups in G are the q hyperplanes from H ;
– the blocks in B are all the 1-dimensional affine subspaces (lines) which do not entirely lie in one of the Hj ,
j ∈ [1, q]. We also say that such lines are secant to the hyperplanes in H .
The design thus defined is a TD(q, q m−1 ), since an affine line is either contained in one of the Hj , or is 1-secant
(i.e. has intersection of size 1) to each of them. To complete the study of the parameters of the induced PIR protocol,
it remains to compute the dimension of Code(TA (m, q)).
First notice that all blocks of TA (m, q) belong to the block set of the affine geometry design AG1 (m, q) — which
is defined as the incidence structure of all points and affine lines in Am (Fq ). Thus, the incidence matrix MTA (m,q)
10
is a sub-matrix of MAG1 (m,q) , which implies that Codep (AG1 (m, q)) ⊆ Codep (TA (m, q)) for any field Fp . The
benefit to consider AG1 (m, q) is that the p-rank of its incidence matrix has been well-studied. Besides, when p
and q are coprime, it has been proved that its code over Fp has dimension 0 or 1 (Theorem 2.4.1 in [1]). Since
our goal is to obtain the largest codes as possible, we now assume that p is the characteristic of the field Fq .
In this setting (q = pe ), Hamada [12] gives a generic formula to compute the p-rank of a design coming from
projective geometry. Yet, as presented in Appendix A, asymptotics are hard to derive from his formula for a generic
value of m.
e ) = p+1 e , which implies that dim Code (T (2, pe )) ≥
However, when
m
=
2
,
we
know
that
rank
AG
(2,
p
p
1
p A
2
e
p2e − p+1
.
Hence
we
obtain
the
following
family
of
PIR
protocols.
2
e
Proposition V.2. Let D be a database with k = p2e − p+1
entries, p a prime, e ≥ 1. There exists a distributed
2
1-private PIR protocol for D with:
`(k) = pe and n(k) = p2e .
For fixed p and k → ∞, we have
√
1
`(k) = k + Θ(k 2 +cp )
and
1
n(k)/k =
1−
where cp =
1
2
1+1/p
2
e = 1 + Θ(k cp ) → 1 ,
(1)
logp ( 1+1/p
2 ) < 0.
Proof. The existence of the PIR protocol is a consequence of the previous discussion. Let us state the asymptotics
of the parameters. Recall we fix the prime p and we let e → ∞. First we have:
!
1
1 + 1/p 2e
p2e
1 + 1/p e
e = 1 +
.
(2)
+O
n(k)/k =
e =
2
2
p2e − p+1
1 − 1+1/p
2
2
Notice that
logp k = 2e + logp 1 −
1 + 1/p
2
e
= 2e + O
1 + 1/p
2
e
.
Hence,
1+1/p e
1
O(( 1+1/p )e )
2
1+1/p
1
1 + 1/p e
1
+
1/p
1 + 1/p 2 logp k+O(( 2 ) )
= k 2 logp ( 2 ) ×
= Θ (k cp ) ,
=
2
2
2
O(( 1+1/p )e )
2
since 1+1/p
→ 1, and with (2) we obtain the asymptotics we claimed on n(k)/k .
2
For `(k), we see that n(k) = `(k)2 . Therefore, we get
√ p
√ p
√
1
`(k) = k n(k)/k = k 1 + Θ(k cp ) = k + Θ(k 2 +cp ) .
We give in Table I lower bounds on the dimension of some practical codes arising from affine transversal designs,
when m is not restricted to 2. Computations show that the lower bound is tight for small values of p, e and m
(only a few of them are presented in Table I), and we conjecture it to be tight for all of them.
Finally, for a better understanding of the parameters we can point out two PIR instances:
– choosing m = 2 and ` = 4096, there exists a PIR protocol on a ' 2.0 MB file with 6 kB of communication
and only 3.2% storage overhead;
– for a ' 46 GB database (m = 3, ` = 8192), we obtain a PIR protocol with 39 kB of communication and 27%
storage overhead.
11
m
2
2
2
2
2
2
2
2
3
3
3
3
3
3
4
4
4
5
5
`=q
8
16
32
64
1024
4096
16 384
65 536
8
16
64
256
1024
8192
8
64
256
8
64
n = s` = q m
64
256
1024
4096
1 048 576
16 777 216
268 435 456
4 294 967 296
512
4096
262 144
16 777 216
1 073 741 824
549 755 813 888
4096
16 777 216
4 294 967 296
32 768
1 073 741 824
lower bound on
k = dim C
37∗
175∗
781∗
3367∗
989 527
16 245 775
263 652 487
4 251 920 575
139∗
1377∗
118 873
9 263 777
680 200 873
400 637 408 211
406
2 717 766
890 445 921
994
44 281 594
lower bound on
R = k/n
0.578∗
0.684∗
0.763∗
0.822∗
0.944
0.968
0.982
0.990
0.271∗
0.336∗
0.453
0.552
0.633
0.729
0.099
0.162
0.207
0.030
0.041
TABLE I: Lower bounds on the dimension and rate of binary codes arising from TA (m, q). A star denotes
the bound is reached (shown by computation). Remind that the rate R of the code is related to the server
storage overhead of the PIR protocol, and that q = ` is essentially the communication complexity and the
number of servers.
B. Transversal designs from projective geometries
Projective geometries are closely related to affine geometries, but contrary to them, there is no partition of the
projective space into hyperplanes, since every pair of distinct projective hyperplanes intersects in a projective space
of co-dimension 2. To tackle this problem, an idea is to consider the hyperplanes Hi which intersect on a fixed
subspace of co-dimension 2 (call it Π∞ ). Then, all the sets Hi \ Π∞ are disjoint, and their union gives exactly
Pm (Fq ) \ Π∞ , where Pm (Fq ) denotes the projective space of dimension m over Fq . Besides, any projective line
disjoint from Π∞ is either contained in one of the Hi , or is 1-secant to all of them. It results to the following
construction:
Construction V.3 (Projective transversal design). Let Pm (Fq ) and Π∞ defined as above. Let us define
– a point set X = Pm (Fq ) \ Π∞ ;
– a group set G = {projective hyperplanes H ⊂ Pm (Fq ), Π∞ ⊂ H};
– a block set B = {projective lines L ⊂ Pm (Fq ), L ∩ Π∞ = ∅ and ∀H ∈ G, L 6⊂ H} .
Finally, denote by TP (m, q) := (X, B, G).
The design TP (m, q) is a TD(q + 1, (q + 1)q m−1 ) and, as in the affine setting, its p-rank is related to that of
PG1 (m, q), the classical design of point-line incidences in the projective space Pm (Fq ). Indeed, the incidence
matrix M of TP (m, q) is a submatrix of MPG1 (m,q) from which we removed:
– the columns corresponding to the points in Π∞ ,
– the rows corresponding to the lines not in B .
Said differently, the code associated to TP (m, q) contains (as a subcode) the Π∞ -shortening of the code associated
to PG1 (m, q). Hence dim Code(TP (m, q)) ≥ dim Code(PG1 (m, q)) − |Π∞ |. Once again, Hamada’s formula gets
simpler for m = 2, and leads to the following proposition.
e
Proposition V.4. Let D be a database with k = p2e + pe − p+1
− 1 entries, p a prime and e ≥ 1. There exists
2
a distributed 1-private PIR protocol for D with:
`(k) = pe + 1
and
n(k) = p2e + pe .
12
Asymptotics are the same as in Equation (1).
In order to emphasize that the two previous constructions are asymptotically the same, we draw the rates of the
codes involved in these two kinds of PIR schemes in Figure 4.
R = k/n (rate)
1
0.9
√
` = Θ( k)
0.8
0.7
0.6
` = Θ(k 1/3 )
0.5
0.4
0.3
` = Θ(k 1/4 )
0.2
0.1
0
` = Θ(k 1/5 )
210
220
230
240
n (code length)
Fig. 4: Rate of binary codes from Construction V.1 (red line) and Construction V.3 (blue line).
C. Orthogonal arrays and the incidence code construction
In this subsection, we first recall a way to produce plenty of transversal designs from other combinatorial constructions called orthogonal arrays.
Definition V.5 (orthogonal array). Let λ, s ≥ 1 and ` ≥ t ≥ 1, and let A be an array with ` columns and λst
rows, whose entries are elements of a set S of size s. We say that A is an orthogonal array OAλ (t, `, s) if, in any
subarray A0 of A formed by t columns and all its rows, every row vector from S t appears exactly λ times in the
rows of A0 . We call λ the index of the orthogonal array, t its strength and ` its degree. If t (resp. λ) is omitted, it
is understood to be 2 (resp. 1). When both these parameters are omitted we write A = OA(`, s).
From now on, for convenience we restrict Definition V.5 to orthogonal arrays with no repeated column and no
repeated row. Next paragraph introduces a link between orthogonal arrays and transversal designs.
1) Construction of transversal designs from orthogonal arrays: We can build a transversal design TD(`, s) from
an orthogonal array OA(`, s) with the following construction, given as a remark in [8, ch.II.2].
Construction V.6 (Transversal designs from orthogonal arrays). Let A be an OA(`, s) of strength t = 2 and index
λ = 1 with symbol set S , |S| = s, and denote by Rows(A) the s2 rows of A. We define the point set X = S ×[1, `].
To each row c ∈ Rows(A) we associate a block
Bc := {(ci , i), i ∈ [1, `]} ,
so that the block set is defined as
B := {Bc , c ∈ Rows(A)} .
Finally, let G := {S × {i}, i ∈ [1, `]}. Then (X, B, G) is a transversal design TD(`, s).
13
Example V.7. A very simple example of this construction is given in Figure 5, where for clarity we use letters
for elements of the symbol set {a, b}, while the columns are indexed by integers. On the left-hand side, A is an
OA1 (2, 3, 2) with symbol set {a, b}. On the right-hand side, the associated transversal design TD(3, 2) is represented
as a hypergraph: the nodes are the points of the design, the “columns” of the graph form the groups, and a block
consists in all nodes linked with a path of a fixed colour. One can check that every pair of nodes either belongs to
the same group or is linked with one path.
a
b
A=
b
a
b
b
a
a
b
a
b
a
(a, 1)
(a, 2)
(a, 3)
(b, 1)
(b, 2)
(b, 3)
⇒
Fig. 5: A representation of the construction of a transversal design from an orthogonal array.
Remark V.8. Listed in rows, all the codewords of a (generic) code C0 give rise to an orthogonal array, whose
strength t is derived from the dual distance d0 of C0 by t = d0 − 1. Notice that for linear codes, the dual distance is
simply the minimum distance of the dual code, but it can also be defined for non-linear codes (see [15, Ch.5.§5.]).
More details about the link between orthogonal arrays and codes can also be found in [8]. For example, the
orthogonal array of Figure 5 comes from the binary parity-check code of length 3 (by replacing a by 0 and b by
1). One can check that its dual distance is 3 and its associated transversal design has strength 2.
Given a code C0 , we denote by AC0 the orthogonal array it defines (see Remark V.8) and by TC0 the transversal
design built from AC0 thanks to Construction V.6.
Example V.9. Let x = (x1 , . . . , x` ) be an `-tuple of pairwise distinct elements of Fq and denote by RS2 (x) the
Reed-Solomon code of length ` and dimension 2 over Fq with evaluation points x:
RS2 (x) := {(f (x1 ), . . . , f (x` )), f ∈ Fq [X], deg f < 2} .
Then, RS2 (x) has dual distance 3, so its codewords form an orthogonal array ARS2 (x) = OA(`, q) of strength 2. Now,
one can use Construction V.6 to obtain a transversal design TRS2 (x) = TD(`, q). The point set is X = Fq × [1, `],
and the blocks are “labeled Reed-Solomon codewords”, that is, sets of the form {(ci , i), i ∈ [1, `]} with c ∈ RS2 (x).
The ` groups correspond to the ` coordinates of the code: Gi = Fq × {i}, 1 ≤ i ≤ `.
We can finally sum up our construction by introducing the code Codeq (TC0 ) arising from C0 , that we call incidence
code. Our motivation is that studying the structure of C0 may lead to properties of Codeq (TC0 ) which is used in
PIR schemes.
Definition V.10 (incidence code). Let C0 be a (generic) code of length ` over an alphabet S of size s. The incidence
code of C0 over Fq , denoted ICq (C0 ), is the Fq -linear code of length n = s` built from the transversal design TC0 ,
that is:
ICq (C0 ) := Code(TC0 ) .
Notice that the field Fq does not need to be the alphabet S of the code C0 .
Incidence codes are introduced in order to design PIR protocols, as summarizes Figure 6. We can show that, if C0
has dual distance more than 3, then the induced PIR protocol is 1-private. A generalisation is formally proved in
Corollary VI.8.
Example V.11. Here we provide a full example of the construction of an incidence code. Let C0 be the full-length
Reed-Solomon code of dimension 2 over the field F4 = {0, 1, α, α2 = α + 1}. The orthogonal array associated to
C0 is composed by the following list of codewords:
14
equivalence (Rem. V.8)
/ Orthogonal array
Base code C0 o
Construction V.6 [8, ch.II.2]
t
Transversal design
incidence
matrix
/ Incidence code of C0
database
encoding
/ Distributed PIR scheme
Fig. 6: A distributed PIR scheme using the incidence code construction.
0, 0, 0, 0
1, 1, 1, 1
α, α, α, α
2 2 2 2
α , α , α , α
0, 1, α, α2
0, α, α2 , 1
0, α2 , 1, α
1, 0, α2 , α
A=
2 , α, 0
1,
α
1, α, 0, α2
α, α2 , 0, 1
α, 0, 1, α2
α, 1, α2 , 0
α2 , α, 1, 0
α2 , 1, 0, α
α2 , 0, α, 1
Using Construction V.6, we get a transversal design TC0 = (X, B, G) with 16 points (4 groups made of 4 points)
and 16 blocks. Let us recall how we map a row of A to a word in {0, 1}16 . For instance, consider the fifth row:
a := A5 = (0, 1, α, α2 ) .
We turn a into a block Ba := {(0, 1), (1, 2), (α, 3), (α, 4)} ∈ B , and we build the incidence vector 1Ba of the block
Ba over the point set X = {(β, i), i ∈ [1, 4], β ∈ F4 }. Of course, in order to see 1Ba as a word in {0, 1}16 , we
need to order elements in X , for instance:
(0, 1), (1, 1), (α, 1), (α2 , 1), (0, 2), (1, 2), (α, 2), (α2 , 2), (0, 3), . . . , (α2 , 4) .
Using this ordering, we get:
1Ba = 1, 0, 0, 0, 0, 1, 0, 0, 0, 0, 1, 0, 0, 0, 0, 1 ∈ {0, 1}16 .
15
By computing all the 1Ba for a ∈ Rows(A), we obtain the
1000 1000
0100 0100
0010 0010
0001 0001
1000 0100
1000 0010
1000 0001
0100 1000
M =
0100 0001
0100 0010
0010 0001
0010 1000
0010 0100
0001 0010
0001 0100
0001 1000
incidence matrix M of the transversal design TC0 :
1000 1000
0100 0100
0010 0010
0001 0001
0010 0001
0001 0100
0100 0010
0001 0010
,
0010 1000
1000 0001
1000 0100
0100 0001
0001 1000
0100 1000
1000 0010
0010 0100
Notice that this matrix can be quickly obtained by respectively replacing entries 0, 1, α and α2 in the array A
by the binary 4-tuples (1000), (0100), (0010) and (0001) in the matrix M (of course this map depends on the
ordering of X we have chosen, but another choice would lead to a column-permutation-equivalent matrix, hence
a permutation-equivalent code). Notice that in matrix M , coordinates lying in the same group of the transversal
design have been distinguished by dashed vertical lines.
Matrix M then defines, over any extension F2e of the prime field F2 , the dual code of the so-called incidence code
IC2e (C0 ). For all values of e, the incidence codes IC2e (C0 ) have the same generator matrix of 2-rank 7, being:
1001 0010 0001 1010
0101 0000 0110 1100
0011 0000 0011 1111
.
0000
1010
0011
0011
G=
0000 0110 0010 1110
0000 0001 0101 1110
0000 0000 1111 0101
2) A deeper analysis of incidence codes coming from linear MDS codes of dimension 2: Incidence codes lead to
an innumerably large family of PIR protocols — as many as there exists codes C0 — but most of them are not
practical for PIR protocols (essentially because the kernel of the incidence matrix is too small). To simplify their
study, one can first remark that intuitively, the more blocks a transversal design, the larger its incidence matrix, and
consequently, the lower the dimension of its associated code. But the number of blocks of TC0 is the cardinality of
C0 . Hence, informally the smaller the code C0 , the larger IC(C0 ).
We recall that a [n, k, d] linear code is said to be maximum distance separable (MDS) if it reaches the Singleton
bound n + 1 = k + d. Besides, the dual code of an MDS code is also MDS, hence its dual distance is k + 1. In
this paragraph we analyse the incidence codes constructed with MDS codes of dimension 2. Their interest lies in
being the smallest codes with dual distance 3, which is the minimal setting for defining 1-private PIR protocols.
Generalised Reed-Solomon codes are the best-known family of MDS codes.
Definition V.12 (generalised Reed-Solomon code). Let ` ≥ k ≥ 1. Let also x ∈ F`q be a tuple of pairwise distinct
`
so-called evaluation points, and y ∈ (F×
q ) be the column multipliers. We associate to x and y the generalised
Reed-Solomon (GRS) code:
GRSk (x, y) := {(y1 f (x1 ), . . . , y` f (x` )), f ∈ Fq [X], deg f < k} .
16
Generalised Reed-Solomon codes GRSk (x, y) are linear MDS codes of dimension k over Fq , and give ReedSolomon codes when y = (1, . . . , 1). Moreover, GRS codes are essentially the only MDS codes of dimension 2,
as shows the following lemma.
Lemma V.13. All [`, 2, ` − 1] MDS codes over Fq with 2 ≤ ` ≤ q are generalised Reed-Solomon codes.
Proof. First we know that GRS codes are MDS.
Let C be an [`, 2, ` − 1]q code with 2 ≤ ` ≤ q . Since C is MDS, it has dual distance d⊥ = 3, and we claim there
exists a codeword c ∈ C with Hamming weight `. Indeed, let G = (P1 , . . . , P` ) be a generator matrix of C , where
Pi ∈ F2q is written in column. Notice that each point Pi is non-zero (otherwise d⊥ = 1) and 0, Pi , Pj are not on
the same line for i 6= j (otherwise d⊥ = 2). Moreover codewords in C are simply evaluations of bilinear maps
µ : F2q → Fq over (P1 , . . . , P` ):
C = {(µ(P1 ), . . . , µ(P` )), µ ∈ L(F2q , Fq )} ,
and the Pi ’s are not all on the same line (otherwise, dim C ≤ 1).
Since ` ≤ q , there exists Q = (Q0 , Q1 ) ∈ F2q \ {0} such that Q does not lie in the (vector) line defined by any
of the Pi ’s. Let now µQ (X, Y ) = Q1 X − Q0 Y : it is a non-zero bilinear form which must vanish on a line of F2q ,
and since µQ (Q) = 0, it vanishes on the one spanned by Q. To sum up, for every i ∈ [1, `], we have µQ (Pi ) 6= 0.
Hence, c = (µQ (P1 ), . . . , µQ (P` )) belongs to C and has Hamming weight `.
Let now u ∈ C such that {c, u} spans C . We denote by c ∗ u the coordinate-wise product (c1 u1 , . . . , c` u` ) and by
1 the all-one vector of length `. Then c = 1 ∗ c and u = c ∗ (c−1 ∗ u), where c−1
is the coordinate-wise inverse of
1
0
0
0
c through ∗. Hence, the code C can be written c ∗ C where C has G = c−1 ∗u as generator matrix. It means that
C is the GRS code with evaluation points x = c−1 ∗ u, multipliers y = c and dimension 2.
We now study the consequences of Lemma V.13 in terms of transversal designs. We say a map φ : X → X 0 is
an isomorphism between transversal designs (X, B, G) and (X 0 , B 0 , G 0 ) if it is one-to-one and if it preserves the
incidence relations, or in other words, if φ is invertible on the points, blocks and groups:
φ(X) = X 0 ,
φ(B) = B 0 ,
φ(G) = G 0 .
`
Lemma V.14. Let C, C 0 be two codes such that C 0 = y ∗ C for some y ∈ (F×
q ) , where ∗ is the coordinatewise product of `-tuples. Recall TC , TC 0 are the transversal designs they respectively define. Then, TC and TC 0 are
isomorphic.
Proof. Write TC = (X, B, G) and TC 0 = (X 0 , B 0 , G 0 ). From the definition it is clear that X = X 0 = Fq × [1, `] and
G = G 0 = {Fq × {i}, 1 ≤ i ≤ `}. Now consider the blocks sets. We see that B = {{(ci , i), 1 ≤ i ≤ `}, c ∈ C} and
B 0 = {{(yi ci , i), 1 ≤ i ≤ `}, c ∈ C}. Let:
φy : Fq × [1, `] → Fq × [1, `]
(x, i)
7→ (yi x, i)
The vector y is ∗-invertible, hence φy is one-to-one on the point set X . It remains to notice that φy maps G to
itself since it only acts on the first coordinate, and that φy (B) is exactly B 0 by definition of C and C 0 .
Proposition V.15. Let 2 ≤ ` ≤ q and C0 be an [`, 2, ` − 1]q linear (MDS) code. Let also Fp be any finite field. The
incidence code ICp (C0 ) is permutation-equivalent to ICp (RS2 (x)), with x ∈ F`q , xi 6= xj .
Proof. Lemma V.13 shows that all [`, 2, ` − 1]q linear codes C0 can be written as y ∗ RS2 (x) for some x ∈ F`q .
Moreover, with the previous notation φy (TRS2 (x) ) = Ty∗RS2 (x) , so we have u ∈ ICp (y ∗ RS2 (x)) if and only if
u ∈ Codep (φy (TRS2 (x) )). Now, let:
φ̃y : FX
p
u = (ux )x∈X
→ FX
p
.
7
→
(uφy (x) )x∈X
Clearly φ̃y (ICp (RS2 (x))) = Codep (φy (TRS2 (x) )) and φ̃y is a permutation of coordinates. So ICp (C0 ) is permutationequivalent to ICp (RS2 (x)) which proves the result.
17
In our study of incidence codes of 2-dimensional MDS codes C0 , the previous proposition allows us to restrict our
work on Reed-Solomon codes C0 = RS2 (x) with x an `-tuple on pairwise distinct Fq -elements.
A first result proves that if x contains all the elements in Fq , then ICq (RS2 (x)) is the code which has been
previously studied in subsection V-A. More precisely,
Proposition V.16. The following two codes are equal up to permutation:
1) C1 = ICq (RS2 (Fq )), the incidence code over Fq of the full-length Reed-Solomon code of dimension 2 over
Fq ;
2) C2 , the code over Fq based on the transversal design TA (2, q).
Proof. It is sufficient to show that the transversal design defined by C0 = RS2 (Fq ) is isomorphic to TA (2, q). Let
us enumerate Fq = {x1 , . . . , xq }. We recall that TC = (X, B, G) where:
X = Fq × [1, q],
B = {{(ci , i), i ∈ [1, q]}, c ∈ C0 },
G = {{(α, i), α ∈ Fq }, i ∈ [1, q]} ,
and that TA (2, q) = (X 0 , B 0 , G 0 ) with:
X 0 = Fq × Fq ,
B 0 = {{(axi + b, xi ), i ∈ [1, q]}, (a, b) ∈ F2q },
G 0 = {{(α, xi ), α ∈ Fq }, i ∈ [1, q]} .
In the light of the above, one defines φ : X → X 0 , (α, i) 7→ (α, xi ), which is clearly one-to-one and satisfies
φ(G) = G 0 . Moreover, a codeword c ∈ C0 is the evaluation of a polynomial of degree ≤ 1 over Fq . Hence for some
(a, b) ∈ F2q , we have ci = axi + b, ∀i. This proves that φ extends to a one-to-one map B → B 0 , giving the desired
isomorphism.
It remains to study the case of tuples x of length ` < q . First, one may notice that ICq (RS2 (x)) is a shortening of
ICq (RS2 (Fq )). Indeed, we have the following property:
Lemma V.17. Let C0 be a linear code of length ` over Fq , and C0 be a puncturing of C0 on s positions. Then
for all prime powers q 0 , ICq0 (C0 ) is a shortening of ICq0 (C0 ) on the coordinates corresponding to s groups of the
transversal design TC0 .
Proof. Without loss of generality, we can assume that C0 is punctured on its s last coordinates in order to give C0 .
Let us analyse the link between TC0 = (X, B, G) and TC0 = (X, B, G). We have:
X = Fq × [1, ` − s]
⊂ X,
G = {Fq × {i}, i ∈ [1, ` − s]} ⊂ G,
B = {B ∩ X, B ∈ B}
Let C = ICq0 (C0 ) and C = ICq0 (C0 ). For clarity, we index words in C (resp. C ) by X (resp. X ). For c ∈ FX
q 0 , we
define ext(c) := c ∈ FX
=
=
0
.
By
definition
of
code’s
puncturing/shortening,
all
we
,
such
that
c
c
and
c
0
q
|X
|X\X
need to prove is:
C = {c ∈ FX
q 0 , ext(c) ∈ C}.
P
Remind that C is defined as the set of c ∈ FX
q 0 satisfying
b∈B cb = 0 for every B ∈ B . Hence we have:
X
c ∈ C ⇐⇒
cb = 0, ∀B ∈ B
b∈B
⇐⇒
X
cb = 0,
∀B ∈ B
b∈B∩X
⇐⇒
X
ext(c)b +
b∈B∩X
⇐⇒
X
ext(c)b = 0,
b∈B
⇐⇒ ext(c) ∈ C
X
b∈B∩(X\X)
∀B ∈ B
ext(c)b = 0,
∀B ∈ B
18
We conclude the proof by pointing out that X \ X is a union of s distinct groups from G .
Despite this result, incidence codes of Reed-Solomon codes RS2 (x) remain hard to classify for |x| = ` < q . Indeed,
for a given length ` < q , some IC(RS(x)) appear to be non-equivalent. Their dimension can even be different, as
shows an exhaustive search on IC16 (RS(x)) with pairwise distinct x ∈ F`q , q = 16 and ` = 5: we observe that
48 of these codes have dimension 24 while the 4320 others have dimension 22. Further interesting research would
then be to understand the values of x leading to the largest codes, for a fixed length |x| = `.
D. High-rate incidence codes from divisible codes
In this subsection, we prove that linear codes C0 satisfying a divisibility condition yield incidence codes whose rate
is roughly greater than 1/2. Let us first define divisible codes.
Definition V.18 (divisibility of a code). Let p ≥ 2. A linear code is p-divisible if p divides the Hamming weight
of all its codewords.
When studying the incidence matrix which defines an incidence code, we exhibit the following property.
Lemma V.19. Let C0 be a code of length ` over a set S , and let T be the transversal design associated to C0 . We
denote by M the incidence matrix of T , where rows of M are indexed by codewords from C0 . Then we have:
(M M T )c,c0 = ` − d(c, c0 )
∀c, c0 ∈ C ,
where d(·, ·) denotes the Hamming distance.
Proof. For clarity we adopt the notation M [c, (α, i)] for the entry of M which is indexed by the codeword c ∈ C0
(for the row), and (α, i) ∈ S × [1, `] (for the column). We also denote by 1U (c,i,α) ∈ {0, 1} the boolean value of
the property U , that is, 1U (c,i,α) = 1 if and only if U(c, i, α) is satisfied. Now, let c, c0 ∈ C0 .
X
X
(M M T )c,c0 =
M [c, (α, i)]M [c0 , (α, i)] =
1ci =α 1c0i =α
α∈S, i∈[1,`]
=
` X
X
α∈S, i∈[1,`]
1ci =c0i =α =
i=1 α∈S
`
X
1ci =c0i = ` − d(c, c0 ) .
i=1
Hence, if some prime p divides ` as well as the weight of all the codewords in C0 , then the product M M T vanishes
over any extension of Fp , and M is a parity-check matrix of a code containing its dual. A more general setting is
analysed in the following proposition.
Proposition V.20. Let C0 be a linear code of length ` over S , |S| = s. Let also C = ICq (C0 ) with char(Fq ) = p.
Denote the length of C by n = `s. If C0 is p-divisible, then
C ⊥ ∩ Cpar ⊆ C ,
where Cpar denotes the parity-check code of length n over Fq . In particular, we get dim C ≥
n−1
2 .
Moreover, if p|`, then C ⊥ ⊆ C and dim C ≥ n2 .
Proof. Let M be the incidence matrix of the transversal design TC0 . Also denote by J and J 0 the all-ones matrices
of respective size |C0 | × n and |C0 | × |C0 |. If we assume that C0 is p-divisible, then Lemma V.19 translates into
M M T = `J 0
mod p
(3)
while an easy computation shows that
M J T = `J 0 .
Hence, over Fq we obtain
M (M − J)T = 0
(4)
19
which brings us to consider the code A of
P length n generated over Fq by the matrix M − J . Equation (4)
indicates that A ⊆ C . Let Cpar := {c ∈ Fnq , i ci = 0} be the parity-check code of length n over Fq . Notice that
c ∈ Cpar ⇐⇒ cJ T = 0 and uJ = 0 ⇐⇒ uJ 0 = 0. If p 6 | `, this leads to:
C ⊥ ∩ Cpar = {c = uM ∈ Fnq , cJ T = 0}
= {c = uM ∈ Fnq , `uJ 0 = 0} = {c = uM ∈ Fnq , uJ = 0}
= {u(M − J) ∈ Fnq , uJ = 0} ⊆ A ⊆ C .
On the other hand, if p|`, then equation (3) turns into M M T = 0, meaning that C ⊥ ⊆ C .
Finally, the first bound on the dimension comes from
dim C ≥ dim(C ⊥ ∩ Cpar ) ≥ dim C ⊥ − 1 = n − dim C − 1 ,
while the second one is straightforward.
In terms of PIR protocols, previous result translates into the following corollary.
Corollary V.21. Let p be a prime, and assume there exists a p-divisible linear code of length `0 over Fq . Then,
there exists k ≥ (`0 q − 1)/2 such that we can build a distributed PIR protocol for a k -entries database over Fq ,
and whose parameters are `(k) = `0 and n(k) = `0 q ≤ 2k + 1.
Divisible codes over small fields have been well-studied, and contain for instance the extended Golay codes [15,
ch.II.6], or the famous MDS codes of dimension 3 and length q + 2 over Fq [15, ch.XI.6].
Example V.22. The extended binary Golay code is a self-dual [24, 12, 8]2 linear code. It produces a transversal
design with 24 groups, each storing 2 points. Its associated incidence code Code2 (Golay) has length n = 24×2 = 48
and dimension ≥ 24, and by computation we can show that this bound is tight.
Remark V.23. In our application for PIR protocols, we would like to find divisible codes C0 defined over large
alphabets (compared to the code length), but these two constraints seem to be inconsistent. For instance, the binary
Golay code presented in Example V.22 leads to a PIR protocol with a too expensive communication cost (24 bits of
communication for an original file of size... 24 bits: that is exactly the communication cost of the trivial PIR protocol
where the whole database is downloaded). Nevertheless, Example V.22 represents the worst possible case for our
construction, in a sense that the rate of IC2 (Golay2 ) is exactly 1/2 (it attains the lower bound), and that each server
stores 2 bits (which is the smallest possible). Codes with better rate and/or with larger server storage capability
would then give PIR protocols with relevant communication complexity. For instance, the extended ternary Golay
code gives better parameters — see Example VI.9.
Divisible codes over large fields seems not to have been thoroughly studied (to the best of our knowledge), since
coding theorists use to consider codes over small alphabets as more practical. We hope that our construction of
PIR protocols based on divisible codes may encourage research in this direction.
VI. PIR PROTOCOLS WITH BETTER PRIVACY
When servers are colluding, the PIR protocol based on a simple transversal design does not ensure a sufficient
privacy, because the knowledge of two points on a block gives some information on it. To solve this issue, we
propose to use orthogonal arrays with higher strength t.
A. Generic construction and analysis
In the previous section, classical (t = 2) orthogonal arrays were used to build transversal designs. Considering
higher values of t, we naturally generalise the latter as follows:
Definition VI.1 (t-transversal designs). Let ` ≥ t ≥ 1. A t-transversal design is a block design D = (X, B)
equipped with a group set G = {G1 , . . . , G` } partitioning X such that:
20
– |X| = s`;
– any group has size s and any block has size `;
– for any T ⊆ [1, `] with |T | = t and for any (x1 , . . . , xt ) ∈ GT1 × . . . × GTt , there exist exactly λ blocks B ∈ B
such that {x1 , . . . , xt } ⊂ B .
A t-transversal design with parameters s, `, t, λ is denoted t-TDλ (`, s), or t-TD(`, s) when λ = 1.
Given a t-transversal design T , we can build a (t − 1)-private PIR protocol with the exactly the same steps as in
section IV. First, we define the code C = Codeq (T ) associated to the design according to Definition III.7, and then
we follow the algorithm given in Figure 2. Since a t-transversal design is also a 2-transversal design for t ≥ 2, the
analysis is identical for every PIR feature, except for the security where it remains very similar.
Security ((t − 1)-privacy). Let T be a collusion of servers of size |T | ≤ t − 1. For varying i ∈ I , the distributions
Q(i)|T are the same because there are exactly λst−1−|T | ≥ λ 6= 0 blocks which contain both i and the queries
known by the servers in T .
To sum up, the following theorem holds:
Theorem VI.2. Let D be a database with k entries over Fq , and T = t-TD(`, s) be a t-transversal design, whose
incidence matrix has rank `s − k over Fq . Then, there exists an `-server (t − 1)-private PIR protocol with:
–
–
–
–
only 1 symbol to read for each server,
` − 1 field operations for the user,
` log(sq) bits of communication,
a (total) storage overhead of (`s − k) log q bits on the servers.
B. Instances and results
1) t-transversal designs from curves of degree ≤ t − 1: Looking for instances of t-transversal designs, it is natural
to try to generalise the transversal designs of Construction V.1. An idea is to turn affine lines into higher degree
curves.
Construction VI.3. Let X be the set of points in the affine plane F2q , and G = {G1 , . . . , Gq } be a partition of X
in q parallel lines. W.l.o.g. we choose the following partition: Gi = {(αi , y), y ∈ Fq } for each αi ∈ Fq . Blocks are
now defined as the sets of the form
BF = {(x, F (x)), x ∈ Fq }, where F ∈ Fq [x], deg F ≤ t − 1.
Lemma VI.4. The design (X, B, G) given in Construction VI.3 forms a t-transversal design t-TD1 (q, q).
Proof. The group set indeed partitions X into q groups, each of size q . It remains to check the incidence property.
Let {GT1 , . . . , GTt } be a set of t distinct groups, and let ((xT1 , yT1 ), . . . , (xTt , yTt )) ∈ GT1 × · · · × GTt . From
Lagrange interpolation theorem, we know there exists a unique polynomial F ∈ Fq [X] of degree ≤ t − 1 such that:
F (xTj ) = yTj
∀1 ≤ j ≤ t .
Said differently, there is a unique block which contains the t points {(xTj , yTj )}1≤j≤t .
We do not yet analyse the rank properties of these designs, since Construction VI.3 corresponds to a particular
case of the generic construction given below.
2) t-transversal designs from orthogonal arrays of strength t: In this paragraph we give a generic construction
of t-transversal designs, which is a simple generalisation of the way we build transversal designs with orthogonal
arrays (Subsection V-C).
Construction VI.5. Let A be an orthogonal array OAλ (t, `, s) on a symbol set S . Recall that the array A is
composed of rows ai = (ai,j )1≤j≤` for 1 ≤ i ≤ λst . We define the following design:
– its point set is X = S × [1, `];
– its group set is G = {S × {i}, 1 ≤ i ≤ `};
21
– its blocks are Bi = {(ai,j , j), 1 ≤ j ≤ `} for all ai ∈ Rows(A).
Proposition VI.6. If A is an OAλ (t, `, s), then the design defined with A by Construction VI.5 is a t-TDλ (`, s).
Proof. It is clear that G is a partition of X and that blocks and groups have the claimed size. Now focus on the
incidence property. Let T ⊂ [1, `] with |T | = t, and let (x1 , . . . , xt ) ∈ GT1 × . . . × GTt . We need to prove that
there are exactly λ blocks B ∈ B such that {x1 , . . . , xt } ⊂ B .
Consider the map from blocks in B to rows of A given by:
ψ:
B
→
Rows(A)
Bi = {(ai,j , j), 1 ≤ j ≤ `} 7→ (ai,1 , . . . , ai,` )
Since we assumed that orthogonal arrays have no repeated row, the map ψ is one-to-one. Denote by x0 =
(x01 , , . . . , x0t ) ∈ S t the vector formed by the first coordinates of (x1 , , . . . , xt ) ∈ X t . From the definition of an
orthogonal array of strength t and index λ, we know that x0 appears exactly λ times in the submatrix of A defined
by the columns indexed by T . Hence this defines λ preimages in B , which proves the result.
Remark VI.7. As we noticed before, Construction VI.3 is a particular case of Construction VI.5. Indeed, a block
BF = {(x, F (x)), x ∈ Fq }, with deg F ≤ t − 1 is in one-to-one correspondence with a codeword cF of a ReedSolomon code of dimension t.
Corollary VI.8. Let C0 be a code of length ` and dual distance t + 2 ≤ ` over a set S of size s. Then, ICq (C0 )
defines a t-private PIR protocol.
Proof. Let A be the orthogonal array defined by C0 . We know that A has strength t + 1 (see e.g. [15]), hence from
Proposition VI.6, the associated transversal design is a (t + 1)-TD(`, s). Theorem VI.2 then ensures that the PIR
protocol induced by this transversal design is t-private.
As in Section V, when the code C0 is divisible, we can lower bound the rate of its incidence code. We provide two
examples in finite (and small) length.
Example VI.9. A first example would be to consider extended Golay codes. Indeed, they are known to be divisible
by their characteristic [15, ch.II.6], they have large dual distance, and Proposition V.20 then ensures their incidence
codes have non-trivial rate. In Remark V.23, we noticed that the binary Golay code does not lead to a practical PIR
protocol due to a large communication complexity. Thus, let us instead consider the [12, 6, 6]3 extended ternary
Golay code, that we denote Golay3 . It is self-dual, hence d⊥ (Golay3 ) = 6. Then, C = IC3e (Golay3 ), e ≥ 1, has
length 36 and Proposition V.20 shows that dim C ≥ 18 (the bound can be proved to be tight by computation).
Hence, the associated PIR protocol works on a raw file of 18 F3e -symbols encoded into 36, uses 12 servers (each
storing 3 F3e -symbols) and resists any collusion of one third (i.e. 4) of them.
Example VI.10. A second example arises from the exceptional [q + 2, 3, q]q MDS codes in characteristic 2 [15,
ch.XI.6]. For instance, for q = 4, we obtain a 2-private PIR protocol with 6 servers, each storing 4 symbols of F2e
for some e ≥ 1. Once again, the dimension of the incidence code attains the lower bound, here k = 12.
Example VI.11. Examples of incidence codes which do not attain the lower bound of Proposition V.20 come from
binary Reed-Muller codes of order 1, denoted RM2 (m, 1). These codes are 2-divisible since they are known to be
equivalent to extended Hamming codes. They also have length n = 2m and dual distance d⊥ = n/2.
For instance, RM2 (3, 1) provides an incidence code of dimension k = 11 > 8, that is, a 2-private 8-server PIR
protocol on a database with 11 F2e -symbols, where each server stores 2 symbols. For m = 4, RM2 (4, 1) gives a
6-private 16-server PIR protocol on a database with 20 F2e -symbols, each server storing 2 symbols. We conjecture
that IC2 (RM2 (m, 1)) leads to a (2m−1 − 2)-private 2m -server PIR protocol on a database with 2m + m symbols,
each server storing 2 symbols.
As pointed out in Subsection V-C, high-rate incidence codes C = IC(C0 ) have the best chance to occur when
the dimension of C0 is small, since the cardinality of C0 is the number of rows in a (non-full-rank) parity-check
matrix which defines C . Besides, in order to define t-private PIR protocols, we need an orthogonal array of strength
22
t + 2, i.e. a code C0 with dual distance t + 2. Conciliating both constraints, we are tempted to pick MDS codes of
dimension t + 1.
A well-known family of MDS codes is the family of Reed-Solomon codes. For C0 = RSt+1 (Fq ) and varying
values of q and t, we were able to compute the√rate of IC(C0 ), and these codes lead to t-private PIR protocols
with communication complexity approximately n, where n is the length of the encoded database. These rates
are presented in Figure 7 and as expected, the rate of our families of incidence codes decreases with t, the privacy
parameter. Figure 7 also shows that Reed-Solomon-based instances cannot expect to reach at the same time constant
information rate and resistance to a constant fraction of colluding servers.
rate R = k/n
1
0.9
t=1
0.8
0.7
t=2
0.6
0.5
t=4
0.4
0.3
t=log(q)
0.2
t=q1/2
t=q/2
0.1
t=q/3
t=q-1
0
6
8
10
12
14
16
18
20
log n = 2 log q
Fig. 7: Rate of incidence codes of C0 that are used for building t-private PIR protocols. Codes C0 are fulllength Reed-Solomon codes of dimension t + 1 (dual distance t + 2) over Fq . Associated PIR protocols then
need q servers, each storing q symbols.
VII. C OMPARISON WITH OTHER WORKS
Our construction fits into the model of distributed (or coded) PIR protocols, which is currently instantiated in a few
schemes, notably the construction of Augot et. al. [2] and all the works involving the use of PIR codes initiated
by Fazeli et. al [11]. We recall that we aimed at building PIR protocols with very low burden for the servers, in
terms of storage and computation. While PIR codes are a very efficient way to reduce the storage overhead, they
do not cut down the computation complexity of the original replication-based PIR protocol used for the emulation.
Hence, for the sake of consistency, we will only compare the parameters of our PIR schemes with those of the
multiplicity code construction presented in [2].
Sketch of the construction [2]. Multiplicity codes C have the property that a codeword c ∈ C can be seen as the
vector of evaluations of a multivariate polynomial fc ∈ Fq [X1 , . . . , Xm ] and its derivatives over the space Fm
q , where
Fq denotes the finite field with q elements. Every affine line of the space Fm
then
induces
linear
relations
between
q
fc and its derivatives, which translates into low-weight parity-check equations for the codewords. This allows to
define to a local decoder for C :when trying to retrieve a symbol Di indexed by i ∈ Fm
q , one can pick random
affine lines going through i and recover Di by computing short linear combinations of the symbols associated to
the evaluations of fc and their derivatives along these lines. We refer to [14] for more details on these codes.
Augot et. al. [2] realized that partitioning Fm
q into q parallel hyperplanes gives rise to storage improvements.
By splitting the encoded database according to these hyperplanes and giving one part to each of the q servers,
23
they obtained a huge cutdown on both the total storage and the number of servers, while keeping an acceptable
communication complexity. Their construction requires a minor modification of the LDC-based PIR protocol of
Figure 1; indeed, in the query generation process, the only server which holds the desired symbol must receive
a random query. Nevertheless, the PIR scheme they built was at that time the only one to let the servers store
less than twice the size of the database. Moreover, the precomputation of the encoding of the database ensures
an optimal computational complexity for the servers. As noticed previously, we emphasize the significance of this
feature when the database is very frequently queried.
Parameters of the distributed PIR protocol [2] based on multiplicity codes. This PIR scheme depends on four
main parameters: the field size q , the dimension m of the underlying affine space, the multiplicity order s and
the maximal degree d of evaluated polynomials.
For an error-free and collusion-free setting, d = s(q − 1) − 1 is
m+s−1
the optimal choice.
Let
σ
=
.
The
associated
PIR protocol uses q servers to store an original database
m
d+m
containing m Fq -symbols, but encoded into codewords of length q m , where each symbol has size σ log q bits.
Hence the redundancy (in bits) of the scheme is:
s(q − 1) + m − 1
ρ = σq m −
log q .
m
Concerning the communication complexity, let us only focus on the download cost (which is often the bottleneck in
practice). Each server answers a symbol of size σ 2 log q bits. Thus it leads to a download communication complexity
of
γ = σ 2 q log q bits.
Note about the comparison strategy. We consider a database D of size 100 MB. Since the protocols may be
initially constructed for databases of smaller size k , we split D into k chunks of size |D|/k . Hence, when running
the PIR protocol, the user is allowed to retrieve a whole chunk, and the chunk size will be precised in our tables.
For instance, in the first row of Table II, one shall understand that the user is able to retrieve 31.1kB of the database
privately, with 1.99MB of communication, while each server only produces 1 operation over the chunks (of size
31.1kB) it holds.
Instance
TA (m = 2, q = 64)
TA (m = 3, q = 64)
Mult(q = 64, m = 3, s = 6)
Mult(q = 64, m = 4, s = 2)
download
communication
1.99 MB
56 kB
2.32 MB
15 kB
complexity
(#op./server)
1
1
56
5
storage overhead
22.7 MB
126 MB
64.8 MB
694 MB
chunk size
31.1 kB
882 B
16 B
13 B
TABLE II: Comparison of some distributed PIR protocols with 64 servers on a 100MB initial database.
Parameters of multiplicity codes have been chosen in order to obtain simultaneously low communication
complexity and storage overhead.
Instance
TA (m = 2, q = 8)
TA (m = 3, q = 8)
Mult(q = 8, m = 4, s = 2)
Mult(q = 8, m = 6, s = 3)
download
communication
22.7 MB
6.03 MB
8.8 MB
2.86 MB
complexity
(#op./server)
1
1
5
28
storage overhead
76 MB
281 MB
797 MB
3.24 GB
chunk size
2.83 MB
754 kB
117 kB
1.2 kB
TABLE III: Comparison of some distributed PIR protocols with 8 servers on a 100MB initial database. We
notice that for q = 8 servers, there only exist a few non-Reed-Muller (s ≥ 2) multiplicity codes whose
associated PIR protocols have communication complexity strictly less than the size of the original database.
Tables II and III first reveal that our PIR schemes are more storage efficient than the PIR schemes relying on
multiplicity codes. Moreover, our constructions provide a better communication rate (defined as the ratio between
communication cost and chunk size), though the multiplicity code PIR protocols allow to retrieve smaller chunks
(hence is more flexible).
24
Remark VII.1. Recent constructions of PIR protocols (for instance results of Sun and Jafar such that [17]) lead
to better parameters in terms of communication complexity. However, we once more emphasize that we aimed at
minimizing the computation carried out by the servers, which is a feature that is mostly not considered in those
works.
VIII. C ONCLUSION
In this paper, we have presented a generic construction of codes which turns out to be appropriate for distributed
PIR protocols. Our construction makes use of transversal designs, whose incidence properties ensure a natural
distribution of the database on the servers, as well as the privacy of the queries. We point out that our scheme is
computationnally optimal for the servers, in a sense that they only have to read one symbol of the file they hold.
Moreover, it also admits an efficient reconstructing step since the user only needs to compute a linear combination
of the symbols it receives. Our PIR protocols also require low storage for the servers and acceptable communication
complexity.
We instantiated our construction with classical transversal designs coming from affine and projective geometries,
and with transversal designs emerging from orthogonal arrays of strength 2. The last construction that we call
incidence code can even be generalized, since stronger orthogonal arrays lead to PIR protocols with better privacy.
The genericity of our construction allows the user to choose appropriate settings according to the context (low
storage capability, few colluding servers, etc.). This genericity also leads to the question of finding transversal
designs with the most practical PIR parameters for a given context. Indeed, while affine and projective geometries
give excellent PIR parameters for the servers (low computation, low storage), there seems to remain room for
improving the communication complexity and the number of servers needed in the PIR scheme.
ACKNOWLEDGEMENTS
This work is partially funded by French ANR-15-CE39-0013-01 “Manta”. The author would like to thank Françoise
Levy-dit-Vehel and Daniel Augot for their valuable comments, and more specifically the first collaborator for her
helpful guidance all along the writing of the paper.
R EFERENCES
[1] Edward F. Assmus and Jennifer D. Key. Designs and Their Codes. Cambridge University Press, 1992.
[2] Daniel Augot, Françoise Levy-dit-Vehel, and Abdullatif Shikfa. A storage-efficient and robust private information retrieval scheme allowing few servers. In Dimitris Gritzalis, Aggelos Kiayias, and Ioannis G. Askoxylakis,
editors, Cryptology and Network Security - 13th International Conference, CANS 2014, Heraklion, Crete,
Greece, October 22-24, 2014. Proceedings, volume 8813 of Lecture Notes in Computer Science, pages 222–
239. Springer, 2014.
[3] Amos Beimel, Yuval Ishai, Eyal Kushilevitz, and Jean-François Raymond. Breaking the O(n1/(2k−1) ) barrier
for information-theoretic private information retrieval. In 43rd Symposium on Foundations of Computer Science
(FOCS 2002), 16-19 November 2002, Vancouver, BC, Canada, Proceedings, pages 261–270. IEEE Computer
Society, 2002.
[4] Amos Beimel, Yuval Ishai, and Tal Malkin. Reducing the servers’ computation in private information retrieval:
PIR with preprocessing. J. Cryptology, 17(2):125–151, 2004.
[5] Benny Chor and Niv Gilboa. Computationally private information retrieval. In Frank Thomson Leighton and
Peter W. Shor, editors, Proceedings of the Twenty-Ninth Annual ACM Symposium on the Theory of Computing,
El Paso, Texas, USA, May 4-6, 1997, pages 304–313. ACM, 1997.
[6] Benny Chor, Oded Goldreich, Eyal Kushilevitz, and Madhu Sudan. Private information retrieval. In 36th
Annual Symposium on Foundations of Computer Science, Milwaukee, Wisconsin, 23-25 October 1995, pages
41–50. IEEE Computer Society, 1995.
[7] Benny Chor, Eyal Kushilevitz, Oded Goldreich, and Madhu Sudan. Private information retrieval. J. ACM,
45(6):965–981, 1998.
25
[8] Charles J. Colbourn and Jeffrey H. Dinitz. Handbook of Combinatorial Designs, Second Edition. Chapman
& Hall/CRC, 2006.
[9] Zeev Dvir and Sivakanth Gopi. 2-server PIR with subpolynomial communication. J. ACM, 63(4):39:1–39:15,
2016.
[10] Klim Efremenko. 3-query locally decodable codes of subexponential length. SIAM J. Comput., 41(6):1694–
1703, 2012.
[11] Arman Fazeli, Alexander Vardy, and Eitan Yaakobi. Codes for distributed PIR with low storage overhead.
In IEEE International Symposium on Information Theory, ISIT 2015, Hong Kong, China, June 14-19, 2015,
pages 2852–2856. IEEE, 2015.
[12] Noboru Hamada. The rank of the incidence matrix of points and d-flats in finite geometries. Journal of
Science of the Hiroshima University, Series A-I (Mathematics), 32(2):381–396, 1968.
[13] Jonathan Katz and Luca Trevisan. On the efficiency of local decoding procedures for error-correcting codes.
In F. Frances Yao and Eugene M. Luks, editors, Proceedings of the Thirty-Second Annual ACM Symposium
on Theory of Computing, May 21-23, 2000, Portland, OR, USA, pages 80–86. ACM, 2000.
[14] Swastik Kopparty, Shubhangi Saraf, and Sergey Yekhanin. High-rate codes with sublinear-time decoding. J.
ACM, 61(5):28:1–28:20, 2014.
[15] F.J. MacWilliams and N.J.A. Sloane. The Theory of Error-Correcting Codes. North Holland, 1977.
[16] Douglas R. Stinson. Combinatorial Designs - Constructions and Analysis. Springer, 2004.
[17] Hua Sun and Syed Ali Jafar. The capacity of private information retrieval. IEEE Trans. Information Theory,
63(7):4075–4088, 2017.
[18] Razan Tajeddine and Salim El Rouayheb. Private information retrieval from MDS coded data in distributed
storage systems. In IEEE International Symposium on Information Theory, ISIT 2016, Barcelona, Spain, July
10-15, 2016, pages 1411–1415. IEEE, 2016.
[19] Sergey Yekhanin. Towards 3-query locally decodable codes of subexponential length. J. ACM, 55(1):1:1–1:16,
2008.
[20] Sergey Yekhanin. Locally decodable codes. Foundations and Trends in Theoretical Computer Science,
6(3):139–255, 2012.
A PPENDIX
Hamada [12] gives a generic formula to compute the p-rank of a projective geometry design PGt (m, q), for q = pe :
j+1 ,sj )
e−1
X
Y L(sX
m + 1 m + sj+1 p − sj − ip
rankp PGt (m, q) =
(−1)i
i
m
(s0 ,..., se )∈S j=0
where L(sj+1 , sj ) =
b sj+1pp−sj c
i=0
and (s0 , . . . , se ) ∈ S ⊂ Ze+1 are such that:
s0 = se
t + 1 ≤ sj ≤ m + 1
0 ≤ sj+1 p − sj ≤ (m + 1)(p − 1) .
The p-rank of the associated affine geometry design AGt (m, q) can be derived from the projective one by:
rankp AGt (m, q) = rankp PGt (m, q) − rankp PGt (m − 1, q) .
Despite its heavy expression, Hamada’s formula can be simplified by picking very specific values of m, p or e.
For instance we have:
p+1 e
e
m = 2 : ∀p, e, rankp AG1 (2, p ) =
,
2
m+p−2
m
e = 1 : ∀p, m, rankp AG1 (m, p) = p −
,
m
p+1 2
p p+1
2
3
m = 3 and e = 2 : ∀p, rankp AG1 (3, p ) = p −
+2
,
3
2
3
26
where the last equality has been found by interpolation (since rankp AG1 (m, pe ) is a polynomial of degree at most
me in p).
| 7cs.IT
|
FROBENIUS AND VALUATION RINGS
arXiv:1507.06009v3 [math.AC] 29 Dec 2016
RANKEYA DATTA AND KAREN E. SMITH
Abstract. The behavior of the Frobenius map is investigated for valuation rings of prime
characteristic. We show that valuation rings are always F-pure. We introduce a generalization of the notion of strong F-regularity, which we call F-pure regularity, and show that a
valuation ring is F-pure regular if and only if it is Noetherian. For valuations on function
fields, we show that the Frobenius map is finite if and only if the valuation is Abhyankar;
in this case the valuation ring is Frobenius split. For Noetherian valuation rings in function
fields, we show that the valuation ring is Frobenius split if and only if Frobenius is finite, or
equivalently, if and only if the valuation ring is excellent.
Contents
Page
1. Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
2. Preliminaries . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
2.1. Valuation Rings . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
2.2. Extension of Valuations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
2.3. Abhyankar Valuations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
2.4. Frobenius . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
2.5. F-purity and Frobenius splitting . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
2.6. F-finiteness and Excellence . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
3. Flatness and Purity of Frobenius in Valuation Rings. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
4. F-finite Valuation Rings . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
4.1. Finiteness and Freeness of Frobenius . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
4.2. Frobenius Splitting in the Noetherian case.. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
4.3. A numerical Criterion for F-finiteness . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
4.4. Examples of Frobenius Split Valuations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
4.5. Example of a non-Frobenius Split Valuation. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
4.6. Finite Extensions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
5. F-finiteness in Function Fields . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
6. F-regularity. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
6.1. Basic Properties of F-pure regularity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
6.2. Relationship of F-pure regularity to other singularities . . . . . . . . . . . . . . . . . . . . . . 20
The first author was partially supported by the Juha Heinonen Memorial Graduate Fellowship at the
University of Michigan. The second author was partially supported by NSF grants DMS-1001764 and DMS1501625.
1
6.3. Connections with Tight Closure. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
6.4. Elements along which F-purity fails . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
6.5. F-pure regularity and Valuation Rings . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
6.6. Split F-regularity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
7. Concluding Remarks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
21
22
23
25
26
27
1. Introduction
Classes of singularities defined using Frobenius—F-purity, Frobenius splitting, and the
various variants of F-regularity—have played a central role in commutative algebra and
algebraic geometry over the past forty years. The goal of this paper is a systematic study
of these F-singularities in the novel, but increasingly important non-Noetherian setting of
valuation rings.
Let R be a commutative ring of prime characteristic p. The Frobenius map is the ring
F
homomorphism R → R sending each element to its p-th power. While simple enough, the
Frobenius map reveals deep structural properties of a Noetherian ring of prime characteristic,
and is a powerful tool for proving theorems for rings containing an arbitrary field (or varieties,
say, over C) by standard reduction to characteristic p techniques. Theories such as Frobenius
splitting [33] and tight closure [22] are well-developed in the Noetherian setting, often under
the additional assumption that the Frobenius map is finite. Since classically most motivating
problems were inspired by algebraic geometry and representation theory, these assumptions
seemed natural and not very restrictive. Now, however, good reasons are emerging to study
F-singularities in certain non-Noetherian settings as well.
One such setting is cluster algebras [14]. An upper cluster algebra over Fp need not be
Noetherian, but recently it was shown that it is always Frobenius split, and indeed, admits a
“cluster canonical” Frobenius splitting [4]. Likewise valuation rings are enjoying a resurgence
of popularity despite rarely being Noetherian, with renewed interest in non-Archimedean
geometry [9], the development of tropical geometry [17], and the valuative tree [13], to name
just a few examples, as well as fresh uses in higher dimensional birational geometry (e.g.
[10, 11, 6]).
For a Noetherian ring R, the Frobenius map is flat if and only if R is regular, by a famous
theorem of Kunz [28]. As we observe in Theorem 3.1, the Frobenius map is always flat for
a valuation ring. So in some sense, a valuation ring of characteristic p might be interpreted
as a “non-Noetherian regular ring.”
On the other hand, some valuation rings are decidedly more like the local rings of smooth
points on varieties than others. For example, for a variety X (over, say, an algebraically
closed field of characteristic p), the Frobenius map is always finite. For valuation rings of the
function field of X, however, we show that the Frobenius is finite if and only the valuation
is Abhyankar; see Theorem 5.1. In particular, for discrete valuations, finiteness of Frobenius
is equivalent to the valuation being divisorial—that is, given by the order of vanishing along
2
a prime divisor on some birational model. Abhyankar valuations might be considered the
geometrically most interesting ones (C.f. [12]), so it is fitting that their valuation rings
behave the most like the rings of smooth points on a variety. Indeed, recently, the local
uniformization problem for Abhyankar valuations was settled in positive characteristic [27].
One can weaken the demand that Frobenius is flat and instead require only that the
Frobenius map is pure (see §2.5). Hochster and Roberts observed that this condition, which
they dubbed F-purity, is often sufficient for controlling singularities of a Noetherian local
ring, an observation at the heart of their famous theorem on the Cohen-Macaulayness of
invariant rings [26, 25]. We show in Corollary 3.3 that any valuation ring of characteristic p
is F-pure. Purity of a map is equivalent to its splitting under suitable finiteness hypotheses,
but at least for valuation rings (which rarely satisfy said hypotheses), the purity of Frobenius
seems to be better behaved and more straightforward than its splitting. Example 4.5.1 shows
that not all valuation rings are Frobenius split, even in the Noetherian case.
Frobenius splitting has well known deep local and global consequences for algebraic varieties. In the local case, Frobenius splitting has been said to be a “characteristic p analog”
of log canonical singularities for complex varieties, whereas related properties correspond to
other singularities in the minimal model program [19, 37, 40, 46]. For projective varieties,
Frobenius splitting is related to positivity of the anticanonical bundle; see [8, 33, 41, 42]. Although valuation rings are always F-pure, the question of their Frobenius splitting is subtle.
Abhyankar valuations in function fields are Frobenius split (Theorem 5.1), but a discrete
valuation ring is Frobenius split if and only if it is excellent in the sense of Grothendieck
(Theorem 4.2.2). Along the way, we prove a simple characterization of the finiteness of Frobenius for a Noetherian domain in terms of excellence, which gives a large class of Noetherian
domains in which Frobenius splitting implies excellence; see §2.6 for details.
Closely related to F-purity and Frobenius splitting are the various variants of F-regularity.
Strong F-regularity was introduced by Hochster and Huneke [21] as a proxy for weak Fregularity—the property that all ideals are tightly closed—because it is easily shown to pass
to localizations. Whether or not a weakly F-regular ring remains so after localization is a long
standing open question in tight closure theory, as is the equivalence of weak F-regularity and
strong F-regularity. Strong F-regularity has found many applications beyond tight closure,
and is closely related to Ramanathan’s notion of “Frobenius split along a divisor" [35, 41].
A smattering of applications might include [2, 4, 5, 8, 15, 18, 34, 39, 36, 42, 43, 44, 41].
Traditionally, strong F-regularity has been defined only for Noetherian rings in which
Frobenius is finite. To clarify the situation for valuation rings, we introduce a new definition
which we call F-pure regularity (see Definition 6.1.1) requiring purity rather than splitting
of certain maps. We show that F-pure regularity is better suited for arbitrary rings, but
equivalent to strong F-regularity under the standard finiteness hypotheses; it also agrees
with another generalization of strong F-regularity proposed by Hochster (using tight closure)
in the local Noetherian case [20]. Likewise, we show that F-pure regularity is a natural and
straightforward generalization of strong F-regularity, satisfying many expected properties—
for example, regular rings are F-pure regular. Returning to valuation rings, in Theorem 6.5.1
we characterize F-pure regular valuation rings as precisely those that are Noetherian.
Finally, in §6.6, we compare our generalization of strong F-regularity with the obvious
competing generalization, in which the standard definition in terms of splitting certain maps
3
is naively extended without assuming any finiteness conditions. To avoid confusion,1 we call
this split F-regularity. We characterize split F-regular valuation rings (at least in a certain
large class of fields) as precisely those that are Frobenius split, or equivalently excellent; see
Corollary 6.6.3. But we also point out that there are regular local rings that fail to be split
F-regular, so perhaps split F-regularity is not a reasonable notion of “singularity.”
The authors gratefully acknowledge the conversation with Karl Schwede and Zsolt Patakfalvi that inspired our investigation of Frobenius in valuation rings, and especially Karl,
with whom we later had many fruitful discussions. The authors also acknowledge helpful
discussions with Mel Hochster, Daniel Hernandez, Mattias Jonsson, Eric Canton, Linquan
Ma, and Juan Felipe Pérez, some of which occurred at the American Mathematical Society’s
2015 Mathematics Research Community in commutative algebra, attended by the first author. In addition, Karl and Linquan made detailed comments on an earlier version, which
greatly improved the paper. In particular, Linquan suggested Corollary 6.5.4. We also thank
the referee for helpful suggestions.
2. Preliminaries
Throughout this paper, all rings are assumed to be commutative, and of prime characteristic p unless explicitly stated otherwise. By a local ring, we mean a ring with a unique
maximal ideal, not necessarily Noetherian.
2.1. Valuation Rings. We recall some basic facts and definitions about valuation rings (of
arbitrary characteristic), while fixing notation. See [7, Chapter VI] or [32, Chapter 4] for
proofs and details.
The symbol Γ denotes an ordered abelian group. Recall that such an abelian group is
torsion free. The rational rank of Γ, denoted rat. rank Γ, is the dimension of the Q-vector
space Q ⊗Z Γ.
Let K be a field. A valuation on K is a homomorphism
v : K× → Γ
from the group of units of K, satisfying
v(x + y) ≥ min{v(x), v(y)}
for all x, y ∈ K × . We say that v is defined over a subfield k of K, or that v is a
valuation on K/k, if v takes the value 0 on elements of k.
There is no loss of generality in assuming that v is surjective, in which case we say Γ (or
Γv ) is the value group of v. Two valuations v1 and v2 on K are said to be equivalent if
there is an order preserving isomorphism of their value groups identifying v1 (x) and v2 (x)
for all x ∈ K × . Throughout this paper, we identify equivalent valuations.
The valuation ring of v is the subring Rv ⊆ K consisting of all elements x ∈ K × such
that v(x) ≥ 0 (together with the zero element of K). Two valuations on a field K are
1An
earlier version of this paper used the terms pure F-regularity and split F-regularity for the two generalizations of classical strong F-regularity, depending upon whether maps were required to be pure or split.
The names were changed at Karl Schwede’s urging to avoid confusion with terminology for pairs in [45].
4
equivalent if and only if they determine the same valuation ring. Hence a valuation ring of
K is essentially the same thing as an equivalence class of valuations on K.
The valuation ring of v is local, with maximal ideal mv consisting of elements of strictly
positive values (and zero). The residue field Rv /mv is denoted κ(v). If v is a valuation
over k, then both Rv and κ(v) are k-algebras.
A valuation ring V of K can be characterized directly, without reference to a valuation, as
a subring with the property that for every x ∈ K, either x ∈ V or x−1 ∈ V . The valuation
ring V uniquely determines a valuation v on K (up to equivalence), whose valuation ring in
turn recovers V . Indeed, it is easy to see that the set of ideals of a valuation ring is totally
ordered by inclusion, so the set of principal ideals Γ+ forms a monoid under multiplication,
ordered by (f ) ≤ (g) whenever f divides g. Thus, Γ can be taken to be the ordered abelian
group generated by the principal ideals, and the valuation v : K × → Γ is induced by the
monoid map sending each non-zero x ∈ V to the ideal generated by x. Clearly, the valuation
ring of v is V . See [32, Chapter 4].
2.2. Extension of Valuations. Consider an extension of fields K ⊆ L. By definition, a
valuation w on L is an extension of a valuation v on K if the restriction of w to the subfield
K is v. Equivalently, w extends v if Rw dominates Rv , meaning that Rv = Rw ∩ K with
mw ∩ Rv = mv . In this case, there is an induced map of residue fields
κ(v) ֒→ κ(w).
The residue degree of w over v, denoted by f (w/v), is the degree of the residue field
extension κ(v) ֒→ κ(w).
If w extends v, there is a natural injection of ordered groups Γv ֒→ Γw , since Γv is the
image of w restricted to the subset K. The ramification index of w over v, denoted by
e(w/v), is the index of Γv in Γw .
If K ֒→ L is a finite extension, then both the ramification index e(w/v) and the residue
degree f (w/v) are finite. Indeed, if K ⊆ L is a degree n extension, then
(2.2.0.1)
e(w/v)f (w/v) ≤ n.
More precisely, we have
Proposition 2.2.1. [7, VI.8]. Let K ⊆ L be an extension of fields of finite degree n. For a
valuation v on K, consider the set S of all extensions (up to equivalence) w of v to L. Then
X
e(wi /v)f (wi /v) ≤ n.
wi ∈S
In particular, the set S is finite. Furthermore, equality holds if and only if the integral closure
of Rv in L is a finitely generated Rv -module.
2.3. Abhyankar Valuations. Fix a field K finitely generated over a fixed ground field k,
and let v be a valuation on K/k. By definition, the transcendence degree of v is the
transcendence degree of the field extension
k ֒→ κ(v).
5
The main result about the transcendence degree of valuations is due to Abhyakar [3]. See
also [7, VI.10.3, Corollary 1].
Theorem 2.3.1 (Abhyankar’s Inequality). Let K be a finitely generated field extension of
k, and let v be a valuation on K/k. Then
(2.3.1.1)
trans. deg v + rat. rank Γv ≤ trans. deg K/k.
Moreover if equality holds, then Γv is a finitely generated abelian group, and κ(v) is a
finitely generated extension of k.
We say v is an Abhyankar valuation if equality holds in Abhyankar’s Inequality (2.3.1.1).
Note that an Abyhankar valuation has a finitely generated value group, and its residue field
is finitely generated over the ground field k.
Example 2.3.2. Let K/k be the function field of a normal algebraic variety X of dimension
n over a ground field k. For a prime divisor Y of X, consider the local ring OX,Y of rational
functions on X regular at Y . The ring OX,Y is a discrete valuation ring, corresponding to
a valuation v (the order of vanishing along Y ) on K/k; this valuation is of rational rank
one and transcendence degree n − 1 over k, hence Abhyankar. Such a valuation is called a
divisorial valuation. Conversely, every rational rank one Abhyankar valuation is divisorial:
for such a v, there exists some normal model X of K/k and a divisor Y such that v is the
order of vanishing along Y [48, VI, §14, Thm 31].
Proposition 2.3.3. Let K ⊆ L be a finite extension of finitely generated field extensions
of k, and suppose that w is valuation on L/k extending a valuation v on K/k. Then w is
Abhyankar if and only if v is Abhyankar.
Proof. Since L/K is finite, L and K have the same transcendence degree over k. On the
other hand, the extension κ(v) ⊆ κ(w) is also finite by (2.2.0.1), and so κ(v) and κ(w) also
have the same transcendence degree over k. Again by (2.2.0.1), since Γw /Γv is a finite abelian
group, Q ⊗Z Γw /Γv = 0. By exactness of
0 → Q ⊗Z Γv → Q ⊗Z Γw → Q ⊗Z Γw /Γz → 0
we conclude that Γw and Γv have the same rational rank. The result is now clear from the
definition of an Abhyankar valuation.
F
2.4. Frobenius. Let R be a ring of prime characteristic p. The Frobenius map R → R
is defined by F (x) = xp . We can denote the target copy of R by F∗ R and view it as an
R-module via restriction of scalars by F ; thus F∗ R is both a ring (indeed, it is precisely R)
and an R-module in which the action of r ∈ R on x ∈ F∗ R produces r p x. With this notation,
the Frobenius map F : R → F∗ R and its iterates F e : R → F∗e R are ring maps, as well as
R-module maps. See [44, 1.0.1] for a further discussion of this notation.
We note that F∗e gives us an exact covariant functor from the category of R-modules to
itself. This is nothing but the usual restriction of scalars functor associated to the ring
homomorphism F e : R → R.
e
For an ideal I ⊂ R, the notation I [p ] denotes the ideal generated by the pe -th powers of
e
the elements of I. Equivalently, I [p ] is the expansion of I under the Frobenius map, that is,
e
I [p ] = IF∗e R as subsets of R.
6
e
The image of F e is the subring Rp ⊂ R of pe -th powers. If R is reduced (which is equivalent to the injectivity of Frobenius), statements about the R-module F∗e R are equivalent to
e
statements about the Rp -module R.
Definition 2.4.1. A ring R of characteristic p is F-finite if F : R → F∗ R is a finite map
of rings, or equivalently, if R is a finitely generated Rp -module. Note that F : R → F∗ R is
a finite map if and only if F e : R → F∗e R is a finite map for all e > 0.
F-finite rings are ubiquitous. For example, every perfect field is F-finite, and a finitely
generated algebra over an F-finite ring is F-finite. Furthermore, F-finiteness is preserved
under homomorphic images, localization and completion. This means that nearly every ring
classically arising in algebraic geometry is F-finite. However, valuation rings even of F-finite
fields are often not F-finite.
2.5. F-purity and Frobenius splitting. We first review purity and splitting for maps
of modules over an arbitrary commutative ring A, not necessarily Noetherian or of prime
ϕ
characteristic. A map of A-modules M → N is pure if for any A-module Q, the induced
map
M ⊗A Q → N ⊗A Q
ϕ
is injective. The map M → N is split if ϕ has a left inverse in the category of A-modules.
Clearly, a split map is always pure. Although it is not obvious, the converse holds under a
weak hypothesis:
ϕ
Lemma 2.5.1. [26, Corollary 5.2] Let M → N be a pure map of A-modules where A is a
commutative ring. Then ϕ is split if the cokernel N/ϕ(M) is finitely presented.
Definition 2.5.2. Let R be an arbitrary commutative ring of prime characteristic p.
(a) The ring R is Frobenius split if the map F : R → F∗ R splits as a map of R-modules,
that is, there exists an R-module map F∗ R → R such that the composition
F
R−
→ F∗ R → R
is the identity map.
(b) The ring R is F-pure if F : R → F∗ R is a pure map of R-modules.
A Frobenius split ring is always F-pure. The converse is also true under modest hypothesis:
Corollary 2.5.3. A Noetherian F-finite ring of characteristic p is Frobenius split if and only
if it is F -pure.
Proof. The F-finiteness hypothesis implies that F∗ R is a finitely generated R-module. So a
quotient of F∗ R is also finitely generated. Since a finitely generated module over a Noetherian
ring is finitely presented, the result follows from Lemma 2.5.1.
2.6. F-finiteness and Excellence. Although we are mainly concerned with non-Noetherian
rings in this paper, it is worth pointing out the following curiosity for readers familiar with
Grothendieck’s concept of an excellent ring, a particular kind of Noetherian ring expected
to be the most general setting for many algebro-geometric statements [16, Définition 7.8.2].
7
Proposition 2.6.1. A Noetherian domain is F-finite if and only if it is excellent and its
fraction field is F-finite.
Proof. If R is F-finite with fraction field K, then also R ⊗Rp K p ∼
= K is finite over K p , so
the fraction field of R is F-finite. Furthermore, Kunz showed that F-finite Noetherian rings
are excellent [29, Theorem 2.5].
We need to show that an excellent Noetherian domain with F-finite fraction field is Ffinite. We make use of the following well-known property2 of an excellent domain A: the
integral closure of A in any finite extension of its fraction field is finite as a A-module [16,
IV, 7.8.3 (vi)]. The ring Rp is excellent because it is isomorphic to R, and its fraction field
is K p . Since K p ֒→ K is finite, the integral closure S of Rp in K is a finite Rp -module. But
clearly R ⊂ S, so R is also a finitely generated Rp module, since submodules of a Noetherian
module over a Noetherian ring are Noetherian. That is, R is F-finite.
Using this observation, we can clarify the relationship between F-purity and Frobenius
splitting in an important class of rings.
Corollary 2.6.2. For an excellent Noetherian domain whose fraction field is F-finite, Frobenius splitting is equivalent to F-purity.
Proof of Corollary. Our hypothesis implies F-finiteness, so splitting and purity are equivalent
by Lemma 2.5.1.
3. Flatness and Purity of Frobenius in Valuation Rings.
Kunz showed that for a Noetherian ring of characteristic p, the Frobenius map is flat if
and only if the ring is regular [28, Theorem 2.1]. In this section, we show how standard
results on valuations yield the following result:
Theorem 3.1. Let V be a valuation ring of characteristic p. Then the Frobenius map
F : V → F∗ V is faithfully flat.
This suggests that we can imagine a valuation ring to be “regular” in some sense. Of course,
a Noetherian valuation ring is either a field or a one dimensional regular local ring— but
because valuation rings are rarely Noetherian, Theorem 3.1 is not a consequence of Kunz’s
theorem.
Theorem 3.1 follows from the following general result, whose proof we include for the sake
of completeness.
Lemma 3.2. [7, VI.3.6, Lemma 1]. A finitely-generated, torsion-free module over a valuation
ring is free. In particular, a torsion free module over a valuation ring is flat.
Proof. Let M 6= 0 be a finitely generated, torsion-free V -module. Choose a minimal set of
generators {m1 , . . . , mn }. If there is a non-trivial relation among these generators, then there
exists v1 , . . . , vn ∈ V (not all zero) such that v1 m1 + · · ·+ vn mn = 0. Re-ordering if necessary,
we may assume that v1 is minimal among (non-zero) coefficients, that is, (vi ) ⊂ (v1 ) for all
2sometimes
called the Japanese or N2 property.
8
i ∈ {1, . . . , n}. Then for each i > 1, there exists ai ∈ V such that vi = ai v1 . This implies
that
v1 (m1 + a2 m2 + · · · + an mn ) = 0.
Since v1 6= 0 and M is torsion free, we get
m1 + a2 m2 + · · · + an mn = 0.
Then m1 = −(a2 m2 +· · ·+an mn ). So M can be generated by the smaller set {m2 , . . . , mn }
which contradicts the minimality of n. Hence {m1 , . . . , mn } must be a free generating set.
The second statement follows by considering a torsion-free module as a directed union of
its finitely generated submodules, since a directed union of flat modules is flat [7, I.2.7 Prop
9]
Proof of Theorem 3.1. Observe that F∗ V is a torsion free V -module. So by Corollary 3.2,
the module F∗ V is flat, which means the Frobenius map is flat. To see that Frobenius is
faithfully flat, we need only check that mF∗ V 6= F∗ V for m the maximal ideal of V [7, I.3.5
Prop 9(e)]. But this is clear: the element 1 ∈ F∗ V is not in mF∗ V , since 1 ∈ V is not in the
ideal m[p] .
Corollary 3.3. Every valuation ring of characteristic p is F -pure.
Proof. Fix a valuation ring V of characteristic p. We have already seen that the Frobenius
map V → F∗ V is faithfully flat (Theorem 3.1). But any faithfully flat map of rings A → B
is pure as a map of A-modules [7, I.3.5 Prop 9(c)].
4. F-finite Valuation Rings
In this section, we investigate F-finiteness in valuation rings. We first prove Theorem 4.1.1
characterizing F-finite valuation rings as those V for which F∗ V is a free V -module. We then
prove a numerical characterization of F-finiteness in terms of ramification index and residue
degree for extensions of valuations under Frobenius in Theorem 4.3.1. This characterization
is useful for constructing interesting examples, and later for showing that F-finite valuations
are Abhyankar.
4.1. Finiteness and Freeness of Frobenius. For any domain R of characteristic p, we
have already observed (see the proof of Proposition 2.6.1) that a necessary condition for Ffiniteness is the F-finiteness of its fraction field. For this reason, we investigate F-finiteness
of valuation rings only in F-finite ambient fields.
Theorem 4.1.1. Let K be an F-finite field. A valuation ring V of K is F-finite if and only
if F∗ V is a free V -module.
Proof. First assume F∗ V is free over V . Since K ⊗R F∗ V ∼
= F∗ K as K-vector spaces, the
rank of F∗ V over V must be the same as the rank of F∗ K over K, namely the degree
[F∗ K : K] = [K : K p ]. Since K is F-finite, this degree is finite, and so F∗ V is a free
V -module of finite rank. In particular, V is F-finite.
9
Conversely, suppose that V is F-finite. Then F∗ V is a finitely generated, torsion-free
V -module. So it is free by Lemma 3.2.
Corollary 4.1.2. An F-finite valuation ring is Frobenius split.
Proof. One of the rank one free summands of F∗ V is the copy of V under F , so this copy
of V splits off F∗ V . Alternatively, since V → F∗ V is pure, we can use Lemma 2.5.1: the
cokernel of V → F∗ V is finitely presented because it is finitely generated (being a quotient
of the finitely generated V -module F∗ V ) and the module of relations is finitely generated
(by 1 ∈ F∗ V ).
Remark 4.1.3. The same argument shows that any module finite extension V ֒→ S splits—
in other words, every valuation ring is a splinter in the sense of [31]; see also [31, Lemma
1.2].
4.2. Frobenius Splitting in the Noetherian case. We can say more for Noetherian
valuation rings. First we make a general observation about F-finiteness in Noetherian rings:
Theorem 4.2.1. For a Noetherian domain whose fraction field is F-finite, Frobenius splitting
implies F-finiteness (and hence excellence).
Before embarking on the proof, we point out a consequence for valuation rings:
Corollary 4.2.2. For a discrete valuation ring V whose fraction field is F-finite, the following are equivalent:
(i) V is Frobenius split;
(ii) V is F-finite;
(iii) V is excellent.
Proof of Corollary. A DVR is Noetherian, so equivalence of (i) and (ii) follows from combining Theorem 4.2.1 and Corollary 4.1.2. The equivalence with excellence follows from
Proposition 2.6.1.
Remark 4.2.3. We have proved that all valuation rings are F-pure. However, not all
valuation rings, even discrete ones on Fp (x, y), are Frobenius split, as Example 4.5.1 below
shows.
The proof of Theorem 4.2.1 relies on the following lemma:
Lemma 4.2.4. A Noetherian domain with F-finite fraction field is F-finite if and only there
exists φ ∈ HomRp (R, Rp ) such that φ(1) 6= 0.
Proof of Lemma. Assuming such φ exists, we first observe that the canonical map to the
double dual
R → R∨∨ := HomRp (HomRp (R, Rp ), Rp )
is injective. Indeed, let x ∈ R be a non-zero element. It suffices to show that there exists
f ∈ R∨ := HomRp (R, Rp ) such that f (x) 6= 0. Let f = φ ◦ xp−1 , where xp−1 is the Rp -linear
map R → R given by multiplication by xp−1 . Then f (x) = φ(xp ) = xp φ(1) 6= 0. This shows
that the double dual map is injective.
10
Now, to show that R is a finitely generated Rp -module, it suffices to show that the larger
module R∨∨ is finitely generated. For this it suffices to show that R∨ is a finitely generated
Rp -module, since the dual of a finitely generated module is finitely generated.
We now show that R∨ is finitely generated. Let M be a maximal free Rp -submodule of
R. Note that M has finite rank (equal to [K : K p ], where K is the fraction field of R) and
that R/M is a torsion Rp -module. Since the dual of a torsion module is zero, dualizing the
exact sequence 0 → M → R → R/M → 0 induces an injection
R∨ := HomRp (R, Rp ) ֒→ HomRp (M, R) = M ∨ .
Since M is a finitely generated Rp -module, also M ∨ , and hence its submodule R∨ is finitely
generated (R is Noetherian). This completes the proof that R is F-finite.
For the converse, fix any K p -linear splitting ψ : K → K p . Restricting to R produces
an Rp -linear map to K p . Since R is finitely generated over Rp , we can multiply by some
non-zero element c of Rp to produce a non-zero map φ : R → Rp such that φ(1) = c 6= 0.
The proof of Lemma 4.2.4 is complete.
Proof of Theorem 4.2.1. Let R be a domain with F-finite fraction field. A Frobenius splitting
is a map φ ∈ HomRp (R, Rp ) such that φ(1) = 1. Theorem 4.2.1 then follows immediately
from Lemma 4.2.4.
4.3. A numerical Criterion for F-finiteness. Consider the extension
Kp ⊆ K
where K any field of characteristic p. For any valuation v on K, let v p denote the restriction
to K p . We next characterize F-finite valuations in terms of the ramification index and residue
degree of v over v p :
Theorem 4.3.1. A valuation ring V of an F-finite field K of prime characteristic p is
F-finite if and only if
e(v/v p )f (v/v p) = [K : K p ],
where v is the corresponding valuation on K and v p is its restriction to K p .
Proof. First note that v is the only valuation of K extending v p . Indeed, v is uniquely
determined by its values on elements of K p , since v(xp ) = pv(x) and the value group of v is
torsion-free. Furthermore, the valuation ring of v p is easily checked to be V p .
Observe that V is the integral closure of V p in K. Indeed, since V is a valuation ring, it
is integrally closed in K, but it is also obviously integral over V p . We now apply Lemma
2.2.1. Since there is only one valuation extending v p , the inequality
e(v/v p )f (v/v p) ≤ [K : K p ]
will be an equality if and only if the integral closure of V p in K, namely V , is finite over
V p.
The following simple consequence has useful applications to the construction of interesting
examples of F-finite and non F-finite valuations:
Corollary 4.3.2. Let V be a valuation ring of an F -finite field K of characteristic p. If
e(v/v p ) = [K : K p ] or f (v/v p) = [K : K p ], then V is F -finite.
11
Remark 4.3.3. Theorem 4.3.1 and its corollary are easy to apply, because the ramification
index and residue degree for the extension V p ֒→ V can be computed in practice. Indeed,
since Γvp is clearly the subgroup pΓv of Γv , we see that
e(v/v p ) = [Γv : pΓv ].
(4.3.3.1)
Also, the local map V p ֒→ V induces the residue field extension κ(v p ) ֒→ κ(v), which
identifies the field κ(v p ) with the subfield (κ(v))p . This means that
(4.3.3.2)
f (v/v p) = [κ(v) : κ(v)p ].
4.4. Examples of Frobenius Split Valuations. We can use our characterization of Ffinite valuations to easily give examples of valuations on Fp (x, y) that are non-discrete but
Frobenius split.
Example 4.4.1. Consider the rational function field K = k(x, y) over a perfect field k of
characteristic p. For an irrational number α ∈ R, let Γ be the ordered additive subgroup of
R generated by 1 and α. Consider the unique valuation v : K × → Γ determined by
v(xi y j ) = i + jα,
and let V be the corresponding valuation ring. Since Γ ∼
= Z ⊕ Z via the map which sends
a + bα 7→ (a, b), we see that the value group of v p is pΓ ∼
= p(Z ⊕ Z). Hence
e(v/v p ) = [Γ : pΓ] = p2 = [K : K p ].
So V is F-finite by Corollary 4.3.2. Thus V is also Frobenius split by Corollary 4.1.2.
Example 4.4.2. Consider the lex valuation on the rational function field K = k(x1 , x2 , . . . , xn )
over a perfect field k of characteristic p. This is the valuation v : K × → Zn on K/k defined
by sending a monomial xa11 . . . xann to (a1 , . . . , an ) ∈ Z⊕n , where Γ = Z⊕n is ordered lexicographically. Let V be the corresponding valuation ring. The value group of V p is pΓ, so
e(v/v p ) = [Γ : pΓ] = pn = [K : K p ]. As in the previous example, Corollary 4.3.2 implies that
V is F -finite, and so again F-split.
4.5. Example of a non-Frobenius Split Valuation. Our next example shows that discrete valuation rings are not always F-finite, even in the rational function field Fp (x, y). This
is adapted from [48, Example on pg 62], where it is credited to F. K. Schmidt.
Example 4.5.1. Let Fp ((t)) be the fraction field of the discrete valuation ring Fp [[t]] of
power series in one variable. Since the field of rational functions Fp (t) is countable, the
uncountable field Fp ((t)) can not be algebraic over Fp (t). So we can find some power series
∞
X
f (t) =
an tn
n=1
in Fp [[t]] transcendental over Fp (t).
Since t and f (t) are algebraically independent, there is an injective ring map
Fp [x, y] ֒→ Fp [[t]]
sending x 7→ t and y 7→ f (t)
which induces an extension of fields
Fp (x, y) ֒→ Fp ((t)).
12
Restricting the t-adic valuation on Fp ((t)) to the subfield Fp (x, y) produces a discrete valuation v of Fp (x, y). Let V denote its valuation ring.
We claim that V is not F-finite, a statement we can verify with Theorem 4.3.1. Note that
L = Fp (x, y) is F-finite, with [L : Lp ] = p2 . Since the value group Γv is Z, we see that
e(v/v p ) = [Γv : pΓv ] = p.
On the other hand, to compute the residue degree f (v/v p), we must understand the field
extension κ(v)p ֒→ κ(v). Observe that for an element u ∈ Fp (x, y) to be in V , its image in
Fp ((t)) must be a power series of the form
∞
X
bn tn
n=0
where bn ∈ Fp . Clearly
v(u − b0 ) > 0
which means that the class of u = n=0 bn tn in κ(v) is equal to the class of b0 in κ(v). This
implies that κ(v) ∼
= Fp , so that [κ(v) : κ(v)p ] = 1. That is, f (v/v p ) = 1.
P∞
Finally, we then have that
e(v/v p )f (v/v p ) = p 6= p2 = [Fp (x, y) : (Fp (x, y))p].
So V cannot be F-finite by Theorem 4.3.1. Thus this Noetherian ring is neither Frobenius
split nor excellent by Corollary 4.2.2.
4.6. Finite Extensions. Frobenius properties of valuations are largely preserved under
finite extension. First note that if K ֒→ L is a finite extension of F-finite fields, then
[L : Lp ] = [K : K p ]; this follows immediately from the commutative diagram of fields
LO o
? _K
O
?
?
? _ K p.
Lp o
To wit, [L : K p ] = [L : K][K : K p ] = [L : Lp ][Lp : K p ] and [L : K] = [Lp : K p ], so that
[L : Lp ] = [K : K p ]. Moreover,
Proposition 4.6.1. Let K ֒→ L be a finite extension of F-finite fields of characteristic p.
Let v be a valuation on K and w an extension of v to L. Then:
(i) The ramification indices e(v/v p ) and e(w/w p ) are equal.
(ii) The residue degrees f (v/v p ) and f (w/w p) are equal.
(iii) The valution ring for v is F -finite if and only if the valuation ring for w is F -finite.
Proof. By (2.2.0.1), we have
[Γw : Γv ][κ(w) : κ(v)] ≤ [L : K],
so both [Γw : Γv ] and [κ(w) : κ(v)] are finite. Of course, we also know that the ramification
indices e(w/w p) = [Γw : pΓw ] and e(v/v p ) = [Γv : pΓv ] are finite, as are the residue degrees
f (w/w p) = [κ(w) : κ(w)p ] and f (v/v p ) = [κ(v) : κ(v)p ].
13
(i) In light of (4.3.3.1), we need to show that [Γw : pΓw ] = [Γv : pΓv ]. Since Γw is torsionfree, multiplication by p induces an isomorphism Γw ∼
= pΓw , under which the subgroup Γv
corresponds to pΓv . Thus [pΓw : pΓv ] = [Γw : Γv ]. Using the commutative diagram of finite
index abelian subgroups
? _ Γv
ΓO w o
O
?
?
? _ pΓv ,
pΓw o
we see that [Γw : pΓw ][pΓw : pΓv ] = [Γw : Γv ][Γv : pΓv ]. Whence [Γw : pΓw ] = [Γv : pΓv ].
(ii) In light of (4.3.3.2), we need to show that [κ(w) : κ(w)p ] = [κ(v) : κ(v)p ]. We have
[κ(w)p : κ(v)p ] = [κ(w) : κ(v)], so the result follows from computing the extension degrees
in the commutative diagram of finite field extensions
κ(w) o
? _ κ(v)
O
O
?
?
? _ κ(v)p .
κ(w)p o
(iii) By (i) and (ii) we get e(w/w p) = e(v/v p ) and f (w/w p) = f (v/v p). Therefore
e(w/w p)f (w/w p) = e(v/v p )f (v/v p ).
Since also [L : Lp ] = [K : K p ], we see using Theorem 4.3.1 that w is F -finite if and only
if v is F -finite.
5. F-finiteness in Function Fields
An important class of fields are function fields over a ground field k. By definition, a field
K is a function field over k if it is a finitely generated field extension of k. These are the
fields that arise as function fields of varieties over a (typically algebraically closed) ground
field k. What more can be said about valuation rings in this important class of fields?
We saw in Example 4.5.1 that not every valuation of an F-finite function field is F-finite.
However, the following theorem gives a nice characterization of those that are.
Theorem 5.1. Let K be a finitely generated field extension of an F-finite ground field k.
The following are equivalent for a valuation v on K/k:
(i) The valuation v is Abhyankar.
(ii) The valuation ring Rv is F-finite.
(iii) The valuation ring Rv is a free Rvp -module.
Furthermore, when these equivalent conditions hold, it is also true that Rv is Frobenius split.
Since Abhyankar valuations have finitely generated value groups and residue fields, the
following corollary holds.
Corollary 5.2. An F-finite valuation of a function field over an F-finite field k has finitely
generated value group and its residue field is a finitely generated field extension of k.
14
For example, valuations whose value groups are Q can never be F-finite.
Remark 5.3. In light of Proposition 2.6.1, we could add a fourth item to the list of equivalent conditions in Theorem 5.1 in the Noetherian case: the valuation Rv is excellent. The
theorem says that the only discrete valuation rings (of function fields) that are F-finite are
the divisorial valuation rings or equivalently, the excellent DVRs.
To prove Theorem 5.1, first recall that the equivalence of (ii) and (iii) was already established in Theorem 4.1.1. The point is to connect these conditions with the Abyhankar
property. Our strategy is to use Theorem 4.3.1, which tells us that a valuation v on K is
F-finite if and only if
e(v/v p )f (v/v p) = [K : K p ].
We do this by proving two propositions—one comparing the rational rank of v to the ramification index e(v/v p ), and the other comparing the transcendence degree of v to the residue
degree f (v/v p ).
Proposition 5.4. Let v be a valuation of rational rank s on an F-finite field K. Then
e(v/v p ) ≤ ps ,
with equality when the value group Γv is finitely generated.
Proof. To see that equality holds when Γv is finitely generated, note that in this case, Γv ∼
=
⊕s
s
p
s
⊕s
∼
Z . So Γv /pΓv = (Z/pZ) , which has cardinality p . That is, e(v/v ) = p .
It remains to consider the case where Γ may not be finitely generated. Nonetheless, since
e(v/v p ) is finite (C.f. 2.2.0.1), we do know that [Γv : pΓv ] = e(v/v p ) is finite. So the proof of
Proposition 5.4 comes down to the following simple lemma about abelian groups:
Lemma 5.5. Let Γ be a torsion free abelian group of rational rank s. Then [Γ : pΓ] ≤ ps .
It suffices to show that Γ/pΓ is a vector space over Z/pZ of dimension ≤ s. For if d
is the dimension of Γ/pΓ, then [Γ : pΓ] = pd . So let t1 , . . . , tn be elements of Γ whose
classes modulo pΓ are linearly independent over Z/pZ. Then we claim that the ti are Zindependent elements of Γ. Assume to the contrary that there is some non-trivial relation
a1 t1 + · · · + an tn = 0, for some integers ai . Since Γ is torsion-free, we can assume without
loss of generality, that at least one aj is not divisible by p. But now modulo pΓ, this relation
produces a non-trivial relation on classes of the ti in Γ/pΓ, contrary to the fact that these are
linearly independent. This shows that any Z/pZ-linearly independent subset of Γ/pΓ must
have cardinality at most s. Then the lemma, and hence Proposition 5.4, are proved.
Proposition 5.6. Let K be a finitely generated field extension of an F-finite ground field
k. Let v be a valuation of transcendence degree t on K over k. Then
f (v/v p ) ≤ pt [k : k p ],
with equality when κ(v) is finitely generated over k.
Proof. The second statement follows immediately from the following well-known fact, whose
proof is an easy computation:
Lemma 5.7. A finitely generated field L of characteristic p and transcendence degree n over
k satisfies [L : Lp ] = [k : k p ]pn .
15
It remains to consider the case where κ(v) may not be finitely generated. Because K/k is
a function field, Abhyankar’s inequality (2.3.1.1) guarantees that the transcendence degree
of κ(v) over k is finite. Let x1 , . . . , xt be a transcendence basis. There is a factorization
k ֒→ k(x1 , . . . , xt ) ֒→ κ(v)
where the second inclusion is algebraic. The proposition follows immediately from:
Lemma 5.8. If L′ ⊆ L is an algebraic extension of F-finite fields, then [L : Lp ] ≤ [L′ : L′p ].
To prove this lemma, recall that Proposition 4.6.1 ensures that [L : Lp ] = [L′ : L′p ] when
L′ ⊆ L is finite. So suppose L is algebraic but not necessarily finite over L′ . Fix a basis
{α1 , . . . , αn } for L over Lp , and consider the intermediate field
L′ ֒→ L′ (α1 , . . . , αn ) ֒→ L.
Since each αi is algebraic over L′ , it follows that L̃ := L′ (α1 , . . . , αn ) is finite over L′ , so
again [L̃ : L̃p ] = [L′ : L′p ] by Proposition 4.6.1. Now observe that L̃p ⊂ Lp , and so the
Lp -linearly independent set {α1 , . . . , αn } is also linearly independent over L̃p . This means
that [L : Lp ] ≤ [L̃ : L̃p ] and hence [L : Lp ] ≤ [L′ : L′p ]. This proves Lemma 5.8.
Finally, Proposition 5.6 is proved by applying Lemma 5.8 to the inclusion
L′ = k(x1 , . . . , xt ) ֒→ L = κ(v).
[Note that κ(v) is F-finite, because [κ(v) : (κ(v))p ] = f (v/v p) ≤ [K : K p ] from the general
inequality (2.2.0.1)).] So we get
f (v/v p) = [κ(v) : (κ(v))p ] ≤ [k(x1 , . . . , xt ) : (k(x1 , . . . , xt ))p ] = pt [k : k p ].
Proof of Theorem 5.1. It only remains to prove the equivalence of (i) and (ii). First assume
v is Abhyankar. Then its value group Γv is finitely generated and its residue field κ(v) is
finitely generated over k. According to Proposition 5.4, we have e(v/v p ) = ps , where s is
the rational rank of v. According to Proposition 5.6, we have f (v/v p ) = pt [k : k p ], where
t is the transcendence degree of v. By definition of Abhyankar, s + t = n, where n is the
transcendence degree of K/k. But then
e(v/v p )f (v/v p ) = (ps )(pt )[k : k p ] = pn [k : k p ] = [K : K p ].
By Theorem 4.3.1, we can conclude that v is F-finite.
Conversely, we want to prove that a valuation v with F-finite valuation ring Rv is Abhyankar. Let s denote the rational rank and t denote the transcendence degree of v. From
Theorem 4.3.1, the F-finiteness of v gives
e(v/v p )f (v/v p ) = [K : K p ] = pn [k : k p ].
Using the bounds ps ≥ e(v/v p ) and pt [k : k p ] ≥ f (v/v p ) provided by Propositions 5.4 and
5.6, respectively, we substitute to get
(ps )(pt [k : k p ]) ≥ e(v/v p )f (v/v p) = pn [k : k p ].
It follows that s + t ≥ n. Then s + t = n by (2.3.1.1), and v is Abhyankar.
16
6. F-regularity
An important class of F-pure rings are the strongly F-regular rings. Originally, strongly
F-regular rings were defined only in the Noetherian F-finite case. By definition, a Noetherian
F-finite reduced ring R of prime characteristic p is strongly F-regular if for every non-zerodivisor c, there exists e such that the map
R → F∗e R
sending 1 7→ c
splits in the category of R-modules [21]. In this section, we show that by replacing the word
"splits" with the words "is pure" in the above definition, we obtain a well-behaved notion
of F-regularity in a broader setting. Hochster and Huneke themselves suggested, but never
pursued, this possibility in [23, Remark 5.3].
Strong F-regularity first arose as a technical tool in the theory of tight closure: Hochster
and Huneke made use of it in their deep proof of the existence of test elements [23]. Indeed,
the original motivation for (and the name of) strong F-regularity was born of a desire to
better understand weak F-regularity, the property of a Noetherian ring that all ideals are
tightly closed. In many contexts, strong and weak F-regularity are known to be equivalent
(see e.g. [30] for the graded case, [21] for the Gorenstein case) but it is becoming clear that
at least for many applications, strong F-regularity is the more useful and flexible notion.
Applications beyond tight closure include commutative algebra more generally [2, 5, 39,
36, 44], algebraic geometry [15, 18, 34, 42, 41], representation theory [8, 33, 35, 43] and
combinatorics [4].
6.1. Basic Properties of F-pure regularity. We propose the following definition, intended to be a generalization of strong F-regularity to arbitrary commutative rings of characteristic p, not necessarily F-finite or Noetherian.
Definition 6.1.1. Let c be an element in a ring R of prime characteristic p. Then R is said
to be F-pure along c if there exists e > 0 such that the R-linear map
λec : R → F∗e R sending 1 7→ c
is a pure map of R-modules. We say R is F-pure regular if it is F-pure along every
non-zerodivisor.
A ring R is F-pure if and only if it is F -pure along the element 1. Thus F-pure regularity is
a substantial strengthening of F-purity, requiring F-purity along all non-zerodivisors instead
of just along the unit.
Remark 6.1.2.
(i) If R is Noetherian and F-finite, then the map λec : R → F∗e R is
pure if and only if it splits (by Lemma 2.5.1). So F-pure regularity for a Noetherian
F-finite ring is the same as strong F-regularity.
(ii) If c is a zerodivisor, then the map λec is never injective for any e ≥ 1. In particular,
a ring is never F -pure along a zerodivisor.
(iii) The terminology “F-pure along c” is chosen to honor Ramanathan’s closely related
notion of “Frobenius splitting along a divisor” [35]. See [41].
The following proposition gathers up some basic properties of F-pure regularity for arbitrary commutative rings.
17
Proposition 6.1.3. Let R be a commutative ring of characteristic p, not necessarily Noetherian or F-finite.
(a) If R is F-pure along some element, then R is F-pure. More generally, if R is F -pure
along a product cd, then R is F -pure along the factors c and d.
(b) If R is F -pure along some element, then R is reduced.
(c) If R is an F-pure regular ring with finitely many minimal primes, and S ⊂ R is a
multiplicative set, then S −1 R is F-pure regular. In particular, F-pure regularity is
preserved under localization in Noetherian rings, as well as in domains.
(d) Let ϕ : R → T be a pure ring map which maps non-zerodivisors of R to nonzerodivisors of T . If T is F-pure regular, then R is F-pure regular. In particular,
if ϕ : R → T is faithfully flat and T is F-pure regular, then R is F-pure regular.
(e) Let R1 , . . . , Rn be rings of characteristic p. If R1 × · · · × Rn is F-pure regular, then
each Ri is F-pure regular.
The proof of Proposition 6.1.3 consists mostly of applying general facts about purity to
the special case of the maps λec . For the convenience of the reader, we gather these basic
facts together in one lemma:
Lemma 6.1.4. Let A be an arbitrary commutative ring A, not necessarily Noetherian nor
of characteristic p.
(a) If M → N and N → Q are pure maps of A-modules, then the composition M →
N → Q is also pure.
(b) If a composition M → N → Q of A-modules is pure, then M → N is pure.
(c) If B is an A-algebra and M → N is pure map of A-modules, then B ⊗A M → B ⊗A N
is a pure map of B-modules.
(d) Let B be an A-algebra. If M → N is a pure map of B-modules, then it is also pure
as a map of A-modules.
(e) An A-module map M → N is pure if and only if for all prime ideals P ⊂ A,
MP → NP is pure.
(f) A faithfully flat map of rings is pure.
(g) If (Λ, ≤) is a directed set with a least element λ0 , and {Nλ }λ∈Λ is a direct limit system
of A-modules indexed by Λ and M → Nλ0 is an A-linear map, then M → limλ Nλ is
−→
pure if and only if M → Nλ is pure for all λ.
(h) A map of modules A → N over a Noetherian local ring (A, m) is pure if and only if
E ⊗A A → E ⊗A N is injective where E is the injective hull of the residue field of R.
Proof of Lemma 6.1.4. Properties (a)-(d) follow easily from the definition of purity and elementary properties of tensor product. As an example, let us prove (d). If P is an A-module,
we want to show that P ⊗A M → P ⊗A N is injective. The map of B-modules
(P ⊗A B) ⊗B M → (P ⊗A B) ⊗B N
is injective by purity of M → N as a map of B-modules. Using the natural A-module
isomorphisms (P ⊗A B) ⊗B M ∼
= P ⊗A N, we conclude that
= P ⊗A M and (P ⊗A B) ⊗B N ∼
P ⊗A M → P ⊗A N is injective in the category of A-modules.
18
Property (e) follows from (c) by tensoring with Ap and the fact that injectivity of a map of
modules is a local property. Property (f) follows from [7, I.3.5, Proposition 9(c)]. Properties
(g) and (h) are proved in [24, Lemma 2.1].
Proof of Proposition 6.1.3. (a) Multiplication by d is an an R-linear map, so by restriction
of scalars also
×d
F∗e R −→ F∗e R
is R-linear. Precomposing with λec we have
λe
×d
c
R −→
F∗e R −→ F∗e R sending 1 7→ cd,
which is λecd . Our hypothesis that R is F-pure along cd means that there is some e for which
this composition is pure. So by Lemma 6.1.4(b), it follows also that λec is pure. That is, R
is F-pure along c (and since R is commutative, along d). The second statement follows since
F-purity along the product c × 1 implies R is F-pure along 1. So some iterate of Frobenius
is a pure map, and so F-purity follows from Lemma 6.1.4(b).
(b) By (a) we see that R is F-pure. In particular, the Frobenius map is pure and hence
injective, so R is reduced.
(c) Note R is reduced by (b). Let α ∈ S −1 R be a non-zerodivisor. Because R has finitely
many minimal primes, a standard prime avoidance argument shows that there exists a nonzerodivisor c ∈ R and s ∈ S such that α = c/s (a minor modification of [20, Proposition on
Pg 57]). By hypothesis, R is F-pure along c. Hence there exists e > 0 such that the map
λec : R → F∗e R is pure. Then the map
λec/1 : S −1 R −→ F∗e (S −1 R) sending 1 7→ c/1
is pure by 6.1.4(e) and the fact that S −1 (F∗e R) ∼
= F∗e (S −1 R) as S −1 R-modules (the isomore
phism S −1 (F∗e R) ∼
= F∗e (S −1 R) is given by r/s 7→ r/sp ). Now the S −1 R-linear map
ℓ1/s : S −1 R → S −1 R sending
is an isomorphism.
Applying F∗e , we see that
F∗e (ℓ1/s ) : F∗e (S −1 R) → F∗e (S −1 R)
−1
is also an isomorphism of S
modules. So purity of
1 7→ 1/s
sending 1 7→ 1/s
R-modules. In particular, F∗e (ℓ1/s ) is a pure map of S −1 R-
F∗e (ℓ1/s ) ◦ λec/1
follows by 6.1.4(a). But F∗e (ℓ1/s ) ◦ λec/1 is precisely the map
λec/s : S −1 R → F∗e (S −1 R) sending 1 7→ c/s.
(d) Let c ∈ R be a non-zerodivisor. Then ϕ(c) is a non-zero divisor in T by hypothesis.
Pick e > 0 such that the map λeϕ(c) : T → F∗e T is a pure map of T -modules. By 6.1.4(f) and
6.1.4(a),
ϕ
λeϕ(c)
R−
→ T −−−→ F∗e T
19
is a pure map of R-modules. We have commutative diagram of R-linear maps
ϕ
R
λeϕ(c)
λec
F∗e R
T
F∗e (ϕ)
F∗e T
The purity of λec follows by 6.1.4(b). Note that if ϕ is faithfully flat, then it is pure by
6.1.4(f) and maps non-zerodivisors to non-zerodivisors.
(e) Let R := R1 × · · · × Rn . Consider the multiplicative set
S := R1 × · · · × Ri−1 × {1} × Ri+1 × · · · × Rn .
Since S −1 R ∼
= Ri , it suffices to show that S −1 R is F-pure regular. So let α ∈ S −1 R be a
non-zerodivisor. Note that we can select u ∈ R and s ∈ S such that u is a non-zerodivisor
and α = u/s. So we can now repeat the proof of (c) verbatim to see that S −1 R must be
pure along α.
Remark 6.1.5. It is worth observing in Definition 6.1.1, that if the map λec is a pure map,
then λfc is also a pure map for all f ≥ e. Indeed, to see this note that it suffices to show that
λe+1
is pure. We know R is F-pure by 6.1.3(a). So Frobenius
c
F : R → F∗ R
is a pure map of R-modules. By hypothesis,
λec : R → F∗e R
is pure. Hence 6.1.4(d) tell us that
F∗ (λec ) : F∗ R → F∗ (F∗e R)
is a pure map of R-modules. Hence the composition
F
F∗ (λe )
c
R−
→ F∗ R −−−→
F∗ (F∗e R) sending 1 7→ c
is a pure map of R-modules by 6.1.4(a). But F∗ (F∗e R) as an R-module is precisely F∗e+1 R.
So
λe+1
: R → F∗e+1 R.
c
is pure.
Example 6.1.6. The polynomial ring over Fp in infinitely many variables (localized at the
obvious maximal ideal) is an example of a F-pure regular ring which is not Noetherian.
6.2. Relationship of F-pure regularity to other singularities. We show that our generalization of strong F-regularity continues to enjoy many important properties of the more
restricted version.
Theorem 6.2.1. (C.f. [21, Theorem 3.1(c)]) A regular local ring, not necessarily F-finite,
is F-pure regular.
Proof. Let (R, m) be a regular local ring. By Krull’s intersection theorem we know that
\
e
m[p ] = 0.
e>0
20
Since R is a domain, the non-zerodivisors are precisely the non-zero elements of R. So let
e
c ∈ R be a non-zero element. Choose e such that c ∈
/ m[p ] . We show that the map
λec : R → F∗e R; 1 7→ c
is pure.
By Lemma 6.1.4, it suffices to check that for the injective hull E of the residue field of R,
the induced map
λec ⊗ idE : R ⊗R E → F∗e R ⊗R E
is injective, and for this, in turn, we need only check that the socle generator is not in the
kernel.
Recall that E is the direct limit of the injective maps
x
x
x
R/(x1 , . . . , xn ) −→ R/(x21 , . . . , x2n ) −→ R/(x31 , . . . , x3n ) −→ R/(x41 , . . . , x4n ) −→ · · ·
where x1 , . . . , xn is a minimal set of generators for m, and the maps are given by multiplication by x = Πdi=1 xi . So the module F∗e R ⊗R E is the direct limit of the maps
e
e
xp
e
e
e
xp
e
e
e
xp
e
e
e
4p
4p
R/(xp1 , . . . , xpn ) −→ R/(x12p , . . . , xn2p ) −→ R/(x13p , . . . , x3p
n ) −→ R/(x1 , . . . , xn ) −→ · · ·
which remains injective by the faithful flatness of F∗e R. The induced map λec ⊗ idE : E →
e
F∗e R ⊗ E sends the socle (namely the image of 1 in R/m) to the class of c in R/m[p ] , so
e
it is non-zero provided c ∈
/ m[p ] . Thus for every non-zero c in a regular local (Noetherian)
ring, we have found an e, such that the map λec is pure. So regular local rings are F-pure
regular.
Proposition 6.2.2. An F-pure regular ring is normal, that is, it is integrally closed in its
total quotient ring.
Proof. Take a fraction r/s in the total quotient ring integral over R. Then clearing denominators in an equation of integral dependence, we have r ∈ (s), the integral closure of the
ideal (s). This implies that there exists an h such that (r, s)n+h = (s)n (r, s)h for all n [32,
p64]. Setting c = sh , this implies cr n ∈ (s)n for all large n. In particular, taking n = pe , we
see that class of r modulo (s) is in the kernel of the map induced by tensoring the map
(6.2.2.1)
R → F∗e R sending 1 7→ c
with the quotient module R/(s). By purity of the map (6.2.2.1), it follows that r ∈ (s). We
conclude that r/s is in R and that R is normal.
6.3. Connections with Tight Closure. In his lecture notes on tight closure [20], Hochster
suggests another way to generalize strong F-regularity to non-F-finite (but Noetherian) rings
using tight closure. We show here that his generalized strong F-regularity is the same as
F-pure regularity for local Noetherian rings.
Although Hochster and Huneke introduced tight closure only in Noetherian rings, we can
make the same definition in general for an arbitrary ring of prime characteristic p. Let
∗
N ֒→ M be R-modules. The tight closure of N in M is an R-module NM
containing
∗
N. By definition, an element x ∈ M is in NM if there exists c ∈ R, not in any minimal
prime, such that for all sufficiently large e, the element c ⊗ x ∈ F∗e R ⊗R M belongs to the
image of the module F∗e R ⊗R N under the natural map F∗e R ⊗R N → F∗e R ⊗R M induced
21
by tensoring the inclusion N ֒→ M with the R-module FRe . We say that N is tightly closed
∗
in M if NM
= N.
Definition 6.3.1. Let R be a Noetherian ring of prime characteristic p. We say that R is
∗
strongly F-regular in the sense of Hochster if, for any pair of R modules N ֒→ M, NM
= N.
The next result compares F-pure regularity with strongly F-regularity in the sense of
Hochster:
Proposition 6.3.2. Let R be an arbitrary commutative ring of prime characteristic. If R
is F-pure regular, then N is tightly closed in M for any pair of R modules N ⊂ M. The
converse also holds if R is Noetherian and local.
∗
Proof. (i) Suppose x ∈ NM
. Equivalently the class x of x in M/N is in 0∗M/N . So there exists
c not in any minimal prime such that c ⊗ x = 0 in F∗e R ⊗R M/N for all large e. But this
means that the map
R → F∗e R sending 1 7→ c
is not pure for any e, since the naturally induced map
R ⊗ M/N → F∗e R ⊗ M/N
has 1 ⊗ x in its kernel.
For the converse, let c ∈ R be not in any minimal prime. We need to show that there
exists some e such that the map R → F∗e R sending 1 to c is pure. Let E be the injective hull
of the residue field of R. According to Lemma 6.1.4(i), it suffices to show that there exists
an e such that after tensoring E, the induced map
R ⊗ E → F∗e R ⊗ E
is injective. But if not, then a generator η for the socle of E is in the kernel for every e, that
is, for all e, c ⊗ η = 0 in F∗e R ⊗ E. In this case, η ∈ 0∗E , contrary to our hypothesis that all
modules are tightly closed.
Remark 6.3.3. We do not know whether Proposition 6.3.2 holds in the non-local case.
Indeed, we do not know if F-pure regularity is a local property: if Rm is F-pure regular for
all maximal ideals m of R, does it follow that R is F-pure regular? If this were the case,
then our argument above extends to arbitrary Noetherian rings.
Remark 6.3.4. A Noetherian ring of characteristic p is weakly F-regular if N is tightly
closed in M for any pair of Noetherian R modules N ⊂ M. Clearly F-pure regular implies
weakly F-regular. The converse is a long standing open question in the F-finite Noetherian
case. For valuation rings, however, our arguments show that weak and pure F-regularity are
equivalent (and both are equivalent to the valuation ring being Noetherian); See Corollary
6.5.4.
6.4. Elements along which F-purity fails. We now observe an analog of the splitting
prime of Aberbach and Enescu [1]; See also [47, 4.7].
Proposition 6.4.1. Let R be a ring of characteristic p, and consider the set
I := {c ∈ R : R is not F-pure along c}.
Then I is closed under multiplication by R, and R−I is multiplicatively closed. In particular,
if I is closed under addition, then I is a prime ideal (or the whole ring R).
22
Proof of Proposition 6.4.1. We first note that I is closed under multiplication by elements
of R. Indeed, suppose that c ∈ I and r ∈ R. Then if rc ∈
/ I, we have that R is F-pure along
rc, but this implies R is F-pure along c by Proposition 6.1.3(a), contrary to c ∈ I.
We next show that the complement R \ I is a multiplicatively closed set (if non-empty).
To wit, take c, d ∈
/ I. Because R is F-pure along both c and d, we have that there exist e
and f such such the maps
λf
λe
d
c
F∗f R sending 1 7→ d
F∗e R sending 1 7→ c, and R −→
R −→
are both pure. Since purity is preserved by restriction of scalars (Lemma 6.1.4(d)), we also
have that
F∗e (λf )
F∗e R −→d F∗e F∗f R = F∗e+f R
is pure. Hence the composition
λe
λf
e
c
d
F∗e R −→
F∗e F∗f R sending 1 7→ cp d
R −→
e
is pure as well (Lemma 6.1.4(a)). This means that cp d is not in I, and since I is closed
under multiplication, neither is cd. Note also that if R \ I is non-empty, then 1 ∈ R \ I by
Proposition 6.1.3(a). Thus R \ I is a multiplicative set.
Finally, if I is closed under addition (and I =
6 R), we conclude that I is a prime ideal
since it is an ideal whose complement is a multiplicative set.
Remark 6.4.2. If R is a Noetherian local domain, then the set I of Proposition 6.4.1 can
be checked to be closed under addition (see, for example, [47, 4.7] for the F-finite case).
Likewise, for valuation rings, the set I is also an ideal: we construct it explicitly in the next
section. However, for an arbitrary ring, I can fail to be an ideal. For example, under suitable
hypothesis, the set I is also the union of the centers of F-purity in the sense of Schwede,
hence in this case, I is a finite union of ideals but not necessarily an ideal in the non-local
case; see [38].
6.5. F-pure regularity and Valuation Rings. In this subsection we characterize valuation rings that are F-pure regular. The main result is:
Theorem 6.5.1. A valuation ring is F-pure regular if and only if it is Noetherian. Equivalently, a valuation ring is F-pure regular if and only if it is a field or a DVR.
A key ingredient in the proof is the following theorem about the set of elements along
which V fails to be F-pure (C.f. Definition 6.1.1):
Theorem 6.5.2. The set of elements c along which a valuation ring (V, m) fails to be F-pure
is the prime ideal
\
e
Q :=
m[p ] .
e>0
Proof. First, take any c ∈ Q. We need to show that V is not F-pure along c, that is, that
the map
λec : V → F∗e V sending 1 7→ c
e
is not pure for any e. Because c ∈ m[p ] , we see that tensoring with κ := V /m produces the
zero map. So λec is not pure for any e, which means V is not F-pure along c.
23
e
For the other inclusion, let c ∈
/ m[p ] for some e > 0. We claim that λec : V → F∗e V is pure.
Apply Lemma 6.1.4(g) to the set Σ of finitely generated submodules of F∗e V which contain
c. Note Σ is a directed set under inclusion with a least element, namely the V -submodule
of F∗e V generated by c, and F∗e V is the direct limit of the elements of Σ. It suffices to show
that if T ∈ Σ, then
λT : V → T sending 1 7→ c
is pure. But T is free since it is a finitely generated, torsion-free module over a valuation
e
ring (Lemma 3.2). Since c ∈
/ m[p ] , by the V module structure on T , we get c ∈
/ mT . By
Nakayama’s Lemma, we know c is part of a free basis for T . So λT splits, and is pure in
particular.
Now that we know that the set of elements along which R is not F-pure is an ideal, it
follows that it is a prime ideal from Proposition 6.4.1.
T
e
Corollary 6.5.3. For a valuation ring (V, m) of characteristic p, define Q := e>0 m[p ] .
Then the quotient V /Q is a F-pure regular valuation ring. Furthermore, V is F-pure regular
if and only if Q is zero.
Proof. The second statement follows immediately from Theorem 6.5.2. For the first, observe
that V /Q is a domain since Q is prime. So ideals of V /Q inherit the total ordering
under
T
[pe ]
inclusion from V , and V /Q is a valuation ring whose maximal m ideal satisfies e>0 m = 0.
So V /Q is F-pure regular.
Corollary 6.5.4. For a valuation ring, F-pure regularity is equivalent to all ideals (equivalently, the maximal ideal) being tightly closed.
Proof. Proposition 6.3.2 ensures that F-pure regularity implies
all ideals are tightly closed.
T
[pe ]
For the converse, note that if there is some non-zero c in e>0 m , then 1 ∈ m∗ . So for any
T
e
proper ideal m, the condition that m∗ = m implies that e>0 m[p ] = 0. In particular, if the
maximal ideal of a valuation ring V is tightly closed, then Corollary 6.5.3 implies that V is
F-pure regular.
Proof of Theorem 6.5.1. First observe that if V is a field or DVR, then it is F-pure regular.
Indeed, every map of modules over a field is pure (since all vector space maps split). And a
DVR is a one dimensional regular local ring, so it is F-pure regular by Theorem 6.2.1.
Conversely, we show that if (V, m) is F-pure regular, its dimension is at most one. Suppose
(V, m) admits a non-zero prime ideal P 6= m. Choose x ∈ m \ P , and a non-zero element
c ∈ P. The element c cannot divide xn in V , since in that case we would have xn ⊂ (c) ⊂ P ,
but P is a prime ideal not containing x. It follows from the definition of a valuation ring,
e
e
then, that xn divides c for all n. This means in particular that c ∈ (x)[p ] ⊂ m[p ] for all e.
So c ∈ Q. According to Theorem 6.5.2, R is not F-pure regular.
It remains to show that an F-pure regular valuation ring V of dimension one is discrete.
Recall that the value group Γ of V is (order isomorphic to) an additive subgroup of R [32,
Theorem 10.7].
We claim that Γ has a least positive element. To see this, let η be the greatest lower
bound of all positive elements in Γ. First observe that η is strictly positive. Indeed, for fixed
24
consists of positive real numbers approaching zero as e gets large. If
c ∈ m, the sequence v(c)
pe
Γ contains elements of arbitrarily small positive values, then we could find x ∈ V such that
v(c)
0 < v(x) < e .
p
e
e
e
But then 0 < v(xp ) < v(c), which says that c ∈ (x)[p ] ⊂ m[p ] for all e. This contradicts
our assumption that V is F-pure along c (again, using Theorem 6.5.2).
Now that we know the greatest lower bound η of Γ is positive, it remains to show that
η ∈ Γ. Choose ǫ such that 0 < ǫ < η. If η ∈
/ Γ, we know η < v(y) for all y ∈ m. Since η is
the greatest lower bound, we can find y such that
η < v(y) < η + ǫ,
as well as x such that
η < v(x) < v(y) < η + ǫ.
Then
0 < v(y/x) < ǫ < η,
contradicting the fact that η is a lower bound for Γ. We conclude that η ∈ Γ, and that Γ
has a least positive element.
It is now easy to see using the Archimedean axiom for real numbers that the ordered
subgroup Γ of R is generated by its least positive element η. In particular, Γ is order
isomorphic to Z. We conclude that V is a DVR.
Remark 6.5.5. For a valuation ring (V, m) of dimension n ≥ 1, our results show that in
general
\
e
Q=
m[p ]
e∈N
is a prime ideal of height at least n − 1. It is easy to see that the situation where V /Q is
a DVR arises if and only if m is principal, which in turn is equivalent to the value group Γ
having a least positive element. For example, this is the case for the lex valuation in Example
4.4.2. It is not hard to check that Q is uniformly F-compatible ideal in the sense of Schwede
[38] (see also [44, 3.1] for further discussion of uniformly F-compatible ideals), generalizing of
course to the non-Noetherian and non-F-finite setting. A general investigation of uniformly
F-compatible ideals appears to be fruitful, and is being undertaken by the first author.
6.6. Split F-regularity. Of course, there is another obvious way3 to adapt Hochster and
Huneke’s definition of strongly F-regular to arbitrary rings of prime characteristic p:
Definition 6.6.1. A ring R is split F-regular if for all non-zero divisors c, there exists e
such that the map R → F∗e R sending 1 to c splits as a map of R-modules.
Since split maps are pure, a split F-regular ring is F-pure regular. Split F-regular rings
are also clearly Frobenius split. On the other hand, Example 4.5.1 shows that a discrete
valuation ring need not be Frobenius split, so split F-regularity is strictly stronger than
F-pure regularity. In particular, not every regular local ring is split F-regular, so split Fregularity should not really be considered a class of "singularities" even for Noetherian rings.
3This
generalization is used for cluster algebras in [4] for example.
25
Remark 6.6.2. In Noetherian rings, split F-regularity is very close to F-pure regularity. For
example, if R is an F-pure regular Noetherian domain whose fraction field is F-finite, then
the only obstruction to split F-regularity is the splitting of Frobenius. This is a consequence
of Theorem 4.2.4, which tells us R is F-finite if it is Frobenius split, and Lemma 2.5.1, which
tells us F-split and F-pure are the same in F-finite Noetherian rings.
Corollary 6.6.3. For a discrete valuation ring V whose fraction field is F-finite, the following are equivalent:
(i)
(ii)
(iii)
(iv)
(v)
V
V
V
V
V
is
is
is
is
is
split F-regular;
Frobenius split;
F-finite;
free over V p ;
excellent.
Moreover, if K is a function field over an F-finite ground field k, and V is a valuation of
K/k, then (i)-(v) are equivalent to V being a divisorial valuation ring.
Proof. All this has been proved already. Recall that a DVR is a regular local ring, so it is
always F-pure regular and hence split F-regular if it is F-finite. Also, the final statement
follows from Theorem 5.1 because an Abyhankar valuation of rational rank one is necessarily
divisorial, and a divisorial valuation of a functional field over an F-finite field is necessarily
F -finite.
To summarize: A valuation ring is F-pure regular if and only if it is Noetherian, and split
F-regular (under the additional assumption that its fraction field is F-finite) if and only if it
is excellent.
7. Concluding Remarks
We have argued that for valuation rings, F-purity and F-pure regularity (a version of
strong F-regularity defined using pure maps instead of split maps) are natural and robust
properties. We have also seen that the conditions of Frobenius splitting and split F-regularity
are more subtle, and that even regular rings can fail to satisfy these.
For Noetherian valuation rings in F-finite fields, we have seen that the Frobenius splitting property is equivalent to F-finiteness and also to excellence, but we do not know what
happens in the non-Noetherian case: does there exist an example of a (necessarily nonNoetherian) Frobenius split valuation ring of an F-finite field that is not F-finite? By
Corollary 5.2, a possible strategy could be to construct a Frobenius split valuation ring
in a function field whose value group is infinitely generated. For example, can one construct
an F-split valuation in Fp (x, y) with value group Q? On the other hand, perhaps Frobenius
splitting is equivalent to F-finiteness (just as in the Noetherian case). One might then ask
whether a generalized version of Theorem 4.1.1 holds for arbitrary fields: is a valuation ring
Frobenius split if and only if Frobenius is free?
We propose that F-pure regularity is a more natural generalization of strong F-regularity
to the non-F-finite case than a suggested generalization of strong F-regularity using tight
closure due to Hochster. We have seen that F-pure regularity implies Hochster’s notion, and
26
that they are equivalent for local Noetherian rings. However, we do not know whether F-pure
regularity is a local notion: if Rm is F-pure regular for all maximal ideals, does it follow that
R is F-pure regular? We expect this to be true but the standard arguments are insufficient
to prove it. (In the Noetherian F-finite case, this is well known; C.f. [23, Theorem 5.5(a)].
Furthermore, the answer is affirmative for excellent rings with F-finite total quotient rings by
Proposition 2.6.1.) If true, then F-pure regularity would be equivalent to all modules being
tightly closed in the Noetherian case. More generally, might F-pure regularity be equivalent
to the property that all modules are tightly closed even in the non-Noetherian case? Or even
that all ideals are tightly closed? An affirmative answer to this last question would imply
that strong and weak F-regularity are equivalent.
References
[1] I. M. Aberbach and F. Enescu. The structure of F-pure rings. Mathematische Zeitschrift, 250(4):791–806,
2005.
[2] I. M. Aberbach and J. G. Leuschke. The F-signature and strong F-regularity.
http://arxiv.org/abs/math/0208067.
[3] S. Abhyankar. On the valuations centered in a local domain. American Journal of Mathematics,
78(2):321–348, 1956.
[4] A. Benito, G. Muller, J. Rajchgot, and K. E. Smith. Singularities of locally acyclic cluster algebras.
Algebra and Number Theory, 9(4):913–936, 2015.
[5] M. Blickle. Minimal γ-sheaves. Algebra and Number Theory, 2(3):347–368, 2008.
[6] S. Boucksom. Corps d’Okounkov. Séminaire Bourbaki, 65(1059), October 2012.
[7] N. Bourbaki. Commutative Algebra, Chapters 1-7. Springer, 1989.
[8] M. Brion and S. Kumar. Frobenius splitting methods in geometry and representation theory. Birkhäuser
Boston, 2005.
[9] B. Conrad. Several approaches to non-Archimedean geometry in p-adic Geometry (Lectures from the
2007 Arizona Winter School). AMS, Providence, RI, 2008.
[10] S. D. Cutkosky. Resolution of Singularities, volume 63 of Graduate Studies in Mathematics. AMS,
Providence, RI, 2004.
[11] J. F. de Bobadilla and M. P. Pereira. The Nash problem for surfaces. Annals of Math., 176:2003–2029,
2012.
[12] L. Ein, R. Lazarsfeld, and K. E. Smith. Uniform approximation of Abhyankar valuation ideals in smooth
function fields. American Journal of Mathematics, 125(2):409–440, 2003.
[13] C. Favre and M. Jonsson. The Valuative Tree. Springer, 2004.
[14] S. Fomin and A. Zelevinsky. Cluster algebras i: Foundations. J. Amer. Math. Soc., 15:497–529, 2002.
[15] Y. Gongyo, Z. Li, Z. Patakfalvi, K. Schwede, H. Tanaka, and R. Zong. On rational connectedness of
globally F-regular threefolds. Advances in Math., 280(6):47–78, 2015.
[16] A. Grothendieck and J. Dieudonné. Eléments de géométrie algébrique: IV, seconde partie. Publications
Mathématiques de l’IHÉS, 24:5–231, 1965.
[17] W.
Gubler,
J.
Rabinoff,
and
A.
Werner.
Skeletons
and
Tropicalizations.
http://arxiv.org/abs/1404.7044.
[18] C. D. Hacon and C. Xu. On the three dimensional minimal model program in positive characteristic. J.
Amer. Math. Soc., 28(3):711–744, 2015.
[19] N. Hara and K Watanabe. F-regular and F-pure rings vs. log terminal and log canonical singularities.
J. Algebraic Geom, 11(2):363–392, 2002.
[20] M. Hochster. Foundations of tight closure theory. http://www.math.lsa.umich.edu/~hochster/711F07/fndtc.pdf.
[21] M. Hochster and C. Huneke. Tight closure and strong F-regularity. Mèmoires de la S.M.F., 38:119–133,
1989.
[22] M. Hochster and C. Huneke. Tight closure, invariant theory and the Briançon-Skoda theorem. J. Amer.
Math. Soc., 3:31–116, 1990.
27
[23] M. Hochster and C. Huneke. F-regularity, test elements, and smooth base change. Trans. Amer. Math.
Soc., 346:1–62, 1994.
[24] M. Hochster and C. Huneke. Applications of the existence of big Cohen-Macaulay algebras. Advances
in Mathematics, 113(1):45–117, 1995.
[25] M. Hochster and J. L. Roberts. Rings of invariants of reductive groups acting on regular rings are
Cohen-Macaulay. Advances in Mathematics, 13:115–175, 1974.
[26] M. Hochster and J. L. Roberts. The purity of the Frobenius and local cohomology. Advances in Mathematics, 21(2):117–172, 1976.
[27] H. Knaf and F. Kuhlmann. Abhyankar places admit local uniformization in any characteristic. Annales
scientifiques de l’École Normale Supérieure, 38(6):833–846, 2005.
[28] E. Kunz. Characterizations of regular local rings of characteristic p. American Journal of Mathematics,
91(3):772–784, 1969.
[29] E. Kunz. On noetherian rings of characteristic p. American Journal of Mathematics, 98(4):999–1013,
1976.
[30] G. Lyubeznik and K. E. Smith. Strong and weak F-regularity are equivalent for graded rings. Amer. J.
Math., 121(6):1279–1290, 1999.
[31] F. Ma. Splitting in integral extensions, Cohen-Macaulay modules and algebras. Journal of Algebra,
116:176–195, 1988.
[32] H. Matsumura. Commutative Ring Theory. Cambridge University Press, 1989.
[33] V. B. Mehta and A. Ramanathan. Frobenius splitting and cohomology vanishing for Schubert varieties.
Ann. of Math, 122(2):27–40, 1985.
[34] Z. Patakfalvi. Semi-positivity in positive characteristics. Algebra and Number Theory, 2(3):347–368,
2008.
[35] A. Ramanathan. Frobenius splitting and Schubert varieties. Proceedings of the Hyderabad Conference
on Algebraic Groups, pages 497–508, 1991.
[36] K. Schwede. F-adjunction. Algebra and Number Theory, 3(8):907–950, 2009.
[37] K. Schwede. F-injective singularities are Du Bois. Amer. J. Math., 131(2):445–473, 2009.
[38] K. Schwede. Centers of F-purity. Mathematische Zeitschrift, 265(3):687–714, 2010.
[39] K. Schwede and K. Tucker. A survey of Test Ideals. http://arxiv.org/abs/1104.2000.
[40] K. E. Smith. F-rational rings have rational singularities. Amer. J. Math., 119(1):159–180, 1997.
[41] K. E. Smith. Globally F-regular varieties: applications to vanishing theorems for quotients of Fano
varieties. Michigan Math J., 48:553–572, 2000.
[42] K. E. Smith and K. Schwede. Globally F-regular and log Fano varieties. Adv. Math., 224(3):863–894,
2010.
[43] K. E. Smith and M. Van den Bergh. Simplicity of rings of differential operators in prime characteristic.
Proc. London Math. Soc., 75(1):32–62, 1997.
[44] K.
E.
Smith
and
W.
Zhang.
Frobenius
splitting
in
commutative
algebra.
http://arxiv.org/abs/1409.1169.
[45] S. Takagi. F-singularities of pairs and Inversion of Adjunction of arbitrary codimension. Inventiones
Mathematicae, 157(1):123–146, 2004.
[46] S. Takagi. A characteristic p analogue of plt singularities and adjoint ideals. Mathematische Zeitschrift,
259(2):321–341, June 2008.
[47] K. Tucker. F-signature exists. Inventiones Mathematicae, 190(3):743–765, 2012.
[48] O. Zariski and P. Samuel. Commutative Algebra, Volume II. Springer-Verlag, 1960.
E-mail address: [email protected], [email protected]
28
ERRATUM TO FROBENIUS AND VALUATION RINGS
arXiv:1507.06009v3 [math.AC] 29 Dec 2016
RANKEYA DATTA AND KAREN E. SMITH
Abstract. Theorem 5.1 is corrected in the paper Frobenius and valuations rings, DOI:
10.2140/ant.2016.10.1057.
The characterization of F-finite valuation rings of function fields (Theorem 5.1) is incorrect. The corrected characterization is
Theorem 0.1. Let K be a finitely generated extension of an F-finite ground field k. Let V
be any non-trivial valuation ring of K/k. Then V is F-finite if and only if V is divisorial.
A proof of this corrected theorem follows a complete accounting of affected statements
in the paper. Notation, Theorem and page numbers are as in [2].
1. List of Affected Statements
The source of the error is Proposition 2.2.1 (pg. 1061), which is misquoted from the
original source, [1]. Specifically, the last sentence in the statement of Proposition 2.2.1
should read: Furthermore, equality holds if the integral closure of Rν in L is a finitely
generated Rν -module, not if and only if. The statement is correct, however, under the
additional assumption that the valuation ring Rν is Noetherian [1, VI, §8.5, Remark (1)].
This error has consequences in the following statements from the paper:
Theorem 4.3.1 should read: Let V be a valuation ring of an F-finite field K of prime
characteristic p. If V is F-finite, then
[Γ : pΓ][κ : κp ] = [K : K p ],
where Γ is the value group and κ is the residue field of V .
The paper had incorrectly stated the converse, which does hold if the valuation ring is
Noetherian [1, VI, §8.5, Remark (1)].
Corollary 4.3.2 should read: With hypotheses as in Thm 4.3.1, if [K : K p ] = [κ : κp ],
then V is F-finite. See Section 2 for the proof. The original statement holds as stated if
V is Noetherian.
Examples 4.4.1 and 4.4.2 (pg. 1069): The computations here are correct but the
conclusions are not. Rather, Theorem 0.1 ensures these valuation rings are not F-finite.
Example 4.5.1 (pg. 1070) is correct as stated, as it follows from the correct implication
of Theorem 4.3.1.
The first author was partially supported by the Juha Heinonen Memorial Graduate Fellowship at the
University of Michigan. The second author was partially supported by NSF grant DMS-1501625.
1
Proposition 4.6.1 (pg. 1071): This statement is correct. However, the proof of (iii) is
not, because it cited the incorrect implication of Proposition 2.2.1. The corrected proof
is in Section 3 below.
Theorem 5.1 (pg. 1072): The equivalence of (ii) and (iii) is correct, and both imply (i).
Under the additional assumption that the valuation is discrete, also (i) implies (ii) and
(iii). The proof of Theorem 5.1 in the paper proves:
Theorem 1.1. Let K be a finitely generated field extension of an F-finite ground field k
of characteristic p. The following are equivalent for a valuation v on K/k:
(i) The valuation v is Abhyankar.
(ii) [Γ : pΓ][κ : κp ] = [K : K p ]
where Γ is the value group, and κ is the residue field of v.
Corollary 5.2 (pg. 1072) is correct as stated; it is a consequence of Theorem 0.1.
Corollary 6.6.3 (pg. 1086) is correct as stated; the proof should invoke Theorem 0.1
instead of Theorem 5.1.
The remaining statements in Section 5 (Remark 5.3, Proposition 5.4, Lemma 5.5,
Proposition 5.6, Lemma 5.7, Lemma 5.8) and Section 6 are correct, because they do
not rely on Lemma 2.2.1 or its consequences.
2. Proof of Theorem 0.1
Lemma 2.1. Let (V, m) be a valuation ring of prime characteristic p. If m is not principal,
then
m = m[p] .
Proof. Take any x ∈ m. It suffices to show that there exists y ∈ m such that y p|x. Note
that since m is not principal, the value group Γ of the corresponding valuation v does not
have a smallest element > 0. This means we can choose β0 ∈ Γ such that 0 < β0 < v(x),
and then also choose β1 ∈ Γ such that
0 < β1 < min{β0 , v(x) − β0 }.
Let y1 ∈ m such that v(y1) = β1 . Then
v(y12) = 2β1 < β0 + (v(x) − β0 ) = v(x).
Thus, y12 |x. Repeating this argument inductively, we can find yn ∈ m such that yn2 |yn−1
n
for all n ∈ N. For n > log2 p, we have yn2 , and hence ynp , divides x.
Lemma 2.2. Let (V, m, κ) be valuation ring of characteristic p. Then the dimension of
V /m[p] over κp is
(a) [κ : κp ] if m is not finitely generated.
(b) p[κ : κp ] if m is finitely generated.
2
Proof. Consider the short exact sequence of κp -vector spaces
(1)
0 → m/m[p] → V /m[p] → κ → 0.
If m is not finitely generated, then Lemma 2.1 implies that m/m[p] = 0, and (a) follows.
Otherwise, m is principal, so m[p] = mp and we have a filtration
m ) m2 ) · · · ) mp−1 ) m[p] = mp .
Since mi /mi+1 ∼
= κ, we see that
dimκp (m/m[p] ) = (p − 1)[κ : κp ].
From the short exact sequence (1), dimκp (V /m[p] ) = p[κ : κp ], proving (b).
Proof of Theorem 0.1. If V is divisorial, it is a localization of a finitely generated algebra
over the F -finite ground field k, hence it is F-finite.
For the converse, let m be the maximal ideal and κ the residue field of V . Since V is
F-finite, by (the corrected) Theorem 4.3.1
(2)
[Γ : pΓ][κ : κp ] = [K : K p ],
where Γ is the value group of V . Thus, the valuation is Abhyankar by Theorem 1.1 above.
To show it is divisorial, we need to show it is discrete.
Because the value group of an Abhyankar valuation is finitely generated, Γ ∼
= Z⊕s , for
some integer s ≥ 1. Hence it suffices to show that s = 1.
Since V is F-finite, we know V is free over V p of rank [K : K p ]. Tensoring with the
residue field κp of V p , we have that V /m[p] is also a free κp -module of rank [K : K p ]. Thus
from (2), we have
(3)
dimκp V /m[p] = [K : K p ] = [Γ : pΓ][κ : κp ] = |Z⊕s /pZ⊕s |[κ : κp ] = ps [κ : κp ].
But now since ps 6= 1, Lemma 2.2 forces s = 1, and the proof is complete.
Remark 2.3. Theorem 0.1 does not hold without some assumption on the field K. For
instance, if K is perfect, Frobenius will be an isomorphism (hence a finite map) for any
valuation ring. The proof of Theorem 0.1 does show that an F-finite valuation ring with
finitely generated value group must be Noetherian (hence discrete), without any restriction
on its fraction field. Furthermore, the proof also shows that a valuation ring cannot be
F-finite if its value group Γ satisfies [Γ : pΓ] > p.
Proof of Revised Corollary 4.3.2: In general, [Γ : pΓ][κ : κp ] ≤ [K : K p ] by [1, VI,
§8.1, Lemma 2]. Hence our hypothesis forces [Γ : pΓ] = 1. Thus Γ = pΓ, so that Γ can
not have a smallest positive element, which means that the maximal ideal m of V is not
finitely generated. Lemma 2.2(a) now ensures dimκp (V /m[p] ) = [κ : κp ] = [K : K p ]. Then
V is F-finite by [1, VI, §8.5, Theorem 2(c)].
3
3. Proof of Proposition 4.6.1(iii):
We recall Proposition 4.6.1(iii): Let K ֒→ L be a finite extension of F-finite fields of
characteristic p. Let w be a valuation on L and v its restriction to K. Then the valuation
ring of v is F-finite if and only if the valuation ring of w is F-finite.
Lemma 3.1. With notation as in Proposition 4.6.1(iii), the maximal ideal of the valuation
ring of v is finitely generated if and only if the maximal ideal of the valuation ring of w
is finitely generated.
Proof. For ideals in a valuation ring, finite generation is the same being principal. Principality of the maximal ideal is equivalent to the value group having a smallest element
> 0. Thus, it suffices to show that the value group Γv of v has this property if and only
if Γw does.
Assume Γw has a smallest element g > 0. We claim that for each t ∈ N, the only
positive elements of Γw less than tg are g, 2g, . . . , (t − 1)g. Indeed, suppose 0 < h < tg.
Since g is smallest, g ≤ h < tg, whence 0 ≤ h − g < (t − 1)g. So by induction, g − h = ig
for some i ∈ {0, 1, . . . , t − 2}, and hence h is among g, 2g, . . . , (t − 1)g.
Now, because [Γw : Γv ] ≤ [L : K] < ∞ by [1, VI, §8.1, Lemma 2], every element of
Γw /Γv is torsion. Let n be the smallest positive integer such that ng ∈ Γv . We claim that
ng is the smallest positive element of Γv . Indeed, the only positive elements smaller than
ng in Γw are g, 2g, . . . , (n − 1)g, and none of these are in Γv by our choice of n.
Conversely, if Γv has a smallest element h > 0, then the set
S := {g ∈ Γw : 0 < g < h}
is finite because for distinct g1 , g2 in this set, their classes in Γw /Γv are also distinct, while
Γw /Γv is a finite group. Then the smallest positive element of Γw is the smallest element
of S, or h if S is empty.
Proof of Proposition 4.6.1(iii). A necessary and sufficient condition for the F-finiteness
of a valuation ring (V, m, κ) with F-finite fraction field K is, by [1, VI, §8.5, Theorem
2(c)], that
(4)
dimκp (V /m[p]) = [K : K p ].
Lemma 2.2 gives a formula for dimκp (V /m[p]) in terms of [κ : κp ] that depends on whether
the maximal ideal is finitely generated, which is the same for v and w by Lemma 3.1.
Proposition 4.6.1 (i) and (ii) tells us that both [K : K p ] = [L : Lp ] and [κv : κpv ] = [κw : κpw ].
Thus Lemma 2.2 and Equation 4 guarantee that the valuation ring of v is F-finite if and
only if the valuation ring of w is F-finite.
References
[1] N. Bourbaki. Commutative Algebra, Chapters 1-7. Springer, 1989.
[2] R. Datta and K.E. Smith. Frobenius and valuation rings. Algebra and Number Theory, 10(5):1057–
1090, 2016.
E-mail address: Rankeya Datta [email protected], [email protected]
4
| 0math.AC
|
Bias Correction with Jackknife, Bootstrap, and Taylor Series
Jiantao Jiao,
Yanjun Han,
and Tsachy Weissman∗
arXiv:1709.06183v1 [math.ST] 18 Sep 2017
September 20, 2017
Abstract
We analyze the bias correction methods using jackknife, bootstrap, and Taylor series. We
focus on the binomial model, and consider the problem of bias correction for estimating f (p),
where f ∈ C[0, 1] is arbitrary. We characterize the supremum norm of the bias of general
jackknife and bootstrap estimators for any continuous functions, and demonstrate the in deleted jackknife, different values of d may lead to drastically different behavior in jackknife. We
show that in the binomial model, iterating the bootstrap bias correction infinitely many times
may lead to divergence of bias and variance, and demonstrate that the bias properties of the
bootstrap bias corrected estimator after r−1 rounds is exactly the same as that of the r-jackknife
estimator if a bounded coefficients condition is satisfied.
1
Introduction
One of the classic problems in statistics is to design procedures to reduce the bias of estimators.
General bias correction methods such as the bootstrap, the jackknife, and the Taylor series have
been widely employed and well studied in the literature. See [Que56, AGW71, Mil74, Efr79, Efr82,
RS02, Wit87, HM88, Hal92, ET94, PRW99, ZI04, ST12, CCN14] and the references therein.
A close inspection of the literature on those general bias correction methods show that they
usually rely on certain expansion (differentiability) properties of the expectation of the estimator
one would like to correct the bias for, and the analysis is pointwise asymptotics [Wit87,Hal92,ST12].
One motivation for this work is that the methods based on series expansions and differentiability
assumptions may not suffice in the analysis of bootstrap and jackknife even in the simplest statistical models, and the practical implementations of bootstrap and jackknife do not require those
differentiability conditions. The Taylor series itself, by definition, is a series expansion method
which we include here for comparison with bootstrap and jackknife.
To illustrate our point, consider one of the simplest statistical models, the binomial model,
where n · p̂n ∼ B(n, p). For any function f : [0, 1] 7→ R, we would like to correct the bias of f (p̂n )
as an estimator of f (p). Let e1,n (p) = f (p) − Ep f (p̂n ) be the bias term. The expectation of the
jackknife bias corrected estimator fˆ2 satisfies
E[fˆ2 ] = E [nf (p̂n ) − (n − 1)f (p̂n−1 )] ,
(1)
∗
Jiantao Jiao, Yanjun Han, and Tsachy Weissman are with the Department of Electrical Engineering, Stanford
University. Email: {jiantao, yjhan, tsachy}@stanford.edu
1
where (n − 1) · p̂n−1 ∼ B(n − 1, p). The textbook argument of the bias reduction property of the
jackknife is the following [ST12]. Suppose that
a(p) b(p)
1
e1,n (p) =
+ 2 + Op
,
n
n
n3
(2)
where a(p), b(p) are unknown functions of p which do no depend on n, and Op (an ) is a sequence
that is elementwise upper bounded by an up to a multiplicative constant for fixed p. We also have
1
b(p)
a(p)
+ Op
.
+
e1,n−1 (p) =
2
n − 1 (n − 1)
(n − 1)3
(3)
Hence, the overall bias of fˆ2 is:
f (p) − Ep fˆ2 = ne1,n (p) − (n − 1)e1,n−1 (p)
b(p)
b(p)
1
=
−
+ Op
n
n−1
n2
1
b(p)
+ Op
,
=−
n(n − 1)
n2
(4)
(5)
(6)
which seems to suggest that the bias has been reduced to order n12 instead of order n1 . However, if
we particularize (2) to f (p) = p ln(1/p), which relates to the Shannon entropy [Sha48], we have
1
1 − p 1 − 3p + 2p2
−
+ Op
.
e1,n (p) =
2
2n
6pn
n3
(7)
One immediately sees that it may not be reasonable to claim that the jackknife has reduced the
bias upon looking at (7) and (6). Indeed, the bias of the jackknife estimator is uniformly upper
bounded by O(n), but the right hand side of (7) and (6) explodes to infinity as p → 0. It shows
that one cannot ignore the dependence on p in the Op (·) notation, but even doing higher order of
Taylor expansion does not help. In fact, it was shown first in [Pan03] that for f (p) = p ln(1/p),
there exist universal constants C1 > 0, C2 > 0 such that
sup |e1,n (p)| ≤
p∈[0,1]
C1
n
C2
.
sup |f (p) − Ep [fˆ2 ]| ≥
n
p∈[0,1]
(8)
(9)
In other words, the jackknife does not change the bias order at all. There exist other estimators that
achieve a smaller order of bias. Indeed, the estimators [VV11, VV13, WY16, JVHW15] that achieve
the optimal minimax sample complexity for Shannon entropy estimation used best approximation
1
polynomials to reduce the bias of of each symbol from n1 to n ln
n.
In this paper, we connect the jackknife and bootstrap to the theory of approximation, and
provide a systematic treatment of the problem of correcting the bias for f (p̂n ) as an estimator of
f (p) for f ∈ C[0, 1] and n · p̂n ∼ B(n, p). Compared with existing literature, we choose to simplify
the statistical model to the extreme, but consider arbitrary functions f . We believe it is an angle
worth investigating due to the following reasons. First of all, it directly leads to analysis of the
bias correction properties of jackknife and bootstrap for important statistical questions such as the
2
Shannon entropy estimation, which existing theory proves insufficient of handling. Second, even in
this simplest statistical model the analysis of jackknife and bootstrap is far non-trivial, and there
still exist abundant open problems as we discuss in this paper. One message we would like to convey
in this work is that the analysis of jackknife and bootstrap in simple statistical models but general
functions could lead to interesting and deep mathematical phenomena that remain fertile ground
for research. Moreover, we emphasize that the insights obtained through analyzing the binomial
model could generalize to other statistical models. 1
We introduce some notations below. The r-th symmetric difference of a function f : [0, 1] 7→ R
is given by
∆rh f (x)
=
r
X
!
k
(−1)
k=0
r
f (x + r(h/2) − kh),
k
(10)
where ∆rh f (x) = 0 if x + rh/2 or x − rh/2 is not in [0, 1]. We introduce the r-th Ditzian–Totik
modulus of smoothness of a function f : [0, 1] 7→ R as
ωϕr (f, t) = sup
0<h≤t
∆rhϕ(x) f ,
(11)
p
where ϕ(x) = x(1 − x), and the norm is the supremum norm.
The ωϕr (f, t) modulus satisfies the following properties.
Lemma 1. [DT87, Chap. 4] The Ditzian–Totik modulus of smoothness ωϕr (f, t) in (11) satisfies
the following:
1. ωϕr (f, t) is a nondecreasing function of t.
2. There exist universal constants K > 0, t0 > 0 such that ωϕr (f, λt) ≤ Kλr ωϕr (f, t) for λ ≥ 1,
and λt ≤ t0 .
3. There exist universal constants K > 0, t0 > 0 such that ωϕr+1 (f, t) ≤ Kωϕr (f, t) for 0 < t ≤ t0 .
4. There exists a universal constant K > 0 such that ωϕr (f, t) ≤ K supx∈[0,1] |f (r) (x)|tr .
ω r (f,t)
5. limt→0+ ϕtr
= 0 ⇒ f is a polynomial with degree r − 1. (f is a polynomial of degree
r
r − 1 ⇒ ωϕ (f, t) = 0).
6. ωϕr (f, t) = O(tr ) for fixed f, r if and only if f (r−1) ∈ A.C.loc and kϕr f (r) k < ∞,
where f (r−1) ∈ A.C.loc means that f is r − 1 times differentiable and f (r−1) is absolutely continuous
in every closed finite interval [c, d] ⊂ (0, 1).
1
We mention that most of the results in this paper could be generalized to the case of natural exponential family of
quadratic variance functions [Mor82, Mor83], which comprises of six families: Gaussian, Poisson, binomial, negative
binomial, gamma, and generalized hyperbolic secant. Coincidentally, these distribution families were also identified as
special in approximation theory literature, where they were named operators of the exponential-type [May76, IM78].
3
We emphasize that Ditzian–Totik modulus of smoothness is easy to compute for various functions. For example, for f (x) = xδ | ln x/2|γ , x ∈ (0, 1). Then for r ≥ 2δ, we have [DT87, Section
3.4]:
ωϕr (f, t) r,δ,γ
t2δ | ln t|γ
δ∈
/ Z, δ ≥ 0, γ ≥ 0
t2δ | ln t|γ−1
δ ∈ Z, δ ≥ 0, γ ≥ 1
.
(12)
An intuitive understanding of ωϕr (f, t) is the following. If the function f is “smoother”, the
modulus is smaller. However, a non-zero ωϕr (f, t) cannot vanish faster than the order tr for any
fixed f .
Notation: All the norms in this paper refer to the supremum norm. Concretely kf k =
supx |f (x)|. For non-negative sequences aγ , bγ , we use the notation aγ .α bγ to denote that there
a
exists a universal constant C that only depends on α such that supγ bγγ ≤ C, and aγ &α bγ is
equivalent to bγ .α aγ . Notation aγ α bγ is equivalent to aγ .α bγ and bγ .α aγ . We write
aγ . bγ if the constant is universal and does not depend on any parameters. Notation aγ bγ
a
means that lim inf γ bγγ = ∞, and aγ bγ is equivalent to bγ aγ . We write a ∧ b = min{a, b} and
a ∨ b = max{a, b}. Moreover, polydn denotes the set of all d-variate polynomials of degree of each
variable no more than n, and En [f ; I] denotes the distance of the function f to the space polydn
in the uniform norm k · k∞,I on I ⊂ Rd . The space poly1n is also abbreviated as polyn . All logarithms are in the natural base. The notation Eθ [X] denotes the mathematical expectation of the
random variable X whose distribution is indexed by the parameter θ. The s-backward difference
of a function defined over integers Gn is
s
∆ Gn ,
s
X
!
k
(−1)
k=0
s
Gn−k .
k
(13)
Remark 1 (Operator view of bias reduction). It was elaborated in [JVHW17] that for any statistical model, the quantity Eθ F (θ̂) could be viewed as an operator that maps the function F (·) to
another function of θ. The operator is obviously linear in F , and is also positive in the sense that
if F ≥ 0 everywhere it is also everywhere non-negative. If we view Eθ F (θ̂) as an approximation of
F (θ), then analyzing the bias of the estimator F (θ̂) is equivalent to analyzing the approximation
error of Eθ F (θ̂).
Casting the bias analysis problem as an approximation problem, a key observation of this work
is that the bootstrap bias correction could be viewed as the iterated Boolean sum approximation,
and the jackknife bias correction could be viewed as a linear combination approximation, and the
Taylor series bias correction corresponds to the Taylor series approximation. The main tool we
use to handle these approximation theoretic questions is the K-functional, which we introduce in
Section A.
We now summarize our main results for jackknife, bootstrap, and Taylor series bias correction.
1.1
Jackknife bias correction
The jackknife is a subsampling technique [ST12] that aims at making the biases of estimators with
different sample sizes cancel each other.
4
Definition 1 (r-jackknife estimator). Fix r ≥ 1, r ∈ Z. Fix K > 0 such that K does not scale
with n. For a collection of sample sizes n1 < n2 < n3 < . . . < nr ≤ Kn1 = n, under the binomial
model the general r-jackknife estimator is defined as
r
X
fˆr =
Ci Un [f (p̂ni )],
(14)
i=1
P
β
f (Xβ1 ,Xβ2 ,...,Xβq )
is the U -statistic with kernel f , and the sum is over all distinct
(nq)
combinations of q elements from total n samples from the Bernoulli distribution Bern(p). Each
ni · p̂ni ∼ B(ni , p). The coefficients {Ci } are given by
where Un [f ] =
Ci =
ni
, 1 ≤ i ≤ r.
n − nj
j6=i i
Y
(15)
If nr = n, ni − ni−1 = d, then it is called the delete-d r-jackknife estimator.
Note that the standard jackknife in (1) corresponds to n1 = n − 1, n2 = n, whose corresponding
coefficients are C1 = −(n − 1), C2 = n. As shown in Lemma 13 in Section B, the coefficients
{Ci }1≤i≤r in (15) satisfy the following:
r
X
i=1
r
X
i=1
Ci = 1
Ci
= 0,
nρi
1 ≤ ρ ≤ r − 1, ρ ∈ Z.
(16)
The intuition behind this condition is clear: only through these equations can one completely
cancel any bias terms of order 1/nρ , ρ ≤ r − 1. It is also clear that (16) corresponds to solving
a linear system with the Vandermonde matrix, for which the solution given in (15) is the unique
solution due to the fact that all the ni ’s are distinct. The rationale above also appeared in [SGO71],
and the corresponding coefficients Ci were given in the form of determinants. Equation (15) shows
that in this special case the coefficients admit a simple expression.
1.1.1
Jackknife with the bounded coefficients condition
We introduce the following condition on {Ci }1≤i≤r which turns out to be crucial for the bias and
variance properties of the general r-jackknife.
Condition 1 (Bounded coefficients condition). We say that the jackknife coefficients Ci in (15)
satisfy the bounded coefficients condition with parameter C if there exists a constant C that only
depends on r such that
r
X
|Ci | ≤ C.
(17)
i=1
One motivation for Condition 1 is the following. Observe that
Ep fˆr =
r
X
Ci Ep [f (p̂ni )].
i=1
5
(18)
Viewing Ep [f (p̂ni )] as an operator that maps f to a polynomial, it is an approximation to f (p),
which, by the Bernstein theorem [AS04, Chap. 7] in fact limn→∞ kEp f (p̂n ) − f (p)k∞ = 0 for any
continuous function f on [0, 1]. Then, one can view the r-jackknife as a linear combination of
operators. In this sense, Condition 1 assures that the linear combination as a new operator has
bounded norm that is independent of n.
The following theorem quantifies the performance of the general r-jackknife under the bounded
coefficients condition in Condition 1.
Theorem 1. Suppose {Ci }1≤i≤r satisfies Condition 1 with parameter C. Suppose r ≥ 1 is a fixed
integer. Let fˆr denote the general r-jackknife in (14). Then, for any f ∈ C[0, 1], the following is
true.
1.
√
kf (p) − Ep fˆr k .r,C ωϕ2r (f, 1/ n) + n−r kf k
(19)
kf (p) − Ep fˆr k .α,r,C n−α/2 ⇔ ωϕ2r (f, t) .α,r,C tα .
(20)
2. Fixing 0 < α < 2r,
3. Suppose there is a constant D < 22r for which
ωϕ2r (f, 2t) ≤ Dωϕ2r (f, t) for t ≤ t0 .
(21)
√
kf (p) − Ep fˆr k r,C,D ωϕ2r (f, 1/ n).
(22)
√
kf (p) − Ep fˆr k ωϕ2r (f, 1/ n).
(23)
Then,
4. For r = 1,
The following corollary of Theorem 1 is immediate given (12).
Corollary 1. Under the conditions in Theorem 1, if f = −p ln p, then
1
kf (p) − Ep fˆr k r,C .
n
(24)
1
.
nα
(25)
If f (p) = pα , 0 < α < 1, then,
kf (p) − Ep fˆr k r,C
Corollary 1 implies that the r-jackknife estimator for fixed r does not improve the bias of f (p̂n )
for f (p) = p ln(1/p), which makes it incapable of achieving the minimax rates of Shannon entropy
estimation [WY16, JVHW15].
6
1.1.2
Jackknife without the bounded coefficients condition
Theorem 1 has excluded the case that n1 = n − 1, n2 = n. Clearly, to satisfy the assumptions of
Theorem 1, we need to require |ni − ni−1 | & n1 , which puts a minimum gap between the different
sample sizes we can use. This begs the question: is this condition necessary to Theorem 1 to hold?
If not, what bad consequences it will lead to?
From a computational perspective, taking d small in the delete-d jackknife may reduce the
computational burden. However, as we now show, the usual delete-1 jackknife does not satisfy
Theorem 1 in general and exhibits drastically different bias and variance properties. 2
We now show that the delete-1 jackknife may have bias and variance both diverging to infinity
in the worst case.
Theorem 2. Let fˆr denote the delete-1 r-jackknife estimator. There exists a fixed function f ∈
C(0, 1] that satisfies kf k∞ ≤ 1 such that
kEp fˆr − f (p)k∞ & nr−1 .
(26)
If we allow the function f to depend on n, then one can have f ∈ C[0, 1]. 3
Meanwhile, for any n ≥ 4, there exists a function f ∈ C[0, 1] depending on n such that
n2
kVarp (fˆ2 )k∞ ≥
.
e
(30)
Theorem 2 shows that in the worst case, the delete-1 r-jackknife may have bad performances
compared to that satisfying Condition 1. Before we delve into the refined analysis of delete-1 rjackknife, we illustrate the connection between various types of r-jackknife estimators. It turns out
that the jackknife is intimately related to the divided differences of functions.
Definition 2 (Divided difference). The divided difference f [x1 , x2 , . . . , xn ] of a function f over n
distinct points {x1 , x2 , . . . , xn } is defined as
f [x1 , x2 , . . . , xn ] =
n
X
f (xi )
.
j6=i (xi − xj )
Q
i=1
(31)
2
It has been observed in the literature [SW89] that in jackknife variance estimation, which is a different area of
application of the jackknife methodology, sometimes it is also necessary to take d large to guarantee consistency.
3
We emphasize that if we restrict f ∈ C[0,
kf k∞ ≤ 1, and do not allow f to depend on n, then one cannot
P1],
r
achieve the bound (26). Indeed, noting that
C = 1, the error term can be written as
i=1 i
r
X
Ci Ep f (p̂n+i−r ) − f (p) =
i=1
r
X
Ci (Ep f (p̂n+i−r ) − f (p))
(27)
i=1
r
≤
X
|Ci |kEp f (p̂n+i−r ) − f (p)k∞ ,
(28)
i=1
It follows from the Bernstein theorem [AS04, Chap. 7] that limn→∞ kEp f (p̂n ) − f (p)k∞ = 0 for any continuous
function f on [0, 1]. Hence, for any f ∈ C[0, 1] one has
kEp fˆr − f (p)k∞ = o(nr−1 ),
since max1≤i≤r |Ci | . nr−1 for the delete-1 r-jackknife.
7
(29)
It follows from (18), (15) and Lemma 13 in Section B that the bias of a general r-jackknife
estimator fˆr can be written as
Ep [fˆr ] − f (p) =
=
r Y
X
ni
(Ep [f (p̂ni )] − f (p))
n − nj
i=1 j6=i i
(32)
r
X
nr−1
(Ep [f (p̂ni )] − f (p))
i
Q
.
(33)
j6=i (ni
i=1
− nj )
Define Gn,f,p = nr−1 (Ep [f (p̂n )] − f (p)). Then, the bias of fˆr can be written as the divided
difference of function G·,f,p :
Ep [fˆr ] − f (p) = G·,f,p [n1 , n2 , . . . , nr ].
(34)
It follows from the mean value theorem of divided differences defined over integers in Lemma 11
of Section B that for every p, f ,
|Ep [fˆr ] − f (p)| ≤
max |G·,f,p [n − r + 1, n − r + 2, . . . , n − 1, n]|,
n1 ≤n≤nr
(35)
and the right hand side of (35) is nothing but the maximum of the bias of the delete-1 r-jackknife
with varying sample sizes.
Equation (35) shows that in terms of the bias, the delete-1 jackknife might be the “worst” among
all r-jackknife estimators. However, what is the precise performance of the delete-1 r-jackknife when
the function f is “smooth”? Does the performance improve compared to Theorem 2? We answer
this question below.
Condition 2 (Condition Ds ). A function f : [0, 1] 7→ R is said to satisfy the condition Ds , s ≥
0, s ∈ Z with parameter L > 0 if the following is true:
1. s = 0: f is Lebesgue integrable on [0, 1] and supx∈[0,1] |f (x)| ≤ L.
2. s ≥ 1:
(a) f (s−1) is absolutely continuous on [0, 1];
(b) supx∈[0,1] |f (i) (x)| ≤ L, 0 ≤ i ≤ s.
Remark 2. We mention that if a function f satisfies condition Ds , it does not necessarily belong
to the space C s [0, 1], where C s [0, 1] denotes the space of continuously differentiable functions on
[0, 1]. Indeed, the function f (x) = x2 sin(1/x)1(x ∈ (0, 1]) satisfies condition D1 as a function
mapping from [0, 1] to R, but it does not belong to C 1 [0, 1].
The performance of the delete-1 r-jackknife in estimating f (p) satisfying condition Ds is summarized in the following theorem.
Theorem 3. For any r ≥ 1, s ≥ 0 and f satisfying condition Ds with parameter L, let fˆr be the
delete-1 r-jackknife. Then,
kEp fˆr − f (p)k .r,s,L
nr−s−1
1
n
−(r− 2 )
n−r
8
if 0 ≤ s ≤ 2r − 2;
if s = 2r − 1;
if s ≥ 2r.
(36)
Theorem 4. For 1 ≤ s ≤ 2r − 3, there exists some universal constant c > 0 such that for any
n ∈ N, there exists some function f ∈ C s [0, 1] such that kf k∞ ≤ 1, kf 0 k∞ ≤ 1, · · · , kf (s) k∞ ≤ 1,
and for delete-1 r-jackknife fˆr :
kEp fˆr (p̂n ) − f (p)k∞ ≥ cnr−1−s .
(37)
Theorem 5. For integer 2r − 2 ≤ s ≤ 2r − 1, there exists some function f ∈ C s [0, 1] such that
kf k∞ ≤ 1, kf 0 k∞ ≤ 1, · · · , kf (s) k∞ ≤ 1, and for delete-1 r-jackknife fˆr ,
lim inf
n→∞
kEp fˆr (p̂n ) − f (p)k∞
> 0.
n−s/2
(38)
Moreover, if s ≥ 2r, then
1
kEp fˆr (p̂n ) − f (p)k∞ & r .
n
(39)
Proof. The first part follows from (35) and Theorem 1. The second part follows from taking f (p)
to be a polynomial of order 2r with leading coefficient one.
Now we compare the performance of the r-jackknife estimator fˆr with and without Condition 1.
Under Condition 1, we know from Theorem 1 that
s
|Ep fˆr (p̂n ) − f (p)| .r,s,L,C n− min{r, 2 }
(40)
for f satisfying condition Ds with parameter L 4 , where the exponent is better than that of Theorem
3. A pictorial illustration is shown in Figure 1.
Remark 3. For general delete-d r-jackknife, the cases of d n and d = 1 exhibit drastically
different behavior. It remains fertile ground for research to analyze what is the minimum d needed
for the delete-d r-jackknife to achieve the bias performance that is of the same order as those
satisfying Condition 1 for a specific function f .
1.1.3
Specific functions
The last part of results pertaining to the jackknife investigates some specific functions f (p). Here we
take f (p) = −p ln p or pα , 0 < α < 1. Those functions even do not belong to D1 under Condition 2.
However, we show that the jackknife applied to these functions exhibits far better convergence rates
than the worst case analysis in Theorem 4 predicted.
We show that for the r-jackknife when r = 2, no matter whether Condition 1 is satisfied or not,
the bias of the jackknife estimator can be universally controlled.
Theorem 6. Let fˆ2 denote a general 2-jackknife in Definition 1. Then,
1. if f (p) = −p ln p,
kEp fˆ2 − f (p)k .
4
1
.
n
(41)
It follows from the proof of Theorem 1 that the first part of Theorem 1 also applies to functions f satisfying
condition Ds .
9
3
r=1,
r=2,
r=3,
r=4,
r=1,
r=2,
r=3,
r=4,
2
error exponent
1
Good
Good
Good
Good
Bad
Bad
Bad
Bad
0
-1
-2
-3
-4
0
1
2
3
4
5
6
7
8
9
10
smoothness of function
Figure 1: Error exponents of “Good" and “Bad" jackknife estimators. Here “Good” refers to the
r-jackknife satisfying Condition 1, and “Bad” refers to the delete-1 r-jackknife.
2. if f (p) = pα , 0 < α < 1,
1
kEp fˆ2 − f (p)k . α .
n
(42)
Meanwhile, let fˆ2 be either the delete-1 2-jackknife, or a 2-jackknife that satisfies Condition 1.
Then,
1. if f (p) = −p ln p,
1
kEp fˆ2 − f (p)k & .
n
(43)
1
kEp fˆ2 − f (p)k & α .
n
(44)
2. if f (p) = pα , 0 < α < 1,
Remark 4. We conjecture that Theorem 6 holds for any fixed r instead of only r = 2.
1.2
Bootstrap bias correction
The rationale behind bootstrap bias correction is to use the plug-in rule to estimate the bias and
then iterate the process [Hal92]. Concretely, suppose we would like to estimate a function f (θ), and
10
we have an estimator for θ, denoted as θ̂. The estimator θ̂(X1n ) is a function of the observations
i.i.d.
(X1 , X2 , . . . , Xn ), and Xi ∼ Pθ . The bias of the plug-in rule fˆ1 = f (θ̂) is defined as
e1 (θ) = f (θ) − Eθ f (θ̂).
(45)
We would like to correct this bias. The additive bootstrap bias correction does this by using
the plug-in rule e1 (θ̂) to estimate e1 (θ), and then use f (θ̂) + e1 (θ̂) to estimate f (θ), hoping that
this bias corrected estimator has a smaller bias.
It is the place that the Monte Carlo approximation principle takes effect: it allows us to compute
the plug-in estimator e1 (θ̂) without knowing the concrete form of the bias function e1 (θ). Indeed,
we have
e1 (θ̂) = f (θ̂) − Eθ̂ f (θ̂∗ ),
(46)
i.i.d.
where θ̂∗ = θ̂(X1∗ , X2∗ , . . . , Xn∗ ), and the samples Xi∗ ∼ Pθ̂ . To compute Eθ̂ f (θ̂∗ ), it suffices to
draw the n-tuple sample (X1∗ , X2∗ , . . . , Xn∗ ) in total B times under Pθ̂ , and use the empirical average
to replace the expectation, hoping that the law of large number would make the empirical average
close to the expectation Eθ̂ f (θ̂∗ ). This argument also shows that it takes B rounds of sampling to
evaluate e1 (·) at one point.
After doing bootstrap bias correction as introduced above once, we obtain fˆ2 = f (θ̂) + e1 (θ̂).
What about its bias? The bias of this new estimator, denoted as e2 (θ), is
e2 (θ) = f (θ) − Eθ f (θ̂) + Eθ e1 (θ̂)
= e1 (θ) − Eθ e1 (θ̂).
(47)
(48)
Clearly, in order to compute e2 (θ̂), we need to evaluate e1 (·) in total B times, which amounts
to a total computation complexity B 2 .
It motivates the general formula: the bias of the bootstrap bias corrected estimator after m − 1
rounds of correction is related to that after m − 2 rounds via
em (θ) = em−1 (θ) − Eθ em−1 (θ̂).
(49)
Indeed, denoting the estimator after m − 2 rounds of bias correction as fˆm−1 , by definition we know
em−1 (θ) = f (θ)−Eθ fˆm−1 . The bias corrected estimator after m−1 rounds is fˆm = fˆm−1 +em−1 (θ̂),
whose bias is
em (θ) = f (θ) − Eθ fˆm−1 + em−1 (θ̂)
= em−1 (θ) − Eθ em−1 (θ̂).
(50)
(51)
The bias corrected estimator after m − 1 rounds of correction is
fˆm = f (θ̂) +
m−1
X
ei (θ̂).
(52)
i=1
It takes B m−1 order computations to compute fˆm if we view the computation of fˆ2 = f (θ̂)+e1 (θ̂)
takes computational time B. We introduce an linear operator An that maps the function f to the
same function space such that
An [f ](θ) = Eθ f (θ̂).
11
(53)
With the help of the operator An , one may view the bias of fˆm in the following succinct way.
Indeed, since
em (θ) = em−1 (θ) − An [em−1 ](θ)
= (I − An )[em−1 ](θ),
(54)
(55)
we have
m−1
X
Eθ fˆm = An I +
!
i
(I − An )
[f ]
(56)
i=1
= An
m−1
X
!
i
(I − An )
[f ]
(57)
i=0
= (I − (I − An )m )[f ].
(58)
The operator I − (I − An )m is known as the iterated Boolean sum in the approximation theory
literature [Nat83, Sev91]. Indeed, defining the Boolean sum as P ⊕ Q = P + Q − P Q, we have
I − (I − An )m = An ⊕ An ⊕ . . . ⊕ An = ⊕m An ,
(59)
where there are m terms on the right hand side.
Let the bias of the bootstrap bias corrected estimator after m − 1 rounds be denoted as em (p),
where
em (p) = em−1 (p) − Ep em−1 (p̂n ),
(60)
and e1 (p) = f (p) − Ep [f (p̂n )]. Here f ∈ C[0, 1], and n · p̂n ∼ B(n, p). Our first result on bootstrap
bias correction is about the limiting behavior of em (p) as m → ∞. In other words, what happens
when we conduct the bootstrap bias correction infinitely many times?
Theorem 7. Denote the unique polynomial of order n that interpolates the function f (p) at n + 1
points {i/n : 0 ≤ i ≤ n} by Ln [f ]. Then, for any f : [0, 1] 7→ R,
lim
sup |em (p) − (f − Ln [f ])| = 0,
m→∞ p∈[0,1]
(61)
where em (p) is defined in (60).
In other words, the bias function converges uniformly to the approximation error of the Lagrange
interpolation polynomial that interpolates the function f at equidistant points on [0, 1]. This
interpolating polynomial is in general known to exhibit bad approximation properties unless the
function is very smooth. The Bernstein example below shows an extreme case.
Lemma 2. [Nat64, Chap. 2, Sec. 2] [Bernstein’s example] Suppose f (p) = |p − 1/2|, and Ln [f ]
denotes the unique polynomial that interpolates the function f (p) at n + 1 points {i/n : 0 ≤ i ≤ n}.
Then,
lim inf |Ln [f ](p)| = ∞
n→∞
for all p ∈ [0, 1] except for p = 0, 1/2, and 1.
(62)
5
This phenomenon has been generalized to other functions such as |p − 21 |α , α > 0 when α is not an even integer.
See [Gan03] for more details.
5
12
For more discussions on the convergence/divergence behavior of Ln [f ], we refer the readers
to [MM08] for more details. We emphasize that it is a highly challenging question. For example, it
was shown in [LS94] that for any p ∈ [0, 1], we have
lim |Ln [f ](p) − f (p)| = 0,
n→∞
(63)
where f (p) = −p ln p or pα , α > 0, α ∈
/ Z, and Ln [f ] is the Lagrange interpolation polynomial at
equi-distant points. However, to our knowledge it is unknown that whether supp∈[0,1] |Ln [f ](p) −
f (p)| converges to zero as n → ∞ for those specific functions, and if so, what the convergence rate
is.
As Theorem 7 and Lemma 2 show, it may not be a wise idea to iterate the bootstrap bias
correction too many times. It is both computationally prohibitive, and even may deteriorate
statistically along the process. In practice, one usually conducts the bootstrap bias correction
a few times. The next theorem provides performance guarantees for the first few iterations of
bootstrap bias correction.
Theorem 8. Fix the number of iterations m ≥ 0, and 0 < α ≤ 2m. Then the following statements
are true for any f ∈ C[0, 1]. Here em (p) is defined in (60).
1.
√
kem (p)k .m ωϕ2m (f, 1/ n) + kf kn−m .
(64)
kem (p)k .α,m n−α/2 if and only if ωϕ2m (f, t) .α,m tα .
(65)
kem (p)k m n−m if and only if f is an affine function
(66)
2.
3.
4. Suppose there is a constant D < 22m for which
ωϕ2m (f, 2t) ≤ Dωϕ2m (f, t) for t ≤ t0 .
(67)
√
kem (p)k m,D ωϕ2m (f, 1/ n).
(68)
√
kem (p)k ωϕ2m (f, 1/ n).
(69)
Then,
5. For m = 1,
Theorem 8 has several interesting implications. First of all, it shows that for a few iterations
of the bootstrap bias correction, we have a decent bound on the bias kem (p)k, which is intimately
√
connected with the 2m-th order Ditzian–Totik modulus of smoothness evaluated at 1/ n. This
bound is tight in various senses. The second statement shows that it captures the bias kem (p)k at
least up to the granularity of the exponent in n, and the third statement shows that it is impossible
13
for the bootstrap bias corrected estimator to achieve bias of order lower than n−m except for the
trivial case of affine functions, which have bias zero. The fourth statement shows that as long as
the modulus ωϕ2m (f, t) is not too close to t2m , the DT modulus bound is tight. The fifth statement
shows that when we do not do any bias correction, the DT modulus bound is tight for any function
in C[0, 1].
The following corollary is immediate given (12).
Corollary 2. If f (p) = −p ln p, then,
kem (p)k m
1
m ke1 (p)k,
n
(70)
If f (p) = pα , 0 < α < 1, then
1
m ke1 (p)k,
(71)
nα
which means that the bootstrap bias correction for the first few rounds does not change the order
of bias at all.
We have shown that the bias of fˆm converges to the approximation error of the Lagrange
kem (p)k m
interpolation polynomial at equi-distant points when m → ∞ (Theorem 7). However, we also know
that for the first few iterations of the bootstrap, the bias of fˆm can be well controlled (Theorem 8).
It begs the question: how does kem (p)k evolve as m → ∞?
We study this problem through the example of f (p) = |p − 1/2|, with the sample size n = 20.
Thanks to the special structure of the Binomial functions, we are able to numerically compute
kem (p)k up to m ≈ 8.5 × 105 . It follows from Theorem 7 that
lim kem (p)k = kf − Ln [f ]k,
m→∞
and for n = 20, we numerically evaluated kf − Ln [f ]k to be 47.5945.
14
(72)
Figure 2: The evolution of kem (p)k as a function of m for 1 ≤ m ≤ 8.5 × 105 .
Figure 3: The evolution of kem (p)k as a function of m for 1 ≤ m ≤ 2 × 104 .
15
Figure 4: The Lagrange interpolation points and the Lagrange interpolation polynomial of the
function |p − 1/2| with equi-distant n + 1 points, where n = 20.
Figure 2, 3, and 4 show that the behavior of kem (p)k could be highly irregular: in fact, for the
specific function f (p) = |p − 1/2|, it continues to decrease until m grows slightly above 2 × 103 , and
then keeps on increasing until m exceeds about 1.2 × 104 , then it continues to drop until it hits
about 3 × 104 , then it keeps on increasing again within the range of computations we conduct. It is
also clear that after about 8.5 × 105 bootstrap iterations, which is by no means practical, kem (p)k
is still far from its limit kf − Ln [f ]k, which is about 47.5945 as shown in Figure 4.
Remark 5 (Connections between bootstrap and jackknife). The interested reader must have observed that the bias properties of the bootstrap bias corrected estimator after r − 1 rounds is
exactly the same as that of the r-jackknife estimator satisfying Condition 1. Concretely, their bi√
ases are both dictated by the modulus ωϕ2r (f, 1/ n). It would be interesting to compare the rate
√
ωϕ2r (f, 1/ n) with that of the best polynomial approximation, upon noting that in the binomial
model, the biases of both the jackknife and bootstrap estimators are polynomial approximation
errors of the function f (p) with degree at most n. It follows [DT87, Thm. 7.2.1.] that for best
polynomial approximation with degree n, the approximation error inf P ∈polyn supp∈[0,1] |f (p) − P (p)|
is upper bounded by ωϕk (f, 1/n) for any k < n. We first observe that one achieves a smaller argu√
ment (1/n compared to 1/ n) in this case, but more importantly, there is essentially no restriction
on the modulus order when n is large. It indicates the best polynomial approximation induces a
much better approximation (smaller bias) for estimating f (p), which, unfortunately has been shown
in [Pan03] to fail to achieve the minimax rates in entropy estimation, since the variance explodes
while the bias is very small. The estimators in [VV11, WY16, JVHW17] only choose to conduct
16
best polynomial approximation in certain regimes of f , which reduces the bias by a logarithmic
factor without increasing too much the variance.
1.3
Taylor series bias correction
The Taylor series can only be applied to functions with certain global differentiablity conditions,
which makes it a less versatile method compared to the bootstrap and jackknife. The Taylor series
bias correction method exhibits various forms in the literature, and we discuss two of them in
this section. We call one approach the iterative first order correction, and the other approach
the sample splitting correction. To illustrate the main ideas behind the methods, we still use the
binomial model n · p̂n ∼ B(n, p).
1.3.1
Iterative first order correction
As shown in [LC98, Chapter 6, Section 1, Pg. 436], suppose for certain f , we have
1
Bn (p)
,
Ep f (p̂n ) − f (p) =
+O
n
n2
(73)
where Bn (p) = 12 f 00 (p)nEp (p̂n − p)2 . Then, the Taylor series bias corrected estimator is defined as
Bn (p̂n )
fˆ2 = f (p̂n ) −
.
n
(74)
We can generalize the approach above to conduct bias correction for multiple rounds [Wit87].
However, the correction formula becomes increasingly more complicated as the correction order
becomes higher. We start with the following lemma.
Lemma 3. Suppose function f : [0, 1] 7→ R satisfies condition Ds with parameter L as in Condition 2, where s = 2k is a positive even integer. Then, if n · p̂n ∼ B(n, p), there exist k − 1 linear
operators denoted as Tj [f ](p), 1 ≤ j ≤ k − 1, independent of n, such that
Ep f (p̂n ) − f (p) −
k−1
X
j=1
1
1
Tj [f ](p) .k,L k .
j
n
n
(75)
Here supp∈[0,1] |Tj [f ](p)| .k,L 1. Concretely, Tj [f ](p) is a linear combination of the derivatives of f
of order from j + 1 to 2j where the combination coefficients are polynomials of p with degree no
more than 2j.
Now we describe the Taylor series bias correction algorithm below [Wit87].
Construction 1 (Taylor series bias correction). [Wit87] Define ti (p) iteratively. Set t0 (p) = f (p),
and for i ≥ 1 define
ti (p) = −
i
X
Tj [ti−j ](p).
(76)
j=1
The final bias corrected estimator is
fˆk =
k−1
X
i=0
1
ti (p̂n ).
ni
17
(77)
Construction 1 may be intuitively understood as the iterative generalization of the order one
Taylor series bias correction (74). Indeed, after we conduct the first order bias correction and
obtain
T1 [f ](p̂)
fˆ2 = f (p̂n ) −
n
t1 (p̂n )
= t0 (p̂n ) +
,
n
we apply Lemma 3 to the function t0 +
t0 (p) +
t1
n
and obtain the expansion up to order
(78)
(79)
1
n2
as
t1 (p) T1 [t0 ](p) T2 [t0 ](p) T1 [t1 ](p)
T2 [t0 ](p) T1 [t1 ](p)
+
+
+
= t0 (p) +
+
,
2
2
n
n
n
n
n2
n2
(80)
where we used the definition of t1 = −T1 [t0 ]. It naturally leads to the further correction
t2 (p) = −T2 [t0 ](p) − T1 [t1 ](p).
(81)
One can repeat this process to obtain the formula in Construction 1.
Now we prove that the estimator fˆk in Construction 1 achieves bias of order O(n−k ) if the
original function f satisfies condition Ds with s = 2k. It can be viewed as one concrete example
of [Wit87].
Theorem 9. Suppose f : [0, 1] 7→ R satisfies condition Ds with parameter L, and s = 2k, k ≥
1, k ∈ Z. Then, the estimator in Construction 1 satisfies
1
kEp [fˆk ] − f (p)k .k,L k .
n
1.3.2
(82)
Sample splitting correction
This method was proposed in [HJW16]. It aims at solving one disadvantage of Construction 1,
which is that the bias correction formula for higher orders may not be easy to manipulate since it is
defined through a recursive formula. The sampling splitting correction method provides an explicit
bias correction formula which is easy to analyze with transparent proofs, but the disadvantage it
has is that it only applies to certain statistical models.
The intuition of the sample splitting correction method is the following, which is taken from [HJW16].
Suppose f satisfies condition D2k with parameter L. Instead of doing Taylor expansion of f (p̂n )
near p, we employ Taylor expansion of f (p) near p̂n :
f (p) ≈
2k−1
X
i=0
f (i) (p̂n )
(p − p̂n )i .
i!
(83)
Now, f (i) (p̂n ) is by definition an unbiased estimator for Ep [f (i) (p̂n )]. However, the unknown
p in the right hand side still prevents us from using this estimator explicitly. Fortunately, this
difficulty can be overcome by the standard sample splitting approach: we split samples to obtain
(1)
(2)
independent p̂n and p̂n , both of which follow the same class of distribution (with possibly different
18
parameters) as p̂n . We remark that sample splitting can be employed for divisible distributions,
including multinomial, Poisson and Gaussian models [Nem00]. Now our bias-corrected estimator is
fˆk =
2k−1
X
i=0
(1)
i
f (i) (p̂n ) X
i
(1) i−j
Sj (p̂(2)
n )(−p̂n )
i!
j
j=0
!
(84)
(2)
where Sj (p̂n ) is an unbiased estimator of pj (which usually exists when sample splitting is doable).
Now it is straightforward to show that
E[fˆk ] − f (p) = Ep
"2k−1
X f (i) (p̂(1)
n )
i=0
kf (2k) k
. Ep
.k,L
(2k)!
#
(p −
i!
i
p̂(1)
n )
− f (p)
(85)
2k
(p̂(1)
n − p)
(86)
1
,
nk
(87)
where in the last step we used the property of the binomial distribution in Lemma 15.
The rest of the paper is organized as follows. Section 2 discusses the key proof ingredients of
the results pertaining to jackknife bias correction. The proofs of main results on bootstrap bias
correction are provided in Section 3. Section A reviews the K-functional approach for bias analysis.
Section B collects auxiliary lemmas used throughout this paper. Proofs of the rest of the theorems
and lemmas in the main text are provided in Section C, and the proofs of the auxiliary lemmas are
presented in Section B.
2
2.1
Jackknife bias correction
Theorem 3
We first present the proof of Theorem 3. We explain the roadmap below, and the key lemmas used
in roadmap are proved in Section C.
The first step to analyze the general r-jackknife for functions f satisfying the Condition Ds in
Condition 2 is to use Taylor expansions. It is reflected in Lemma 4.
Lemma 4. Suppose f satisfies Condition 2 for fixed s ≥ 1 with parameter L. Then, for the general
r-jackknife estimator with fixed r ≥ 1 in Definition 1,
|Ep fˆr − f (p)| .r,s,L n−r +
+
Z 1 X
r
p
Z p X
r
0
Ci Ep (p̂ni − t)s−1
dt
+
i=1
Ci Ep (p̂ni − t)s−1
dt.
−
(88)
i=1
Here the coefficients {Ci }1≤i≤r are given in Definition 1.
Lemma 4 shows that it suffices to analyze the behavior of the quantities ri=1 Ci Ep (p̂ni − t)s−1
+
P
s−1
and ri=1 Ci Ep (p̂ni − t)−
. These quantities can be viewed as divided differences (see (34)), and
to analyze the worst case we analyze the following backward difference sequences.
P
19
For t ≥ p, u ≥ 0, s ≥ 0, define
An,u (t) = Ep (p̂n − t)u+
(89)
and consider its s-backward difference defined as
s
∆ An,u (t) ,
s
X
!
k
(−1)
k=0
s
An−k,u (t).
k
(90)
We have
Lemma 5. For s, u ≥ 0, t ≥ p and n ≥ 2s, we have
|∆s An,u (t)| ≤ c1 ·
−(u+s−1) p + n−u pu∧1 ( √1 ∧ 1) · exp(−c nt)
n
2
nt
−( u +s) u
c2 n(t−p)2
−u u∧1 1
n
2
p2 + n
p
∧ 1) · exp(−
(√
nt
u
u
n−( 2 +s) (1 − p) 2 + n−u (1 − p)u∧1 ( √ 1
n(1−t)
if p ≤ n1 ,
if
1
n
< p ≤ 21 ,
∧ 1) · exp(− c2 n(t−p)
) if
1−p
1
2
< p ≤ 1 − n1 .
t
)
2
(91)
where the universal constants c1 , c2 > 0 only depend on u, s (not on n or p). Moreover, if t > 1 − n1 ,
we have
|∆s An,u (t)| ≤ c1 (1 − t)u (1 − p)s pn−s .
(92)
Note that in Lemma 5, the case where p > 1− n1 has already been included in the case t > 1− n1 ,
for t ≥ p. Hence, Lemma 5 has completely characterized an upper bound on the dependence of
|∆s An,u (t)| on all n, p and t. By symmetry, we have the following corollary regarding
u
A−
n,u (t) , Ep (p̂n − t)− .
(93)
Corollary 3. For s, u ≥ 0, t ≤ p and n ≥ 2s, we have
|∆s A−
n,u (t)| ≤ c1 ·
1
−(u+s−1)
−u
u∧1
n
(1 − p) + n (1 − p) ( √
∧ 1) · exp(−c2 n(1 − t))
n(1−t)
u
u
n−( 2 +s) (1 − p) 2 + n−u (1 − p)u∧1 ( √ 1
∧ 1) · exp(−
n(1−t)
u
u
c2 n(t−p)2
−u u∧1 √1
−( 2 +s) 2
n
p +n
p
(
nt
∧ 1) · exp(−
p
)
c2 n(t−p)2
)
1−t
if p ≥ 1 − n1 ,
if
1
2
≤ p ≤ 1 − n1 ,
if
1
n
≤ p < 12 .
(94)
where the universal constants c1 , c2 > 0 only depend on u, s (not on n or p). Moreover, if t <
we have
|∆s An,u (t)| ≤ c1 tu ps (1 − p)n−s .
1
n,
(95)
Furthermore, in most cases we do not need the dependence on p, and Lemma 5 implies the
following corollary.
20
Corollary 4. For s, u ≥ 0, t ≥ p ≥ t0 and n ≥ 2s, we have
s
−( u
+s)
2
exp(−c2 n(t − p) ) + n
0
A−
n,u (t )|
−( u
+s)
2
0
|∆ An,u (t)| ≤ c1 n
s
|∆
≤ c1 n
2
!
c2 n(t − p)2
1
exp(−
·√
)
t
t + n−1
−(u+ 21 )
2
exp(−c2 n(t − p) ) + n
−(u+ 21 )
(96)
!
1
c2 n(t0 − p)2
·√
exp(−
)
1−t
1 − t0 + n−1
(97)
where the universal constants c1 , c2 > 0 only depend on u, s (not on n or p).
Now we can start the proof of Theorem 3.
Proof of Theorem 3. We split into three cases. When s = 0, we use the triangle inequality to
conclude that
|Ep fˆr (p̂n ) − f (p)| =
r
X
Ci Ep f (p̂ni ) − f (p)
(98)
|Ci | + 1)kf k∞
(99)
i=1
r
X
≤(
i=1
r−1
.L n
.
(100)
For 1 ≤ s ≤ 2r, it follows from Lemma 4 that
|Ep fˆr (p̂n ) − f (p)| .r,s,L n−r +
+
Z 1 X
r
Z p X
r
0
p
Ci Ep (p̂ni − t)s−1
dt
+
i=1
Ci Ep (p̂ni − t)s−1
dt.
−
(101)
i=1
For t ≥ p, it follows from Corollary 4 that there exist universal constants c1 , c2 depending on
r, s only such that
c2 n(t−p)2
1
2
t
e−
+ e−c2 n(t−p) )
−1
t+n
|∆u An,s−1 (t)| ≤ c1 ( √
·
n−(
s−1
+u)
2
n−(s− 12 )
if 0 ≤ u ≤ b 2s c,
(102)
if u ≥ d 2s e.
Now define
Bn,r,s (t) = nr−1 An,s (t).
(103)
By the product rule of backward difference we obtain
X
|∆r−1 Bn,r,s (t)| .r,s
|∆i nr−1 | · |∆j An,s (t)|
(104)
0≤i,j≤r−1,i+j≥r−1
X
.r,s
nr−1−i · ( √
0≤i,j≤r−1,i+j≥r−1
c2 n(t−p)2
s−1
1
1
2
t
+ e−c2 n(t−p) )n− min{ 2 +j,s− 2 }
e−
t + n−1
(105)
.r,s ( √
1
e
t + n−1
c n(t−p)2
− 2 t
2
+ e−c2 n(t−p) )n− min{
21
s−1
,s−r+ 21 }
2
.
(106)
As a result of Lemma 11,
r
X
r
X
s−1
=
Ci Ep (p̂ni − t)+
Ci Ani ,s−1 (t)
(107)
i=1
i=1
= |B·,r,s [n1 , · · · , nr ](t)|
1
≤
max |∆r Bm,r,s (t)|
(r − 1)! m∈[n1 ,nr ]
c2 n(t−p)2
s−1
1
1
2
t
e−
.r,s ( √
+ e−c2 n(t−p) )n− min{ 2 ,s−r+ 2 } .
−1
t+n
(108)
(109)
(110)
Using this inequality, finally we arrive at
Z 1 X
r
p
s−1
dt
Ci Ep (p̂ni − t)+
i=1
Z 1
c2 n(t−p)2
s−1
1
1
2
t
e−
+ e−c2 n(t−p) )n− min{ 2 ,s−r+ 2 } dt
−1
t+n
p
!
Z ∞
Z ∞
c nu2
1
1
2
− 2u+p
− min{ s−1
−c
nu
,s−r+
}
2
2
p
e 2 du
du +
e
≤n
u + p + n−1
0
0
.r,s
≤n
(√
− min{ s−1
,s−r+ 12 }
2
Z p
1
0
≤n
− min{ s−1
,s−r+ 12 }
2
1
√
p
√ e
p
Z ∞
−
c2 nu2
2p
du +
Z ∞
√
ne
−
c2 nu
2
Z ∞
−
e
c2 nu2
2p
(112)
!
du
(113)
0
√ Z
du + n
∞
e
−
c2 nu
2
Z ∞
−c2 nu2
e
du +
0
0
e
du +
p
−c2 nu2
(111)
!
du
(114)
0
s
.r,s n− min{ 2 ,s−r+1}
(115)
as desired. The remaining part can be dealt with analogously.
When s ≥ 2r, the desired result follows from applying Lemma 4 with s = 2r.
2.2
Theorem 4
We consider the case where s > 0 and s ≤ 2r − 3. To come up with an example which matches the
upper bound in Theorem 3, we first need to prove a “converse" of Lemma 5. Recall that for t ≥ p,
An,u (t) = Ep (p̂n − t)u+ .
(116)
1
}
Lemma 6. For any 0 ≤ u ≤ 2(s−1), there exists some p0 > 0 such that for any 0 < p < min{p0 , 4s
1
1
and any n ≥ p2 , whenever t ∈ [p, p + √n ] satisfies
1.
k
n−s+1
2.
k
n−s
<t<
<t<
k−p
n−s
k+p
n−s
for some k ∈ N if u < s;
for some k ∈ N if u ≥ s,
we have
1
|∆s An,u (t)| ≥ cn−(u+ 2 )
where c > 0 is a universal constant which only depends on u, s and p.
22
(117)
Now we start the proof of Theorem 4. The basic idea of the proof is to construct functions f
such that Lemma 4 is nearly tight.
Proof of Theorem 4. Pick an arbitrary p > 0 which satisfies Lemma 6, and we define
g(t) = sign
r
X
!
t)s−1
+
Ci Ep (p̂ni −
· 1(t ≥ p).
(118)
i=1
Note that g is not continuous, but we can find some h ∈ C[0, 1] such that kg − hk1 ≤ n−2r . Now
choose any f with f (s) = h , we know that f ∈ C s [0, 1], and the norm conditions are satisfied under
proper scaling.
It follows from Lemma 4 that
|Ep fˆr (p̂n ) − f (p)| O(n
−r
r
X
Z 1
h(t)
)+
p
r
X
h(t)
0
!
Ci Ep (p̂ni −
r
X
Z 1
g(t)
0
= O(n−r ) +
Z 1 X
r
= o(nr−1−s ) +
dt
(119)
r−1−s
!
Ci Ep (p̂ni − t)s−1
dt + O(kg − hk1 ) ·
+
r
X
|Ci |
(120)
i=1
Ci Ep (p̂ni − t)s−1
dt
+
(121)
i=1
Z 1 X
r
p
Z
)+
i=1
r
X
G i=1
√1 ]
n
t)s−1
−
i=1
p
where G ⊂ [p, p +
to see
dt
i=1
= O(n−r ) +
≥ o(n
Ci Ep (p̂ni −
i=1
Z p
+
!
t)s−1
+
Ci Ani ,s−1 (t) dt
(122)
Ci Ani ,s−1 (t) dt
(123)
is the set of all “good" t’s which satisfy the condition of Lemma 6. It’s easy
1
m(G) n− 2
(124)
where m(·) denotes the Lebesgue measure. Moreover, by our choice of delete-1 jackknife, for t ∈ G
we have
r
X
Ci Ani ,s−1 (t) = ∆r−1 (nr−1 An,s−1 (t))
(125)
i=1
& nr−1 |∆r−1 An,s−1 (t)|
X
−
|∆i nr−1 | · |∆j An,s−1 (t)|
(126)
1≤i≤r−1,0≤j≤r−1,i+j≥r−1
1
X
& nr−1−(s− 2 ) −
nr−1−i−min{
s−1
+j,s− 12 }
2
(127)
1≤i≤r−1,0≤j≤r−1,i+j≥r−1
1
& nr−s− 2
(128)
23
where we have used Lemma 5, Lemma 6 and the assumption that 0 ≤ s − 1 ≤ 2(r − 2). As a result,
we conclude that
|Ep fˆr (p̂n ) − f (p)| & o(nr−1−s ) +
Z
r
X
G i=1
Ci Ani ,s−1 (t) dt
1
& o(nr−1−s ) + m(G) · nr−s− 2
&n
r−s−1
(129)
(130)
(131)
as desired.
2.3
Theorem 6
The following lemma characterizes the difference Ep f (p̂n ) − Ep f (p̂n−1 ) for certain functions.
Lemma 7. [ST77] Suppose f (p) = −p ln p, p ∈ [0, 1]. Then,
1 − pn − (1 − p)n
.
n(n − 1)
(132)
(1 − α)(1 − pn − (1 − p)n )
.
n(n − 1)α
(133)
0 ≤ Ep f (p̂n ) − Ep f (p̂n−1 ) ≤
Suppose f (p) = pα , p ∈ [0, 1], 0 < α < 1. Then,
0 ≤ Ep f (p̂n ) − Ep f (p̂n−1 ) ≤
The next lemma characterizes the lower bound for the bias of the jackknife estimate fˆ2 .
Lemma 8. Suppose fˆ2 is the delete-1 2-jackknife. Then,
1. for f (p) = −p ln p,
1
kEp fˆ2 − f (p)k & .
n
(134)
1
kEp fˆ2 − f (p)k & α .
n
(135)
2. for f (p) = pα , 0 < α < 1,
Suppose fˆ2 is a 2-jackknife that satisfies Condition 1. Then,
1. for f (p) = −p ln p,
1
kEp fˆ2 − f (p)k & .
n
(136)
1
.
nα
(137)
2. for f (p) = pα , 0 < α < 1,
kEp fˆ2 − f (p)k &
Now we can start the proof of Theorem 6.
24
Proof of Theorem 6. The lower bounds follow from Lemma 8. Now we prove the upper bounds.
For a general 2-jackknife and general function f , we have
Ep [fˆ2 ] − f (p) =
n2
n1
(Ep [f (p̂n1 )] − f (p)) +
(Ep [f (p̂n2 )] − f (p)) .
n1 − n2
n2 − n1
(138)
Define Hn = Ep [f (p̂n )] − f (p). Then,
n2 Hn2 − n1 Hn1
Ep [fˆ2 ] − f (p) =
n2 − n1
n1
(Hn2 − Hn1 ) .
= Hn2 +
n2 − n1
(139)
(140)
For any f ∈ C[0, 1], we have limn→∞ Hn = 0, which implies
Hn2 = Hn2 − H∞
=
∞
X
(141)
(Hj − Hj+1 )
(142)
j=n2
and
Hn2 − Hn1 =
n2
X
(Hj − Hj−1 )
(143)
j=n1 +1
It follows from Lemma 8 that for f (p) = −p ln p, 0 ≤ Hj − Hj−1 .
kEp [fˆ2 ] − f (p)k .
∞
X
1
j=n2
1
.
j2
n
Hence,
2
X
n1
1
+
·
2
j
n2 − n1
j2
j=n +1
(144)
1
1
n1
.
+
n2 n2 − n1
1
.
n2
1
. ,
n
1
1
−
n1 n2
(145)
(146)
(147)
where in the last step we used Definition 1.
The case of f (p) = pα , 0 < α < 1 can be proved analogously.
3
3.1
Bootstrap bias correction
Theorem 7
Define the Bernstein operator Bn : C[0, 1] 7→ C[0, 1] as
Bn [f ](p) =
n
X
i=0
i
n
f
25
!
n i
p (1 − p)n−i .
i
(148)
Theorem 7 can be proved using the eigenstructure of the Bernstein operator [Sev95]. We give
a concrete proof as follows.
It was shown in [CW00] that the Bernstein operator Bn [f ] admits a clean eigenstructure. Con(n)
cretely, it has n + 1 linearly independent eigenfunctions pk , 0 ≤ k ≤ n, which are polynomials
(n)
n!
with order k, with k simple zeros on [0, 1]. The corresponding eigenvalues are λk = n1k (n−k)!
.
The Bernstein operator is also degree reducing in the sense that it maps a k-degree polynomial to
another polynomial with degree no more than k.
Decomposing f (p) as f (p) = Ln [f ](p) + g(p). It follows the definition of Ln [f ] that g(i/n) =
0, 0 ≤ i ≤ n. Hence, Bn [g] ≡ 0.
Since Ln [f ](p) is a polynomial with degree no more than n, it admits a unique expansion
Ln [f ](p) =
n
X
(n)
ak p k .
(149)
k=0
Applying I − Bn on the decomposition of f , we have
(I − Bn )[f ] = f − Bn [f ]
n
X
=
=
k=0
n
X
(150)
(n)
ak pk + g −
n
X
(n) (n)
ak λk pk
(151)
k=0
(n)
(n)
ak (1 − λk )pk + g
(152)
(n)
(153)
k=0
It follows by induction that
em (p) =
n
X
(n)
ak (1 − λk )m pk + g
k=0
We have
sup |em (p) − g| ≤
p
n
X
(n)
smallest
(n)
λk ,
= 1· 1−
(n)
1
n
... 1 −
k−1
n
(154)
p
k=0
→0
Note that λk
(n)
(1 − λk )m |ak | sup |pk |
as m → ∞.
(155)
(n)
, k ≥ 2, and λk
= 1 when k = 0, 1. Hence, the
which is the slowest to vanish corresponds to k = n. Since
√
n!
2πn
(n)
,
λn = n ≈
n
en
n
(156)
(n)
e
rounds of iteration to make (1 − λn )m vanish, which is a prohibitively large
it takes m & √
n
number in practice.
4
Acknowledgment
We are grateful to Yihong Wu for pointing out the the connection of Theorem 10 with f -divergence
inequalities [SV16].
26
A
The K-functional approach to bias analysis
We introduce the r-th modulus of smoothness of a function f : [0, 1] 7→ R as
ω r (f, t) = sup k∆rh f k,
(157)
0<h≤t
where ∆rh f is defined in (10). The r-th Ditzian–Totik modulus of smoothness of a function f :
[0, 1] 7→ R is defined in (11).
Intuitively, the smoother the function is, the smaller is its moduli of smoothness. For f ∈ C[0, 1],
we define the K-functional Kr (f, tr ) as follows:
Kr (f, tr ) = inf {kf − gk + tr kg (r) k : g (r−1) ∈ A.C.loc },
g
(158)
and the K-functional Kr,ϕ (f, tr ) as
Kr,ϕ (f, tr ) = inf {kf − gk + tr kϕr g (r) k : g (r−1) ∈ A.C.loc },
g
(159)
where g (r−1) ∈ A.C.loc means that g is r − 1 times differentiable and g (r−1) is absolutely continuous
in every closed finite interval [c, d] ⊂ (0, 1).
The remarkable fact is, the K-functionals are equivalent to the corresponding moduli of smoothness for any function f ∈ C[0, 1]. Concretely, we have the following lemma:
Lemma 9. [DL93, Chap. 6, Thm. 2.4, Thm. 6.2] [DT87, Thm. 2.1.1.] There exist constants
c1 > 0, c2 > 0 which depend only on r such that for all f ∈ C[0, 1],
c1 ω r (f, t) ≤ Kr (f, tr ) ≤ c2 ω r (f, t)
c1 ωϕr (f, t)
r
≤ Kr,ϕ (f, t ) ≤
c2 ωϕr (f, t)
for all t > 0
for all t ≤ t0 ,
(160)
(161)
where t0 > 0 is a constant that only depends on r.
We emphasize that the K-functionals and moduli of smoothness introduced here are tailored for
the interval [0, 1] and the supremum norm, which can be generalized to general finite intervals and
infinite intervals, and Lp norms. The corresponding equivalence results also hold in those settings.
We refer the interested readers to [DL93, Chap. 6] and [DT87, Chap. 2] for details. For other
use of K-functionals in statistics and machine learning, we refer the readers to the theory of Besov
spaces as interpolation spaces [HKPT12] and distribution testing [BCG16].
Now we illustrate the K-functional approach to bias analysis, which is well known in the approximation theory literature, see, e.g. [Tot88]. Suppose X is a random variable taking values in
[0, 1], and we would like to bound the quantity |E[f (X)] − f (E[X])| for any f ∈ C[0, 1]. Clearly,
f may not be differentiable, so we introduce another function g ∈ C[0, 1] such that g (1) ∈ A.C.loc .
We proceed as follows:
|E[f (X)] − f (E[X])| = |E [f (X) − g(X) + g(X) − g(EX) + g(EX) − f (EX)] |
≤ 2kf − gk + |E[g(X) − g(EX)]|
1
≤ 2kf − gk + kg 00 kVar(X)
2
= 2 kf − gk + t2 kg 00 k ,
27
(162)
(163)
(164)
(165)
where t2 =
Var(X)
.
4
Since g (1) ∈ A.C.loc is arbitrary, we know
n
|E[f (X)] − f (E[X])| ≤ 2 inf kf − gk + t2 kg 00 k
o
(166)
g
= 2K2 (f, t2 )
Var(X)
= 2K2 f,
4
!
p
Var(X)
≤ 2c2 ω 2 f,
,
2
(167)
(168)
(169)
where the constant c2 is introduced in Lemma 9.
It was shown in [DL93, Chap. 2, Sec. 9, Example 1] that if f (x) = x ln x, x ∈ [0, 1], then
2
ω (f, t) ≤ 2(ln 2)t. Hence, we have shown that for any random variable X ∈ [0, 1],
|E[X ln X] − E[X] ln(E[X])| .
q
Var(X).
(170)
Specializing to the case where X = p̂n , where n · p̂n ∼ B(n, p), we have proved that for f (p) =
−p ln p, we have
s
|E[f (p̂n )] − p ln p| .
p(1 − p)
.
n
(171)
q
The upper bound p(1−p)
is a pointwise bound that becomes smaller when p is close to 0 or
n
1. When p lies in the middle of the interval [0, 1], say p ≈ 21 , the bound is of scale √1n . We now
show that using the K-functional Kr,ϕ instead of Kr results in a better uniform bound in this case,
which is of order n1 . 6
For any f ∈ C[0, 1], n · p̂n ∼ B(n, p), and any g ∈ C[0, 1] such that g (1) ∈ A.C.loc , we have
|Ep [f (p̂n )] − f (p)| ≤ |Ep [f (p̂n ) − g(p̂n ) + g(p̂n ) − g(p) + g(p) − f (p)|
(172)
= 2kf − gk + |Ep g(p̂n ) − g(p)|
(173)
"
= 2kf − gk + Ep g 0 (p)(p̂n − p) +
= 2kf − gk + Ep
≤ 2kf − gk + Ep
= 2kf − gk + Ep
"Z
Z p̂n
#
(p̂n − t)g 00 (t)dt
(174)
p
#
p̂n
00
(p̂n − t)g (t)dt
(175)
p̂n − t
· t(1 − t)g 00 (t) dt
t(1 − t)
(176)
p
Z p̂n
p
Z p̂n
|p̂n − p|
p
p(1 − p)
≤ 2kf − gk + kϕ2 g 00 kEp
1
≤ 2kf − gk + kϕ2 g 00 k ,
n
6
|t(1 − t)g 00 (t)|dt
(p̂n − p)2
p(1 − p)
(177)
(178)
(179)
One remarkable fact is that, the K-functional approach with Kr,ϕ in bias analysis provides the tight norm bound
for any f ∈ C[0, 1] under the binomial model [Tot94].
28
p
where ϕ(t) = t(1 − t), and we used the elementary inequality that
between p and p̂n . Taking the infimum over all g, we have
|p̂n −t|
t(1−t)
≤
|p̂n −p|
p(1−p)
for any t
1
)
2n
(180)
1
. ωϕ2 (f, √ ).
2n
(181)
|Ep [f (p̂n )] − f (p)| ≤ 2Kr,ϕ (f,
It follows from [JVHW17] that for f (p) = −p ln p, ωϕ2 (f, t) t2 , which implies that
|Ep [p̂n ln p̂n ] − p ln p| .
1
.
n
(182)
The functional Ent(X) , E[X ln X]−E[X] ln(E[X]), which is also called entropy, plays a crucial
role in the theory of concentration inequalities. Concretely, the Herbst argument [BLM13] shows
2
that if Ent(eλf (X) ) ≤ λ2 E[eλf (X) ], we have sub-Gaussian type concentration P (f (X) − E[f (X)] ≥ t) ≤
2
e−t /2 . Due to the significance of the functional Ent(X), we now present a theorem providing upper and lower bounds of Ent(X). The key idea in the following proof is to relate the Ent(X)
functional to the KL divergence, whose functional inequalities have been well studied in the literature. Conceivably, they are stronger bounds than those obtained using the general K-functional
approach (Lemma 17 in Section B).
Theorem 10. Suppose X is a non-negative random variable. Denote Ent(X) = E[X ln X] −
E[X] ln(E[X]). Then,
q
Var(X)
Ent(X) ≤ E[X] ln 1 +
≤
Var(X)
(E[X])2
q
√
√
Ent(X) ≥ 2 E[X] − E[X]E[ X] ≥ Var( X)
1
X
Ent(X) ≥ E[X] E
−1
2
E[X]
(183)
(184)
2
.
(185)
√
Remark 6. It was shown in [LO00] that Ent(X) ≥ Var( X). Theorem 10 strengthens [LO00].
Proof. Without loss of generality we assume E[X] > 0. Then, we have
X
Ent(X) = E X ln
E[X]
X
X
= E[X] · E
ln
E[X] E[X]
(186)
(187)
(188)
Denote the distribution of X as Q, and introduce a new probability measure P via the Radon–
Nikodym derivative
dP
X
=
,
dQ
E[X]
29
(189)
we have
dP
dP
ln
dQ dQ
= EQ [X] · D(P, Q),
(190)
Ent(X) = EQ [X]EQ
(191)
where D(P, Q) is the KL divergence between P and Q.
Applying Lemma 18 in Section B, we have
Ent(X) ≤ EQ [X] · ln EQ
X
EQ [X]
!2
(192)
Var(X)
= E[X] · ln 1 +
,
(E[X])2
(193)
where in the last step we used the fact that Var(X) = E[X 2 ] − (E[X])2 . Using the fact that
√
< 1, we have
supx≥0 ln(1+x)
x
Ent(X) ≤
q
Var(X).
(194)
Now we prove the lower bounds. Applying the Hellinger distance part of Lemma 18 in Section B,
we have
Ent(X) ≥ EQ [X] · H 2 (P, Q)
s
(195)
!2
X
−1
EQ [X]
q
√
≥ 2 EQ [X] − EQ [X]EQ [ X]
√
≥ EQ [X] − (EQ [ X])2
√
= Var( X).
√
√
p
Here in the last inequality we used the fact that E[X] + (E[ X])2 − 2 E[X]E[ X] ≥ 0.
Applying the total variation distance part of Lemma 18 in Section B, we have
= EQ [X] · EQ
Ent(X) ≥ EQ [X] · 2V 2 (P, Q)
X
1
= EQ [X] · EQ
−1
2
EQ [X]
B
(196)
(197)
(198)
(199)
(200)
!2
.
(201)
Auxiliary lemmas
Lemma 10 (Mean value theorem for divided difference). Suppose the function f is n − 1 times
differentiable in the interval determined by the smallest and the largest of the xi ’s, we have
f [x1 , x2 , . . . , xn ] =
30
f (n−1) (ξ)
,
(n − 1)!
(202)
where ξ is in the open interval (mini xi , maxi xi ), and f [x1 , x2 , . . . , xn ] is the divided difference in
Definition 2.
The following lemma which is closely related to the mean value theorem for divided differences
in the continuous case.
Lemma 11. For integers x0 < x1 < · · · < xr and any function f defined on Z, the following holds:
|f [x0 , · · · , xr ]| ≤
1
max |∆r f (x)|.
r! x∈[x0 ,xr ]
(203)
Here ∆r f (x) denotes the r-th order backward difference of f , which is defined as
r
X
r
∆ f (x) ,
!
k
(−1)
k=0
r
f (x − k).
k
(204)
Lemma 12. Suppose one observes X ∼ B(n, p). Then, the r-jackknife estimator with n1 =
n − 1, n2 = n, r = 2 in estimating f (p) in (14) can be represented as
fˆ2 = nf
X
n
n−1
(n − X)f
n
−
X
n−1
+ Xf
X −1
n−1
,
(205)
where one conveniently sets f (x) = 0 if x < 0.
Lemma 13. Let r ≥ 2. Then, for the coefficients given in (15), we have the following.
1. If ρ = 0, then
r
X
Ci
i=1
nρi
= 1.
(206)
= 0.
(207)
2. If 1 ≤ ρ ≤ r − 1, then
r
X
Ci
i=1
nρi
3. If ρ ≥ r, then
r
X
Ci
nρ
i=1 i
≤
r−2
Y
(r − 1 − ρ − s)
s=0
≤
1
1
(r − 1)! nρ1
(ρ − 1)r−1 1
.
(r − 1)! nρ1
(208)
(209)
Define Tn,s (p) = ns Ep (p̂n − p)s , n = 1, 2, . . . , s = 0, 1, . . .. We have Tn,0 = 1, Tn,1 = 0. Upon
observing the recurrence relation
0
Tn,s+1 (x) = x(1 − x) Tn,s
(x) + nsTn,s−1 (x) ,
one obtains the following result.
31
(210)
Lemma 14. [DL93, Chapter 10, Theorem 1.1.] For a fixed s = 0, 1, . . ., Tn,s (p) is a polynomial in
p of degree ≤ s, and in n of degree bs/2c. Moreover, for ϕ2 = p(1 − p), we have
Tn,2s (p) =
s
X
aj,s (ϕ2 )nj ϕ2j
(211)
j=1
Tn,2s+1 (p) = (1 − 2p)
s
X
bj,s (ϕ2 )nj ϕ2j ,
(212)
j=1
where aj,s , bj,s are polynomials of degree ≤ s − j, with coefficients independent of n.
Lemma 15. The central moments of p̂n where n · p̂n ∼ B(n, p) satisfy the following:
bs/2c
ns Ep (p̂n − p)s =
X
hj,s (p)nj ,
(213)
j=1
where
(4es)s
.
j!
khj,s (p)k∞ ≤
(214)
Lemma 16 (Chernoff bound). [AV79] Let X1 , X2 , . . . , Xn be independent {0, 1} valued random
P
variables with P(Xi = 1) = pi . Denote X = ni=1 Xi , µ = E[X]. Then,
P(X ≤ (1 − β)µ) ≤ e−β
2 µ/2
0<β≤1
β2 µ
e− 2+β ≤ e− βµ
3
P(X ≥ (1 + β)µ) ≤
2
−β µ
e
(215)
β>1
(216)
0<β≤1
3
Lemma 17. [ST77] For any continuous function f : R 7→ R and any random variable X taking
values in R, we have
p
|E[f (X)] − f (E[X])| ≤ 3 · ω
2
f,
!
Var(X)
.
2
(217)
If f is only defined on an interval [a, b] that is a strict subset of R, the result holdswith the constant 3
P
replaced by 15. Here ω r (f, t) , sup0<h≤t k∆rh f k, where ∆rh f (x) = rk=0 (−1)k kr f (x+r(h/2)−kh),
and ∆rh f (x) = 0 if x + rh/2 or x − rh/2 is not inside the domain of f .
Lemma 18. [Tsy08, Section 2.4] Suppose P, Q are both probability measures, and P Q.
Introduce the following divergence functionals:
1. Total variation distance:
1
dP
V (P, Q) = EQ
−1 ;
2
dQ
(218)
2. Hellinger distance:
s
H(P, Q) = EQ
32
dP
−1
dQ
!2 1/2
;
(219)
3. Kullback–Leibler (KL) divergence:
D(P, Q) = EQ
dP
dP
;
ln
dQ dQ
(220)
4. χ2 divergence:
2
dP
χ (P, Q) = EQ
−1
dQ
dP 2
= EQ
− 1.
dQ
2
(221)
(222)
Then, we have the following upper and lower bounds on the KL divergence:
D(P, Q) ≤ ln 1 + χ2 (P, Q)
(223)
D(P, Q) ≥ 2V 2 (P, Q)
(224)
2
D(P, Q) ≥ H (P, Q).
C
C.1
(225)
Proofs of main theorems and lemmas
Proof of Theorem 1
Recognizing Ep fˆr as linear combination of operators, the first and second parts of Theorem 1 follow
from [DT87, Theorem 9.3.2.], the third part follows from [DT87, Corollary 9.3.8.], and the last part
follows from [Tot94].
C.2
Proof of Theorem 2
n
m
o
Define f ∈ C(0, 1] to be the piecewise linear interpolation function at nodes m−1 , 1+(−1)
, m ∈ N+ .
2
Clearly f (m−1 ) = 1 when m is even, f (m−1 ) = 0 when m is odd, and kf k∞ ≤ 1. We set f (0) = 0.
We have
|Ep fˆr − f (p)| =
r
X
Ci Ep f (p̂n+i−r ) − f (p)
(226)
i=1
≥
X
X
Ci Ep f (p̂n+i−r ) −
n+i−r is even
≥
X
Ci Ep f (p̂n+i−r ) − |f (p)|
(227)
n+i−r is odd
X
|Ci |Ep f (p̂n+i−r ) −
n+i−r is even
|Ci |Ep f (p̂n+i−r ) − 1
(228)
n+i−r is odd
where in the last step we have used the fact that C1 , · · · , Cr have alternating signs, and f ≥ 0.
When n + i − r is even and p = n−1 ,
1
Ep f (p̂n+i−r ) ≥ f
· P(B(n + i − r, p) = 1)
n+i−r
= (n + i − r)p (1 − p)n+i−r−1
1
≥ (1 − o(1)).
e
33
(229)
(230)
(231)
When n + i − r is odd and p = n−1 , noting that f (0) = f
1
n+i−r
= 0, we have
Ep f (p̂n+i−r ) ≤ kf k∞ · P(B(n + i − r, p) ≥ 2)
2
(1 + o(1)).
≤ 1−
e
Since ri=1 Ci = 1, we have
these together, we arrive at
P
n+i−r is odd |Ci |
P
|Ep fˆr − f (p)| &
= (1+o(1))
P
(233)
n+i−r is even |Ci |
2
1
− 1−
|Ci |
e
e
n+i−r is even
X
(232)
nr−1 . Combining
− o(1)
(234)
& nr−1 .
(235)
which completes the proof of the first claim.
As for the second claim, it suffices to replace the function f on interval [0, 1/n] by the linear
n
interpolation function interpolating f (0) = 0 and f (1/n) = 1+(−1)
and keep other parts of the
2
function intact. Consequently, after this modification f ∈ C[0, 1].
Now we prove the variance part.
Construct
interpolation function at the following nodes:
f ∈ C[0,
1] to be a piecewise
linear
1
2
1
2
f (0) = 0, f n = f n−1 = 1, f n−1 = f n = −1, f (1) = 0.
It follows from straightforward algebra and Lemma 12 that
fˆ2 =
0
X=0
2n − 2 + n−1
−2n + 5 − 4
n
X=1.
(236)
X=2
It follows from the definition of variance that
Varp (fˆ2 ) = Ep fˆ2 − Ep fˆ2
= inf Ep fˆ2 − a
2
(237)
2
(238)
a
≥ inf P(B(n, p) = 1)(2n − 2 + n−1 − a)2 + P(B(n, p) = 2)(−2n + 5 − 4/n − a)2
a
= inf np(1 − p)n−1 (2n − 2 + n−1 − a)2 +
a
(239)
n(n − 1) 2
p (1 − p)n−2 (−2n + 5 − 4/n − a)2
2
(240)
Setting p = 1/n, we have
1
1
Varp (fˆ2 ) ≥
1−
2
n
n−1
inf (2n − 2 + n−1 − a)2 + (−2n + 5 − 4/n − a)2 .
a
(241)
The infimum is achieved when
(2n − 2 + n−1 ) + (−2n + 5 − 4/n)
2
3
1
=
1−
.
2
n
a=
34
(242)
(243)
Hence,
1 n−1
1 3
1
1
1−
· 2 · 2n − 2 + −
1−
kVarp (fˆ2 )k ≥
2
n
n 2
n
n−1
1
≥ 1−
n2
n
n2
≥
.
e
where we have used n ≥ 4 and 1 −
C.3
1
n
n−1
2
(244)
(245)
(246)
≥ e−1 .
Proof of Lemma 4
Since f satisfies condition Ds , it admits the Taylor expansion:
f (x) = f (p) +
s−1
X
f (u) (p)
(x − p)u + Rs (x; p)
u!
u=1
(247)
Applying the r-jackknife estimator on it, we have
Ep fˆr = f (p) +
= f (p) +
r
r
X
f (u) (p) X
Ci Ep Rs (p̂ni ; p)
Ci Ep (p̂ni − p)u +
u! i=1
u=1
i=1
s−1
X
(248)
s−1
X
r
r
X
f (u) (p) X
Ci Ep Rs (p̂ni ; p),
Ci Ep (p̂ni − p)u +
u! i=1
u=r+1
i=1
Pb
where in the last step we have used Lemma 13 and Lemma 14. By convention
P
Denote Eu = ri=1 Ci Ep (p̂ni − p)u , u ≥ r + 1, it follows from Lemma 15 that
Eu =
r
X
bu/2c
Ci
i=1
X
hj,u (p)
j=1
1
nu−j
i
bu/2c
X
=
hj,u (p)
r
X
Ci
i=1
j=1
nu−j
i
.
a
(249)
= 0 if a > b.
(250)
Note that
r
X
Ci
i=1
nu−j
i
6= 0
(251)
if and only if u − j ≥ r, and when that is the case, we have
r
X
Ci
i=1
nu−j
≤ (u − j − 1)r−1
1
nu−j
1
≤
ur−1
nu−j
1
≤
ur−1
nr1
(252)
Since r is fixed, it follows from (252), Lemma 15 and condition Ds with parameter L that
s−1
X
f (u) (p)
1
Eu .r,s,L r .
u!
n
u=r+1
35
(253)
It follows from the integral form of Taylor remainder that
Rs (x; p) =
=
1
(s − 1)!
Z x
(x − t)s−1 f (s) (t)dt
p
R1
1
(s−1)! 0 |x
1 R 1 |x
(s−1)! 0
(254)
− t|s−1 f (s) (t)1t∈[min{x,p},max{x,p}] (−1)1x<p dt
s is odd
− t|s−1 f (s) (t)1t∈[min{x,p},max{x,p}] dt
s is even
(255)
When s is even, we have
Ep |p̂ni − t|s−1 1t∈[min{p̂ni ,p},max{p̂ni ,p}] =
Ep (p̂n − t)s−1
+
i
t≥p
t)s−1
−
t<p
E (p̂ −
p ni
(256)
where (x)+ = max{x, 0}, (x)− = max{−x, 0}.
When s is odd,
Ep |p̂ni − t|s−1 1t∈[min{p̂ni ,p},max{p̂ni ,p}] (−1)1p̂ni <p =
Ep (p̂n − t)s−1
+
i
t≥p
−E (p̂ − t)s−1
p ni
−
t<p
(257)
Hence,
r
X
1
Ci Ep Rs (p̂ni ; p) =
(s − 1)!
i=1
Z 1
+
Z p
f
0
f (s) (t)
p
.r,s,L
(s)
(t)
r
X
!
Ci Ep (p̂ni −
s
t)s−1
− (−1)
dt
i=1
r
X
!
Ci Ep (p̂ni − t)s−1
dt
+
(258)
i=1
Z p X
r
0
Ci Ep (p̂ni − t)s−1
dt +
−
Z 1 X
r
i=1
p
Ci Ep (p̂ni − t)s−1
dt,
+
(259)
i=1
where we have used the assumption that kf (s) k ≤ L.
C.4
Proof of Lemma 5
First we recall the following additive Chernoff bound: for np̂n ∼ B(n, p) and t ≥ p, we have
P(p̂n > t) ≤ exp(−nD(tkp))
(260)
where D(tkp) denotes the KL divergence between the binary distributions (t, 1 − t) and (p, 1 − p).
Note that
D(tkp) ≥
1
d2 D(ukp)
min
2 ξ∈[p,t]
du2
(t − p)2
ξ∈[p,t] 2ξ(1 − ξ)
(t − p)2
≥ min
2ξ
ξ∈[p,t]
2
(t − p)
=
2t
= min
36
· (t − p)2
(261)
u=ξ
(262)
(263)
(264)
we arrive at the following inequality:
P(p̂n > t) ≤ exp(−nD(tkp)) ≤ exp(−
n(t − p)2
).
2t
(265)
Now we prove the lemma. Let X1 , · · · , Xn be i.i.d Bern(p) random variables, and consider the
following coupling between p̂n−s , · · · , p̂n : for k = 0, · · · , s, define
p̂n−k =
X
1 n−k
Xi .
n − k i=1
(266)
In other words, we have
p̂n−k = p̂n−s +
X
1 s−1
(Xn−i − p̂n−s ).
n − k i=k
(267)
Now define gu,t (x) = (x − t)u+ , by Taylor expansion we have
s
∆ An,u (t) = Ep
s
X
!
s
gu,t (p̂n−k )
k
k
(−1)
k=0
= Ep
s
X
!
(268)
(j)
X gu,t (p̂n−s )
s u−1
(p̂n−k − p̂n−s )j
j!
k
j=0
k
(−1)
k=0
(u)
gu,t (ξk )
+
(p̂n−k − p̂n−s )u
u!
= Ep
s
X
!
(269)
(j)
u
gu,t (p̂n−s )
s X
(p̂n−k − p̂n−s )j
j!
k
j=0
k
(−1)
k=0
(u)
(u)
gu,t (ξk ) − gu,t (p̂n−s )
+
(p̂n−k − p̂n−s )u
u!
= Ep
(j)
u
s
X
gu,t (p̂n−s ) X
j=0
j!
+ Ep
s
X
(j)
u
X
gu,t (p̂n−s )
j=0
j!
(−1)
k
! (u)
k=0
≡ Ep
X
s
1 s−1
Ep [(
(Xn−i − p̂n−s ))j |X n−s ]
k
n − k i=k
!
k=0
k
(−1)
(270)
(u)
s gu,t (ξk ) − gu,t (p̂n−s )
(p̂n−k − p̂n−s )u
k
u!
Aj + Ep Bu .
(271)
(272)
(u)
Note that gu,t (x) is in fact not u-times differentiable at x = t, but with the convention that gu,t (t)
can stand for any number in [0, u!], the previous formula remains valid.
C.4.1
Non-remainder term Aj
Further define
aj (t) , Ep (Xn − t)j = p(1 − t)j + (1 − p)(−t)j
37
(273)
we have a0 (t) = 1, a1 (t) = p − t, and
Aj =
s−1
X
!
s
1
k (n − k)j
(−1)k
k=0
j
i1 · · · is−k
X
i1 +···+is−k =j
! s−k
Y
Denote by Ij the set of all multi-indices i = (i1 , · · · , it(i) ) with t(i) ≤ s,
for any i ∈ Ij , the coefficient of
Qt(i)
l=1 ail (p̂n−s )
ail (p̂n−s ).
Pt(i)
l=1 il
= j and il ≥ 1, then
in Aj is
s−1
X
j
s
s−k
1
P0
(−1)k
j
i1 · · · it(i) k=0
k (n − k) d1 · · · dt0 s − k − tl0 =1 dl0
!
bi =
=
=
=
j
i1 · · · it(i)
!
j
i1 · · · it(i)
!
j
i1 · · · it(i)
!
!
t(i)
d1 · · · dt0
! s−1
X
t(i)
d1 · · · dt0
!
t(i)
d1 · · · dt0
!
!
s−k
s
1
j
t(i)
k (n − k)
k
(−1)
k=0
! s−t(i)
X
s
t(i)
max
!
(275)
!
(276)
!
k
(−1)
k=0
s − t(i)
1
(n − k)j
k
(277)
!
s
· ∆s−t(i) n−j
t(i)
(278)
where d1 , · · · , dt0 is the nonzero histograms of i = (i1 , · · · , it(i) ) with
value theorem of backward differences, we have
|∆s−t(i) n−j | ≤
(274)
l=1
x∈[n−s+t(i),n]
Pt0
l0 =1 dl0
= t(i). By the mean
ds−t(i) −j
(x ) .s n−(j+s−t(i))
dxs−t(i)
(279)
and thus
|bi | ≤ c(i)n−(j+s−t(i))
(280)
where the constant c(i) does not depend on n or p.
Now for any j = 0, 1, · · · , u, we have
(j)
t(i)
X Y
gu,t (p̂n−s )
b
·
ail (p̂n−s )
Ep
Aj Ep (p̂n−s − t)u−j
i
+
j!
i∈I
l=1
(281)
j
≤
X
t(i)
Y
u−j
c(i)n−(j+s−t(i)) · Ep (p̂n−s − t)+
i∈Ij
|ail (p̂n−s )| .
(282)
l=1
For any i ∈ Ij , denote by v(i) the number of ones in i = (i1 , · · · , it ), using a1 (t) = p − t and
|aj (t)| ≤ p + t yields
t(i)
Y
u−j
Ep (p̂n−s − t)+
h
i
v(i)
|ail (p̂n−s )| ≤ Ep (p̂n−s − t)u−j
|p̂n−s + p|t(i)−v(i) .
+ |p̂n−s − p|
l=1
38
(283)
We distinguish into two cases. If np ≥ 1, by Holder’s inequality E[XY Z]3 ≤ E[|X|3 ]E[|Y |3 ]E[|Z|3 ]
we obtain
Ep (p̂n−s −
t)u−j
+
t(i)
Y
h
v(i)
|p̂n−s + p|t(i)−v(i)
|ail (p̂n−s )| ≤ Ep (p̂n−s − t)u−j
+ |p̂n−s − p|
i
(284)
l=1
≤ Ep |p̂n−s − p|u−j+v(i) 1(p̂n−s ≥ t) · |p̂n−s + p|t(i)−v(i)
h
≤ Ep |p̂n−s − p|3(u−j+v(i))
1
3
Ep |p̂n−s + p|3(t(i)−v(i))
i
(285)
1 q
3
· 3 P(p̂n−s > t)
(286)
p)2
p u−j+v(i) t(i)−v(i)
c2 n(t −
.u,s ( ) 2
)
·p
· exp(−
n
t
u−j+v(i) u−j+2t(i)−v(i)
c2 n(t − p)2
2
= n− 2 p
· exp(−
)
t
(287)
(288)
where in the last step we have used (265), and the fact that when np ≥ 1, we have [HJW16]
p k
Ep |p̂n − p|k .k ( ) 2
n
Ep |p̂n + p|k .k pk .
(289)
(290)
As a result, in this case we conclude that
Ep
(j)
X
u−j+v(i) u−j+2t(i)−v(i)
gu,t (p̂n−s )
c2 n(t − p)2
2
Aj .u,s
c(i)n−(j+s−t(i)) · n− 2 p
exp(−
)
j!
t
i∈I
(291)
j
=
X
u
u
c(i)n−( 2 +s) p 2 · (np)−
j+v(i)−2t(i)
2
exp(−
i∈Ij
Note that j =
Pt(i)
k=1 ik
c2 n(t − p)2
).
t
(292)
≥ v(i) + 2(t(i) − v(i)), we have j + v(i) ≥ 2t(i), and thus by np ≥ 1,
(j)
X
gu,t (p̂n−s )
u
u
c2 n(t − p)2
Ep
Aj .u,s
)
c(i)n−( 2 +s) p 2 · exp(−
j!
t
i∈I
(293)
j
u
u
.u,s n−( 2 +s) p 2 · exp(−
c2 n(t − p)2
)
t
(294)
where we have used that both |Ij | and c(i) do not depend on n or p in the last step.
If np < 1, we first show that
Ep [p̂kn 1(p̂n > t)] .k
p
nk−1
exp(−
c2 n(t − p)2
).
t
(295)
In fact, the MGF of p̂n gives
λ
Ep [eλp̂n ] = (pe n + 1 − p)n .
Differentiating w.r.t λ for k times, for λ > 0 we arrive at
39
(296)
Ep [p̂kn eλp̂n ] =
i
dk h λ
(pe n + 1 − p)n
k
dλ
≤ Ck
k
X
λ
(297)
λ
(pe n )j nj−k (pe n + 1 − p)n−j
(298)
j=1
λ
λ
pe n (pe n + 1 − p)n−1
1+
.k
nk−1
!k−1
λ
npe n
(299)
λ
n
pe + 1 − p
λ
kλ
pe n (pe n + 1 − p)n
.k
,
nk−1
(300)
where Ck is a universal constant depending on k only, and in the last step we have used np < 1.
As a result, by Markov’s inequality we have for any λ > 0,
λ
kλ
Ep [p̂kn
Specifically, when t >
1(p̂n > t)] ≤
2k
n,
Ep [p̂kn eλ(p̂n −t) ]
.k e
−λt
pe n (pe n + 1 − p)n
.
·
nk−1
(301)
choosing λ = n ln (1−p)t
(2−t)p yields
Ep [p̂kn
1(p̂n > t)] .k
p
nk−1
p
(2 − t)p
(1 − p)t
(2 − t)p
(1 − p)t
nt−k
1−p
1 − 2t
·
nt
1−p
1 − 2t
2
!n
(302)
!n
·
nk−1
t
p
≤ k−1 exp(−nD( kp))
n
2
p
c2 n(t − p)2
≤ k−1 exp(−
)
n
t
≤
as desired. When p < t ≤
2k
n,
we have
c2 n(t−p)2
t
(303)
(304)
(305)
= O(1), and (295) follows from
Ep [p̂kn 1(p̂n < t)] ≤ Ep p̂kn .k
p
nk−1
.
(306)
Now based on (295), we have
Ep (p̂n−s −
t)u−j
+
t(i)
Y
h
v(i)
|ail (p̂n−s )| ≤ Ep (p̂n−s − t)u−j
|p̂n−s + p|t(i)−v(i)
+ |p̂n−s − p|
i
(307)
l=1
≤ Ep |p̂n−s − p|u−j+v(i) 1(p̂n−s ≥ t) · |p̂n−s + p|t(i)−v(i)
h
h
u−j+t(i)
.u,s Ep (p̂n−s
.u,s
.u,s
p
+ pu−j+t(i) )1(p̂n > t)
i
+ pu−j+t(i) · exp(−
nu−j+t(i)−1
p
c2 n(t − p)2
exp(−
).
t
nu−j+t(i)−1
40
c2 n(t − p)2
)
t
i
(308)
(309)
(310)
(311)
As a result, in this case we have
(j)
X
gu,t (p̂n−s )
p
c2 n(t − p)2
c(i)n−(j+s−t(i)) · u−j+t(i)−1 exp(−
Ep
Aj .u,s
)
j!
t
n
i∈I
(312)
j
=
X
c(i)n−(u+s−1) p · exp(−
i∈Ij
.u,s n−(u+s−1) p · exp(−
c2 n(t − p)2
)
t
c2 n(t − p)2
).
t
(313)
(314)
Moreover, when 21 < p ≤ 1 − n1 , we can use the symmetry n(1 − p̂n ) ∼ B(n, 1 − p) and |aj (t)| ≤
(1 − p) + (1 − t), and then adapt the proof of the first case to conclude that
(j)
Ep
gu,t (p̂n−s )
u
u
c2 n(t − p)2
Aj .u,s n−( 2 +s) (1 − p) 2 · exp(−
).
j!
1−p
(315)
In summary, for the non-remainder terms we have
2
n−(u+s−1) p · exp(− c2 n(t−p)
)
t
2
u
u
if p < n1
(j)
gu,t (p̂n−s )
Ep
Aj .u,s n−( 2 +s) p 2 · exp(− c2 n(t−p)
)
if n1 ≤ p ≤ 12
t
j!
2
u
−( u2 +s)
) if 21 < p ≤ 1 − n1 .
n
(1 − p) 2 · exp(− c2 n(t−p)
1−p
C.4.2
(316)
The remainder term Bu
(u)
(u)
(u)
By our convention on gu,t (·) we observe that gu,t (ξk ) − gu,t (p̂n−s ) is non-zero only if
(ξk − t)(p̂n−s − t) ≤ 0.
(317)
However, since min{p̂n−k , p̂n−s } ≤ ξk ≤ max{p̂n−k , p̂n−s }, the previous inequality implies that the
“path" consisting of p̂n−s , · · · , p̂n under our coupling “walks across" t. Let’s call it “good path".
Moreover, note that
|p̂n−k − p̂n−s | ≤
s
,
n−s
k = 0, · · · , s
(318)
under our coupling, for a good path we must have
|p̂n−s − t| ≤
s
.
n−s
(319)
1
Since p̂n−s must be an integral multiple of n−s
, we conclude that the number of good paths is O(1).
Let’s call p̂n−k → p̂n−k+1 a “right step" if p̂n−k < p̂n−k+1 or p̂n−k = p̂n−k+1 = 1, and a “left step"
if p̂n−k > p̂n−k+1 or p̂n−k = p̂n−k+1 = 0.
First we consider the case where p ≤ 12 , and consider any good path L. The idea is that, each
good path L gives rise to a realization of Bu , and Ep Bu is the expectation averages over all good
paths. Hence, to evaluate Ep Bu , it suffices to compute the value of Bu given L, and the probability
41
of the path L. Denote by r, l the number of right and left steps in L, and by q the starting point
of L, it is easy to see that the probability of this path is
P(L) = pr (1 − p)l · P(B(n − s, p) = (n − s)q).
(320)
We first take a look at the quantity P(B(n, p) = nq) for any q with |q − t| = O(n−1 ). By
Stirling’s approximation
n! =
√
2πn
n
n
e
(1 + o(1)),
(321)
we have
!
P(B(n, p) = nq) =
.
.
≤
n nq
p (1 − p)n−nq
nq
√
2πn( ne )n
nq
n−nq
√
nq nq p
n−nq n−nq p (1 − p)
2πnq( e ) · 2π(n − nq)( e )
1
p
· e−nD(qkp)
nq(1 − q)
n(q − p)2+
1
p
· exp(−
)
2q
nq(1 − q)
(322)
(323)
(324)
(325)
where the last step is given by (265). Moreover, for q > 1 − n1 , (265) gives a better bound
P(B(n, p) = nq) ≤ exp(−
n(q − p)2
).
2q
(326)
Combining them together, we conclude that
n(q − p)2+
1
)
P(B(n, p) = nq) . √ exp(−
nq
4q
(327)
Now we show that if q = t+O(n−1 ), we can replace q by t without loss in the previous inequality.
In fact, if nq ≥ 1, it is easy to verify that
n(q − p)2+
1
1
n(t − p)2
P(B(n, p) = nq) . √ exp(−
) . ( √ ∧ 1) exp(−
)
nq
4q
4t
nt
(328)
and if nq < 1, we use the trivial bound
1
n(t − p)2
).
P(B(n, p) = nq) ≤ 1 . ( √ ∧ 1) exp(−
4t
nt
(329)
As a result, for all good paths with starting point q we conclude that
1
n(t − p)2
P(B(n, p) = nq) . ( √ ∧ 1) exp(−
)
4t
nt
42
(330)
Now we evaluate the quantity Bu (L) given L. In fact, it is easy to see
|Bu (L)| .u
n−u
if r > 0
( p )u
if r = 0.
n
(331)
As a result,
p
|Bu (L)| · P(L) .u n−u · p + ( )u · P(B(n, p) = nq)
n
1
c2 n(t − p)2
.u n−u pu∧1 · ( √ ∧ 1) exp(−
)
t
nt
Finally, there are only O(1) good paths, and for p ≤
|Ep Bu | ≤
X
L
1
2
c2 n(t − p)2
1
).
|Bu (L)| · P(L) .u n−u pu∧1 · ( √ ∧ 1) exp(−
t
nt
1
2
n−u (1 −
(√
1
n(1−t)
(334)
< p ≤ 1 − n1 , and we conclude
2
n−u pu∧1 · ( √1 ∧ 1) exp(− c2 n(t−p)
)
t
nt
u∧1
p)
·
(333)
we arrive at
The previous approach can also be applied to the case where
that
|Ep Bu | .u
(332)
if p ≤
2
) if
∧ 1) exp(− c2 n(t−p)
1−p
1
2
1
2
< p ≤ 1 − n1 .
(335)
Finally, when t > 1 − n1 , it’s easy to see
An,u (t) = (1 − t)u · P(p̂n = 1) = (1 − t)u pn
(336)
|∆s An,u (t)| = (1 − t)u |∆s pn | .u,s (1 − t)u pn−s (1 − p)s .
(337)
and
Now the combination of (316), (335) and (337) completes the proof of Lemma 5.
C.5
Proof of Lemma 6
First note that by assumption we have
k+1
1 + O(p)
−t=
n−j
n−j
(338)
for any j = 0, · · · , s in both cases, and when u < s we have
k
p
−t& .
n−s
n
(339)
Adopt the same coupling as in the proof of Lemma 5, and write
|∆s An,u (t)| = |Ep ∆s (p̂n − t)u+ (1(p̂n−s ∈ A) + 1(p̂n−s ∈
/ A)|
s
≥ |Ep ∆ (p̂n −
t)u+
1(p̂n−s ∈ A)| − |Ep ∆ (p̂n −
s
43
t)u+
(340)
1(p̂n−s ∈/ A)|.
(341)
where
k−s
k
.
A,
,
n−s n−s
(342)
u
By the proof of Lemma 5, the non-remainder terms in the Taylor expansion is at most O(n−( 2 +s) ) =
1
o(n−(u+ 2 ) ), and by our coupling the remainder term is non-zero only if p̂n−s ∈ A, we conclude that
|Ep ∆s (p̂n − t)u+ 1(p̂n−s ∈
/ A)| = o(n−(u+ 2 ) ).
1
(343)
Now we deal with the first term. It’s easy to verify that for any x ∈ A, we have
1
1
1
|x − p| = |t − p| + O( ) ≤ √ + O( ).
n
n
n
(344)
As a result, by the Stirling approximation formula in the proof of Lemma 5, we conclude that for
m
any n−s
∈ A with m ∈ N,
q , P(p̂n−s =
m
1
k
) = (1 + o(1)) · P(p̂n−s =
)& √ .
n−s
n−s
n
(345)
In other words, p̂n−s is almost uniformly distributed restricted to A with mass at least Θ( √1n ) on
each point. Moreover, for j = 0, 1, · · · , s and i = 1, · · · , s − j, it follows from the coupling that
k+i
, p̂n−s ∈ A)
n−j
s
X
k+i
k−m
k−m
P(p̂n−j =
=
|p̂n−s =
)P(p̂n−s =
)
n
−
j
n
−
s
n−s
m=0
P(p̂n−j =
=
s
X
m=0
(346)
!
s − j m+i
p
(1 − p)s−j−m−i · q(1 + o(1))
m+i
44
(347)
As a result, when u < s, we have
|Ep ∆s (p̂n − t)u+ 1(p̂n−s ∈ A)|
= Ep
s
X
!
s
(p̂n−j − t)u+ 1(p̂n−s ∈ A)
j
j
(−1)
j=0
s−j
(348)
s
XX
p
k+i
s k+i
= (−1) ( )u +
− t)u+ · P(p̂n−j =
, p̂n−s ∈ A)
(−1)j
(
n
n
−
j
n
−j
j
i=1 j=0
!
s
s−j X
s s−i−j
X
X
p
s
= q(1 + o(1)) (−1) ( )u +
(−1)j
n
j
i=1 j=0 m=0
!
s
(349)
!
s − j m+i
k+i
− t)u
p
(1 − p)s−j−m−i (
n−j
m+i
(350)
s X
s−i
X
!
p
s
= q(1 + o(1)) (−1)s ( )u +
pm+i
n
m
+
i
i=1 m=0
·
s−i−m
X
!
(−1)
j
j=0
s−m−i
k+i
− t)u
(1 − p)s−m−i−j (
n−j
j
(351)
s−i
s X
p u X
s
= q(1 + o(1)) (−1) ( ) +
pm+i (i + O(p))u
m
+
i
n
i=1 m=0
!
s
·
s−i−m
X
!
(−1)
j=0
j
s−m−i
1 u
(1 − p)s−m−i−j (
)
j
n−j
(352)
s X
s−i
s
(1 − p)s−m−i−x
p u X
pm+i (i + O(p))u ∆s−i−m
= q(1 + o(1)) (−1) ( ) +
.
m+i
n
(n − x)u
i=1 m=0
!
s
(353)
By the product rule of differentiation we know that
∆s−i−m
(−p)s−i−m (1 + O(p))
(1 − p)s−m−i−x
=
· (1 + o(1)).
(n − x)u
nu
(354)
Plugging into the previous expression, we arrive at
|Ep ∆s (p̂n − t)u+ 1(p̂n−s ∈ A)|
s X
s−i
X
q(1 + o(1))(1 + O(p))
s
s
u
=
(−1)
Θ(p
)
+
pm+i iu (−p)s−i−m
nu
m
+
i
i=1 m=0
!
s X
s−i
X
q(1 + o(1))(1 + O(p))
s
s
u
s
=
(−1)
Θ(p
)
+
p
(−1)m+i iu
nu
m
+
i
i=1 m=0
q
u
& u · p (1 + O(p))
n
1
n−(u+ 2 ) · pu (1 + O(p))
(355)
!
(356)
(357)
(358)
where we have used u < s. Hence, in this case by choosing p0 small enough we arrive at the desired
result.
45
When u ≥ s, similarly we have
|Ep ∆s (p̂n − t)u+ 1(p̂n−s ∈ A)|
s X
s−i
s
q(1 + o(1))(1 + O(p)) s X
p
(−1)m+i iu
=
nu
m
+
i
i=1 m=0
!
ps (1 + O(p))
1
nu+ 2
n
u+ 21
r
s X
iu
r i=1
!
r
(−1)
r=0
ps (1 + O(p))
=
·
s
X
· ∆s
(359)
r
X
(360)
iu .
(361)
i=1
Since ri=1 iu is a polynomial of r with degree u + 1 > s, its s-backward difference is not zero, and
thus the desired result also follows by choosing p0 small enough.
P
C.6
Proof of Lemma 8
We first prove the lower bound for the delete-1 jackknife. It suffices to fix p = nc , where c > 0 is a
positive constant, and prove
(362)
(363)
lim n Ep fˆ2 − f (p) 6= 0
n→∞
when f (p) = −p ln p, and
lim nα Ep fˆ2 − f (p) 6= 0
n→∞
for f (p) = pα , 0 < α < 1.
It follows from Lemma 12 that it suffices to analyze the expectation of the following quantity:
gn (X) = nf
X
n
n−1
(n − X)f
n
−
X
n−1
+ Xf
X −1
n−1
c
.
n
−f
(364)
As n → ∞, the measure µn = B(n, p) converges weakly to µ = Poi(c).
We will show that limn→∞ nEµn gn (X) 6= 0 when f (p) = −p ln p and limn→∞ nα Eµn gn (X) 6= 0
when f (p) = pα .
For f (p) = −p ln p,
X
n
n − 1 (n − X)X n − 1 n − 1 X(X − 1)
n−1
c n
ln
−
ln
−
ln
− ln
n
X
n
n−1
X
n
n−1
X −1 n c
(365)
n
X(n − X) n − 1 X(X − 1)
n−1
c n
= Eµn n · X ln
−
ln
−
ln
− ln
(366)
X
n
X
n
X −1 n c
X(n − X)
X(X − 1)
c n
= Eµn n · X ln n −
ln(n − 1) −
ln(n − 1) − ln
(367)
n
n
n c
1
X(n − X)
1
X(X − 1)
1
+ Eµn n · X ln
−
ln
−
ln
(368)
X
n
X
n
X −1
1
1
1
= (c − cn) ln 1 −
+ c ln c + Eµn X 2 ln
− X(X − 1) ln
(369)
n
X
X −1
= An + Bn .
(370)
Eµn n · gn (X) = Eµn n · n
46
1
Since limn→∞ An = c + c ln c, limc→0+ c−2 (c + c ln c) = ∞, and c−2 Eµ X 2 ln X1 − X(X − 1) ln X−1
has a finite limit at c = 0+ , we know there exists some c > 0 such that
lim Eµn n · gn (X) 6= 0.
(371)
n→∞
For f (p) = pα , 0 < α < 1, we have
α
n−1
n−1
c α
X
X −1 α
Eµn n · gn (X) = n
−
−
−
(372)
(n − X)
X
n
n−1
n
n−1
n
α
n−1
= Eµn nX α −
(X α (n − X) + X(X − 1)α ) − cα
(373)
n
1 α
n − 1 1−α
= Eµn nX α 1 − 1 −
− cα +
Eµn X α+1 − X(X − 1)α .
n
n
(374)
α
X
n
n
α
α
Hence,
lim Eµn nα · gn (X) = Eµ αX α + X α+1 − X(X − 1)α − cα .
(375)
n→∞
Noting that c−1 Eµ αX α + X α+1 − X(X − 1)α has a finite limit at c = 0 but cα−1 does not, we
know there exists some c > 0 such that
lim Eµn nα · gn (X) 6= 0
(376)
n→∞
when f (p) = pα , 0 < α < 1.
For a 2-jackknife that satisfies Condition 1, the lower bound is given by Theorem 1 and (12).
C.7
Proof of Theorem 8
The first three statements follow from [GZ94]. The last statement follows from [Tot94]. Now we
work to prove the fourth statement.
It follows from [DT87, Inequality 9.3.5., 9.3.7.] that for f ∈ C[0, 1], we have kϕ2m (Bk [f ])(2m) k ≤
m
(2m) ∈ C[0, 1], we have kϕ2m (B [f ])(2m) k ≤ Ckϕ2m f (2m) k. Here
Ck kf k,
k
p and for f such that f
ϕ(x) = x(1 − x).
It follows from the definition of the K-functional K2m,ϕ (f, 1/nm ) that
K2m,ϕ (f, 1/nm ) ≤ kf − ⊕m Bk [f ]k + n−m kϕ2m (⊕m Bk [f ])(2m) k
(377)
Noting that ⊕m Bk [f ] is nothing but linear combinations of the powers of Bk up to degree m,
we can apply the estimates on kϕ2m (Bk [f ])(2m) k to kϕ2m (⊕m Bk [f ])(2m) k. For any g ∈ A.C.loc , we
have
kϕ2m (⊕m Bk [f ])(2m) k ≤ kϕ2m (⊕m Bk [f − g])(2m) k + kϕ2m (⊕m Bk [g])(2m) k
m
2m (2m)
≤ Ck kf − gk + Ckϕ
g
k.
(378)
(379)
Hence,
n
−m
2m
kϕ
m
(2m)
(⊕ Bk [f ])
m
k
k ≤ C inf
n
g
m
=C
k
n
47
kf − gk + k −m kϕ2m g (2m) k
K2m,ϕ (f, k −m ).
(380)
(381)
We have showed that
m
m
K2m,ϕ (f, 1/n ) ≤ kf − ⊕ Bk [f ]k + C
m
k
n
K2m,ϕ (f, k −m ).
(382)
Now we utilize the assumption that ωϕ2m (f, 2t) ≤ Dωϕ2m (f, t). We have
√
D−q ωϕ2m (f, 1/ k) ≤ ωϕ2m (f, 2−q k −1/2 )
(383)
= ωϕ2m (f, (22q k)−1/2 )
(384)
≤ M1 K2m,ϕ (f, 1/(22q(m) k m )),
(385)
where in the last step we have used the equivalence between K-functional and the Ditzian–Totik
modulus.
Setting n = 22q k and applying (382), we have
√
√
D−q ωϕ2m (f, 1/ k) ≤ M1 kf − ⊕m Bk [f ]k + M1 2−2qm ωϕ2m (f, 1/ k).
(386)
We now choose q so that D−q > 2M1 2−2qm , which is possible since D−1 > 2−2m , and obtain
√
kf − ⊕m Bk [f ]k ≥ 2−2qm ωϕ2m (f, 1/ k).
(387)
The proof is complete.
C.8
Proof of Lemma 3
It follows from [Hur76, Theorem 1] that
Ep f (p̂n ) − f (p) −
2k−1
X
j=1
f (j) (p)
1
Ep (p̂n − p)j .k,L k .
j!
n
(388)
It follows from Lemma 14 and 15 that for any j ≥ 1 integer there exist polynomials hm,j (p) with
degree no more than j, and coefficients independent of n such that
bj/2c
Ep (p̂n − p)j =
X
m=1
hm,j (p)
1
nj−m
.
(389)
Define
Tj [f ](p) =
2k−1
X
i=1
f (i) (p)
hi−j,i (p)1 (i − bi/2c ≤ j ≤ i − 1) .
i!
(390)
It is clear that Tj [f ](p) depends on f only through the derivatives of f of order from j + 1 to 2j.
Concretely, Tj [f ](p) is a linear combination of the derivatives of f of order from j + 1 to 2j where
the combination coefficients are polynomials of p with degree no more than 2j.
48
C.9
Proof of Theorem 9
Proof. We first show, through induction, that each ti , 0 ≤ i ≤ k − 1 satisfies condition D2(k−i) with
parameter .k L. Indeed, it is true for i = 0. Assuming that it is true for i = m, 1 ≤ m ≤ k − 2, we
now show that it is true for i = m + 1. We have
tm+1 (p) = −
m+1
X
Tj [tm+1−j ](p).
(391)
j=1
Since Tj [f ] involves the derivatives of f up to order 2j (Lemma 3), for each j-th term Tj [tm+1−j ](p),
1 ≤ j ≤ m + 1, it involves the derivatives of f up to order
2(m + 1 − j) + 2j = 2(m + 1),
(392)
which implies that tm+1 satisfies the condition D2k−2(m+1) with parameter .k L.
Now we apply Lemma 3 to each term in the formula of fˆk . For the i-th term, 0 ≤ i ≤ k − 1, we
have
Ep
k−i−1
X
1
1
1
1
t
(p̂
)
=
t
(p)
+
Tj [ti ](p) + O
i n
i
i
i
j+i
n
n
n
nk
j=1
k−1
X 1
1
1
= i ti (p) +
Tm−i [ti ](p) + O
m
n
n
nk
m=i+1
(393)
.
(394)
.
(395)
Sum over 0 ≤ i ≤ k − 1, we have
k−1
X
k−1
X 1
1
1
ˆ
Ep [fk ] =
t
(p)
+
Tm−i [ti ](p) + O
i
i
m
n
n
nk
i=0
m=i+1
It suffices to show that
k−1
X
k−1
X 1
1
ti (p) +
Tm−i [ti ](p) = f (p).
i
n
nm
i=0
m=i+1
(396)
We have
k−1
X
i=0
k−1
X 1
1
t
(p)
=
f
(p)
+
ti (p),
i
ni
ni
i=1
(397)
and
k−1
X 1 m−1
X
1
T
[t
](p)
=
Tm−i [ti ](p)
m−i
i
m
m
n
n
m=1
i=0 m=i+1
i=0
k−1
X
k−1
X
=−
k−1
X
1
tm (p),
nm
m=1
where in the last step we used the definition of ti (p). The proof is now complete.
49
(398)
(399)
D
Proofs of auxiliary lemmas
D.1
Proof of Lemma 11
Let p(x) be the Lagrangian interpolating polynomial of f at points x0 , · · · , xr , and define g(x) =
f (x) − p(x). It is easy to see that g(x) has at least (r + 1) zeros x0 , · · · , xr . Since
x1
X
0 = g(x1 ) − g(x0 ) =
∆g(x)
(400)
x=x0 +1
there must be some y0 ∈ (x0 , x1 ] such that ∆g(y0 ) ≤ 0. Similarly, there exists some y1 ∈ (x1 , x2 ]
such that ∆g(y1 ) ≥ 0, and y2 ∈ (x2 , x3 ] such that ∆g(y2 ) ≤ 0, so on and so forth. Next observe
that
y1
X
0 ≤ ∆g(y1 ) − ∆g(y0 ) =
∆2 g(y)
(401)
y=y0 +1
there must exist some z0 ∈ (y0 , y1 ] such that ∆2 g(z0 ) ≥ 0. Similarly, there exists some z1 ∈ (y1 , y2 ]
such that ∆2 g(z1 ) ≤ 0, and z2 ∈ (y2 , y3 ] such that ∆2 g(z2 ) ≥ 0, so on and so forth. Repeating this
process, there must be some x ∈ [x0 , xr ] such that (−1)r ∆r g(x) ≥ 0. Note that p(x) is a degree-r
polynomial with leading coefficient f [x0 , · · · , xr ], we conclude that
0 ≤ (−1)r ∆r g(x)
r
(402)
r
= (−1) ∆ (f (x) − p(x))
r
r
= (−1) (∆ f (x) − r!f [x0 , · · · , xr ]).
(403)
(404)
Similarly, we can also show that there exists x0 ∈ [x0 , xr ] such that
0 ≤ (−1)r+1 (∆r f (x0 ) − r!f [x0 , · · · , xr ]).
(405)
Combining these two inequalities completes the proof.
D.2
Proof of Lemma 13
The key idea is to connect the problem of solving the matrix equations (16) to the notion of divided
difference in approximation theory.
For ρ ≥ 0, ρ ∈ Z,
r
X
Ci
i=1
nρi
=
=
r
X
1 Y
i=1
r
X
nρi
ni
n
− nj
j6=i i
nr−1−ρ
i
i=1
r−1−ρ
=x
1
n − nj
j6=i i
Y
[n1 , n2 , . . . , nr ],
(406)
(407)
(408)
where f [x1 , x2 , . . . , xr ] denotes the divided difference in Definition 2. The lemma is proved using
the mean value theorem (Lemma 10) of the divided difference for function xr−1−ρ .
50
D.3
Proof of Lemma 15
The first part follows from Lemma 14. Regarding the second part, the moment generating function
of p̂n − p is given by
n
E[exp(z(p̂n − p))] = e−zp 1 + p(ez/n − 1)
.
(409)
Written as formal power series of z, the previous identity becomes
∞
X
E(p̂n − p)s s
z =
∞
X
(−p)i i
z
s=0
i=0
s!
i!
! n
X
k=0
!
n k
p
k
!k
∞
X
1 z l
.
( )
l=1
l! n
(410)
Hence, by comparing the coefficient of nj−s z s at both sides, we obtain
j
s−i
tk,j−i k
hj,s (p) X (−p)i X
=
p
s!
i!
k!
i=0
a
k=j−i
k
Y
1
X
1 +···+ak =s−i,a1 ,··· ,ak ≥1 l=1
al !
(411)
where tk,r is the coefficient of xr in x(x − 1) · · · (x − k + 1). It’s easy to see
|tk,r | ≤ k
k
r
k−r
!
≤
kk
.
r!
(412)
Moreover, it’s easy to see when k ≤ s − i, we have
k
Y
1
X
a1 +···+ak =s−i,a1 ,··· ,ak ≥1 l=1
al !
X
≤
k
Y
1
a1 +···+ak =s−i l=1
X
≤
k+i
Y
a1 +···+ak+i =s l=1
al !
1
(k + i)s
=
al !
s!
(413)
(414)
and this quantity is zero when k > s − i.
Then, applying k! ≥
k
k
e
yields
|hj,s (p)| ≤
j
s−i
X
1 X
ek (k + i)s
i=0
≤
j
X
i=0
i! k=j−i (j − i)!
ses ss
i!(j − i)!
2j
j!
s(2es)s
≤
j!
(4es)s
≤
j!
= s(es)s
51
(415)
(416)
(417)
(418)
(419)
D.4
Proof of Lemma 17
We first prove the statement when the domain of f is the whole real line. We introduce the first
and second Stekolv functions fh (x), fhh (x) as follows:
fh (x) = f ∗ Kh
(420)
fh (x) = fh ∗ Kh = f ∗ Kh ∗ Kh ,
(421)
where Kh = h1 1(x ∈ [−h/2, h/2]) is the box kernel, and the operation ∗ denotes convolution.
The Steklov functions have the following nice properties [Akh56, Chap. V, Sec. 83]:
1 h/2
(f (x) − f (x + t))dt
(422)
h −h/2
1
h
h
0
fh (x) =
f x+
−f x−
(423)
h
2
2
Z
Z
1 h/2 h/2
f (x) − fhh (x) = 2
[4f (x) − f (x + s + t) − f (x + s − t) − f (x − s + t) − f (x − s − t)]dsdt
h 0
0
(424)
1
00
fhh
(x) = 2 (f (x + h) + f (x − h) − 2f (x)) .
(425)
h
f (x) − fh (x) =
Z
Hence, we have
|f (x) − fhh (x)| ≤
1
h2
Z h/2 Z h/2
0
[|2f (x) − f (x + s + t) − f (x − s − t)| + |2f (x) − f (x + s − t) − f (x − s + t)|]dsdt
0
(426)
≤
Z
Z
1 h/2 h/2
h2 0
0
1 2
= ω (f, h),
2
[2ω 2 (f, h)]dsdt
(427)
(428)
and
00
kfhh
(x)k ≤
1 2
ω (f, h).
h2
(429)
We use
|E[f (X)] − f (E[X])| ≤ |E[f (X) − fhh (X) + fhh (X) − fhh (E[X]) + fhh (E[X]) − f (E[X])]| (430)
≤ 2kf − fhh k + |E[fhh (X)] − fhh (E[X])|
1 00
≤ 2kf − fhh k + kfhh
kVar(X)
2
1
≤ ω 2 (f, h) + 2 ω 2 (f, h)Var(X)
2h
= 3 · ω 2 (f, h),
(431)
(432)
(433)
(434)
√
Var(X)
where h =
.
2
Now we argue that when the domain of f is an interval [a, b] that is a strict subset of R, we can
replace the constant 3 by 15. Indeed, as argued in [Tim14, Sec. 3.5.71, pg. 121], for any continuous
52
(a)
function f ∈ [a, b], one can extend φ(x) = f (x) − f (b)−f
(x − a) − f (a) to the whole real line while
b−a
ensuring the second order modulus of the extension is upper bounded by five times the ω 2 (f, t) of
the original function f . Indeed, one achieves this by extending φ so as to be odd with respect to
the ends of the [a, b] and then periodically with period 2(b − a) on the whole real line.
References
[AGW71]
JE Adams, HL Gray, and TA Watkins. An asymptotic characterization of bias reduction
by jackknifing. The Annals of Mathematical Statistics, pages 1606–1612, 1971.
[Akh56]
Naum Il’ich Akhiezer. Theory of approximation. New York, 1956.
[AS04]
Noga Alon and Joel H Spencer. The probabilistic method. John Wiley & Sons, 2004.
[AV79]
Dana Angluin and Leslie G Valiant. Fast probabilistic algorithms for Hamiltonian
circuits and matchings. Journal of Computer and system Sciences, 18(2):155–193, 1979.
[BCG16]
Eric Blais, Clément L Canonne, and Tom Gur. Alice and bob show distribution testing
lower bounds. 2016.
[BLM13]
Stéphane Boucheron, Gábor Lugosi, and Pascal Massart. Concentration inequalities:
A nonasymptotic theory of independence. Oxford university press, 2013.
[CCN14]
Gauss M Cordeiro and Francisco Cribari-Neto. An introduction to Bartlett correction
and bias reduction. Springer, 2014.
[CW00]
Shaun Cooper and Shayne Waldron. The eigenstructure of the Bernstein operator.
Journal of Approximation Theory, 105(1):133–165, 2000.
[DL93]
Ronald A DeVore and George G Lorentz. Constructive approximation, volume 303.
Springer Science & Business Media, 1993.
[DT87]
Zeev Ditzian and Vilmos Totik. Moduli of smoothness. Springer, 1987.
[Efr79]
Bradley Efron. Bootstrap methods: another look at the Jackknife. The Annals of
Statistics, pages 1–26, 1979.
[Efr82]
Bradley Efron. The jackknife, the bootstrap and other resampling plans. SIAM, 1982.
[ET94]
Bradley Efron and Robert J Tibshirani. An introduction to the bootstrap, volume 57.
CRC press, 1994.
[Gan03]
Michael I Ganzburg. Strong asymptotics in lagrange interpolation with equidistant
nodes. Journal of approximation theory, 122(2):224–240, 2003.
[GZ94]
Heinz H Gonska and Xin-long Zhou. Approximation theorems for the iterated Boolean
sums of Bernstein operators. Journal of Computational and applied mathematics,
53(1):21–31, 1994.
[Hal92]
Peter Hall. The bootstrap and Edgeworth expansion. Springer Science & Business Media,
1992.
53
[HJW16]
Yanjun Han, Jiantao Jiao, and Tsachy Weissman. Minimax rate-optimal estimation of
divergences between discrete distributions. arXiv preprint arXiv:1605.09124, 2016.
[HKPT12] Wolfgang Härdle, Gerard Kerkyacharian, Dominique Picard, and Alexander Tsybakov.
Wavelets, approximation, and statistical applications, volume 129. Springer Science &
Business Media, 2012.
[HM88]
Peter Hall and Michael A Martin. On bootstrap resampling and iteration. Biometrika,
75(4):661–671, 1988.
[Hur76]
Jan Hurt. Asymptotic expansions of functions of statistics. Aplikace matematiky,
21(6):444–456, 1976.
[IM78]
Mourad EH Ismail and C Ping May. On a family of approximation operators. Journal
of Mathematical Analysis and Applications, 63(2):446–462, 1978.
[JVHW15] Jiantao Jiao, Kartik Venkat, Yanjun Han, and Tsachy Weissman. Minimax estimation
of functionals of discrete distributions. IEEE Transactions on Information Theory,
61(5):2835–2885, 2015.
[JVHW17] Jiantao Jiao, Kartik Venkat, Yanjun Han, and Tsachy Weissman. Maximum likelihood
estimation of functionals of discrete distributions. IEEE Transactions on Information
Theory, 63(10):6774–6798, Oct 2017.
[LC98]
Erich Leo Lehmann and George Casella.
Springer, 1998.
[LO00]
Rafał Latała and Krzysztof Oleszkiewicz. Between Sobolev and Poincaré. In Geometric
aspects of functional analysis, pages 147–168. Springer, 2000.
[LS94]
Xin Li and EB Saff. Local convergence of Lagrange interpolation associated with
equidistant nodes. Journal of Approximation Theory, 78(2):213–225, 1994.
[May76]
CP May. Saturation and inverse theorems for combinations of a class of exponential
type operators. Canad. J. Math, 28(6):1224–1250, 1976.
[Mil74]
Rupert G Miller. The jackknife-a review. Biometrika, 61(1):1–15, 1974.
[MM08]
Giuseppe Mastroianni and Gradimir Milovanovic. Interpolation processes: Basic theory
and applications. Springer Science & Business Media, 2008.
[Mor82]
Carl N Morris. Natural exponential families with quadratic variance functions. The
Annals of Statistics, pages 65–80, 1982.
[Mor83]
Carl N Morris. Natural exponential families with quadratic variance functions: statistical theory. The Annals of Statistics, pages 515–529, 1983.
[Nat64]
Isidor Pavlovich Natanson. Constructive function theory: interpolation and approximation quadratures, volume 3. Ungar, 1964.
[Nat83]
GI Natanson. Application of the method of IP Natanson and I. Yu. Kharrik in the
algebraic case. Operator Theory and Function Theory, (1):166–170, 1983.
54
Theory of point estimation, volume 31.
[Nem00]
A Nemirovski. Topics in non-parametric statistics. lectures on probability theory and
statistics, saint flour 1998, 1738, 2000.
[Pan03]
Liam Paninski. Estimation of entropy and mutual information. Neural Computation,
15(6):1191–1253, 2003.
[PRW99]
Dimitris N Politis, Joseph P Romano, and Michael Wolf. Subsampling. Springer Series
in Statistics, 1999.
[Que56]
Maurice H Quenouille. Notes on bias in estimation. Biometrika, 43(3/4):353–360, 1956.
[RS02]
C Rose and MD Smith. h-statistics: Unbiased estimators of central moments. Mathematical Statistics with Mathematica, pages 253–256, 2002.
[Sev91]
Joathan Carl Sevy. Acceleration of convergence of sequences of simultaneous approximants. 1991.
[Sev95]
JC Sevy. Lagrange and least-squares polynomials as limits of linear combinations of
iterates of Bernstein and Durrmeyer polynomials. Journal of Approximation Theory,
80(2):267–271, 1995.
[SGO71]
WR Schucany, HL Gray, and DB Owen. On bias reduction in estimation. Journal of
the American Statistical Association, 66(335):524–533, 1971.
[Sha48]
Claude Elwood Shannon. A mathematical theory of communication. The Bell System
Technical Journal, 27:379–423, 623–656, 1948.
[ST77]
LI Strukov and AF Timan. Mathematical expectation of continuous functions of random
variables. smoothness and variance. Siberian Mathematical Journal, 18(3):469–474,
1977.
[ST12]
Jun Shao and Dongsheng Tu. The jackknife and bootstrap. Springer Science & Business
Media, 2012.
[SV16]
Igal Sason and Sergio Verdú. f -divergence inequalities. IEEE Transactions on Information Theory, 62(11):5973–6006, 2016.
[SW89]
Jun Shao and CF Jeff Wu. A general theory for jackknife variance estimation. The
Annals of Statistics, pages 1176–1197, 1989.
[Tim14]
Aleksandr Filippovich Timan. Theory of approximation of functions of a real variable,
volume 34. Elsevier, 2014.
[Tot88]
V Totik. Uniform approximation by exponential-type operators. Journal of mathematical analysis and applications, 132(1):238–246, 1988.
[Tot94]
Vilmos Totik. Approximation by Bernstein polynomials. American Journal of Mathematics, 116(4):995–1018, 1994.
[Tsy08]
A. Tsybakov. Introduction to Nonparametric Estimation. Springer-Verlag, 2008.
55
[VV11]
Gregory Valiant and Paul Valiant. The power of linear estimators. In Foundations of
Computer Science (FOCS), 2011 IEEE 52nd Annual Symposium on, pages 403–412.
IEEE, 2011.
[VV13]
Paul Valiant and Gregory Valiant. Estimating the unseen: improved estimators for
entropy and other properties. In Advances in Neural Information Processing Systems,
pages 2157–2165, 2013.
[Wit87]
Christopher Stroude Withers. Bias reduction by Taylor series. Communications in
Statistics-Theory and Methods, 16(8):2369–2383, 1987.
[WY16]
Yihong Wu and Pengkun Yang. Minimax rates of entropy estimation on large alphabets via best polynomial approximation. IEEE Transactions on Information Theory,
62(6):3702–3720, 2016.
[ZI04]
Abdelhak M Zoubir and D Robert Iskander. Bootstrap techniques for signal processing.
Cambridge University Press, 2004.
56
| 10math.ST
|
arXiv:1802.10178v1 [math.AC] 27 Feb 2018
ON THE CONTAINMENT PROBLEM FOR FAT POINTS
IMAN BAHMANI JAFARLOO
GIUSEPPE ZITO
Abstract. Given an ideal I, the containment problem is concerned about finding
the values m and n such that the m-th symbolic power of I is contained in its n-th
ordinary power. In this paper we consider this problem focusing on two classes of
ideals of fat points. In particular we study the ideal of n points on a line in PN
and the ideal of three nonlinear points in PN for N ≥ 2.
1. introduction
Let us consider the polynomial ring K[x0 , . . . , xN ] = K[PN ] = R, where K is an
algebraically closed field of any characteristic. In general,
T if I is a homogeneous
ideal of R, the m-th symbolic power of I is I (m) = R ∩ ( p∈ASS(I) (I m Rp )). However
in this
T paper we will always deal with ideals of fat points that are ideals of the form
I = i I(Pi )mi , where Pi are distinct points in PN , I(Pi ) is the ideal of all the forms
that vanish at Pi and the multiplicity mi is a non-negative integer. For
T ideals of
this type, the m-th symbolic power can be simply defined as I (m) = i I(Pi )mmi .
During the last decades, there has been a lot of interest comparing powers of ideals
with symbolic powers in various ways; see for example, [Hoc73], [Swa00], [Kod00],
[ELS01], [HH02], and [LS06]. It is easy to see that I r ⊆ I (r) ⊆ I (m) if and only if
r ≥ m. Furthermore I (m) ⊆ I r implies m ≥ r but the converse is not true in general.
Therefore it makes sense to ask the containment question: given an ideal I, for which
m and r is the symbolic power I (m) contained in the ordinary power I r ? In [BH10b]
and [BH10a], Bocci and Harbourne introduced and studied an asymptotic quantity,
known as the resurgence, whose computation is clearly linked to our containment
problem. Specifically, given an ideal I of fat points
in R, the resurgence of I, denoted
by ρ(I), is defined as the quantity ρ(I) = sup mr : I (m) 6⊆ I r .
From the results of [ELS01] and [HH02], it follows that I (m) ⊆ I r whenever
m ≥ Nr. Thus, we can conclude that ρ(I) ≤ N for any homogeneous ideal in R.
In general, directly computing ρ(I) is quite difficult and it has been determined
only in very special cases. For example, it is known that ρ(I) = 1 when I is
generated by a regular sequence [BH10a]. The resurgence is also known for some
kind of monomial ideals [GHM13, Theorem 4.11] or for the ideal of a projective cone
[BH10a, Proposition 2.5.1].
Somewhat more is known for ideals I defining zero-dimensional subschemes of
projective space. For example, if I defines a zerodimensional subscheme such that
2010 Mathematics Subject Classification. 14N20, 13A15, 13F20.
Key words and phrases. Containment, fat points scheme, symbolic powers, resurgence.
1
2
I. BAHMANI JAFARLOO
G. ZITO
α(I) = reg(I), where reg(I) is the Castelnuovo-Mumford regularity of I and α(I)
is the degree of a nonzero element of I of least degree, then the resurgence can be
completely described in terms of numerical invariants of I ( [BH10a, Corollary 2.3.7]
and [BH10b, Corollary 1.2]). One of these invariants is the Waldschmidt constant
(m)
. In particular, in [BH10b, Theorem 1.2]
of I that is α̂(I(Z)) = limm→∞ α(I(Z))
m
α(I(Z))
the authors proved that ρ(I(Z)) ≥ α̂(I(Z)) .
T
P
We denote by Z = ni=1 mi Pi ⊆ PN the subscheme defined by I(Z) = i I(Pi )mi ,
where the Pi is distinct point in PN . In this paper we compute the resurgence of
I(Z) P
for some classes of fat point subschemes. In section 2, we study the subscheme
Z = ni mi Pi in PN , where the points Pi belong to the same line and give a precise
answer to the value of the resurgence.
Theorem 1.1. Let Z be a fat points scheme, where P1 , . . . , Pn are distinct collinear
points in PN . Then I(Z)(m) = I(Z)m for all m ∈ N, thus ρ(I(Z)) = 1.
The strategy used in the proof consists in splitting the ideal I(Z) into a product
of symbolic powers of complete intersection schemes, and using the fact, showed
in [ZS75, Lemma 5 and Theorem 2 of Appendix 6], that the ideal of a complete
intersection is such that I (m) = I m for all m, we manage to prove Theorem 1.1. In
section 3, we study the subschemes of the type Z = m1 P1 + m2 P2 + m3 P3 , where
the Pi are nonlinear points in PN and m1 ≤ m2 ≤ m3 . For this class of ideals
we find that the value of the resurgence ρ(I(Z)) depends on the assignment of the
multiplicities mi .
Theorem 1.2. Let P1 , P2 and P3 be nonlinear points in PN and m1 ≤ m2 ≤ m3 and
suppose that m1 + m2 > m3 and m1 + m2 + m3 is odd. If Z = m1 P1 + m2 P2 + m3 P3 ,
then
m1 + m2 + m3 + 1
ρ(I(Z)) =
.
m1 + m2 + m3
We split the ideal I(Z) in a proper way and using the results of previous section,
first
P we show that ρ(I(Z)) = 1 whenever m1 + m2 < m3 or m1 + m2 ≥ m3 where
mi is even (Propositions 3.2 and 3.6). Finally, using the bound given by the
inequality ρ(I(Z)) ≥ α(I(Z))
and the consequences of several lemmas, we manage to
α̂(I(Z))
prove Theorem 1.2.
Acknowledgement
This project started during the summer school PRAGMATIC 2017. The authors
would like to thank Enrico Carlini, Tai Huy Ha, Brian Harbourne and Adam Van
Tuyl, for giving very interesting lectures and for sharing open problems. They
also wish to thank all the organizers of PRAGMATIC 2017 for giving them the
opportunity to attend the school.
ON THE CONTAINMENT PROBLEM FOR FAT POINTS
3
2. Fat points on a line in PN
Let L be a line in PN . We
P choose {P1 , . . . , Pn } as a set of points which lie on L.
Consider the scheme Z = ni=1 mi Pi , where the multiplicities m1 ≤ m2 ≤ · · · ≤ mn
are nonnegative integers. In this section, we study the containment problem for this
case in order to find the value of the resurgence and we prove Theorem 1.1.
Proving the main theorem in this section, requires some lemmas in order to represent the case in a convenient way. The following lemma plays a significant role
throughout this section.
Lemma 2.1. Let F ∈ R be a homogeneous form. Then F ∈ hL, x2 , . . . , xN im , where
L ∈ K[x0 , x1 ] is a homogeneous linear form, if and only if it can be written as
F =
d i2 +···+i
XN =k
X
gk,i2 ,...,iN · xi22 · · · xiNN
(1)
i2 ,...,iN
k=0
m−k
where gk,i2 ,...,iN ∈ K[x0 , x1 ] and L
|gk,i2,...,iN for all k ≤ m and i2 , . . . , iN .
′
Proof. Let F ∈ hL, x2 , . . . , xN im be a degree d form as (1), where gk,i
is ho2 ,...,iN
mogeneous forms of degree d − k in K[x0 , x1 ]. Consider a linear form H ∈ K[x0 , x1 ]
such that H ∈
/ hL, x2 , . . . , xN i. Hence, K[H, L][x2 , . . . , xN ] = R and it follows that
R has a basis, as a K-vector space, which consists of all monomials H i0 Li1 xi22 · · · xiNN
generated by H, L, x2 , . . . , xN . We conclude that hL, x2 , . . . , xN im is a K-vector
space spanned by the monomials H i0 Li1 xi22 . . . xiNN in K[H, L][x2 , . . . , xN ], such that
PN
j=1 ij ≥ m. Assume L = ax0 + bx1 , H = cx0 + dx1 such that ad − bc 6= 0.
We suppose that x0 = αL + βH and x1 = γL + δH, therefore we can write
′
gk,i
(x0 , x1 ) = gk,i2,...,iN (L, H) with same degree. Thus,
2 ,...,iN
F =
d i2 +···+i
XN =k
X
k=0
Given
that F
n
the basis H Li1 xi22
∈ hL, x2 , . . . , xN im if
i0
gk,i2,...,iN (L, H) · xi22 · · · xiNN .
i2 ,...,iN
. . . xiNN
:
o
for hL, x2 , . . . , xN im , it follows
j=1 ij ≥ m
PN
and only if Lm−k |gk,i2,...,iN ∀k ≤ m, i2 , . . . , iN .
Remark 1. Considering the previous proof, since hL = ax0 + bx1 , x2 , . . . , xN i is the
ideal of the point P = [−b : a : 0 : · · · : 0], indeed, we showed that
F ∈ I(mP ) if and only if (ax0 + bx1 )m−k |gk,i2,...,iN ∀ k ≤ m, ∀i2 , . . . , iN .
The following corollary is immediately deduced from the previous lemma.
Corollary 2.2. Given distinct points Pi = [−di : ci : 0 : · · · : 0], i = 1,
P. n. . , n, on
the line x2 = x3 = . . . = xN = 0, let F be a form as (1). Then F ∈ I( i=1 mi Pi )
if and only if (ci x0 + di x1 )mi −k |fk,i2 ,...,iN ∀ i = 1, . . . , n, k ≤ mi and ∀i2 , . . . , iN .
InP
other words, we have shown that a set of generators of the homogeneous ideal
I( ni=1 mi Pi ) consists of “monomials” of the type La11 · · · · Lann · xb22 · · · xbNN , where
Lj = cj x0 + dj x1 and aj + b2 + · · · + bN ≥ mj , j = 1, .., n.
4
I. BAHMANI JAFARLOO
G. ZITO
All things considered, we are ready to prove the main theorem of this section.
Proof of Theorem 1.1.
We can assume that the points lie on the line x2 = x3 = . . . = xN = 0 and
m1 ≤ · · · ≤ mn . We denote by Li the linear form in K[x0 , x1 ] such that we have
I(Pi ) = hLi , x2 , . . . , xN i for all i = 1, . . . , n. Furthermore, we denote the fat points
scheme Zi = (mi − mi−1 )(Pi + · · · + Pn ) for i = 1, . . . , n and we define m0 = 0.
Notice that
I(Zi )(m) = I(Zi )m for all m.
(2)
In fact, the ideal I(Zi ) is a power of a complete intersection scheme (a set of simple
point on a line), and by [ZS75, Lemma 5 and Theorem 2 of Appendix 6], its
Q symbolic
powers and regular powers are always equal. Now, we claim that I(Z) = ni=1 I(Zi ).
Let us to prove the claim: the inclusion “ ⊇ ” is immediately concluded from the
definition of I(Z). For proving the
Qn other inclusion “ ⊆ ”, it suffices to consider
Corollary 2.2 and show that M ∈ i=1 I(Zi ) for all monomials of the following type
M=
La11
·
· · · Lann
·
xb22
· · · xbNN
∈ I(Z) iff aj +
N
X
bi ≥ mj , ∀1 ≤ j ≤ n.
(3)
i=2
By using the definition of Zi , it is easy to see for all 1 ≤ i ≤ n,
a′
L11
·
′
· · · Lann
·
b′
x22
b′
· · · xNN
∈ I(Zi ) iff
a′j
+
N
X
b′t ≥ mi − mi−1 , ∀i ≤ j ≤ n.
(4)
t=2
Thus, our goal is splitting the monomial M ∈ I(Z) in a proper way in order to
Q
P
show that M ∈ ni=1 I(Zi ). Set b = N
i=2 bi and consider the following cases:
2 −b
n −b b2
· · · Lm
x2 · · · xbNN divides
• Case b ≤ m1 : by (3) the polynomial P = L1m1 −b Lm
2
n
M,
mi − b ≤ ai forQall i. Furthermore P can be written as,
Qn since
mn−1 −mn−2 mn−1 −mn−2
m1 −b QN
n−1
Ln
)Lnmn −mQ
,
· i=2 xbi i )( ni=2 Lim2 −m1 ) · · · (Ln−1
( i=1 Li
where the j-th part of the last product is in I(Zj ) by (4). Thus M ∈ ni=1 I(Zi ).
• Case mi−1 ≤ b ≤ mi : since b ≥ mi−1 ≥ m1 , it can be found integers c1,i for
P
i = 2, . . . , N, such that 0 ≤ c1,i ≤ bi and N
i=2 c1,i = m1 .
P
c
≥
m
−
m
≥
m
Because b − N
i−1
1
2 − m1 , we can find some integers c2,i
i=2 1,i
PN
for i = 2, . . . , N such that 0 ≤ c2,i ≤ bi − c1,i and
i=2 c2,i = m2 − m1 . Following this way, for k ≤ i − 1, there exist integers ck,i for i = 2, . . . , N such that
Pk−1
P
0 ≤ ck,i ≤ bi − j=1
cj,i and N
i=2 ck,i = mk − mk−1 . The condition (3) implies that
mi+1 −b
n −b
mi − b ≤ ai for all i. So, the polynomial Q = Limi −b Li+1
· · · Lm
· xb22 · · · xbNN
n
divides M. Furthermore Q can be written as:
(
N
Y
t=2
c
xt 1,t ) · · · (
N
Y
t=2
c
xt i−1,t )(
n
Y
t=i
Ltmi −b
·
N
Y
t=2
bt −
xt
Pi−1
j=1 cj,t
)···(
n
Y
mn−1 −mn−2
Lt
n −mn−1
)Lm
,
n
t=n−1
Q
where the j-th part of the product is in I(Zj ) by (4). Thus M ∈ ni=1 I(Zi ).
• Case mn ≤ b: same as the previous case, P
for each k ≤ n,
P we can find integers
k−1
ck,i ∀ i = 2, . . . , N such that 0 ≤ ck,i ≤ bi − j=1
cj,i and N
i=2 cki = mk − mk−1 .
cn−1,N
c
cn−1,2
c
c1N
c2N
c12
c22
The monomial N = (x2 · · · xN )(x2 · · · xN ) · · · (x2
· · · xN
)·(x2n,2 · · · xNn,N )
ON THE CONTAINMENT PROBLEM FOR FAT POINTS
5
Q
divides M. Furthermore, N ∈ ni=1 I(Zi ), since by (4), its j-th factor belongs to
I(Zj ). Thus theQ
proof of the claim
By previous claim and (2) we
Qn is completed.
n
(m)
m
m
= i=1 I(Zi ) . Then by applying the claim to the
obtain I(Z) Q= i=1 I(Zi ) Q
n
(m)
scheme mZ, i=1 I(Zi )
= ni=1 I(mZi ) = I(mm1 P1 + · · · + mmn Pn ) = I(Z)(m) .
Hence, we conclude I(Z)(m) = I(Z)m for all m ∈ N.
3. Three nonlinear points in PN
P3
N
In this section we study the fat point schemes Z =
i=1 mi Pi in P , where
P1 , P2 and P3 are not collinear and mi is nonnegative integer. We can assume
P1 = [0 : 0 : 1 : 0 : · · · : 0], P2 = [0 : 1 : 0 : · · · : 0], P3 = [1 : 0 : 0 : · · · : 0] and
m1 ≤ m2 ≤ m3 . Notice that ideal I(Pi ) is a square-free monomial ideal. We are
interested in computing the resurgence ρ(I(Z)) of the ideal I(Z). In particular, we
want to understand how the resurgence of scheme Z depends on the values of the
multiplicities mi .
The following lemma gives some conditions for a monomial to belong to I(Z).
Lemma 3.1. Let P1 , P2 and P3 be nonlinear points in PN and mi ≥ 0. We
define the fat points scheme Z = m1 P1 + m2 P2 + m3 P3 . Then the monomial
N = xa00 xa11 · · · xaNN ∈ I(Z) if and only if (a0 , . . . , aN ) satisfies the following system of inequalities
y0 + y1 + y3 + · · · + yN ≥ m1
Cond(Z) := y0 + y2 + y3 + · · · + yN ≥ m2
(5)
y + y + y + · · · + y ≥ m .
1
2
3
N
3
Proof. The thesis easily follows from the fact that the ideal I(Z) is,
T3
mi
= hx0 , x1 , x3 , . . . , xN im1 ∩ hx0 , x2 , x3 , . . . , xN im2 ∩ hx1 , x2 , x3 , . . . , xN im3
i=1 I(Pi )
which is an intersection of powers of square-free monomial ideals.
Notice that in previous lemma, in order to simplify the notation, we made implicit
the dependence of Cond(Z) on m1 , m2 , m3 . We divide this section in two subsections
where we study distinct configuration for the multiplicities mi .
3.1. Case m1 +m2 < m3 . The aim of this subsection is to prove the following result.
Proposition 3.2. Let P1 , P2 and P3 be nonlinear points in PN and m1 ≤ m2 ≤ m3 .
Let Z = m1 P1 + m2 P2 + m3 P3 be a fat points scheme. If m1 + m2 < m3 , Then
I(Z)(m) = I(Z)m for all m ∈ N and consequently ρ(I(Z)) = 1.
In order to prove Proposition 3.2, we want to show that the ideal I(Z) can be
seen as a product of ideals I(Zi ), where Zi is a scheme of fat points on a line which
is easier to study by means of the results obtained in the previous section. The
following lemma helps us in this direction.
6
I. BAHMANI JAFARLOO
G. ZITO
Lemma 3.3. Let P1 , P2 and P3 be nonlinear points in PN and m1 ≤ m2 ≤ m3 .
Consider fat points scheme Z = m1 P1 + m2 P2 + m3 P3 . If m1 + m2 < m3 , then
I(Z) = I(m1 (P1 + P3 )) · I(m2 (P2 + P3 )) · I((m3 − m1 − m2 )P3 ).
Proof. Set Z1 = m1 (P1 + P3 ), Z2 = m2 (P2 + P3 ) and Z3 = (m3 − m1 − m2 )P3 . The
inclusion I(Z1 ) · I(Z2 ) · I(Z3 ) ⊆ I(Z) is trivial from the definition. Now, we show the
other inclusion holds. Thus, let us consider a monomial N = xa00 xa11 · · · xaNN ∈ I(Z)
P
and ai satisfies the system Cond(Z), and set b = N
i=3 ai . We have the following
cases:
• Case a1 + b < m1 : by Cond(Z) it follows that a0 ≥ m1 − a1 − b > 0 and
a2 ≥ m3 − a1 − b = (m1 − a1 − b) + m2 + (m3 − m1 − m2 ). Then, the monomial
m3 −m1 −m2
2
(x0m1 −a1 −b xa11 x2m1 −a1 −b xa33 · · · xaNN ) · (xm
)
2 ) · (x2
divides N and belongs to I(Z1 ) · I(Z2 ) · I(Z3 ) because the j-th part of the above
product is in I(Zj ) by Cond(Zj ) for j = 1, 2, 3.
• Case a1 + b ≥ m1 and a2 + b < m2 : proof is similar to the previous case by using
a2 + b < m2 .
• Case a1 + b ≥ m1 and a2 + b ≥ m2 : we have four subcases
◮ Subcase a1 ≥ m1 and a2 ≥ m2 : we can write N as
m2
a0 a1 −m1 a2 −m2 a3
1
(xm
x2
x3 · · · xaNN ).
1 ) · (x2 ) · (x0 x1
The first two factors belong respectively to I(Z1 ) and I(Z2 ) while the third one is in
I(Z3 ) because Cond(Z) implies a1 + a2 + a3 + · · · + aN − m1 − m2 ≥ m3 − m1 − m2 .
Hence, Cond(Z3 ) are satisfied.
P
◮ Subcase a1 < m1 and a2 ≥ m2 : by a1 + b ≥ m1 , we deduce N
ai ≥ m1 − a1 .
Pi=3
N
Then, for each i = 3, . . . , N, we can choose 0 ≤ bi ≤ ai such that i=3 bi = m1 − a1 .
a0 a2 −m2 a3 −b3
2
It can be written N = (xa11 xb33 · · · xbNN ) · (xm
x3
· · · xaNN −bN ), where
2 ) · (x0 x2
it is easy to check the first two factors belong respectively to I(Z1 ) and I(Z2 ) while
the third term is in I(Z3 ) because Cond(Z) implies that
PN
P
a2 + a3 + · · · + aN − m2 − N
i=1 ai − m1 − m2 ≥ m3 − m1 − m2 . So,
i=3 bi =
Cond(Z3 ) are satisfied.
◮ Subcase a1 ≥ m1 and a2 < m2 : proof of this case is similar to the proof of previous
one.
◮ Subcase a1 < m1 and a2 < m2 : Cond(Z) implies that
N
X
ai = b = (a1 + a2 + b − m1 − m2 ) + (m1 − a1 ) + (m2 − a2 )
i=3
≥ (m3 − m1 − m2 ) + (m1 − a1 ) + (m2 − a2 ).
Because the three last summands are all positive we can choose for all i = 3, . . . , N,
some integers 0 ≤ ci , di , ei ≤ ai such that
(
PN
d = m2 − a2
• ci + di + ei ≤ ai ∀ i = 3, . . . , N
•
PN i=3 i
PN
•
•
i=3 ei = m3 − m1 − m2 .
i=3 ci = m1 − a1
ON THE CONTAINMENT PROBLEM FOR FAT POINTS
7
So, the monomial M = (xa11 xc33 · · · xcNN ) · (xa22 xd33 · · · xdNN ) · (xe3i · · · xeNN ) divides N and
belongs to I(Z1 ) · I(Z2 ) · I(Z3 ) because its j-th factor belongs to I(Zj ) by Cond(Zj )
for j = 1, 2, 3.
Now we can give a proof of Proposition 3.2.
Proof of Proposition 3.2. By Lemma 3.3 and Theorem 1.1, it follows that
I(Z)m =I(m1 (P1 + P3 ))m · I(m2 (P2 + P3 ))m · I((m3 − m1 − m2 )P3 )m
=I(m1 (P1 + P3 ))(m) · I(m2 (P2 + P3 ))(m) · I((m3 − m1 − m2 )P3 )(m)
=I(mm1 (P1 + P3 )) · I(mm2 (P2 + P3 )) · I(m(m3 − m1 − m2 )P3 ).
Now, by applying Lemma 3.3 to scheme Z ′ = mm1 P1 + mm2 P2 + mm3 P3 ,
I(Z)(m) = I(Z ′ ) = I(mm1 (P1 + P3 )) · I(mm2 (P2 + P3 )) · I(m(m3 − m1 − m2 )P3 ).
So, the proof is complete.
3.2. Case m1 + m2 ≥ m3 . In this subsection, we deal with the case m1 + m
P2 ≥ m3
showing how the value of the resurgence depends on the parity of the sum 3i=1 mi .
Using the same approach of the previous subsection, we want to split in a convenient
way the ideal I(Z) as a product of ideals I(Zi ).
Lemma 3.4. Let P1 , P2 and P3 be nonlinear points in PN and m1 ≤ m2 ≤ m3 . We
consider scheme Z = m1 P1 + m2 P2 + m3 P3 . If m1 + m2 ≥ m3 , then
I(Z) = I((m1 +m2 −m3 )(P1 +P2 +P3 ))·I((m3 −m2 )(P1 +P3 ))·I((m3 −m1 )(P2 +P3)).
Proof. Set Z1 = (m1 + m2 − m3 )(P1 + P2 + P3 ), Z2 = (m3 − m2 )(P1 + P3 ) and
Z3 = (m3 − m1 )(P2 + P3 ). The inclusion I(Z1 ) · I(Z2 ) · I(Z3 ) ⊆ I(Z) is trivial by
the definition. We just need to show that if a monomial N = xa00 xa11 · · · xaNN ∈ I(Z),
P
then N ∈ I(Z1 ) · I(Z2 ) · I(Z3 ). Thus, suppose N ∈ I(Z), and set b = N
i=3 ai . We
have the following cases:
• Case a1 + b < m3 − m2 : considering the system Cond(Z)
(
a0 ≥ m1 − a1 − b = (m1 + m2 − m3 ) + (m3 − m2 − a1 − b)
a2 ≥ m3 − a1 − b = (m1 + m2 − m3 ) + (m3 − m2 − a1 − b) + (m3 − m1 ),
where all the numbers between parenthesis are nonnegative.
So, the
monomial
M = ((x0 x2 )m1 +m2 −m3 ) · (x0 x2 )m3 −m2 −a1 −b xa11 xa33 · · · xaNN · x2m3 −m1 divides N .
Furthermore, M belongs to I(Z1 ) · I(Z2 ) · I(Z3 ) because the j-th factor belongs to
I(Zj ) by Cond(Zj ) for j = 1, 2, 3. Thus N ∈ I(Z1 ) · I(Z2 ) · I(Z3 ).
• Case a1 + b ≥ m3 − m2 and a2 + b < m3 − m1 : proof is similar to previous case
using a2 + b < m3 − m1 .
• Case a1 + b ≥ m3 − m2 and a2 + b ≥ m3 − m1 : we have four subcases,
◮ Subcase a1 ≥ m3 − m2 and a2 ≥ m3 − m1 : we can write
8
I. BAHMANI JAFARLOO
G. ZITO
N = (xa00 xa11 −m3 +m2 xa22 −m3 +m1 xa33 · · · xaNN ) · (x1m3 −m2 ) · (x2m3 −m1 ) where the first factor
is in I(Z1 ) because Cond(Z) implies
a0 + a1 + a3 + · · · + aN − m3 + m2 ≥ m1 + m2 − m3
a0 + a2 + a3 + · · · + aN − m3 + m1 ≥ m1 + m2 − m3
a + a + a + · · · + a − 2m + m + m ≥ m + m − m .
1
2
3
N
3
1
2
1
2
3
So, Cond(Z1 ) are satisfied. Furthermore, it is easy to check that x1m3 −m2 ∈ I(Z2 )
and x2m3 −m1 ∈ I(Z3 ). Thus N ∈ I(Z1 ) · I(Z2 ) · I(Z3 ).
◮ Subcase a1 ≥ m3 − m2 and a2 < m3 − m1 :
PN
a2 + b ≥ m3 − m1 implies b ≥ m3 − m1 − a2 > 0 then
ai ≥ m3 − m1 − a2 . For
Pi=3
N
each i = 3, . . . , N we can choose 0 ≤ bi ≤ ai such that i=3 bi = m3 − m1 − a2 . We
can write N = (xa00 xa11 −m3 +m2 x3a3 −b3 · · · xaNN −bN ) · (x1m3 −m2 ) · (xa22 xb33 · · · xbNN ), where
the first factor is in I(Z1 ) because by Cond(Z), it follows
PN
P
b − m3 + m2 − N
a0 + a1 +P
i=0 ai − 2m3 + m1 + m2 ≥ m1 + m2 − m3
i=3 bi =
PN
N
a0 + b − i=3 bi = a0 + i=2 ai + m1 − m3 ≥ m1 + m2 − m3
a + b + m − m − PN b = PN a + m + m − 2m ≥ m + m − m .
1
2
3
1
2
3
1
2
3
i=1 i
i=3 i
So, the conditions at Cond(Z1 ) are satisfied. As we have seen in the previous subcase,
the second factor belongs to I(Z2 ). Furthermore, it is easy to check, using Cond(Z3 ),
that xa22 xb33 · · · xbNN ∈ I(Z3 ). Thus N ∈ I(Z1 ) · I(Z2 ) · I(Z3 ).
◮ Subcase a1 < m3 − m2 and a2 ≥ m3 − m1 : proof is similar to the previous one.
◮ Subcase a1 < m3 − m2 and a2 < m3 − m1 : by Cond(Z), we have
N
X
ai = b = (a1 + a2 + b + m1 + m2 − 2m3 ) + (m3 − m2 − a1 ) + (m3 − m1 − a2 )
i=3
≥ (m1 + m2 − m3 ) + (m3 − m2 − a1 ) + (m3 − m1 − a2 ).
Because the last three summands are all positive, it is possible to choose for all
i = 3, . . . , N some integers 0 ≤ ci , di , ei ≤ ai such that
(
PN
• ci + di + ei ≤ ai ∀ i = 3, . . . , N •
c = m1 + m2 − m3
PNi=3 i
PN
• i=3 ei = m3 − m1 − a2 .
•
i=3 di = m3 − m2 − a1
By Cond(Zi ), it follows that M = (xc33 · · · xcNN ) · (xa11 xd33 · · · xdNN ) · (xa22 xe3i · · · xeNN ) belongs to I(Z1 )·I(Z2)·I(Z3 ). Since M divides N , we deduce N ∈ I(Z1 )·I(Z2)·I(Z3).
So, in all the possible cases, N ∈ I(Z1 ) · I(Z2 ) · I(Z3 ) and the proof of the lemma is
complete.
The next lemma helps us to deal with subschemes of the type (2q+r)(P1 +P2 +P3 )
that appeared as a factor in the splitting presented in Lemma 3.4.
Lemma 3.5. Let P1 , P2 and P3 be nonlinear points in PN . If q, r ∈ N with 0 ≤ r < 2,
then I((2q + r)(P1 + P2 + P3 )) = I(2(P1 + P2 + P3 ))q · I(P1 + P2 + P3 )r .
ON THE CONTAINMENT PROBLEM FOR FAT POINTS
9
Proof. We give a proof by induction on q. For the base case q = 0, we need to
show that I(r(P1 + P2 + P3 )) = I(P1 + P2 + P3 )r . But this is trivial for r = 0, 1.
For the induction, we suppose that the lemma is true for q − 1 and we prove that it
holds for q. We claim that
I((2q + r)(P1 + P2 + P3 )) = I(2(P1 + P2 + P3 )) · I((2(q − 1) + r)(P1 + P2 + P3 )).
Proof of the claim. Set Z1 = (2q + r)(P1 + P2 + P3 ), Z2 = 2(P1 + P2 + P3 ) and
Z3 = (2(q − 1) + r)(P1 + P2 + P3 ). The inclusion I(Z2 ) · I(Z3) ⊆ I(Z1 ), is trivial from
the definition. Therefore, we show that if a monomial N = xa00 xa11 · · · xaNN ∈ I(Z1 )
P
then N ∈ I(Z2 ) · I(Z3 ). Thus, consider N ∈ I(Z1 ). Set b = N
i=3 ai . We have the
following cases:
•PCase b ≥ 2: for each i = 3, . . . , N, it can be chosen 0 ≤ bi ≤ ai such that
N
bN
b3
a0 a1 a2 a3 −b3
· · · xaNN −bN ), then we can
i=3 bi = 2. If we write N = (x3 · · · xN ) · (x0 x1 x2 x3
easily deduce by Cond(Z2 ) and Cond(Z3 ), that N ∈ I(Z2 ) · I(Z3 ).
• Case b = 1: we have three subcases.
◮ Subcase a0 = 0: by Cond(Z1 ) it follows that
a1 ≥ 2q + r − 1 ≥ 1 and a2 ≥ 2q + r − 1 ≥ 1.
(6)
Therefore, N = (x1 x2 xa33 · · · xaNN ) · (xa11 −1 xa22 −1 ), where it is easy to see that the first
factor is in I(Z2 ). So we need to show that the second one is in I(Z3 ). By Cond(Z1 )
it follows
(
aj − 1 ≥ 2(q − 1) + r
for j = 1, 2 by (6)
a1 + a2 − 2 ≥ 2q + r − 3
by Cond(Z1 ).
Notice that, we cannot have a1 +a2 −2 = 2q +r −3 which implies a1 +a2 = 2q +r −1.
In fact, by (6), it follows 4q + 2r − 2 ≤ a1 + a2 = 2q + r − 1 so 2q + r ≤ 1, which is a
contradiction. Then, also a1 + a2 − 2 ≥ 2(q − 1) + r and the conditions at Cond(Z3 )
are satisfied.
◮ Subcase a1 = 0 and a2 = 0: these cases are similar to the previous one.
◮ Subcase a0 , a1 , a2 > 0: we can write N = (x0 x1 x2 ) · (xa00 −1 xa11 −1 xa22 −1 xa33 · · · xaNN ),
where it is easy to prove that the two factors belong to I(Z2 ) and I(Z3 ) respectively.
• Case b = 0: we have the following subcases.
◮ Subcase a0 = 0: by Cond(Z1 ) it follows that
a1 ≥ 2q + r ≥ 2 and a2 ≥ 2q + r ≥ 2.
(7)
Therefore, N = (x21 x22 ) · (xa11 −2 xa22 −2 ), where it is immediately concluded the first
factor belongs to I(Z2 ), while xa11 −2 xa22 −2 ∈ I(Z3 ) because Cond(Z1 ) implies
(
aj − 2 ≥ 2(q − 1) + r
for j = 1, 2 by (7)
a1 + a2 − 4 ≥ 2q + r − 4
by Cond(Z1 ).
Notice that we cannot have a1 + a2 − 4 = 2q + r − 4 which implies a1 + a2 = 2q + r.
In fact, by (7), it follows that 4q + 2r ≤ a1 + a2 = 2q + r, thus 2q + r ≤ 0, which is
a contradiction. Similarly, we can show that the equality a1 + a2 − 4 = 2q + r − 3
does not hold. Then, also a1 + a2 − 4 ≥ 2(q − 1) + r and the conditions at Cond(Z3 )
are satisfied.
◮ Subcase a1 = 0 and a2 = 0: these cases are similar to the previous one.
10
I. BAHMANI JAFARLOO
G. ZITO
◮ Subcase a0 , a1 , a2 > 0: if we write N = (x0 x1 x2 ) · (xa00 −1 x1a1 −1 x2a2 −1 ), then we can
deduce N ∈ I(Z2 ) · I(Z3 ) by Cond(Z2 ) and Cond(Z3 ). So, the proof of the claim is
complete. By the inductive step
I((2q + r)(P1 + P2 + P3 )) = I(2(P1 + P2 + P3 )) · I((2(q − 1) + r)(P1 + P2 + P3 ))
= I(2(P1 + P2 + P3 )) · I(2(P1 + P2 + P3 ))q−1 · I(P1 + P2 + P3 )r ,
and proof of the lemma is complete.
Now, we can solve our main problem when
P3
i=1
mi is even.
Proposition 3.6. Let P1 , P2 and P3 be nonlinear points in PN and m1 ≤ m2 ≤ m3 .
Denote by Z the corresponding fat points scheme Z = m1 P1 + m2 P2 + m3 P3 . If
m1 + m2 ≥ m3 and m1 + m2 + m3 is even, then I(Z)(m) = I(Z)m for all m ∈ N and
ρ(I(Z)) = 1.
Proof. Since m1 + m2 + m3 is even, we can set m1 + m2 − m3 = 2q. By Lemma
3.4, Theorem 1.1 and Lemma 3.5:
I(Z)m = I(2q(P1 + P2 + P3 ))m I((m3 − m2 )(P1 + P3 ))m I((m3 − m1 )(P2 + P3 ))m
= I(2(P1 + P2 + P3 ))mq I((m3 − m2 )(P1 + P3 ))(m) I((m3 − m1 )(P2 + P3 ))(m)
= I(2mq(P1 + P2 + P3 ))I(m(m3 − m2 )(P1 + P3 ))I(m(m3 − m1 )(P2 + P3 )).
Now, by applying Lemma 3.4 to the fat point scheme Z ′ = mm1 P1 +mm2 P2 +mm3 P3 ,
we obtain
I(Z ′ ) = I(2mq(P1 + P2 + P3 )) · I(m(m3 − m2 )(P1 + P3 )) · I(m(m3 − m1 )(P2 + P3 )).
So, the proof is complete, since I(Z ′ ) = I(Z)(m) .
P
Let 3i=1 mi be odd. Our aim is proving
Theorem 1.2. We present
a proof divided
P
P
m
+1
m
+1
i
i
in two steps, showing that ρ(I(Z)) ≥ P mi and ρ(I(Z)) ≤ P mi . We find the
lower bound by using results from the literature, generalizing in a proper way the
case N = 2, while the upper bound is proved by directly considering the definition
of resurgence and using further preliminary lemmas on the splitting of the symbolic
powers. By Lemma 3.4 and Lemma 3.5, we can deduce the following corollary.
Corollary 3.7. Let P1 , P2 and P3 be nonlinear points in PN and m1 ≤ m2 ≤ m3 .
Denote by Z the corresponding fat points scheme Z = m1 P1 + m2 P2 + m3 P3 . If
m1 + m2 > m3 and m1 + m2 + m3 is odd, then
I(Z)(k) = I(P1 +P2 +P3 )(k) ·I((m1 −1)P1 +(m2 −1)P2 +(m3 −1)P3 )(k) for all k ∈ N.
Proof. Consider k ∈ N. We write k = 2q1 + r where r < 2. Since m1 + m2 + m3
is odd, it follows m1 + m2 − m3 is odd and we can write m1 + m2 − m3 = 2q2 + 1.
ON THE CONTAINMENT PROBLEM FOR FAT POINTS
11
Set Z1 = P1 + P2 + P3 . Because km1 + km2 = k(m1 + m2 ) > km3 , we can apply
Lemma 3.4 to the scheme kZ and we obtain that I(Z)(k) is equal to
I(k(m1 + m2 − m3 )Z1 )I(k(m3 − m2 )(P1 + P3 ))I(k(m3 − m1 )(P2 + P3 ))
= I((2(2q1 q2 + q1 + rq2 ) + r)Z1 )I(k(m3 − m2 )(P1 + P3 ))I(k(m3 − m1 )(P2 + P3 ))
= I(2Z1 )2q1 q2 +q1 +rq2 I(Z1 )r I(k(m3 − m2 )(P1 + P3 ))I(k(m3 − m1 )(P2 + P3 )),
where the last equality holds by Lemma 3.5. By Lemma 3.5:
I(Z1 )(k) = I((2q1 + r)Z1 ) = I(2Z1 )q1 I(Z1 )r .
By applying Lemma 3.4 to the scheme Z ′ = k(m1 −1)P1 +k(m2 −1)P2 +k(m3 −1)P3
(we can use it because k(m1 − 1) + k(m2 − 1) = k(m1 + m2 − 2) ≥ k(m3 − 1) since
m1 + m2 ≥ m3 + 1).
I(Z ′ ) = I((m1 − 1)P1 + (m2 − 1)P2 + (m3 − 1)P3 )(k)
= I(k(m1 + m2 − m3 − 1)Z1 )I(k(m3 − m2 )(P1 + P3 ))I(k(m3 − m1 )(P2 + P3 ))
= I((2(2q1 q2 + rq2 )Z1 )I(k(m3 − m2 )(P1 + P3 ))I(k(m3 − m1 )(P2 + P3 ))
= I(2Z1 )2q1 q2 rq2 I(k(m3 − m2 )(P1 + P3 ))I(k(m3 − m1 )(P2 + P3 )),
where the last equality holds by Lemma 3.5. Thus, using the previous equation, we
have
I(Z)(k) = I(P1 + P2 + P3 )(k) · I((m1 − 1)P1 + (m2 − 1)P2 + (m3 − 1)P3 )(k) .
Notice that in general the equality I(Z)(a+b) = I(Z)(a) · I(Z)(b) is not satisfied.
However, the previous results imply the following corollary which tells us when this
splitting is possible for I(Z).
Corollary 3.8. Let P1 , P2 and P3 be nonlinear points in PN and m1 ≤ m2 ≤ m3 .
Denote by Z the fat points scheme Z = m1 P1 + m2 P2 + m3 P3 . If m1 + m2 > m3 and
m1 + m2 + m3 is odd, then
(
I(Z)(k) = I(Z)(2i) · I(Z)(k−2i) for 1 ≤ i ≤ k2 − 1 if k is even
I(Z)(k) = I(Z)(i) · I(Z)(k−i) for 1 ≤ i ≤ k − 1 if k is odd.
Proof. • Case k even: suppose that k = 2q. Then I(Z)(2q) = I(2Z)(q) , where 2Z is
a fat points scheme that satisfies the condition of the Proposition 3.6. Therefore
I(2Z)(q) = I(2Z)q = I(2Z)i · I(2Z)q−i = I(2Z)(i) · I(2Z)(q−i) = I(Z)(2i) · I(Z)(k−2i) .
• Case k odd: suppose that k = 2q + 1. By Proposition 3.6, Lemma 3.5, Corollary
3.7 and the even case, it follows that
I(Z)(2q+1) = I(P1 + P2 + P3 )(2q+1) · I((m1 − 1)P1 + (m2 − 1)P2 + (m3 − 1)P3 )(2q+1)
= I(P1 + P2 + P3 )(2q) I(P1 + P2 + P3 )I((m1 − 1)P1 + (m2 − 1)P2 + (m3 − 1)P3 )2q+1
= I(Z) · I(Z)(2q) = I(Z) · I(Z)(2i) · I(Z)(2q−2i) = I(Z)(2i+1) · I(Z)(2q−2i)
and the desired result follows.
12
I. BAHMANI JAFARLOO
G. ZITO
The following proposition and corollary are proved in P2 in order to generalize
them for any N.
Proposition 3.9. Let P1 , P2 and P3 be three nonlinear points in P2 and consider
Z = m1 P1 + m2 P2 + m3 P3 . Without loss of generality we assume P1 = [0 : 0 : 1],
P2 = [0 : 1 : 0], P3 = [1 : 0 : 0]. PSuppose m1 ≤ m2 ≤ m3 , if m1 + m2 > m3 and
3
P3
i=1 mi +1
P
.
3
i=1 mi is odd. Then ρ(I(Z)) ≥
m
i=1
i
Proof. We want to use the following inequality
ρ(I(Z)) ≥ α(I(Z))/α̂(I(Z)),
(8)
which was proved in [BH10b, Theorem 1.2]. To compute α(I(Z)), we focus on the
degree of the monomials xa y bz c in I(Z) (because I(Z) is a monomial ideal). A
monomial xa y b z c belongs to I(Z) if and only if a + b ≥ m1 , a + c ≥
m , b + c ≥ m3 .
P3 2
P3
mi
a b c
y z ) =ma + b + c ≥ i=1
It follows 2(a + b + c) ≥ i=1 mi . So, deg(xl
. Then, all
2
P3
P
m
3
i
i=1
monomials in I(Z) have degree greater than
. Since i=1 mi is odd, hence
2
m P3
l P3
m +1
i=1 mi
= i=12 i . Thus, we proved that
2
α(I(Z)) ≥ (
3
X
mi + 1)/2.
(9)
i=1
Now, consider m ∈ N and the 2m-th symbolic power I(Z)(2m) . Considering the
definition of symbolic powers,
I(Z)(2m) = hx, yi2m·m1 ∩ hx, zi2m·m2 ∩ hy, zi2m·m3 .
Then xa y bz c ∈ I(Z)(2m) if and only if
a + b ≥ 2m · m1 , a + c ≥ 2m · m2 , b + c ≥ 2m · m3 .
(10)
Hence
xm(m1 +m2 −m3 ) y m(m1 +m3 −m2 ) z m(m2 +m3 −m1 ) ∈ I(Z)(2m) .
(11)
The monomial in (11) has degree m(m1 + m2 + m3 ).
So, α(I(Z)(2m) ) ≤ m(m1 + m2 + m3 ) for all m ∈ N. Notice also that by (10), if
P
xa y b z c ∈ I(Z)(2m) then deg(xa y bz c ) = a + b + c ≥ m( 3i=1 mi ), and we conclude
P
α(I(Z)(2m) ) = m( 3i=1 mi ). Thus we obtain the Waldschmidt constant of I(Z) as
follow:
P
P3
α(I(Z))(2m)
m( 3i=1 mi )
mi
α(I(Z))(m)
= lim
= lim
= i=1 .
α̂(I(Z)) = lim
m→∞
m→∞
m→∞
m
2m
2m
2
3
X
mi )/2.
(12)
α̂(I(Z)) = (
i=1
Eventually by (9), (12) and (8),
P3
α(I(Z))
i=1 mi + 1
≤
≤ ρ(I(Z)),
P
3
α̂(I(Z))
i=1 mi
ON THE CONTAINMENT PROBLEM FOR FAT POINTS
and the desired result is given.
13
As a consequence of the results which were proved in [BH10b, Theorem 3.4], we
can deduce the following corollary for three simple points in P2 .
Corollary 3.10. Let P1 = [0 : 0 : 1], P2 = [0 : 1 : 0], P3 = [1 : 0 : 0]. Then
ρ(I(P1 + P2 + P3 )) = 4/3.
From the previous corollary we can deduce the following useful lemma.
Lemma 3.11. Let P1 , P2 and P3 be nonlinear points in PN . Then
I(P1 + P2 + P3 )(r) ⊆ I(P1 + P2 + P3 )r−1 for 1 ≤ r ≤ 4.
Proof. We work by induction on r. It is trivial for r = 1 . For the induction
suppose that it is true for r − 1 and we prove it for r.
Consider N = xa00 xa11 · · · xaNN ∈ I(P1 + P2 + P3 )(r) then
a0 + a1 + a3 + · · · + aN ≥ r
(13)
a0 + a2 + a3 + · · · + aN ≥ r
a + a + a + · · · + a ≥ r.
1
2
3
N
PN
Set b = i=3 ai . We have the following cases:
• Case b = 0: we can see the monomial N as an element of the ideal
I(P1∗ + P2∗ + P3∗ )(r) , where P1∗ = [0 : 0 : 1], P2∗ = [0 : 1 : 0], P3∗ = [1 : 0 : 0]
in P2 . In fact, it is easy to see that a monomial xa00 xa11 xa22 in K[x0 , x1 , x2 ] is in
I(P1∗ + P2∗ + P3∗ )(r) if and only if it belongs to I(P1 + P2 + P3 )(r) . By Corollary 3.10,
ρ(I(P1∗ + P2∗ + P3∗ )) = 4/3. Furthermore, r ≤ 4 implies 4r − 4 ≤ 3r, therefore
r/(r − 1) ≥ 4/3 = ρ(I(P1∗ + P2∗ + P3∗ )). Then, using the definition of resurgence
I(P1∗ + P2∗ + P3∗)(r) ⊆ I(P1∗ + P2∗ + P3∗ )r−1 and N ∈ I(P1∗ + P2∗ + P3∗ )r−1 . Hence,
N ∈ I(P1 + P2 + P3 )r−1 .
PN
• Case
i=3 ai = b > 0: there exists i ∈ {3, . . . , N} such that ai > 0. We assume
i = 3. We write N = (x3 ) · (xa00 xa11 xa22 xa33 −1 · · · xaNN ), where x3 ∈ I(P1 + P2 + P3 ). By
(13) it follows that xa00 xa11 xa22 x3a3 −1 · · · xaNN ∈ I(P1 +P2 +P3 )(r−1) ⊆ I(P1 +P2 +P3 )r−2 ,
where the last inclusion holds for the induction. Hence, N ∈ I(P1 + P2 + P3 )r−1 .
Now we can prove the following important lemma.
Lemma 3.12. Let P1 , P2 and P3 be nonlinear points in PN and m1 ≤ m2 ≤ m3 and
suppose that m1 + m2 > m3 and m1 + m2 + m3 is odd. Let Z = m1 P1 + m2 P2 + m3 P3
be a scheme of fat points. Then
P
P
i)
(a). I(Z)(q(P mi +1)) ⊆ I(Z)q( m
for all q ∈ N,
P
P
(b). I(Z)(q( mi +1)+r) ⊆ I(Z)q( mi )+r−1 for all q ∈ N and 0 < r <
mi + 1.
Proof. Let us start with proving
on q. First, we let q = 1 as
P (a) by induction
P
( mi +1)
mi
the base case. Therefore, I(Z)
⊆ I(Z)
. Set Z1 = P1 + P2 + P3 and
Z2 = (m1 − 1)P1 + (m2 − 1)P2 + (m3 − 1)P3 , and we define
14
I. BAHMANI JAFARLOO
G. ZITO
P
P
P
W (n1 , n2 , n3 ) = (n1 + 3i=1 ni )P1 + (n2 +P 3i=1 ni )P2 + (n3 + 3i=1 ni )P3 , for ni ≥ 1.
3
We claim that I(W (n1 , n2 , n3 )) ⊆ I(Z1 ) i=1 ni , for all ni ≥ 1.
P3
Proof of the claim. We prove by induction on sum
i=1 ni . The base case
P3
by Lemma 3.11 it holds. Now,
is
i=1 ni = 3, with n1 = n2 = n3 = 1 and
P3
P3
′
we suppose the claim holds for ni such that i=1 n′i <
i=1 ni and we prove it
for ni . Because we have already considered the case n1 = n2 = n3 = 1, there
must exist an i such that ni > 1. We can assume
that n3 > 1. We consider
PN
aN
a0 a1
N = x0 x1 · · · xN ∈ I(W (n1 , n2 , n3 )). Set b = i=3 ai .
• Case b = 0: we have the following subcases.
◮ Subcase a1 = 0: by Cond(W (n1 , n2 , n3 )), it follows
P
P
≥ n1 + n2 + 2n3 ≥ 3i=1 ni , then it can be writa0 ≥ 2n1 + n2 + n3 ≥ 3i=1Pni and a2 P
P3
a −
3
n
3
a −
P3
n
ten N = (x0 x2 ) i=1 ni x00 i=1 i x22 i=1 i ∈ I(Z1 ) i=1 ni , because x0 x2 ∈ I(Z1 ).
◮ Subcase a2 = 0: similar to the subcase a1 = 0.
◮ Subcase a1 , a2 > 0: we can write, N = (x1 x2 )xa00 xa11 −1 xa22 −1 , where x1 x2 ∈ I(Z1 ).
It is easy, using the fact that ai satisfy the conditions at Cond(W (n1 , n2 , n3 )), to
check that xa00 xa11 −1 x2a2 −1 ∈ I(W (n1 , n2 , n3 − P
1)). So, using induction P
(n3 − 1 ≥ 1)
3
3
a0 a1 −1 a2 −1
n
−1
i
x0 x1 x2
∈ I(W (n1 , n2 , n3 − 1)) ⊆ I(Z1 ) i=1
, and N ∈ I(Z1 ) i=1 ni .
PN
• Case
i=3 ai = b > 0: without loss of generality, let a3 > 0.
We can write N = (x3 ) · (xa00 xa11 xa22 x3a3 −1 · · · xaNN ), where x3 ∈ I(Z1 ). By
using
P3
Cond(W (n1 , n2 , n3 )), the second factor is in I(W (n1 , n2 , n3 − 1))
⊆ I(Z1 ) i=1 ni −1 ,
P3
where the last inclusion holds by induction.P Hence N ∈ I(Z1 ) i=1 ni . So the claim
is proved. Now, from the definition I(Z1 ) mi +1 · I(Z2 ) ⊆ I(W (m1 , m2 , m3 )). By
Corollary 3.7 and the fact ρ(I(Z2 )) = 1, by Proposition 3.6 it follows
P
I(Z)(
mi +1)
= I(Z1 )(
P
mi +1)
P
· I(Z2 )(
= I(Z1 )(
mi +1)
· I(Z2 ) · I(Z2 )
⊆ I(Z1 )
P
mi
P
· I(Z2 )
P
mi
mi +1)
P
= I(Z1 )(
P
mi
mi +1)
· I(Z2 )
P
mi +1
⊆ I(W (m1 , m2 , m3 )) · I(Z2 )
= (I(Z1 ) · I(Z2 ))
P
mi
= I(Z)
P
mi
P
mi
,
and the base case is proved.
We suppose that (a) is true forP
q − 1, then we prove it for q. By induction and
mi + 1 is even,
Corollary 3.8, using the fact that
P
I(Z)(q(
mi +1))
P
= I(Z)((q−1)(
P
⊆ I(Z)(q−1)(
mi +1))
mi )
P
· I(Z)(
· I(Z)
P
mi
mi +1)
P
= I(Z)q(
mi )
.
For proving (b), we work by induction on q as before. First of all, we
Pneed to prove
(r)
r−1
the base case for q = 0. Hence, I(Z) ⊆ I(Z)
for 1 < r <
mi + 1. Set
Z1 = P1 + P2 + P3 and Z2 = (m1 − 1)P1 + (m2 − 1)P2 + (m3 − 1)P3 . We define
V (n1 , n2 ,P
n3 , r) = (r + n1 − 1)P1 + (r + n2 − 1)P2 + (r + n3 − 1)P3 for ni ≥ 1 and
1<r<
ni + 1. We claim that I(V (n1 , n2 , n3 , r)) ⊆ I(Z1 )r−1 always holds.
P
Proof of the claim. We work by induction on the sum
ni . The base case is
(r)
ni = 1. Then we have to prove I(P1 + P2 + P3 ) ⊆ I(P1 + P2 + P3)r−1 , for 1 < r < 4
and this is true by Lemma 3.11. We suppose that the claim is true for assignment n′i
ON THE CONTAINMENT PROBLEM FOR FAT POINTS
15
P ′ P
such that
ni < ni , then we prove it for ni . Because we have already considered
the case n1 = n2 = n3 = 1, there must exist an i such that ni > 1. We canP
assume
that n3 > 1. We consider N = xa00 xa11 · · · xaNN ∈ I(V (n1 , n2 , n3 , r)). Set b = N
i=3 ai ,
and we consider cases depending upon b.
• Case b = 0: we have the following subcases.
◮ Subcase a1 = 0: by Cond(V (n1 , n2 , n3 , r)), it follows that
a0 ≥ r + n1 − 1 ≥ r − 1 and a2 ≥ r + n3 − 1 ≥ r − 1.
So we can write N = (x0 x2 )r−1 xa00 −r+1 xa22 −r+1 ∈ I(Z1 )r−1 , because x0 x2 ∈ I(Z1 ).
◮ Subcase a2 = 0: similar to the case a1 = 0.
◮ Subcase a1 , a2 > 0: we write, N = (x1 x2 )xa00 x1a1 −1 x2a2 −1 , where x1 x2 ∈ I(Z1 ).
By Cond(V (n1 , n2 , n3 , r)), we deduce xa00 x1a1 −1 x2a2 −1 ∈ I(V (n1 , n2 , n3 − 1, r − 1)).
So for induction (n3 − 1 ≥ 1 and r − 1 < (n1 + n2 + n3 − 1) + 1) we conclude
xa00 xa11 −1 xa22 −1 ∈ I(V (n1 , n2 , n3 − 1, r − 1)) ⊆ I(Z1 )r−2 , and N ∈ I(Z1 )r−1 .
PN
• Case
i=3 ai = b > 0: we can assume that a3 > 0.
Then we can write N = (x3 )(xa00 xa11 xa22 xa33 −1 · · · xaNN ), where x3 ∈ I(Z1 ). Using
Cond(V (n1 , n2 , n3 , r)), we deduce xa00 xa11 xa22 xa33 −1 · · · xaNN ∈ I(V (n1 , n2 , n3 −1, r −1)).
So for induction (n3 − 1 ≥ 1 and r − 1 < (n1 + n2 + n3 − 1) + 1) we see
xa00 xa11 xa22 x3a3 −1 · · · xaNN ∈ I(V (n1 , n2 , n3 − 1, r − 1)) ⊆ I(Z1 )r−2 , and N ∈ I(Z1 )r−1 .
So the claim is true. From the definition I(Z1 )r · I(Z2 ) ⊆ I(V (m1 , m2 , m3 , r)). By
Corollary 3.7 and the fact that by Proposition 3.6, ρ(I(Z2 )) = 1,
I(Z)(r) = I(Z1 )(r) · I(Z2 )(r) = I(Z1 )(r) · I(Z2 )r = I(Z1 )(r) · I(Z2 ) · I(Z2 )r−1
⊆ I(V (m1 , m2 , m3 , r)) · I(Z2 )r−1 ⊆ I(Z1 )r−1 · I(Z2 )r−1
= (I(Z1 ) · I(Z2 ))r−1 = I(Z)r−1 ,
and the base case is proved. Now we can proceed with the inductive step. We
suppose that (b) is true for q − 1, thenP
we prove it for q. By induction and Corollary
3.8 we can write, using the fact that
mi + 1 is even,
P
I(Z)(q(
mi +1)+r)
P
= I(Z)((q−1)(
⊆ I(Z)
mi +1)+r)
P
(q−1)( mi )+r−1
P
· I(Z)(
· I(Z)
P
mi +1)
mi
P
= I(Z)q(
Thus the proof of lemma is complete.
mi )+r−1
.
By Lemma 3.12 we can deduce the following crucial corollary.
Corollary 3.13. Let P1 , P2 and P3 be
points in PN and m1 ≤ m2 ≤ m3 ,
Pnonlinear
and suppose that m1 + m2 > m3 and 3i=1 mi is odd. If Z = m1 P1 + m2 P2 + m3 P3 ,
P
P
then ρ(I(Z)) ≤ ( 3i=1 mi + 1)/( 3i=1 mi ).
P
P
(m)
Proof. It is enough to show that if m/n ≥ ( P
mi +1)/( mP
⊆ I(Z)n .
i ) then I(Z)
Suppose that
1)/( mi ). Then we can del P m andm n are such that m/n ≥ ( mi + P
P
m
+1
i
duce m ≥ P mi n . Now, n can be written as n = q mi + r with 0 ≤ r < mi .
16
I. BAHMANI JAFARLOO
G. ZITO
Thus
X
( P
r
q( mi + 1) if r = 0
P
m ≥ q(
mi + 1) + r + P
=
mi
q( mi + 1) + r + 1 if r 6= 0.
If r = 0, by Lemma 3.12
P
P
I(Z)(m) ⊆ I(Z)(q( mi +1)) ⊆ I(Z)q mi = I(Z)n .
P
If r =
6 0, then r ′ = r + 1 < mi + 1. By Lemma 3.12
P
′
I(Z)(m) ⊆ I(Z)(q( mi +1)+r ) ⊆ I(Z)q
P
P
Then ρ(I(Z)) ≤ ( 3i=1 mi + 1)/( 3i=1 mi ).
P
mi +r ′ −1
= I(Z)n .
By Corollary 3.13 we deduce the following result which generalizes Proposition
3.9 to the case N > 2.
Corollary 3.14. Let P1 , P2 and P3 be nonlinear points in PN and m1 ≤ m2 ≤ m3
and suppose that P
m1 +m2 > m3 and
P3 m1 +m2 +m3 is odd. If Z = m1 P1 +m2 P2 +m3 P3 ,
3
then ρ(I(Z)) ≥ ( i=1 mi + 1)/( i=1 mi ).
Proof. Consider Corollary 3.13 when N = 2. By Proposition 3.9
P
mi + 1
∗
∗
∗
,
ρ(I(m1 P1 + m2 P2 + m3 P3 )) = P
mi
where P1∗ = [0 : 0 : 1], P2∗ = [0 : 1 : 0], P3∗ = [1 : 0 : 0]. By contradiction, suppose
that
P
mi + 1
ρ(I(Z)) < P
= ρ(I(m1 P1∗ + m2 P2∗ + m3 P3∗ ))
mi
o
nm
= sup
: I(m1 P1∗ + m2 P2∗ + m3 P3∗ )(m) 6⊆ I(m1 P1∗ + m2 P2∗ + m3 P3∗)n .
n
Hence, it follows that there exists m/n such that
(
m/n ≥ ρ(I(Z)) then I(Z)(m) ⊆ I(Z)n
I(m1 P1∗ + m2 P2∗ + m3 P3∗ )(m) 6⊆ I(m1 P1∗ + m2 P2∗ + m3 P3∗ )n .
Thus, there exists a monomial N = xa00 xa11 xa22 ∈ I(m1 P1∗ + m2 P2∗ + m3 P3∗ )(m) such
that N ∈
/ I(m1 P1∗ + m2 P2∗ + m3 P3∗ )n . For the monomial N in K[x0 , x1 , x2 ], we have
N ∈ I(m1 P1∗ + m2 P2∗ + m3 P3∗)(m) if and only if N ∈ I(Z)(m) . Then N ∈ I(Z)n , and
∗
∗
∗ n
similarly this implies N ∈ I(m1 PP
3 P3 ) . Therefore we have found a
1 + m2 P2 + mP
3
3
contradiction. Then, ρ(I(Z)) ≥ ( i=1 mi + 1)/( i=1 mi ).
Now we can give a proof of Theorem 1.2.
Proof of Theorem 1.2. The desired result immediately follows from Corollary
3.13 and Corollary 3.14.
ON THE CONTAINMENT PROBLEM FOR FAT POINTS
17
References
[BH10a] Cristiano Bocci and Brian Harbourne. Comparing powers and symbolic powers of ideals.
J. Algebraic Geom., 19(3):399–417, 2010.
[BH10b] Cristiano Bocci and Brian Harbourne. The resurgence of ideals of points and the containment problem. Proc. Amer. Math. Soc., 138(4):1175–1190, 2010.
[ELS01] Lawrence Ein, Robert Lazarsfeld, and Karen E. Smith. Uniform bounds and symbolic
powers on smooth varieties. Invent. Math., 144(2):241–252, 2001.
[GHM13] A. V. Geramita, B. Harbourne, and J. Migliore. Star configurations in Pn . J. Algebra,
376:279–299, 2013.
[HH02] Melvin Hochster and Craig Huneke. Comparison of symbolic and ordinary powers of
ideals. Invent. Math., 147(2):349–369, 2002.
[Hoc73] Melvin Hochster. Criteria for equality of ordinary and symbolic powers of primes. Math.
Z., 133:53–65, 1973.
[Kod00] Vijay Kodiyalam. Asymptotic behaviour of Castelnuovo-Mumford regularity. Proc.
Amer. Math. Soc., 128(2):407–411, 2000.
[LS06]
Aihua Li and Irena Swanson. Symbolic powers of radical ideals. Rocky Mountain J.
Math., 36(3):997–1009, 2006.
[Swa00] Irena Swanson. Linear equivalence of ideal topologies. Math. Z., 234(4):755–775, 2000.
[ZS75]
Oscar Zariski and Pierre Samuel. Commutative algebra. Vol. II. Springer-Verlag, New
York-Heidelberg, 1975. Reprint of the 1960 edition, Graduate Texts in Mathematics,
Vol. 29.
i. bahmani jafarloo, dipartimento di scienze matematiche, politecnico di torino,
italy
E-mail address: [email protected]
i. bahmani jafarloo, dipartimento di matematica, università di torino, italy
E-mail address: [email protected]
g. zito, dipartimento di matematica e informatica, università di catania, italy
E-mail address: [email protected]
| 0math.AC
|
arXiv:1605.08534v1 [math.ST] 27 May 2016
On the two-filter approximations of marginal smoothing
distributions in general state space models
Thi Ngoc Minh Nguyen∗
Sylvain Le Corff†
Eric Moulines‡
Abstract
A prevalent problem in general state space models is the approximation of the
smoothing distribution of a state conditional on the observations from the past, the
present, and the future. The aim of this paper is to provide a rigorous analysis of
such approximations of smoothed distributions provided by the two-filter algorithms.
We extend the results available for the approximation of smoothing distributions to
these two-filter approaches which combine a forward filter approximating the filtering
distributions with a backward information filter approximating a quantity proportional
to the posterior distribution of the state given future observations.
1
Introduction
State-space models play a key role in a large variety of disciplines such as engineering,
econometrics, computational biology or signal processing, see [9, 8] and references therein.
This paper provides a nonasymptotic analysis of a Sequential Monte Carlo Method (SMC)
which aims at performing optimal smoothing in nonlinear and non Gaussian state space
models. Given two measurable spaces (X, X ) and (Y, Y), consider a bivariate stochastic
process {(Xt , Yt )}t≥0 taking values in the product space (X × Y, X ⊗ Y), where the hidden
state sequence {Xt }t≥0 is observed only through the observation process {Yt }t≥0 . Statistical inference in general state space models usually involves the computation of conditional
distributions of some unobserved states given a set of observations. These posterior distributions are crucial to compute smoothed expectations of additive functionals which appear
naturally for maximum likelihood parameter inference in hidden Markov models (computation of the Fisher score or of the intermediate quantity of the Expectation Maximization
algorithm), see [3, Chapter 10 and 11], [17, 25, 20, 21].
Nevertheless, exact computation of the filtering and smoothing distributions is possible
only for linear and Gaussian state spaces or when the state space X is finite. This paper
focuses on particular instances of Sequential Monte Carlo methods which approximate sequences of distributions in a general state space X with random samples, named particles,
associated with nonnegative importance weights. Those particle filters and smoothers rely
on the combination of sequential importance sampling steps to propagate particles in the
state space and importance resampling steps to duplicate or discard particles according to
their importance weights. The first implementation of these SMC methods, introduced in
∗ LTCI,
CNRS and Télécom ParisTech.
de Mathématiques d’Orsay, Univ. Paris-Sud, CNRS, Université Paris-Saclay.
‡ Centre de Mathématiques Appliquées, Ecole Polytechnique.
† Laboratoire
1
Nguyen et al.
Two-filter approximations of marginal smoothing distributions
[14, 18], propagates the particles using the Markov kernel of the hidden process {Xt }t≥0
and uses a multinomial resampling step based on the importance weights to select particles
at each time step. An interesting feature of this Poor man’s smoother is that it provides
an approximation of the joint smoothing distribution by storing the ancestral line of each
particle with a complexity growing only linearly with the number N of particles, see for instance [4]. However, this smoothing algorithm has a major shortcoming since the successive
resampling steps induce an important depletion of the particle trajectories. This degeneracy
of the particle sequences leads to trajectories sharing a common ancestor path; see [25, 16]
for a discussion.
Approximations of the smoothing distributions may also be obtained using the forward
filtering backward smoothing decomposition in general state space models. The Forward
Filtering Backward Smoothing algorithm (FFBS) and the Forward Filtering Backward Simulation algorithm (FFBSi) developed respectively in [18, 15, 10] and [13] avoid the path
degeneracy issue of the Poor man’s smoother at the cost of a computational complexity
growing with N 2 . Both algorithms rely on a forward pass which produces a set of particles
and weights approximating the sequence of filtering distributions up to time T . Then, the
backward pass of the FFBS algorithm modifies all the weights computed in the forward pass
according to the so-called backward decomposition of the smoothing distribution keeping
all the particles fixed. On the other hand, the FFBSi algorithm samples independently
particle trajectories among all the possible paths produced by the forward pass. It is shown
in [22, 2, 5] that the FFBS algorithm can be implemented using only a forward pass when
approximating smoothed expectations of additive functionals but with a complexity still
growing quadratically with N . Under the mild assumption that the transition density of the
hidden chain {Xt }t≥0 is uniformly bounded above, [6] proposed an accept-reject mechanism
to implement the FFBSi algorithm with a complexity growing only linearly with N . Concentration inequalities, controls of the Lq -norm of the deviation between smoothed functionals
and their approximations and Central Limit Theorems (CLT) for the FFBS and the FFBSi
algorithms have been established in [5, 6, 11].
Recently, [23] proposed a new SMC algorithm, the particle-based rapid incremental
smoother (PaRIS), to approximate online, using only a forward pass, smoothed expectations of additive functionals. The crucial feature of this algorithm is that its complexity
grows only linearly with N as it samples on-the-fly particles distributed according to the
backward dynamics of the hidden chain conditionally on the observations Y0 , . . . , YT . The
authors show concentration inequalities and CLT for the estimators provided by the PaRIS
algorithm.
In this paper, we extend the theoretical results available for the SMC approximations of
smoothing distributions to the estimators given by the two-filter algorithms. These methods were first introduced in the particle filter literature by [18] and developed further by
[1] and [12]. The two-filter approach combines the output of two independent filters, one
that evolves forward in time and approximates the filtering distributions and another that
evolves backward in time approximating a quantity proportional to the posterior distribution
of a state given future observations. In [12], the authors introduced a proposal mechanism
leading to algorithms whose complexity grows linearly with the number of particles. An algorithm similar to the algorithm of [1] may also be implemented with an O(N ) computational
complexity following the same idea. We analyze all these algorithms which approximate
the marginal smoothing distributions (smoothing distributions of one state given all the
observations) and provide concentration inequalities as well as CLT.
This paper is organized as follows. Section 2 introduces the different particle approxima2
Nguyen et al.
Two-filter approximations of marginal smoothing distributions
tions of the marginal smoothing distributions given by the two-filter algorithms. Sections 3
and 4 provide exponential deviation inequalities and CLT for the particle approximations
under mild assumptions on the hidden Markov chain. Under additional strong mixing assumptions, it is shown that the results of Section 3 are uniform in time and that the asymptotic variance in Section 4 may be uniformly bounded in time. All proofs are postponed to
Section 5.
Notations and conventions
Let X and Y be two general state-spaces endowed with countably generated σ-fields X and Y.
Fb (X, X ) is the set of all real valued bounded measurable functions on (X, X ). Q is a Markov
transition kernel defined on X×X and {gt }t≥0 a family of positive functions defined on X. For
any x ∈ X, Q(x, ·) has a density q(x, ·) with respect to a measure λ on (X, X ). The oscillation
of a real valued function defined on a space Z is given by: osc (h) := supz,z′ ∈Z |h(z) − h(z ′ )|.
2
The two-filter algorithms
For any measurable function h on Xt−s+1 , probability distribution χ on (X, X ), T ≥ 0 and
0 ≤ s ≤ t ≤ T , define the joint smoothing distribution by:
R
QT
χ(dx0 )g0 (x0 ) u=1 Q(xu−1 , dxu )gu (xu )h(xs:t )
φχ,s:t|T [h] :=
,
(1)
R
QT
χ(dx0 )g0 (x0 ) u=1 Q(xu−1 , dxu )gu (xu )
where au:v is a short-hand notation for {as }vs=u . In the following we use the notations
φχ,s|T := φχ,s:s|T and φχ,t := φχ,t:t|t . The aim of this paper is to provide a rigorous analysis
of the performance of SMC algorithms approximating the sequence φχ,s|T for 0 ≤ s ≤ T .
The algorithms analyzed in this paper are based on the two-filter formula introduced in
[1, 12], which we now detail.
2.1
Forward filter
Let {ξ0ℓ }N
ℓ=1 be i.i.d. and distributed according to the instrumental distribution ρ0 and define
the importance weights
dχ ℓ
(ξ ) g0 (ξ0ℓ ) .
ω0ℓ :=
dρ0 0
For any h ∈ Fb (X, X ),
−1
φN
χ,0 [h] := Ω0
N
X
ω0ℓ h(ξ0ℓ ) ,
where Ω0 :=
ℓ=1
N
X
ω0ℓ ,
ℓ=1
ℓ
ℓ
is a consistent estimator of φχ,0 [h], see for instance [4]. Then, based on {(ξs−1
, ωs−1
)}N
ℓ=1 a
new set of particles and importance weights is obtained using the auxiliary sampler introduced in [24]. Pairs {(Isℓ , ξsℓ )}N
ℓ=1 of indices and particles are simulated independently from
the instrumental distribution with density on {1, . . . , N } × X:
ℓ
ℓ
ℓ
πs|s (ℓ, x) ∝ ωs−1
ϑs (ξs−1
)ps (ξs−1
, x) ,
3
(2)
Nguyen et al.
Two-filter approximations of marginal smoothing distributions
where ϑs is the adjustment multiplier weight function and ps is a Markovian transition
density. For any ℓ ∈ {1, . . . , N }, ξsℓ is associated with the importance weight defined by:
Iℓ
ωsℓ :=
s
q(ξs−1
, ξsℓ )gs (ξsℓ )
Iℓ
(3)
Iℓ
s
s
ϑs (ξs−1
)ps (ξs−1
, ξsℓ )
to produce the following approximation of φχ,s [h]:
−1
φN
χ,s [h] := Ωs
N
X
ωsℓ h(ξsℓ ) ,
where Ωs :=
ℓ=1
2.2
N
X
ωsℓ .
ℓ=1
Backward filter
Let {γt }t≥0 be a family of positive measurable functions such that, for all t ∈ {0, . . . , T },
" T
#
Z
Y
γt (xt ) dxt
gu−1 (xu−1 ) Q(xu−1 , dxu ) gT (xT ) < ∞ .
(4)
u=t+1
Following [1], for any 0 ≤ t ≤ T we introduce the backward filtering distribution ψγ,t|T on X
(referred to as the backward information filter in [18] and [1]) defined, for any h ∈ Fb (X, X ),
by:
i
hQ
R
T
g
(x
)
Q(x
,
dx
)
gT (xT )h(xt )
γt (xt ) dxt
u−1
u
u=t+1 u−1 u−1
hQ
i
.
ψγ,t|T [h] :=
R
T
γt (xt ) dxt
u=t+1 gu−1 (xu−1 ) Q(xu−1 , dxu ) gT (xT )
If the distribution of Xt has probability density function
γt , then ψγ,t|T is the conditional
R
distribution of Xt given Yt:T . Contrary to [1] or [12], γt (xt )dxt may be infinite. The only
requirement about the nonnegative functions {γt }t≥0 is the condition (4) and the fact that
γt should be available in closed form. Here γt is a possibly improper prior introduced to
make ψγ,t|T a proper posterior distribution, which is of key importance when producing
particle approximations of such quantities. For 0 ≤ t ≤ T − 1, the backward information
filter is computed by the recursion
Z
q(xt , xt+1 )
h(xt ) dxt ,
(5)
ψγ,t|T [h] ∝ ψγ,t+1|T (dxt+1 ) γt (xt )gt (xt )
γt+1 (xt+1 )
in the backward time direction. (5) is analogous to the forward filter recursion and particle
approximations of the backward information filter can be obtained similarly. Using the
definition of the forward filtering distribution at time s − 1 and the backward information
filter at time s + 1, the marginal smoothing distribution may be expressed as
Z
q(xs , xs+1 )
h(xs )dxs . (6)
φχ,s|T [h] ∝ φχ,s−1 (dxs−1 )ψγ,s+1|T (dxs+1 )q(xs−1 , xs )gs (xs )
γs+1 (xs+1 )
We now describe the Sequential Monte Carlo methods used to approximate the recursion
(5) in [1], [12]. Let ρ̌T be an instrumental probability density on X and {ξˇTi |T }N
i=1 be i.i.d.
i
random variables such that ξ̌T |T ∼ ρ̌T and define
ω̌Ti |T
gT (ξˇTi |T )γT (ξˇTi |T )
.
:=
ρ̌T (ξˇi )
T |T
4
Nguyen et al.
Two-filter approximations of marginal smoothing distributions
i
i
Let now {(ξˇt+1|T
, ω̌t+1|T
)}N
i=1 be a weighted sample targeting the backward information
filter distribution ψγ,t+1|T [h] at time t + 1:
N
[h] := Ω̌−1
ψγ,t+1|T
t+1|T
N
X
i
i
),
h(ξˇt+1|T
ω̌t+1|T
where
Ω̌t+1|T :=
N
X
i
.
ω̌t+1|T
i=1
i=1
Plugging this approximation into (5) yields the target probability density
"
#
i
N
X
q(xt , ξ̌t+1|T
)
i
tar
ω̌t+1|T γt (xt )gt (xt )
ψ̂γ,t|T (xt ) ∝
,
i
)
γt+1 (ξˇt+1|T
i=1
which is the marginal probability density function of xt of the joint density
aux
(i, xt ) ∝
ψ̂γ,t|T
i
ω̌t+1|T
γt+1 (ξˇi
t+1|T )
i
).
γt (xt )gt (xt )q(xt , ξˇt+1|T
A particle approximation of the backward information filter at time t can be derived by
choosing an adjustment weight function ϑt|T and an instrumental density kernel rt|T , and
i
simulating {(Iˇti , ξˇt|T
)}N
i=1 from the instrumental probability density on {1, . . . , N } × X given
by
i
i
ω̌t+1|T
ϑt|T (ξˇt+1|T
)
i
πt|T (i, xt ) ∝
, xt ) .
(7)
rt|T (ξˇt+1|T
)
γt+1 (ξˇi
t+1|T
Subsequently, the particles are associated with the importance weights
Iˇi
i
:=
ω̌t|T
i
i
i
t
)
γt (ξˇt|T
)gt (ξˇt|T
)q(ξˇt|T
, ξˇt+1|T
Iˇi
Iˇi
i )
t
t
, ξˇt|T
)rt|T (ξˇt+1|T
ϑt|T (ξˇt+1|T
.
(8)
Ideally, a fully adapted version of the auxiliary
backward information filter is obtained by
R
using the adjustment weights ϑ⋆t|T (x) = γt (xt )gt (xt )q(xt , x) dxt and the proposal kernel
density
q(xt , x)
⋆
rt|T
(x, xt ) = γt (xt )gt (xt ) ⋆
,
ϑt|T (x)
yielding uniform importance weights. Such a solution is most likely to be cumbersome from
a computational perspective.
2.3
Two-filter approximations of the marginal smoothing distributions
Plugging the particle approximations of the forward and backward filter distributions into
(6) provides the following mixture approximation of the smoothing distribution:
φ̂tar
χ,s|T (xs )
j
N X
N ωi
X
s−1 ω̌s+1|T
j
i
q(ξs−1
, xs )gs (xs )q(xs , ξˇs+1|T
).
∝
γs+1 (ξˇj
)
i=1 j=1
s+1|T
5
(9)
Nguyen et al.
Two-filter approximations of marginal smoothing distributions
Following the TwoFiltf wt algorithm of Fearnhead, Wyncoll and Tawn [12], the probability
density (9) might be seen as the marginal density of xs obtained from the joint density on
the product space {1, . . . , N }2 × X given by
φ̂aux
χ,s|T (i, j, xs ) ∝
j
i
ωs−1
ω̌s+1|T
j
i
q(ξs−1
, xs )gs (xs )q(xs , ξ̌s+1|T
).
γs+1 (ξˇj
)
(10)
s+1|T
ℓ
The TwoFiltf wt algorithm draws a set {(Isℓ , Iˇsℓ , ξ˜s|T
)}N
ℓ=1 of indices and particle positions
from the instrumental density
j
j
i
i
ωs−1
ϑ̃s|T (ξs−1
, ξˇs+1|T
)ω̌s+1|T
j
i
r̃s|T (ξs−1
, ξˇs+1|T
; xs ) ,
πs|T (i, j, xs ) ∝
j
ˇ
γs+1 (ξ
)
(11)
s+1|T
where, as above, ϑ̃s|T (x, x′ ) is an adjustment multiplier weight function (which now depends
on the forward and backward particles) and r̃s|T is an instrumental kernel. We then associate
ℓ
with each draw (Isℓ , Iˇsℓ , ξ˜s|T
) the importance weight
Iˇℓ
Iℓ
ℓ
ω̃s|T
:=
ℓ
ℓ
ℓ
s
s
)
q(ξs−1
, ξ̃s|T
)gs (ξ˜s|T
)q(ξ˜s|T
, ξˇs+1|T
Iℓ
Iˇℓ
Iˇℓ
Iℓ
s
s
s
ℓ )
s
ϑ̃s|T (ξs−1
, ξ̌s+1|T
)r̃s|T (ξs−1
, ξˇs+1|T
; ξ̃s|T
,
Ω̃s|T :=
N
X
ℓ
.
ω̃s|T
(12)
ℓ=1
ℓ
ℓ
N
Then, the auxiliary indices {(Isℓ , Iˇsℓ )}N
ℓ=1 are discarded and {(ω̃s|T , ξ̃s|T )}ℓ=1 approximate the
target smoothing density φ̂tar
χ,s|T . Mimicking the arguments in [15] and further developed
in [19], the auxiliary
particle
filter is fully adapted if the adjustment weight function is
R
ϑ⋆s|T (x, x′ ) = q(x, xs )gs (xs )q(xs , x′ ) dxs and the instrumental kernel is
⋆
(x, x′ ; xs ) = q(x, xs )gs (xs )q(xs , x′ )/ϑ⋆s|T (x, x′ ) .
rs|T
Except in simple scenarios, simulating from the fully adapted auxiliary filter is computationally intractable.
Instead of considering the target distribution (9) as the marginal of the auxiliary distribution (10) over pairs of indices, the TwoFiltbdm algorithm of [1] uses the following partial
auxiliary distributions having densities,
φ̂aux,f
s|T (i, xs )
∝
φ̂aux,b
s|T (j, xs ) ∝
i
i
ωs−1
q(ξs−1
, xs )gs (xs )
N
X
j=1
j
ω̌s+1|T
γs+1 (ξˇj
s+1|T )
j
q(xs , ξˇs+1|T
),
j
N
X
ω̌s+1|T
j
i
i
ωs−1
q(ξs−1
, xs )
q(x
,
ξ̌
)g
(x
)
s
s
s
s+1|T
j
γs+1 (ξˇs+1|T
)
i=1
.
aux,f
Since φ̂tar
χ,s|T is the marginal probability density of the partial auxiliary distributions φ̂s|T
and φ̂aux,b
with respect to the forward and the backward particle indices, respectively,
s|T
ℓ ℓ N
N
ˇℓ ˇℓ
we may sample from φ̂tar
χ,s|T by simulating instead {(Is , ξs )}ℓ=1 or {(Is , ξs|T )}ℓ=1 from the
instrumental probability density functions
f
i
i
i
πs|T
(i, xs ) ∝ ωs−1
ϑs (ξs−1
)ps (ξs−1
, xs ) ,
j
j
j
j
b
πs|T
(j, xs ) ∝ ϑs|T (ξˇs+1|T
)ω̌s+1|T
rs|T (ξˇs+1|T
, xs )/γs+1 (ξˇs+1|T
),
6
Nguyen et al.
Two-filter approximations of marginal smoothing distributions
where (ϑs , ps ) and (ϑs+1|T , rs|T ) are the adjustment multiplier weight functions and the
instrumental kernels used in the forward and backward passes. In this case the algorithm
uses the particles obtained when approximating the forward filter and backward information
i,f
i,b
i
N
filter to provide two different weighted samples {(ω̃s|T
, ξsi )}N
i=1 and {(ω̃s|T , ξ̌s|T )}i=1 targeting
i,b N
i,f N
}i=1
the marginal smoothing distribution, where the forward {ω̃s|T
}i=1 and backward {ω̃s|T
importance weights are given by
i,f
ω̃s|T
:= ωsi
N
X
j
j
j
ω̌s+1|T
q(ξsi , ξ̌s+1|T
)/γs+1 (ξˇs+1|T
),
N
X
Ω̃fs|T :=
(13)
j=1
j=1
j,b
j
ω̃s|T
:= ω̌s|T
j,f
ω̃s|T
,
N
X
j
j
i
i
ωs−1
q(ξs−1
, ξ̌s|T
)/γs (ξˇs|T
),
Ω̃bs|T :=
N
X
j,b
ω̃s|T
.
(14)
j=1
i=1
An important drawback of these algorithms is that the computation of the forward and
backward importance weights grows quadratically with the number N of particles.
2.4
O(N) approximations of the marginal smoothing distributions
In [12], the authors introduced a proposal mechanism in (11) such that the indices (Is , Iˇs )
of the forward and backward particles chosen at time s − 1 and s + 1 are sampled independently. Such choices lead to algorithms whose complexity grows linearly with the number
of particles. The O(N ) algorithm displayed in [12] suggests to use an adjustment multiplier
weight function in (11) such that Is and Iˇs are chosen according to the same distributions as
the indices sampled in the forward filter and in the backward information filter. It is done
in [12] by choosing ϑ̃s|T (x, x′ ) = ϑs (x)ϑs|T (x′ ) so that (11) becomes
πs|T (i, j, xs ) ∝
i
i
ωs−1
ϑs (ξs−1
)
j
j
ϑs|T (ξˇs+1|T
)ω̌s+1|T
j
i
r̃s|T (ξs−1
, ξˇs+1|T
; xs ) .
j
ˇ
γs+1 (ξ
)
(15)
s+1|T
ℓ
In this case, the importance weight (12) associated with each draw (Isℓ , Iˇsℓ , ξ˜s|T
) is given by
Is
Iˇs
ℓ
ℓ
ℓ
)
q(ξs−1
, ξ˜s|T
)gs (ξ˜s|T
)q(ξ˜s|T
, ξ̌s+1|T
ℓ
ℓ
ω̃s|T
:=
ℓ
Isℓ
Iˇsℓ
Isℓ
Iˇsℓ
ℓ )
ϑs (ξs−1
)ϑs|T (ξˇs+1|T
)r̃s|T (ξs−1
, ξ̌s+1|T
; ξ̃s|T
.
(16)
Instead of sampling new particles at time s, an algorithm similar to the TwoFiltbdm algorithm
ˇℓ N
of [1] which uses the forward particles {ξsℓ }N
ℓ=1 or backward particles {ξs|T }ℓ=1 may also be
implemented with an O(N ) computational complexity.
Iℓ
Iˇℓ
Iˇℓ
s
s
s
, xs ) in (15), the smoothing distribution
; xs ) = rs|T (ξˇs+1|T
, ξ̌s+1|T
(a) Choosing r̃s|T (ξs−1
approximation is obtained by reweighting the particles obtained in the backward pass.
ℓ
The backward particles {ξˇs|T
}N
ℓ=1 are associated with the importance weights:
Iˇs
ℓ
ℓ
ℓ
)
γs (ξˇs|T
)gs (ξˇs|T
)q(ξˇs|T
, ξˇs+1|T
ℓ
ℓ
:=
ω̃s|T
ˇℓ
ˇℓ
Iℓ
ℓ
s
, ξ̌s|T
)
q(ξs−1
Iℓ
Is
Is
ℓ ) γ (ξˇℓ )ϑ (ξ s )
ϑs|T (ξˇs+1|T
)rs|T (ξˇs+1|T
, ξˇs|T
s s−1
s s|T
,
Iℓ
=
ℓ
ω̌s|T
ℓ
s
, ξ̌s|T
)
q(ξs−1
Iℓ
ℓ )ϑ (ξ s )
γs (ξˇs|T
s s−1
7
.
(17)
Nguyen et al.
Two-filter approximations of marginal smoothing distributions
Iˇℓ
Iℓ
Iℓ
s
s
s
; xs ) = ps (ξs−1
, xs ) in (15), the smoothing distribution ap(b) Choosing r̃s|T (ξs−1
, ξ̌s+1|T
proximation is obtained by reweighting the particles obtained in the forward filtering
pass. The forward particles {ξsℓ }N
ℓ=1 are associated with the importance weights:
Iˇℓ
Iℓ
ℓ
ω̃s|T
3
:=
s
)
q(ξsℓ , ξˇs+1|T
s
q(ξs−1
, ξsℓ )gs (ξsℓ )
Is
Is
Iˇsℓ
ϑs (ξs−1
)ps (ξs−1
, ξsℓ ) ϑs|T (ξˇs+1|T
)
ℓ
ℓ
Iˇℓ
=
ωsℓ
s
)
q(ξsℓ , ξ̌s+1|T
Iˇsℓ
ϑs|T (ξˇs+1|T
)
.
(18)
Exponential deviation inequality for the two-filter algorithms
In this section, we establish exponential deviation inequalities for the two-filter algorithms
introduced in Section 2. Before stating the results, some additional notations are required.
Define, for all (x, x′ , x′′ ) ∈ X3 ,
q [2] (x, x′ ; x′′ ) = q(x, x′′ )q(x′′ , x′ )
and for any functions f : X2 → R and g : X → R,
f ⊙ g(x, x′ ) := f (x, x′ )g(x′ ) .
Consider the following assumptions:
A1. |q|∞ < ∞ and for all 0 ≤ t ≤ T , gt is positive and |gt |∞ < ∞.
A2. For all 0 ≤ t ≤ T , |ϑt |∞ < ∞, |pt |∞ < ∞ and |ωt |∞ < ∞ where
ω0 (x) :=
dχ
(x)g0 (x)
dρ0
and for all t ≥ 1
A3. - For all 0 ≤ t ≤ T − 1, ϑt|T /γt+1
ω̌t|T ∞ < ∞, where
ω̌T |T (x) :=
∞
q(x, x′ )gt (x′ )
.
ϑt (x)pt (x, x′ )
< ∞ and |rt|T |∞ < ∞. For all 0 ≤ t ≤ T
gT (x)γT (x)
γt (x)gt (x)q(x, x′ )
and for all 0 ≤ t < T, ω̌t|T (x, x′ ) :=
.
ρ̌T (x)
ϑt|T (x′ )rt|T (x′ , x)
−1
- For all 1 ≤ t ≤ T − 1, ϑ̃t|T ⊙ γt+1
r̃t|T
∞
ωt (x, x′ ) :=
∞
< ∞ where
ω̃t|T (x, x′ ; x′′ ) :=
−1
< ∞, q ⊙ γt+1
∞
< ∞, ω̃t|T
q [2] (x, x′ ; x′′ )gs (x′′ )
ϑ̃t|T (x, x′′ )r̃t|T (x, x′ ; x′′ )
∞
< ∞ and
.
j
j
We first show that the weighted sample {(ωsi ω̌t|T
), (ξsi , ξˇt|T
)}N
i,j=1 targets the product
distribution φχ,s ⊗ ψγ,t|T .
Theorem 1. Assume that A1, A2 and A3 hold for some T < ∞. Then, for all 0 ≤
s < t ≤ T , there exist 0 < Bs,t|T , Cs,t|T < ∞ such that for all N ≥ 1, ǫ > 0 and all
h ∈ Fb (X × X, X ⊗ X ),
N
i ω̌ j
X
2
2
ω
t|T
j
s
P
h(ξsi , ξˇt|T
) − φχ,s ⊗ ψγ,t|T [h] > ǫ ≤ Bs,t|T e−Cs,t|T N ǫ / osc (h) .
Ω
Ω̌
i,j=1 s t|T
8
Nguyen et al.
Two-filter approximations of marginal smoothing distributions
Proof. The proof is postponed to Section 5.1.
i
ℓ
We now study the weighted sample {(ω̃s|T
, ξ˜s|T
)}N
ℓ=1 produced by the TwoFiltf wt algorithm of Fearnhead, Wyncoll and Tawn [12] defined in (11) and (12) and targeting the
marginal smoothing distribution φχ,s|T .
Theorem 2 (deviation inequality for TwoFiltf wt of [12]). Assume that A1, A2 and A3 hold
for some T < ∞. Then, for all s < T , there exist 0 < Bs|T , Cs|T < ∞ such that for all
N ≥ 1, ε > 0 and all h ∈ Fb (X, X ),
P
N
i
X
ω̃s|T
i=1
Ω̃s|T
i
)
h(ξ˜s|T
− φχ,s|T [h] > ǫ
!
≤ Bs|T e−Cs|T N ǫ
2
/ osc2(h)
.
Proof. The proof is postponed to Section 5.2.
Using Theorem 1 and Lemma 7, we may derive an exponential inequality for the weighted
i,f
i,b
N
ˇi
samples {(ξsi , ω̃s|T
)}N
i=1 and {(ξs|T , ω̃s|T )}i=1 produced by the TwoFiltbdm algorithm of [1],
i,f
i,b
where ω̃s|T
and ω̃s|T
are defined in (13) and (14). Therefore, both the forward and the
backward particle approximations of the smoothing distribution converge to the marginal
smoothing distribution, and these two approximations satisfy an exponential inequality.
Theorem 3 (deviation inequality for the TwoFiltbdm algorithm of [1]). Assume that A1, A2
and A3 hold for some T < ∞. Then, for all 1 ≤ s ≤ T − 1, there exist 0 < Bs|T , Cs|T < ∞
such that for all N ≥ 1, ε > 0 and all h ∈ Fb (X, X ),
P
P
N ω̃ i,f
X
s|T
Ω̃fs|T
i=1
N ω̃ i,b
X
s|T
b
Ω̃
s|T
i=1
h(ξsi )
− φχ,s|T [h] > ǫ
!
i
) − φχ,s|T [h] > ǫ
h(ξˇs|T
≤ Bs|T e−Cs|T N ǫ
!
2
/ osc2(h)
≤ Bs|T e−Cs|T N ǫ
2
,
/ osc2(h)
(19)
.
(20)
Proof. The proof is postponed to Section 5.3.
Remark 1. Following [23, 6, 11], time uniform exponential inequalities for the two-filter
approximations of the marginal smoothing distributions may be obtained using strong mixing
assumptions which are standard in the SMC literature:
H1. There exist 0 < σ− < σ+ < ∞ and c− > 0 such that for all x, x′ ∈ X, σ− ≤ q(x, x′ ) ≤
σ+ and for all t ≥ 0,
Z
Z
χ(dx0 )g0 (x0 ) ≥ c− and inf
Q(x, dx′ )gt (x′ ) ≥ c− .
x∈X
H 2. There exist 0 < γ− < γ+ < ∞ and č− > 0 such that for all x ∈ X and all t ≥ 0,
γ− ≤ γt (x) ≤ γ+ and for all t ≥ 0,
Z
Z
−1
γT (xT )gT (xT )dxT ≥ č− and inf
γt (xt )gt (xt )q(xt , x)γt+1
(x)dxt ≥ č− .
x∈X
9
Nguyen et al.
Two-filter approximations of marginal smoothing distributions
(i) If A1 and A2 hold uniformly in T and if H1 holds, then, it is proved in [6] that
Proposition 8 holds with constants that are uniform in time : there exist 0 < B, C < ∞
such that for all s ≥ 0, N > 0, ǫ > 0 and all h ∈ Fb (X, X ),
!
N
X
2
2
ωsi h(ξsi ) − φχ,s [h] ≥ ǫ ≤ Be−CN ǫ / osc(h) .
P Ω−1
s
i=1
(ii) It can be shown following the exact same steps that if A1 and A3 hold uniformly in
T and if H1 and H2 hold then Proposition 9 holds with constants that are uniform
in time: there exist 0 < B, C < ∞ such that for all t ≥ 0, N ≥ 1, ǫ > 0, and all
h ∈ Fb (X, X ),
"
#
N
X
2
2
−1
i
i
ˇ
P Ω̌t|T
ω̌t|T h(ξt|T ) − ψγ,t|T [h] ≥ ǫ ≤ Be−CN ǫ / osc(h) .
i=1
(iii) Therefore, if A1, A2 and A3 hold uniformly in T and if H1 and H2 hold, then Theorem 1
holds with constants that are uniform in time. As a direct consequence, Theorems 2
and 3 hold also with constants that are uniform in time.
4
Asymptotic normality of the two-filter algorithms
We now establish CLT for the two-filter algorithms. Note first that under assumptions
A1, A2 and A3, for all 0 ≤ s, t ≤ T a CLT may be derived for the weighted samples
N
i
ˇi
{(ξsℓ , ωsℓ )}N
ℓ=1 and {(ξt|T , ω̌t|T )}i=1 which target respectively the filtering distribution φχ,s
and the backward information filter ψγ,t|T . By Propositions 10 and 11, there exist Γχ,s and
Γ̌γ,t|T such that for any h ∈ Fb (X, X ),
N 1/2
N 1/2
N
X
D
ωsi
h(ξsi ) − φχ,s [h] −→N →∞ N (0, Γχ,s [h − φχ,s [h]]) ,
Ωs
i=1
N ω̌ j
X
t|T
j=1
Ω̌t|T
D
j
h(ξˇt|T
) − ψγ,t|T [h] −→N →∞ N 0, Γ̌γ,t|T h − ψγ,t|T [h] .
(21)
(22)
j
j
Theorem 4 establishes a CLT for the weighted sample {ωsi ω̌t|T
, (ξsi , ξ̌t|T
)}N
i,j=1 which targets
the product distribution φχ,s ⊗ψγ,t|T . As an important consequence, the asymptotic variance
of the weighted sample {ω i ω̌ j , (ξ i , ξˇj )}N
is the sum of two contributions, the first one
s t|T
s
t|T
i,j=1
involves Γχ,s and the second one Γ̌γ,t|T . Intuitively, this may be explained by the fact that
N
the estimator φN
χ,s ⊗ψγ,t|T [h] is obtained by mixing two independent weighted samples which
suggests the following decomposition:
j
N
N ω̌ j
√ X
√ X
ωsi ω̌t|T
t|T
j
j
N
φχ,s [h̃s,t (·, ξˇt|T
)]
h̃s,t (ξsi , ξ̌t|T
)= N
Ωs Ω̌t|T
Ω̌t|T
i,j=1
j=1
+
10
N
√ X
ωsi
N
N
ψγ,t|T [h̃s,t (ξsi , ·)] + Es,T
|t (h̃s,t ) ,
Ω
s
i=1
Nguyen et al.
Two-filter approximations of marginal smoothing distributions
where h̃s,t = h − φχ,s ⊗ ψγ,t|T [h] and
N
Es,T
|t (h)
j
N
o
√ X
ωsi ω̌t|T n i j
j
h(ξs , ξ̌t|T ) − φχ,s [h(·, ξˇt|T
)] − ψγ,t|T [h(ξsi , ·)] .
:= N
Ω Ω̌
i,j=1 s t|T
A CLT for the two independent first terms is obtained by (21) and (22). It remains then to
N
prove that Es,T
|t (h) converges in probability to 0. However, this cannot be obtained directly
from the exponential deviation inequality derived in Theorem 1 and requires sharper controls
of the smoothing error (for instance nonasymptotic Lp -mean error bounds). Theorem 4
provides a direct proof following the asymptotic theory of weighted system of particles
developed in [7].
Theorem 4. Assume that A1, A2 and A3 hold for some T < ∞. Then, for all 0 ≤ s < t ≤
T and all h ∈ Fb (X × X, X × X ),
N
i ω̌ j
X
√
ω
t|T
j
s
h(ξsi , ξˇt|T
) − φχ,s ⊗ ψγ,t|T [h]
N
Ω
Ω̌
s
t|T
i,j=1
D
−→N →∞ N 0, Γ̃s,t|T h − φχ,s ⊗ ψγ,t|T [h] ,
where Γ̃s,t|T [h] is defined by:
Z
Z
Γ̃s,t|T [h] := Γχ,s
ψγ,t|T (dxt )h(·, xt ) + Γ̌γ,t|T
φχ,s (dxs )h(xs , ·) ,
(23)
with Γχ,s and Γ̌γ,t|T are given in Proposition 10 and Proposition 11.
Proof. The proof is postponed to Section 5.4.
Define
Z
[2]
−1
γs+1
σs := φχ,s−1 ⊗ ψγ,s+1|T
,
q (·, x)gs (x)dx ⊙
Z
−1
Σs [h] := Γ̃s−1,s+1|T
.
q [2] (·; x)gs (x)h(x)dx ⊙ γs+1
Theorem 5 provides a CLT for the TwoFiltf wt algorithm of [12]
Theorem 5 (CLT for the TwoFiltf wt algorithm of [12]). Assume that A1, A2 and A3 hold
for some T < ∞. Then, for all 1 ≤ s ≤ T − 1 and all h ∈ Fb (X, X ),
!
N
i
X
√
ω̃s|T
D
i
N
h(ξ˜s|T ) − φχ,s|T [h] −→N →∞ N 0, Υχ,s|T h − φχ,s|T [h] .
Ω̃s|T
i=1
where
i
h
−1
Υχ,s|T [h] = σs−2 Σs [h] + φχ,s−1 ⊗ ψγ,s+1|T ϑ̃s|T ⊙ γs+1
Z
−1
× φχ,s−1 ⊗ ψγ,s+1|T
. (24)
ω̃s|T (·; x)q [2] (·, x)gs (x)h2 (x)dx ⊙ γs+1
11
Nguyen et al.
Two-filter approximations of marginal smoothing distributions
Proof. The proof is postponed to Section 5.5.
The decompositions (27) and (28) together with Theorem 4 allow to prove a CLT form
the forward and the backward approximations of the marginal smoothing distribution. Theorem 6 is a direct consequence of Proposition 11, Theorem 4 and Slutsky Lemma.
Theorem 6 (CLT for the TwoFiltbdm algorithm of [1]). Assume that A1, A2 and A3 hold
for some T < ∞. Then, for all 1 ≤ s ≤ T − 1 and all h ∈ Fb (X, X ),
√
N
N ω̃ i,f
X
s|T
i=1
Ω̃fs|T
h(ξsi )
− φχ,s|T [h]
!
D
−→N →∞ N 0, ∆fχ,s|T h − φχ,s|T [h] ,
where
−1
]}2 ,
∆fχ,s|T [h] := Γ̃s,s+1|T Hsf /{φχ,s ⊗ ψγ,s+1|T [q ⊙ γs+1
−1
Hsf (x, x′ ) := h(x)q(x, x′ )γs+1
(x′ ) .
Similarly,
√
N
N ω̃ i,b
X
s|T
i=1
Ω̃bs|T
h(ξsi )
− φχ,s|T [h]
!
D
−→N →∞ N 0, ∆bχ,s|T h − φχ,s|T [h] ,
where
∆bχ,s|T [h] := Γ̃s−1,s|T Hsb /{φχ,s−1 ⊗ ψγ,s|T [q ⊙ γs−1 ]}2 ,
Hsb (x, x′ ) := q(x, x′ )γs−1 (x′ )h(x′ ) .
Note that σs and Σs [h] may be written as:
−1
σs = φχ,s ⊗ ψγ,s+1|T q ⊙ γs+1
× φχ,s−1
Z
q(·, x)gs (x)dx
and by Theorem 4,
Σs [h] = Γχ,s−1
Z
q(·, x)gs (x)h1s+1 (x)dx
+
φ2χ,s−1
Z
q(·, x)gs (x)dx Γ̌γ,s+1|T h2s+1 ,
−1
−1
with h1s+1 (x) := h(x)ψγ,s+1|T [q(x, ·)γs+1
] and h2s+1 (x) := γs+1
(x)φχ,s [h(·)q(·, x)]. In the
case where r̃s|T (xs , xs+1 ; xs ) = ps (xs−1 , xs ) in (15) and ϑ̃s|T (x, x′ ) = ϑs (x)ϑs|T (x′ ), the
smoothing distribution approximation given by the TwoFiltf wt algorithm is obtained by
reweighting the particles obtained in the forward filtering pass and Υχ,s|T [h] may be compared to ∆fχ,s|T [h] as both approximations of φχ,s|T [h] are based on the same particles
(associated with different importance weights). In this case, the two last terms in (24) are
easily interpreted in the case ϑs|T = γs+1 :
i
h
−1
−1
= φχ,s−1 [ϑs ]ψγ,s+1|T [ϑs|T γs+1
] = φχ,s−1 [ϑs ]
φχ,s−1 ⊗ ψγ,s+1|T ϑ̃s|T ⊙ γs+1
12
Nguyen et al.
Two-filter approximations of marginal smoothing distributions
and by Jensen’s inequality,
Z
−1
[2]
2
φχ,s−1 ⊗ ψγ,s+1|T
ω̃s|T (·; x)q (·, x)gs (x)h (x)dx ⊙ γs+1
Z
−2
= φχ,s−1 (dxs−1 )ωs (xs−1 , x)gs (x)q(xs−1 , x)ψγ,s+1|T [q 2 (x, ·)γs+1
]h2 (x)dx ,
Z
≥ φχ,s−1 (dxs−1 )ωs (xs−1 , x)gs (x)q(xs−1 , x)(h1s+1 (x))2 dx .
Therefore, by Proposition 11 and Theorem 6
Γχ,s h1s+1 + Γ̌γ,s+1|T h2s+1
f
Υχ,s|T [h] ≥
2 = ∆χ,s|T [h] ,
−1
φχ,s ⊗ ψγ,s+1|T q ⊙ γs+1
where the last inequality comes from Theorem 4. The same inequality holds for ∆bχ,s|T [h]
when r̃s|T (xs−1 , xs+1 ; xs ) = rs|T (xs+1 , xs ) in (15).
Remark 2. Under the strong mixing assumptions H1 and H2, time uniform bounds for the
asymptotic variances of the two-filter approximations of the marginal smoothing distributions may be obtained.
(i) If A1 and A2 hold uniformly in T and if H1 holds, then it is proved in [6] that there
exists C > 0 such that for all s ≥ 0 and all h ∈ Fb (X, X ), the asymptotic variance
Γχ,s [h] defined in Proposition 10 satisfies:
2
Γχ,s [h] ≤ C |h|∞ .
(ii) Following the same steps, if A1 and A3 hold uniformly in T and if H1 and H2 hold,
there exists C > 0 such that for all 0 ≤ t ≤ T and all h ∈ Fb (X, X ), the asymptotic
variance Γ̌t|T [h] defined in Proposition 11 satisfies:
2
Γ̌γ,t|T [h] ≤ C |h|∞ .
(iii) As a consequence, if A1, A2 and A3 hold uniformly in T and if H1 and H2 hold, the
asymptotic variances Γ̃s,t|T [h], ∆fχ,s|T [h], ∆bχ,s|T [h] and Υχ,s|T [h] defined in Theorem 4, Theorem 5 and Theorem 6 are all uniformly bounded.
5
5.1
Proofs
Proof of Theorem 1
j
j
N
, 1 ≤ j ≤ N ) and
Define Gt|T
:= σ(ξˇt|T
, ω̌t|T
ft|T (x) := Ω̌−1
t|T
N
X
j=1
13
j
j
ω̌t|T
h(x, ξˇt|T
)
Nguyen et al.
Two-filter approximations of marginal smoothing distributions
whose oscillation is bounded by osc (h). By the exponential inequality for the auxiliary
particle filter (Proposition 8), there exist constants Bs and Cs such that
Z
N ω̌ j
N
i ω̌ j
X
X
ω
t|T
t|T
j
j
s
h(ξsi , ξ̌t|T
)−
φχ,s (dxs )h(xs , ξˇt|T
) > ǫ
P
Ω
Ω̌
Ω̌
s
t|T
t|T
j=1
i,j=1
!#
"
N
X ωi
2
2
s
i
N
≤ Bs e−Cs N ǫ / osc (h) . (25)
=E P
ft|T (ξs ) − φχ,s (ft|T ) > ǫ Gt|T
Ωs
i=1
R
Since the oscillation of the function x 7→ φχ,s (dxs )h(xs , x) is bounded by osc (h), by
Proposition 9 there exist constants Bt|T and Ct|T such that
P
Z
N ω̌ j
X
t|T
j=1
Ω̌t|T
j
φχ,s (dxs )h(xs , ξ̌t|T
) − φχ,s ⊗ ψγ,t|T [h] > ǫ
≤ Bt|T e−Ct|T N ǫ
2
/ osc2(h)
, (26)
which concludes the proof.
5.2
Proof of Theorem 2
Define h̃s|T := h − φχ,s|T [h]. Lemma 7 is used with
aN := N −1
N
X
i
i
h̃s|T (ξ˜s|T
),
ω̃s|T
bN := N −1 Ω̃s|T ,
i=1
b :=
−1
q [2] (·; xs )gs (xs )dxs ⊙ γs+1
i
h
.
−1
φχ,s ⊗ ψγ,s+1|T ϑ̃s|T ⊙ γs+1
φχ,s ⊗ ψγ,s+1|T
R
Lemma 7-(i) is satisfied using β := b and |aN |/|bN | ≤ osc (h). To prove Lemma 7-(ii) for
aN , note that Hoeffding inequality implies that, for any ǫ > 0,
(
)
h
i
2
N
ǫ
1
N
1
N
≥ ǫ Gs,T
h̃s|T (ξ˜s|T
) Gs,T
≤ 2 exp −
P aN − E ω̃s|T
,
2
8 ω̃s|T ∞ osc2 (h)
N,+
N,−
N
where Gs,T
:= Gs−1
∨ Gs+1,T
and
GsN,+ := σ {(ωui , ξui )}N
i=1 , u = 1, . . . , s − 1 ,
o
n
N,−
i
i
Gs,T := σ {(ω̌u|T , ξˇu|T )}N
.
i=1 , u = s + 1, . . . , T
On the other hand, for all ℓ ∈ {1, . . . , N },
i
h
N
ℓ
ℓ
) Gs,T
h̃s|T (ξ˜s|T
E ω̃s|T
R [2] i
PN
j
−1 ˇj
i
ˇj
i,j=1 ωs−1 ω̌s+1|T γs+1 (ξs+1|T ) q (ξs−1 , ξs+1|T ; xs )gs (xs )h̃s|T (xs )dxs
=
.
PN
j
j
−1 ˇj
i
i
i,j=1 ωs−1 ω̌s+1|T γs+1 (ξs+1|T )ϑ̃s|T (ξs−1 , ξ̌s+1|T )
14
Nguyen et al.
Two-filter approximations of marginal smoothing distributions
The proof of Lemma 7-(ii) is then completed by applying Lemma 7 to a′N , b′N and b′ defined
by:
a′N
:=
j
i
N
X
ωs−1
ω̌s+1|T
i,j=1
b′N
:=
Ωs−1 Ω̌s+1|T
j
i
N
X
ωs−1
ω̌s+1|T
i,j=1
Ωs−1 Ω̌s+1|T
−1 ˇj
γs+1
(ξs+1|T )
Z
j
i
q [2] (ξs−1
, ξˇs+1|T
; xs )gs (xs )h̃s|T (xs )dxs ,
j
−1 ˇj
i
),
γs+1
(ξs+1|T )ϑ̃s|T (ξs−1
, ξˇs+1|T
−1
b := φχ,s ⊗ ψγ,s+1|T [ϑ̃s|T ⊙ γs+1
].
′
Note first that Lemma 7-(i) is satisfied using β ′ := b′ and |a′N /b′N | ≤ ω̃s|T
addition, by (6),
φχ,s ⊗ ψγ,s+1|T [h̄s|T ] ∝ φχ,s|T [h̃s|T ] = 0 ,
where
h̄s|T (x, x′ ) :=
Z
∞
osc (h). In
−1
q [2] (·; xs )gs (xs )h̃s|T (xs )dxs ⊙ γs+1
(x, x′ ) .
Theorem 1 ensures that Lemma 7-(ii) is satisfied for a′N as
−1
osc h̄s|T ≤ 2 ϑ̃s|T ⊙ γs+1
∞
ω̃s|T
∞
osc (h) .
Similarly, Theorem 1 yields:
P (|b′N − b′ | ≥ ǫ) ≤ Bs e−Cs N ǫ
2
−1
/ osc2(ϑ̃s|T ⊙γs+1
)
,
which proves Lemma 7-(iii) for b′N and concludes the proof of Lemma 7-(ii) for aN . The
proof of Lemma 7-(iii) for bN is along the same lines.
5.3
Proof of Theorem 3
Define
−1
hs (x, x′ ) := γs+1
(x′ )h(x)q(x, x′ ) and hs (x, x′ ) := γs−1 (x′ )q(x, x′ )h(x′ ) .
It follows from the definition of the forward and backward smoothing weights (13) and (14)
that,
PN
i j
−1 −1
i j
N ω̃ i,f
X
Ω̌
Ω
s
i,j=1 ωs ω̌s+1|T hs (ξs , ξ̌s+1|T )
s|T
s+1|T
i
h(ξs ) = −1 −1 PN
,
(27)
j
j
Ω̃fs|T
Ωs Ω̌s+1|T i,j=1 ωsi ω̌s+1|T
1s (ξsi , ξ̌s+1|T
)
i=1
j
j
−1 PN
i
i
N ω̃ i,b
X
Ω−1
s−1 Ω̌s|T
i,j=1 ωs−1 ω̌s|T hs (ξs−1 , ξ̌s|T )
s|T
i
ˇ
h(ξs|T ) = −1 −1 PN
.
(28)
j
j
i
i
Ω̃bs|T
Ωs−1 Ω̌s|T i,j=1 ωs−1
ω̌s|T
1s (ξs−1
, ξ̌s|T
)
i=1
On the other hand, from the definition of the filtering distribution and of the backward
information filter
φχ,s|T [h] = φχ,s ⊗ ψs+1|T hs /φχ,s ⊗ ψs+1|T 1s ,
φχ,s|T [h] = φχ,s−1 ⊗ ψs|T [hs ] /φχ,s−1 ⊗ ψs|T [1s ] .
15
Nguyen et al.
Two-filter approximations of marginal smoothing distributions
Then, (19) is established by writing:
N ω̃ i,f
X
s|T
i=1
Ω̃fs|T
i,f
h(ξsi ) − φχ,s|T [h] = ai,f
N /bN ,
where
ai,f
N
:=
j
N
X
ωsi ω̌s+1|T
i,j=1
bi,f
N :=
Ωs Ω̌s+1|T
j
N
X
ωsi ω̌s+1|T
i,j=1
Ωs Ω̌s+1|T
j
1s (ξsi , ξˇs+1|T
)
(
j
1s (ξsi , ξˇs+1|T
),
j
hs (ξsi , ξˇs+1|T
) φχ,s ⊗ ψs+1|T
−
j
i
ˇ
1s (ξs , ξs+1|T ) φχ,s ⊗ ψs+1|T
b := φχ,s ⊗ ψs+1|T 1s .
)
hs
,
1s
Lemma 7 may then be applied with β := b. Note that
j
hs (ξsi , ξ̌s+1|T
)
φχ,s ⊗ ψs+1|T hs
= h(ξsi ) − φχ,s|T [h] ,
−
j
1s (ξsi , ξ̌s+1|T
) φχ,s ⊗ ψs+1|T 1s
i,f
which ensures that ai,f
N /bN ≤ osc (h) and that Lemma 7-(i) is satisfied. By
−1
,
osc 1s = osc q ⊙ γs+1
i
−1
osc 1s ⊙ h(ξs ) − φχ,s|T [h] ≤ 2 q ⊙ γs+1
∞
osc (h) ,
Theorem 1 shows that Lemma 7-(ii) and (iii) are satisfied. The proof of (20) follows the
exact same lines.
5.4
Proof of Theorem 4
For all 1 ≤ t ≤ T , the result is shown by induction on s where s ∈ {0, . . . , t − 1}. Write
h̃0,t := h − φχ,0 ⊗ ψγ,t|T [h] and set, for i ∈ {1, . . . , N },
UN,i := N
−1/2
ω0i
N ω̌ j
X
t|T
j=1
Ω̌t|T
j
h̃0,t (ξ0i , ξ̌t|T
).
Then,
N
N
i ω̌ j
X
X
√
ω0 t|T
−1
j
UN,i .
h(ξ0i , ξˇt|T
) − φχ,0 ⊗ ψγ,t|T [h] = (Ω0 /N )
N
Ω Ω̌
i,j=1 0 t|T
i=1
j
Define GN,i := σ {ξ0ℓ }ℓ≤i , {ξˇu|T
}t≤u≤T , j = 1, . . . , N . Then,
N
X
i=1
E [ UN,i | GN,i−1 ] = N
1/2
N ω̌ j
X
t|T
j=1
16
Ω̌t|T
h
i
j
ρ0 ω0 h̃0,t (·, ξˇt|T
) .
Nguyen et al.
Two-filter approximations of marginal smoothing distributions
R
As ψγ,t|T (dxt )ρ0 (dx0 )ω0 (x0 )h̃0,t (x0 , xt ) = 0, by the CLT for the backward information
filter (Proposition 11),
N
X
i=1
D
E [ UN,i | GN,i−1 ] −→N →∞ N 0, Γ̌t|T [H0,t ] ,
R
where H0,t (xt ) := ρ0 (dx0 )ω0 (x0 )h̃0,t (x0 , xt ). We now prove that
!
"
!
#
N
2
X
u2 σ0,t|T
[h]
P
{UN,i − E [ UN,i | GN,i−1 ]} GN,0 −→N →∞ exp −
E exp iu
,
2
i=1
where
2
[h] :=
σ0,t|T
Z
2
[h̃0,t (x, ·)] .
ρ0 (dx)ω02 (x)ψγ,t|T
This is done by applying [7, Theorem A.3] which requires to show that
N
X
2
P
2
2
E UN,i
[h] ,
GN,i−1 − E [ UN,i | GN,i−1 ] −→N →∞ σ0,t|T
(29)
i=1
N
X
P
2
E UN,i
1{|UN,i | > ε} GN,i−1 −→N →∞ 0 .
(30)
i=1
By Proposition 9,
2
N ω̌ j
X
t|T
P
2
j
2
E [ UN,i | GN,i−1 ] =
H0,t (ξˇt|T
) −→N →∞ ψγ,t|T
[H0,t ] = 0 .
Ω̌
t|T
i=1
j=1
N
X
On the other hand,
#
" N
X
2
2
E UN,i GN,i−1 − σ0,t|T [h]
E
i=1
2
N ω̌ j
X
t|T
j
2
= ρ0 (dx)ω02 (x)E
h̃0,t (x, ξˇt|T
) − ψγ,t|T
[h̃0,t (x, ·)] ,
Ω̌
j=1 t|T
Z
≤ 2 osc (h) ρ0 (dx)ω02 (x)E [AN (x)] ,
Z
where
AN (x) :=
N ω̌ j
X
t|T
j=1
Ω̌t|T
j
h̃0,t (x, ξˇt|T
) − ψγ,t|T [h̃0,t (x, ·)] .
By Proposition 9, there exist Bt|T and Ct|T such that for all x ∈ X,
E [AN (x)] =
Z
0
∞
P (AN (x) ≥ ε) dε
≤ Bt|T
Z
∞
e−Ct|T N ǫ
0
17
2
/ osc(h)2
dε ≤ Dt|T osc (h) N −1/2 , (31)
Nguyen et al.
Two-filter approximations of marginal smoothing distributions
which shows that
N
X
i=1
2
P
2
E UN,i
[h]
GN,i−1 −→N →∞ σ0,t|T
and concludes the proof of (29). For all N ≥ 1,
n
o
−1
{|UN,i | ≥ ε} ⊆ ω0i ≥ εN 1/2 osc (h)
,
which implies that
N
X
i=1
Z
2
2
E UN,i
1{|UN,i| ≥ ε} GN,i−1 ≤ osc (h)
o
n
−1
ρ0 (dx)ω02 (x)1 ω0 (x) ≥ N 1/2 osc (h)
and (30) follows by letting N → ∞. Note that
Z
P
−1
N Ω0 −→N →∞ χ(dx0 )g0 (x0 ) ,
which shows (23) since
−2
i Z
h
χ(dx0 )g0 (x0 )
Γ̃0,t|T h̃0,t =
Z
2
× Γ̌γ,t|T [H0,t ] + ρ0 (dx)ω02 (x)ψγ,t|T
[h̃0,t (x, ·)] ,
Z
Z
= Γ̌γ,t|T
φχ,0 (dx0 )h̃0,t (x0 , ·) + Γχ,0
ψγ,t|T (dxt )h̃0,t (·, xt ) .
Assume now that the result holds for some s − 1. Write h̃s,t := h − φχ,s ⊗ ψγ,t|T [h] and set,
for i ∈ {1, . . . , N },
N ω̌ j
X
t|T
j
−1/2 i
h̃s,t (ξsi , ξˇt|T
).
UN,i := N
ωs
Ω̌
j=1 t|T
Then,
N
N
i ω̌ j
X
X
√
ωs t|T
−1
j
N
UN,i .
h(ξsi , ξˇt|T
) − φχ,0 ⊗ ψγ,t|T [h] = (Ωs /N )
Ω Ω̌
i,j=1 s t|T
i=1
Define, for 1 ≤ i ≤ N ,
GN,i
Then,
N
X
i=1
j
ξs
:= σ
i
j=1
, ξuℓ
n
oN
j
ˇ
, ξv|T
, 1 ≤ u < s, t ≤ v ≤ T .
ℓ=1
N
j=1
E [ UN,i | GN,i−1 ] = φN
χ,s−1 [ϑs ]
where
Hs,t (xs−1 , xt ) :=
Z
−1
N 1/2
j
N
i
X
ω̌t|T
ωs−1
j
i
Hs (ξs−1
, ξ̌t|T
),
Ω
Ω̌
i,j=1 s−1 t|T
q(xs−1 , x)gs (x)h̃s,t (x, xt )dx .
18
(32)
Nguyen et al.
Two-filter approximations of marginal smoothing distributions
Since φχ,s−1 ⊗ ψγ,t|T [Hs,t ] = 0, by the induction assumption,
N
X
i=1
D
E [ UN,i | GN,i−1 ] −→N →∞ N 0, Γ̃s−1,t|T [Hs,t ] /φ2χ,s−1 [ϑs ] .
We will now prove that
#
!
"
!
2
N
X
u2 σs,t|T
[h]
P
{UN,i − E [ UN,i | GN,i−1 ]} GN,0 −→N →∞ exp −
E exp iu
,
2
i=1
where
2
σs,t|T
[h] := φχ,s−1 [ϑs ]−1 φχ,s−1 [fs−1,t ] ,
Z
2
[h̃s,t (xs , ·)]gs (xs )dxs .
fs−1,t (xs−1 ) := q(xs−1 , xs )ωs (xs−1 , xs )ψγ,t|T
2
This is done using again [7, Theorem A.3] and proving that (29) and (30) hold with σ0,t|T
[h]
2
replaced by σs,t|T [h]. Note that
2
j
N
i
X
ω̌
ω
2
j
s−1 t|T
i
2
E [ UN,i | GN,i−1 ] =
Hs,t (ξs−1
, ξˇt|T
) /(φN
χ,s−1 [ϑs ]) ,
Ω
Ω̌
s−1
t|T
i=1
i,j=1
N
X
which converges in probability to 0 by Theorem 1 and the fact that φχ,s−1 ⊗ ψγ,t|T [Hs,t ] = 0.
In addition,
Z
N
N
i
i
X
2
−1 X
q 2 (ξs−1
, xs )gs (xs )
ωs−1
N
E UN,i GN,i−1 = φχ,s−1 [ϑs ]
i
i
Ωs−1
ϑs (ξs−1 )ps (ξs−1
, xs )
i=1
i=1
2
N
[h̃s,t (xs , ·)] gs (xs )dxs ,
× ψγ,t|T
Z
N
i
X
ωs−1
i
i
=
ωs (ξs−1
, xs )q(ξs−1
, xs )
Ω
s−1
i=1
2
N
[h̃s,t (xs , ·)] gs (xs )dxs ,
× ψγ,t|T
N
−1 N
,
= φN
φχ,s−1 fs−1,t
χ,s−1 [ϑs ]
−1
φN
χ,s−1 [ϑs ]
where
N
fs−1,t
(xs−1 ) :=
Z
2
N
[h̃s,t (xs , ·)] gs (xs )dxs .
q(xs−1 , xs )ωs (xs−1 , xs ) ψγ,t|T
P
First note that φN
χ,s−1 [ϑs ] −→N →∞ φχ,s−1 [ϑs ] and write
N
N
N
φN
χ,s−1 fs−1,t − φχ,s−1 [fs−1,t ] ≤ As,t + Bs,t ,
N
N
N
N
N
where AN
s,t := |φχ,s−1 [fs−1,t ] − φχ,s−1 [fs−1,t ]| and Bs,t := |φχ,s−1 [fs−1,t ] − φχ,s−1 [fs−1,t ]|.
j
j
i
i
N
As (ωs−1
, ξs−1
)N
i=1 and (ω̌t|T , ξ̌t|T )j=1 are independent,
"N
#
Z
h
i
i
X ωs−1
N
N
i
E As,t ≤ |ωs |∞ |gs |∞ E
[h̃s,t ](xs ) dxs ,
q(ξs−1
, xs )E ∆ψγ,t|T
Ω
s−1
i=1
19
Nguyen et al.
Two-filter approximations of marginal smoothing distributions
N
N
2
where ∆ψγ,t|T
[h̃s,t ](xs ) := (ψγ,t|T
[h̃s,t (xs , ·)])2 − ψγ,t|T
[h̃s,t (xs , ·)]. Following the same steps
as in (31), there exists Dt|T such that
h
i
√
2
N
E ∆ψγ,t|T
[h̃s,t ](xs ) ≤ 2Dt|T osc (h) / N ,
which yields
E AN
s,t
#
Z
N
i
X
√
ωs−1
i
≤ 2 osc (h) |ωs |∞ |gs |∞ DT |t E
q(ξs−1
, xs )dxs / N ,
Ωs−1
i=1
√
≤ 2 osc (h)2 |ωs |∞ |gs |∞ DT |t / N
"
2
P
2
N
and E AN
s,t −→N →∞ 0. On the other hand, as osc (fs−1,t ) ≤ osc (h) |ωs |∞ |gs |∞ , Bs,t −→N →∞
0 by Proposition 8. Finally, the tightness condition (30) holds since |UN,i | ≤ N −1/2 |ωs |∞ osc (h).
Note that,
Z
P
N −1 Ωs −→N →∞ φχ,s−1
q(·, xs )gs (xs )dxs /φχ,s−1 [ϑs ] .
Therefore (23) holds with
Γ̃s,t|T
i
h
h̃s,t =
=
φ2χ,s−1 [ϑs ]
R
φ2χ,s−1 q(·, xs )gs (xs )dxs
φ2χ,s−1
(
Γ̃s−1,t|T [Hs,t ] φχ,s−1 [fs−1,t ]
+
φ2χ,s−1 [ϑs ]
φχ,s−1 [ϑs ]
)
,
Γ̃s−1,t|T [Hs,t ]
φ
[f
]φ
[ϑ ]
+ 2 χ,s−1R s−1,t χ,s−1 s ,
R
q(·, xs )gs (xs )dxs
φχ,s−1 q(·, xs )gs (xs )dxs
where, by induction assumption,
Γ̃s−1,t|T [Hs,t ] = Γχ,s−1
Z
ψγ,t|T (dxt )Hs,t (·, xt )
+ Γ̌γ,t|T
Z
φχ,s−1 (dxs−1 )Hs,t (xs−1 , ·) .
The proof is completed upon noting that
R
R
φχ,s−1 (dxs−1 )q(xs−1 , xs )gs (xs )h̃s,t (xs , ·)dxs
φχ,s−1 (dxs−1 )Hs (xs−1 , ·)
=
R
R
,
φχ,s−1 q(·, xs )gs (xs )dxs
φχ,s−1 q(·, xs )gs (xs )dxs
Z
= φχ,s (dxs )h̃s,t (xs , ·)
and, by Proposition 10,
Γχ,s−1
"
#
ψγ,t|T (dxt )Hs,t (·, xt )
φ
[f
]φ
[ϑ ]
+ 2 χ,s−1R s−1,t χ,s−1 s
R
φχ,s−1 q(·, xs )gs (xs )dxs
φχ,s−1 q(·, xs )gs (xs )dxs
Z
= Γχ,s
ψγ,t|T (dxt )h̃s,t (·, xt ) .
R
20
Nguyen et al.
5.5
Two-filter approximations of marginal smoothing distributions
Proof of Theorem 5
Write h̃s,T = h − φχ,s|T (h). Note that
√
N
N
i
X
ω̃s|T
i=1
where UN,ℓ := N
−1/2
Ω̃s|T
N
−1 X
i
) = Ω̃s|T /N
UN,i ,
h̃s,T (ξ˜s|T
ℓ
ℓ
h̃s,T (ξ˜s|T
).
ω̃s|T
i=1
Set, for i ∈ {1, . . . , N },
n
ℓ
ℓ
)}iℓ=1 , {(ωuℓ , ξuℓ )}N
, ξ̃s|T
GN,i := σ {(ω̃s|T
ℓ=1 , u = 0, . . . , s − 1,
ℓ
ℓ
{(ω̌v|T
, ξˇv|T
)}N
ℓ=1 , v = s + 1, . . . , T
By the proof of Theorem 2,
N
−1
P
Ω̃s|T −→N →∞
o
.
−1
q [2] (·, x)gs (x)dx ⊙ γs+1
i
h
.
−1
φχ,s−1 ⊗ ψs+1|T ϑ̃s|T ⊙ γs+1
φχ,s−1 ⊗ ψs+1|T
R
PN
The proof therefore amounts to establish a CLT for
ℓ=1 UN,ℓ and then to use Slutsky
PN
Lemma. The limit distribution of ℓ=1 UN,ℓ is again obtained using the invariance principle
for triangular array of dependent random variables derived in [7]. As
N
X
i=1
E [ UN,i | GN,i−1 ]
√
= N
PN
i,j=1
R [2]
j
j
−1
i
i
ωs−1
ω̌s+1|T
q (·; xs )gs (xs )h̃s,T (xs )dxs ⊙ γs+1
(ξs−1
, ξˇs+1|T
)
,
PN
j
j
−1 ˇj
i
i
ˇ
i,j=1 ωs−1 ω̌s+1|T γs+1 (ξs+1|T )ϑ̃s|T (ξs−1 , ξs+1|T )
it follows from Theorems 1 and 4 that
N
X
i=1
Using that
P
E [ UN,i | GN,i−1 ] −→N →∞ N 0,
φχ,s|T [h̃s,T ] =
i=1
−1
φχ,s−1 ⊗ ψs+1|T [ϑ̃s|T ⊙ γs+1
]
2 .
i
−1
q [2] (·; x)gs (x)h̃s,T (x)dx ⊙ γs+1
R
=0,
−1
φχ,s−1 ⊗ ψs+1|T
q [2] (·; x)gs (x)dx ⊙ γs+1
φχ,s−1 ⊗ ψs+1|T
Theorem 1 yields
N
X
Σs [h̃s,T ]
hR
2
E [ UN,i | GN,i−1 ] =
R
!2
j
j
−1 ˇj
i
i
ωs−1
ω̌s+1|T
γs+1
(ξs+1|T ) q [2] (ξs−1
, ξˇs+1|T
; x)gs (x)h̃s,T (x)
PN
j
j
−1 ˇj
i
i
i,j=1 ωs−1 ω̌s+1|T γs+1 (ξs+1|T )ϑ̃s|T (ξs−1 , ξ̌s+1|T )
i 2
hR
−1
[2]
φ
⊗
ψ
q
(·;
x)g
(x)
h̃
(x)dx
⊙
γ
χ,s−1
s
s,T
s+1|T
s+1
P
=0.
−→N →∞
−1
φχ,s−1 ⊗ ψs+1|T [ϑ̃s|T ⊙ γs+1 ]
PN
i,j=1
21
Nguyen et al.
Two-filter approximations of marginal smoothing distributions
Similarly, using again Theorem 1,
N
X
i=1
2
P
E UN,i
GN,i−1 −→N →∞
φχ,s−1 ⊗ ψs+1|T
hR
−1
ω̃s|T (·; x)q [2] (·; x)gs (x)h̃2s,T (x)dx ⊙ γs+1
−1
φχ,s−1 ⊗ ψs+1|T [ϑ̃s|T ⊙ γs+1
]
Since under A2, |UN,i | ≤ N −1/2 ω̃s|T
N
X
i=1
∞
osc (h), for any ǫ > 0,
P
2
E UN,i
1{|UN,i | ≥ ǫ} GN,i−1 −→N →∞ 0 ,
which concludes the proof.
22
i
.
Nguyen et al.
A
Two-filter approximations of marginal smoothing distributions
Exponential deviation inequalities for the forward filter and the backward information filter
The following result is proved in [6].
Lemma 7. Assume that aN , bN , and b are random variables defined on the same probability
space such that there exist positive constants β, B, C, and M satisfying
(i) |aN /bN | ≤ M , P-a.s. and b ≥ β, P-a.s.,
2
(ii) For all ǫ > 0 and all N ≥ 1, P [|aN | > ǫ] ≤ Be−CN (ǫ/M) ,
2
(iii) For all ǫ > 0 and all N ≥ 1, P [|bN − b| > ǫ] ≤ Be−CN ǫ .
Then, for all ε > 0,
P
(
2 )
aN
ǫβ
.
> ǫ ≤ B exp −CN
bN
2M
Proposition 8 provides an exponential deviation inequality for the forward filter and is
proved in [6].
Proposition 8. Assume that A1 and A2 hold for some T > 0. Then, for all s ≥ 1, there
exist 0 < Bs , Cs < ∞ such that for all N ≥ 1, ǫ > 0, and all h ∈ Fb (X, X ),
!
N
X
2
2
ωsi h(ξsi ) − φχ,s [h] ≥ ǫ ≤ Bs e−Cs N ǫ / osc(h) .
P Ω−1
s
i=1
Proposition 9 provides an exponential inequality for the backward information filter
ψγ,t|T and its unnormalized approximation. Its proof is similar to the proof [6, Theorem 5]
and is omitted.
Proposition 9. Assume that A1 and A3 hold for some T > 0. Then, for all 0 ≤ t ≤ T ,
there exist 0 < Bt|T , Ct|T < ∞ such that for all N ≥ 1, ǫ > 0, and all h ∈ Fb (X, X ),
P
B
"
−1
Ω̌t|T
N
X
i=1
i
i
)
h(ξˇt|T
ω̌t|T
#
− ψγ,t|T [h] ≥ ǫ ≤ Bt|T e−Ct|T N ǫ
2
/ osc(h)2
.
Asymptotic normality of the forward filter and the
backward information filter
Proposition 10 provides a CLT for the weighted particles {(ωsi , ξsi )}N
i=1 approximating the
filtering distribution φχ,s and is proved for instance in [4].
Proposition 10. Assume that A1 and A2 hold for some T > 0. Then, for all 0 ≤ s ≤ T
and all h ∈ Fb (X, X ),
N
1/2
N
X
ωsi
h(ξsi ) − φχ,s [h]
Ω
s
i=1
!
D
−→N →∞ N (0, Γχ,s [h − φχ,s [h]]) ,
23
Nguyen et al.
Two-filter approximations of marginal smoothing distributions
where
ρ0 (dx0 )ω02 (x0 )h2 (x0 )
and for all s ≥ 1 ,
2
R
ρ0 (dx0 )ω0 (x0 )
R
Γχ,s−1 q(·, xs )gs (xs )h(xs )dxs
R
Γχ,s [h] :=
φ2χ,s−1 q(·, xs )gs (xs )dxs
R
φχ,s−1 ωs (·, xs )q(·, xs )gs (xs )h2 (xs )dxs φχ,s−1 [ϑs ]
R
+
.
φ2χ,s−1 q(·, xs )gs (xs )dxs
Γχ,0 [h] :=
R
j
j
Proposition 11 provides a CLT for the weighted particles {(ω̌t|T
, ξ̌t|T
)}N
j=1 approximating
the backward information filter. Its proof follows the same lines as the proof of Proposition 10
and is omitted for brevity.
Proposition 11. Assume that A1 and A3 hold. Then, for all 0 ≤ t ≤ T and all h ∈
Fb (X, X ),
N ω̌ j
X
t|T
D
j
N 1/2
h(ξˇt|T
) − ψγ,t|T [h] −→N →∞ N 0, Γ̌γ,t|T h − ψγ,t|T [h] ,
Ω̌t|T
j=1
where
ρ̌T (dxT )ω̌T2 |T (xT )h2 (xT )
and for all t ≤ T − 1 ,
R
2
ρ̌T (dxT )ω̌T |T (xT )
R
−1
Γ̌γ,t+1|T
γt (xt )gt (xt )q(xt , ·)γt+1
(·)h(xt )dxt
R
[h] :=
−1
2
ψγ,t+1|T
γt (xt )gt (xt )q(xt , ·)γt+1
(·)dxt
R
−1
−1
ψγ,t+1|T
ω̌t (xt , ·)q(xt , ·)gt (xt )γt (xt )γt+1
(·)h2 (xt )dxt ψγ,t+1|T ϑt|T γt+1
R
.
+
−1
2
γt (xt )gt (xt )q(xt , ·)γt+1
ψγ,t+1|T
(·)dxt
Γ̌γ,T |T [h] :=
Γ̌γ,t|T
R
References
[1] M. Briers, A. Doucet, and S. Maskell. Smoothing algorithms for state-space models.
Annals Institute Statistical Mathematics, 62(1):61–89, 2010.
[2] O. Cappe. Online EM algorithm for hidden Markov models. Journal of Computational
and Graphical Statistics, 20(73):728–749, 2011.
[3] O. Cappé, E. Moulines, and T. Rydén. Inference in Hidden Markov Models. Springer,
2005.
[4] P. Del Moral. Feynman-Kac Formulae. Genealogical and Interacting Particle Systems
with Applications. Springer, 2004.
[5] P. Del Moral, A. Doucet, and S. Singh. A Backward Particle Interpretation of FeynmanKac Formulae. ESAIM M2AN, 44(5):947–975, 2010.
24
Nguyen et al.
Two-filter approximations of marginal smoothing distributions
[6] R. Douc, A. Garivier, E. Moulines, and J. Olsson. Sequential Monte Carlo smoothing
for general state space hidden Markov models. Ann. Appl. Probab., 21(6):2109–2145,
2011.
[7] R. Douc and E. Moulines. Limit theorems for weighted samples with applications to
sequential Monte Carlo methods. Ann. Statist., 36(5):2344–2376, 2008.
[8] R. Douc, E. Moulines, and D.S. Stoffer. Nonlinear time series: theory, methods and
applications with R examples. CRC Press, 2014.
[9] A. Doucet, N. de Freitas, and N. Gordon, editors. Sequential Monte Carlo methods in
practice. Springer, 2001.
[10] A. Doucet, S. Godsill, and C. Andrieu. On sequential Monte-Carlo sampling methods
for Bayesian filtering. Stat. Comput., 10:197–208, 2000.
[11] C. Dubarry and S. Le Corff. Non-asymptotic deviation inequalities for smoothed additive functionals in nonlinear state-space models. Bernoulli, 19(5B):2222–2249, 2013.
[12] P. Fearnhead, D. Wyncoll, and J. Tawn. A sequential smoothing algorithm with linear
computational cost. Biometrika, 97(2):447–464, 2010.
[13] S. J. Godsill, A. Doucet, and M. West. Monte Carlo smoothing for non-linear time
series. J. Am. Statist. Assoc., 50:438–449, 2004.
[14] N. Gordon, D. Salmond, and A.F. Smith. Novel approach to nonlinear/non-Gaussian
bayesian state estimation. IEE Proc. F, Radar Signal Process, 140:107–113, 1993.
[15] M. Hürzeler and H. R. Künsch. Monte Carlo approximations for general state-space
models. J. Comput. Graph. Statist., 7:175–193, 1998.
[16] P.E. Jacob, L.M. Murray, and S. Rubenthaler. Path storage in the particle filter.
Statistics and Computing, pages 1–10, 2013.
[17] N. Kantas, A. Doucet, S.S. Singh, J. Maciejowski, and N. Chopin. On particle methods
for parameter estimation in state-space models. Statist. Sci., 30(3):328–351, 2015.
[18] G. Kitagawa. Monte-Carlo filter and smoother for non-Gaussian nonlinear state space
models. J. Comput. Graph. Statist., 1:1–25, 1996.
[19] H. R. Künsch. Recursive Monte Carlo filters: Algorithms and theoretical analysis. Ann.
Statist., 33(5):1983–2021, 2005.
[20] S. Le Corff and G. Fort. Convergence of a particle-based approximation of the block
online Expectation Maximization algorithm. ACM Transactions on Modeling and Computer Simulation, 23(1):2, 2013.
[21] S. Le Corff and G. Fort. Online Expectation Maximization based algorithms for inference in hidden Markov models. Electronic Journal of Statistics, 7:763–792, 2013.
[22] G. Mongillo and S. Deneve. Online learning with hidden Markov models. Neural
Computation, 20(7):1706–1716, 2008.
25
Nguyen et al.
Two-filter approximations of marginal smoothing distributions
[23] J. Olsson and J. Westerborn. Efficient particle-based online smoothing in general hidden
Markov models: the PaRIS algorithm. ArXiv:1412.7550, 2015.
[24] M. K. Pitt and N. Shephard. Filtering via simulation: Auxiliary particle filters. J. Am.
Statist. Assoc., 94(446):590–599, 1999.
[25] G. Poyiadjis, A. Doucet, and S.S Singh. Particle approximations of the score and
observed information matrix in state space models with application to parameter estimation. Biometrika, 98:65–80, 2011.
26
| 10math.ST
|
arXiv:0708.0505v1 [cs.AI] 3 Aug 2007
Alma Mater Studiorum
Università degli Studi di Bologna
DEIS
A preliminary analysis on
metaheuristics methods applied to
the Haplotype Inference Problem
Luca Di Gaspero
Andrea Roli
August 3, 2007
DEIS Technical Report no. DEIS-LIA-006-07
LIA Series no. 84
A preliminary analysis on metaheuristics methods
applied to the Haplotype Inference Problem
Luca Di Gaspero 1
1
DIEGM, University of Udine
via delle Scienze 208, I-33100, Udine, Italy
[email protected]
Andrea Roli 2
2
DEIS, Campus of Cesena
University of Bologna
via Venezia 52, I-47023 Cesena, Italy
[email protected]
August 3, 2007
Abstract. Haplotype Inference is a challenging problem in bioinformatics that
consists in inferring the basic genetic constitution of diploid organisms on the
basis of their genotype. This information allows researchers to perform association studies for the genetic variants involved in diseases and the individual
responses to therapeutic agents.
A notable approach to the problem is to encode it as a combinatorial problem
(under certain hypotheses, such as the pure parsimony criterion) and to solve
it using off-the-shelf combinatorial optimization techniques. The main methods
applied to Haplotype Inference are either simple greedy heuristic or exact methods (Integer Linear Programming, Semidefinite Programming, SAT encoding)
that, at present, are adequate only for moderate size instances.
We believe that metaheuristic and hybrid approaches could provide a better
scalability. Moreover, metaheuristics can be very easily combined with problem specific heuristics and they can also be integrated with tree-based search
techniques, thus providing a promising framework for hybrid systems in which
a good trade-off between effectiveness and efficiency can be reached.
In this paper we illustrate a feasibility study of the approach and discuss
some relevant design issues, such as modeling and design of approximate solvers
that combine constructive heuristics, local search-based improvement strategies
and learning mechanisms. Besides the relevance of the Haplotype Inference
problem itself, this preliminary analysis is also an interesting case study because
the formulation of the problem poses some challenges in modeling and hybrid
metaheuristic solver design that can be generalized to other problems.
Keywords: Metaheuristics, haplotype inference, bioinformatics
Contents
1 Introduction
3
2 The Haplotype Inference problem
2.1 Genotype resolution . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.2 Compatibility and complementarity . . . . . . . . . . . . . . . . . . .
4
4
5
3 Local Search models for Haplotype Inference
3.1 Search space . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.2 Cost function . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
6
7
8
4 Metaheuristic approaches
4.1 Iterative minimization of the number of haplotypes
4.1.1 Hamming neighborhoods . . . . . . . . . .
4.1.2 Deletion/insertion . . . . . . . . . . . . . .
4.1.3 A hybrid strategy: Iterated local search . .
4.2 Minimization and feasibility approach . . . . . . .
4.2.1 Dynamic local search . . . . . . . . . . . . .
4.2.2 Adaptive constructive approach . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
9
9
10
10
10
11
11
12
5 Constructive procedures for initial solutions
13
6 Discussion and future work
6.1 Haplotype cardinality reduction . . . . . . . . . . . . . . . . . . . . .
13
14
DEIS Technical Report no. DEIS-LIA-006-07
LIA Series no. 84
1
Introduction
A fundamental tool of analysis to investigate the genetic variations in a population
is based on haplotype data. A haplotype is a copy of a chromosome of a diploid
organism (i.e., an organism that has two copies of each chromosome, one inherited
from the father and one from the mother). The haplotype information allows to
perform association studies for the genetic variants involved in diseases and the
individual responses to therapeutic agents. The assessment of a full Haplotype Map
of the human genome is indeed one of the current high priority tasks of human
genomics [21].
Instead of dealing with complete DNA sequences, usually the researchers are
focusing on Single Nucleotide Polymorphisms (SNPs), which are the most common
mutations among haplotypes. A SNP is a single nucleotide site (allele) where exactly
two (out of four) different nucleotides occur in a large percentage of the population.
The haplotype collection is not an easy task: in fact, due to technological limitations it is currently infeasible to directly collect haplotypes in an experimental
way, but rather it is possible to collect genotypes, i.e., the conflation of a pair of
haplotypes. Moreover, instruments can only identify whether the individual is homozygous (i.e., the alleles are the same) or heterozygous (i.e., the alleles are different)
at a given site. Therefore, haplotypes have to be inferred from genotypes in order
to reconstruct the detailed information and trace the precise structure of human
populations. This process is called Haplotype Inference and the goal is to find a set
of haplotype pairs so that all the genotypes are resolved.
The main approaches to solve the Haplotype Inference are either combinatorial
or statistical methods. However, both of them, being of non-experimental nature,
need some genetic model of haplotype evolution, which poses some hypotheses to
constrain the possible inferences to the ones compatible with Nature. In the case
of the combinatorial methods, which are the subject of the present work, a reasonable criterion is the pure parsimony approach [7], which searches for the smallest
collection of distinct haplotypes that solves the Haplotype Inference problem. This
criterion is consistent with current observations in natural populations for which
the actual number of haplotypes is vastly smaller than the total number of possible
haplotypes.
Current approaches for solving the problem include simple greedy heuristic [4]
and exact methods such as Integer Linear Programming [3, 7, 9, 14], Semidefinite
Programming [11, 13] and SAT models [16, 17]. These approaches, however, at
present are adequate only for moderate size instances.
To the best of our knowledge, the only attempt to employ metaheuristic techniques for the problem is a recently proposed Genetic Algorithm [22]. However, also
the cited paper does not report results on real size instances. Anyway, we believe
that metaheuristic approaches could be effective to this problem since they could
3
provide a better scalability.
In this work we make a preliminary conceptual analysis on the use of metaheuristics for the Haplotype Inference problem. We start introducing the Haplotype
Inference problem in Section 2 and then we present two possible local search models
for the problem (Section 3) highlighting the possible benefits and drawbacks of each
model. Section 4 contains the description of metaheuristic approaches that, in our
opinion, could be adequate for Haplotype Inference. In Section 5 we consider the
role of constructive techniques in the hybridization with metaheuristics and, finally,
in Section 6 we discuss our proposals and outline future developments.
2
The Haplotype Inference problem
In the Haplotype Inference problem we deal with genotypes, that is, strings of length
m that corresponds to a chromosome with m sites. Each value in the string belongs
to the alphabet {0, 1, 2}. A position in the genotype is associated with a site of
interest on the chromosome (e.g., a SNP) and it has value 0 (wild type) or 1 (mutant)
if the corresponding chromosome site is a homozygous site (i.e., it has that state on
both copies) or the value 2 if the chromosome site is heterozygous. A haplotype is a
string of length m that corresponds to only one copy of the chromosome (in diploid
organisms) and whose positions can assume the symbols 0 or 1.
2.1
Genotype resolution
Given a chromosome, we are interested in finding an unordered1 pair of haplotypes
that can explain the chromosome according to the following definition:
Definition 1 (Genotype resolution). Given a chromosome g, we say that the pair
hh, ki resolves g, and we write hh, ki ⊲ g (or g = h ⊕ k), if the following conditions
hold (for j = 1, . . . , m):
g[j] = 0 ⇒ h[j] = 0 ∧ k[j] = 0
(1a)
g[j] = 1 ⇒ h[j] = 1 ∧ k[j] = 1
(1b)
g[j] = 2 ⇒ (h[j] = 0 ∧ k[j] = 1) ∨ (h[j] = 1 ∧ k[j] = 0)
(1c)
If hh, ki ⊲ g we indicate the fact that the haplotype h (respectively, k) contributes
in the resolution of the genotype g writing h E g (resp., k E g). We extend this
notation to set of haplotypes and we write H = {h1 , . . . , hl } E g, meaning that
hi E g for all i = 1, . . . , l.
1
In the problem there is no distinction between the maternal and paternal haplotypes.
4
Conditions (1a) and (1b) require that both haplotypes must have the same value
in all homozygous sites, while condition (1c) states that in heterozygous sites the
haplotypes must have different values.
Observe that, according to the definition, for a single genotype string the haplotype values at a given site are predetermined in the case of homozygous sites, whereas
there is a freedom to choose between two possibilities at heterozygous places. This
means that for a genotype string with l heterozygous sites there are 2l−1 possible
pairs of haplotypes that resolve it.
As an example, consider the genotype g = (0212), then the possible pairs of
haplotypes that resolve it are h(0110), (0011)i and h(0010), (0111)i.
After these preliminaries we can state the Haplotype Inference problem as follows:
Definition 2 (Haplotype Inference problem). Given a population of n individuals, each of them represented by a genotype string gi of length m we are interested in finding a set R of n pairs of (not necessarily distinct) haplotypes R =
{hh1 , k1 i, . . . , hhn , kn i}, so that hhi , ki i ⊲ gi , i = 1, . . . , n. We call H the set of haplotypes used in the construction of R, i.e., H = {h1 , . . . , hn , k1 , . . . , kn }.
Of course, from the mathematical point of view, there are many possibility in
constructing the set R since there is an exponential number of possible haplotypes for
each genotype. However, the knowledge about the biological phenomenons should
guide the selection of the haplotypes used by constraining it through a genetic model.
One natural model to the Haplotype Inference problem is the already mentioned
pure parsimony approach that consists in searching for a solution that minimizes the
total number of distinct haplotypes used or, in other words, |H|, the cardinality of
the set H. A trivial upper bound for |H| is 2n in the case of all genotypes resolved
by a pair of distinct haplotypes.
It has been shown that the Haplotype Inference problem under the pure parsimony hypothesis is APX-hard [14] and therefore NP-hard.
2.2
Compatibility and complementarity
It is possible to define a graph that express the compatibility between genotypes,
so as to avoid unnecessary checks in the determination of the resolvents.2 Let us
build the graph G = (G, E), in which the set of vertices coincides with the set of the
genotypes; in the graph, a pair of genotypes g1 , g2 are connected by an edge whether
they are compatible, i.e., one or more common haplotypes can resolve both of them.
For example, the genotypes (2210) and (1220) are compatible, whereas genotypes
(2210) and (1102) are not compatible. The formal definition of this property is as
follows.
2
In some cases, also a graph representing incompatibilities between genotypes can provide useful
information.
5
Definition 3 (Genotypes compatibility). Let g1 and g2 be two genotypes, g1 and
g2 are compatible if, for all j = 1, . . . , m, the following conditions hold:
g1 [j] = 0 ⇒ g2 [j] ∈ {0, 2}
(2a)
g1 [j] = 1 ⇒ g2 [j] ∈ {1, 2}
(2b)
g2 [j] = 2 ⇒ g2 [j] ∈ {0, 1, 2}
(2c)
The same concept can be expressed also between a genotype and a haplotype as
in the following definition.
Definition 4 (Compatibility between genotypes and haplotypes). Let g be a genotype and h a haplotype, g and h are compatible if, for all j = 1, . . . , m, the following
conditions hold:
g[j] = 0 ⇒ h[j] = 0
(3a)
g[j] = 1 ⇒ h[j] = 1
(3b)
g[j] = 2 ⇒ h[j] ∈ {0, 1}
(3c)
We denote this relation with h 7→ g. Moreover with an abuse of notation we indicate
with h 7→ {g1 , g2 , . . . } the set of all the genotypes that are compatible with haplotype
h.
Observe that the set of compatible genotypes of a haplotype can contain only
mutually compatible genotypes (i.e., they form a clique in the compatibility graph).
Another interesting observation is the following. Due to the resolution definition,
when one of the two haplotypes composing the pair, say h, has been selected, then
the other haplotype can be directly inferred from h and the genotype g thanks to
the resolution conditions.
Proposition 1 (Haplotype complement). Given a genotype g and a haplotype h 7→
g, there exists a unique haplotype k such that h ⊕ k = g. The haplotype k is called
the complement of h with respect to g and is denoted with k = g ⊖ h.
Proof. The existence and uniqueness of k is a direct consequence of Conditions (1a)–
(1c).
3
Local Search models for Haplotype Inference
We start our conceptual analysis of metaheuristic approaches for Haplotype Inference
with the basic building blocks of local search methods. Indeed, in order to apply
this class of methods to a given problem we need to specify three entities, namely
the search space, the cost function and the neighborhood relation, that constitute
the so-called local search model of the problem.
6
In the following subsections we consider some natural choices for the search space
and the cost function, and we try to envisage their benefits and drawbacks on the
basis of the general experience on different problem domains. Instead, the discussion
of the possible neighborhood relations is provided in the next section together with
the general strategies for tackling the problem.
3.1
Search space
As the search space for the Haplotype Inference problem we could either adopt a
complete or incomplete representation as explained below.
Complete representation: This search space is based on a straightforward encoding of the problem. In this representation we consider, for each genotype
g, the pair of haplotypes hh, ki that resolves it. Therefore in this representation
all the genotypes are fully resolved at each state by construction. The search
space is therefore the collection of sets R defined as in the problem statement.
Notice that, thanks to complementarity w.r.t. a genotype just one haplotype
can be selected to represent a pair of resolvents. Moreover, since the pair is
unordered, the haplotype pair hh, ki is equal to hk, hi, therefore we can break
this symmetry by selecting the haplotype h as the representative for the state
if h ≺ k, where the symbol ≺ indicates the lexicographic precedence of the two
strings, otherwise we select k as the representative.
Incomplete representation: In the incomplete representation, instead, we deal
with sets of haplotypes that are the elements of the set H. An element h ∈ H
is a potential resolvent of a genotype g, and it actually resolves it only whether
its companion k, so that hh, ki ⊲ g, also belongs to H. As a consequence, in
this representation not necessarily
Sn all the genotypes have a resolvent and the
search space is the powerset P( i=1 Ai ), where Ai = {h|∃khh, ki⊲gi } is the set
of all potential resolvents of a genotype gi of the problem. This formulation is
an incomplete representation of the problem, since the genotype resolution is
only potential and a solution of the problem must be constructed on the basis
of the haplotypes selected in a given state.
The complete representation is in general more adequate for hybridization with
constructive methods since it maintains the feasibility (i.e., the resolution of all
genotypes) at every state of the search. For example it makes it possible to employ
local search in a GRASP-like manner [20] or to hybridize it with more informed
strategies making use problem-specific knowledge.
On the other hand, in some problem domains, this type of representation could
constrain too much the search so that it is difficult to explore different areas of the
search space because of the impossibility to relax feasibility constraints. For this
7
reason, sophisticated metaheuristic approaches must be employed to improve the
effectiveness of methods based on this representation.
Moreover, this formulation makes the neighborhood design more difficult since
the search can move only from and to fully resolved sets of genotypes. A possibility
to deal with this complication could be the fine-grain hybridization with declarative
neighborhood formulations that explicitly encode the feasibility constraints (e.g.,
those expressed through Constraint Programming or Integer Linear Programming).
On the contrary, the incomplete representation does not guarantee feasibility
at every search step since in a given state it is not mandatory to resolve all the
genotypes. As a consequence, it commonly happens that in this kind of search
spaces a great part of the search effort focuses in uninteresting areas of the search
space just to pursue a feasible solution.
Nevertheless, this representation permits more freedom to explore different areas
of the search space and a higher degree of flexibility. For example, it allows an easier
design of neighborhood relations.
3.2
Cost function
As for the cost function we identify different components related either to optimality
or to feasibility of the problem.
A natural component is the objective function of the original problem, that is
the cardinality |H| of the set of haplotypes employed in the resolution.
f1 = |H|
(4)
In the incomplete representation, however, we must also include the so-called
distance to feasibility, which can simply be measured as the number of genotypes
not resolved in the current solution.
f2 = |{g|g is not resolved from haplotypes in H}|
(5)
Moreover in both formulations we might want to include some heuristic measure
related to the potential quality of the solution. To this respect, a possible measure
could be the number of genotypes resolved by each haplotype.
f3 =
X
|{g|g can be resolved by haplotype h}|
(6)
h∈H
Components f1 and f2 have to be minimized in the cost function whereas, in
contrast, f3 must be maximized. To simplify the formulation we prefer to deal only
with minimization components, therefore we consider the function f3′ = n|H| − f3
(n|H| is an upper bound on function f3 ) as the component to be included in the
cost function.
8
The cost function F is then the weighted sum of the three components:
F = α1 f1 + α2 f2 + α3 f3′
(7)
in which the weights α1 , α2 and α3 must be chosen for the problem at hand to reflect
the trade-offs among the different components.
Notice that α2 can be zero in the case of the complete representation of the
search space. Conversely, α2 must be a value much greater than the possible values
of the other components in the case of the incomplete representation in order to
reflect the importance of feasibility.
4
Metaheuristic approaches
In this section we discuss some possible metaheuristic approaches to tackle the problem and we emphasize their strengths and weaknesses.
In the following discussion we will assume states encoded with the incomplete
formulation, even if most of the considerations holds also for the complete formulation. Indeed, the difference between the two is that the latter is associated to a
cost function in which the distance from a feasible solution is not considered, as all
states are feasible.
In general, the problem can be solved either by trying to solve it for a given
cardinality k = |H|, progressively decreased during search, or by trying to minimize |H| while searching at the same time a set of haplotypes that resolve all the
genotypes. The first approach is often preferred over the former as it can be easily
reduced to a sequence of feasibility problems (see, for example, the case of graph
coloring problem family [10]). Nevertheless, this method might require a very high
execution time, because a quite long series of feasibility problems has often to be
solved. In the next subsections we detail some promising formulations, using either
of the mentioned approaches, which can be effectively employed by metaheuristic
strategies such as Tabu Search, Iterated local search, Variable neighborhood search,
etc. [2]. Indeed, the choice of the actual metaheuristic is in general orthogonal with
respect to the formulation.
4.1
Iterative minimization of the number of haplotypes
This approach is based on the classical method of iteratively solving the problem
by decreasing values of k = |H|. Hence, the problem is reduced to find k different
haplotypes such that all the genotypes can be resolved. A solution to the problem is
represented by a set of distinct haplotypes. Neighborhoods can be defined in many
ways; in the following we discuss what we believe are the most promising ones.
9
4.1.1
Hamming neighborhoods
As it is often the case with solutions encoded as boolean variables, also in this
case we can consider a neighborhood based on Hamming distance. In particular, a
good trade-off between exploration and execution time is the 1-Hamming distance
neighborhood w.r.t. each haplotype in the current solution. The exploration of such
a neighborhood has a time complexity of O(nk), where k is the number of haplotypes
and n the number of sites per haplotype. Besides its low time complexity, this
neighborhood has the advantage of making it very easy to generate more complex
neighborhoods just by concatenating 1-Hamming distance moves. The drawback of
this choice relies in its local character of exploring neighboring solutions, that might
not help search escape from areas in which it has been attracted.
4.1.2
Deletion/insertion
A complementary neighborhood structure is the one that does not pose restrictions
on the distance among solutions. The neighborhoods based on deletion/insertion
consist in deleting an haplotype in the solution and inserting a new one. The criteria upon which haplotypes are deleted and inserted define the actual neighborhood;
moreover, the way the neighborhood is explored represents another degree of freedom.
The most general neighborhood relation would be to try the move deletion/insertion for all the possible combinations, i.e., for each haplotype in the solution, try
to substitute it with each possible new haplotype and check the corresponding cost
function. Since the number of haplotypes with n sites is 2n , this choice is in general
not feasible. A more practical solution, instead, consists of choosing the new haplotype among the ones stored in a kind of candidate set of given cardinality. Among
the possible ways of generating the candidate set we may distinguish between random and heuristically generated. The general neighborhood relation can thus be
defined as follows:
N (s) = {M |M = N (s) \ s′ ∪ z, s′ ∈ N (s), z ∈ C}
(8)
where s is the current solution and C the candidate set. Both the elements and
cardinality of C can be kept constant or varied during search. Since the complexity
of enumerating all the neighborhood is O(|H| · |C|), an efficient balance between
exploration and time complexity has to be set.
4.1.3
A hybrid strategy: Iterated local search
The characteristics of the previous described neighborhoods can be combined in
a single metaheuristic such as Iterated local search [15]. This local search metaheuristic uses the first neighborhood (1-Hamming) with a local search metaheuristic
10
(e.g., hill climbing, tabu search, etc.) until stagnation, then a run of another local
search method with the insertion/deletion neighborhood is performed in order to
move away from the attractor (perturbation phase). The strategy is described in
algorithm 1.
Algorithm 1 Two neighborhoods Iterated Local Search
s0 ← GenerateInitialSolution()
s∗ ← LocalSearch(s0 ) with 1-Hamming neighborhood
while termination conditions not met do
s′ ← LocalSearch(s∗ ) with I/D neighborhood {perturbation phase}
′
s∗ ← LocalSearch(s′ ) with 1-Hamming neighborhood
′
s∗ ← ApplyAcceptanceCriterion(s∗ , s∗ , history)
end while
4.2
Minimization and feasibility approach
The second approach for tackling the Haplotype Inference problem defines a search
strategy that tries to minimize |H| and resolve all the genotypes at the same time.
In such a case, it is possible that some genotypes are not resolved during search,
therefore also states which are infeasible w.r.t. the original problem formulations are
explored during search. We will illustrate two possible strategies for implementing
metaheuristics based on this problem formulation.
4.2.1
Dynamic local search
The core of a possible local search strategy for this model strongly relies on an effective penalty mechanism that should be able to guide the search toward ‘good’
(i.e., at low values of |H|) and feasible (i.e., all genotypes are resolved) states. A
very effective technique for achieving such a search guidance is what is usually called
Dynamic local search [10], in which the cost function is composed of two (or more)
components, each with a weight that is dynamically varied during search. In formulas, referring to the cost functions defined in section 3.2:
F
= w1 f1 + w2 f2
(9)
= w1 |H| + w2 |{g|g is not resolved from haplotypes in H}|
The coefficients w1 and w2 are varied during search by trying to explore feasible
states with |H| as low as possible. Effective techniques for achieving this goal can be
inspired by Guided local search, as discussed in [2] and, more in general, Dynamic
local search strategies described in [10]. The high level description of this metaheuristic is reported in algorithm 2. A very effective scheme used for weight update
11
is the shifting penalty method, that, in this case, can be implemented as follows.
First of all, w1 and w2 have to be initialized in such a way that the contribution
of w2 f2 (feasibility) is higher than the optimization contribution of w1 f1 . Then,
during search, if the set of haplotypes resolve all the genotypes for K consecutive
iterations, w2 is reduced (by a small random factor γ > 1), otherwise it is increased
(by a comparable value).
Algorithm 2 Dynamic Local Search for Haplotype Inference problem
s ← GenerateInitialSolution()
while termination conditions not met do
s ← LocalSearch(s)
Update(weights w1 , w2 )
end while
4.2.2
Adaptive constructive approach
Another very promising approach, which deals with both complete and incomplete
representation of the search space, is inspired by Strategic oscillation [6] and Squeaky
wheel search [12]. The idea is to expand and reduce the set of haplotypes in solution
by trying to minimize the combined cost function F defined above. The difference
between dynamic local search is that the focus of this algorithm is based on the
selection of elements to be deleted and to be inserted, i.e., the critical decision is
on the delete and insert sets. Several effective techniques can be employed, such
as the already mentioned Strategic oscillation and Squeaky wheel search; moreover,
some incomplete iterative constructive heuristics could be effectively applied, such
as Incomplete dynamic backtracking [19].
As an example, we briefly discuss one of the possible implementations of such
a strategy for tackling the Haplotype Inference problem. In this implementation,
haplotypes are added one at a time and we use a probabilistic criterion for accepting
the new haplotype. As shown in algorithm 3, we start with an empty set of haplotypes and we incrementally add one element at a time. The criteria upon which the
candidate set is built and the new haplotype is chosen can be based on heuristics or
randomness or both. The new element is always added to the current solution if it
decreases the cost function (i.e., the overall quality of the solution increases); even
if it does not improve the current solution, it can be added anyway in a simulated
annealing fashion.
When the cardinality of |H| reaches its maximum value (parameter of the algorithm) some haplotypes are removed from the solution and the insertion restarted
on this smaller set. As one can easily see, there are many choices a designer has
to make and the implementation and comparison of these techniques is subject of
ongoing work.
12
Algorithm 3 Adaptive constructive metaheuristic for the Haplotype Inference problem
s ← ∅ {Initial solution is an empty set}
while termination conditions not met do
s′ ← s ∪ h′ , h′ ∈ C {Choose haplotype to add from the candidate set}
∆ ← F (s′ ) − F (s)
if ∆ < 0 then
s ← s′
else
With probability p(∆, history): s ← s′
end if
if |H| > maxCard then
Delete q haplotypes from s
end if
end while
5
Constructive procedures for initial solutions
Constructive heuristics produce often an important contribution to the effectiveness
of the metaheuristic. Initial sets of haplotypes can be constructed on the basis of
several heuristics. For instance, variations of the Clark’s algorithm and the other
rule-based methods described in [8] can be employed. Moreover, some SAT/Constraint Programming-based techniques can be used to reduce the search space for
the initial solution construction.
One critical factor the designer has to take into account is whether to prefer
small |H| over number of resolved genotypes or viceversa. In the first case, a better
choice would be to start from an empty set, while in the second a complete set
(corresponding to the complete formulation) would be preferable. Furthermore,
constructive procedures should also be employed for generating the candidate sets,
as this could make the search more effective.
6
Discussion and future work
We have presented a feasibility study on the application of metaheuristics to the
Haplotype Inference problem. The main purpose of this work was to point out
critical design issues about the problem in order to guide future developments and
to foster further research on metaheuristic approaches to this problem. Indeed,
we believe that the Haplotype Inference problem could become a relevant problem
subject of application of metaheuristic techniques. However, besides the relevance
of the Haplotype Inference problem itself, this preliminary analysis has posed some
13
challenges in modeling hybrid metaheuristic solver design that can be generalized to
other problems.
The result of this analysis is in some sense a roadmap of the work that has still
to be done. The first steps consist in implementing the described metaheuristic
solvers for the problem and evaluating them on some benchmarks. To this respect we plan to use a set of publicly available datasets (simulated or extracted
from the HapMap project), which are provided by J. Marchini at the website
http://www.stats.ox.ac.uk/~marchini/phaseoff.html.
Afterwards, we will investigate the ways in which domain specific knowledge can
be integrated in the metaheuristics. As an example, it is possible to easily construct
(in)compatibility relations between genotypes and it will be interesting to exploit
this information to make the search more effective. Other ways to integrate domain
specific information is through the hybridization with constructive techniques or
with other search paradigms (e.g., Constraint Programming [1]).
Finally, in future work we intend also to take into consideration different kind of
metaheuristics such as population-based algorithms (e.g., Ant-Colony Optimization
[5] or Memetic algorithms [18]).
Some of the metaheuristics described in this paper have been already implemented and subject of ongoing testing. We are currently designing and implementing
advanced metaheuristics for the complete and incomplete formulation.
Appendix: Further observations on the problem
To the best of our knowledge, there have been no attempts to exploit structural
properties of the problem which can be deduced from compatibility graphs, or other
problem representations. In this section, we present a reduction procedure that
starts from a set of haplotypes in the complete representation and tries to reduce its
cardinality by exploiting compatibility properties of the instance. Other heuristics
based on graph representation of the problem are subject of ongoing work.
6.1
Haplotype cardinality reduction
Let us illustrate this property with an example. Consider the following set of genotypes, which corresponds to the compatibility graph in Figure 1.
g1 : (2210212)
g3 : (1212122)
g5 : (1202201)
g2 : (2112110)
g4 : (1222122)
Now consider the following set of haplotypes that resolves all the genotypes
g1 , . . . , g5 . The set consists of 8 distinct haplotypes; genotypes currently resolved by
a given haplotype are denoted by bold typeface in the resolvent list (i.e., we write g
if g D h ∈ H).
14
g2
g1
g4
g5
g3
Figure 1: Compatibility graph
g1
g2
g3
(
a = (1110110)
⊳
b = (0010010)
(
c = (0111110)
⊳
a = (1110110)
(
d = (1111101)
⊳
e = (1010110)
7 {g1 , g2 , g3 , g4 }
→
7→ {g1 }
g4
7→ {g2 }
7→ {g1 , g2 , g3 , g4 }
g5
(
f = (1101101)
⊳
p = (1010101)
(
f = (1101101)
⊳
q = (1000001)
7 {g4 , g5 }
→
7→ {g3 , g4 }
7 {g4 , g5 }
→
7→ {g5 }
7 {g3 , g4 }
→
7→ {g1 , g3 , g4 }
Notice that the haplotype a = (1110110), resolving g1 , is compatible also with
genotypes g2 , g3 and g4 and, as a matter of fact, it resolves g2 . This configuration
is depicted in the bipartite graph of Figure 2 (dashed edges between genotypes
represent genotype compatibility) in which the haplotypes are represented by square
nodes, the compatibility between haplotypes and genotypes is represented by solid
edges and a bold edge represents the current resolution of a genotype by a haplotype.
We will call this graph extended (compatibility) graph. The constraint on genotypes
resolution is mapped onto the extended graph by imposing that every genotype
node must have (at least) two bold edges. The goal of the reduction procedure is to
try to decrease the number of distinct haplotypes, i.e., the number of square nodes
while satisfying the resolution constraint. The intuition behind the procedure is
that a possible way of reducing the haplotype number is to resolve a genotype by
a haplotype that is compatible, but not currently resolving it, i.e., changing an arc
from solid to bold. Of course, this move must be followed by repairing moves in the
graph so that the state is still feasible.These moves consists in adding one or more
haplotypes and relinking some nodes.
From the situation described above, we can use the resolvent of g1 to resolve g3
or g4 , however the situation is different in the two cases. For example, if we use the
15
c
a
g2
f
g1
b
g4
g5
p
q
g3
e
d
Figure 2: Haplotype resolution graph
resolvent a of g1 to resolve g3 , then the situation will become:
g3
(
a = (1110110)
⊳
r = (1011101)
7 {g1 , g2 , g3 , g4 }
→
7→ {g3 , g4 }
As an effect of the reduction, the total number of haplotypes employed in the
solution has been decreased by 1, since now haplotypes d and e resolving g3 are
replaced by a, which was already a member of the haplotype set, and a new haplotype
r (see Figure 3).
Instead, if we use the resolvent a to resolve g4 we obtain the following situation
(see Figure 4):
g4
(
a = (1110110)
⊳
s = (1001101)
7 {g1 , g2 , g3 , g4 }
→
7→ {g4 }
Differently from the previous reduction, in this case the number of haplotypes
in the solution has not changed. The reason of this is that the haplotype f was
16
c
a
g2
b
f
g1
g4
g5
p
q
g3
r
Figure 3: The haplotype resolution graph after the reduction of g3
already shared between the resolutions of g4 and g5 , therefore the reduction operation
removes one haplotype but it introduces also a new one, leaving the total number
of haplotypes unchanged.
The situation of the example can be generalized in the following proposition.
Proposition 2 (Haplotype local reduction). Given n genotypes G = {g1 , . . . , gn }
and the resolvent set R = {hh1 , bi, . . . , hhn , kn i}, so that hhi , ki i ⊲ gi . Suppose there
exist two genotypes g,(
g′ ∈ G such that:
(
h′ 7→ {g′ , . . .}
h 7→ {g, g ′ , . . .}
(10)
, g′ ⊳
g⊳
k′ 7→ {g′ , . . .}
k 7→ {g, . . .}
and h 6= h′ , h 6= k′ , h′ E A, k′ E B.
The replacement of hh′ , k′ i with hh, g ′ ⊖ hi in the resolution of g′ is a correct
resolution that employs a number of distinct haplotypes according to the following
criteria:
• if |A| = 1 and |B| = 1, the new resolution uses at most one less distinct
haplotype;
17
c
a
g2
b
f
g1
g4
g5
s
g3
q
r
Figure 4: The haplotype resolution graph after the reduction of g4
• if |A| > 1 and |B| = 1 (or symmetrically, |A| = 1 and |B| > 1), the new
resolution uses at most the same number of distinct haplotypes;
• in the remaining case the new resolution uses at most one more distinct haplotype.
Proof. The proof of the proposition is straightforward. The resolution is obviously
correct because h is compatible with g′ and g′ ⊖ h is the complement of h with
respect to g′ .
Concerning the validity of the conditions on the cardinality, let us proceed by
cases and first consider the situation in which g′ ⊖ h does not resolve any other
genotype but g′ .
If |A| = |B| = 1, then h′ and k′ are not shared with other genotype resolutions
so they will not appear in the set H after the replacement, therefore since in the
new resolution h is shared between g and g′ the cardinality of H is decreased by
one.
Conversely, if one of the sets |A| or |B| consists of more than a genotype and the
other set of just one genotype, there is no guarantee of obtaining an improvement
from the replacement. Indeed, since one of the two haplotypes is already shared with
18
another genotype there is just a replacement of the shared haplotype with another
one in the set H.
Finally, when |A| > 1 and |B| > 1 both h′ and k′ are shared with other genotypes
therefore the replacement introduces the new haplotype g ′ ⊖ h in the set H.
Moving to the situation in which g ′ ⊖ h resolves also other genotypes, the same
considerations apply; additionally, given that g′ ⊖ h is already present in H, the
number of distinct haplotypes employed in the resolution is decreased by one. For
this reason the estimation of the changes of |H| is conservative.
Note 1. Perhaps some additional observation could be made in the case of homozygous genotypes g ⊳ hh, ki.
A preliminary experimental analysis on the application of the haplotype local
reduction shows that the results achieved are very promising.
References
[1] Krzysztof R. Apt. Principles of Constraint Programming. Cambridge University
Press, Cambridge, UK, 2003. ISBN 978-0-521-82583-0.
[2] Christian Blum and Andrea Roli. Metaheuristics in combinatorial optimization:
Overview and conceptual comparison. ACM Computing Surveys, 35(3):268–308,
September 2003.
[3] Daniel G. Brown and Ian M. Harrower. Integer programming approaches to
haplotype inference by pure parsimony. IEEE/ACM Transactions on Computational Biology and Bioinformatics, 3(2):141–154, 2006.
[4] Andrew G. Clark. Inference of haplotypes from PCR-amplified samples of
diploid populations. Molecular Biology and Evolution, 7:111–122, 1990.
[5] Marco Dorigo and Thomas Stützle. Ant Colony Optimization. MIT Press,
Cambridge, MA, USA, 2004. ISBN 978-0-262-04219-2.
[6] Fred W. Glover and Manuel Laguna. Tabu Search. Kluwer Academic Publishers,
Norwell, MA, USA, 1997. ISBN 978-0-7923-9965-0.
[7] Dan Gusfield. Haplotype inference by pure parsimony. In Ricardo A. BaezaYates, Edgar Chávez, and Maxime Crochemore, editors, Combinatorial Pattern
Matching (CPM 2003), Proceedings of the 14th Annual Symposium, volume
2676 of Lecture Notes in Computer Science, pages 144–155, Berlin-Heidelberg,
Germany, 2003. Springer-Verlag. ISBN 3-540-40311-6.
[8] Dan Gusfield. Inference of haplotypes from sample diploid populations: complexity and algorithms. Journal of Computational Biology, 8:305–323, 2001.
19
[9] Bjarni V. Halldórsson, Vineet Bafna, Nathan Edwards, Ross Lippert, Shibu
Yooseph, and Sorin Istrail. A survey of computational methods for determining
haplotypes. In Sorin Istrail, Michael S. Waterman, and Andrew G. Clark,
editors, Computational Methods for SNPs and Haplotype Inference, volume 2983
of Lecture Notes in Computer Science, pages 26–47. Springer, 2002. ISBN 3540-21249-3.
[10] H.H. Hoos and T. Stützle. Stochastic Local Search Foundations and Applications. Morgan Kaufmann Publishers, San Francisco, CA (USA), 2005. ISBN
978-1-55860-872-6.
[11] Yao-Ting Huang, Kun-Mao Chao, and Ting Chen. An approximation algorithm
for haplotype inference by maximum parsimony. In Hisham Haddad, Lorie M.
Liebrock, Andrea Omicini, and Roger L. Wainwright, editors, Proceedings of
the 2005 ACM Symposium on Applied Computing (SAC 2005), pages 146–150.
ACM, 2005. ISBN 1-58113-964-0.
[12] D.E. Joslin and D.P. Clements. “Squeaky Wheel” Optimization. Journal of
Artificial Intelligence Research, 10:353–373, 1999.
[13] Konstantinos Kalpakis and Parag Namjoshi. Haplotype phasing using semidefinite programming. In BIBE, pages 145–152. IEEE Computer Society, 2005.
ISBN 0-7695-2476-1.
[14] Giuseppe Lancia, Maria Cristina Pinotti, and Romeo Rizzi. Haplotyping populations by pure parsimony: Complexity of exact and approximation algorithms.
INFORMS Journal on Computing, 16(4):348–359, 2004.
[15] H. Ramalhino Lourenço, O. Martin, and T. Stützle. Iterated Local Search. To
appear in: F. Glover, G. Kochenberger, Handbook of Metaheuristics, 2001.
[16] Inês Lynce and João Marques-Silva. Efficient haplotype inference with boolean
satisfiability. In Proceedings of the 21st National Conference on Artificial Intelligence and the Eighteenth Innovative Applications of Artificial Intelligence
Conference, Menlo Park, CA, USA, 2006. AAAI Press.
[17] Inês Lynce and João Marques-Silva. SAT in bioinformatics: Making the case
with haplotype inference. In Armin Biere and Carla P. Gomes, editors, SAT,
volume 4121 of Lecture Notes in Computer Science, pages 136–141, BerlinHeidelberg, Germany, 2006. Springer-Verlag. ISBN 3-540-37206-7.
[18] Pablo Moscato. On evolution, search, optimization, genetic algorithms and
martial arts: Towards memetic algorithms. Technical Report 826, Caltech
Concurrent Computation Program (C3P), 1989.
20
[19] S. Prestwich. Combining the scalability of local search with the pruning techniques of systematic search. Annals of Operations Research, 115:51–72, 2002.
[20] Mauricio G. C. Resende and Celso C. Ribeiro. Greedy randomized adaptive
search procedures. In Fred Glover and Gary A. Kocheberger, editors, Handbook
of Metaheuristics, volume 57 of International Series in Operations Research &
Management Science, pages 219–249. Kluwer Academic Publishers, Norwell,
MA, USA, 2003. ISBN 1-4020-7263-5.
[21] The International HapMap Consortium. The international HapMap project.
Nature, 426:789–796, 2003.
[22] Rui-Sheng Wang, Xiang-Sun Zhang, and Li Sheng. Haplotype inference by
pure parsimony via genetic algorithm. In Xiang-Sun Zhang, De-Gang Liu,
and Ling-Yun Wu, editors, Operations Research and Its Applications: the Fifth
International Symposium (ISORA’05), Tibet, China, August 8–13, volume 5 of
Lecture Notes in Operations Research, pages 308–318. Beijing World Publishing
Corporation, Beijing, People Republic of China, 2005.
21
| 5cs.CE
|
Fast Meta-Learning for Adaptive Hierarchical Classifier
Design
Gerrit J.J. van den Burg1
Alfred O. Hero2
1
2
Econometric Institute, Erasmus University Rotterdam, The Netherlands
Electrical Engineering and Computer Science, University of Michigan, USA
arXiv:1711.03512v1 [cs.LG] 9 Nov 2017
November 10, 2017
Abstract
We propose a new splitting criterion for a meta-learning approach to multiclass
classifier design that adaptively merges the classes into a tree-structured hierarchy of increasingly difficult binary classification problems. The classification tree is
constructed from empirical estimates of the Henze-Penrose bounds on the pairwise
Bayes misclassification rates that rank the binary subproblems in terms of difficulty of classification. The proposed empirical estimates of the Bayes error rate are
computed from the minimal spanning tree (MST) of the samples from each pair of
classes. Moreover, a meta-learning technique is presented for quantifying the onevs-rest Bayes error rate for each individual class from a single MST on the entire
dataset. Extensive simulations on benchmark datasets show that the proposed hierarchical method can often be learned much faster than competing methods, while
achieving competitive accuracy.
Keywords: classifier performance, meta-learning, multiclass classification, hierarchical classifier, classifier comparison
1
Introduction
The Bayes error rate (BER) is a central concept in the statistical theory of classification.
It represents the error rate of the Bayes classifier, which assigns a label to an object
corresponding to the class with the highest posterior probability. By definition, the
Bayes error represents the smallest possible average error rate that can be achieved by
any decision rule (Wald, 1947). Because of these properties, the BER is of great interest
both for benchmarking classification algorithms as well as for the practical design of
classification algorithms. For example, an accurate approximation of the BER can be
used for classifier parameter selection, data dimensionality reduction, or variable selection. However, accurate BER approximation is difficult, especially in high dimension,
and thus much attention has focused on tight and tractable BER bounds. This paper
proposes a model-free approach to designing multiclass classifiers using a bias-corrected
BER bound estimated directly from the multiclass data.
There exists several useful bounds on the BER that are functions of the classdependent feature distributions. These include information theoretic divergence measures such as the Chernoff α-divergence (Chernoff, 1952), the Bhattacharyya divergence
(Kailath, 1967), or the Jensen-Shannon divergence (Lin, 1991). Alternatively, arbitrarily tight bounds on performance can be constructed using sinusoidal or hyperbolic
1
approximations (Avi-Itzhak and Diep, 1996, Hashlamoun et al., 1994). These bounds
are functions of the unknown class-dependent feature distributions.
Recently, Berisha et al. (2016) introduced a divergence measure belonging to the
family of f -divergences which tightly bounds the Bayes error rate in the binary classification problem. The bounds on the BER obtained with this measure are tighter than
bounds derived from the Bhattacharyya or Chernoff bounds. Moreover, this divergence
measure can be estimated nonparametrically from the data without resorting to density
estimates of the distribution functions. Inspired by the Friedman-Rafsky multivariate
runs test (Friedman and Rafsky, 1979), estimation is based on computing the Euclidean
minimal spanning tree (MST) of the data, which can be done in approximately O(n log n)
time. In this paper we propose improvements to this estimator for problems when there
are unequal class priors and apply the improved estimator to the adaptive design of a
hierarchical multiclass classifier. Furthermore, a fast method is proposed for bounding
the Bayes error rate of individual classes which only requires computing a single minimal spanning tree over the entire set of samples. Thus our proposed method is faster
than competing methods that use density plug-in estimation of divergence or observed
misclassification rates of algorithms, such as SVM or logistic regression, which involve
expensive parameter tuning.
Quantifying the complexity of a classification problem has been of significant interest (Ho and Basu, 2002) and it is clear that a fast and accurate estimate of this
complexity has many practical applications. For instance, an accurate complexity estimator allows the researcher to assess a priori whether a given classification problem is
difficult to classify or not. In a multiclass problem, a pair of classes which are difficult
to disambiguate could potentially be merged or could be designated for additional data
collection. Moreover, an accurate estimate of the BER could be used for variable selection, an application that was explored previously in Berisha et al. (2016). In Section 3
further applications of the BER estimates to multiclass classification are presented and
evaluated.
There are many methods available for the design of multiclass classification algorithms, including: logistic regression (Cox, 1958); support vector machines (Cortes and
Vapnik, 1995); and neural networks (McCulloch and Pitts, 1943). It is often the case
that classifier performance will be better for some classes than for others, for instance
due to sample imbalance in the training set. Often classifier designs apply weights to
the different classes in order to reduce the effect of such imbalances on average classifier
accuracy (Lu et al., 1998, Qiao and Liu, 2009). We take a different and more general approach that incorporates an empirical determination of the relative difficulties of
classifying between different classes. Accurate empirical estimates of the BER are used
for this purpose. A multiclass classifier is presented in Section 4 that uses MST-based
BER estimates to create a hierarchy of binary subproblems that increase in difficulty
as the algorithm progresses. This way, the classifier initially works on easily decidable
subproblems before moving on to more difficult multiclass classification problems.
The paper is organized as follows. The theory of the nonparametric Bayes error
estimator of Berisha et al. (2016) will be reviewed in Section 2. We will introduce a bias
correction for this estimator, motivate the use of the estimator for multiclass classification, and discuss computational complexity. Section 3 will introduce applications of the
estimator to meta-learning in multiclass classification. A novel hierarchical classification
method will be introduced and evaluated in Section 4. Section 5 provides concluding
remarks.
2
2
An Improved BER Estimator
Here the motivation and theory for the estimator of the Bayes error rate is reviewed,
as introduced by Berisha et al. (2016). An improvement on this estimator is proposed
for the case where class prior probabilities are unequal. Next, the application of the
estimator in multiclass classification problems is considered. Finally, computational
considerations and robustness analyses are presented.
2.1
Estimating the Bayes Error Rate
Consider the binary classification problem on a dataset D = {(xi , yi )}i=1,...,n , where
xi ∈ X ⊆ Rd and yi ∈ {1, 2}. Denote the multivariate density functions for the two
classes by f1 (x) and f2 (x) and the prior probabilities by p1 and p2 = 1−p1 , respectively.
The Bayes error rate of this binary classification problem can be expressed as (Fukunaga,
1990)
Z
Pe (f1 , f2 ) =
min {p1 f1 (x), p2 f2 (x)} dx.
(1)
Recently, Berisha et al. (2016) derived a tight bound on the BER which can be
estimated directly from the data without a parametric model for the density or density
estimation. This bound is based on a divergence measure introduced by Berisha and
Hero (2015), defined as
"Z
#
1
(p1 f1 (x) − p2 f2 (x))2
DHP (f1 , f2 ) =
dx − (p1 − p2 )2 ,
(2)
4p1 p2
p1 f1 (x) + p2 f2 (x)
and called the Henze-Penrose divergence, as it is motivated by an affinity measure defined
by Henze and Penrose (1999). In Berisha and Hero (2015) it was shown that (2) is a
proper f -divergence as defined by Csiszár (1975).
Estimation of the Henze-Penrose (HP) divergence is based on the multivariate runs
test proposed by Friedman and Rafsky (1979) and convergence of this test was studied
by Henze and Penrose (1999). Let Xk = {xi ∈ X ⊆ Rd : yi = k} with k ∈ {1, 2} denote
multidimensional features from two classes. Define the class sample sizes nk = |Xk |,
k ∈ {1, 2}. Let the combined sample be denoted by X = X1 ∪X2 where n = |X| = n1 +n2
is the total number of samples from both classes. Define the complete graph G over X
as the graph connecting all n nodes {xi }ni=1 with edge weights |ekj | = kxk − xj k equal
to Euclidean distances. The Euclidean minimal spanning tree that spans X, denoted
by T , is defined as the subgraph of G that is both connected and whose sum of edge
weights is the smallest possible. The Friedman-Rafsky test statistic equals the number
of edges in T that connect an observation from class 1 to an observation from class 2
and is denoted by R1,2 .
Building on the work by Henze and Penrose (1999), Berisha et al. (2016) show that
if n1 → ∞ and n2 → ∞ in a linked manner such that n1 /(n1 + n2 ) → δ ∈ (0, 1) then,
1−
nR1,2
→ DHP (f1 , f2 )
2n1 n2
a.s.
(3)
Thus, the number of cross connections between the classes in the Euclidean MST is
inversely proportional to the divergence between the respective probability density functions of these classes.
Finally, the HP-divergence can be used to bound the Bayes error rate, Pe (f1 , f2 ),
following Theorem 2 of Berisha et al. (2016)
p
1
1
uHP (f1 , f2 ) ≤ Pe (f1 , f2 ) ≤ 21 − 12 uHP (f1 , f2 ),
(4)
2 − 2
3
0.6
0.6
BER, p = 0.15
P̂e , p = 0.33
PBC , p = 0.33
0.4
0.3
0.2
0.1
0
P̂e , p = 0.15
PBC , p = 0.15
BER, p = 0.33
0.5
Error Probability
0.5
Error Probability
BER, p = 0.15
P̂e , p = 0.15
PBC , p = 0.15
BER, p = 0.33
P̂e , p = 0.33
PBC , p = 0.33
0.4
0.3
0.2
0.1
0
1
2
3
Distance
4
0
5
(a) without bias correction
0
1
2
3
Distance
4
5
(b) with bias correction
Figure 1: Illustration of the estimates of the Bayes error rate for two different values
of the prior probability p and for spherical bivariate Gaussian distributions with increasing separation between the classes as measured by Euclidean distance between the
class means, kµ1 − µ2 k. We compare the true BER, the Bhattacharyya bound (PBC )
requiring the true class distributions, and the HP-estimator (P̂e ) based on an empirical
sample. Figures (a) and (b) show the estimates for p = 0.15 and p = 1/3, respectively
with and without the proposed bias correction of the HP-estimator. The HP-estimator
was implemented using n = 1000 samples (n1 = pn and n2 = n − n1 ) and averaged over
200 trials.
where
uHP (f1 , f2 ) = 4p1 p2 DHP (f1 , f2 ) + (p1 − p2 )2 .
Averaging these bounds yields an estimate of the BER given by
p
P̂e (f1 , f2 ) = 21 − 41 uHP (f1 , f2 ) − 14 uHP (f1 , f2 ).
(5)
(6)
In the following, this estimator will be referred to as the HP-estimator of the BER.
2.2
A modified HP-estimator for unequal class priors
To illustrate the performance of the HP-estimator, and to motivate the proposed modification, consider a binary classification problem where the samples are drawn from two
independent bivariate Gaussian distributions with equal covariance matrices. For this
example the BER and associated bounds can be computed exactly (Fukunaga, 1990).
In Figure 1a we compare the BER, the HP-estimator of the BER (6) and the popular
Bhattacharyya bound on the BER (Bhattacharyya, 1946, Kailath, 1967). Figure 1a
shows that the HP-estimator is closer to the true BER than the Bhattacharyya bound.
This result was illustrated by Berisha et al. (2016) for the case where p1 = p2 and is
confirmed here for the case where p1 6= p2 .
However, Figure 1a also shows that a significant bias occurs in the HP-estimate of
the BER when the distance between classes is small. Considering that Pe (f1 , f2 ) =
4
min{p1 , p2 } if f1 → f2 , the solution of the equation P̂e (f1 , f2 ) = min{p̂1 , p̂2 } for R1,2
suggests a bias corrected version of the HP-estimate:1
0
R1,2
= min{γ, R1,2 },
(7)
with
γ = 2n min{p̂1 , p̂2 } − 43 n + 14 n
p
9 − 16 min{p̂1 , p̂2 },
(8)
where p̂1 = n1 /n and p̂2 = n2 /n are estimates of the true prior probabilities. Figure 1b
shows the effect of this bias correction on the accuracy of the HP-estimator. As can
be seen, the bias correction significantly improves the accuracy of the HP-estimator for
when the class distributions are not well separated.
2.3
Multiclass classification
Here we apply the HP-estimate to multiclass classification problems by extending the
bias corrected HP-estimator to a multiclass Bayes error rate. The original multiclass
HP-estimator has been defined by Wisler et al. (2016) and we show how the framework
can be applied to hierarchical multiclassifier design.
Consider a multiclass problem with K classes with xi ∈ X ⊆ Rd and yi ∈ {1, . . . , K},
with
P prior probabilities pk and density functions fk (x) for k = 1, . . . , K such that
k pk = 1. Then, the BER can be estimated for each pair of classes using the biascorrected HP-estimator P̂e (fk , fj ) using (7). The binary classification problem with the
largest BER estimate is defined as most difficult.
Recall that the largest BER that can be achieved in a binary classification problem
with unequal class priors is equal to the value of the smallest prior probability. This
makes it difficult to compare empirical estimates of the BER when class sizes are imbalanced. To correct for this, the HP-estimator for pairwise classification BERs can be
normalized for class sizes using
P̂e0 (fk , fl ) =
P̂e (fk , fl )
.
min{p̂k , p̂l }
(9)
This normalization places the HP-estimate in the interval [0, 1] and makes it possible to
more accurately compare the BER estimates of different binary problems.
In practice it can also be of interest to understand how difficult it is to discriminate
each individual class. By reducing the multiclass problem to a One-vs-Rest classification
problem, it is straightforward to define a confusion rate for a given class k. This represents the fraction of instances that are erroneously assigned to class k and the fraction
of instances which are truly from class k that are assigned to a different class. Formally,
define the confusion rate for class k as
|{i : ŷi = k, yi 6= k}| + |{i : ŷi 6= k, yi = k}|
Ck (y, ŷ) =
,
(10)
n
with ŷi the predicted class for instance i. Recall that the Bayes error rate is the error
rate of the Bayes classifier, which assigns an instance x to class k = arg maxl pl fl (x).
Hence, the BER for a single class k equals the error of assigning to a class l 6= k when
the true class is k and the total error of assigning to class k when the true class is c 6= k,
thus
Z
Z
X
Pe,k =
pk fk (x) dx +
pc fc (x) dx
(11)
c6=k
max{pl fl (x)}≥pk fk (x)
l6=k
max{pl fl (x)}<pk fk (x)
l6=k
1
This correction can readily be derived by using the fact that 1 −
under the same conditions as for (3).
5
2R1,2
n
→ uHP (f1 , f2 ), which holds
·10−4
1
2
3
5
7
11
2.5
Variance
2
1.5
1
0.5
0
0
1
2
3
4
5
Distance
Figure 2: Variance of the HP-estimator of the BER for varying number of orthogonal
MSTs as a function of the separation between the classes. Results are based on 500
repetitions of a binary classification problem with n = 1000, d = 2, p1 = 0.33, where
class distributions are bivariate spherical Gaussians with different means.
We make two observations about this One-vs-Rest Bayes error rate (OvR-BER).
First, the OvR-BER for class k is smaller than the sum of the binary BERs for the
problems involving class k (see Appendix A). Second,
the OvR-BER can be estimated
S
using the Henze-Penrose divergence with R(Xk , l6=k Xl ), which yields the estimate P̂e,k .
A computational advantage of using the OvR-BER in multiclass problems is that the
MST only has to be computed only
S once on the set X, since the union of Xk and ∪l6=k Xl
is equal to X. Therefore, R(Xk , l6=k Xl ) can be computed for all k from the single MST
on X by keeping track of the labels of each instance.
2.4
Computational Considerations
The construction of the minimal spanning tree lies at the heart of the HP-estimator of
the BER, so it is important to use a fast algorithm for the MST construction. Since
the HP-estimator is based on the Euclidean MST the dual-tree algorithm by March
et al. (2010) can be applied. This algorithm is based on the construction of Borůvka
(1926) and implements the Euclidean MST in approximately O(n log n) time. For larger
datasets it can be beneficial to partition the space into hypercubes and construct the
MST in each partition.
A simple way to improve the robustness of the HP-estimator is to use multiple
orthogonal MSTs and average the number of cross-connections (Friedman and Rafsky,
1979). Computing orthogonal MSTs is not straightforward in the dual-tree algorithm
of March et al. (2010), but is easy to implement in MST algorithms that use a pairwise
distance matrix such as that of Whitney (1972). Figure 2 shows the empirical variance
of the HP-estimator for different numbers of orthogonal MSTs as a function of the
separation between the classes. As expected, the variance decreases as the number of
orthogonal MSTs increases, although the benefit of including more orthogonal MSTs
also decreases when adding more MSTs. Therefore, 3 orthogonal MSTs are typically
used in practice.
6
0
1
2
3
4
5
6
7
8
9
0
0
0
1
1
2
2
3
3
4
4
5
5
6
6
7
7
8
8
9
9
(a)
1
2
3
4
5
6
7
8
9
(b)
Figure 3: Heat maps illustrating the difficulty of distinguishing between different handwritten digits in the MNIST dataset of LeCun et al. (1998). Brighter squares correspond
to higher values. In (a) the heat map of the BER estimates on the training data is shown.
In (b) the heat map is shown of the 82 misclassifications made by LeNet-5 on the test
data (LeCun et al., 1998). From (a) it can be readily identified that the numbers 3 and
5 are difficult to distinguish, as well as 4 and 9. Easy to distinguish number pairs are
for instance 6 and 7, and 0 and 1. This pattern is reflected in the results on the test
data shown in (b).
3
Meta-Learning of optimal classifier accuracy
Applying the HP-estimator to meta-learning problems creates a number of opportunities
to assess the difficulty of a classification problem before training a classifier. For example, given a multiclass classification problem it may be useful to know which classes
are difficult to distinguish from each other and which classes are easy to distinguish.
Figure 3a shows an illustration of this for handwritten digits in the well-known MNIST
dataset (LeCun et al., 1998). This figure shows a heat map where each square corresponds to an estimate of the BER for a binary problem in the training set. From this
figure it can be seen that the digits 4 and 9 are difficult to distinguish, as well as the
digits 3 and 5. This information can be useful for the design of a classifier, to ensure for
instance that higher weights are placed on misclassifications of the more difficult number
pairs if correct classification of these pairs is of importance to the end-task. In Figure 3b
a similar heat map is shown based on misclassified instances of LeNet-5 (LeCun et al.,
1998) on the test set. This figure shows the symmetric confusion matrix based on the
82 misclassified instances. As can be seen, this figure closely corresponds to the heat
map on the training data, which confirms the predictive accuracy of the HP-estimator
for real data.
Another example of the accuracy of the BER estimates for multiclass classification
problems is given in Figure 4. In this figure, OvR-BER estimates P̂e,k and class accuracy scores Ck (y, ŷ) are shown for the Chess dataset (n = 28056, d = 34, K = 18)
obtained from the UCI repository (Bache and Lichman, 2013). This dataset was split
into a training dataset (70%) and a test dataset (30%) and the OvR-BER estimates
were computed on the training dataset. These estimates are compared with the class
error rates obtained from out-of-sample predictions of the test dataset using GenSVM
(Van den Burg and Groenen, 2016). This figure shows that the OvR-BER estimates are
accurate predictors of classification performance. The classes that are relatively difficult
7
class error rate
0.2
0.15
0.1
0.05
0
1
2
3
4
5
6
7
8 9 10 11 12 13 14 15 16 17 18
class index
Ck (y, ŷ)
P̂e,k
Figure 4: OvR-BER estimates and test set error rates for each class in the Chess
dataset. It can be seen that for most classes the OvR-BER estimate is an accurate
predictor for test set performance. For classes where there is a difference between the
BER and the test set error the BER generally gives a lower bound on the test set
performance. Test set results were obtained with the GenSVM classifier (Van den Burg
and Groenen, 2016).
to classify may benefit from increasing misclassification weights.
The BER estimates can also be applied to feature selection and, in particular, to
the identification of useful feature transformations of the data. A feature selection
strategy based on forward selection was outlined in Berisha et al. (2016). At each
feature selection stage, this algorithm adds the feature which gives the smallest increase
in the BER estimate. Berisha et al. (2016) show that this feature selection strategy
quickly yields a subset of useful features for the classification problem.
Because the BER estimate is a fast and asymptotically consistent estimate of a
bound on classification performance, it is easy to try a number of potential feature
transformations and use the one with the smallest BER estimate in the classifier. This
can be useful both for traditional feature transformations such as PCA (Pearson, 1901)
and Laplacian Eigenmaps (Belkin and Niyogi, 2003), but also for commonly used kernel
transformations in SVMs. For a researcher this can significantly reduce the time needed
to train a classifier on different transformations of the data. In a multiclass setting where
the One-vs-One strategy is used, one can even consider a different feature transformation
for each binary subproblem. When using a unified classification method one can consider
feature transformations which reduce the average BER estimate or the worst-case BER
estimate.
Note that a feature transformation which reduces the dimensionality of the dataset
without increasing the BER estimate can be considered beneficial, as many classification
methods are faster for low-dimensional datasets. For instance, applying PCA with 2
components on the Chess dataset only slightly increases the BER estimates for two
classes, while remaining the same for the other classes. Thus, a classifier will likely
achieve comparable accuracy with this transformed dataset, but will be much faster to
train since the dimensionality can be reduced from d = 34 to d = 2.
4
Hierarchical Multiclass Classification
In this section a novel hierarchical multiclass SVM classifier is introduced which is based
on uncertainty clustering. The BER estimate can be considered a measure of the irre-
8
1
1
cut
7
2
7
2
cut
6
3
5
4
6
V
V0
(a)
3
cut
5
4
V1
(b)
Figure 5: Illustration of the min-cut procedure for a multiclass classification problem
with K = 7 classes. The first cut in (a) splits the vertices in the set V into two sets, V0
and V1 , which are used in a binary classifier. Figure (b) illustrates the next step where
both V0 and V1 are split further by cuts in the complete graph for each subset.
ducible uncertainty of a classification problem, as a high BER indicates an intrinsically
difficult problem. This can be used to construct a tree of binary classification problems
that increase in difficulty along the depth of the tree. By fitting a binary classifier (such
as an SVM) at each internal node of the tree, a classification method is obtained which
proceeds from the easier binary subproblems to the more difficult binary problems.
Similar divide-and-conquer algorithms have been proposed (Frank and Kramer, 2004,
Schwenker and Palm, 2001, Takahashi and Abe, 2002, Tibshirani and Hastie, 2007, Vural
and Dy, 2004, among others). See Lorena et al. (2008) for a review. These approaches
often apply a clustering method to create a grouping of the dataset into two clusters,
repeating this process recursively to form a binary tree of classification problems. In
Lorena and De Carvalho (2010) several empirical distance measures are used as indicators of separation difficulty between classes, which are applied in a bottom-up procedure
to construct a classification tree. Finally, in El-Yaniv and Etzion-Rosenberg (2010) the
Jensen-Shannon divergence is used to bound the BER with inequalities from Lin (1991)
and a classification tree is constructed using a randomized heuristic procedure. Unfortunately, the Jensen-Shannon divergence implementation requires parametric estimation
of distribution functions. Moreover, for the equiprobable case the upper bound on the
BER obtained with the Jensen-Shannon divergence can be shown to be less tight than
that obtained with the HP-divergence (see Appendix B). Because of this, these estimates
of the BER may be less accurate than those obtained with the proposed HP-estimator.
To construct the hierarchical classification tree a complete weighted graph G =
(V, E, w) is created where the vertices correspond to the classes and the weight of the
edges equals the HP-estimate for that binary problem. Formally, let V = {1, . . . , K},
E = {(i, j) : i, j ∈ V, i 6= j} and define the edge weight w(e) for e ∈ E as w(e) =
w(i, j) = P̂e0 (fi , fj ). In the HP-estimator P̂e0 (fi , fj ) the bias correction (7) and the
normalization (9) are used. By recursively applying min-cuts to this graph a tree of
binary classification problems is obtained which increase in difficulty along the depth
of the tree. Min-cuts on this weighted graph can be computed using for instance the
method of Stoer and Wagner (1997). Figure 5 illustrates this process for a multiclass
classification problem with K = 7.
9
V
{1, 2, 3, 4, 5, 6, 7}
V0
{2, 3, 4}
V00
{2}
V1
{1, 5, 6, 7}
V01
{3, 4}
V010
{3}
V10
{1, 7}
V011
{4}
V100
{1}
V11
{5, 6}
V101
{7}
V110
{5}
V111
{6}
Figure 6: Illustration of the tree induced by the graph cutting procedure illustrated in
Figure 5, for a hypothetical problem with K = 7 classes. Each dashed line in the tree
indicates a point where a binary classifier will be trained.
The tree construction can be described formally as follows. Starting with the complete weighted graph with vertices V , apply a min-cut algorithm to obtain the disjoint vertex sets V0 and V1 such that V0 ∪ V1 = V . This pair of vertex sets then
forms a binary classification problem with datasets X0 = {xi ∈ X : yi ∈ V0 } and
X1 = {xi ∈ X : yi ∈ V1 }. Recursively applying this procedure to the sets V0 and
V1 until no further splits are possible yields a tree of binary classification problems, as
illustrated in Figure 6.
In the remainder of this section the results of an extensive simulation study are presented, which aims to evaluate the performance of this hierarchical classifier on multiclass
classification problems. The classifier that will be used in each binary problem in the
tree will be a linear support vector machine, but in practice any binary classifier could
be used in the algorithm. The implementation of the hierarchical classifier based on the
linear binary SVM will be called SmartSVM.2
The experimental setup is comparable to that used in Van den Burg and Groenen
(2016), where a nested cross-validation (CV) approach is used to reduce bias in classifier
performance (Stone, 1974). From each original dataset 5 independent training and test
datasets are generated. Subsequently, each classification method is trained using 10
fold CV on each of the training datasets. Finally, the model is retrained on the entire
training dataset using the optimal hyperparameters and this model is used to predict
the test set. In the experiments 16 datasets are used of varying dimensions from which
80 independent test sets are constructed. The train and test datasets were generated
using a stratified split, such that the proportions of the classes correspond to those in
the full dataset. Table 1 shows the descriptive statistics of each of the datasets used.
Datasets are collected from the UCI repository (Bache and Lichman, 2013) and the
KEEL repository (Alcalá et al., 2010).
SmartSVM will be compared to five other linear multiclass SVMs in these experiments. Three of these alternatives are heuristic methods which use the binary SVM
as underlying classifier, while two others are single-machine multiclass SVMs. One of
the most commonly used heuristic approaches to multiclass SVMs is the One vs. One
(OvO) method (Kreßel, 1999) which solves a binary SVM for each of the K(K − 1) pairs
of classes. An alternative is the One vs. Rest (OvR) method (Vapnik, 1998) in which
2
The SmartSVM classifier and the meta-learning and BER estimation techniques presented in the
previous sections have been implemented in the smartsvm Python package, available at: https://
github.com/HeroResearchGroup/SmartSVM.
10
Table 1: Dataset summary statistics for the datasets used in the experimental study.
The final two columns denote the size of the smallest and the largest class, respectively.
Datasets marked with an asterisk are collected from the KEEL dataset repository, all
others are from the UCI repository.
Dataset
abalone
fars*
flare
kr-vs-k
letter
nursery
optdigits
pageblocks
pendigits
satimage
segment
shuttle
texture*
wine red
wine white
yeast
Instances
4177
100959
1066
28056
20000
12960
5620
5473
10992
6435
2310
58000
5500
1599
4898
1479
Features
8
338
19
34
16
19
64
10
16
36
19
9
40
12
12
8
Classes
20
7
6
18
26
4
10
5
10
6
7
6
11
5
6
9
min nk
14
299
43
27
734
330
554
28
1055
626
330
23
500
18
20
20
max nk
689
42116
331
4553
813
4320
572
4913
1144
1533
330
45586
500
681
2198
463
a binary SVM is solved for each of the K − 1 binary problems obtained by separating
a single class from the others. The directed acyclic graph (DAG) SVM was proposed
by Platt et al. (2000) as an extension of the OvO approach. It has a similar training
procedure as OvO, but uses a different prediction strategy. In the OvO method a voting
scheme is used where the class with the most votes from each binary classifier becomes
the predicted label. In contrast, the DAGSVM method uses a voting scheme where
the least likely class is voted away until only one remains. Finally, two single-machine
multiclass SVMs are also compared: the method by Crammer and Singer (2002) and
GenSVM (Van den Burg and Groenen, 2016).
All methods are implemented in either C or C++, to ensure that speed of the methods
can be accurately compared. The methods that use a binary SVM internally are implemented with LibLinear (Fan et al., 2008). LibLinear also implements a fast solver for the
method by Crammer and Singer (2002) using the algorithm proposed by Keerthi et al.
(2008). For SmartSVM the Bayes error rates and the corresponding classification tree
were calculated once for each training dataset as a preprocessing step. For most datasets
the BERs were computed based on 3 orthogonal MSTs using the algorithm of Whitney
(1972). For the two largest datasets (fars and shuttle) the BER was computed based on
a single MST using the algorithm of March et al. (2010). Computing these MSTs was
done in parallel using at most 10 cores. In the results on training time presented below
the training time of SmartSVM is augmented with the preprocessing time.
The binary SVM has a cost parameter for the regularization term, which is optimized
using cross validation. The range considered for this parameter is C ∈ {2−18 , 2−16 , . . . , 218 }.
The GenSVM method has additional hyperparameters which were varied in the same
way as in the experiments of Van den Burg and Groenen (2016). All experiments were
11
Table 2: Total training time per dataset in seconds, averaged over the 5 nested CV folds.
Minimal values per dataset are underlined. As can be seen, SmartSVM is the fastest
method on 10 out of 16 datasets.
Dataset
abalone
fars
flare
krkopt
letter
nursery
optdigits
pageblocks
pendigits
satimage
segment
shuttle
texture
winered
winewhite
yeast
C&S
13332
158242
467.5
86000
31684
2267
88.0
523.4
2499
5184
377.1
7658
588.1
706.4
2570
1209
DAG
146.5
3108
7.0
728.4
407.3
74.7
20.0
26.8
30.2
74.7
7.1
570.8
34.5
12.3
59.2
16.7
GenSVM
86787
205540
501.7
56554
44960
1363
31615
2001
6591
2563
1542
14618
14774
542.7
2570
1138
OvO
134.3
2866
6.5
680.4
381.5
68.7
18.5
25.2
28.5
70.0
7.0
561.9
32.8
11.1
55.1
15.3
OvR
214.4
5630
8.2
1388
1805
94.9
24.3
48.6
151.3
188.4
18.7
1996
69.6
18.3
70.9
24.2
SmartSVM
75.0
3158
5.0
664.0
652.1
56.8
10.9
27.6
53.7
74.5
5.4
832.0
19.3
8.8
36.7
13.2
performed on the Dutch National LISA Compute Cluster using the abed utility.3
The experiments are compared on training time and out-of-sample predictive performance. Table 2 shows the results for training time, averaged over the 5 nested cross
validation folds for each dataset. As can be seen SmartSVM is the fastest method on
10 out of 16 datasets. This can be attributed to the smaller number of binary problems
that SmartSVM needs to solve compared to OvO and the fact that the binary problems
are smaller than those solved by OvR. The OvO method is the fastest classification
method on the remaining 6 datasets. The single-machine multiclass SVMs by Crammer
and Singer (2002) and Van den Burg and Groenen (2016) both have larger computation
times than the heuristic methods. Since GenSVM has a larger number of hyperparameters, it is interesting to look at the average time per hyperparameter configuration as
well. In this case, GenSVM is on average faster than Crammer and Singer (2002) due
to the use of warm starts (see Appendix C for additional simulation results).
Classification performance of the methods is reported using the adjusted Rand index
(ARI) which corrects for chance (Hubert and Arabie, 1985). Use of this index as a classification metric has been proposed previously by Santos and Embrechts (2009). Table 3
shows the predictive performance as measured with the ARI. As can be seen, SmartSVM
obtains the maximum performance on two of the sixteen datasets. However, SmartSVM
outperforms One vs. One on 3 datasets and outperforms One vs. Rest on 10 out of 16
datasets. The OvO and OvR methods are often used as default heuristic approaches
for multiclass SVMs and are respectively the default strategies in the popular LibSVM
(Chang and Lin, 2011) and LibLinear (Fan et al., 2008) libraries. Since SmartSVM is
often faster than these methods, our results indicate a clear practical benefit to using
SmartSVM for multiclass classification.
3
See https://github.com/GjjvdBurg/abed.
12
Table 3: Out-of-sample predictive performance as measured by the adjusted Rand index.
Although SmartSVM doesn’t often achieve the maximum performance, there are several
datasets for which SmartSVM outperforms One vs. One, or differs from the maximum
performance in only the second or third decimal.
Dataset
abalone
fars
flare
krkopt
letter
nursery
optdigits
pageblocks
pendigits
satimage
segment
shuttle
texture
winered
winewhite
yeast
5
C&S
0.0788
0.8127
0.4274
0.1300
0.6173
0.8279
0.9721
0.7681
0.9068
0.7420
0.9013
0.9275
0.9936
1.0000
1.0000
0.2595
DAG
0.0898
0.8143
0.6480
0.1988
0.6991
0.8175
0.9854
0.7335
0.9566
0.7652
0.9000
0.8543
0.9950
1.0000
1.0000
0.2540
GenSVM
0.0895
0.8085
0.6687
0.1779
0.5909
0.8303
0.9732
0.6847
0.8971
0.7403
0.8812
0.8887
0.9888
0.9985
0.9998
0.2521
OvO
0.0898
0.8146
0.6496
0.2022
0.7118
0.8157
0.9850
0.7353
0.9597
0.7672
0.8982
0.8543
0.9947
1.0000
1.0000
0.2587
OvR
0.0791
0.8134
0.4084
0.1512
0.5155
0.8095
0.9686
0.6696
0.8607
0.7107
0.8354
0.6925
0.9865
0.8369
0.8131
0.2433
SmartSVM
0.0595
0.8115
0.6544
0.1585
0.4533
0.8138
0.9640
0.7214
0.8817
0.7607
0.8986
0.8038
0.8977
1.0000
1.0000
0.2088
Discussion
In this work the practical applicability of nonparametric Bayes error estimates to metalearning and hierarchical classifier design has been investigated. For the BER estimate
introduced by Berisha et al. (2016) a bias correction was derived which improves the
accuracy of the estimator for classification problems with unequal class priors. Furthermore, a normalization term was proposed which makes the BER estimates comparable
in multiclass problems. An expression of the OvR-BER was given which represents the
exact Bayes error for a single class in the multiclass problem and it was shown that this
error can be efficiently estimated using the HP-estimator as well. A robustness analysis
of the HP-estimator was performed which showed the benefit of using orthogonal MSTs
in the estimator.
There are many potential applications of the BER estimates to meta-learning problems. Above, several possibilities were explored including the prediction of which pairs of
classes are most difficult to distinguish and which individual classes will yield the highest
error rate. Preliminary experiments with feature transformations were also performed,
which showed that the BER estimates can be a useful tool in determining beneficial
transformations before a classifier is trained.
Based on the weighted graph of pairwise BER estimates, a hierarchical multiclass
classification method was proposed. The classifier uses a top-down splitting approach
to create a tree of binary classification problems which increase in difficulty along the
depth of the tree. By using a linear SVM for each classification problem, a hierarchical
multiclass SVM was obtained which was named SmartSVM. Extensive simulation studies
showed that SmartSVM is often faster than existing approaches and yields competitive
predictive performance on several datasets.
Note that the SmartSVM classifier is only one example of how the BER estimates
13
can be used to construct better classification methods. As discussed in Section 3, BER
estimates could also be used to define class weights in a multiclass classifier. Moreover,
the min-cut strategy used for SmartSVM may not be the optimal way to construct the
classification tree. Evaluating different approaches to constructing classification hierarchies and other applications of the BER estimates to multiclass classification problems
are topics for further research.
Acknowledgements
The computational experiments of this work were performed on the Dutch National LISA
Compute Cluster, and supported by the Dutch National Science Foundation (NWO).
The authors thank SURFsara (www.surfsara.nl) for the support in using the LISA
cluster. This research was partially supported by the US Army Research Office, grant
W911NF-15-1-0479 and US Dept. of Energy grant DE-NA0002534.
A
One vs. Rest Bayes Error Rate
In this section bounds for the One vs. Rest Bayes error rate will be derived, which
measures the error of the Bayes classifier in correctly identifying an individual class.
Definition 1 (OvR-BER). Let f1 , f2 , . . . , fK and p1 , p2 , . . . , pK denoteP
density functions
and prior probabilities for the classes 1 through K respectively, with c pc = 1. Then,
the Bayes error rate between a class k and the remaining classes is given by
Z
Z
X
Pe,k =
pk fk (x) dx +
pc fc (x) dx.
(12)
c6=k
max{pl fl (x)}≥pk fk (x)
l6=k
max{pl fl (x)}<pk fk (x)
l6=k
Below it will be shown that the OvR-BER can
S be bounded using the FriedmanRafsky statistic in the One-vs-Rest setting, R(Xk , l6=k Xl ). Let the mixture distribution of the classes l 6= k be given by
P
l6=k pl fl (x)
gk (x) = P
,
(13)
l6=k pl
P
S
with prior probability pg = l6=k pl . Then Xgk = l6=k Xl can be seen as a draw from
this mixture distribution. By Theorem 2 of Berisha et al. (2016) it holds that
Z
p
1
1
u(fk , gk ) ≤ min{pk fk (x), pg gk (x)} dx ≤ 12 − 12 u(fk , gk ).
(14)
2 − 2
The following theorem relates this error to the OvR-BER defined above.
Theorem 2. The error rate between class k and the mixture distribution without class
k is bounded from above by the OvR-BER,
Z
Qe,k = min{pk fk (x), pg gk (x)} dx ≤ Pe,k .
(15)
Proof. Note that
Z
Qe,k =
Z
pk fk (x) dx
pk fk (x)≤pg gk (x)
+
pg gk (x) dx.
pk fk (x)>pg gk (x)
14
(16)
To simplify the notation, introduce the sets
n
o
T = x ∈ Rd : pk fk (x) ≤ pg gk (x)
d
S = x ∈ R : pk fk (x) ≤ max{pl fl (x)}
(17)
and denote their respective complements by T 0 and S 0 . Then,
Z
Z
pg gk (x) dx
pk fk (x) dx +
Qe,k =
0
T
T
Z
Z
pg gk (x) dx.
pk fk (x) dx +
Pe,k =
(19)
l6=k
P
l6=k
pl fl (x) ≤ maxl6=k pl fl (x) it holds that S ⊆ T and T 0 ⊆ S 0 . Hence,
Z
Z
pk fk (x) dx =
Z
pk fk (x) dx +
T
pk fk (x) dx
Z
Z
pg gk (x) dx =
T0
(21)
T \S
S
Z
S0
(20)
S0
S
Since pg gk (x) =
(18)
pg gk (x) dx +
S 0 \T 0
pg gk (x) dx
However, the sets T \ S and S 0 \ T 0 both equal
d
U = x ∈ R : max{pl fl (x)} < pk fk (x) ≤ pg gk (x) ,
l6=k
(22)
(23)
so it follows that
Z
Qe,k = Pe,k +
U
pk fk (x) dx −
Z
U
pg gk (x) dx ≤ Pe,k
(24)
by definition of the set U .
This provides a lower bound for Pe,k in terms of u(fk , gk ). What remains to be
shown is that Pe,k has an upper bound in terms of u(fk , gk ). No such bound has yet
been found. However, the following result can be presented which does bound Pe,k from
above.
Theorem 3. For a single class k the OvR-BER is smaller than or equal to the sum of
the pairwise BER estimates involving class k.
Proof. Recall that the OvR-BER for class k is given by
Z
X
Pe,k =
pk fk (x) dx +
Z
pc fc (x) dx,
c6=k
pk fk (x)>max{pl fl (x)}
pk fk (x)≤max{pl fl (x)}
l6=k
(25)
l6=k
and denote the sum of the pairwise BERs involving k as Fe,k given by,
XZ
Fe,k =
min{pk fk (x), pc fc (x)} dx
(26)
c6=k
=
X
Z
pk fk (x) dx
c6=k
pk fk (x)<pc fc (x)
15
+
X
Z
pc fc (x) dx.
c6=k
pk fk (x)>pc fc (x)
(27)
Then comparing the first term of Fe,k with that of Pe,k shows
Z
X Z
pk fk (x) dx,
pk fk (x) dx ≥
c6=k
pk fk (x)<pc fc (x)
(28)
pk fk (x)≤max{pl fl (x)}
l6=k
since the area of integration on the left is larger than on the right. Similarly,
X Z
X Z
pc fc (x) dx
pc fc (x) dx ≥
(29)
c6=k
pk fk (x)>max{pl fl (x)}
c6=k
pk fk (x)>pc fc (x)
l6=k
for the same reason. This completes the proof.
B
Jensen-Shannon Bound Inequality
In this section a proof is given for the statement that the Henze-Penrose upper bound
on the Bayes error rate is tighter than the Jensen-Shannon upper bound derived by Lin
(1991). Before presenting the proof, the following lemma is presented.
Lemma 4. For x, y > 0 it holds that
y
x
4 log(2)xy
x log 1 +
+ y log 1 +
≥
.
x
y
x+y
Proof. Let t =
y
x
(30)
and multiply both sides by 1t + 1 > 0, then the inequality reduces to
1
1
+ 1 log (1 + t) + (1 + t) log 1 +
≥ 4 log(2).
(31)
t
t
Denote the left hand side by f (t). The proof will now proceed by showing that f (t) ≥
f (1) = 4 log(2) for all t > 0. The derivatives of f (t) are given by
1
log(1 + t)
2 log(1 + t) − t
f 0 (t) = log 1 +
−
and
f 00 (t) =
.
(32)
t
t2
t3
Write the numerator of f 00 (t) as g(t) such that
g(t) = 2 log(1 + t) − t,
and
g 0 (t) =
1−t
.
1+t
(33)
Then it is clear that g 0 (t) > 0 for 0 < t < 1 and g 0 (t) < 0 for t > 1. Furthermore
limt→0+ g(t) = 0 and limt→∞ g(t) = −∞. Thus, it follows that g(t) increases on 0 <
t < 1 and decreases for t > 1. Let t = α > 1 be such that g(α) = 0, then g(t) > 0 for
0 < t < α and g(t) < 0 for t > α.
From this it follows that f 00 (t) > 0 for 0 < t < α and f 00 (t) < 0 for t > α. Hence,
f 0 (t) is increasing on 0 < t < α and decreasing for t > α. Moreover, limt→0+ f 0 (t) = −∞
and f 0 (1) = 0. Thus, it follows that f 0 (t) is negative on 0 < t < 1, positive for t > 1,
and attains a maximum at t = α after which it decreases to limt→∞ f 0 (t) = 0. Since
limt→0+ f (t) = ∞ it follows that f (t) is decreasing on 0 < t < 1 and increasing for
t > 1.
16
Definition 5 (Kullback-Leibler Divergence). For probability density functions f1 (x)
and f2 (x) the Kullback-Leibler divergence is given by
Z
f1 (x)
DKL (f1 kf2 ) = f1 (x) log2
dx,
(34)
f2 (x)
Kullback and Leibler (1951).
Definition 6 (Jensen-Shannon Divergence). According to El-Yaniv et al. (1998) the
Jensen-Shannon divergence for two probability density functions f1 (x) and f2 (x) with
prior probabilities p1 and p2 , can be stated in terms of the Kullback-Leibler divergence
as
JS(f1 , f2 ) = p1 DKL (f1 kM ) + p2 DKL (f2 kM )
(35)
with M (x) = p1 f1 (x) + p2 f2 (x) the mixture distribution and p2 = 1 − p1 .
Theorem 7. For p1 = p2 = 21 the Henze-Penrose upper bound on the BER is tighter
than the Jensen-Shannon upper bound of Lin (1991),
Pe (f1 , f2 ) ≤
1
2
− 12 uHP ≤ 12 J,
(36)
where J = H(p1 ) − JS(f1 , f2 ) with H(p1 ) the binary entropy and JS(f1 , f2 ) the JensenShannon divergence.
Proof. First, note that with p1 = p2 = 21 the binary entropy H(p1 ) = H( 12 ) = 1. Second,
for the equiprobable case it holds that
Z
f1 (x)f2 (x)
1
1
dx.
(37)
2 − 2 uHP (f1 , f2 ) =
f1 (x) + f2 (x)
The Jensen-Shannon upper bound can be written as
Z
Z
2f1 (x)
2f2 (x)
1
1
1
1
f1 (x) log2
dx − 4 f2 (x) log2
dx
2J = 2 − 4
f1 (x) + f2 (x)
f1 (x) + f2 (x)
Z
Z
2f1 (x)
dx
= 41 f1 (x) + f2 (x) dx − 14 f1 (x) log2
f1 (x) + f2 (x)
Z
2f2 (x)
1
− 4 f2 (x) log2
dx
f1 (x) + f2 (x)
Z
2f1 (x)
1
= 4 f1 (x) 1 − log2
dx
f1 (x) + f2 (x)
Z
2f2 (x)
1
+ 4 f2 (x) 1 − log2
dx
f1 (x) + f2 (x)
Z
f2 (x)
f1 (x)
1
= 4 f1 (x) log2 1 +
+ f2 (x) log2 1 +
dx
f1 (x
f2 (x)
By Lemma 4 it follows that
f2 (x)
f1 (x)
4f1 (x)f2 (x)
f1 (x) log2 1 +
+ f2 (x) log2 1 +
≥
,
f1 (x
f2 (x)
f1 (x) + f2 (x)
and therefore
1
2J
≥
1
4
Z
4f1 (x)f2 (x)
dx =
f1 (x) + f2 (x)
17
1
2
− 12 uHP (f1 , f2 ).
(38)
(39)
(40)
(41)
(42)
(43)
(44)
Table 4: Average training time per hyperparameter configuration in seconds, averaged
over the 5 nested CV folds. Minimal values per dataset are underlined.
Dataset
C&S
DAG GenSVM
OvO
OvR SmartSVM
abalone
702
7.712
254
7.067 11.282
3.960
fars
8329
164
601
151
296
184
flare
24.606
0.369
1.467
0.344
0.431
0.263
krkopt
4526 38.337
165 35.813 73.053
35.627
letter
1668 21.438
131 20.076 94.977
34.454
nursery
119
3.932
3.985
3.617
4.994
3.133
optdigits
4.630
1.053
92.441
0.971
1.278
0.590
pageblocks 27.550
1.408
5.852
1.326
2.559
1.504
pendigits
132
1.589
19.273
1.498
7.965
2.872
satimage
273
3.931
7.494
3.685
9.918
3.942
segment
19.846
0.376
4.509
0.370
0.985
0.285
shuttle
403 30.042
42.744 29.571
105
43.928
texture
30.954
1.816
43.200
1.727
3.665
1.028
winered
37.179
0.646
1.587
0.586
0.966
0.465
winewhite
135
3.115
7.514
2.903
3.732
1.954
yeast
63.655
0.877
3.327
0.807
1.272
0.699
C
Additional Simulation Results
In this section some additional simulation results are presented for the SmartSVM experiments presented in Section 4. Table 4 shows the average time per hyperparameter
configuration for each of the methods. This is especially useful for comparing GenSVM
(Van den Burg and Groenen, 2016) with the other methods, as it has a larger set of
hyperparameters to consider.
A commonly used tool to summarize results of simulation experiments is to use rank
plots (Demšar, 2006). For each dataset the methods are ranked, with the best method
receiving rank 1 and the worst method receiving rank 6 (since there are 6 methods in
this experiment). In case of ties fractional ranks are used. By averaging the ranks over
all datasets, a visual summary of the results can be obtained. Figures 7a, 7b and 7c
show these average ranks for predictive performance, total training time, and average
training time respectively.
The ordering of OvO and SmartSVM in the rank plots for training time may seem
counterintuitive, considering that SmartSVM is more often the fastest method. This
can be explained by the fact that in the cases where SmartSVM is slower than OvO it
is usually also slower than DAG. In contrast, where SmartSVM is the fastest method
OvO is usually the second fastest method. Because of this, SmartSVM obtains a slightly
higher average rank than OvO.
18
OvO
1.0
1.5
DAG
2.0
C&S
2.5
GenSVM
3.0
3.5
SmartSVM
4.0
4.5
OvR
5.0
5.5
6.0
(a) Predictive performance
OvO
1.0
SmartSVM
1.5
2.0
DAG
2.5
OvR
3.0
3.5
C&S
4.0
4.5
5.0
GenSVM
5.5
6.0
(b) Total training time
OvO
1.0
1.5
SmartSVM
2.0
DAG
2.5
OvR
3.0
3.5
4.0
GenSVM
4.5
5.0
C&S
5.5
6.0
(c) Average training time
Figure 7: Rank plots of classifier performance in the simulation study. Figure (a) shows
the average ranks for out-of-sample predictive performance as measured by the ARI.
Figures (b) and (c) respectively show the average ranks for total training time and
average training time per hyperparameter configuration.
References
Alcalá, J., Fernández, A., Luengo, J., Derrac, J., Garcı́a, S., Sánchez, L., and Herrera, F.
Keel data-mining software tool: data set repository, integration of algorithms and experimental analysis framework. Journal of Multiple-Valued Logic and Soft Computing,
17(2-3):255–287, 2010.
Avi-Itzhak, H. and Diep, T. Arbitrarily tight upper and lower bounds on the bayesian
probability of error. IEEE Transactions on Pattern Analysis and Machine Intelligence,
18(1):89–91, 1996.
Bache, K. and Lichman, M. UCI machine learning repository. 2013.
Belkin, M. and Niyogi, P. Laplacian eigenmaps for dimensionality reduction and data
representation. Neural computation, 15(6):1373–1396, 2003.
Berisha, V. and Hero, A. Empirical non-parametric estimation of the Fisher information.
Signal Processing Letters, IEEE, 22(7):988–992, 2015.
Berisha, V., Wisler, A., Hero, A., and Spanias, A. Empirically estimable classification
bounds based on a nonparametric divergence measure. IEEE Transactions on Signal
Processing, 64(3):580–591, 2016.
Bhattacharyya, A. On a measure of divergence between two multinomial populations.
Sankhyā: The Indian Journal of Statistics, 7(4):401–406, 1946.
Borůvka, O. O jistém problému minimálnı́m. Práce Moravské Pridovedecké Spolecnosti,
3:37–58, 1926.
Chang, C. and Lin, C. LIBSVM: A library for support vector machines. ACM Transactions on Intelligent Systems and Technology, 2(3):27:1–27:27, 2011.
19
Chernoff, H. A measure of asymptotic efficiency for tests of a hypothesis based on the
sum of observations. The Annals of Mathematical Statistics, 23(4):493–507, 1952.
Cortes, C. and Vapnik, V. Support-vector networks. Machine Learning, 20(3):273–297,
1995.
Cox, D. The regression analysis of binary sequences. Journal of the Royal Statistical
Society. Series B (Methodological), 20(2):215–242, 1958.
Crammer, K. and Singer, Y. On the algorithmic implementation of multiclass kernelbased vector machines. The Journal of Machine Learning Research, 2(Dec):265–292,
2002.
Csiszár, I. I-divergence geometry of probability distributions and minimization problems.
The Annals of Probability, 3(1):146–158, 1975.
Demšar, J. Statistical comparisons of classifiers over multiple data sets. The Journal of
Machine Learning Research, 7(Jan):1–30, 2006.
El-Yaniv, R. and Etzion-Rosenberg, N. Hierarchical multiclass decompositions with
application to authorship determination. arXiv preprint arXiv:1010.2102, 2010.
El-Yaniv, R., Fine, S., and Tishby, N. Agnostic classification of markovian sequences.
In: Advances of Neural Information Processing Systems 10, pp. 465–471. MIT Press,
1998.
Fan, R.E., Chang, K.W., Hsieh, C.J., Wang, X.R., and Lin, C.J. LIBLINEAR: A
library for large linear classification. The Journal of Machine Learning Research,
9(Aug):1871–1874, 2008.
Frank, E. and Kramer, S. Ensembles of nested dichotomies for multi-class problems.
In: Proceedings of the 21st International Conference on Machine Learning, pp. 39–46.
ACM, 2004.
Friedman, J. and Rafsky, L. Multivariate generalizations of the Wald-Wolfowitz and
Smirnov two-sample tests. The Annals of Statistics, 7(4):697–717, 1979.
Fukunaga, K. Introduction to Statistical Pattern Recognition. Academic Press, 1990.
Hashlamoun, W., Varshney, P., and Samarasooriya, V. A tight upper bound on the
bayesian probability of error. IEEE Transactions on Pattern Analysis and Machine
Intelligence, 16(2):220–224, 1994.
Henze, N. and Penrose, M. On the multivariate runs test. Annals of statistics, 27(1):290–
298, 1999.
Ho, T. and Basu, M. Complexity measures of supervised classification problems. IEEE
Transactions on Pattern Analysis and Machine Intelligence, 24(3):289–300, 2002.
Hubert, L. and Arabie, P. Comparing partitions. Journal of Classification, 2(1):193–218,
1985.
Kailath, T. The divergence and Bhattacharyya distance measures in signal selection.
IEEE Transactions on Communication Technology, 15(1):52–60, 1967.
20
Keerthi, S., Sundararajan, S., Chang, K.W., Hsieh, C.J., and Lin, C.J. A sequential
dual method for large scale multi-class linear SVMs. In: Proceedings of the 14th ACM
SIGKDD International Conference on Knowledge Discovery and Data Mining, pp.
408–416. 2008.
Kreßel, U. Pairwise classification and support vector machines. In: B. Schölkopf, C.J.C.
Burges, and A.J. Smola, editors, Advances in Kernel Methods, pp. 255–268. MIT
Press, 1999.
Kullback, S. and Leibler, R. On information and sufficiency. The Annals of Mathematical
Statistics, 22(1):79–86, 1951.
LeCun, Y., Bottou, L., Bengio, Y., and Haffner, P. Gradient-based learning applied to
document recognition. Proceedings of the IEEE, 86(11):2278–2324, 1998.
Lin, J. Divergence measures based on the shannon entropy. IEEE Transactions on
Information Theory, 37(1):145–151, 1991.
Lorena, A. and De Carvalho, A. Building binary-tree-based multiclass classifiers using
separability measures. Neurocomputing, 73(16):2837–2845, 2010.
Lorena, A., De Carvalho, A., and Gama, J. A review on the combination of binary
classifiers in multiclass problems. Artificial Intelligence Review, 30(1):19–37, 2008.
Lu, Y., Guo, H., and Feldkamp, L. Robust neural learning from unbalanced data samples. In: Proceedings of the IEEE International Joint Conference on Neural Networks,
volume 3, pp. 1816–1821. IEEE, 1998.
March, W., Ram, P., and Gray, A. Fast Euclidean minimum spanning tree: algorithm,
analysis, and applications. In: Proceedings of the 16th ACM SIGKDD International
Conference on Knowledge Discovery and Data Mining, pp. 603–612. ACM, 2010.
McCulloch, W. and Pitts, W. A logical calculus of the ideas immanent in nervous
activity. The Bulletin of Mathematical Biophysics, 5(4):115–133, 1943.
Pearson, K. On lines and planes of closest fit to systems of points in space. The London,
Edinburgh, and Dublin Philosophical Magazine and Journal of Science, 2(11):559–572,
1901.
Platt, J., Cristianini, N., and Shawe-Taylor, J. Large margin DAGs for multiclass
classification. In: S.A. Solla, T.K. Leen, and K. Müller, editors, Advances in Neural
Information Processing Systems 12, pp. 547–553. MIT Press, 2000.
Qiao, X. and Liu, Y. Adaptive weighted learning for unbalanced multicategory classification. Biometrics, 65(1):159–168, 2009.
Santos, J. and Embrechts, M. On the use of the adjusted rand index as a metric
for evaluating supervised classification. In: Proceedings of the 19th International
Conference on Artificial Neural Networks: Part II, pp. 175–184. Springer-Verlag, 2009.
Schwenker, F. and Palm, G. Tree-structured support vector machines for multi-class
pattern recognition. In: International Workshop on Multiple Classifier Systems, pp.
409–417. Springer, 2001.
Stoer, M. and Wagner, F. A simple min-cut algorithm. Journal of the ACM, 44(4):585–
591, 1997.
21
Stone, M. Cross-validatory choice and assessment of statistical predictions. Journal of
the Royal Statistical Society. Series B (Methodological), 36(2):111–147, 1974.
Takahashi, F. and Abe, S. Decision-tree-based multiclass support vector machines. In:
Neural Information Processing, 2002. ICONIP’02. Proceedings of the 9th International Conference on, volume 3, pp. 1418–1422. IEEE, 2002.
Tibshirani, R. and Hastie, T. Margin trees for high-dimensional classification. Journal
of Machine Learning Research, 8(Mar):637–652, 2007.
Van den Burg, G. and Groenen, P. GenSVM: A generalized multiclass support vector
machine. Journal of Machine Learning Research, 17(225):1–42, 2016.
Vapnik, V. Statistical learning theory. Wiley, New York, 1998.
Vural, V. and Dy, J. A hierarchical method for multi-class support vector machines. In:
Proceedings of the 21st International Conference on Machine Learning, pp. 105–112.
ACM, 2004.
Wald, A. Foundations of a general theory of sequential decision functions. Econometrica,
15(4):279–313, 1947.
Whitney, V. Algorithm 422: minimal spanning tree. Communications of the ACM,
15(4):273–274, 1972.
Wisler, A., Berisha, V., Wei, D., Ramamurthy, K., and Spanias, A. Empiricallyestimable multi-class classification bounds. In: 2016 IEEE International Conference
on Acoustics, Speech and Signal Processing (ICASSP), pp. 2594–2598. IEEE, 2016.
22
| 2cs.AI
|
Planar Disjoint-Paths Completionab
arXiv:1511.04952v2 [cs.DS] 17 Nov 2015
Isolde Adlerc
Stavros G. Kolliopoulosd
Dimitrios M. Thilikosef
Abstract
We introduce Planar Disjoint Paths Completion, a completion counterpart of
the Disjoint Paths problem, and study its parameterized complexity. The problem
can be stated as follows: given a, not necessarily connected, plane graph G, k pairs
of terminals, and a face F of G, find a minimum-size set of edges, if one exists,
to be added inside F so that the embedding remains planar and the pairs become
connected by k disjoint paths in the augmented network. Our results are twofold:
first, we give an upper bound on the number of necessary additional edges when a
solution exists. This bound is a function of k, independent of the size of G. Second,
we show that the problem is fixed-parameter tractable, in particular, it can be solved
in time f (k) · n2 .
Keywords: Completion Problems, Disjoint Paths, Planar Graphs.
1
Introduction
Suppose we are given a planar road network with n cities and a set of k pairs of them.
An empty area of the network is specified and we wish to add a minimum-size set of
a
The second and the third author where co-financed by the European Union (European Social Fund
– ESF) and Greek national funds through the Operational Program “Education and Lifelong Learning”
of the National Strategic Reference Framework (NSRF) - Research Funding Program: “Thalis. Investing
in knowledge society through the European Social Fund”.
b
Emails: Isolde Adler: [email protected], Stavros Kolliopoulos: [email protected], Dimitrios M. Thilikos: [email protected]
c
Institut für Informatik, Goethe-Universität, Frankfurt, Germany.
d
Department of Informatics and Telecommunications, National and Kapodistrian University of
Athens, Athens, Greece.
e
Department of Mathematics, National and Kapodistrian University of Athens, Athens, Greece.
f
AlGCo project-team, CNRS, LIRMM, France.
1
intercity roads in that area so that the augmented network remains planar and the pairs
are connected by k internally disjoint roads. In graph-theoretic terms, we are looking for
a minimum-size edge-completion of a plane graph so that an infeasible instance of the
Disjoint Paths problem becomes feasible while maintaining planarity. In this paper
we give an algorithm that solves this problem in f (k) · n2 steps. Our algorithm uses
a combinatorial lemma stating that, whenever such a solution exists, its size depends
exclusively on k.
The renowned Disjoint Paths Problem (DP) is defined as follows.
DP(G, s1 , t1 , . . . , sk , tk )
Input: An undirected graph G and k pairs of terminals
s1 , t1 , . . . , sk , tk ∈ V (G).
Question: are there k pairwise internally vertex-disjoint paths
Q1 , . . . Qk in G such that path Qi connects si to ti ?
(By pairwise internally vertex-disjoint we mean that two paths can
only intersect at a vertex which is a terminal for both.)
DP is NP-complete even on planar graphs [14] but, when parameterized by k, the
problem belongs to the parameterized complexity class FPT, i.e., it can be solved in time
f (k) · nO(1) , for some function f. More precisely, it can be solved in f (k) · n3 time by the
celebrated algorithm of Robertson and Seymour [19] from the Graph Minors project.
For planar graphs, the same problem can be solved in f (k) · n [16].
We introduce a completion counterpart of this problem, Planar Disjoint Paths
Completion (PDPC), which is of interest on infeasible instances of DP, and we
study its parameterized complexity, when parameterized by k. We are given an embedding of a, possibly disconnected, planar graph G in the sphere, k pairs of terminals
s1 , t1 , . . . , sk , tk ∈ V (G), a positive integer `, and an open connected subset F of the
surface of the sphere, such that F and G do not intersect (we stress that the boundary
of F is not necessarily a cycle). We want to determine whether there is a set of at most
` edges to add, the so-called patch, so that
(i) the new edges lie inside F and are incident only to vertices of G on the boundary of
F,
(ii) the new edges do not cross with each other or with G, and
(iii) in the resulting graph, which consists of G plus the patch, DP has a solution.
PDPC is NP-complete even when ` is not a part of the input and G is planar by
the following simple reduction from DP: add a triangle T to G and let F be the interior
2
of T. That way, we force the set of additional edges to be empty and obtain DP as a
special case.
Notice that our problem is polynomially equivalent to the minimization problem
where we ask for a minimum-size patch: simply solve the problem for all possible values
of `. Requiring the size of the patch to be at most ` is the primary source of difficulty.
In case there is no restriction on the size of the patch and we simply ask whether one
exists, the problem is in FPT by a reduction to DP, which is summarized as follows.
For simplicity, let F be an open disk. Let G0 be the graph obtained by “sewing” along
the boundary of F an O(n) × O(n)-grid. By standard arguments, PDPC has a solution
on G if and only if DP has a solution on G0 . A similar, but more involved, construction
applies when F is not an open disk.
Parameterizing completion problems. Completion problems are natural to define:
take any graph property, represented by a collection of graphs P, and ask whether it is
possible to add edges to a graph so that the new graph is in P. Such problems have been
studied for a long time and some of the most prominent are the following: Hamiltonian
Completion [8, GT34], Path Graph Completion [8, GT36] Proper Interval
Graph Completion [9] Minimum Fill-In [20] Interval Graph Completion [8,
GT35].
Kaplan et al. in their seminal paper [12] initiated the study of the parameterized
complexity of completion problems and showed that Minimum Fill-In, Proper Interval Graph Completion and Strongly Chordal Graph Completion are in
FPT when parameterized by the number of edges to add. Recently, the problem left
open by [12], namely Interval Graph Completion was also shown to be in FPT [11].
Certainly, for all these problems the testing of the corresponding property is in P, while
for problems such as Hamiltonian Completion, where P is the class of Hamiltonian
graphs, there is no FPT algorithm, unless P=NP. For the same reason, one cannot expect
an FPT-algorithm when P contains all YES-instances of DP, even on planar graphs. We
consider an alternative way to parameterize completion problems, which is appropriate
for the hard case, i.e., when testing P is intractable: we parameterize the property itself.
In this paper, we initiate this line of research, by considering the parameterized property
Pk that contains all YES-instances of DP on planar graphs with k pairs of terminals.
Basic concepts. As open sets are not discrete structures, we introduce some formalism
that will allow us to move seamlessly from topological to combinatorial arguments. The
definitions may look involved at first reading, but this is warranted if one considers, as we
do, the problem in its full generality where the input graph is not necessarily connected.
3
Let G be a graph embedded in the sphere Σ0 . Embeddings are always without crossings and we view the graph G as a subset of Σ0 . Given a set X ⊆ Σ0 , let clos(X),
int(X), and ∂X denote the closure, the interior, and the boundary of X, respectively.
We define V (X) = V (G) ∩ ∂X. A noose is a Jordan curve of Σ0 that meets G only on
vertices. Let D be a finite collection of mutually non-intersecting open disks of Σ0 whose
boundaries are nooses and such that each point that belongs to at least two such nooses
S
is a vertex of G. We define ID = D∈D D and define ΓD as the Σ0 -embedded graph
whose vertex set is V (ID ) and whose edge set consists of the connected components of
the set ∂ID \V (ID ). Notice that, in the definition of ΓD , we permit multiple edges, loops,
or vertex-less edges.
Let J be an open subset of Σ0 . J is a cactus set of G if there is a collection D as
above such that J = ID , all biconnected components of the graph ΓD are cycles, V (G) ⊆
clos(J), and each edge of E(G) is a subset of J. Given such a J, we define ΓJ = ΓD . Two
cactus sets J and J0 of G are isomorphic if ΓJ and ΓJ0 are topologically isomorphic.
Throughout this paper, we use the standard notion of topological isomorphism between
planar embeddings, see Section 2.
Figure 1: A cactus set J and the vertices of V (J). Next to each vertex v we give its multiplicity
µ(v).
Given a cactus set J, we define for each vertex v ∈ V (J) its multiplicity µ(v) to be
equal to the number of connected components of the set ΓJ \ {v}, minus the number of
P
connected components of ΓJ , plus one. We also define µ(J) = v∈V (J) µ(v). Observe
that, given a cactus set J of G, the edges of G lie entirely within the interior of J. See
Figure 1. The boundary of J corresponds to a collection of simple closed curves such
that
4
(i) no two of them intersect at more than one point and
(ii) they intersect with G only at (some of) the vertices in V (G).
Cactus sets are useful throughout our paper as “capsule” structures that surround G
and thus they abstract the interface of a graph embedding with the rest of the sphere
surface.
We say that an open set F of Σ0 is an outer-cactus set of G if Σ0 \ clos(F) is a cactus
set of G. See Fig. 2.(ii). For example, if G is planar, any face F of G can be used to
define an outer-cactus set, whose boundary meets G only at the vertices incident to F.
Our definition of an outer-cactus set is more general: it can be a subset of a face F,
meeting the boundary of F only at some of its vertices.
Let G be an input graph to DP, see Figure 2.(i). Given an outer-cactus set F of G,
an F-patch of G is a pair (P, J) where (i) J is a cactus set of G, where Σ0 \ clos(J) ⊆ F
and (ii) P is a graph embedded in Σ0 without crossings such that E(P ) ⊆ Σ0 \ clos(J),
V (P ) = V (J) (see Figures 2.(iii) and 2.(iv)). Observe that the edges of P do not cross
any edge in E(G). In the definition of the F-patch, the graph P contains the new edges we
add. The vertices in V (F) define the vertices of G which we are allowed to include in P.
V (J) is meant to contain those vertices of V (F) that become vertices (possibly isolated)
of P. In terms of data structures, we assume that a cactus set J is represented by the
(embedded) graph ΓJ . Similarly an outer-cactus set F is represented by the (embedded)
graph ΓΣ0 \clos(F) .
We restate now the definition of the Planar Disjoint Paths Completion problem
as follows:
PDPC(G, s1 , t1 , . . . , sk , tk , `, F)
Input: A graph G embedded in Σ0 without crossings, terminals
s1 , t1 , . . . , sk , tk ∈ V (G), a positive integer `, and an outer-cactus set F of
G.
Parameter: k
Question: Is there an F-patch (P, J) of G, such that |P | ≤ ` and
DP(G ∪ P, s1 , t1 , . . . , sk , tk ) has a solution? Compute such an F-patch
if it exists.
If such an F-patch exists, we call it a solution for PDPC. In the corresponding optimization problem, denoted by min-PDPC, one asks for the minimum ` for which PDPC has
a solution, if one exists. See Fig. 2 for an example input of PDPC and a solution to it.
Our results. Notice that in the definition of PDPC the size of the patch does not
depend on the parameter k. Thus, it is not even obvious that PDPC belongs to the
5
Figure 2: An example input of the PDPC problem and a solution to it when ` = 8: (i) The
graph embedding in the input and the terminals s1 , t1 , s2 , t2 , s3 , t3 . The closure of the grey area
contains the graph G and the big vertices are the terminals. The white area is a face of G. (ii)
The input of the problem, consisting of G, the terminals and the outer-cactus set F. The solid
black vertices are the vertices of G that are also vertices of V (F). (iii) The solution of the problem
consists of the F-patch (P, J) where the edges of P are the dashed lines and J = J1 ∪ J2 ∪ J3 .
(iv) The input and the solution together where the validity of the patch is certified by 3 disjoint
paths.
parameterized complexity class XP, i.e., it has an algorithm of time nf (k) for some function f. Our first contribution, Theorem 2, is a combinatorial one: we prove that if a
k
patch exists, then its size is bounded by k 2 . Therefore, we can always assume that ` is
bounded by a function of k. This bound is the departure point for the proof of the main
algorithmic result of this paper:
Theorem 1. PDPC ∈ FPT. In particular, PDPC can be solved in f (k)·n2 steps, where
f is a function that depends only on k. Therefore, min-PDPC can be solved in g(k) · n2
steps.
We present now the proof strategy and the ideas underlying our results.
6
1.1
Proof strategy
Combinatorial Theorem. In Theorem 2, we prove that every patch whose size is
k
larger than k 2 , can be replaced by another one of strictly smaller size. In particular,
we identify a region B of F that is traversed by a large number of segments of different
paths of the DP solution. Within that region, we apply a global topological transformation that replaces the old patch by a new, strictly smaller one, while preserving its
embeddability in F. The planarity of the new patch is based on the fact that the new
segments are reflections in B of a set of segments of the feasible DP solution that previously lied outside B. This combinatorial result allows us to reduce the search space of
k
the problem to one whose size is bounded by min{`, k 2 }. Therefore, the construction of
the corresponding collection of “candidate solutions” can be done in advance, for each
given k, without requiring any a priori knowledge of the input graph G.
We note that the proof of our combinatorial theorem could be of independent interest.
In fact, it is one of the ingredients of the proof of the main result of [1].
The algorithm for PDPC. As the number of patches is bounded by a function of k,
we need to determine whether there is a correct way to glue one of them on vertices of
the boundary of the open set F so that the resulting graph is a YES-instance of the DP
problem. For each candidate patch P̃ , together with its corresponding candidate cactus
set J̃, we define the set of compatible graphs embedded in J̃. Each compatible graph
H̃ consists of unit-length paths and has the property that P̃ ∪ H̃ contains k disjoint
paths. Intuitively, each H̃ is a certificate of the part of the DP solution that lies within
G when the patch in F is isomorphic to P̃ . It therefore remains to check for each H̃
whether it can be realized by a collection of actual paths within G. For this, we set up
a collection H of all such certificates. Checking for a suitable realization of a member
of H in G is still a topological problem that depends on the embedding of G: graphs
that are isomorphic, but not topologically isomorphic, may certify different completions.
For this reason, our next step is to enhance the structure of the members of H so that
their realization in G reduces to a purely combinatorial check. (Cf. Section 4.1 for the
definition of the enhancement operation). We show in Lemma 6 that for the enhanced
certificates, this check can be implemented by rooted topological minor testing. For this
check, we can apply the recent algorithm of [10] that runs in h1 (k) · n3 steps and obtain
an algorithm of overall complexity h2 (k) · n3 .
We note that the use of the complicated machinery of the algorithm in [10] can be
bypassed towards obtaining a simpler and faster f (k) · n2 algorithm. This is possible
because the generated instances of the rooted topological minor problem satisfy certain
structural properties. This allows the direct application of the Irrelevant Vertex Tech7
nique introduced in [19] for solving, among others, the Disjoint Paths Problem. The
details of this improvement are in Section 5.
2
Preliminaries
We consider finite graphs. For a graph G we denote the vertex set by V (G) and the
edge set by E(G). If G is embedded in the sphere Σ0 , the edges of G and the graph G
refer also to the corresponding sets of points in Σ0 . Clearly the edges of G correspond
to open sets and G itself is a closed set. We denote by F (G) the set of all the faces of
G, i.e., all connected components of Σ0 \ G. Given a set S ⊆ V (G), we say that the pair
(G, S) is a graph rooted at S. We also denote as P(G) the set of all paths in G with at
least one edge. Given a path P ∈ P(G), we denote by I(P ) the set of internal vertices of
P. Given a vertex v ∈ V (G) and a positive integer r, we denote by NGr (v) the set of all
vertices in G that are within distance at most r from v. Given a vertex v of a graph G
with exactly two neighbors x and y, the result of the dissolution of v in G is the graph
obtained if we remove v from G and add, if it does not already exist, the edge {x, y}.
Rooted topological minors. Let H and G be graphs, SH be a subset of vertices
in V (H), SG be a subset of vertices in V (G), and ρ be a bijection from SH to SG .
We say that (H, SH ) is a ρ-rooted topological minor of (G, SG ), if there exist injections
ψ0 : V (H) → V (G) and ψ1 : E(H) → P(G) such that
1. ρ ⊆ ψ0 ,
2. for every e = {x, y} ∈ E(H), ψ1 (e) is a (ψ0 (x), ψ0 (y))-path in P(G), and
3. all e1 , e2 ∈ E(H) with e1 6= e2 satisfy I(ψ1 (e1 )) ∩ V (ψ1 (e2 )) = ∅.
In words, when H is a topological minor of G, G contains a subgraph which is
isomorphic to a subdivision of H. In addition, when (H, SH ) is a ρ-rooted topological
minor of (G, SG ) then this isomorphism respects the bijection ρ between the vertex sets
SH and SG .
Contractions. Let G and H be graphs and let σ : V (G) → V (H) be a surjective mapping such that
1. for every vertex v ∈ V (H), the graph G[σ −1 (v)] is connected;
2. for every edge {v, u} ∈ E(H), the graph G[σ −1 (v) ∪ σ −1 (u)] is connected;
3. for every {v, u} ∈ E(G), either σ(v) = σ(u), or {σ(v), σ(u)} ∈ E(H).
8
We say that H is a σ-contraction of G or simply that H is a contraction of G if such a
σ exists.
Observation 1. Let H and G be graphs such that H is a σ-contraction of G. If x, y ∈
V (G), then the distance in G between x and y is at least the distance in H between σ(x)
and σ(y).
We also need the following topological lemma.
Lemma 1. Let G be a Σ0 -embedded graph and let J be a cactus set of it. Let also M
be a Σ0 -embedded graph such that M ∩ J = ∅ and V (M ) ⊆ V (J). Then there is a closed
curve K in Σ0 \ clos(J) meeting each edge of M twice.
Proof. We consider the dual graph of ΓJ ∪ M and we remove from it all vertices lying
inside J. We denote by Q the resulting graph and notice that Q is connected because the
set Σ0 \ clos(J) is connected. Next, construct the graph Q0 by subdividing once every
edge of Q such that the subdivision vertex is the intersection point of the edge and its
dual. Let T be a spanning tree of Q0 and let K be a closed curve such that T is inside
one of the connected components of the set Σ0 \ K and J in the other. If we further
require K to intersect M a minimum number of times, we obtain the claimed curve.
Topological isomorphism. Given a graph G embedded in Σ0 , let f be a face in F (G)
whose boundary has ξ connected components A1 , . . . , Aξ . We define the set π(f ) =
{π1 , . . . , πξ } such that each πi is the cyclic ordering of V (Ai ), possible with repetitions,
defined by the way vertices are met while walking along Ai in a way that the face f is
always on our left side. Clearly, repeated vertices in this walk are cut-vertices of G.
Let G and H be graphs embedded in Σ0 . We say that G and H are topologically
isomorphic if there exist bijections φ : V (G) → V (H) and θ : F (G) → F (H) such that
1. φ is an isomorphism from G to H, i.e. for every pair {x, y} of distinct vertices in
V (G), {x, y} ∈ E(G) iff {φ(x), φ(y)} ∈ E(H).
2. For every face f ∈ F (G), φ(π(f )) = π(θ(f )).
In the definition above, by φ(π(f )) we mean {φ(π1 ), . . . , φ(πξ )}, where, if πi = (x1 , . . . ,
xζi , x1 ), then by φ(πi ), we mean (φ(x1 ), . . . , φ(xζi ), φ(x1 )). Notice that it is possible for
two isomorphic planar graphs to have embeddings that are not topologically isomorphic
(see [6, page 93] for such an example and further discussion on this topic).
Treewidth. A tree decomposition of a graph G is a pair (X , T ) where T is a tree with
nodes {1, . . . , m} and X = {Xi | i ∈ V (T )} is a collection of subsets of V (G) (called
bags) such that:
9
S
1. i∈V (T ) Xi = V (G),
2. for each edge {x, y} ∈ E(G), {x, y} ⊆ Xi for some i ∈ V (T ), and
3. for each x ∈ V (G) the set {i | x ∈ Xi } induces a connected subtree of T.
The width of a tree decomposition ({Xi | i ∈ V (T )}, T ) is maxi∈V (T ) {|Xi | − 1}. The
treewidth of a graph G denoted tw(G) is the minimum width over all tree decompositions
of G.
3
Bounding the size of the completion
In this section we show the following.
Theorem 2. If there is a solution for PDPC(G, s1 , t1 , . . . , sk , tk , `, F), then there is a
k
solution (P, J) with |E(P )| ≤ k 2 .
For the proof, we need the following combinatorial lemma.
Lemma 2. Let Σ be an alphabet of size |Σ| = k. Let w ∈ Σ∗ be a string over Σ. If
|w| > 2k , then w contains an infix y with |y| ≥ 2, such that every letter occurring in y,
occurs an even number of times in y.
Proof. Let Σ = {a1 , . . . , ak }, and let w = w1 · · · wn with n > 2k . Define vectors
zi ∈ {0, 1}k for i ∈ {1, . . . , n}, and we let the jth entry of vector zi be 0 if and only if
letter aj occurs an even number of times in the prefix w1 · · · wi of w and 1 otherwise. Since
n > 2k , there exist i, i0 ∈ {1, . . . , n} with i 6= i0 , such that zi = zi0 . Then y = wi+1 · · · wi0
proves the lemma.
We also need the following easy topological lemma.
Lemma 3. Let G be a connected outerplanar graph that may have loops but no multiple
edges (or loops) and let Γ be an embedding of G in Σ0 such that all its vertices belong to
the boundary of a face f . Let also Ein be the set of all the edges or loops of G that do
not belong to the boundary of f . Then there is a unique (up to topological isomorphism)
embedding Γ0 of G such that Γ ∩ Γ0 is the boundary of f and that all edges of Ein are
embedded inside f .
Proof. Let Γ̃ be the boundary of f . Let F be the set of all connected components of
Σ0 \ Γ̃, except from f .
For each e ∈ Ein we denote by Se the set of endpoints of e. For each edge e ∈ Ein
we define a set ē that will be the image of e in the new embedding Γ0 . We distinguish
the following two cases.
10
Case 1. |Se | = 2. Notice that Se is a separator of G. Notice also that G \ Se contains
two connected components. Also there exists a (unique) cycle Ce containing the edge e
and no other points of Γ such that each of these connected components belong each to a
distinct disk among the two disks that are bounded by Ce . We define ē = Ce \ clos(e).
Case 2. |Se | = 1. Notice that there exists a unique cycle Ce defining two disks D and
D0 such that (i) Ce ∩ Γ = Se and (ii) Γe ⊆ D where Γe is the (unique) member of F that
contains the loop e. We define ē = Ce \ Se .
Notice that it is possible to define all ē’s, one by one, such that they do not intersect.
S
Given the above definitions, it follows that Γ̃ ∪ e∈Ein ē is the required embedding.
Proof of Theorem 2. Let (P, J) be a solution for PDPC(G, s1 , t1 , . . . , sk , tk , `, F) with
|E(P )| minimal. Consider the embedding of G ∪ P in the sphere Σ0 , and let Q1 , . . . , Qk
be the paths of a DP solution in G ∪ P. By the minimality of |E(P )| we can assume that
S
the edges of P are exactly the edges of i∈{1,...,k} Qi that are not in G. For the same
reason, two edges in P have a common endpoint x that is not a terminal only if x is a
cut-vertex of ΓJ .
Let P ∗ denote the graph obtained from the dual of P ∪ ΓJ , after removing the
vertices corresponding to the faces of ΓJ that are disjoint from F. We first show that the
maximum degree of P ∗ is bounded by k. Assume, to the contrary, that P ∗ has a vertex
incident to two edges e∗1 and e∗2 such that e1 and e2 belong to the same path, say Qx ,
in {Q1 , . . . , Qk }. Then it is possible to choose an endpoint p1 of e1 and an endpoint p2
of e2 such that, (P 0 , J0 ) is also a solution such that |E(P 0 )| < |E(P )|. Indeed, p1 and
p2 are the vertices incident to e1 and e2 belonging to the two connected components of
Qx \ {e1 , e2 } that contain respectively sx and tx . Therefore we established that for every
i ∈ [k], no vertex of P ∗ is incident to two distinct edges of Qi , thus all vertices of P ∗
have degree at most k.
Here P 0 = P \ {e1 , e2 } ∪ {{p1 , p2 }} and J0 is defined such that ΓJ0 is obtained from
ΓJ after dissolving the vertices that became isolated during the construction of P 0 .
Our next aim is to prove that the diameter of P ∗ is bounded by 2k . Then |E(P ∗ )| =
k
|E(P )| ≤ k 2 and we are done.
Note that every edge in E(P ∗ ) corresponds to an
edge in exactly one path of Q1 , . . . , Qk . Hence, given a path R = (r0 , . . . , rζ ) in P ∗ , it
corresponds to a string w ∈ {Q1 , . . . , Qk }∗ in a natural way. It is enough to prove the
following claim.
Claim. The string w contains no infix y with |y| ≥ 2, such that every letter occurring
in y occurs an even number of times in y.
11
Proof of the Claim. Towards a contradiction, suppose that w contains such an infix
y. We may assume that w = y. Let ER ⊆ E(P ) be the set of edges corresponding to the
edges of the path R ⊆ P ∗ . Then |ER | ≥ 2 because w (and hence R) has length at least
2. Let B ⊆ Σ0 be the open set defined by the union of all edges in ER and all faces of
the graph P ∪ ΓJ that are incident to them. Clearly, B is an open connected subset of
F with the following properties:
(a) B contains all edges in ER and no other edges of P ,
(b) the ends of every edge in ER lie on the boundary ∂B, and
(c) every edge in E(P ) \ ER has empty intersection with B.
We consider an ‘up-and-down’ partition (U = {u1 , . . . , ur }, D = {d1 , . . . , dr }) of the
endpoints of the edges in ER as follows: traverse the path R in P ∗ in some arbitrary
direction and when the ith edge ei ∈ ER is met, the endpoint ui of ei on the left of this
direction is added to U and the right endpoint di is added to D. Notice that U and D
may be multisets because it is not necessary that all vertices in P have degree 1. For
each i ∈ {1, . . . , r} we say that ui is the counterpart of di and vice versa.
Because the paths Q1 , . . . , Qk are vertex-disjoint, the following holds.
Observation: If x ∈ V (P ) has degree larger than one, then either x is a terminal and
has degree at most k or x is a cutpoint of ΓJ and has degree exactly 2.
By assumption, every path Qi crosses R an even, say 2ni , number of times. Now
for every path Qi satisfying E(Qi ) ∩ ER 6= ∅, we number the edges in E(Qi ) ∩ ER
by ei1 , . . . , ei2ni in the order of their appearance when traversing Qi from si to ti and
we orient them from si to ti . We introduce shortcuts for Qi as follows: for every odd
number j ∈ {1, . . . , 2ni }, we replace the subpath of Qi from tail(eij ) to head(eij+1 ) by a
new edge fji in D (see Figure 3).
After having done this for all odd numbers j ∈ {1, . . . , 2ni }, we obtain a new path
0
Qi from si to ti that uses strictly less edges in B than Qi . Having replaced all paths
Qi with E(Qi ) ∩ ER 6= ∅ in this way by a new path Q0i , we obtain from P a new graph
P 0 by replacing every pair of edges eij , eij+1 ∈ E(P ) by fji for all i ∈ {1, . . . , k} with
0 the set of all
E(Qi ) ∩ ER 6= ∅, and for all j ∈ {1, . . . , 2ni }, j odd. We denote by ER
replacement edges fji . We also remove every vertex that becomes isolated in P 0 during
this operation.
Then it is easy to verify that:
• None of the edges of ER survives in E(P 0 ).
• |E(P 0 )| < |E(P )| .
12
Figure 3: Example of the transformation in the proof of the Claim in the proof of
Theorem 2; P is on the left and P 0 is shown on the right. The dashed lines represent
the edges of C.
• DP(G ∪ P 0 , s1 , t1 , . . . , sk , tk ) has a solution.
If we show that, for some suitable cactus set J0 of G, (P 0 , J0 ) is an F-patch, then we are
done, because |E(P 0 )| < |E(P )|. In what follows, we prove that P 0 can also be embedded
without crossings in clos(F) such that E(P 0 ) ⊆ Σ0 \ ∂F. For this it suffices to prove that
0 can be embedded in B without crossings.
the edges in ER
For every path Qi with E(Qi )∩ER 6= ∅ let Fji denote the subpath of Qi from head(eij )
to tail(eij+1 ), for j ∈ {1, . . . , 2ni }, j odd (this path may be edgeless only in the case where
head(eij ) = tail(eij+1 ) is a cut-vertex of ΓJ ). We replace Fji by a single edge cij (when
the corresponding path is edgeless, the edge cij is a loop outside B). We consider the
graph C with vertex set V (P ) and edge set {cij | i ∈ {1, . . . , k}, E(Qi ) ∩ ER 6= ∅, j ∈
{1, . . . , 2ni }, j odd}.
Our strategy consists of a two-step transformation of this embedding. The first step
creates an embedding of C inside clos(B) without moving the vertices. Indeed, notice
that C is embedded in Σ0 \ B without crossings such that all the endpoints of the edges
in C lie on the boundary of B. As for every odd j in [2ni ], j + 1 also belongs to [2ni ],
it follows that none of the endpoints of Fji can be a terminal. This, together with the
observation above, implies that no two edges of C have a common endpoint.
By applying Lemma 3 on the Σ0 -embedded outerplanar graph Γ = ΓJ ∪ C where B
plays the role of the face f , we obtain a new non-crossing embedding Γ0 . Notice that
Γ0 \ ΓJ is a new non-crossing embedding of C where all of its edges lie inside B. (Recall
that none of the edges of ER is present in this embedding.) This transformation maps
every edge cij to a new edge inside B with the same endpoints.
The second step “reflects” the resulting embedding along the axis defined by the
path R such that each vertex is exchanged with its counterpart. Now define (cij )0 so
13
that it connects tail(eij ) and head(eij+1 ) – these are exactly the counterparts of head(eij )
and tail(eij+1 ). Due to symmetry, the (cij )0 are pairwise non-crossing, and none of them
crosses a drawing of an edge in E(P ) \ ER . Hence the (cij )0 together with the drawing
of edges in E(P ) \ ER provide a planar drawing of P 0 (where (cij )0 is the drawing of fji ).
We finally define (up to isomorphism) the cactus set J0 of G such that ΓJ0 is obtained
from ΓJ after dissolving the vertices of P ∩ ΓJ that are isolated in P 0 . It is easy to verify
that (P 0 , J0 ) is an F-patch of G. This concludes the proof of the Claim.
Using the Claim above and from Lemma 2, it follows that |w| ≤ 2k , and hence the
diameter of P ∗ is bounded by 2k . The proof of Theorem 2 is now complete.
k
Let L be a list of all simple plane graphs with at most min{`, k 2 } edges and no
isolated vertices. We call a graph in L a completion. As a first step, our algorithm for
PDPC computes the list L. Obviously, the running time of this process is bounded by
a function depending only on k.
4
The algorithm for Planar-Dpc
The fact that the size of L is bounded by a function of k implies that PDPC is in XP.
Indeed, given the list L, for each completion P̃ ∈ L we define the graph QP̃ = (V (P̃ ), ∅)
and we consider all cactus sets J̃ of QP̃ where (P̃ , J̃) is a (Σ0 \ clos(J̃))-patch of QP̃
and V (J̃) = V (P̃ ). We denote the set of all such pairs (P̃ , J̃) by J and observe that the
number of its elements (up to topological isomorphism of the graph P̃ ∪ ΓJ̃ ) is bounded
by a function of k.
For each pair (P̃ , J̃) ∈ J , we check whether there exists an F-patch (P, J) of G such
that P̃ ∪ ΓJ̃ and P ∪ ΓJ are topologically isomorphic and DP has a solution in the graph
G∪P. As there are nz(k) ways to choose (P, J) and each check can be done in O(z1 (k)·n3 )
steps, we conclude that PDPC can be solved in nz2 (k) steps. In the remainder of the
paper, we will prove that the problem is actually in FPT.
The main bottleneck is that there are too many ways to identify V (J̃) with vertices
of V (F), because we cannot bound |V (F)| by a function of k. To overcome this, we
associate with a positive instance of PDPC a rooted topological minor (H̃, T̃ ) of the
original graph G, that witnesses the fact that (P̃ , J̃) corresponds to the desired F-patch
of G. For convenience, we assume from now on that Γint(Σ0 \F) does not contain any
loops or multiple edges. In other words, none of the open disks corresponding to the
connected components of Σ0 \ F contains less than 3 vertices in its boundary. If the
input instance violates this property we may enforce it by adding isolated vertices in G
14
Figure 4: In the leftmost image the dotted lines are the edges of P̃ . Together with the interior of
the grey areas they form a pair (P̃ , J̃) ∈ J . J̃ has two weakly connected components. V (J̃) consists of the 12 vertices on the boundary of the grey areas. The solid lines that intersect the open
set J̃ are the edges of the graph H̃, which is compatible with (P̃ , J̃). Let T̃ = {s1 , t1 , . . . , s4 , t4 }.
The triple τ = (J̃, H̃, T̃ ) ∈ H has two parts. The middle and the leftmost pictures show how
each of these parts is enhanced in order to construct the graphs Rτ1 and Rτ2 .
and modify F such that the new vertices belong to its boundary. Similarly, we restrict
J to contain only pairs (P̃ , J̃) where ΓJ̃ does not contain any loops or multiple edges.
Given a pair (P̃ , J̃) ∈ J , we say that a rooted simple graph (H̃, T̃ = {a1 , b1 , . . . ,
ak , bk }) embedded in Σ0 , is compatible with (P̃ , J̃) when
1. for every e ∈ E(H̃), e ⊆ J̃,
2. the vertices of H̃ that are also vertices of ΓJ̃ are |V (P̃ )|, while the vertices of H̃
that are not vertices of ΓJ̃ are all vertices of T̃ ,
3. V (H̃) \ T̃ ⊆ V (J̃) ⊆ V (H̃),
4. DP(P̃ ∪ H̃, a1 , b1 , . . . , ak , bk ) has a solution.
We define
H = {(J̃, H̃, T̃ ) | there exists a (P̃ , J̃) ∈ J
such that (H̃, T̃ ) is compatible with (P̃ , J̃)
and notice that |H| is bounded by some function of k. See the leftmost part of Fig. 4 for
an example of a triple in H.
Assuming that (P, J) is a solution for PDPC(G, s1 , t1 , . . . , sk , tk , `, F), consider the
parts of the corresponding disjoint paths that lie within G. The intuition behind the
15
definition above is that H̃ is a certificate of these “partial paths” in G. Clearly, the
number of these certificates is bounded by |H| and they can be enumerated in f0 (k)
steps, for some suitable function f0 . For example, for the solution depicted in Fig. 4,
H̃ consists of 7 disjoint edges, one for each subpath within G. Our task is to find an
FPT-algorithm that for every such certificate checks whether the corresponding partial
paths exist in G.
Given an open set O, a weakly connected component of O is the interior of some
connected component of the set clos(O). Notice that a weakly connected component is
not necessarily a connected set.
Let F̄1 , . . . , F̄λ be the weakly connected components of the set Σ0 \ clos(F). We call
such a component F̄i active if clos(F̄i ) ∩ T 6= ∅. We denote the collection of all active
components by FF . A crucial observation is that if an F-patch exists we can always
replace it by one that bypasses the inactive components.
Lemma 4. Let (G, s1 , t1 , . . . , sk , tk , F) be an instance for the PDPC problem and
S
S
let G0 = G[ F̄i ∈FF clos(F̄i ) ∩ V (G)] and F0 = Σ0 \ F̄i ∈FF clos(F̄i ). Then
(G0 , s1 , t1 , . . . , sk , tk , F0 ) is an equivalent instance.
Proof. For the non-trivial direction, we assume that G has an F-patch (P, J) such
that G ∪ P contains a collection {P1 , . . . , Pk } paths that certify the feasibility of DP(G ∪
S
P, s1 , t1 , . . . , sk , tk ). Let P 0 be the graph obtained if in ( i=1,...,k Pi )∩clos(F0 ) we dissolve
all vertices that are in F0 . Let also J0 be the union of all weakly connected components
of the set J that contain some open set from FF . Observe that (P 0 , J0 ) is an F0 -patch of
G0 .
By Lemma 4, we can assume from now on that the number λ of the weakly connected
components of the set Σ0 \ clos(F) is at most 2k. Also we restrict H so that it contains
only triples (J̃, H̃, T̃ ) such that the weakly connected components of the set J̃ are exactly
λ.
4.1
The enhancement operation
Consider the triple τ = (J̃, H̃, T̃ ) ∈ H. Let J̃1 , . . . , J̃λ be the weakly connected components of the set J̃. Then we define C̃ i = ΓJ̃i ∪ (clos(J̃i ) ∩ H̃) for i ∈ {1, . . . , λ} and
we call them parts of τ. Also we set T̃ i = T̃ ∩ V (C̃ i ), 1 ≤ i ≤ λ. We now apply the
following enhancement operation on each part of τ : For i = 1, . . . , λ, we consider the
sequence Rτ = (Rτ1 , . . . , Rτλ ) where Rτi is the rooted graph (Rτ0i , T̃ i ∪ {xinew }) such that
Rτ0i is defined as follows. Take the disjoint union of the graph C̃ i and a copy of the wheel
Wµ(J̃i ) with center xinew and add µ(J̃i ) edges, called i-external, between the vertices of
16
V (J̃i ) and the peripheral vertices of Wµ(J̃i ) such that the resulting graph remains Σ0 embedded and each vertex v ∈ V (J̃i ) is incident to µ(v) non-homotopic edges not in J̃.
As the graph ΓJ̃i is connected and planar, the construction of Rτ0i is possible. Observe
also that Rτ0i \ J̃i is unique up to topological isomorphism. To see this, it is enough to
verify that for every two vertices in Rτ0i \ J̃i of degree ≥ 3 there are always 3 disjoint
paths connecting them (here we use our assumption that ΓJ̃ does not contain any loops
or multiple edges).
We define R = {Rτ | τ ∈ H} and observe that |R| is bounded by a function of k
because the same holds for |H|. (For an example of the construction of Rτ , see Fig. 4)
We now define (C 1 , . . . , C λ ) such that C i = ΓF̄i ∪ (clos(F̄i ) ∩ G), i ∈ {1, . . . , λ}. We
call the graphs in (C 1 , . . . , C λ ) parts of G and let T i = T ∩ V (C i ), 1 ≤ i ≤ λ, where
T = {s1 , t1 , . . . , sk , tk }. As above we define the enhancement of the parts of G as follows.
0i
For each i = 1, . . . , λ we define the rooted graph G∗i = (G0i , T i ∪ {x∗i
new }) where G is
∗
defined as follows: take the disjoint union of C i and the wheel Wµ(
with center x∗i
new
F̄i )
i
i
and add µ(F̄ ) edges, called ∗i-external, between the vertices of V (F̄ ) and the peripheral
∗
vertices of Wµ(
such that the resulting graph remains Σ0 -embedded and each vertex
F̄i )
i
v ∈ V (F̄ ) is incident to µ(v) non-homotopic edges. As above, each G0i is possible to
construct and G0i \ F̄i is unique up to topological isomorphism.
The purpose of the above definitions is twofold. First, they help us to treat separately
each of the parts of G and try to match them with the correct parts of τ. Second, the
addition of the wheels to each part gives rise to a single, uniquely embeddable interface,
between the part and its “exterior” and this helps us to treat embeddings as abstract
graphs. Therefore, to check whether a part of τ is realizable within the corresponding
part of G, we can use the rooted version of the topological minor relation on graphs as
defined in Section 2.
4.2
The stretching lemma
A bijection ρ from T̃ = {a1 , b1 , . . . , ak , bk } to T = {s1 , t1 , . . . , sk , tk } is legal if for every
i ∈ {1, . . . , k}, there exists some j ∈ {1, . . . , k} such that ρ((ai , bi )) = (sj , tj ).
Let τ ∈ H and let ρ be a legal bijection from T̃ to T and let ρi be the restriction
of ρ in T̃i . We say that Rτ = (Rτ1 , . . . , Rτλ ) is ρ-realizable in G if there exists a bijection
φ : {1, . . . , λ} → {1, . . . , λ} such that for i = 1, . . . , λ, Rτi is a ρ̂i -rooted topological minor
∗φ(i)
of G∗φ(i) were ρ̂i = ρi ∪ {(xinew , xnew )}.
By enumerating all possible bijections φ, we enumerate all possible correspondences
between the parts of of G and the parts of τ. In order to simplify notation, we assume
in the remainder of this section that φ is the identity function.
The following lemma is crucial. It shows that when Rτi is a topological minor of G∗i
17
we can always assume that all vertices and edges of C̃ i are mapped via ψ0 and ψ1 to
vertices and paths in clos(F̄i ); the wheel Wµ(J̃i ) is mapped to a “sub-wheel” of Wµ(F̄i )
while i-external edges are mapped to ∗i-external edges. This will be useful in the proof
of Lemma 6, as the i-external edges represent the interface of the completion P̃ with
C̃ i . The topological minor relation certifies that the same interface is feasible between
the corresponding part C i of G and its “exterior”. Lemma 5 establishes also that the
image of ΓJ̃i can be “stretched” so that it falls on ΓF̄i . As all the vertices in V (F̄i ) are
∗i
within distance 2 from the artificial terminal x∗i
new in G , this allows us later in the
proof of Lemma 9 to locate within G∗i the possible images of V (Rτ0i ) in a neighborhood
of the terminals. It is then safe to look for an irrelevant vertex “far away” from this
neighborhood.
Lemma 5. Let Rτi be a ρ̂i -rooted topological minor of G∗i where ρ̂i = ρi ∪{(xinew , x∗i
new )},
i
i
for i = 1, . . . , λ. Let also ψ0 and ψ1 be the functions (cf. Section 2) certifying this topological minor relation. Then ψ0i and ψ1i can be modified so that the following properties
are satisfied.
1. if ẽ is an edge of the wheel Wµ(J̃i ) incident to xinew , then ψ1i (ẽ) is an edge incident
to x∗i
new .
2. if ẽ is an edge of the wheel Wµ(J̃i ) not incident to xinew , then ψ1i (ẽ) is an x∗i
new ∗
∗
avoiding path of Wµ(F̄i ) . Moreover, each edge in Wµ(F̄i ) is contained in a path
ψ i (f˜) for some edge f˜ of the cycle W ∗
\ xi .
1
µ(J̃i )
new
3. if ẽ is an i-external edge between V (J̃i ) and V (Wµ(J̃i ) ) \ {xinew }, then ψ1 (e) is a
path consisting of an ∗i-external edge between V (F̄i ) and V (Wµ(F̄i ) ) \ {x∗i
new }.
4. ψ0i (V (J̃i )) ⊆ V (F̄i ).
Proof. Our first step is to enforce Properties 1 and 2. For each edge ẽ = {xinew , x} of
Rτi , let ẽ0 = {x, x0 } be the unique edge of Rτi that is not an edge of Wµ(J̃i ) . Consider the
path Pẽ that is formed by the concatenation of ψ1i (ẽ0 ) and ψ1i (ẽ) and let x̂ be the neighbor
i
of ψ0i (xinew ) = x∗i
new in the path ψ1 (Pẽ ). For each such ẽ and its incident vertex x, we
simultaneously update ψ0i such that ψ0i (x) = x̂. Accordingly, for every ẽ = {xinew , x}
we simultaneously update ψ1i so that ψ1i ({xinew , x}) is the path consisting of the edge
i
0
i 0
{x∗i
new , x̂} (enforcing Property 1) and ψ1 (x, x ) is the subpath of Pẽ between x̂ and ψ0 (x ).
We also update ψ1i such that for each ẽ1 = {xinew , x1 } and ẽ2 = {xinew , x2 } where x1 and
∗
x2 are adjacent, we set ψ1i ({x1 , x2 }) to be the (x̂1 , x̂2 )-path of Wµ(
that does not meet
F̄i )
i
any other vertex of ψ0 (V (Wµ(J̃i ) )), enforcing Property 2.
18
So far, we have enforced that for each i ∈ {1, . . . , λ}, the wheel Wµ(F̄i ) is mapped
via ψ0i and ψ1i to the wheel Wµ(J̃i ) (by slightly abusing the notation, we can say that
ψ1 (Wµ(J̃i ) ) = Wµ(F̄i ) ). Notice that Properties 1 and 2 imply that all vertices and edges
of C̃ i are mapped via ψ0 and ψ1 to vertices and paths of clos(F̄i ).
Our second step is to enforce Properties 3 and 4. The transformation that we describe
next, essentially “stretches” the image of J̃i until it hits from within the boundary of
F̄i . Let ẽ = {x, x0 } be an edge of Rτi where x ∈ V (J̃i ) and x0 ∈ V (Wµ(J̃i ) ) \ {xinew }.
We also set Pẽ = ψ1i (ẽ). By definition, Pẽ is a path in G∗i which starts from ψ0i (x) and
ends at ψ0i (x0 ). As ψ0i (x) is a point of clos(F̄i ) and ψ0i (x0 ) is not, we can define x̂ as the
first vertex of Pẽ that is a vertex of V (F̄i ). For each such ẽ and its incident vertex x,
we simultaneously update ψ0i such that ψ0i (x) = x̂. Accordingly, for every ẽ = {x, x0 } we
simultaneously update ψ1i so that ψ1i ({x, x0 }) is the path consisting of the single edge
{x̂, ψ0i (x0 )} – thus enforcing Property 3 – and for each edge {x1 , x2 } of ΓJ̃i we update ψ1i
so that ψ1i ({x1 , x2 }) is a (x̂1 , x̂2 )-path avoiding any other vertex of ψ0i (V (J̃i )). As now
all images of the vertices and the edges of ΓJ̃i lie on ΓF̄i , Property 4 holds.
4.3
Reducing PDPC to topological minor testing
Lemma 6. PDPC(G, t1 , s1 , . . . , tk , sk , `, F) has a solution if and only if there exists a
τ = (J̃, H̃, T̃ ) ∈ H and a legal bijection ρ : T̃ → T such that Rτ is ρ-realizable in G.
Proof. Suppose that (P, J) is a solution for PDPC(G, t1 , s1 , . . . , tk , sk , `, F) giving rise
to a collection P = {P1 , . . . , Pk } of disjoint paths in G ∪ P where Pi connects si with ti .
From Theorem 2, we can assume that |V (J)| is bounded by a suitable function of k
and therefore, (P, J) is isomorphic to a member (P̃ , J̃) of J , in the sense that P̃ ∪ ΓJ̃
and P ∪ ΓJ are topologically isomorphic. Let (H, T ) be the rooted subgraph of G formed
by the edges of the paths in P that are also edges of G. We now define the rooted
graph (H 0 , T ) by dissolving all non-terminal vertices in the interior of J and observe
that (H 0 , T ) is isomorphic, with respect to some bijection ω, to some (H̃, T̃ ) that is
compatible with (P̃ , J̃). Therefore τ = (J̃, H̃, T̃ ) ∈ H and let ρ be the correspondence
between T̃ and T , induced by ω. By partitioning ω with respect to the weakly connected
components of the set J̃, we generate a bijection φ between the parts {C̃ 1 , . . . , C̃ λ } of τ
and the parts {C 1 , . . . , C λ } of G such that a subdivision of C̃ i is topologically isomorphic
to a subgraph of C φ(i) , i ∈ {1, . . . , λ}. This topological isomorphism can be extended in
the obvious manner to the graphs obtained after the enhancement operation applied to
the parts of τ and the parts of G. This translates to the fact that each of the resulting Rτi
is a ρi -rooted topological minor of G∗φ(i) where each ρi is obtained by the restriction of
19
ρ to the vertices of V (Rτi ) plus the pair (xinew , x∗i
new ). We conclude that Rτ is ρ-realizable
in G.
For the opposite direction, let τ ∈ H. By renumbering if necessary the parts of G, let
be a ρ̂i -rooted topological minor of G∗i were ρ̂i = ρi ∪ {(xinew , x∗i
new )}, for i = 1, . . . , λ.
i
i
Let also ψ0 and ψ1 be the functions satisfying the four properties of Lemma 5.
Rτi
+
when H̃ + is the leftmost graph of Figure 4.
Figure 5: The graph H̃K
For the given τ = (J̃, H̃, T̃ ) ∈ H we assume w.l.o.g. that |V (J̃)| is minimal. Recall
that there exists a P̃ such that (H̃, T̃ ) is compatible with (P̃ , J̃). Let H̃ + = H̃ ∪ ΓJ̃ ∪ P̃ .
From Lemma 1, there exists a curve K where K ⊆ Σ0 \ clos(J̃) and such that K
intersects each edge of P̃ twice. Let ∆in and ∆out be the two connected components
of the set Σ0 \ K and w.l.o.g. we may assume that clos(J̃) ⊆ ∆out . If we consider all
connected components of the set K \ H̃ + as edges and take their union with the graph
obtained by H̃ + after subdividing the edges of P̃ at the points of K ∩ H̃ + , we create
+
+
+
∩ ∆in
= H̃K
a new graph H̃K
(see Figure 5 for an example). We also define H̃K,in
+(i)
+
+
and H̃K,out
= H̃K
∩ clos(∆out ). For i ∈ {1, . . . , λ}, we define H̃K,out as follows: first
+
take the graph obtained by H̃K,out
if we remove all edges except those that have at
least one endpoint in C̃ i and those that belong in K and, second, dissolve in this graph
all vertices that have degree 2. If we now add a new vertex xinew ∈ ∆in and make it
adjacent to all remaining vertices in K, it follows, by the minimality of the choice of τ ,
that the resulting graph is isomorphic to Rτi . In the above construction, we considered
+
the vertex set Q = K ∩ H̃ + ⊆ V (HK
) and we slightly abuse notation by denoting by Q
i
also its counterpart in Rτ . We also denote by πQ the cyclic ordering of Q defined by the
20
order of appearance of the vertices in Q along K. Similarly, for i ∈ {1, . . . , λ}, we set
+(i)
Qi = Q ∩ V (H̃K,out ) and we denote by πQi the induced sub-ordering of πQ . We denote by
Ẽ i = {ẽi1 , . . . , ẽiµ(J̃i ) , ẽi1 } the edges of Rτi , not in Wµ(J̃i ) , that are incident to the vertices
of Qi , cyclically ordered according to πQi .
We now take the graph G∗i and remove from it all edges outside clos(F̄i ) except from
the images of the edges of Ẽ i (observe that clos(F̄i ) is always a connected set while this
is not necessarily the case for F̄i ). From Property 3 in Lemma 5, these images are just
∗i−
edges between V (F̄i ) and V (Wµ(F̄i ) ) \ {x∗i
new }. We denote the resulting graph by G
and observe that it can be drawn in Σ0 in a way that the vertices in ψ0 (Qi ) lie on a
virtual closed curve in accordance with the cyclic ordering πQi of their ψ0i -preimages.
We consider the disjoint union G∗out of all G∗i− and we observe that it is possible to
S
embed it in Σ0 such that all vertices in Q∗ = i∈{1,...,λ} ψ0i (Qi ) lie on the same closed
curve K ∗ . By the way each πQi is defined from πQ and the bijection between Q and
Q∗ we directly obtain that the following properties are satisfied: (i) the new embedding
is planar, (ii) G∗out is a subset of one of the connected components of the set Σ0 ∩ K ∗ ,
(iii) for all i ∈ {1, . . . , λ}, the cyclic ordering of ψ0i (Qi ) is an induced sub-ordering of
the cyclic ordering defined by the way the vertices of Q∗ are being arranged along K ∗ .
S
Note that ψ0 = i∈{1,...,λ} ψ0i defines a bijection from the vertices of Q to the vertices of
+
and renaming each vertex
Q∗ . Our next step is to obtain the graph G∗in by taking H̃K,in
+
∗
∗
x ∈ Q to ψ0 (x). Now the graph G = Gin ∪ Gout consists of the original graph G and a
collection Z of paths of length 3.
We are now ready to define the desired F-patch (P, J) of G. The edges of P are
created by dissolving all internal vertices of each path in Z. Also, J can be any cactus
set for which the embedding of ΓJ results from the embedding of ΓF after dissolving all
vertices of V (F) that are not in ψ0 (V (J̃)).
In [13], Grohe, Kawarabayashi, Marx, and Wollan gave an h1 (k) · n3 algorithm for
checking rooted topological minor testing, where h1 is some computable function. Combining their algorithm with Lemma 6, we obtain an h2 (k)·n3 algorithm for PDPC (here,
again, h2 is some computable function). Therefore, PDPC ∈ FPT.
In the next section, we show how to obtain the improved running time claimed in
Theorem 1.
5
Applying the irrelevant vertex technique
Lemma 6 established that for every candidate patch P̃ , certifying its feasibility for PDPC
on G reduces to finding a rooted topological minor. By the results of [10], PDPC is in
FPT by an h2 (k) · n3 algorithm. In this section we show that this running time can be
21
improved. We first explain the intuition behind this improvement.
5.1
Proof strategy
At a high level, we have reduced the validity of P̃ to the problem of determining whether
a q-vertex rooted graph H is a rooted topological minor of an n-vertex graph G, where
q is bounded by some function of k. By Lemma 5, we can assume that the images of
the vertices of H are either terminals of G or lie on the boundary of the same face of G.
This observation makes it possible to directly employ, in Lemma 9, the irrelevant vertex
technique [19].
In particular, if the treewidth of G is big enough, one can detect a sufficiently large
set of concentric cycles that are away from the images of the vertices of H in G; this
is possible due to Lemma 5. Using the “vital linkage” theorem of Robertson and Seymour [18, 17] (see also [13]), we obtain that the topological minor mapping can be
updated so that the realization of H avoids the inner cycle of this collection. Therefore,
the removal of any of the vertices of this cycle creates an equivalent instance of the
problem with a smaller number of vertices. By repeating this vertex-removal operation,
we end up with a graph whose treewidth is bounded by some function of q. In this
case, since the rooted variant of the topological minor checking problem is definable in
Monadic Second Order Logic (msol) (see Lemma 10 in the Appendix), the problem can
be solved in a linear number of steps according to Courcelle’s Theorem.
For the running time of our algorithm, we use the fact that the detection of an
irrelevant vertex in planar graphs requires to find a vertex that is “far enough” from all
the terminals. As this can be done by standard BFS in O(n) steps and at most n such
vertices are deleted, the overall complexity of the algorithm is f (k) · n2 .
5.2
Treewidth and linkages
Let G be a graph. A linkage in G is a set of pairwise disjoint paths of it. The endpoints
of a linkage L are the endpoints of the paths in L. The pattern of L is defined as
π(L) = {{s, t} | L contains a path from s to t}
Consider now the rooted graph G = (G0 , T ) where G0 is a Σ0 -embedded graph. We
call a cycle of G0 T -respectful if all the vertices of T are inside one of the two connected
components of the set Σ0 \ C. Given a T -respectful cycle C of G0 we denote by ∆ext (C)
the connected component of the set Σ0 \ C that contains T and by ∆int (C) the other.
A sequence C1 , . . . , Ck of T -respectful cycles in G is T -concentric if ∆ext (C1 ) ⊆ · · · ⊆
∆ext (Ck ).
22
Figure 6: The leftmost graph is Γ6 . The insulation of x in the rightmost graph is equal
to 2 (the terminals are the blue vertices).
Let Γk (k ≥ 2) be the graph obtained from the (k × k)-grid by triangulating internal
faces of the (k × k)-grid such that all internal vertices become of degree 6, all non-corner
external vertices are of degree 4, and then one corner of degree two is joined by edges –
we call them external – with all vertices of the external face (the corners are the vertices
that in the underlying grid have degree two). Graph Γ6 is shown in Fig. 6.(i). We also
define the graph Γ∗k as the graph obtained from Γk if we remove all its external edges.
In Γ∗k we call all vertices incident to its unique non-triangle face perimetric.
Let G = (G0 , T ) be a rooted Σ0 -embedded graph. Given x ∈ V (G0 ) we define the
insulation between x and T , denoted insT (x) as the maximum length of a sequence of
S
T -concentric cycles C1 , . . . , Cl in G such that (i) {x} ∩ i∈{1,...,l} Ci = ∅ and (ii) every
line of Σ0 connecting x and some vertex of T meets every Ci , i ∈ {1, . . . , l} For an
example, see Figure 6.(ii). We define the planar thickness of a rooted graph G = (G0 , T )
as follows:
pthT (G) = max{insT (x) | x ∈ V (G0 )}.
It is easy to verify that pth is closed under contractions. In other words, the following
holds.
Lemma 7. Let G1 = (G01 , T1 ) and G2 = (G02 , T2 ) be two rooted graphs where G01 is a
σ-contraction of G02 and T1 = σ(T2 ). Then pthT1 (G1 ) ≤ pthT2 (G2 ).
According to the following lemma, if the treewidth of G is big enough then every
constant radius ball around T is sufficiently insulated from some vertex of G.
23
Lemma 8. Let G = (G0 , T ) be a rooted plane graph where tw(G0 ) ≥ 24 · (2l + 2r +
p
S
2)(d |T | + 1e) + 49. Let also T 0 = t∈T NGr 0 (t). Then pthT 0 (G) ≥ l.
Proof. By using [7, Lemma 6], the graph G0 contains as a σ-contraction the graph
H 0 = Γ(2l+2r+2)(d√|T |+1e) . Define the rooted graph H = (H 0 , TH ) where TH = σ(T ). We
consider a vertex packing of H 0 in |T | + 1 copies of Γ∗2l+2r+2 . By the pigeonhole principle
and the fact that |TH | ≤ |T |, one, say Z, of these copies does not contain any vertex in
r (T ).
TH . By Observation 1, contractions do not increase distances, hence σ(T 0 ) ⊆ NH
0
H
0
∗
Let Z be the subgraph of Z that is isomorphic to Γ2l+2 whose vertices have distance
at least r from the perimetric vertices of Z and therefore are also at distance strictly
greater than r from TH . We conclude that V (Z 0 ) ∩ σ(T 0 ) = ∅. Notice that Z 0 contains
l + 1 σ(T 0 )-concentric cycles C1 , . . . , Cl , Cl+1 that are also σ(T 0 )-concentric cycles of H.
Let x ∈ Cl+1 . Then the cycles C1 , . . . , Cl certify that insσ(T 0 ) (x) ≥ l. Therefore
pthσ(T 0 ) (H 0 ) ≥ l. By Lemma 7, we conclude that pthT 0 (G) ≥ l.
We need the following theorem from [17]. We present it using the terminology of [5].
Proposition 1. There is a computable function
Γ be a Σ0 -embedded plane graph, L be a linkage
in the pairs of π(L). Let also C1 , . . . , Cg(|π(L)|) be
is a linkage L0 with the same pattern as L such
∆ext (Cg(|π(L)|) ).
5.3
g such that the following holds: Let
of Γ and let T be the set of vertices
T -concentric cycles of Γ. Then there
that all paths in L are contained in
The algorithm
Using Lemmata 5, 7, 8, and Proposition 1, we can prove the following result.
Lemma 9. There is an FPTalgorithm, running in f2 (k) · n2 , for some function f2 ,
that given a τ = (J̃, H̃, T̃ ) ∈ H and a legal bijection ρ : T̃ → T, checks whether Rτ is
ρ-realizable in G.
Proof. As the number of bijections φ : {1, . . . , λ} → {1, . . . , λ} is bounded by a function
of k it is enough to show how to check in FPT time whether, for i = 1, . . . , λ, Rτi is a ρˆi rooted topological minor of G∗φ(i) were ρ̂i = ρi ∪ {(xinew , x∗i
new )}. To simplify notation, we
∗φ(i)
i
0i
i
i
drop indices and we denote Rτ = (Rτ , T̃ ∪ {xnew }) and G∗φ(i) = (G0φ(i) , T φ(i) ∪ {xnew })
by R = (R0 , T̃ ∪ {xnew }) and G∗ = (G0 , T ∪ {x∗new }) respectively. We also use ρ̂ instead
of ρ̂i and xnew instead of xinew .
We now apply the irrelevant vertex technique, introduced in [19] as follows. Using
the algorithm from [2] (or, alternatively, the one from [15]) one can either compute a
p
tree-decomposition of G0 of width at most q = 4 · (24 · (2 · g(|E(R0 )|) + 8)( |T | + 2) + 49)
or prove that no tree decomposition exists with width less than q/4. In the case where
24
tw(G0 ) ≤ q, we recall that |E(R0 )| is a function of k and |T | ≤ 2k. Consequently, there
exists a function f3 such that tw(G0 ) ≤ f3 (k) and the result follows from Lemma 10 in
the Appendix.
p
Suppose now that tw(G0 ) ≥ q/4 = 24 · (2 · g(|E(R0 )|) + 8)( |T | + 2) + 49. Applying
Lemma 8 for r = 3 we have that pthT 0 (G∗ ) ≥ g(|E(R0 )|) where T 0 = NG3 0 (T ∪ {x∗new }).
We now prove that there is a vertex x ∈ V (G∗ ) such that if R is a ρ̂-rooted topological
minor of G∗ then R is a ρ̂-rooted topological minor of G∗ \ {x}. Let ψ0 and ψ1 be the
functions certifying that R is a ρ̂-rooted topological minor of G∗ . We apply on ψ0 and
ψ1 the modifications of the Lemma 5 so that they satisfy Properties 1–4. An important
consequence is that the images of all vertices of R under ψ0 are close to terminals
in T. Indeed, from Properties 1 and 2, all neighbors of xnew are mapped via ψ0 to
vertices that belong in NG1 0 (x∗new ). Moreover, from Properties 3 and 4 it follows that
ψ0 (V (J̃)) ⊆ V (F) ⊆ NG2 0 (x∗new ). So far we have proved that all vertices of R, except
those inside J̃, are mapped via ψ0 to vertices in NG2 0 (x∗new ). As all vertices of R that are
inside J̃ belong to T̃ , it follows that ψ0 (V (R)) ⊆ T 0 .
The set L = {ψ1 (e) | e ∈ E(R)} is a set of paths in G∗ . Let L1 ⊆ L be the set of those
paths that have length at most 2 and define L2 = L \ L1 . For each path Q ∈ L2 define its
interior, denoted int(Q), as the subpath of Q consisting of all vertices of I(Q). Clearly,
L = {int(Q) | Q ∈ L2 } is a linkage in G∗ and π(L) ⊆ T 0 . Consider the collection C of
T 0 -concentric cycles C1 , . . . , Cg(|E(R0 )|) certifying the fact that pthT 0 (G∗ ) ≥ g(|E(R0 )|).
S
Define the graph Γ = (clos(∆in (C1 )) ∩ G∗ ) ∪
Q∈L2 int(Q) . We have that (i) L is a
linkage of Γ, and (ii) |E(R0 )| ≥ |π(L)|. Let x be a vertex of Cg(|π(L)|) . By Proposition 1
there is another linkage L0 with the same pattern as L such that all paths in L0 avoid x.
The vertices of all paths in L1 belong to T 0 since they are at distance at most 3 from
T ∪ x∗ new . The endpoints of all paths in L2 also belong T 0 since they are at distance
at most 2 from T ∪ x∗ new . Therefore all paths in L1 and the endpoints of all paths in
L2 avoid Γ altogether. We now show that all paths in L can be rerouted so that they
avoid x, while they remain internally vertex-disjoint. The paths in L1 stay the same.
For the paths in L2 , we only have to reroute their interiors within Γ. This is achieved by
connecting the pairs in π(L) via the linkage L0 . After this substitution all paths in the
updated set L avoid x. By updating ψ1 to reflect the new interiors of the paths in L2 ,
we obtain that R is a ρ̂-rooted topological minor of G∗ \ {x}.
Notice that x can be found in linear time applying BFS starting from T 0 . After deleting x, we create an equivalent instance of the problem with smaller size. By recursively
applying the same reduction to the new instance at most |V (G0 )| times, at some point
the treewidth will drop below q and then we solve the problem by applying Lemma 10
in the Appendix as above.
25
From Lemmata 6 and 9 we obtain Theorem 1.
6
Further extensions and open problems
We chose to tackle the disjoint-paths completion problem with the topological restriction
of having non-crossing patch edges. A natural extension of this problem is to allow a
fixed number ξ > 0 of crossings in the patch. We believe that, using the same techniques,
one may devise an f (k) · n2 algorithm for this problem as well. The only substantial
difference is a generalization of our combinatorial result (Theorem 2) under the presence
of crossings.
An interesting topic for future work is to define and solve the disjoint-paths completion problem for graphs embedded in surfaces of higher genus. A necessary step in
this direction is to extend Theorem 2 for the case where the face to be patched contains
handles.
Another issue is to extend the whole approach for the case where the faces to be
patched are more than one. This aim can be achieved without significant deviation from
our methodology, in case the number of these faces is bounded. However, when this
restriction does not apply, the problem seems challenging and, in our opinion, it is not
even clear whether it belongs to FPT.
Acknowledgement. We wish to thank the anonymous reviewers of an earlier version
of this paper for valuable comments and suggestions.
References
[1] I. Adler, S. G. Kolliopoulos, P. K. Krause, D. Lokshtanov, S. Saurabh,
and D. M. Thilikos, Tight bounds for linkages in planar graphs, in Proceedings
of the 38th International Colloquium on Automata, Languages and Programming
(ICALP 2011), 2011.
[2] E. Amir, Efficient approximation for triangulation of minimum treewidth, in Uncertainty in Artificial Intelligence: Proc. of the Seventeenth Conference (UAI-2001),
San Francisco, CA, Morgan Kaufmann Publishers, 2001, pp. 7–15.
[3] B. Courcelle, The monadic second-order logic of graphs. I. Recognizable sets of
finite graphs, Information and Computation, 85 (1990), pp. 12–75.
[4] B. Courcelle and M. Mosbah, Monadic second-order evaluations on treedecomposable graphs, Theoretical Computer Science, 109 (1993), pp. 49–82.
26
[5] A. Dawar, M. Grohe, and S. Kreutzer, Locally excluding a minor, in Proc.
of the 21st IEEE Symposium on Logic in Computer Science (LICS’07), New York,
2007, IEEE, pp. 270–279.
[6] R. Diestel, Graph theory, vol. 173 of Graduate Texts in Mathematics, SpringerVerlag, Berlin, third ed., 2005.
[7] F. V. Fomin, P. Golovach, and D. M. Thilikos, Contraction bidimensionality:
the accurate picture, in 17th Annual European Symposium on Algorithms (ESA
’09), Lecture Notes in Computer Science, Springer Verlag, 2009, pp. 706–717.
[8] M. R. Garey and D. S. Johnson, Computers and Intractability, A Guide to the
Theory of NP-Completeness, W.H. Freeman and Company, New York, 1979.
[9] M. C. Golumbic, H. Kaplan, and R. Shamir, On the complexity of DNA
physical mapping, Adv. in Appl. Math., 15 (1994), pp. 251–261.
[10] M. Grohe, K. Kawarabayashi, D. Marx, and P. Wollan, Finding topological subgraphs is fixed-parameter tractable, in 43rd ACM Symposium on Theory of
Computing (STOC 2011), San Jose, California, June 6-8 2011.
[11] P. Heggernes, C. Paul, J. A. Telle, and Y. Villanger, Interval completion
with few edges, in Proceedings of the 39th Annual ACM Symposium on the Theory
of Computing (STOC 2007), San Diego, California, American Mathematical Society,
2007, pp. 374–381.
[12] H. Kaplan, R. Shamir, and R. E. Tarjan, Tractability of parameterized completion problems on chordal, strongly chordal, and proper interval graphs, SIAM
Journal on Computing, 28 (1999), pp. 1906–1922.
[13] K. Kawarabayashi and P. Wollan, A shorter proof of the graph minor algorithm - the unique linkage theorem, in Proc. of the 42nd annual ACM Symposium
on Theory of Computing (STOC 2010), 2010.
[14] M. R. Kramer and J. van Leeuven, The complexity of wire-routing and finding
minimum area layouts for arbitrary VLSI circuits, Advances in Comp. Research, 2
(1984), pp. 129–146.
[15] B. A. Reed, Finding approximate separators and computing tree width quickly,
in Proc. of the Twenty-fourth annual ACM Symposium on Theory of Computing,
ACM Press, 1992, pp. 221–228.
27
[16] B. A. Reed, N. Robertson, A. Schrijver, and P. D. Seymour, Finding
disjoint trees in planar graphs in linear time, in Graph structure theory (Seattle,
WA, 1991), vol. 147 of Contemp. Math., Amer. Math. Soc., Providence, RI, 1993,
pp. 295–301.
[17] N. Robertson and P. Seymour, Graph minors. XXII. Irrelevant vertices in
linkage problems, preprint, 1992.
[18] N. Robertson and P. Seymour, Graph minors. XXI. Graphs with unique linkages, J. Combin. Theory Ser. B, 99 (2009), pp. 583–616.
[19] N. Robertson and P. D. Seymour, Graph minors. XIII. The disjoint paths
problem, J. Combin. Theory Ser. B, 63 (1995), pp. 65–110.
[20] M. Yannakakis, Computing the minimum fill-in is NP-complete, SIAM J. Algebraic Discrete Methods, 2 (1981), pp. 77–79.
28
Appendix
A.1 MSOL and rooted topological minors
In this section we define a parameterized version of Rooted Topological Minor Testing
that uses treewidth as a parameter and we show, using Monadic Second Order Logic
(msol), that it is in FPT. The corresponding algorithm will be used as a subroutine in
the proof of Theorem 1.
p-Bounded Treewidth Rooted Topological Minor Testing
Input: a rooted graph (H, SH ), a rooted graph (H, SG ), and a
bijection ρ : SH → SG .
Parameter: k = tw(G) + |V (H)|
Question: is (H, SH ) is a ρ-rooted topological minor of the rooted
graph (G, SG )?
Lemma 10. p-Bounded Treewidth Rooted Topological Minor Testing can
be solved in f (k) · |V (G)| steps.
Proof. Let SH = {a1 , . . . , as }, let V (H)\SH = {u1 , . . . , ut }, and let E(H) = {e1 , . . . , e }.
Let τs = {VERT, EDGE, I, c1 , . . . cs } be the vocabulary of graphs (as incidence structures) with s constants. We give an msol[τs ] formula φH,SH such that for every graph
G with SG = {b1 , . . . , bn } and ρ(ai ) = bi we have that (G, b1 , . . . , bs ) |= φH,SH ⇐⇒
(H, SH ) is a ρ-rooted topological minor of (G, SG ).
Let path(x, y, Z) be the msol[τs ] formula stating that Z is a path from x to y. (This
can be easily done by saying that Z is a set of edges with the property that for every
vertex v incident to an edge in Z, the vertex v is either incident to precisely two edges
in Z, or v is incident to one edge in Z and v = x or v = y. Finally, we can express that
the path Z is connected.)
29
Let
φH,SH :=∃Z1 . . . Z ∃x1 . . . xt
^
xi 6= xj ∧
^
path(xk , c` , Zi ) ∧
ei ={uk ,a` }
xk ∈V (H)\SH
a` ∈SH
^
path(xk , x` , Zi ) ∧
ei ={uk ,u` }⊆V (H)\SH
i6=j;i,j≤t
^
^
path(ck , c` , Zi ) ∧
ei ={ak ,a` }⊆SH
∃y (VERTy ∧ ‘y is incident to an edge in Zi
i6=j; i,j≤
and to an edge in Zj ’)
→(
s
_
i=1
y = ci ∨
t
_
y = xi ) .
i=1
The constants ci are interpreted by the bi , hence they make sure that ai is mapped to
bi , and Condition 1 of rooted topological minors is satisfied. In addition we make sure
that every edge of H is mapped to a path in G. Finally we make sure that Condition 3
is satisfied. (The statement ‘x is incident to an edge in Zi and to an edge in Zj ’ can be
easily expressed in msol.)
Observe that the length of the formula φH,SH only depends on H. Hence by Courcelle’s Theorem [3, 4] there is a computable function f1 such that p-Bounded Treewidth
Rooted Topological Minor Testing can be solved in time f1 (tw(G) + |V (H)|) ·
|V (G)|.
30
| 8cs.DS
|
Forking and JSJ decompositions in the free group II
Chloé Perin
Rizos Sklinos
arXiv:1801.07908v1 [math.GR] 24 Jan 2018
January 25, 2018
Abstract
We give a complete characterization of the forking independence relation over any set
of parameters in the free groups of finite rank, in terms of the JSJ decomposition relative
to those parameters.
1
Introduction
In [Sel13], Sela proves that the first order theory of torsion free hyperbolic groups, and thus
in particular that of free groups, is stable. This amazing result implies the existence of a
good notion of independence of tuples of elements over any set of parameters, akin to that
of algebraic independence in algebraically closed fields. It was thus natural to ask whether
there was a way to characterize this notion of independence (called forking independence) in
a purely group theoretic way.
In [PS16] we gave such a description for two families of parameter sets: when the parameter
set A is a free factor of the free group F, and at the other extreme, when A is not contained
in a proper free factor of F.
In the first case, we showed two tuples b and c are independent if and only if there is a free
product decomposition F = Fb ∗ A ∗ Fc such that b ∈ Fb ∗ A and c ∈ A ∗ Fc , in other words if
and only if the tuples live in "different parts" of the Grushko decomposition of F relative to
A. In the second case, we proved that two tuples b and c are independent if and only if they
live in "different parts" of the pointed cyclic JSJ decomposition of F relative to A. Note that
under the assumption that A is not contained in a proper free factor of F, there is a canonical
JSJ decomposition for F relative to A, namely the tree of cylinders of the deformation space.
The aim of the present paper is to complete this description for any set of parameters. In
this setting, there is no canonical JSJ decomposition - we thus give a condition in terms of
all the pointed cyclic JSJ decomposition for which the minimal subtree of the smallest free
factor FA is the tree of cylinders (we call such decomposition of F standard pointed cyclic
JSJ decompositions). Our main result is:
Theorem 1: Let A ⊂ F be a set of parameters and b, c be tuples from F. Then b is independent
from c over A if and only if there exists a standard cyclic JSJ decomposition Λ of F relative
min
to A in which the minimal subgraphs Λmin
Ab , ΛAc of hA, bi respectively hA, ci intersect at most
in a disjoint union of envelopes of rigid vertices.
(For a precise definition of minimal subgraphs and envelope of a rigid vertex, see Definition
3.4, for standard JSJ decompositions see Definition 3.12).
1
We also give in Theorem 3.15 an even more general characterization of forking independence, stated in terms of all pointed cyclic JSJ decompositions, not just standard ones (for
definitions of the various JSJ decompositions see Section 3).
Before Sela’s work proving the stability of torsion free hyperbolic groups, only the families of
abelian groups and algebraic groups (over algebraically closed fields) were known to be stable.
For these families it was fairly easy to understand and characterize forking independence.
But there is a qualitative difference between the already known examples of stable groups
and torsion-free hyperbolic groups: a group elementarily equivalent to an abelian group is an
abelian group and likewise a group elementarily equivalent to an algebraic group is an algbraic
group. This is certainly false in the case of torsion-free hyperbolic groups: an ultrapower of a
non abelian free group under a non-principal ultrafilter is not free. Even more, by definition,
a non finitely generated group cannot be hyperbolic.
The difficulty in describing forking independence arising from this easy observation is
significant. A priori, even for understanding the independence relation (provided by stability)
in natural models of our theory, one has to move to a saturated model of the theory. The
main problem with that is that we have very little knowledge of what a saturated model of the
theory of nonabelian free groups look like. In [PS16] we managed to overcome this difficulty
by using the assumptions we imposed on the parameter set A. When A is not contained in
any proper free factor, there are a number of useful results available. Model theoretically: F is
atomic over A (see [PS16], [OH11]). This enables us to transfer a sequence witnessing forking
from a big model to F. More importantly we have a natural candidate for a formula witnessing
forking: in an atomic model every type is isolated, and by homogeneity (see [PS12], [OH11]),
this means that the orbit of any tuple under AutA (F) is definable. Geometrically, under our
assumption, we have a very good understanding of AutA (F): the canonical JSJ decomposition
of F relative to A enables one to describe the restricted automorphisms up to a finite index,
i.e. one understands the modular automorpshism group M odA (F).
The paper is structured as follows: in Section2, we recall the essential facts and results
about stability and forking we use in the sequel. In Section 3, we give definitions and results
regarding actions on trees and JSJ decompositions. This enables us to state the two main
results, Theorem 3.14 and Theorem 3.15. At the beginning of Section 4, we show that the
conditions outlined in Theorem 3.15 are in fact equivalent to those of Theorem 3.14. We then
proceed to show one direction of Theorem 3.14, namely that if two tuples lie in essentially
disjoint parts of some standard pointed cyclic JSJ decomposition for F relative to A, they
are independent. Finally, in Section 5, we show that if in any standard pointed cyclic JSJ
decomposition of F relative to A, the minimal subgraphs of Ab and Ac intersect in more than
a disjoint union of envelopes of rigid vertices, then b and c are not independent over A.
2
Stability
A detailed account of the background on stability needed for this article has been already
developed in [PS16, Section 2], thus for avoiding repetition we only state here the notions and
results of fundamental importance to this article.
We fix a stable first-order theory T and we work in a “big” saturated model M of T , which
is usually called the monster model (see [Mar02, p.218]). Capital letters A, B, C, . . . denote
parameter sets of small cardinality, i.e. cardinality strictly less than |M|.
Definition 2.1: A formula φ(x̄, b) forks over A if there are n < ω and an infinite sequence
2
(bi )i<ω of tuples in M such that tp(b/A) = tp(b¯i /A) for i < ω, and the set {φ(x̄, bi ) : i < ω} is
n-inconsistent (i.e, any subset of n formulas is inconsistent).
| B) if there is no formula in tp(ā/B)
A tuple ā is independent from B over A (denoted ā ^
A
which forks over A.
Forking independence satisfies certain axioms, a full list of which we have recorded in
[PS16, Section 2]. We give next those important for this paper.
| C if and only if ā ^
| B and ā ^
| C;
Fact 2.2: Let A ⊆ B ⊆ C. Then ā ^
A
A
B
| B if and only if acl(āA) ^
| acl(B).
Fact 2.3: Let A ⊆ B. Then ā ^
A
acl(A)
In the above fact and for the rest of the article the algebraic closure acl is understood as
the imaginary algebraic closure acleq .
Definition 2.4: Let A ⊆ B be sets of parameter, and let p ∈ Sn (A). A non forking extension of p over B is a complete type q ∈ Sn (B) containing p such that no formula φ(x̄, b) ∈ q
forks over A. A type p over A is stationary if for any A ⊂ B, the type p admits a unique
non-forking extension over B.
Fact 2.5: Every type over an algebraically closed set A, i.e. A = acl(A), is stationary.
The following lemma has been proved in [PS16, Lemma 2.10].
Lemma 2.6: Let M |= T . Let b, A ⊂ M, and M be countable and atomic over A. Suppose
that the set X ⊂ M defined by a formula φ(x̄, b) contains a non empty almost A-invariant
subset (i.e. a subset that has finitely many images under Aut(M/A)). Then φ(x̄, b) does not
fork over A.
Fact 2.7: Let M ≺ M. Let X ⊂ M be a set definable over M. Then X ∩ M is definable over
M.
3
Trees
For this section we use the notations and framework of [GL17].
Let G be a finitely generated group. A G-tree is a simplicial tree T together with an action
of G on T . A cyclic G-tree is a G-tree whose edge stabilizers are cyclic (possibly finite or
trivial). We will only consider cyclic G-trees in this paper.
We say a G-tree T is minimal if it admits no proper G-invariant subtree. Note that unless
otherwise mentioned, the G-trees we will consider are not minimal.
If H is a finitely generated subgroup of G, it is a standard result that there is a unique
subtree of T preserved by H on which the action of H is minimal. We denote it by THmin .
Definition 3.1: Let T be a G-tree, with corresponding graph of groups Λ, and let H be a
finitely generated subgroup of G. The minimal subgraph Λmin
of H in Λ is the image of THmin
H
under the quotient map.
If A is a subset of G, a (G, A)-tree is a G-tree in which A fixes a point. A (not necessarily simplicial) surjective equivariant map d : T1 → T2 between two (G, A)-trees is called a
domination map. It sends edges to paths in T2 . We then say that T1 dominates T2 . This is
equivalent to saying that any subgroup of G which is elliptic in T1 is also elliptic in T2 .
3
A surjective simplicial map p : T1 → T2 which consists in collapsing some orbits of edges
to points is called a collapse map. In this case, we also say that T1 refines T2 .
3.1
Cylinders and envelopes
Definition 3.2: Let T be a G-tree with infinite cyclic edge stabilizers. A cylinder in T is an
equivalence class of edges under the equivalence relation given by e ∼ f iff Stab(e) and Stab(f )
are commensurable (see Example (3) of Definition 7.1 in [GL17]).
Note that cylinders are subtrees. The stabilizer of a cylinder is the commensurator of the
stabilizer of any one of its edges e, i.e. the set of elements g such that gStab(e)g −1 ∩ Stab(e)
has finite index both in Stabe and in gStab(e)g −1 .
Figure 1: Two cylinders, in blue and red, the cyclic vertices belong to the boundary of each
cylinder while the star vertices do not.
Definition 3.3: The boundary of a cylinder C is the set ∂C of vertices of C which either
are of valence 1 in T or belong to at least two cylinders.
Note that the stabilizer of a vertex v in the boundary ∂C of a cylinder C is infinite cyclic
if and only if v has valence 1 in T .
Given a G-tree T with infinite cyclic edge stabilizers, we can construct the associated tree
of cylinders Tc (Definition 7.2 in [GL17] or Definition 4.3 in [GL11]). It is obtained from T as
follows: the vertex set is the union V0 (Tc ) ∪ V1 (Tc ) where V0 (Tc ) contains a vertex w0 for each
vertex w of T contained in the boundary of some cylinder, and V1 (Tc ) contains a vertex vc for
each cylinder c of T . There is an edge between vertices w0 and vc lying in V0 (Tc ) and V1 (Tc )
respectively if and only if w belongs to the cylinder c.
We get a tree which is bipartite: every edge in the tree of cylinders joins a vertex from
V0 (Tc ) to a vertex of V1 (Tc ). Since the action of G on T sends cylinders to cylinders, the tree
of cylinder admits an obvious G action. The stabilizer of a vertex w0 ∈ V0 (Tc ) is the stabilizer
of the original vertex w, while the stabilizer of a vertex vC ∈ V1 (Tc ) is the setwise stabilizer
of the cylinder C. In a tree of cylinders Tc , cylinders are subtrees of diameter exactly 2 (star
graphs) whose center belongs to V1 (Tc ) and is not contained in any other cylinder.
For a general group G, the tree of cylinders need not have infinite cyclic edge stabilizers.
But if G is torsion-free hyperbolic for example (which will be our case in the sequel since we
deal exclusively with free groups), then the commensurator of an infinite cyclic group is itself
infinite cyclic, so the tree of cylinders does have infinite cyclic edge groups. In this case, by
[GL17, Lemma 7.3(5)], Tc is its own tree of cylinders, i.e. (Tc )c = Tc .
Definition 3.4: Let T be a G-tree with infinite cyclic edge stabilizers which is its own tree of
cylinders. Let v be a vertex in the boundary of some cylinder of T . An envelope of v is a
4
union of edges adjacent to v which meets at most one orbit of edge of each cylinder of T . Let
p : T → Λ be the quotient map to the associated graph of groups. The image of an envelope of
v is a union of edges adjacent to p(v), we call it an envelope of p(v) in Λ.
Remark 3.5: Note that the definition implies that an envelope of p(v) is a star graph, i.e. no
two edges in it have both endpoints in common.
Figure 2: An envelope of v in the tree of cylinders.
3.2
Pointed JSJ tree
We recall the following from Section 4.1 of [PS16].
Deformation space. The deformation space of a minimal cyclic (G, A)-tree T is the set of
all minimal cyclic (G, A)-trees T 0 such that T dominates T 0 and T 0 dominates T . The tree
of cylinders is an invariant of the deformation space [GL17, Lemma 7.3(3)] [GL11, Corollary
4.10].
A minimal cyclic (G, A)-tree is universally elliptic if its edge stabilizers are elliptic in every
minimal cyclic (G, A)-tree. If T is a universally elliptic cyclic (G, A)-tree, and T 0 is any minimal
cyclic (G, A)-tree, it is easy to see that there is a tree T̂ which refines T and dominates T 0 (see
[GL17, Lemma ]).
JSJ trees. A cyclic relative JSJ tree for G with respect to A is a minimal universally elliptic
cyclic (G, A)-tree which dominates any other minimal universally elliptic cyclic (G, A)-tree. All
these JSJ trees belong to a common deformation space, that we denote JSJA (G). Guirardel
and Levitt show that if G is finitely presented and A is finitely generated, the JSJ deformation
space always exists (see [GL17, Corollary 2.21]). It is easily seen to be unique.
Rigid and flexible vertices. A vertex stabilizer in a (relative) JSJ tree is said to be rigid
if it is elliptic in any cyclic (G, A)-tree, and flexible if not. In the case where G is a torsion-free
hyperbolic group and A is a finitely generated subgroup of G with respect to which G is freely
indecomposable, the flexible vertices of a cyclic JSJ tree of G with respect to A are surface
type vertices [GL17, Theorem 6.6], i.e. their stabilizers are fundamental groups of hyperbolic
5
surfaces with boundary, any adjacent edge group is contained in a maximal boundary subgroup,
and any maximal boudary subgroup contains either exactly one adjacent edge group, or exactly
one conjugate of A [GL17, Remark ].
We now restrict our attention to the case where G is torsion free hyperbolic.
Pointed trees of cylinders. We will define pointed JSJ trees - they are not minimal in
general.
Definition 3.6: Let G be a torsion free hyperbolic group and A a non trivial subgroup. We say
(T, vA ) is a pointed cyclic JSJ tree for G relative to A if one of the following alternatives
holds
(i) A is not infinite cyclic, the tree T is in JSJA (G), and vA is the unique vertex of T fixed
by A,
(ii) A is cyclic, the vertex vA has a unique adjacent edge e in T , the stabilizers of both vA
and e are equal to the centralizer C(A) of A, and removing the orbit of e from T gives a
G-tree which lies in JSJA (G).
When we consider the minimal subtree of a subgroup H containing A in such a pointed
tree, we require that vA belongs to THmin . A vertex of a pointed cyclic JSJ tree T for G relative
to A is called rigid if either it is the vertex vA , or it belongs to the subtree of T which lies in
JSJA (G) and is rigid in this subtree. A vertex is called flexible if it belongs to the subtree of
T which lies in JSJA (G) and is flexible in this subtree.
Definition 3.7: A vertex of a pointed cyclic JSJ tree (T, vA ) is said to be a Z-type vertex if
it is distinct from the base vertex vA and its stabilizer is infinite cyclic.
Assume G torsion-free hyperbolic and freely indecomposable with respect to A.
Then, there is a preferred tree in the cyclic JSJ deformation space JSJA (G) - indeed, in this
case the tree of cylinders of JSJA (G) itself lies in JSJA (G) ([GL11, Theorem 2] or [GL17,
Theorem 9.18]).
This enables us to define a preferred pointed cyclic JSJ tree for G relative to A.
Definition 3.8: Let G be a torsion free hyperbolic group which is freely indecomposable relative
to a finitely generated subgroup A. The pointed cyclic JSJ tree of cylinders for G relative
to A is the tree defined by:
(i) if A is not infinite cyclic, T is the tree of cylinders Tc of JSJA (G), and vA is the unique
vertex of Tc fixed by A;
(i) if A is cyclic, there is a unique edge e = (va , p) adjacent to vA , and removing its orbit
from T gives the tree of cylinder Tc of JSJA (G). Moreover, if C(A) stabilizes a cylinder
C in Tc , the vertex p is the center of C.
It will be useful to construct the tree of cylinders from any pointed JSJ tree by a refinement
and a collapse, as the following remark explains.
Remark 3.9: Let G be a torsion free hyperbolic group which is freely indecomposable relative
to a finitely generated subgroup A. If (T̃ , vA ) is any pointed cyclic JSJ tree for G with respect to
A, the pointed cyclic JSJ tree of cylinders (T, vA ) can be obtained from (T̃ , ṽA ) by the following
procedure:
6
(i) we build a refinement T 0 of T̃ by replacing each non Z-type vertex v of T̃ (i.e. v is in
the orbit of ṽA or belongs to at least two cylinders of T̃ ), by the star graph with a central
vertex v 0 and an edge (v 0 , vC ) for each cylinder C containing v. If e was an edge which
was adjacent to v, we attach it to vC where C is the cylinder containing e. Denote the
collapse map T 0 → T̃ by p.
(ii) we build the collapse map q : T 0 → T by collapsing all the edges of T 0 which do not come
from the refinement dual to p.
Assume now that G is torsion free hyperbolic but not necessarily freely indecomposable relative to A. We may consider the Grushko decomposition G = GA ∗ G1 ∗ . . . ∗
Gk ∗ Fl for G relative to A, where GA is the smallest free factor of G containing A and Fl is a
free group of rank l.
Lemma 3.10: If T ∈ JSJA (G), the minimal subtree TGmin
of GA in T is in the cyclic JSJ
A
deformation space for GA relative to A, and the minimal subtree TGmin
of each factor Gi for
i
i = 1, . . . , k is in the cyclic JSJ deformation space for Gi .
Proof. Suppose T ∈ JSJA (G). Let E1 (T ) be the union of interior of trivially stabilized edges
of T . Let T0 , T1 , . . . , Tr be representatives of the orbits of connected components of T − E1 (T )
where vA ∈ T0 . Denote by Hi the stabilizer of Ti . The group G admits a decomposition as
H0 ∗ . . . ∗ Hr ∗ Fs for some s, and A ≤ G0 .
Note that any refinement Ti0 of one of the trees Ti can be extended to a refinement T 0 of
T , since the only edges of T − Ti which are adjacent to Ti are trivially stabilized.
Show first that Ti is universally elliptic among cyclic Hi -trees (respectively (H0 , A)-tree if
i = 0). If T̃i is another cyclic Hi -tree (respectively (H0 , A)-tree if i = 0), we extend it to a
cyclic (G, A)-tree T̃ by adding to the graph of group Γ̃i corresponding to T̃i a set of k trivially
stabilized edges with one endpoint in Γ̃i and the other endpoint with vertex group Hj for j 6= i,
and s trivially stabilized loop attached to some vertex of Γi . Since T is universally elliptic for
(G, A)-trees, the stabilizer of an edge in Ti fixes a vertex of T̃ . Since all the edges of T̃ which
are not in the orbit of T̃i are trivially stabilized, and Stab(e) ≤ Hi = Stab(T̃i ), it must in fact
fix a vertex in T̃i .
On the other hand, each Hi is freely indecomposable: suppose by contradiction that Hi acts
minimally and non trivially on a tree Ti∗ with trivial edge stabilizers. Since Ti is universally
elliptic, there is a refinement Ti0 of Ti which dominates Ti∗ , obtained by replacing vertices
in Ti by the minimal subtree of their stabilizer in Ti∗ (in particular, we only add trivially
stabilized edges). Note that Ti0 must have some trivially stabilized edges, in particular Ti does
not dominate Ti0 - there must be a vertex v in Ti whose stabilizer Stab(v) is not elliptic in Ti0 .
Now Ti0 can be extended to a proper refinement T 0 of T . Note that to get T 0 from T , we only
added trivially stabilized edges, hence T 0 is still universally elliptic. But T cannot dominate
T 0 since Stab(v) is elliptic in T and not in T - this contradicts the maximality of T .
Therefore the decomposition G = H0 ∗ . . . ∗ Hr ∗ Fs of G is in fact a Grushko decomposition
for G, in particular H0 = GA and TGmin
⊆ T0 , s = l, r = k, and we may assume Hi = Gi for
A
i = 1, . . . , k.
We now prove that the minimal subtree of Gi = Hi in Ti is a cyclic JSJ tree for Gi (relative
to A if i = 0). We already know it is universally elliptic since Ti is. By Proposition 4.15 in
[GL17], it is enough to check that the JSJ of any vertex stabilizer in Ti relative to the adjacent
edge groups is trivial. Suppose by contradiction that some vertex v ∈ Ti is such that Stab(v)
7
admits a non trivial JSJ tree Tv relative to the adjacent edge groups. Note that Stab(v) is not
elliptic in Tv . By Proposition 4.15 of [GL17] refining T at v by Tv gives a tree which should
lie in the cyclic JSJ deformation space of G relative to A - a contradiction since in this refined
tree, Stab(v) is no longer elliptic.
Corollary 3.11: If (T, vA ) is a pointed cyclic JSJ tree for G relative to A, and TGmin
is the
A
min
minimal subtree of GA in T , then (TGA , vA ) is a pointed cyclic JSJ tree for GA relative to A.
Proof. If A is not cyclic, pointed JSJ tree is just a regular JSJ tree with a specified vertex, so
this follows immediately from the previous lemma. Suppose thus that A is cyclic. By definition
of pointed JSJ decomposition there is a valence one edge e = (vA , p) in T such that the tree
T 0 obtained from T by removing the orbit of e belongs to JSJA (G).
We now prove that the minimal subtree of GA in T 0 must contain the vertex p. Note that
vertex stabilizers of T belong to conjugates of factors of the Grushko decomposition of G.
Thus for any vertex v, if the stabilizer of v contains a non trivial subgroup of A, it is contained
in GA and thus it stabilizes the path between v and (T 0 )min
GA , in particular if it is not cyclic v
0
min
is in fact contained in (T )GA . If A stabilizes only the vertex p in T 0 , this vertex has noncyclic
stabilizer and must be in (T 0 )min
GA . If A stabilizes a cylinder C, then all the vertices other than
vA in the boundary of C have non cyclic stabilizers and contain a non trivial subgroup of A,
0
thus they all lie in (T 0 )min
GA . By minimality of T , any vertex of the cylinder C lies on a path
joining two vertices of ∂C - this implies that p lies in (T 0 )min
GA .
Hence the minimal subtree of GA in the pointed tree (T, vA ) is exactly the subtree obtained
0 min
from (T 0 )min
GA by adding the orbit of e under GA . By the previous lemma, (T )GA lies in
JSJA (GA ). Hence the minimal subtree of GA in the pointed tree (T, vA ) is the pointed JSJ
tree of GA relative to A.
Standard JSJ decompositions. We finally define standard JSJ decompositions appearing
in Theorem 3.14:
Definition 3.12: Let G be torsion free hyperbolic, let A be a subset of G and let G = GA ∗
G1 ∗ . . . ∗ Gk ∗ Fl be the standard Grushko decomposition of G relative to A. A pointed cyclic
JSJ tree T relative to A is said to be standard if the minimal subtree TGmin
of GA in T is
A
the pointed cyclic JSJ tree of cylinders of GA relative to A, and for each i the minimal subtree
TGmin
of Gi in T is the cyclic JSJ tree of cylinders of Gi .
i
The decomposition Λ associated to T is called a standard cyclic JSJ decomposition for G
relative to A.
We now want to generalize the notion of envelopes in the standard JSJ trees, which have
possibly trivial edge stabilizers (recall cylinders and envelopes were defined for G-trees with
infinite cyclic stabilizers).
Definition 3.13: If G, A and T are as in the definition above, a subtree E of T is an envelope
of a vertex v of T if v ∈ E and there exists and element g of G such that g · E is an envelope
of g · v in one of the trees TGmin
or TGmin
.
i
A
We can now state the main results:
Theorem 3.14: Let A ⊂ F be a set of parameters and b, c be tuples from F. Then b is
independent from c over A if and only if there exists a standard cyclic JSJ decomposition
8
min
Λ of F relative to A in which the minimal subgraphs Λmin
Ab , ΛAc of hA, bi, hA, ci respectively
intersect at most in a disjoint union of envelopes of rigid vertices.
The following is a more general, if slightly more cumbersome, equivalent condition for
independence:
Theorem 3.15: Let A ⊂ F be a set of parameters and b, c be tuples from F. Then b is
independent from c over A if and only if there exists a pointed cyclic JSJ tree T of F relative
to A in which
min and in
1. no surface type vertex or trivially stabilized edges lies both in a translate of TAb
min
a translate of TAc ,
2. if some cylinder C of T contains boundary vertices v1 , v2 such that for each i, there are
min , T min containing an edge of C adjacent to v , then v , v are
translates of each of TAb
i
1
2
Ac
in the same orbit under Stab(C) .
Remark 3.16: Suppose that the conditions of Theorem 3.14 hold, i.e. there exists a standard
min
JSJ decomposition Λ for F relative to A in which the minimal subgraphs Λmin
Ab and ΛAc
intersect at most in a disjoint union of envelopes of rigid vertices. Then the tree T associated
to Λ satisfies condition 1. of Theorem 3.15. To see that the second condition also holds, note
is a pointed tree of cylinders, any boundary vertex of a cylinder is adjacent to
that since TFmin
A
exactly one edge of the cylinder, and if two boundary vertices v1 , v2 of a cylinder C are in the
same orbit under Stab(C), the corresponding edges of C adjacent to v1 , v2 respectively are also
in the same orbit.
In Proposition 4.2, we will show that the converse holds, namely that if the conditions of
Theorem 3.15 are satisfied then the conditions of Theorem 3.14 are satisfied.
4
Proving independence
We start by showing that if there exists a pointed cyclic JSJ tree which satisfies the conditions
of Theorem 3.15, then there exists a standard JSJ tree which satisfies the conditions of Theorem
3.14, plus some additional criterions.
4.1
Normalizing the tree
The following proposition shows that in a JSJ tree for F relative to A, we can assume that
every trivially stabilized edge is adjacent to a translate of vA without changing the intersection
pattern of the minimal subtrees of Ab and Ac.
Proposition 4.1: Let b, c be tuples in a free group F, and let A ⊆ F. Denote by FA the
smallest free factor containing A. Suppose there exists a pointed JSJ tree (T̂ , v̂A ) for F relative
min and T̂ min contains
to A in which no intersection of translates of the minimal subtrees T̂Ab
Ac
trivially stabilized edges. Then there exists a pointed JSJ tree (T, vA ) for F relative to A such
that
1. TFmin
is isomorphic to T̂Fmin
,
A
A
min and T min is the image by
2. the intersection of any translates of the minimal subtrees TAb
Ac
min and T̂ min ;
this isomorphism of the intersection of the corresponding translates of T̂Ab
Ac
9
3. any trivially stabilized edge joins a translate of the distinguished vertex vA to a vertex
which lies in a translate of TFmin
.
A
Proof. We assume without loss of generality that T̂ has no valence 2 vertices adjacent to two
trivially stabilized edges. Note that a trivially stabilized edge of T̂ can belong to a translate
min , or to a translate of T̂ min , but not to both.
of T̂Ab
Ac
We build iteratively a new tree in JSJA (F): set T̂ (0) = T̂ . Suppose we have built the
tree T̂ (k), and suppose there is in T̂ (k) an orbit of trivially stabilized edges both of whose
endpoints are not in the orbit of vA . We build T̂ (k + 1) as follows:
1.
(i) if the orbit does not meet any of the minimal subtrees of Ab and Ac in T̂ (k), we pick
an edge e = (p, q) in this orbit minimizing the length m of a path p, p1 , . . . , pm = vA
to the distinguished vertex vA ;
min (k) of Ab (respectively the minimal
(ii) if the orbit meets the minimal subtree T̂Ab
min (k) of Ac) in T̂ (k), pick an edge e = (p, q) in this orbit which minimizes
subtree T̂Ac
min (k) (respectively in T̂ min (k)) to
the length m of a path p, p1 , . . . , pm = vA in T̂Ab
Ac
the distinguished vertex vA ;
2. subdivide e in m + 1 sub-edges by vertices p, p01 , p02 , . . . , p0m , p0m+1 = q;
3. fold the first m edges thus created along the path between p and vA , that is, identify
(p, p01 ) with (p, p1 ), (p01 , p02 ) with (p1 , p2 ), until (p0m−1 , p0m ) is identified with (pm−1 , pm );
4. subdivide and fold equivariantly all the edges in the orbit of e;
5. if two trivially stabilized edges in the tree thus created are adjacent to a common valence
2 vertex, merge them into one by erasing the vertex.
Note that T̂ (k + 1) is in the same deformation space as T̂ (k), that it has at most as many
orbits of trivially stabilized edges, and that it has strictly more orbits of trivially stabilized
edges one of whose endpoints is in the orbit of vA . Note moreover that the intersection of
min (k + 1) and T̂ min (k + 1) is isomorphic to
translates in T̂ (k + 1) of the minimal subtrees T̂Ab
Ac
the intersection of the corresponding translates of minimal subtrees in T̂ (k) : this is because
whenever we fold an edge of one of these two minimal subtrees, we fold it onto a path in the
same subtree so no new intersection is created.
Thus eventually we get a tree T in which all trivially stabilized edges have at least one
endpoint in the orbit of vA . Suppose the other endpoint v of one of these edges e does not
lie in a translate of TFmin
- we can collapse e equivariantly and the resulting tree still satisfies
A
1 − 2 and every trivially stabilized edge is adjacent to a vertex in the orbit of vA (recall that
min , T min ). If there remain edges of this type in the
e lies in a translate of at most one of TAb
Ac
new tree, repeat. Eventually we get a tree in which 1 − 3 hold.
Now we show that if there exists a JSJ tree T in which the conditions of Theorem 3.15
hold, there is a standard JSJ tree in which the conditions for Theorem 3.14 hold.
Proposition 4.2: Let b, c be tuples in a free group F, and let A ⊆ F. Denote by FA the
minimal free factor containing A. Suppose there exists a pointed cyclic JSJ tree (T̂ , v̂A ) for
F relative to A in which the conditions of Theorem 3.15 hold. Then there exists a pointed
standard JSJ tree (T, vA ) for F relative to A with corresponding graph of groups Λ such that
10
1. any trivially stabilized edge joins a translate of vA to a vertex which lies in a translate of
TFmin
;
A
min
2. the minimal subgraphs Λmin
Ab and ΛAc intersect at most in a disjoint union of envelopes
of rigid vertices.
For this, the following lemma will be useful
Lemma 4.3: Let FA be a free group which is freely indecomposable with respect to a subgroup
A. Suppose (T̃ , vA ) is any pointed JSJ tree for FA relative to A, and (T, vA ) is the pointed JSJ
tree of cylinders for FA relative to A.
To each cylinder C of T , we can associate a cylinder C̃ of T̃ such that the map C 7→ C̃
is one-to-one and equivariant. Moreover, to each boundary vertex v of C we can associate
a boundary vertex ṽ of C̃ of the same type (rigid or surface) such that the map v 7→ ṽ is
equivariant, and if g ∈ FA is such that [vA , g · vA ] contains
• the vertex v, then [ṽA , g · ṽA ] contains ṽ;
• the edge e of C adjacent to v, then [ṽA , g · ṽA ] contains ṽ and some edge of C̃ adjacent
to ṽ.
Proof. Let T 0 , p : T 0 → T̃ and q : T 0 → T be as in Remark 3.9.
Let C be a cylinder in T : if e is an edge of C, Stab(C) is the centralizer of Stab(e). Now
there exists a unique edge e0 ∈ T 0 such that q(e0 ) = e, and Stab(e0 ) = Stab(e). Let C 0 be the
cylinder of e0 : then Stab(C 0 ) is the centralizer of Stab(e0 ) so Stab(C 0 ) = Stab(C). Note that
q(C 0 ) = C. We set C̃ = p(C 0 ) (the image of a cylinder by a collapse map is a cylinder). In
particular we have Stab(C) = Stab(C̃).
Now let v be a boundary vertex of C - in other words, it is a non Z-type vertex of C.
Note that there is a unique preimage v 0 of v by q, since q collapses only edges both of whose
endpoints are Z-type vertices. Define ṽ = p(v 0 ), it is unique by construction. Note that
Stab(v) = Stab(v 0 ) = Stab(ṽ), and that adjacent edge groups of ṽ and adjacent edge groups
of v are commensurable. In particular this implies that ṽ has the same type as v (rigid if there
is no non trivial splitting relative to adjacent edge groups, flexible(surface) otherwise).
Let us prove injectivity: suppose (C̃1 , ṽ1 ) = (C̃2 , ṽ2 ). Then we must have Stab(C1 ) =
Stab(C2 ) so C1 = C2 . We have thus that v10 and v20 belong to the same cylinder. Now
p(v10 ) = p(v20 ), but p does not identify distinct non Z-vertices of a same cylinder hence v10 = v20 .
Thus v1 = q(v10 ) = q(v20 ) = v2 .
^
Equivariance: note that Stab(C) = Stab(C̃) and Stab(v) = Stab(ṽ). Thus Stab((g
· C)) =
Stab(g · C) = gStab(C)g −1 = Stab(g · C̃). Since cylinders are uniquely defined by their
^
stabilizers, we must have (g
· C) = g · C̃. Similarly, non Z-type vertices are uniquely defined
by their stabilizers, so we get that gg
· v = g · ṽ.
Now to see the last property: suppose the boundary vertex v of C lies on [vA , g · vA ]. Note
0
that v, vA are non Z-type vertex, so they have unique preimages by q, denote them by v 0 , vA
0
0
respectively. Now [ṽA , g · ṽA ] = p([vA , g · vA ]) (this is a general property of collapse maps), and
0 , g · v 0 ]). Therefore the unique preimage v 0 of v by q must lie on [v 0 , g · v 0 ].
[vA , g · vA ] = q([vA
A
A
A
But then ṽ = p(v 0 ) lies on [ṽA , g · ṽA ] as requested.
Suppose that the edge e of C adjacent to v lies on [vA , g · vA ]. The unique preimage
0
0 , g · v 0 ], and w 0 is a Z-type vertex so it is distinct from
e = (v 0 , w0 ) of e by q must lie on [vA
A
0
0
0 , g · v 0 ] contains an edge e00 which is adjacent to w 0
both vA and g · vA . Thus the path [vA
A
11
but distinct from e0 - such an edge is not collapsed by p, and its image p(e00 ) lies in C̃ and in
0 , g · v 0 ]) = [ṽ , g · ṽ ]. This finishes the proof of the claim.
p([vA
A
A
A
Proof. (of Proposition 4.2) By applying Proposition 4.1 we get a tree (T̃ , ṽA ) which satisfies
the first condition.
By Remark 3.9, we can replace translates of T̃Fmin
by copies of the pointed JSJ tree of
A
cylinders for FA relative to A by the following construction. First, we build a refinement T 0
of T̃ as follows: for each vertex ṽ of T̃Fmin
which belongs to at least two cylinders of T̃Fmin
or
A
A
0
0
is equal to vA , we replace ṽ by a star graph with central vertex v and an edge (v , vC̃ ) for
each cylinder C̃ of T̃Fmin
containing ṽ. If an edge e of T̃Fmin
was attached to ṽ, se attach it
A
A
to vC̃ , where C̃ is the cylinder containing e. If an edge e of T̃ − T̃Fmin
was attached to ṽ, we
A
0
attach it to v . We extend this refinement equivariantly to the orbits of T̃Fmin
. Call the p the
A
0
0
collapse map T → T̃ . We then build a collapse map q : T → T by collapsing all the edges of
min
translates (T 0 )min
FA which do not come from the refinement of T̃FA .
The tree T clearly is a standard JSJ decomposition for F relative to A. It is not hard to
see that a trivially stabilized edge attached to vA gives a trivially stabilized edge attached to
vA , so the second condition also holds.
Let us now see that the tree T satisfies the last condition. Suppose that the minimal
min
subgraphs Λmin
Ab and ΛAc intersect in more than a disjoint union of envelopes of rigid vertices
- in other words, the intersection contains either (i) a trivially stabilized edge, or (ii) a surface
type vertex, or (iii) two distinct edges which belong to a same cylinder of Λ.
min
Recall that the minimal subgraphs Λmin
Ab and ΛAc are the images by the quotient map
min and T min . These trees are unions of translates of paths
T → Λ of the minimal subtrees TAb
Ac
of the form [vA , γ · · · vA ] and [vA , β · · · vA ] by elements of hA, bi and hA, ci respectively where
β ∈ hA, bi and γ ∈ hA, ci.
Case (i) cannot occur since translates of the minimal subtrees of hA, bi and hA, ci in T̃
never intersect in a trivially stabilized edge.
In case (ii), there exist β ∈ hA, bi and γ ∈ hA, bi and g ∈ F such that the intersection
[vA , β · vA ] ∩ g[vA , γ · vA ] contains a surface type vertex v. By Lemma 4.3, the intersection
[ṽA , β · ṽA ] ∩ g[ṽA , γ · ṽA ] contains a vertex ṽ which is also of surface type - a contradiction.
In case (iii), denote by e1 , e2 edges of a common cylinder C of T whose images in Λ are
min
distinct, and lie in the intersection Λmin
Ab ∩ΛAc . Then there exist β1 , β2 ∈ hA, bi; γ1 , γ2 ∈ hA, ci
and g1 , g2 , h1 , h2 ∈ F such that ei lies in gi · [vA , βi · vA ] ∩ hi [vA , γi · vA ] for i = 1, 2. Denote by
v1 , v2 the endpoints of e1 , e2 respectively.
We know by Lemma 4.3 that there exists a cylinder C̃ and boundary vertices ṽ1 , ṽ2 of C̃
such that each of gi · [ṽA , β · ṽA ] and hi · [ṽA , γ · ṽA ] contains ṽi and some edge of C̃ adjacent
to ṽi . Note that since e1 , e2 were not in the same orbit, the vertices ṽ1 and ṽ2 are not in the
same orbit under Stab(C̃). This is in contradiction with the hypotheses of Theorem 3.15.
From Proposition 4.2 and Remark 3.16 we deduce
Corollary 4.4: Let b, c be tuples in a free group F, and let A ⊆ F. Denote by FA the minimal
free factor containing A. There exists a pointed cyclic JSJ tree (T̂ , v̂A ) for F relative to A in
which the conditions of Theorem 3.15 hold if and only if there exists a pointed standard JSJ
min
decomposition Λ for F relative to A in which the minimal subgraphs Λmin
Ab and ΛAc intersect
at most in a disjoint union of envelopes of rigid vertices.
12
In other words, the conditions of Theorem 3.14 are satisfied if and only if the conditions
of Theorem 3.15 are satisfied. Therefore, there remains only to prove Theorem 3.14.
4.2
A special case
We fix a nonabelian finite rank free group F. Let A ⊂ F be a set of parameters and b, c be
tuples coming from F. In this section we will prove one direction of Theorem 3.14 under the
assumption that at least one of the tuples, say b, lives in the smallest free factor FA of F
containing A.
In this case, we may assume that F is freely indecomposable relative to Ac, thus by Theorem
5.3 of [PS16] the orbit X := Aut(F/Ac).b is definable over Ac. But X ⊂ FA , thus by Fact 2.7
the set X is definable over FA .
We first show:
Proposition 4.5: Let b, c be tuples in a free group F, let A ⊆ F, and denote by FA the minimal
free factor containing A. Assume that F is freely indecomposable with respect to Ac, and that
b ∈ FA .
Suppose that there exists a standard pointed cyclic JSJ decomposition Λ for F relative to
min
A in which the minimal subgraphs Λmin
Ab and ΛAc intersect at most in a disjoint union of
envelopes of rigid vertices.
Then the orbit of b under AutAc (F) contains the orbit of b under ModA (FA ).
Proof. We can assume that T satisfies the conclusions of Proposition 4.2. Since T is standard,
is the cyclic JSJ tree of cylinders for FA relative to A.
the minimal subtree TFmin
A
By Lemmata 4.21 and 4.22 of [PS16], we can thus write any θ ∈ ModA (FA ) as
θ = Conj (z) ◦ ρ1 ◦ . . . ◦ ρr
where each ρi is a Dehn twist or a surface vertex automorphism supported on an edge or a
. We want to show that there exists α ∈ AutAc (F)
surface vertex of the pointed JSJ tree TFmin
A
such that θ(b) = α(b).
Lemma 4.21 tells us that up to changing z we can permute the list of support of the ρj min and ρ
in particular we may assume that ρ1 , . . . , ρk all have support in TAb
k+1 , . . . , ρr all have
min
support outside of TAb .
S
min ∩
The hypothesis on the minimal subgraphs for Ab and Ac ensures that g∈F g · TAb
S
min
g∈F g · TAc does not contain any surface type vertex or trivially stabilized edges, and does
not contain two edges of a common cylinder which lie in different orbits. Thus if for some i,
ρi is a Dehn twist about an edge which lies in this intersection, we can apply Lemma 4.9 of
[PS16] to replace it by a product of Dehn twist about edges of the same cylinder which are
representatives of the other orbits of edges, and therefore must all lie outside the intersection.
As a result, we may assume that none of the ρi ’s have support which lies both in a translate
min and a translate of T min .
of TAb
Ac
But now by Lemma 4.26 in [PS16], the image of b by θ is, up to conjugation, the same as
the image of b by ρ1 ◦. . . ◦ρk since ρk+1 ◦. . .◦ρr is the composition of an element of ModAb (FA )
min , and
with a conjugation. Thus we may in fact assume that all the ρj have support in TAb
min
thus outside of any translate of TAc .
Now we want to extend θ to an automorphism α : F → F fixing Ac.
First, we choose an adequate presentation of F: denote by p : T → T 0 the map defined by
min , and write Λ0 for the graph
collapsing all the edges of TFmin
which are not in a translate of TAb
A
13
of groups associated to T 0 . Pick a maximal subtree of Λ0 which does not include any trivially
stabilized edges, lift it to a subtree T 0 of T 0 , and extend this to a Bass-Serre presentation
(T 0 , T 1 , {te }e∈E(T 1 −T 0 ) ) for Λ0 . Note that F = FA ∗ h{te }e∈Etriv+ | i where Etriv+ consists of
all the trivially stabilized edges of T 1 of the form (vA , w) (to fix an orientation). Let v 1 , . . . , v s
min or whose inverse image by p is
be the vertices of T 0 which come from rigid vertices of TAb
not reduced to a point, and let H 1 , . . . , H s be the stabilizers of v 1 , . . . , v s .
Then θ restricts to a conjugation by some element γi on each H i and to the identity
on Stab(vA ) in T 0 . Indeed, the elementary automorphisms ρj are supported on edges or on
min . By Lemma 4.24 of [PS16], the ρ are also elementary
surface type vertices which lie in TAb
j
0
automorphisms associated to T , and moreover they are not supported on translates of any of
the vertices v 1 , . . . , v s .
Define now Ĥ i to be the subgroup generated by H i and all the Z-vertex groups adjacent
to H i by a non trivially stabilized edge (so Ĥ i is generated by H i together with some roots of
elements of H i ). It is easy to see that θ also restricts to a conjugation on Ĥ i .
min ), we see that it intersects
Now since the minimal subtree of Ac in T 0 is contained in p(TAc
the union of translates of (T 0 )min
FA at most in an envelope of a rigid vertex. As a consequence,
each element in the tuple c can be written as a product of the form
−1
−1
g1 (te1 h1 t−1
f1 )g2 (te2 h2 tf2 ) . . . gm (tem hm tfm )gm+1 (∗)
where for each j,
1. gj ∈ Stab(vA ) in T 0 ;
2. there exists an index lj such that ej , fj are trivially stabilized edges, each joining vA to
a vertex in the orbit of v lj ;
3. hj ∈ Ĥ lj .
We now define α ∈ AutA (F) by setting α |FA = θ, and for each i and each edge e ∈ Etriv+
joining vA to a vertex in the orbit corresponding to v i , we set α(te ) = te γi−1 (recall γi is the
element of FA such that θ |H i = Conj (γi )). As we saw, F is the free product of FA together with
the free group generated by the elements {te }e∈Etriv+ so this indeed defines an automorphism
of F. But now it is easy to check using (∗) that α(c) = c, hence α ∈ AutAc (F).
We can now prove:
Proposition 4.6: Let b, c be tuples in a free group F, let A ⊆ F, and denote by FA the minimal
free factor containing A. Assume that F is freely indecomposable with respect to Ac, and that
b ∈ FA .
Suppose that there exists a standard pointed cyclic JSJ decomposition Λ for F relative to
min
A in which the minimal subgraphs Λmin
Ab and ΛAc intersect in a disjoint union of envelopes of
rigid vertices.
Then b is independent from c over A.
Proof. We assume, for the sake of contradiction, that tp(b/Ac) forks over A.
By Theorem 5.3 of [PS16], since F is freely indecomposable with respect to Ac, the orbit
X = AutAc (F) · b of b under automorphisms of F fixing Ac is definable over Ac. Since it is
contained in any set definable over Ac which contains b, we must have that X forks over A.
By Proposition 4.5, the set X contains ModA (FA ) · b. Now ModA (FA ) has finite index
in AutA (FA ) (see [RS94, Theorem 4.4]) so ModA (FA ) · b is a non trivial almost A-invariant
subset. By Lemma 2.6 it cannot fork over A, a contradiction.
14
4.3
Proof of independence in the general case
We want to prove the first direction of Theorem 3.14, namely
Theorem 4.7: Let b, c be tuples in a free group F, and let A ⊆ F. Denote by FA the minimal
free factor containing A. Suppose there exists a standard pointed JSJ decomposition Λ for F
min
relative to A in which the minimal subgraphs Λmin
Ab and ΛAc intersect at most in an envelope
| A c.
of a rigid vertex. Then b ^
Note that by Proposition 4.1, we can assume that T is a standard pointed cyclic JSJ tree
in which any trivially stabilized edge joins a translate of vA to a translate of TFmin
.
A
We first prove the result for particularly simple elements b, c satisfying the following
Definition 4.8: Let T be standard pointed cyclic JSJ tree for F with respect to A in which any
trivially stabilized edge joins a translate of vA to a translate of TFmin
. We say that an element
A
β of F is a sandwich term if the path [vA , β · vA ] does not contain any translates of vA other
than its endpoint, and can be subdivided into three subpaths [vA , u] ∪ [u, v] ∪ [v, g · vA ] where
[vA , u] and [v, g · vA ] are either empty or consist of a single trivially stabilized edge, and [u, v]
lies entirely in a translate of TFmin
.
A
The image of the path [u, v] in ΛFA is called the imprint of the sandwich term β in ΛFA
and denoted by imp(β).
The trivially stabilized edges crossed by [vA , β · vA ] (by extension, their images in Λ) are
called the trivially stabilized edges of the sandwich term β.
For sandwich terms, Theorem 4.7 holds.
min
Proposition 4.9: If β, γ are sandwich terms such that the minimal subgraphs Λmin
Aβ and ΛAγ
intersect at most in an envelope of a rigid vertex, then β and γ are independent over A.
We will in fact prove a more general version of this Proposition, for some specific tuples of
sandwich terms.
Proposition 4.10: Let T be a JSJ tree for F with respect to A which satisfy the conclusions of
Proposition 4.2. Suppose β̄ = (β 1 , . . . , β q ) and γ̄ = (γ 1 , . . . , γ r ) are tuples of sandwich terms.
Let ∆ be a connected subgraph of groups of ΛFA which contains the vertex vA . Assume that
• the trivially stabilized edges associated to elements β j , γ k of β̄, γ̄ respectively are distinct;
• the imprint of any element β j of β̄ lies in ∆;
• the intersection of ∆ with ∪rk=1 imp(γk ) is a disjoint union of envelopes of rigid vertices.
Then the tuples β̄ and γ̄ are independent over A.
Proof. Pick a maximal subtree of ΛFA which extends a maximal subtree of ∆. Lift it to a subtree T 0 of T which contains vA , and extend this to a Bass-Serre presentation (T 0 , T 1 , {te }e∈E(T 1 −T 0 ) )
for Λ which contains a (connected) lift of ∆.
Let T∆ be the connected component of the inverse image of ∆ in T containing vA , and let
H0 be the stabilizer of T∆ .
Now each element of the tuple β̄ = (β 1 , . . . , β q ) can be written as β j = gj tej β0j t−1
g 0 where
e0 j
j
• ej , e0j are trivially stabilized edges of T 1 − T 0 joining vA to a vertex in a translate of T∆ ,
• β0j ∈ H0 ;
15
• gj , gj0 ∈ Stab(vA ).
min . In particular
Moreover, the edges {ej , e0j | j = 1, . . . , q} do not lie in any translate of TAγ̄
0
00
this means F admits a free product decomposition F = F ∗ F where FA ⊆ F0 , γ̄ ∈ F0 and
| ∅ FA γ̄ by Theorem 1 of [PS16].
tej , te0j ∈ F00 for all j. In particular, we get te1 te01 . . . teq te0q ^
q
1
0
0
Let β̄0 = (β0 , . . . , β0 , g1 , g1 , . . . , gq , gq ). The minimal subgraph Λmin
Aβ0 lies in ∆, hence its
min
intersection with ΛAγ̄ is a disjoint union of envelopes of rigid vertice.
| A γ̄. On the other hand,
Since β̄0 ∈ FA we can apply Proposition 4.6 to get that β̄0 ^
| Aβ̄ γ̄. Applying transitivity on
| ∅ FA γ̄ so te1 te0 . . . teq te0q ^
we have that te1 te01 . . . teq te0q ^
1
0
| A γ̄, which implies
A ⊆ Aβ̄0 ⊆ Aβ̄0 te1 te01 . . . teq te0q and γ̄, we get that Aβ̄0 te1 te01 . . . teq te0q ^
| A γ̄ as wanted.
β̄ ^
Remark 4.11: Proposition 4.9 follows from Proposition 4.10. Indeed, denote ∆β (respectively
∆γ ) the imprint of the (single) sandwich term β (respectively γ), and let P be a (possibly trivial)
path from vA to ∆β ∪ ∆γ . Without loss of generality the endpoint of P lies in ∆β - if we let
∆ = P ∪ ∆β , the hypotheses of Proposition 4.10 are satisfied and as a conclusion we get that
| A γ.
β^
Now we will prove that if b, c satisfy the hypotheses of Theorem 4.7, their elements can be
written as a product of sandwich terms which have "disjoint" imprints.
Proposition 4.12: Let F be a free group, let A ⊆ F and let b, c be tuples in F.
Suppose there exists a standard cyclic pointed JSJ decomposition Λ for F relative to A in
which trivially stabilized edges join vA to a vertex of Λmin
FA , and in which the intersection of the
min consists of a disjoint union of envelopes of rigid vertices.
and
Λ
minimal subgraphs Λmin
Ac
Ab
Then there exists sets B and C of sandwich terms such that
1. the intersection
(∪β∈B imp(β)) ∩ (∪γ∈C imp(γ))
is a disjoint union of envelopes of rigid vertices;
2. each element bi of b (respectively cj of c) can be witten as a product of elements of B
(respectively of C);
3. for any β ∈ B and γ ∈ C, the trivially stabilized edges associated to β are distinct from
the trivially stabilized edges associated to γ.
min . We
Proof. Let bj be an element of the tuple b. Consider the path [vA , bj · vA ] - it lies in TAb
subdivide it into finitely many subpaths with endpoints the translates of vA that appear in
[vA , bj · vA ]. Each of these subpaths is thus of the form [g · vA , h · vA ] and contains at most two
trivially stabilized edges which must appear as the first or the last edge of the path (because
any trivially stabilized edge in T is adjacent to a translate of vA ). This means precisely that
min , thus imp(β)
β = g −1 h is a sandwich term, and moreover [vA , β · vA ] lies in a translate of TAb
min
lies in ΛAb . Similarly, each term in the tuple c can be written as a product of sandwich terms
γ such that imp(γ) lies in a translate of Λmin
Ac . This proves the claim.
To prove Theorem 4.7, the idea is to divide up the sets of sandwich terms B and C obtained
above so that they are alternatingly contained in a growing chain of connected subgraphs of
Λ, and then to use Proposition 4.10 inductively to prove the result. The next proposition
summarizes the result obtained from such an inductive application of 4.10.
16
Proposition 4.13: Suppose ∆0 = {vA } ⊆ ∆1 ⊆ . . . ⊆ ∆s = ΛFA is a chain of connected
subgraphs of groups of ΛFA .
Fb(s−1)/2c
Let B, C be sets of sandwich terms which can be partitioned as B = i=0
Bi and
Fbs/2c
C = i=0 Ci in such a way that
• no trivially stabilized edge is associated to two elements which belong to distinct sets from
the list {Bi , Cj | 0 ≤ i ≤ b(s − 1)/2c, 0 ≤ j ≤ bs/2c};
S
• for each i, β∈Bi imp(β) lies in ∆2i+1 , and its intersection with ∆2i is a disjoint union
of envelopes of rigid vertices;
S
• for each i, γ∈Ci imp(γ) lies in ∆2i , and its intersection with ∆2i−1 is a disjoint union
of envelopes of rigid vertices.
| A C.
Then B ^
Proof. We prove this by induction on the length of the chain of subgraphs of groups. If s = 1,
we have that the imprint of any term in C is contained in {vA }, hence applying Proposition
4.10 to ∆ = {vA } gives the result.
Suppose it holds for chains of length at most s, and let us prove it for a chain ∆0 = {vA } ⊆
∆1 ⊆ . . . ⊆ ∆s ⊆ ∆s+1 = Λ.
F
F
We prove this for s = 2k, the odd case is very similar. Thus B = ki=0 Bi and C = ki=0 Ci .
We have
| AB ...B
1. B0 . . . Bk−1 Bk ^
C0 . . . Ck . Indeed, the subgraph of groups ∆s satisfies that
0
k−1
the union of the imprints of terms in B0 . . . Bk−1 C0 . . . Ck is in ∆s , and the union of imprints of terms of Bk intersects ∆s in a disjoint union of envelopes of rigid vertices, so the
|A
hypotheses of Proposition 4.10 are satisfied: we get that B0 . . . Bk−1 C0 . . . Ck−1 Ck ^
| AB ...B
Bk so by transitivity C0 . . . Ck ^
B
.
k
0
k−1
| A C0 . . . Ck . This follows from applying the induction hypothesis to the
2. B0 . . . Bk−1 ^
F
chain ∆0 ⊆ ∆1 ⊆ . . . ⊆ ∆s−1 ⊆ ∆s+1 = ΛFA (we skip ∆s ) and to the sets k−1
i=0 Bi and
| A C0 . . . Ck .
C, we get that B = B0 . . . Bk−1 ^
| A C0 . . . Ck .
From 1. and 2. we get by transitivity that B0 . . . Bk ^
We can finally prove Theorem 4.7.
Proof. We may assume that Λ is a standard pointed cyclic JSJ decomposition, in which any
trivially stabilized edge joins a vA to a vertex in Λmin
FA . Let B and C be the sets of sandwich
terms given by Proposition 4.12. Let ΛB and ΛC be the (possibly disconnected) subgraph of
groups obtained by taking the union of the imprints in ΛFA of all the sandwich terms in B and
B
C respectively. Suppose ΛB (respectively ΛC ) is the union of connected subgraphs ΛB
1 , . . . , Λp
C
(respectively ΛC
1 , . . . , Λq ). Note that for any sandwich term β of B whose imprint lies in
B
Λi , the trivially stabilized edges associated to β join vA to vertices in ΛB
i . In particular, two
terms of B ∪ C whose imprints lie in different connected components do not have a common
associated trivially stabilized edge.
We build a sequence of connected subgraphs ∆0 = {vA } ⊆ ∆1 ⊆ . . . ⊆ ∆r = Λ as follows.
C
Suppose we have built ∆i : if ∆i contains all the subgraphs ΛB
j and Λj , we set ∆i+1 = Λ
and we are done.
17
If not, consider all the components ΛB
j which are not contained in ∆i and such that there
is a path in ΛFA between ∆i and ΛB
which
does not intersect any ΛC
j
k in more than a disjoint
union of envelopes of rigid vertices. If such components exist, we build ∆i+1 by adding them
to ∆i together with the appropriate paths joining them to ∆i , and we set Bbi/2c to be the set
of all terms of B whose imprint lies in those components we have added. If not, we consider
components ΛC
j which are not contained in ∆i and for which there are paths in ΛFA between
C
∆i and Λj which does not intersect any ΛB
k in more than a disjoint union of envelopes rigid
vertices (there must be some), and we proceed similarly.
By construction the chain ∆0 = {vA } ⊆ ∆1 ⊆ . . . ⊆ ∆r = Λ and the partitions given for
the sets B and C satisfy the hypotheses of Lemma 4.13 (up to switching between B and C),
| A C.
hence B ^
5
Proving forking
In this subsection we prove the left to right direction of Theorem 3.14, namely:
Proposition 5.1: Let A ⊂ F be a set of parameters and b, c be tuples from F. Suppose that
in any standard pointed cyclic JSJ decomposition Λ for F relative to A, the intersection of
min
the minimal subgraphs Λmin
Ab and ΛAc is not a (possibly empty) disjoint union of envelopes of
rigid vertices.
Then b forks with c over A.
We will prove the following
Proposition 5.2: Let F be a free group, and A ⊆ F a set of parameters. Let b be a tuple in
F, such that F is freely indecomposable with respect to Ab.
There exists a standard pointed cyclic JSJ decomposition Λ for F relative to A with corresponding tree T such that
• for any pair of adjacent non trivially stabilized edge e = (z, x) and e0 = (z, y) lying in the
minimal subgraph Λmin
Ab for which x, y are neither Z-type nor surface type vertices, and
for any adjacent lifts ê = (x̂, ẑ), ê0 = (ẑ, ŷ) of e, e0 in T , if sx , sy are generating tuples for
the stabilizers of x̂, ŷ, then the conjugacy class of (sx , sy ) is in acleq (Ab);
• for any surface type vertex v of Λmin
Ab , and for any lift v̂ of v in T , there is an element g
in the stabilizer of v̂ which corresponds to a non boundary parallel simple closed curve,
and such that the conjugacy class of g is in acleq (Ab).
Proof. Consider the pointed cyclic JSJ decomposition Λ0 for F relative to Ab corresponding to
the pointed cyclic tree of cylinders T 0 . Denote by v̂Ab the vertex stabilized by Ab in T 0 . Note
that by [OHV11], the stabilizer of v̂Ab is exactly the algebraic closure acl(Ab) of Ab. Let H be
the collection of subgroups of acl(Ab) which are either stabilizers of edges adjacent to v̂Ab , or
conjugates of A.
Consider now the free product decomposition of acl(Ab) relative to H, it is of the form
acl(Ab) = U1 ∗ . . . ∗ Us ∗ F
Let T0 be a JSJ tree for acl(Ab) relative to H in which the minimal subtree for each Ui
is isomorphic to the pointed cyclic JSJ tree of cylinders of Ui relative to the subgroups in H
which are contained in Ui .
18
By Lemma in [GL17], a JSJ tree T for F relative to A can be obtained from T 0 by refining
equivariantly the vertex v̂Ab by T0 and then possibly collapsing pairs of edges joining two
surface type vertices. Thus any rigid vertex stabilizer of the image T 0 of T0 under the collapse
map lies in acl(Ab).
min in T is contained in a translate of the image T of T under the
Any translate of TAb
0
0
collapse map. Note also that distinct translates of T 0 intersect at most in a surface type
vertex by construction.
Suppose e = (z, x) and e0 = (z, y) are two non trivially stabilized edge lying in the minimal
0
subgraph Λmin
Ab for which x, y are non Z-type vertices. Then any adjacent lifts ê = (x̂, ẑ), ê =
0
min
(ẑ, ŷ) of e, e in T are contained in the union of the translates of TAb . Up to conjugating the
min
pair (sx , sy ) of generating tuples for the stabilizers of x̂ and ŷ, we may assume ê lies in TAb
itself, so in particular it lies in T 0 . Now ê and ê0 intersect in a Z-type vertex so ê0 also lies
in T 0 . Thus we see that (sx , sy ) ∈ acl(Ab). So in general, we get that the conjugacy class of
(sx , sy ) is in acleq (Ab).
Suppose v̂ is the lift of a surface type vertex lying in a translate of Λmin
Ab . Up to conjugation
we may assume it lies in the image T 0 of T0 by the collapse map. Let v̂ 0 be a vertex in T0
whose image by the collapse map is v̂. Its stabilizer is a subsurface group of the vertex
group corresponding to v̂, and it is a vertex group of the tree T0 hence it lies in acl(Ab). It
must contain an element g corresponding to a non parallel simple closed curve on the surface
associated to v̂. This proves the claim.
The following proposition will help us conclude
Proposition 5.3: Let F be a free group, and A ⊆ F a set of parameters. Let T be a pointed
cyclic JSJ tree for F relative to A. If e = (z, x) and e0 = (z, y) are two non trivially stabilized
edges of T , the conjugacy class of (sx , sy ) is not in acleq (A).
Proof. Up to conjugating (sx , sy ), we can assume that it lies in FA . It is thus enough to show
that the orbit of (sx , sy ) under AutA (FA ) contains infinitely many distinct conjugacy classes.
Without loss of generality the stabilizer of x is not cylic. Let τe be a Dehn twist about e
by some element of Stab(e). For any m, the equation
γ(sx , sy )γ −1 = (sx , m sy −m )
implies that γ = 1 since hsx i is not abelian. Thus the pairs τem (sx , sy ) lie in distinct conjugacy
classes.
We can now prove Proposition 5.1.
Proof. The proof is split into cases depending on whether there exists a free product decomposition of F separating b and c. Denote by FA the smallest free factor of F that contains A.
Case 1: Suppose that there exists a decomposition F = Fb ∗ FA ∗ Fc for which b ∈ Fb ∗ FA and
c ∈ FA ∗ Fc .
In this case, we build a JSJ tree T for F relative to A as follows: first, apply Proposition
5.2 to the tuple Ab and the group Fb ∗ FA : this gives a JSJ tree Tb for Fb ∗ FA relative to A,
and denote by ΛAb the corresponding graph of groups. It is nothing more than the pointed
19
cyclic JSJ decomposition ΛFA associated to the tree of cylinders, to which have been added
some trivially stabilized edges.
Again apply Proposition 5.2, this time to the tuple Ac and the group FA ∗ Fc : this gives a
JSJ tree Tc for FA ∗ Fc relative to A, and denote by ΛAc the corresponding graph of groups.
Here again, this is nothing more than the pointed JSJ decomposition ΛFA with some added
trivially stabilized edges.
We can thus get a JSJ decomposition Λ for F relative to A by adding to ΛFA both sets of
trivially stabilized edges. Both ΛAb and ΛAc are subgraph of groups of Λ. Denote by T the
associated tree.
min
Now since the intersection of the minimal subgraphs Λmin
Ab and ΛAc intersect in more than
a disjoint union of envelopes of rigid vertices, we can assume that one of the following holds:
• there exists adjacent non trivially stabilized edges ê = (ẑ, x̂) and ê0 = (ẑ, ŷ) of T with
x̂, ŷ neither Z type nor surface type vertices, such that if sx , sy are generating tuples
for the stabilizers of x̂, ŷ then the conjugacy class of (sx , sy ) is both in acleq (Ab) and
acleq (Ac);
• there exists a surface type vertex of T and elements g, g 0 in the surface vertex group
corresponding to non boundary parallel simple closed curves such that the conjugacy
classes of g, g 0 respectively are in acleq (Ab), acleq (Ac) respectively.
In the first case, we get that acleq (Ab) ∩ acleq (Ac) contains an element which is not in
| A g 0 by the proof of Lemma 8.7 of
| A c. In the second case, we have that g 6^
acleq (A), so b 6^
| A c.
[PS16], hence we must have b 6^
Case 2: Suppose that we are not in Case 1, that is, there is no JSJ decomposition in which Λmin
Ab
and Λmin
have
no
trivially
stabilized
edge
in
common.
Suppose
moreover
that
there
exists
a
Ac
min ∩ Λmin is
standard pointed cyclic JSJ decomposition Λ in which the intersection Λmin
∩
Λ
Ac
FA
Ab
a (possibly empty) disjoint union of envelopes of rigid vertices.
min
In other words, Λmin
Ab and ΛAc intersect only in trivially stabilized edges and envelopes of
rigid vertices.
Assume for a contradiction that b is independent from c over A. We can write F as
FA ∗ F1 with basis ha1 , . . . , an i ∗ he1 , . . . , em i. Consider the tuple c as a tuple of words
c(a1 , . . . , an , e1 , . . . , em ) in this basis. We now double the free group F over FA and we obtain
F̂ := F2 ∗ FA ∗ F1 where F2 is a copy of F1 with basis he01 , . . . , e0m i. Now consider the tuple c0
defined by c0 = c(a1 , . . . , an , e01 , . . . , e0m ).
There is a natural standard pointed cyclic JSJ decomposition Λ̂ for F̂ relative to A: it is
the "double" of Λ obtained by adding to Λ an extra copy of every trivially stabilized edge.
min
min
Because Λmin
Ab ∩ ΛAc ∩ ΛFA is a disjoint union of envelopes of rigid vertices, we see that
min
the intersection of Λ̂min
Ab and Λ̂Ac0 is a disjoint union of envelopes of rigid vertices.
By the other direction, given by Theorem 4.7, we see that b is independent from c0 over A.
By Facts 2.2 and 2.3 we get that b is independent from c0 over acl(A). By our contradictory
assumption, similarly b is independent from c over acl(A). But by Fact 2.5 every type over an
algebraically closed set is stationary and since c and c0 have the same type over acl(A), they do
have the same type over acl(A)b. Therefore, by homogeneity, they are in the same orbit under
Aut(F̂/Ab). Since Aut(F̂/Ab) preserves FA , the automorphism sending c0 to c and fixing Ab
20
Figure 3: A JSJ decomposition and its double.
sends the decomposition F̂ = F2 ∗ FA ∗ F1 to a decomposition F̂ = F̃2 ∗ FA ∗ F̃1 with b ∈ FA ∗ F̃1
and c ∈ F̃2 ∗ FA . By Kurosh’s Theorem, this induces on F a decomposition F = Fb ∗ FA ∗ Fc
for which b ∈ Fb ∗ FA and c ∈ FA ∗ Fc . But then, we are in Case 1, a contradiction.
Case 3: We place ourselves in the general case. We will find a tuple b0 satisfying:
(i) b0 ∈ acl(Ab);
(ii) there exists a standard pointed cyclic JSJ decomposition Λ for F relative to A in which
min
min
the intersection Λmin
Ab ∩ ΛAc ∩ ΛFA consists only of trivially stabilized edges;
(iii) the minimal free factor containing Ab is the same as the minimal factor containing Ab0 .
Then either we can find a decomposition F = Fb0 ∗ FA ∗ Fc with Ab0 ∈ Fb0 ∗ FA and Ac
in FA ∗ Fc , but Ab is then contained in Fb0 ∗ FA as well, which means we are in Case 1 (for
the original tuple b) - thus b forks with c over A. Or, there is no such decomposition, that is,
Case 1 does not apply to b0 . But then, Case 2 applies by (ii), so b0 and c fork over A. Since
b0 ∈ acl(Ab), we are done.
We now explain how to find such a tuple b0 .
Consider the smallest free factor FAb of F containing Ab. Let Λ0 be a pointed cyclic JSJ
decomposition for F relative to Ab, denote by T 0 its associated tree.
Let H be the collection of subgroups of acl(Ab) consisting of conjugates of A together with
the stabilizers of edges adjacent to the base vertex in T 0 . We want to build a JSJ decomposition
Λ for F relative to A by refining the vertex corresponding to acl(Ab) in Λ0 by a pointed cyclic
JSJ decomposition ∆ of acl(Ab) relative to A and to the incident edge groups.
We choose ∆ as follows: consider the Grushko decomposition U0 ∗ . . . ∗ Ur ∗ Hb of acl(Ab)
relative to H, where A ≤ U0 , the Ui ’s are freely indecomposable, and non-cyclic if i > 0, and
Hb is a free group. For each i, let ∆i be the pointed cyclic JSJ decomposition of Ui relative to
the subgroups in H which lie in Ui . Choose a non surface type vertex vi in each ∆i (it always
contains such a vertex - a Z-vertex adjacent to a surface vertex is enough), with v0 = vA . We
build ∆ from the disjoint union of the ∆i ’s, by adding a trivially stabilized edge (v0 , vi ) for
each i, and rk(Hb ) extra trivially stabilized edges as loops based at v0 .
Let b0 be a concatenation of tuples b00 , . . . , b0r , β̄ where
21
• bj ∈ Uj is a generating set for the non surface type vertex group associated to vj ;
• β̄ is a basis for Hb .
Then the subgroup generated by Ab0 is V0 ∗ V1t1 . . . ∗ Vrtr ∗ Hb where Vj is a subgroup of FA
which is a vertex group of vj and tj corresponds to the trivially stabilized edge joining vj to
the vertex stabilized by A.
We now prove b0 has the required properties:
(i) By construction, bj ∈ Uj ≤ acl(Ab) and β̄ ∈ Hb ≤ acl(Ab). Thus b0 ∈ acl(Ab).
min
min
(ii) In Λ, the intersection Λmin
Ab0 ∩ ΛAc ∩ ΛFA is a disjoint union of envelopes of rigid vertices
min
- this is simply because the intersection Λmin
Ab0 ∩ ΛFA is a disjoint union of envelopes of
rigid vertices.
(iii) If we collapse in Λ all the trivially stabilized edges of ∆, the minimal subgraph Λmin
FAb is
mapped onto a graph of groups for FAb (the smallest free factor containing Ab) without
trivially stabilized edges, and all of whose vertex groups are freely indecomposable with
respect to Ab0 and adjacent edge groups (vertices which already existed before the collapse
are vertices of a JSJ decomposition hence they are rigid relative to adjacent edge groups,
and the unique new vertex has vertex group exactly Ab0 ). This proves that FAb is freely
indecomposable relative to Ab0 , hence it is exactly FAb0 .
This finishes the proof.
References
[GL11]
[GL17]
Vincent Guirardel and Gilbert Levitt, Trees of cylinders and canonical splittings,
Geom. Topol. 15 (2011), no. 2, 977–1012.
, JSJ decompositions of groups, Astérisque 395 (2017).
[Mar02] David Marker, Model theory: an introduction, Graduate Texts in Mathematics, vol.
217, Springer, 2002.
[OH11]
Abderezak Ould Houcine, Homogeneity and prime models in torsion-free hyperbolic
groups, Confluentes Mathematici 3 (2011), no. 1, 121–155.
[OHV11] Abderezak Ould Houcine and Daniele Vallino, Algebraic and definable closure in free
groups, Annales de l’institut Fourier 66 (2011).
[PS12]
Chloé Perin and Rizos Sklinos, Homogeneity in the Free Group, Duke Math. J. 161
(2012), no. 13, 2635–2668.
[PS16]
Chloé Perin and Rizos Sklinos, Forking and JSJ decompositions in the free group, J.
Eur. Math. Soc. (JEMS) 18 (2016), no. 3, 1983–2017.
[RS94]
Eliyahu Rips and Zlil Sela, Structure and rigidity in hyperbolic groups I, Geom.
Funct. Anal. 4 (1994), 337–371.
[Sel13]
Zlil Sela, Diophantine geometry over groups VIII: Stability, Ann. of Math. (2) 177
(2013), no. 3, 787–868.
22
| 4math.GR
|
Evolutionary Optimization for Decision Making
under Uncertainty
Ronald Hochreiter
arXiv:1401.4696v1 [cs.NE] 19 Jan 2014
May 2011
Abstract
Optimizing decision problems under uncertainty can be done using a variety of solution methods. Soft computing and heuristic approaches tend to
be powerful for solving such problems. In this overview article, we survey
Evolutionary Optimization techniques to solve Stochastic Programming
problems - both for the single-stage and multi-stage case.
Keywords: Optimization under Uncertainty, Optimal Decision Making, Evolutionary Optimization, Stochastic Programming, Multi-stage Stochastic Programming.
1
Introduction
There are many approaches towards solving decision problems under uncertainty. Throughout this paper, we will focus on solving Stochastic Programming problems - both the classical single-stage as well as the multi-stage case.
This article intends to provide ideas on how to use Evolutionary Optimization
techniques to solve this class of optimization problems.
To summarize the concept of Stochastic Programming, we will refer to the
summary of [10]: consider a deterministic decision optimization problem, where
a decision maker aims at finding an optimal (numerical) decision x ∈ Rn by minimizing a deterministic cost function f (·) (or by maximizing a profit function
respectively) given a set X of constraints, which generally consists of various
physical, organizational, and regulatory restrictions. The mathematical formulation of this problem can be simplified to the formulation shown in Equ. (1).
optimize x : f (x)
subject to
x ∈ X.
(1)
During the 1950s Stochastic Programming was initiated by Dantzig [5] and
Beale [2]. The idea is to replace deterministic parameters by probability distributions on some probability space (Ω, F, P), which will be denoted by Ξ in
the following, and to optimize a stochastic cost (or profit) function f (·, ·) over
some probability functional F. A common choice regarding this functional is
the expectation E. As Rockafellar [25] points out, expectations are only suitable for situations where the interest lies in long-range operation, and stochastic
ups and downs can safely average out, which is not the case e.g. for managing
financial risks. The recent progress of unifying probabilistic risk measures, as
1
presented in the seminal paper by Artzner et al. [1] on coherent risk measures,
motivated for using probability functionals based on risk measures. See the
book [23] for more details on modeling, measuring, and managing risk for this
class of optimization applications. A different view on integration of risk measures using the concept of deviation measures is shown in [26]. In summary, the
resulting mathematical meta-formulation of a stochastic program for arbitrary
probability functionals is shown in Equ. (2).
optimize x : F(f (x, Ξ))
subject to
(x, Ξ) ∈ X .
(2)
However, a concrete reformulation of this meta-model into some model,
which can be solved with a numerical optimizer depends to a high degree on
the chosen probability functional, as well as on the structure of the underlying
probability space.
The interested reader is referred to [27] for a theoretical overview of the area
of Stochastic Programming, and to [35] for Stochastic Programming languages,
environments, and applications. Interestingly, evolutionary approaches have not
been applied to a wide range of real stochastic programming problems so far,
only scarce examples are available, e.g. recent works in the field of chemical
batch processing, see [34] and [31].
This paper is organized as follows. Section 2 briefly describes the heuristic
(soft computing) approach, which was chosen to solve the stochastic programs.
Section 3 and Section 4 consider the single-stage and the multi-stage case respectively. In both Sections the optimization as well as the scenario generation
aspect will be discussed. Section 5 concludes the paper.
2
Evolutionary stochastic optimization
All stochastic optimization problems throughout this paper will be solved by
adapting a standard Evolutionary Optimization algorithm, e.g. as surveyed by
[3] and summarized below, to handle the stochastic meta-model directly without
complex reformulations.
P ← GenerateInitialP opulation
Evaluate(P )
while termination conditions not met do
P 0 ← Recombine(P )
P 00 ← M utate(P 0 )
Evaluate(P 00 )
P ← Select(P ∪ P 00 )
end while
As mentioned above, this meta-formulation of Evolutionary Optimization
can be applied to the meta-formulation of Stochastic Programming shown in
Equ. (2) in different ways. We consider the case of finite, discrete scenario sets
in this paper. This is not a serious restriction as many real-world problems can
be satisfyingly described with such sets. In addition, analytical solutions and
continuous descriptions of the uncertainty often imply a gross underestimation
of the complexity of the respective problem.
2
3
Single-stage stochastic optimization
In single-stage stochastic optimization, the input is given by a multi-variate
probability distribution - a discrete one in our special application-related approach. In the following, issues raised by using Evolutionary Optimization techniques are shown using the well-known application of financial risk-return portfolio selection. In a second step, scenario generation for the single-stage case is
boiled down to being solvable using an evolutionary approach.
3.1
Single-stage optimization: Portfolio selection
An ideal application example is financial risk-return portfolio optimization. For
various reasons, this optimization problem naturally fits into the general problem structure usually handled with evolutionary techniques. Therefore, evolutionary approaches have been successfully applied to different classes of portfolio
optimization problems, see e.g. [29], [28], [30], [8], [17], as well as the references
therein, or refer to [18]. Some of the approaches mention the field of Stochastic
Programming explicitly, e.g. [32], and [36].
The classical bi-criteria portfolio optimization problem based on Markowitz
[19] can be summarized as follows: An investor has to choose a portfolio from a
set of (financial) assets A with finite cardinality a = |A| to invest her available
budget. The bi-criteria problem stems from the fact that the investor aims
at maximizing her return while aiming at minimizing the risk of the chosen
portfolio at the same time.
In the single-stage stochastic programming case, the underlying uncertainty
is represented as a multi-variate probability distribution on the respective probability space. The distribution may either be continuous or discrete. The discretized probability distribution, which is used to compute the optimal decision
is called scenario set. This discrete set of scenarios S has finite cardinality
s =P
|S|, where each si is equipped with a non-negative probability pi ≥ 0,
s
and j=1 pj = 1. From the financial market viewpoint, each scenario contains
one possible set of joint future returns of all a assets under consideration for
the portfolio. Using the terminology of Markowitz, each scenario contains the
discounted anticipated return of each asset.
Let x ∈ Ra be some
P portfolio. Without loss of generality, we use budget
normalization, i.e.
a∈A xa = 1. Each component xi of the portfolio vector
denotes the fraction of the available budget B invested into the respective asset
i. We may now rewrite the scenario set S as a matrix S to calculate the discrete
Profit & Loss (P&L) distribution ` for some portfolio x, which is simply the
cross product `x = x, S . We will denote ` the loss distribution in the following.
Finally, let x∗ρ ∈ Ra denote the optimal portfolio given some risk measure ρ and
`∗ρ denote the respective ρ-optimal discrete loss distribution.
While this model clearly represents a multi-criteria optimization model, we
apply a criteria-weighted model, where an additional risk-aversion parameter κ
is defined, i.e.
maximize x : E(`x ) − κρ(`x ),
(3)
subject to
x ∈ X.
An important fact is that this problem already shows the issue that one
really has to care about the most is the genotype-phenotype encoding. The most
straightforward approach to design a usable genotype-phenotype representation
3
of the portfolio selection problem is to use a vector of real values between 0
and 1 and normalize the resulting vector to the sum of the available budget, i.e.
in the most simple case to a budget of 1. However, it can be shown that this
approach rarely leads to stable solutions, i.e. two runs almost never converge
to the same portfolio composition, see also [12].
Hence, we need a special genetic encoding of a portfolio, see [29]. Each gene
thereby consists of two parts: One that determines the amount of budget to
be distributed to each selected asset and one part which determines in which
assets to invest. The first part g1 consists of a predefined number b of real values
between 0 and 1 and the second part g2 is encoded as a bit-string of the size of
the amount of assets.
Example (from [12]). Let us define a bucket size of b = 10 and consider that
want to select the optimal portfolio out of a = 5 assets. A random chromosome
with a fixed number of 3 asset picks may then consist of the following two parts
g1 and g2 .
g1 = (0.4893, 0.3377, 0.9001, 0.3692, 0.1112, 0.7803, 0.3897, 0.2417, 0.4039, 0.0965)
g2 = (1, 1, 0, 0, 1)
If we re-map these two parts, we receive the following portfolio x
x = (0.3, 0.5, 0, 0, 0.2),
which is a valid portfolio composition, and can be used to calculate the loss
function and to conduct a full evaluation of the respective portfolio optimization
formulation within the evolutionary optimization process.
With this Evolutionary Optimization strategy a full-fledged portfolio optimization tool for a plethora of risk measures can be created, as the evolutionary approach only needs to evaluate the risk and not directly optimize over
the complete (probably non-convex) problem. Furthermore, all kinds of constraints, even non-convex ones, may be included. Constraint handling can be
easily added using a penalty structure within the objective function evaluation.
3.2
Single-stage scenario generation
The problem of single-stage scenario generation, i.e. an optimal approximation
of a multi-variate probability distribution can be done via various sampling as
well as clustering techniques. Optimal single-stage approximation have been
done by using e.g. Principal Component Analysis (see e.g. [33]), K-Means
clustering (see e.g. [14]), or moment matching ([15]). The result is directly
affected by the chosen approximation method - both a distance and a heuristic
needs to be specified.
Of course, this problem can be solved with Evolutionary Optimization techniques as well. Consider the following example (from [13]) of a full-fledged
scenario generation procedure: let’s assume that we do have 10 input scenarios
(asset returns), each equipped with the same probability p = 0.1, which might
be the output of some sophisticated asset price sampling procedure, e.g.
(0.017, −0.023, −0.008, −0.022, −0.019, 0.024, 0.016, −0.006, 0.032, −0.023).
4
We want to separate those values optimally into 2 clusters, which then represent
our output scenarios and take a random chromosome, which might look as
follows:
(0.4387, 0.3816, 0.7655, 0.7952, 0.1869, 0.4898, 0.4456, 0.6463, 0.7094, 0.7547)
If we map this vector to represent 2 centers we obtain: (1, 1, 2, 2, 1, 1, 1, 2, 2, 2).
Now we need to calculate a center value, e.g. the mean, and have to calculate
the distance for each value of each cluster to its center, e.g. we obtain center
means (0.0032, −0.0055), which represent the resulting scenarios, each with a
probability of 0.5. The l1 distance for each cluster is (0.0975, 0.0750), so the
objective function value is 0.1725. Now flip-mutate chromosome 9, i.e. (1 −
0.7094) = 0.2906, such that input scenario 9 (return = 0.032) will now be part
of cluster 1 instead of cluster 2. We obtain new scenarios (0.0080, −0.0149) with
probabilities (0.6, 0.4). The objective function value is 0.1475 (or 0.1646 if you
weight the distances with the corresponding output scenario probability), i.e.
this mutation led to a better objective value.
3.3
Wrapping up the single-stage case
In this Section we have shown that Evolutionary Optimization can be conveniently be applied to single-stage stochastic optimization problems - both to the
solution of decision models as well as to the creation of scenario sets, i.e. for the
process of scenario generation. However, one has to be careful with choosing
an appropriate genotype-phenotype representation, as the most straightforward
approach might not lead to a stable solution.
4
Multi-stage stochastic optimization
To summarize the case of multi-stage stochastic optimization, we base the following summary on [11]: we consider algorithmic issues of the computation of
multi-stage scenario-based stochastic decision processes for decision optimization models under uncertainty. We consider that given a multi-stage stochastic
programming problem specific discrete-time stochastic process on the decision
horizon t = 1, . . . , T , a decision maker observes the realization of this random
process ξt at each decision stage t, and takes a decision xt based on all observed
values up to t (ξ1 , . . . , ξt ). Let there now be a sequence of decisions x1 , . . . , xT .
At the terminal stage T we observe a sequence of decisions x = (x1 , . . . , xT )
with realizations ξ = (ξ1 , . . . , ξT ), which lead to cost f (x, ξ) (or likewise profit).
The stochastic optimization task is to find the sequence of decisions x(ξ), which
minimizes some probability functional (most commonly the expectation, a risk
measure, or a combination of these two) of the respective cost function f (x(ξ), ξ)
- see e.g. [7] for a classification of risk measures in this context.
We consider the multi-stage case in such a way that there is at least one
intermediary stage between root and terminal stage, i.e. T > 2. Unfortunately,
the topic of scenario generation is not sufficiently treated in most text books on
stochastic programming. Let us consider multi-stage stochastic programming
5
problems as defined in Equ. (4).
minimize x : F f (x(ξ), ξ)
subject to
(x(ξ), ξ) ∈ X
x∈N
(4)
The multi-variate, multi-stage stochastic process ξ describes the future uncertainty, i.e. the subjective part of the stochastic program, and the constraint
set X defines feasible combinations of x and ξ. This constraint set is used to
model the underlying real-world decision problem. Furthermore, a set of nonanticipativity constraints N , consisting of functions ξ 7→ x which make sure
that xt is only based on realizations up to stage t (ξ1 , . . . , ξt ), is necessary. To
solve multi-stage programs with numerical optimization solvers, the underlying
stochastic process has to be discretized into a scenario tree, and this scenario
tree approximation will inherently fulfill these non-anticipativity constraints.
Two issues negatively affect the application of multi-stage stochastic programs for real-world decision problems: First, modeling the underlying decision
problem is a non-trivial task. Multi-stage optimization models and stochastic
scenario models require stable scenario tree handling procedures, which are considered to be too cumbersome to be applied to real-world applications, and the
communication of tree-based models to non-experts is complicated. Secondly,
modeling the underlying uncertainty is complex and messy. As discussed above,
a good discrete-time, discrete-space scenario tree approximation of the underlying stochastic process has to be generated in order to numerically compute a
solution and the quality of the scenario model severely affects the quality of the
solution.
4.1
Solving the scenario tree optimization problem
The quality of the scenario tree severely affects the quality of the solution of the
multi-stage stochastic decision model, such that any approximation should be
done in consideration of optimality criteria, i.e. before a stochastic optimization
model is solved, a scenario optimization problem has to be solved independently
of the optimization model. It should be noted, that there are also scenario generation approaches where the generation is not decoupled from the optimization
procedure, see e.g. [4]. A major drawback is that these methods are often limited to a certain restricted set of models, and cannot be generalized easily.
In the context of separated scenario optimization, optimality can be defined
as the minimization of the distance between the original (continuous or highly
discrete) stochastic process and the approximated scenario tree. Choosing an
appropriate distance may be based on subjective taste, e.g. Moment Matching
as proposed by [15], selected due to theoretical stability considerations (see [24]
and [9]), which leads to probability metric minimization problems as shown by
[22] and [6], or it may be predetermined by chosen approximation method, e.g.
by using different sampling schemes like QMC in [21] or RQMC in [16], see also
[20]. It is important to remark that once the appropriate distance has been
selected, an appropriate heuristic to approximate the chosen distance has to be
applied, which affects the result significantly. The real algorithmic challenge
of multi-stage scenario generation is caring about the tree structure while still
minimizing the distance. Only in rare cases, this problem can be solved without
heuristics.
6
The following evolutionary approach has been presented by [13]. We assume
that there is a finite set S of multi-stage, multi-variate scenario paths, which are
sampled using the preferred scenario sampling engine selected by the decision
taker. Stages will be denoted by t = 1, . . . , T where t = 1 represents the (deterministic) root stage (root node), and T denotes the terminal stage. Therefore,
the input consists of a scenario path matrix of size |S| × (T − 1). Furthermore,
the desired number of nodes of the tree in each stage is required, i.e. a vector n
of size (T − 1).
We will focus on the uni-variate case. However, the extension to the multivariate case does not pose any structural difficulties besides that a dimensionweighting function for calculating the total distance on which the optimality of
the scenario tree approximation is based on has to be defined.
A crucial part in designing a multi-stage scenario tree generator based on
evolutionary techniques is finding a scalable genotype representation of a tree,
as this is always the case for Evolutionary Optimization approaches - specifically
both in terms of the numbers of stages as well as the number of input scenarios.
The approach surveyed here is using a real-valued vector in the range [0, 1] and
mapping it to a scenario tree given the respective node format n. The length
of the vector is equal to the number of input scenarios s = |S| plus the number
of terminal nodes nT . Thus, the presented algorithm is somewhat limited by
the number of input scenarios. This means that input scenarios should not
simply be a standard set of mindlessly sampled scenario paths, but rather a
thoughtfully simulated view on the future uncertainty. This should not be seen
as a drawback, as it draws attention to this often neglected part of the decision
optimization process.
To map the real-valued vector to a scenario tree, which can be used for
a subsequent stochastic optimization, two steps have to be fulfilled. First, the
real-valued numbers are mapped to their respective node-set given the structure
n of the tree, and secondly, values have to be assigned to the nodes. It should
be noted, that a random chromosome does not necessarily lead to a valid tree.
This is the case if the number of mapped nodes is lower than the number of
nodes necessary given by nt of the respective stage t. If an uniform random
variable generator and a thoughtful node structure is used, which depends on
the number of input scenarios, invalid trees should not appear frequently, and
can be easily discarded if they do appear during the Evolutionary Optimization
process.
We may now conveniently use the Evolutionary Optimization clustering
strategy, which has been shown for creating single-stage scenarios above. This
approach seems to be rather trivial for the single-stage case, but this simple
approach leads to a powerful method for the tedious task of constructing multistage scenario trees for stochastic programming problems, because the nested
probability structure of the stochastic process is implicitly generated.
For multi-stage trees, a crucial point is finding a representative value for the
node-sets determined in the first step of the mapping. There exists a range of
methods, which can be used for the determination of centers, i.e. mapping all
values of a node-sets to one node. The distance of the approximation, which is
used for the calculation of the objective function, will also be affected by this
method. A straight-forward solution is to use the median of the values - see [13]
for other methods and their differences. Fig. 1 shows an example of the algorithm, i.e. 200 sampled paths are used to create a multi-stage [1, 10, 40] scenario
7
1.04
1.025
1.02
1.03
1.015
1.02
1.01
Value
Value
1.01
1.005
1
1
0.99
0.995
0.98
0.97
0.99
1
2
Stage
0.985
3
1
2
Stage
3
Figure 1: Input scenarios (n = 200) (left) and a scenario tree with n = [10, 40]
(right).
tree. In summary, the complex task of generating multi-stage scenario trees can
be handled conveniently by using an Evolutionary Optimization approach.
4.2
Solving the multi-stage decision problem
Solving a multi-stage stochastic programming problem with Evolutionary Optimization techniques is not straightforward. Especially when the underlying
scenario tree is huge, because one would have to calculate an optimal decision
on every node, such that the chromosome would simply be too large, i.e. for
each node of the tree the respective amount of genes (related to the multi-variate
structure of the decision problem) would have to be added to the chromosome.
New simplification strategies have to be devised, which are out of the scope of
this paper.
5
Conclusion
In this paper, several methods on how to apply Evolutionary Optimization
methods to solve stochastic programs have been surveyed and outlined. It is
shown that this class of heuristic optimization solvers can be applied to these
problems conveniently. However, one has to take care of a few pitfalls during
the creation of specific solution methods, which were discussed in this paper.
Acknowledgement: The author is grateful to the MENDEL conference organizing committee for being invited to present this research to the soft computing
community.
References
[1] P. Artzner, F. Delbaen, J-M. Eber, and D. Heath. Coherent measures of
risk. Mathematical Finance, 9(3):203–228, 1999.
[2] E. M. L. Beale. On Minimizing a Convex Function Subject to Linear
Inequalities. Journal of the Royal Statistical Society, Series B, 17:173–184,
1955.
8
[3] C. Blum and A. Roli.
Metaheuristics in combinatorial optimization: Overview and conceptual comparison. ACM Computing Surveys,
35(3):268–308, 2003.
[4] M. S. Casey and S. Sen. The scenario generation algorithm for multistage stochastic linear programming. Mathematics of Operations Research,
30(3):615–631, 2005.
[5] G. B. Dantzig. Linear Programming under Uncertainty. Management Science, 1:197–206, 1955.
[6] J. Dupačová, N. Gröwe-Kuska, and W. Römisch. Scenario reduction in
stochastic programming. An approach using probability metrics. Mathematical Programming, 95(3, Ser. A):493–511, 2003.
[7] A. Eichhorn and W. Römisch. Polyhedral risk measures in stochastic programming. SIAM Journal on Optimization, 16(1):69–95, 2005.
[8] M. A. Gomez, C. X. Flores, and M. A. Osorio. Hybrid search for cardinality
constrained portfolio optimization. In GECCO ’06: Proceedings of the 8th
annual conference on Genetic and evolutionary computation, pages 1865–
1866. ACM Press, 2006.
[9] H. Heitsch, W. Römisch, and C. Strugarek. Stability of multistage stochastic programs. SIAM Journal on Optimization, 17(2):511–525, 2006.
[10] R. Hochreiter.
Evolutionary stochastic portfolio optimization.
In
A. Brabazon and M. O’Neill, editors, Natural Computing in Computational
Finance, volume 100 of Studies in Computational Intelligence, pages 67–87.
Springer, 2008.
[11] R. Hochreiter. Algorithmic aspects of scenario-based multi-stage decision
process optimization. In F. Rossi and A. Tsoukiàs, editors, Algorithmic
Decision Theory 2009, volume 5783 of Lecture Notes in Computer Science,
pages 365–376. Springer, 2009.
[12] R. Hochreiter. An algorithm for evolutionary stochastic portfolio optimization with probabilistic constraints. In R. Matousek, editor, 16th International Conference on Soft Computing (MENDEL 2010), volume 16 of
Mendel Conference, pages 1–6. VUT Press, 2010.
[13] R. Hochreiter. Evolutionary multi-stage financial scenario tree generation.
In C. Di Chio et al., editor, Applications of Evolutionary Computation,
Part II, volume 6025 of Lecture Notes in Computer Science, pages 182–
191. Springer, 2010.
[14] R. Hochreiter and G. Ch. Pflug. Financial scenario generation for stochastic multi-stage decision processes as facility location problems. Annals of
Operations Research, 152(1):257–272, 2007.
[15] K. Høyland and S. W. Wallace. Generating scenario trees for multistage
decision problems. Management Science, 47(2):295–307, 2001.
[16] M. Koivu. Variance reduction in sample approximations of stochastic programs. Mathematical Programming, 103(3, Ser. A):463–485, 2005.
9
[17] D. Lin, X. Li, and M. Li. A genetic algorithm for solving portfolio optimization problems with transaction costs and minimum transaction lots.
In L. Wang, K. Chen, and Y.-S. Ong, editors, Advances in Natural Computation, First International Conference, ICNC 2005, Changsha, China,
August 27-29, 2005, Proceedings, Part III, volume 3612 of Lecture Notes
in Computer Science, pages 808–811. Springer, 2005.
[18] D. Maringer. Portfolio Management with Heuristic Optimization, volume 8
of Advances in Computational Management Science. Springer, 2005.
[19] H. M. Markowitz. Portfolio selection. The Journal of Finance, 7(1):77–91,
1952.
[20] T. Pennanen. Epi-convergent discretizations of multistage stochastic programs via integration quadratures. Mathematical Programming, 116(1-2,
Ser. B):461–479, 2009.
[21] T. Pennanen and M. Koivu. Epi-convergent discretizations of stochastic programs via integration quadratures.
Numerische Mathematik,
100(1):141–163, 2005.
[22] G. Ch. Pflug. Scenario tree generation for multiperiod financial optimization by optimal discretization. Mathematical Programming, 89(2, Ser.
B):251–271, 2001.
[23] G. Ch. Pflug and W. Römisch. Modeling, Measuring and Managing Risk.
World Scientific, 2007.
[24] S. T. Rachev and W. Römisch. Quantitative stability in stochastic programming: the method of probability metrics. Mathematics of Operations
Research, 27(4):792–818, 2002.
[25] R. T. Rockafellar. Coherent approaches to risk in optimization under uncertainty. Tutorial Notes, Version May 14th, 2007, 2007.
[26] R. T. Rockafellar, S. Uryasev, and M. Zabarankin. Generalized deviations
in risk analysis. Finance and Stochastics, 10(1):51–74, 2006.
[27] A. Ruszczyński and A. Shapiro, editors. Stochastic programming, volume 10
of Handbooks in Operations Research and Management Science. Elsevier
Science B.V., Amsterdam, 2003.
[28] F. Schlottmann, A. Mitschele, and D. Seese. A multi-objective approach
to integrated risk management. In C. A. Coello Coello, A. H. Aguirre,
and E.Zitzler, editors, Proceedings of the Evolutionary Multi-Criterion Optimization Conference (EMO 2005), volume 3410 of Lecture Notes in Computer Science, pages 692–706. Springer, 2005.
[29] F. Streichert, H. Ulmer, and A. Zell. Evolutionary algorithms and the
cardinality constrained portfolio selection problem. In Selected Papers of
the International Conference on Operations Research (OR 2003), pages
253–260. Springer, 2003.
10
[30] R. Subbu, P.P. Bonissone, N. Eklund, S. Bollapragada, and
K. Chalermkraivuth. Multiobjective financial portfolio design: a hybrid
evolutionary approach. In The 2005 IEEE Congress on Evolutionary Computation, volume 2, pages 1722–1729. IEEE Press, 2005.
[31] J. Till, G. Sand, M. Urselmann, and S. Engell. A hybrid evolutionary
algorithm for solving two-stage stochastic integer programs in chemical
batch scheduling. Computers & Chemical Engineering, 31(5-6):630–647,
2007.
[32] K.-i. Tokoro. A statistical selection mechanism of ga for stochastic programming problems. In Proceedings of the 2001 Congress on Evolutionary
Computation, volume 1, pages 487–492, 2001.
[33] N. Topaloglou, H. Vladimirou, and S.A. Zenios. Cvar models with selective
hedging for international asset allocation. Journal of Banking and Finance,
26(7):1535–1561, 2002.
[34] M. Urselmann, M.T.M. Emmerich, J. Till, G. Sand, and S. Engell. Design
of problem-specific evolutionary algorithm/mixed-integer programming hybrids: Two-stage stochastic integer programming applied to chemical batch
scheduling. Engineering Optimization, 39(5):529–549, 2007.
[35] S. W. Wallace and W. T. Ziemba, editors. Applications of stochastic programming, volume 5 of MPS/SIAM Series on Optimization. Society for
Industrial and Applied Mathematics (SIAM), 2005.
[36] S.-M. Wang, J.-C. Chen, H.-M. Wee, and K.-J. Wang. Non-linear stochastic
optimization using genetic algorithm for portfolio selection. International
Journal of Operations Research, 3(1):16–22, 2006.
11
| 9cs.NE
|
EWENS SAMPLING AND INVARIABLE GENERATION
arXiv:1610.04212v2 [math.PR] 15 Oct 2016
GERANDY BRITO, CHRISTOPHER FOWLER, MATTHEW JUNGE, AND AVI LEVY
Abstract. We study the number of random permutations needed to invariably generate the symmetric group, Sn , when the distribution of cycle counts has the strong
α-logarithmic property. The canonical example is the Ewens sampling formula, for which
the number of k-cycles relates to a conditioned Poisson random variable with mean α/k.
The special case α = 1 corresponds to uniformly random permutations, for which it was
recently shown that exactly four are needed.
measures, and almost every α, we show that precisely
For strong−1α-logarithmic
(1 − α log 2)
permutations are needed to invariably generate Sn . A corollary is that
for many other probability measures on Sn no bounded number of permutations will
invariably generate Sn with positive probability. Along the way we generalize classic
theorems of Erdős, Tehran, Pyber, Luczak and Bovey to permutations obtained from the
Ewens sampling formula.
Contents
1. Introduction
1.1. Ewens sampling and the logarithmic property
1.2. Statement of Theorem 1
1.3. Generalizations for the Ewens sampling formula
1.4. Overview of proof and result for sumsets
1.5. Further questions
1.6. Notation
1.7. Outline of paper
1.8. Acknowledgements
2. Proof of Theorem 1
3. The upper bound for mα
3.1. The Feller Coupling
3.2. Proofs of Theorem 3, Theorem 4, and Theorem 5
3.3. Proving Proposition 7
4. The upper bound for sα
5. The lower bound for sα
References
1
2
3
5
6
7
8
9
9
9
11
11
17
24
25
28
33
1. Introduction
In 1934, van der Waerden raised the question of the minimal number of uniformly random
permutations required to invariably generate Sn with positive probability [Wae34]. Recent
breakthroughs by Pemantle, Peres, Rivin [PPR15] and Eberhard, Ford, Green [EFG15a]
resolved van der Waerden’s question, showing that precisely four permutations are required.
1
2
G. BRITO, C. FOWLER, M. JUNGE, AND A. LEVY
We consider the same question in the context of more general permutation measures related
to the cycle structure, primarily the Ewens sampling formula (1) introduced in [Ewe72].
It aligns a permutation’s cycle structure with conditioned independent Poisson random
variables.
Elements g1 , . . . , gm of a group G generate the group if the smallest subgroup containing
them is G itself. The symmetric group Sn is the group of permutations of the set [n] :=
{1, . . . , n}. Permutations π1 , . . . , πm invariably generate Sn if for all σ1 , . . . , σm ∈ Sn the
−1
elements σ1 π1 σ1−1 , . . . , σm πm σm
generate Sn . The definition of invariable generation is
conjugacy invariant. Conjugacy classes of the symmetric group are parametrized by the cycle
type of a permutation, defined to be the multiset of cycle lengths in the cycle decomposition
of a permutation. For this reason, invariable generation concerns only the cycle structure of
permutations, rather than the action on [n].
Recent developments show that cycle counts converge to independent Poisson random
variables for many measures on Sn as n tends to infinity. Arratia, Barbour and Tavare have
shown that strong α-logarithmic measures have this property and numerous further results
[AT92, ABT92, ABT00, ABT16]. Measures with general cycle weights are considered in
[BUV11, EU14] and related permutons are introduced in [KKRW15]. Following this work,
Mukherjee used Stein’s method to deduce that a variety of permutation measures have
Poisson limiting cycle counts [Muk16]. Furthermore, the structure of the short cycles in the
Mallows measure was characterized in [GP16].
In addition to its intrinsic interest, the minimal number of permutations required for
invariable generation has applications to computational Galois theory. One may use this
quantity to estimate the runtime of a Monte Carlo polynomial factorization algorithm
[Mus78, Hei86, DS00, PPR15]. This establishes a connection between the irreducible factors
of polynomials and the cycle structure of permutations. The cycle structure of a permutation
is also compared with prime factorization of integers in the survey [Gra08]. An analogy
between divisors of large integers and invariable generation of permutations is then mentioned
in the introduction of [EFG15a], postulating a link between the threshold for invariable
generation and small divisors of a set of random integers.
Despite the age of van der Waerden’s question, it was only recently shown that the number
of uniformly
random permutations required to invariably generate Sn is bounded. A bound
√
of O( log n) was first established in [Dix92]. This was improved to O(1) in [LP93], but
the constant was large (≈ 2100 ). Pemantle, Peres and Rivin reduced the bound to four in
[PPR15], which was then shown to be sharp by Eberhard, Ford and Green in [EFG15a].
We extend the results for uniform permutations to the Ewens’ sampling formula and
beyond. Along the way, we extend results in [ET67, LP93, Bov80], replacing counting
arguments with a tool known as the Feller coupling (8) introduced in [Fel45, p. 815].
1.1. Ewens sampling and the logarithmic property. The Ewens sampling formula
appears throughout mathematics, statistics and the sciences. In the words of Harry Crane
it, “exemplifies the harmony of mathematical theory, statistical application, and scientific
discovery.” This is stated in his survey [Cra16b]. It and its followup [Cra16a] give a nice
tour of the formula’s universal character, describing applications to evolutionary molecular
genetics, the neutral theory of biodiversity, Bayesian nonparametrics, combinatorial stochastic
processes, and inductive inference, to name a few. The sampling formula also underpins
foundational mathematics in number theory, stochastic processes, and algebra.
We start by formally defining the Ewens sampling formula, as well as the strong αlogarithmic property. Let Poi(λ) denote the law of a Poisson random variable with mean
EWENS SAMPLING AND INVARIABLE GENERATION
3
λ. Fix α > 0 and let X1 , . . . , Xn be independent random variables in which Xk has law
Poi(α/k). For a permutation π ∈ Sn let Ck denote the number of k-cycles in π. More
precisely, Ck is the number of disjoint k-element subsets of [n] on which π restricts to a
cyclic permutation. We say that the vector C = (C1 , . . . , Cn ) has distribution ESF(α, n) if
it satisfies the joint distribution:
(1)
h
P[C1 = x1 , . . . , Cn = xn ] = P X1 = x1 , . . . , Xn = xn
n
X
i
kXk = n .
1
The tuple (C1 , . . . , Cn ) uniquely specifies the cycle type of a permutation. Furthermore,
it is easy to see that as π ranges over Sn , the tuple (C1 , . . . , Cn ) associated to π ranges over
the set
(
)
n
X
n
(C1 , . . . , Cn ) ∈ {0, . . . , n} :
kCk = n .
k=1
Thus, (1) specifies a measure on the conjugacy classes of Sn . We extend this to a measure
on Sn (also denoted by ESF(α, n)) by weighting all elements of the same class uniformly.
Note that ESF(1, n) is the uniform measure, and therefore the results we establish regarding
ESF(α, n) will generalize the corresponding results about uniform permutations appearing
in [PPR15, EFG15a].
A more general family that includes ESF(α, n) is obtained by replacing the Xk in (1) with
an arbitrary sequence of independent nonnegative integerP
valued random variables Z1 , Z2 , . . ..
kZk = n, logarithmic growth of
To
avoid
intractable
distortions
when
conditioning
on
P
Zk is traditionally assumed (see [ABT00]). For our purposes, we require the Zk satisfy
the strong α-logarithmic condition:
(2)
|iP[Zi = 1] − α| < e(i)c1 ,
iP[Zi = `] ≤ e(i)c` ,
l ≥ 2,
P −1
P
where
i e(i) and
`c` are both finite.
Besides Poisson, the most commonly used distributions for the Zi are the negative-binomial
and binomial distributions. These are called assemblies, multisets, and selections, respectively
(see [ABT00] for more discussion). The proof that all three of these different distributional
choices for the Zi satisfy the strong α-logarithmic property is in [ABT00, Proposition 1.1].
It is no trouble to work at this level of generality, because the limiting cycle structure of
any family satisfying the strong α-logarithmic property is the same as that for an ESF(α, n)
permutation. We will describe this in more detail in Section 1.4, but in short, the number of
`-cycles converges to an independent Poisson with mean α/`.
1.2. Statement of Theorem 1. We state our result in terms of random variables Z1 , Z2 , . . .
satisfying the strong α-logarithmic property at (2). The example to keep in mind though is
the measure ESF(α, n), since all of these have the same limiting cycle counts.
Given such a collection let µn (α, Z1 , Z2 , . . . , Zn ) be the measure induced on Sn via the
relation at (1) with Xk replaced by Zk . We define mα = mα (Z1 , Z2 , . . .) to be the minimum
number of permutations sampled according to µn to invariably generate Sn with positive
probability as n → ∞:
n
o
(3) mα = inf m : inf P {π1 , . . . , πm } ⊂ Sn , πi ∼ µn , invariably generate Sn > 0 .
m≥2
n≥2
4
G. BRITO, C. FOWLER, M. JUNGE, AND A. LEVY
20
15
10
5
0.0
0.2
0.4
0.6
0.8
1.0
1.2
1.4
Figure 1. The graph h(α) defined at (4), and thus the behavior of mα and
sα . We are not sure about the value at the points of discontinuity in h. See
Question 1 for more discussion.
Our theorem gives a closed formula for mα (see Figure 1), save for a countable exceptional
set. Because we will reference it several times, we write the expression here:
(
(1 − α log 2)−1 , 0 < α < 1/ log 2
(4)
h(α) =
.
∞,
α ≥ 1/ log 2
Theorem 1. Let Z1 , Z2 , . . . be a collection of random variables satisfying the strong αlogarithmic condition at (2). Let mα = mα (Z1 , Z2 , . . .), as in (3), be the minimum number
of permutations to invariably generate Sn with positive probability. Let h(α) be as in (4).
For points of continuity of h it holds that mα = h(α). At points of discontinuity we have
h(α) ≤ mα ≤ h(α) + 1.
This has broader implications beyond measures that have the strong α-logarithmic property.
We use it to deduce that any finite collection of random permutations with cycle counts
asymptotically dominating a Poi(1/j log 2) random variable will fail to generate a transitive
subgroup of Sn .
Corollary 2. Suppose that cj ≥ 1/ log 2 and Xk ∼ Poi(cj /j). Let κn be a sequence of
probability measures on Sn such that for random permutations π ∼ κn the cycle structure, C
satisfies
d ((C1 , . . . , Ck0 ), (X1 , . . . , Xk0 ))TV → 0
for all fixed k0 . Then, the probability that any fixed number of permutations sampled according
to κn invariably generate a transitive subgroup of Sn goes to 0 as n → ∞.
This corollary is particularly relevant to the results in [BUV11] and [EU14]. They describe
measures on Sn that are formed by weighting cycle lengths by parameters cj . The limiting
cycle structure has Cj ∼ Poi(cj /j). Thus, we can create any limiting Poisson cycle structure
we like. In particular, those satisfying Corollary 2.
Another source of alternate measures with limiting Poisson cycle counts comes from
[Muk16]. This fits into the larger body of work on permutons initiated in [KKRW15]. Of
EWENS SAMPLING AND INVARIABLE GENERATION
5
particular interest is Mallow’s measure (introduced in [Mal57]). This measure specifies a
parameter qn > 0 that biases towards more or less inversions in a permutation. A permutation
inv(π)
π has probability proportional to qn
, where inv(π) := |{(s, t)|s < t and πs > πt}|.
The recent article ([GP16]) begins to characterize the cycle structure of Mallow’s measure.
They find that for (1 − qn )−2 n all cycles are on the order of o(n). The high density of
small cycles, along with [Muk16, Theorem 1.4], which proves a limiting Poisson cycle profile,
suggests that no finite collection of permutations sampled according to Mallow’s measure in
this regime will invariable generate Sn . Note that in the regime (1 − qn )−2 n the cycle
counts converge to those of a uniformly random permutation. This is a new and exciting
area. We are hopeful our result will find more applications as these objects become better
understood.
1.3. Generalizations for the Ewens sampling formula. In proving Theorem 1 we
connect an approximation sumset model (described in Section 1.4) back to fixed-set sizes in
random permutations. This also takes place in [PPR15] and [EFG15a]. These two articles
use a small-cycle limit theorem for uniformly random permutations ([AT92, Theorem 1]).
Namely, that the number of `-cycles in a uniformly random permutation converges to a
Poisson random variable with mean 1/` that is independent of the other small cycles.
Permutations sampled from µn have an analogous limit theorem ([ABT00, Theorem 3.2]).
Except now the `-cycles behave like Poisson random variables with mean α/`. This lets
us use similar ideas to analyze the corresponding sumset model. However, our approach
diverges significantly when we connect back to random permutations. Especially in obtaining
an upper bound on mα . The difference is that both [PPR15] and [EFG15a] have access to
a large canon of results for uniformly random permutations. We do not. This requires a
deep excavation where we extend classical results to π ∼ ESF(α, n). Along the way we prove
many new estimates (see Lemma 10, Lemma 11, Lemma 12, Proposition 13, Lemma 17, and
Lemma 19) for larger cycle lengths in ESF(α, n) permutations.
We begin by showing that with high probability the only transitive subgroups containing
a permutation π sampled with distribution ESF(α, n) are An and Sn . This is the analogue
of what [LP93, Theorem 1] proves for uniformly random permutations. This was recently
studied in more detail by Eberhard, Ford, Green in [EFG15b]. We plan to explore this vein
for ESF(α, n) permutations in future work. For our current purposes, the following result
suffices.
Theorem 3. Let π ∼ ESF(α, n). Then with probability 1 − o(1) the only transitive subgroups
containing π are An and Sn .
Pyber and Luczak rely on an important theorem of Bovey regarding primitive subgroups.
We need the analogue of it for ESF(α, n) permutations. The minimal degree of hπi =
{π k : k ∈ Z} is the minimum number of elements of {1, . . . , n} displaced by some power of π.
[Bov80, Theorem 1] says that if π is a uniformly random permutation, then for each ε > 0
and 0 < β < 1 we have
P[minimal degree of hπi > nβ ] < Cε,β nε−β .
We establish a weaker analogue of this result.
Theorem 4. Let π ∼ ESF(α, n). For each 0 < β < 1 and any α it holds that
P[minimal degree of hπi > nβ ] = o(1).
Establishing Theorem 3 also requires a generalization of a classical theorem of Erdös and
Turán (see [ET67, Theorem V]).
6
G. BRITO, C. FOWLER, M. JUNGE, AND A. LEVY
Theorem 5. Q
Let π ∼ ESF(α, n) and ω(n) → ∞.√The probability that the largest prime
n
which divides `=1 `C` is larger than n exp(−ω(n) log n) is 1 − o(1).
The proof of Theorem 3 uses both Theorem 4 and Theorem 5. For all three generalizations
we follow a somewhat similar blueprint to their predecessors. However, the previous work
often uses special features of uniformly random permutations. Counting arguments are
heavily employed. In general, these techniques do not translate to ESF(α, n) permutations.
Our approach is to use the Feller coupling (see Section 3.1) to directly relate the cycle
structure to independent Poisson random variables. In some places this yields more elegant
proofs, in others it becomes a bit technical. Keeping in mind that ESF(1, n) is the uniform
measure, this is a nice high-level approach to extending these results.
1.4. Overview of proof and result for sumsets. Because a fixed set’s size is a sum of
cycle lengths of a permutation, invariable generation is related to sumsets formed from
random multisets. The link is developed over a collaborative arc of Arratia, Barbour and
Tavaré, [ABT92, ABT00, AT13, ABT16]. They study the relationship between α-logarithmic
structures and Poisson random variables. Most relevant for our purposes are the descriptions
of the cycle counts C = (C1 , . . . , Cn ) for permutations sampled according to µn .
They prove in [ABT00, Theorem 3.1, Theorem 3.2] that, for permutations induced by
the strong α-logarithmic property, the small cycle counts evolve to be independent Poisson
random variables, and the large cycle counts follow Ewens sampling formula. Furthermore,
[ABT92, Lemma 1] (also discussed in [ABT16]) shows that the Ewens sampling formula can
be cleanly related via the Feller coupling to independent Poi(α/k) random variables. The
payoff is that we can model the sizes of fixed sets in a permutation with independent Poisson
random variables.
Let X(α) = (X1 , X2 , . . .) be a sequence of Poisson random variables where Xj has mean
α/j. The coupling discussed in the previous paragraph allows us to model the sizes of fixed
sets in a permutation with the random sumset
X
L (X(α)) =
jxj : 0 ≤ xj ≤ Xj .
j≥1
Returning to our opening question, we can bound the required number of permutations
from above if no fixed set size is common to all permutations. And we can bound the number
from below if there is guaranteed to be a common fixed set size. Consider independent
realizations X(1) (α), . . . , X(m) (α). The upper and lower bounds correspond to finding
extremal values of m such that
Tm
(i) i=1 L (X(i) (α)) = {0} with positive probability, and
Tm
(ii) | i=1 L (X(i) (α))| = ∞ almost surely.
Since common elements among independent sumsets correspond to common fixed-set sizes of
independent permutations, invariable generation is analogous to a trivial sumset intersection.
On the other hand, an infinite intersection means that common fixed-sets persist among
the permutations. For the case α = 1, the results of [PPR15, Theorem 1.6] and [EFG15a,
Corollary 2.5] each imply that four sets suffice in (i). That (ii) holds with three sets is due
to [EFG15a, Corollary 3.9]. Here is the analogue for general α > 0.
(i)
Theorem 6. Let sα := inf{m : P[∩m
1 L (X (α)) = {0}] > 0} be the smallest number of
i.i.d. sumset intersections so that the resulting set is trivial with positive probability. Let h(α)
EWENS SAMPLING AND INVARIABLE GENERATION
7
be as in (4), and mα as in Theorem 1. At points where h is continuous we have sα = h(α),
and at the discontinuities h(α) ≤ sα ≤ h(α) + 1.
To establish this result for sumsets, we build upon the ideas in [PPR15] and [EFG15a] to
prove matching lower and upper bounds for sα . These bounds can be found in Proposition 23
and Proposition 24. Parts of the argument are as simple as swapping 1’s for α’s, but in
others, particularly the lower bound, some care is required.
A relevant quantity for establishing the upper bound is pk = pk (α), the probability the
element n appears in L (X). Estimating pk (1) is the major probabilistic hurdle in [PPR15].
There are two difficulties. The first is called a lottery effect; certain unlikely events greatly
skew pk . This is circumvented by using
P quenched
P probabilities p̃k , which ignore an o(1)
portion of the probability space (that
Xj and
jXj are uncharacteristically large). With
this restriction, it is shown in [PPR15, Lemma 2.3] that p̃k (1) ≤ k log(2)−1 . Notice that
4(log 2 − 1) < −1 and a Borel-Cantelli type argument implies finiteness of the intersection.
Establishing the quenched formula is the second hurdle. It requires a clever counting and
partitioning of L (X). We generalize these ideas in Lemma 22 to obtain the upper bound
p̃k (α) ≤ Cε k −1+α log 2+ε
for any ε > 0. Taking into account the lottery effect, and ignoring
Qm a vanishing porm
(i)
tion
of
the
probability
space,
we
can
write
P[k
∈
∩
L
(X
)]
=
1
1 p̃k . When m ≥
(1 − α log 2 + ε)−1 , the product is smaller than n−1 ,and the probabilities are summable.
With this, the Borel-Cantelli lemma implies the intersection is almost surely finite, and thus
it is trivial with positive probability.
The lower bound in Proposition 24 requires more care. We generalize the framework in
[EFG15a, Section 3] to higher dimensions to incorporate more than three sumsets as well as
different values of α. Given I ⊆ N and m sumsets we consider the set of differences
S(I; X(1) , . . . , X(m) ) := {(n1 − nm , n2 − nm , . . . , nm−1 − nm ) : ni ∈ L (X(i) ) ∩ I}.
We use Fourier analysis (as in [EFG15a] and [MT84]) to study a smoothed version of
the indicator function of this set. In particular, we will work with its Fourier transform
F : Tm−1 → C. After obtaining a pointwise bound on F , we integrate around that torus to
obtain, for sufficiently large k and I = (k 1−β , k], that |S(I, M1 , . . . , Mm )| k m−1 and lies
in the cube [−ck, ck]m−1 for some c > 0. The density of these sets is sufficiently high that if
we look in a larger interval, with positive probability we can find j1 , . . . , jm ∈ (k, Dk] such
(1)
(m)
that Xj1 , . . . , Xjm > 0 and
(jm − j1 , . . . , jm − jm−1 ) ∈ S(I; X(1) , . . . , X(m) ).
Combining this with the corresponding point in the set of differences then produces a number
in L (X(1) ) ∩ · · · ∩ L (X (m) ) made solely from values in (k 1−β , Dk] for some constants
β, D > 0. By partitioning N into infinitely many disjoint sets of this form and repeating this
argument for each of them, we conclude that the set of intersections is almost surely infinite.
1.5. Further questions. There is still much to be done for both components of this paper—
random permutations and random sumsets. Recall that Theorem 1 and Theorem 6 do not pin
down the exact behavior at points of discontinuity of h. A glaring question is to characterize
mα and sα at these points. We are doubtful our approach generalizes. Characterizing the
behavior at the critical values is likely difficult, and will require a new idea.
Question 1. What are mα and sα at discontinuity points of h?
8
G. BRITO, C. FOWLER, M. JUNGE, AND A. LEVY
For random permutations we would like a more complete characterization than in Corollary 2. In particular an upper bound.
Question 2. Show that if a permutations has cycle counts that are asymptotically dominated
by Poisson random variables with mean h−1 (2)/k, then two will invariably generate Sn with
positive probability.
Corollary 2 shows that for many permutation measures no finite number will invariably
generate Sn . Possibly if we let the number of permutations grow with n we will see different
behavior.
Question 3. Let α > 1/ log 2. Prove that there are constants β < β 0 such that the probability
β log n Ewens-α permutations generate Sn is bounded away from 0, while β 0 log n fail to
generate Sn with high probability?
We believe that log n is the correct order by the following approximation model. When
α > 1/ log 2 the fixed set sizes in a random permutation are dense in [n]. So, we can model
the occurrence of each fixed size by independent Bernoulli random variables with some
parameter pα > 0. If one performs N independent thinnings of [n] by including each integer
k with probability pα , then the probability k belongs to all N subsets is pN
α . Hence, the
n
probability there are no common elements in the N -fold intersection is (1 − pN
α ) . If we
choose N = c log n this will converge to either 0 or a positive number as we increase c. The
difficulty in answering the question for permutations is we do not have independence for
fixing different fixed set sizes.
We also have a question in the simpler setting of generating Sn . Dixon’s theorem [Dix92]
implies that two uniformly random permutations will generate either An or Sn with high
probability. Does the analogue hold for ESF(α, n) permutations? Because generation relies
on more than just cycle structure, this may be a hard question.
Question 4. How many ESF(α, n) permutations are needed to generate one of An or Sn
with high probability?
1.6. Notation. Note that α > 0 is fixed throughout. As already discussed, we denote
cycle counts of a permutation by C = (C1 , . . . , Cn ) with Cj the number of j-cycles. Unless
otherwise noted, our permutations and cycle vectors come from an ESF(α, n) distribution.
We will let Y(α) denote a vector of independent Poi(α/k) random variables that are obtained
from the Feller coupling in (8). We will also take X(α) to be a vector of independent Poi(α/k)
random variables. This is done to distinguish the permutation and random sumset settings.
Henceforth we will suppress the α dependence unless there is reason to call attention to it.
For an infinite vector, such as Y, we will use the notation Y[i, j] to denote the sub-vector
(Yi , Yi+1 , . . . , Yj ).
For I an interval in N define the sumset
X
L (I, X) :=
jxj : 0 ≤ xj ≤ Xj ,
j∈I
and set L (X) := L (N, X). Given X define
X
(5)
Xj ,
f`,k (X) =
`<j≤k
g`,k (X) =
X
jXj .
`<j≤k
It will be convenient to abbreviate fk := f1,k (X) and gk := g1,k (X). Thus, fk is the number
of available summands smaller than k, and gk is the largest sum attainable with them.
EWENS SAMPLING AND INVARIABLE GENERATION
9
We will use the notation f g and f (n) = O(g(n)) interchangeably for the existence of c
such that for all large enough n it holds that f (n) ≤ cg(n). When the c depends on our choice
of ε we will use f ε g, similarly f m,ε denotes dependence of the constant on both ε and
the number of intersections m. We use the little-o notation f (n) = o(g(n)) to mean that
f (n)/g(n) → 0. Also, f (n) = Õ(g(n)) means there is k > 0 such that f (n) = O(logk (n)g(n)).
The symbol f (n) ≈ g(n) means that f (n)/g(n) → c for some c > 0. A sequence of events
En occurs with high probability if lim inf P[En ] = 1. Also i | j means that i divides j (i.e.
there is an integer k with ik = j.)
1.7. Outline of paper. We will lead off with a proof of Theorem 1. This comes in two
parts: an upper bound from Theorem 6, and a matching lower bound in Proposition 9.
The upper bound in Theorem 6 has two distinct components. We connect random sumsets
to ESF(α, n) permutations in Section 3.1. This section includes the proofs of Theorem 3,
Theorem 4, and Theorem 5. The second component is analyzing the random sumset model.
We do this in Section 4 and provide an upper bound on sα at Proposition 23. The lower
bound is Proposition 9 and proven in Section 5.
1.8. Acknowledgements. We would like to thank Robin Pemantle for several helpful
conversations. We are very grateful for Richard Arratia’s correspondence. He made us aware
of the indispensable Feller coupling. Also, thanks to Clayton Barnes, Emily Dinan, and
Jacob Richey for their assistance in the formative stages of this project.
2. Proof of Theorem 1
First we show that h(α) + 1 is an upper bound for the number of ESF(α, n) permutations
to invariably generate Sn .
Proposition 7. Suppose α is a continuity point of h. A collection of h(α) independent permutations sampled according to an ESF(α, n) invariably generates Sn with positive probability
as n → ∞.
Proof. Let m = h(α). Call the permutations π (1) , . . . , π (m) . Condition π (1) to be odd.
Lemma 19 guarantees this happens with probability bounded away from zero. Define the
events
En,k = {π (1) , . . . , π (m) fix an `-set for some k ≤ ` ≤ n/2},
Fn,k = {π (2) has no `-cycles with ` ≤ k}.
The probability of a common fixed set size amongst all of the permutations is bounded above
by
(6)
c
c
P[En,k ∪ Fn,k
] ≤ P[En,k ] + P[Fn,k
].
Therefore, the π (i) will invariably generate a transitive subgroup of Sn if this quantity is less
than 1.
Now α is a continuity point of h, so
m = h(α) = d(1 − α log 2)−1 )e > (1 − α log 2)−1 .
It follows that there exists some ε > 0 such that
m = (1 − (α + ε) log 2)−1 .
Fix ε0 such that 0 < ε0 < ε. By Lemma 18 for any k and i = 2, . . . , m,
P[π (i) fixes a k-set] ≤ ck −δ ,
10
G. BRITO, C. FOWLER, M. JUNGE, AND A. LEVY
where
δ = 1 − (α + ε0 ) log 2.
Similarly, by Corollary 21 for any k,
α+1 α+1
(1)
−δ
P[π fixes a k-set] ≤ ck · max
,
.
α
1
By independence of the π (i) , we conclude
P[En,k ] ≤ (ck −δ )m · max
α+1 α+1
,
α
1
.
By construction we have that δm > 1. A union bound for k0 ≤ k ≤ n/2 ensures that, given
some β > 0, we can choose k0 such that P[En,k0 ] < β. Moreover, [ABT00, Theorem 3.1]
(2)
ensures that the cycle counts C[1, k0 ] of πn converge in total variation to Y[1, k0 ]. Hence,
lim inf P[Fn,k0 ] > 0.
n→∞
Set β equal to half this limit infimum. Now, for an appropriate k0 and large enough n, it
holds that (6) is strictly less than 1. This means that with positive probability the collection
invariably generates a transitive subgroup of Sn . By Theorem 3 we know this transitive
subgroup is with high probability either An or Sn . Because π (1) is odd, it must be Sn .
Remark 8. The argument above implies that mα ≤ mα+η for all η > 0, since any appropriate
choice of ε0 in the latter case will hold for the former. It follows that h(α) + 1 permutations
will invariably generate with positive probability at discontinuity points of h.
Now we prove a matching lower bound.
Proposition 9. A collection of m < h(α) permutations sampled independently according to
an ESF (α, n) with high probability do not invariably generate Sn , or any transitive subgroup
of Sn .
Proof. Let m < h(α), and fix
Tmsome ε > 0. By Proposition 24 the intersection of the i.i.d.
sumsets with parameter α, i=1 L (X(i) ), is almost surely infinite. It follows that there is
Tm
some k0 = k0 (ε) such that if k ≥ k0 then i=1 L (X(i) ) ∩ [1, k] 6= ∅ with probability at least
1 − ε/2.
Fix some k ≥ k0 and call the random permutations π (1) , . . . , π (m) . By [ABT00, Theorem
3.1] the cycle counts C(i) [1, k] of π (i) converge in total variation to X(i) ∩ [1, k]. Then there
exists some n0 = n0 (ε, k) such that if n ≥ n0 , with probability greater than 1 − ε/2 we have
C(i) [1, k] = X(i) ∩ [1, k] for all i = 1, . . . , m. Since
m
m
\
\
L (X(i) ) ∩ [1, k] ⊂
L X(i) ∩ [1, k] ,
i=1
i=1
with probability at least 1 − ε we have that π (1) , . . . , π (m) each fix a set of size ` for some
` ≤ k. As ε was arbitrary, this completes our proof.
Theorem 1 is a straightforward consequence of the last two propositions.
Proof of Theorem 1. Together [ABT00, Theorem 3.1, Theorem 3.2] guarantee that the cycle
counts of any measure satisfying the strong α-logarithmic property converge in total variation
to the same distribution as an ESF(α, n) permutation. Hence it suffices to consider ESF(α, n)
permutations.
EWENS SAMPLING AND INVARIABLE GENERATION
11
By Proposition 7 we have mα ≤ h(α) at point of continuity of h. At points of discontinuity,
we use Remark 8 to deduce that mα ≤ mα+η ≤ h(α) + 1. Then, by Proposition 9 we have
that h(α) ≤ mα for all α, completing our proof.
3. The upper bound for mα
The main goal of this section is to prove that mα ≤ sα . Our primary tool for relating
Poisson sumsets to permutations is the Feller coupling. This is described at the onset of
Section 3.1. We then use it in that section to prove several estimates on the cycle structure
of ESF(α, n) permutations. With these we prove Theorem 3, Theorem 4, and Theorem 5 in
Section 3.2. We put all of this together and establish Proposition 7 in Section 3.3.
3.1. The Feller Coupling. The Ewens sampling formula can be obtained via an elegant
coupling attributed to Feller [Fel45, p. 815] for uniform permutations. The articles [ABT92,
ABT16] have nice descriptions. We start with a sequence ξ1 , ξ2 , . . . of mixed Bernoulli
α
i−1
random variables with P[ξi = 1] = α+i−1
and P[ξi = 0] = α+i−1
. Define an `-spacing to
be any occurrence of ` − 1 zeros enclosed by a 1 on the left and right (e.g. 1, 0, 0, 0, 1 is
a 4 spacing). Remarkably, the total number of `-spacings, Y` , is distributed as Poi(α/`).
Moreover, the collection Y1 , Y2 , . . . are independent (see [ABT16, ABT92]).
The counts of spacings in ξ1 , . . . , ξn , 1 generate a partition of n, which can be filled in
uniformly randomly to form a permutation with C` cycles of size `, where C` are sampled
according to an ESF (α, n) distribution. Explicitly,
C` = the number of `-spacings in ξ1 , . . . , ξn , 1
(7)
= ξn−`+1 (1 − ξn−`+2 ) · · · (1 − ξn ) +
n−`
X
ξk (1 − ξk+1 ) · · · (1 − ξk+`−1 )ξk+` .
k=1
This gives an intuition for why ESF (1, n) corresponds to a uniformly random permutation.
Indeed, we can inductively construct such a permutation by letting ξi indicate the decision
to complete a cycle, when there is an i-way choice for the next element.
Letting Rn be the index of the rightmost 1 in ξ1 , . . . , ξn and Jn = n + 1 − Rn it follows
from the previous discussion that
(8)
C` ≤ Y` + 1{Jn = `}.
This says that the cycle counts can be obtained from independent Poisson random variables
through a random number of deletions, and possibly one insertion. We require a few estimates
on the behavior of these perturbations.
Lemma 10. Let Jn be as defined above, and Dn =
of deletions. The following four inequalities hold:
Pn
`=1
Y` − 1{Jn = `} − C` be the number
12
G. BRITO, C. FOWLER, M. JUNGE, AND A. LEVY
α
,
n−`
(9)
P[Jn = `] ≤
(10)
P[Jn = `] ≤ cn−α(1−γ) ,
(11)
P[Jn = `] = O(n−α/(1+α) ),
(12)
` < n.
n − nγ < ` < n,
0 < γ < 1, for some c = c(γ) > 0.
` < n.
EDn = O(1).
Proof. The inequality in (9) follows from the formula given in [ABT92, Remark following
(n)
(17)], which says P[Jn = `] = `EC` /n. The coupling at (8) ensures that P[Jn = `] ≤
(α + `P[Jn = `])/n. Solving for P[Jn = `] gives (9).
Fix some γ ∈ (0, 1). For the case ` > n − nγ in (10) we use that when Jn = ` we must
have ξn−`+1 = 1 and ξk = 0 for all larger index terms up to ξn . The probability of this can
be computed explicitly as
P[Jn = `] =
n
n
Y
Y
α
i−1
i−1
≤
.
α+n−`
α+i−1
α+i−1
n−`+2
n−`+1
We estimate the product by converting it to a sum
X
Z n
n
n
Y
i−1
α+i−1
α+x−1
log
= exp −
≤ exp −
dx .
log
α+i−1
i−1
x−1
n−`+1
n−`+1
n−`+1
The last inequality follows from the fact that log α+x−1
x−1 is a decreasing function, and the sum
can be viewed as a right-sided Riemann (over-)approximation. This integral has a closed
form
Z n
α+x−1
(α + n − 1)α+n−1
(α + n − `)α+n−`
−
log
dx = − log
+
log
.
x−1
(n − 1)n−1
(n − `)n−`
n−`+1
Hence, exponentiating the above line gives
(13)
n
Y
n−`+1
(n − 1)n−1
(α + n − `)α+n−`
i−1
≤
×
.
α+n−1
α+i−1
(α + n − 1)
(n − ` − 1)n−`
The leftmost product is less than n−α . The rightmost product can be written as
!n−`
n−`
α
1 + n−`
n−`+α
α
α
(14)
= (α + n − `)
.
(α + n − `)
1
n−`−1
1 − n−`
The rightmost product in (14) is asymptotic to eα /e−1 as n − ` → ∞, and thus may be
bounded by some constant for all values of n − `. Using the hypothesis that n − ` ≤ nγ , it
follows that (14) is universally bounded by cnαγ for some constant c. We conclude that (13)
is less than cn−α+αγ = cn−α(1−γ) . This establishes (10).
To obtain the universal bound at (11), observe when ` ≤ n − nα/(1+α) we can input this
into (9) to obtain
α
P[Jn = `] ≤
= O(n−α/(1+α) ).
n − (n − nα/(1+α) )
EWENS SAMPLING AND INVARIABLE GENERATION
13
And, for ` > n − nα we substitute γ = α/(1 + α) into (10) to obtain the same asymptotic
inequality: O(n−α/(1+α) ).
Lastly, (12) is proven in [ABT92, Theorem 2].
We will also require a bound on the joint probability that two cycles occur simultaneously.
There will be some poly-log terms that, besides being a minor nuisance, do not effect the tack
of our proof. Recall the Õ notation that ignores logarithmic contribution. It is described
carefully in Section 1.6.
Lemma 11. For any i < j ≤ n it holds that
P[Ci Cj > 0] = Õ(i−1 j −α/(1+α) ).
Proof. We start by using the formula at (7) to bound Ci with the ξk ,
Ci ≤ ξn−i+1 (1 − ξn−i+2 ) · · · (1 − ξn ) +
n−i
X
ξk ξk+i .
k=1
Thus,
n−i
X
Ci Cj ≤ ξn−i+1 (1 − ξn−i+2 ) · · · (1 − ξn ) +
ξk ξk+i
k=1
× ξn−j+1 (1 − ξn−j+2 ) · · · (1 − ξn ) +
n−j
X
ξ` ξ`+j
`=1
= (x + y)(z + w).
The last line is just labeling the terms in the product immediately above in the natural
Pn−i
way (i.e. y = k=1 ξk ξk+i ). This is so we can bound E[xz + xw + yw + yz] in an organized
fashion, one term at a time.
Term xz: First, consider ξn−i+1 (1−ξn−i+2 ) · · · (1−ξn )ξn−j+1 (1−ξn−j+2 ) · · · (1−ξn ). When
i < j this product contains the term ξn−i+1 (1 − ξn−i+1 ) ≡ 0, thus it is always zero.
We conclude that
(15)
E[xz] ≡ 0.
Pn−j
Term xw: Next, we consider the cross-term ξn−i+1 (1 − ξn−i+2 ) · · · (1 − ξn ) `=1 ξ` ξ`+j .
Many of the summands are zero. In fact, the overlapping terms with n − j − i + 1 <
` ≤ n − j are identically zero. So, for the sake of obtaining disjoint terms, we consider
the smaller sum
ξn−i+1 (1 − ξn−i+2 ) · · · (1 − ξn )
n−j−i+1
X
ξ` ξ`+j .
`=1
Aside from the ` = n − j − i + 1 term, all the indices are all disjoint. Ignoring this
term for now, the expected value is
(16)
n−j−i
X
α
n−i+1
n−1
α2
···
.
α+n−iα+n−i+1
α+n−1
(α + ` − 1)(α + ` − 1 + j)
`=1
Now i and j correspond to cycle lengths, so we must have i + j ≤ n to have a
nonzero expectation at all. Since i < j we conclude that i < n/2. A simple bound
14
G. BRITO, C. FOWLER, M. JUNGE, AND A. LEVY
on the product outside the sum is to plug this value into the first term and ignore
the rest. This gives the bound
α
n−i+1
n−1
α
2α
···
<
<
.
α+n−iα+n−i+1
α+n−1
n − (n/2) + 1
n
Ignoring
R n 1 the product outside of it, the sum can be compared to the integral
dx to prove it is O(log(j)/j). Thus the expression in (16) is Õ(1/nj).
1 x(x+j)
Rather than using the rough bound from above, the full ` = n − j − i + 1 term is
2
ξn−j−i+1 (1 − ξn−j−i+2 ) · · · (1 − ξn−i )ξn−i+1
(1 − ξn−i+2 ) · · · (1 − ξn ).
The expectation of this term is
n−i
Y
α
α
k−1
α+n−j−i
α+k−1 α+n−i
k=n−j−i+2
n
Y
k=n−i+2
k−1
α+k−1
!
.
But this is just
α
α+n−i
α
P[Jn = j + i] =
P[Jn = j + i]
α+n−i n−i
n−i
2α
≤
P[Jn = j + i].
n
Then by Lemma 10 (11) we have the bound
P[Jn = j + i] = O(n−α/(1+α) ).
Thus this term is O(i−1 j −α/(1+α) ), and
(17)
E[xw] = Õ(i−1 j −α/(1+α) ).
Pn−i
Term yw: Because i < j the term ξn−j+1 (1 − ξn−j+2 ) · · · (1 − ξn ) `=1 ξ` ξ`+i behaves
differently than the case just prior. In fact, it is the largest order term amongst the
four. The first step is the same as before though. A few of the terms are identically
zero. So, we can write yw as
ξn−j+1 (1 − ξn−j+2 ) · · · (1 − ξn )
n−i−j+1
X
ξ` ξ`+i .
`=1
Except for when ` = n − i − j + 1, all of the indices now disjoint. When we take
expectation of the terms with disjoint indices we obtain
n−j−i
X
n−j+1
n−1
α2
α
···
.
α+n−j α+n−j+1
α+n−1
(α + ` − 1)(α + ` + i − 1)
`=1
Notice that j could be quite large. The best bound we have for the product outside
the sum comes from (11). This implies it is O(n−α/(1+α) ). As with (16), the sum
term is O(log i/i). This yields that the whole expression is Õ(i−1 j −α/(1+α) ).
As before, the full ` = n − i − j + 1 term is
2
ξn−i−j+1 (1 − ξn−j−i+2 ) · · · (1 − ξn−j )ξn−j+1
(1 − ξn−i+2 ) · · · (1 − ξn ).
The expectation of this term is
n−j
Y
α
k
−
1
α
α+n−j−i
α+k−1 α+n−i
k=n−j−i+2
n
Y
k=n−j+2
k−1
.
α+k−1
EWENS SAMPLING AND INVARIABLE GENERATION
15
This simplies to
α
α+n−j
α
P[Jn = j + i] =
P[Jn = j + i].
α+n−j n−j
n−j
By construction i ≤ n − j, and again by Lemma 10 (11), we have that P[Jn =
j + i] = O(n−α/(1+α) ). Thus this term has expectation that is O(i−1 j −α/(1+α) ). It
follows that
E[yw] = Õ(i−1 n−α/(1+α) ).
(18)
Term yz: The last step is to bound the product
n−i
X
(19)
ξk ξk+i
n−j
X
k=1
n−i,n−j
X
ξ` ξ`+j =
ξk ξk+i ξ` ξ`+j .
k,`=1
`=1
Call the index set Λ = {(k, `) : k ∈ [1, . . . , n − i], ` ∈ [1, . . . , n − j]}. As i < j, we can
divide Λ into disjoint sets
Λ0 = {k, k + i, `, and ` + j are all distinct},
Λ1 = {k = `},
Λ2 = {k = ` + j},
Λ3 = {k + i = `},
Λ4 = {k + i = ` + j}.
This yields the following upper bound on (19):
X
X
X
X
X
ξk ξk+i ξ` ξ`+j +
ξk ξk+i ξ`+j +
ξk ξk+i ξ` +
ξk ξk+i ξ`+j +
ξk ξk+i ξ` .
Λ0
Λ1
Λ2
Λ3
Λ4
The payoff of this partition is that each product of ξ above is of distinct, independent
terms. This lets us tidily compute the expectation of each. For (k, `) ∈ Λ0 we have
the k, k + i, `, ` + j are all distinct. This gives
E[ξk ξk+i ξ` ξ`+j ] ≤
A
k(k + i)`(` + j)
for some constant A depending only on α. We can take the expectation of the entire
sum
E
X
ξk ξk+i ξ` ξ`+j ≤
Λ0
n
X
E[ξk ξk+i ξk+j ] ≤
k=1,`=1
n
X
k=1,`=1
A
= Õ(1/ij).
k(k + i)`(` + j)
We can parametrize Λ1 by k to write it as Λ1 = {(k, k) : k = 1, . . . , n − j}, this
gives the bound
E
X
Λ1
ξk ξk+i ξ`+j ≤
n
X
k=1
E[ξk ξk+i ξk+j ] ≤
n
X
k=1
A
= Õ(1/ij).
k(k + i)(k + j)
We can similarly parametrize the sums over Λ2 , Λ3 and Λ4 to obtain
(20)
E[yz] = Õ(1/ij)
16
G. BRITO, C. FOWLER, M. JUNGE, AND A. LEVY
All of the expressions at (15), (17), (18), and (20) are Õ(i−1 j −α/(1+α) ). Markov’s
inequality yields
P[Ci Cj > 0] = Õ(i−1 j −α/(1+α) ).
(21)
This proves the claimed inequality.
We also need a bound on the probability that cycles of the same size occur.
Lemma 12. For any i ≤ n it holds that
P[Ci ≥ 2] = Õ(i−2 + n−1−α/(1+α) ).
Proof. We start by bounding P[Ci > 2]. By (8) we have
P[Yi + 1{Jn = i} > 2] ≤ P[Yi > 1] = O(1/i2 ).
(22)
It remains to bound P[Ci = 2]. Conditioning on the value of Jn , we obtain the decomposition
(23)
P[Ci = 2] = P[Ci = 2 | Jn = i]P[Jn = i] + P[Ci = 2 | Jn 6= i]P[Jn 6= i].
First we estimate the first summand on the right hand side of (23). Because two cycles larger
than n/2 cannot occur, we must have i ≤ n/2. Then by Lemma 10 (9), P[Jn = i] ≤ 2α/n.
When Jn = i, it ensures that the sequence ξ1 , . . . , ξn has ξn−i+1 = 1 and the larger index ξk
are zero. The smaller index ξi are independent of this event. The value of Ci conditioned on
Jn = i is thus one more than the number of i-spacings in the sequence
ξ1 , . . . , ξn−i , 1,
with ξ1 , . . . ξn−i independent and distributed ESF (α, n − i). Call the number of i-spacings
in the above sequence Ci− . So, conditional on Jn = i we have Ci = Ci− + 1. By writing
explicitly what it means for an i-spacing to arise, we can bound Ci− in terms of the ξk :
Ci− ≤ ξn−2i+1 (1 − ξn−2i+2 ) · · · (1 − ξn−i ) +
n−2i
X
ξk ξk+i .
k=1
Looking to apply Markov’s inequality, we take the expectation of the above line and obtain
ECi− ≤
n−2i
X
n − 2i + 1
n−i−1
α2
α
···
+
.
α + n − 2i α + n − 2i + 1
α+n−i−1
(α + k − 1)(α + k − 1 + i)
k=1
The product above is equivalent to P[Jn−i = i], then we use Lemma 10 (11) and the fact
that i < n/2 to bound it by Õ((n − i)−α/(1+α) ) = Õ(n−α/(1+α) ). The sum, as we have seen
from (16), has order Õ(1/i). Recalling that (9) implies P[Jn = i] = O(1/n), it follows that
(24)
P[Ci = 2 | Jn = i]P[Jn = i] = Õ(n−1−α/(1+α) + 1/in) = Õ(i−2 + n−1−α/(1+α) ).
Now we estimate the second summand on the right hand side of (23). Under the event
{Jn 6= i}, for Ci to equal 2, it is necessary that two i-spacings occur in the infinite sequence
ξ1 , ξ2 , . . . . The count of i-spacings is distributed as Yi from (8). Therefore,
(25)
P[Ci = 2 and Jn 6= i] ≤ P[Yi ≥ 2] = O(1/i2 ).
Then (22), (24), and (25) imply that
(26)
P[Ci ≥ 2] = Õ(i−2 + n−1−α/(1+α) )
as desired.
Using the previous two lemmas we may prove the following.
EWENS SAMPLING AND INVARIABLE GENERATION
17
α
Proposition 13. Let π ∼ ESF (α, n) and a = 1 − 4α+3
. The probability that π has two
cycles whose lengths have a common divisor larger than na is o(1).
Proof. Call Ea the event that π has two cycles with common divisor larger than na . We
utilize a union bound, then apply Lemma 11 and Lemma 12 to write
P[Ea ] ≤
1−a
bnX
c
n
X
d=dna e
= Õ
bn1−a c
P[Cid Cjd > 0] +
P[Cid ≥ 2]
i=1
1≤i<j
X
n
X
bn1−a c
X
(id)
−1
−α/(1+α)
(jd)
n
X
+
d=dna e 1≤i<j
d=dna e
bn1−a c
X
i=1
1
+ n−1−α/(1+α)
(id)2
.
Since d ≥ na the above line is
Õ n1 n2(1−a) n−a−aα/(1+α) + n1 n1−a n−2a + n1−a−α/(1+α) .
We can simplify the exponents in each term to
3 − 3a − aα/(1 + α),
2 − 3a,
and
1 − a − α/(1 + α),
respectively. Some algebra shows that all are negative as long as a > 1 −
we have P[Ea ] = o(1).
α
4α+3 .
For such a
3.2. Proofs of Theorem 3, Theorem 4, and Theorem 5. We now have enough to
establish our bound on the minimal degree of an ESF(α, n) permutation.
Lemma 14. Let π ∼ ESF(α, n). For each 0 < β < 1, there exists λ ∈ (0, 1) such that if we
define the set
Qn = {d < nβ : there exists a prime p > nλβ with p | d},
then
(27)
P[∃Cd = 1 with d ∈ Qn ] = 1 − o(1).
Proof. We start with [ABT00, Theorem 3.1] which implies
X
TV
(28)
Cd → Poi(κn ),
d∈Qn
where
κn =
X α
.
d
d∈Qn
Borrowing an argument from [Bov80, p. 50] we have
X 1
X 1
X
1
=
−
d
d
d
β
β
d∈Qn
d<n
≥
d<n
p|d⇒p≤nλβ
X 1
Y
−
(1 − 1/p)−1
d
β
λβ
d<n
p≤n
∗
= β log n − eγ λβ log n + O(1)
18
G. BRITO, C. FOWLER, M. JUNGE, AND A. LEVY
∗
where γ ∗ is Euler’s constant. When λ is such that eγ λβ < β/2 we have λn > β2 log n for all
n. It follows from (28) and standard estimates on the concentration of a Poisson random
variable that
X
P[
Cd > (β/4) log n] → P[Poi(λn ) > (β/4) log n] = 1 − o(1).
d∈Qn
To obtain (27), observe that the convergence statement in [ABT00, Theorem 3.1] implies
that P[Cd > 1] → P[Yd > 1] = O(1/d2 ). Here Yd is a Poisson random variable as inP
(8). The
Borel-Cantelli lemma implies that only finitely many Yd are larger than 1. Hence Yi >1 Yi
is almost surely bounded. We have seen that, for n large, the sum of Cd with d ∈ Qn is at
least Poi( β4 log n) with probability 1 − o(1). This diverges, and so one or more of the positive
Cd must be equal to one. Our claim follows.
Qn C`
For a random permutation π ∼ ESF(α, n), we define Φ(π) = `=1 ` . To prove
Theorem 4, it suffices to show that with high probability π has a cycle of length d such that
(29)
d ≤ nβ ,
(30)
d2 - Φ(π).
To that end we establish one more lemma.
Lemma 15. Let π, β, λ, and Qn be as in Lemma 14, then
(31)
P[∃d ∈ Qn such that d2 - Φ(π) and Cd = 1] = 1 − o(1).
Proof. Call the event in (31) Fn . Conditioning on the event that Cd = 1 for some d ∈ Qn ,
P[Fn ] ≥ P[Fn | ∃Cd = 1, d ∈ Qn ]P[∃Cd = 1, d ∈ Qn ].
By Lemma 14 we have that P[∃Cd = 1, d ∈ Qn ] = 1 − o(1), so it suffices to show that
P[Fn | ∃Cd = 1, d ∈ Qn ] = 1 − o(1).
Fix d ∈ Qn with Cd = 1, we will prove
P[d2 | Φ(π)] = o(1).
From our construction of Qn , there exists a prime p > nλβ such that
Φ(π)
}.
d
For such prime number p, the event on the right hand side above satisfies
{d2 | Φ(π)} ⊂ {p |
n/p
X
Φ(π)
{p |
}⊂{
Cip > 1}
d
i=1
Call the last event Gn . The Feller coupling at (8) allows us to bound the probability by
(32)
P[Gn ] ≤ P
hdn/de
X
i
Yip + 1{Jn = ip} > 1 .
i=1
This, in turn, is bounded by
dn/pe
h
i
hdn/pe
i
X α
X
P Poi
>0 +P
1{Jn = ip} > 0 .
ip
i=1
i=1
EWENS SAMPLING AND INVARIABLE GENERATION
19
The mean of the above Poisson random variable is asymptotically bounded by α log(n/p)/p,
which is o(1) since p ≥ nλβ . This ensures that the left summand is o(1). To bound the right
summand we write
hdn/pe
i
X
X
X
P
1{Jn = ip} > 0 ≤
P [Jn = ip] +
P [Jn = ip] .
ip>n−nγ
ip≤n−nγ
i=1
We begin by bounding the first summand of the right hand side. By Lemma 10 (9),
b(n−nγ )/pc
X
X
P [Jn = ip] ≤
ip≤n−nγ
i=1
Z
n−nγ
p
α
n − ip
+1
≤
1
log n
p
= o(1).
α
dx
n − px
=O
The last line follows since p ≥ nλβ . For the second summand, from Lemma 10 (10) there
exists some constant c > 0 such that
X
X
P [Jn = ip] ≤
cn−α(1−γ)
ip>n−nγ
ip>n−nγ
≤ cnγ−λβ n−α(1−γ) .
The last line follows as there are at most nγ−λβ different values of i such that n−nγ < ip ≤ n.
We therefore have that
P[Gn ] = O nγ−λβ−α(1−γ) .
An easy calculation confirms that γ satisfying
γ<
α + λβ
1+α
is sufficient to ensure
P[Gn ] = o(1).
Choosing a suitable γ, it follows that
P[Fn ] = 1 − o(1),
completing our proof.
Proof of Theorem 4. By Lemma 15 with probability 1 − o(1) there exists a cycle length of
length d satisfying (29) and (30). When these conditions are met, for K = Φ(π)/d, we have
π K displaces d ≤ nβ elements. Thus, with high probability π has minimal degree no more
than nβ . This completes our proof.
it also
Proof of Theorem 5. The theorem statement is such that if it holds for ω(n), then
√
4
holds for any ω 0 (n) ≥ ω(n). So, it suffices to prove
the
statement
for
all
ω(n)
≤
log
n.
√
Fix such an ω(n) and set bn = n exp(−ω(n) log n). Let P be the set of all primes and
define Pn = P ∩ [bn , n]. Given p ∈ Pn let Ip = [1, 2, . . . , bn/pc] be the set of all i for which pi
20
G. BRITO, C. FOWLER, M. JUNGE, AND A. LEVY
is a multiple of p between p and n. We also introduce the entire collection of such multiples
smaller than n
[
Wn =
{(p, i) : i ∈ Ip }.
p∈Pn
It suffices to show that the event
En = {∃(p, i) ∈ Wn : ip | Φ(π)}
occurs with high probability. We can write En as a union of events,
[
(n)
{Cip > 0}.
En =
(p,i)∈Wn
Notice that any two numbers in the set of products
{ip : (p, i) ∈ Wn }
are different. To see this we argue by contradiction, suppose that ip = i0 p0 with p < p0 .
0
Because of primality,
this could only happen
if p0 | i. Thus
√
√
√ i ≥ p . Also note that our
4
0
assumption ω(n) ≤ log n ensures bn > n. Thus, p, p > n. These two inequalities give
√ √
ip ≥ p0 p > n n = n.
But ip > n is a contradiction, since our construction of Ip ensures that ip ≤ n.
We have now shownPthat the occurrence of En is equivalent to the sum of cycle counts
being positive: En = { Wn Cip > 0}. The Feller coupling (8) ensures that
X
X
(33)
Cip ≥ −Dn +
Yip ,
Wn
Wn
with Dn the total number of deletions. It follows that
X
(34)
P[En ] ≥ P[−Dn +
Yip > 0].
Wn
Using additivity of independent Poisson random variables, the sum on the right of (33) is
distributed as a Poisson with mean
bn/(p)c
µn,α =
(35)
X
X
p∈Pn
i=1
X log(n/(p))
p
p∈Pn
X 1
X log p
= α log n
−
.
p
p
α/ip ≈ α
p∈Pn
p∈Pn
Lemma 16 shows that µn,α → ∞. To establish that (34) occurs with high probability, we
will show that the number of deletions, Dn , in the Feller coupling is unlikely to remove every
successful Yip in (33).
As µn,α → ∞, we take any h(n) = o(µn,α ) with h(n) → ∞. By Lemma 10 (12) we know
that EDn ≤ c for some c = c(α). It follows from Markov’s inequality that
c
P[Dn ≥ h(n)] ≤
= o(1).
h(n)
Standard estimates on a Poisson random variable tell us that
P[Poi(µn,α ) ≤ h(n)] = 1 − o(1).
P
So Dn is with high probability smaller than Wn Yip , and combined with (33) we have
established that the right side of (34) occurs with high probability.
EWENS SAMPLING AND INVARIABLE GENERATION
21
Lemma 16. Let µn,α be as in the proof of Theorem 5. It holds that µn,α → ∞.
Proof. It suffices to show that (35) tends to infinity. Two classical theorems attributed to
Euler ([Eul37]) and Chebyshev ([Apo76]), respectively, state that for sums of primes smaller
than n
X1
X log p
= log log n + O(1) and
= log n + O(1).
p
p
p≤n
p≤n
Recalling that Pn = P ∩ [bn , n], it follows that (35) is asymptotic to
log n log
n
log n
− log .
log bn
bn
We can express the second term precisely
p
p
n
(36)
log
= log exp ω(n) log n = ω(n) log n.
bn
Similarly we compute the order of the first term. Notice
log n
log n
√
log
= log
log bn
log n − ω(n) log n
√
ω(n) log n
√
= log 1 +
(37)
.
log n − ω(n) log n
√
As ω(n) ≤ 4 log n the above term is log(1 − o(1)). Using the fact that log(1 − x) ≈ −x as
x → 0 we have (37) is asymptotic to
√
ω(n) log n
√
.
log n − ω(n) log n
Reintroducing the log n factor, we now have
(38)
log n
√
ω(n) log n
√
≈ log n
.
p
log n − ω(n) log n
X1
Pn
Combining (36) and (38) gives
√
p
ω(n) log n
√
− ω(n) log n
log n − ω(n) log n
log n
√
= ω(n)2
log n − ω(n) log n
µn,α ≈ log n
= ω(n)2 (1 − o(1))
Since ω(n) → ∞ we conclude that µn,α → ∞.
Now we set our sights on proving Theorem 3. This will require Theorem 4 and Theorem 5,
as well as a few additional lemmas.
Lemma 17. Let 0 < γ < 1 and ψ(n) → ∞ such that ψ(n) = o(log n). The probability that
simultaneously for all 2 ≤ r ≤ exp(ψ(n)) there exists Cjr > 0 such that jr > nγ and r - jr is
1 − o(1).
22
G. BRITO, C. FOWLER, M. JUNGE, AND A. LEVY
Proof. Let r0 = exp(ψ(n)) and consider the complement of the event in question:
E = {∃r < r0 : r | j for all nγ < j such that Cj > 0}.
It suffices to show that P[E] = o(1). In order to model the values Cj with independent
Poisson random variables we will work with a larger event. For any γ < γ 0 < 1, we consider
0
Eγ 0 = {∃r < r0 : r | j for all nγ < j < nγ such that Cj > 0}.
A union bound leads to
P[Eγ 0 ] ≤
r0
X
0
P[r | j for all nγ < j < nγ such that Cj > 0]
r=2
(39)
=
r0
X
0
P[Cj = 0 for all nγ < j < nγ such that r - j].
r=2
0
Let Ar be the set of integers in (nγ , nγ ] that r fails to divide. Formally,
0
Ar = [nγ , nγ ] ∩ ([n] \ {ir : i ≤ bn/rc}).
0
0
TV
0
As nγ = o(n) we can apply [ABT00, Theorem 3.1] to conclude that C[nγ , nγ ] → Y [nγ , nγ ].
Thus, the probabilities at (39) converge to
r0 Y
r0
X α
X
X
P[Yj = 0] =
exp −
(40)
.
j
r=2 j∈Ar
r=2
j∈Ar
P
P
The quantity j∈Ar αj is minimized at r = 2, and in this case j∈A2 1j log n. We can
apply this to (40) to obtain
P[Eγ 0 ] ≤
r0
X
n−δ ≤ r0 n−δ = o(1),
r=2
for some δ > 0. As E ⊆ E , this completes the proof.
γ0
Lemma 18. Let δ = δ(ε, α) = 1 − (α + ε) log 2. For any ε > 0, there exists a constant
c = c(ε, α) such that the probability of π fixing a set of size k is bounded by ck −δ uniformly
for all 1 ≤ k ≤ n/2.
Proof. In Lemma 22 we show that P[k ∈ L(X(α))] ≤ Ck −δ for some C and all k ≤ n. The
Feller coupling lets us relate the probability of a fixing a set size to this quantity, the only
complication is a possible contribution from 1{Jn = `}. However, the location of this extra
cycle is unconcentrated enough to not alter the order of these probabilities.
α
Lemma 10 (9) ensures that P[Jn = `] ≤ n−`
. We make use of this as we condition on the
value of Jn .
X
(41)
P[π has a k-cycle] ≤ P[k ∈ L(X(α))] +
P[k − ` ∈ L(X(α))]P[Jn = `]
`≤k
(42)
≤ Ck −δ +
X
C(k − `)−δ
`≤k
(43)
≤ Ck −δ +
2α X −δ
C`
n
`≤k
≤ Ck −δ + C 0 n−δ .
α
n−`
EWENS SAMPLING AND INVARIABLE GENERATION
23
Where at (42) we use our bound on P[k ∈ L(X(α))] from Lemma 22 and bound P[Jn = `]
with Lemma 10 (9). The next line, (43), reindexes the sum and uses that ` ≤ n/2 to bound
α
−δ
time some
n−` ≤ 2α/n. Our claim then follows by bounding the leftmost sum by n
0
−δ
−δ
constant C , and the fact that k > n .
We now prove a result analogous to [LP93, Theorem 1]. Namely, that if a permutation
sampled according to Ewen’s sampling formula is in a transitive subgroup, this subgroup is
likely to be either all of Sn or the alternating group, An . Though the lemmas culminating
to this used rather different arguments, our proof from here closely follows that in [LP93].
Recall that a primitive subgroup is a transitive subgroup that does not fix any partition of
[n].
Proof of Theorem 3. Let π ∈ G for a transitive subgroup G. According to [GM98], it follows
from results√of [Bab81] that the minimal degree of a primitive subgroup not containing An
is at least ( n − 1)/2. On the other hand, by Theorem 4 with high probability the minimal
degree of hπi is less than n0.4 . Thus, the probability that G is primitive is o(1).
Assume G is a non-primitive, transitive subgroup. Then it must fix some partition {Bi }ri=1
of [n]. By transitivity every block in such a partition will be mapped to every other, thus
the partition must be into blocks of equal size, that is
n
|Bi | = .
r
Note that any cycle of a permutation that preserves this partition must have the same
number of elements in each block it acts on. We will use this to establish that any π ∈ G
with high probability cannot fix this partition for any r.
We consider three cases for the possibly amount, r, of blocks:
√
Case 1: 2 ≤ r ≤ r0 = exp(log log n log n) √
By Lemma 17 with ψ(n) = log log n log n, with high probability for every r there
exists there exists jr > n.99 such that π contains a cycle of length jr and r - jr . Fix
some r and jr , and let B be the union of all blocks this cycle acts upon. Because r
does not divide jr , B is a proper invariant subset of [n] with size sn/r, where s is
the number of blocks in B. By Lemma 18 this occurs with probability bounded by
r0 X
r
X
c(sn/r)−δ ≤ cr02 (n/r0 )−δ = o(1).
r=2 s=1
Case 2: r0 ≤ r ≤ n/r0
By Theorem 5, with high probability there exists a prime q > n/r0 that divides
Ω(π). Thus there exists a cycle of length j such that q | j. Let t ≤ r be the number
of blocks this cycle intersects. Because of the bounds on r we have
j
n
n
≤ |Bi | = ≤
< q.
t
r
r0
Since q divides j = t jt and jt < q by the above inequality, we must have that q | t.
However, the total number of blocks r satisfies
n
r<
<q
r0
so this is impossible.
24
G. BRITO, C. FOWLER, M. JUNGE, AND A. LEVY
Case 3: n/r0 ≤ r < n
Let a be as in√Proposition 13 and ε > 0 such that a + ε < 1. Using Lemma 17 with
ψ(n) = log log n log n one more time, we now find j > na+ε such that π contains a
cycle of length j and the blocks size, s = n/r does not divide j. Again, let B be the
union of blocks which intersect this cycle. Because the size of the blocks does not
divide j, this intersection is a proper subset of B. We conclude that π must contain
another cycle of length j 0 that intersects each block in B. Hence both cycle lengths j
and j 0 are divisible by the number of blocks in B. We bound the number of blocks,
|B|
j
na+ε
> >
> na .
s
s
s
However, by Proposition 13 with probability 1 − o(1) no two cycle lengths have a
common divisor greater than na . This completes our proof.
3.3. Proving Proposition 7. To rule out generating An we need an estimate that shows
an ESF(α, n) permutation is odd with probability bounded away from zero. Of course this
probability ought to converge to 1/2, but we were unable to find a proof. We make due with
an absolute lower bound.
Lemma 19. Let π ∼ ESF (α, n), then
inf min{P[π is odd], P[π is even]} ≥ min
n≥2
1
α
,
α+1 α+1
.
Proof. We will proceed by induction. It is straightforward to check via the Feller coupling
α
that an ESF(α, 2) permutation is odd with probability α+1
, and it is even with probability
1
α
1 − α+1 = α+1 .
Suppose that for 2 ≤ m ≤ n − 1, if πm ∼ ESF (α, m) then
min{P[πm is odd], P[πm is even]} ≥ ε,
for some yet to be determined ε > 0. Now, it suffices to prove that
min{P[π is odd], P[π is even]} ≥ ε.
(44)
Consider the sequence ξ1 , ξ2 , . . . of Bernoulli random variables for the Feller coupling. Recall
that Jn = (n + 1) − Rn , with Rn the index of the rightmost 1 in ξ1 , . . . , ξn . We will condition
on the value of Jn , but need to also take into account the parity of n.
n is odd: We condition on the value of Jn . For convenience let q` = P[Jn = `]. If Jn =
`, then ξn+1−` = 1 and we can view the beginning sequence ξ1 , . . . , ξn−` , 1 as
corresponding to a random permutation sampled with distribution ESF (α, n − `).
We denote such a permutation by πn−` . Noting that when Jn = n, π is a single
n-cycle and therefore even, it follows that
P[π is odd] =
n−3
X
P[πn−` is even]q` +
` odd
n−2
X
P[πn−` is odd]q` + qn−1 .
` even
By the inductive hypothesis
P[π is odd] ≥
n−2
X
`=1
εq` + qn−1 .
EWENS SAMPLING AND INVARIABLE GENERATION
25
Pn
Pn
Since 1 q` = 1 P[Jn = `] = 1, it suffices to prove that qn−1 ≥ ε(qn−1 + qn ).
Equivalently, this requires that (1 − ε)qn−1 ≥ εqn . Notice that
n
n
Y
Y
qn−1 = P[ξ2 = 1]
P[ξi = 0], and qn =
P[ξi = 0].
i=3
i=2
Much of this cancels when we solve (1−ε)qn−1 ≥ εqn , so we arrive at the requirement
(1 − ε)
The ratio
(45)
P[ξ2 =1]
P[ξ2 =0]
P[ξ2 = 1]
≥ ε.
P[ξ2 = 0]
= α, and thus we require
(1 − ε)α ≥ ε.
n even: The argument is similar. The key difference is that we instead need qn ≥ ε(qn−1 +qn ).
Rewriting as before this is the same as
(46)
εα ≤ (1 − ε).
After solving for ε in (45) and (46), we see that (44) holds whenever 0 < ε ≤ min
n
1
α
α+1 , α+1
o
.
Remark 20. When α = 1 this confirms the intuition that even and odd permutations occur
with probability 1/2 for all n ≥ 2.
Corollary 21. Let δ = δ(α, ε) and c = c(α, ε) be as in Lemma 18. It holds for all k ≤ n/2
that
α+1 α+1
−δ
P[π fixes a k-set | π is odd] ≤ ck · max
,
.
α
1
Proof. We start with the reformulation
P[π fixes a k-set | π is odd] =
P[π fixes a k-set and π is odd]
.
P[π is odd]
The numerator is bounded by P[π fixes a k-set] which is less than ck −δ by Lemma 18. We
bound the denominator way from zero by Lemma 19, completing our proof.
4. The upper bound for sα
We carry out the plan described in Section 1.4. First we bound the quenched probability
p̃k defined formally just below here. The proof is similar to [PPR15, Lemma 3.1]. We repeat
a selection of the details for clarity, and point out where the argument differs with general α.
Subsequently, we mirror [PPR15, Theorem 3.2] as we apply the Borel Cantelli lemma.
Recall the functions fk and gk from (5). Namely that fk is the number of available
summands smaller than k, and gk is the largest sum attainable with them. Also, define
τε := sup{k : fk ≥ (α + ε) log k},
τ := sup{n : gλk ≥ k},
j
k
k
where ε > 0 is yet to be chosen, but small, and λk := α log
k . The proofs of [PPR15, Lemma
2.1, Lemma 2.2] generalize in a straightforward way, so that T = T (ε, δ) := max{τε , τ } is
almost surely finite. As P[T > λk ] → 0, we define the quenched probabilities to avoid this
diminishing sequence of events
p̃k := P[T < λk and k ∈ L (X)].
26
G. BRITO, C. FOWLER, M. JUNGE, AND A. LEVY
Lemma 22. Suppose that α < 1/ log 2. For each ε > 0 there exists C = C(ε) such that for
all k ≥ 1,
p̃k (α) ≤ Ck −1+(α+ε) log 2 .
Proof. Fix ε > 0. Let Gk be the event that T < λk while also n ∈ L (X). Call a sequence
y = (y1 , y2 , . . .) admissible if it is coordinate-wise
less than or equal to X. When Gk occurs,
P
it is not possible for n to be a sum of j jyj for an admissible y with yj vanishing for j > λk .
P
Indeed, on the event Gk , we have j≤λk jXj < n. The event satisfying the following three
conditions contains Gk :
(i) fλk ≤ (α + ε) log λk .
(ii) gλk < n.
P
(iii) There is some k with k = j (yj0 + yj00 ) with y0 supported on [0, λk ] and y00 nonzero
and supported on [λk + 1, k] with both y0 and y00 admissible.
P
Define the probability p0k = P[fλk ≤ (α + ε) log λk and gλk < n and ` = j jyj0 ] where y0 is
P
admissible and supported on [1, λn ]. Also, define p00` = P[` = j jyj00 ] where y00 is admissible
and supported on [λk + 1, k]. Using independence we can obtain an upper bound on p̃k by
decomposing into these two admissible vectors:
k
X
p̃k ≤
p0k−` p00`
k=λk +1
(47)
k
X
≤
`=λk +1
!
p0k−`
max p00` .
λk ≤`≤k
The first factor above is (by Fubini’s theorem) equal to E|L (X) ∩ [0, λk ]|. Using the trivial
bound where we assume each sum from X is distinct, and that we are working on the event
fλk ≤ (α + ε) log λk we have
(48)
(α+ε) log 2
|L (X) ∩ [0, λk ]| ≤ 2Zm ≤ λk
.
We now bound the second term. We will show there exists a constant C such that
(49)
p00` ≤ C
log2 k
, for all ` ∈ [λk + 1, k].
k
Let H be the event that Xj ≥ 2 for some j ∈ [λk , k]. As P[Xj ≥ 2] ≤ α/j 2 we can write
(50)
P[H] ≤
k
X
j=λk
α/j 2 ≤ α/λk
log k
.
k
P
The event that ` = j jyj00 for an admissible ` supported on [λk , k], but that H does not occur
P
is contained in the union of events Ej that Xj = 1 and ` = j = i iyi00 for some admissible
y00 supported on [λk , k] \ {j}. Using independence of Xj from the other coordinates of X,
EWENS SAMPLING AND INVARIABLE GENERATION
27
along with our description of what must happen if H does not, we obtain
k
X
1 00
p
j `−j
p00` ≤ P[H] +
j=λk
≤ P[H] +
k
1 X 00
p`−j
λk
j=λk
≤
(51)
log k
1+
k
k
X
p00`−j .
j=λk
It remains to bound the summation in the last factor. Here taking α 6= 1 is a minor
nuisance, but we circumvent any difficulties with the crude bound α ≤ 2. The idea is to use
the generating function
Y
F (z, X) :=
(1 + z j ).
j∈[λk ,k]:Xj =1
j
Writing [z ]F for the jth coefficient of F . Observe that this corresponds to forming j using
the j ∈ [λk , k] for which Xj = 1. It follows that
k
X
k
X
p00`−j ≤
j=λk
E [z `−j ]F (z, X) .
j=λk
Using that [z j ]F (z, X) ≤ F (1, X) we can bound the above line by EF (1, X). This is tractable
using independence of the Xj . Carrying the calculation out we obtain
EF (1, X) =
n
Y
E[1 + 1{Xj = 1}]
j=λk
k
Y
α
j
j=λk
n
Y
2
1+
.
≤
j
=
1+
j=λk
The last term can be expanded and written as
k 2 + 3k + 2
k2
2 log2 k.
2
λk + λk
λk
Combine this with (51) and we have
log k
log3 k
1 + log2 k
.
k
k
This yields (49). To conclude we plug (48) and (49) into (47) and arrive at
p00`
p̃k ≤ C(k − λk )(α+ε) log 2
log3 k
.
k
This is bounded above by a constant multiple of k −1+(α+2ε) log 2 , which, after swapping all
ε’s for ε/2, establishes the lemma.
28
G. BRITO, C. FOWLER, M. JUNGE, AND A. LEVY
Proposition 23. Let sα be as in Theorem 6 and h(α) be as in (4). For continuity points
of h it holds that sα ≤ h(α), and sα ≤ h(α) + 1 for all α.
Proof of Proposition 23. Fix a small ε > 0 and let sα,ε = (1 − (α − 2ε) log 2)−1 . Sample
sα,ε vectors X(1) , . . . , X(sα,ε ) and let T (X(1) ), . . . , T (X(sα,ε ) ) be the a.s. random times diss
tributed as T . Take T ∗ = max{T (X(i) )}1α,ε . By independence of the X(i) and the bound in
Lemma 22 we have the probability that T ∗ < λk and k ∈ L (X(i) ) for all 1 ≤ i ≤ sα,ε is at
most a constant multiple of k sα,ε (−1+(α+ε) log 2) . Our choice of sα,ε ensures that the exponent
is less than −1. The series is thus summable. It follows that finitely many k > T ∗ belong to
the intersection of the L (X(i) ). As T ∗ is almost surely finite (the proof of this can be found
in [PPR15, Section 2]), the result follows by letting ε → 0 so that sα,ε → h(α).
5. The lower bound for sα
Here we establish the lower bounds on mα .
Proposition 24. Let sα be as in Theorem 6 and h(α) be as in (4). For 0 ≤ α < 1/ log 2, it
holds that sα ≥ h(α). When α ≥ 1/ log 2 it holds that sα = ∞.
We prove this by a Fourier-analytic argument. We begin by considering continuity points
α of h, and set m = h(α). We then fix m i.i.d. Poisson vectors X(1) , . . . , X(m) .
Lemma 25. Let I = (0, k] and let ε > 0. Then the event
(m
)
[
mk
(i)
A=
L (I, X ) ⊂ 0,
αε
i=1
holds with probability P[A] ≥ 1 − ε.
Proof. This is a straightforward application of Markov’s inequality.
(i) ∞
Recall our notation that X(i) = Xj i=1 .
Lemma 26. Fix ε ∈ (0, 1/2). For any δ > 0, there is a constant C = C(ε, δ, m, α) such that
the event
X
(i)
(52)
E = E(ε, δ, m, α) := min
Xj ≥ (1 − δ)α log(k/`) − C ∀1 ≤ ` ≤ k
i
`<j≤k
holds with probability P[E] ≥ 1 − ε.
We remark that this is a straightforward generalization of [EFG15a, Lemma 3.3].
Proof. In fact, we will take C αδ1 2 log(m/ε) while ensuring that C ≥ 1. Note that the
claim holds vacuously when ` ≥ e−C k, so we need only consider the case ` < e−C k.
P
(i)
Fix 1 ≤ i ≤ m and let B denote the event that `<j≤k Xj < (1 − δ)α log(k/`) − 1 for
some ` ≤ e−C k. Writing `0 for the smallest power of 2 with `0 > `, we thus have
X
(i)
(53)
Xj ≤ α(1 − δ) log(k/`0 )
`0 <j≤k
on B.
Let D = {2i }i≥0 denote the set of non-negative powers of 2. From (53) it follows that
P
X
(i)
0
(54)
1B ≤
(1 − δ) `0 <j≤` Xj −α(1−δ) log(k/` ) .
`0 ≤2e−C k,
`0 ∈D.
EWENS SAMPLING AND INVARIABLE GENERATION
(i)
29
(i)
are independent and Poisson with mean EXj = αj . For a Poisson
P
(i)
variable P with mean λ, we have EaP = eλ(a−1) . Moreover, the sum `0 <j≤k Xj is Poisson
0
with mean α log(k/` ) + O(1). Taking expectations on both sides of (54) therefore yields
X
P[B]
exp [−αδ − α(1 − δ) log(1 − δ)] log(k/`0 )
Each of the variables Xj
`0 ≤2e−C k,
`0 ∈D.
Near zero, −δ − (1 − δ) log(1 − δ) = −(1 + o(1))δ 2 . Therefore
X
2
2
P[B]
(k/`0 )−αδ e−αδ C ,
`0 ≤2e−C k,
`0 ∈D.
since the sum of the geometric series is bounded by a constant multiple of its first term.
ε
Choosing C αδ1 2 log(m/ε) ensures that P[B] ≤ m
, and now the claim follows by the union
of events bound.
An insight in [EFG15a] (to which they credit the idea to Maier and Tenenbaum from
[MT84]) is that Fourier analysis may be used to obtain a lower bound on the size of the set
n
o
(55)
S(I; X(1) , . . . , X(m) ) := (ni − nm )1≤i<m : ni ∈ L (I, X(i) ) .
We show that the technique extends to the present level of generality. Note that the analysis
of [EFG15a] may be regarded as the special case m = 3 and α = 1 of our arguments.
Let T = R/Z be the unit torus. For θ ∈ T, let kθk denote the distance to Z and set
e(z) = e2πiz . Write θ = (θ1 , . . . , θm ) ∈ Tm . Define
m
Tm
0 = {θ ∈ T : θ1 + · · · + θm = 0}.
Fix an interval I. We define the function F (θ) = FI θ, X(1) , . . . , X(m) via the formula
(i)
(56)
F (θ) :=
Xj
m
YY
1 + e(jθi )
j∈I i=1
2
.
Writing θm = −θ1 − · · · − θm−1 , we regard F (θ) as a function on Tm−1 . It is straightforward
to verify that its Fourier transform is a function F̂ : Zm−1 → C supported on the set
S(I; X(1) , . . . , X(m) ) defined above in (55).
The Cauchy-Schwarz inequality yields that
"
#2
X
X
X
2
(57)
|F̂ (a)|
1≥
F̂ (a) .
a∈Zm−1
a : F̂ (a)6=0
a∈Zm−1
P
Recognizing that S(I; X(1) , . . . , X(m) ) ≥ {a : F̂ (a) 6= 0} , that a∈Zm−1 F̂ (a) = F (0) = 1,
R
P
and that by Parseval’s identity a∈Zm−1 |F̂ (a)|2 = Tm |F (θ)|2 dθ, it follows from (57) that
0
!−1
Z
(58)
|S(I; X(1) , . . . , X(m) )| ≥
|F (θ)|2 dθ
.
Tm
0
Our next task is to bound the integral appearing in (58). We will achieve this in two
steps: first, by obtaining a pointwise bound with respect to θ in Lemma 27, and second, by
integrating out the θ dependence in Corollary 28.
30
G. BRITO, C. FOWLER, M. JUNGE, AND A. LEVY
Let β ∈ (0, 1) be a real parameter satisfying
βα log 2 = 1 − m−1 + δ2 ,
where δ2 > 0 and δ2 = oδ (1) is the parameter occurring in (52). (Note that this is possible
since α is a continuity point of h, so m < (1 − α log 2)−1 .) Recall the event E = E(ε, δ, m, α)
from Lemma 26.
Lemma 27. Fix ε ∈ (0, 1/2) and let I denote the interval I := (k 1−β , k]. For a tuple
θ = (θ1 , . . . , θm ), let j(θ) denote an index for which kθj k is maximal. Then
Y
(59)
EX(1) ,...,X(m) 1E |F (θ)|2 ε,m
(kk θi k ∨ 1)−1+δ2 .
i : 1≤i≤m,
i6=j(θ)
Proof. For each 1 ≤ i ≤ m, set ti := k θi and
β
k ,
ki := kti k,
1,
define the cutoff parameter
kti k ≥ k β
1 < kti k < k β
kti k ≤ 1.
Note that ki ≥ 1 and ki ≥ kti kβ . For each 1 ≤ i ≤ m, let Yi denote the expression
X
(i)
Yi :=
Xj .
ki <j≤k
By definition of the event E, when it occurs we have that
m
m
X
X
Yi ≥ α(1 − δ)
log ki − C(ε).
i=1
i=1
It will be convenient to rewrite this bound in the form
m
m
Y
Y
−α(1−δ) log 2
(60)
1E ε
ki
2 Yi .
i=1
i=1
We define the quantity
(61)
m
Y
R = R θ, X(1) , . . . , X(m) , ε, δ, α := |F (θ)|2
2 Yi .
i=1
To establish the bound (59), we will first show that EX(1) ,...,X(m) R ε,m 1. Once this has
been established, it will follow directly from (60) that
m
Y
−α(1−δ) log 2
(62)
E1E |F (θ)|2 ε,m
ki
.
i=1
Obtaining (59) from (62) is a simple exercise in lower bounding the maximum of a sequence
by its geometric mean. Recalling the definition of j(θ) above, we have that
Y
1
kj(θ) ≥
kim−1 .
i : 1≤i≤m,
i6=j(θ)
Consequently
(63)
m
Y
i=1
−α(1−δ) log 2
ki
≤
Y
i : 1≤i≤m,
i6=j(θ)
m
− m−1
α(1−δ) log 2
ki
.
EWENS SAMPLING AND INVARIABLE GENERATION
31
β
Substituting the lower bound ki ≥ kk θi k ∨ 1 into (63), we see that (62) implies
Y
− m αβ(1−δ) log 2
E 1E |F (θ)|2 ε,m
,
kk θi k ∨ 1 m−1
i : 1≤i≤m,
i6=j
m
αβ(1 − δ) log 2 = 1 + oδ (1) by definition of β, the desired bound follows.
and since m−1
Finally we justify that EX(1) ,...,X(m) R ε,m 1. Substituting (56) into (61) yields that
(i)
m
Y
Y 1 + e(jθi ) 2Xj
.
2Yi
R=
2
i=1
j∈I
Recall that Yi =
(i)
P
R=
ki <j≤k
m
Y
Xj
and that I = (k 1−β , k]. Thus
(i)
Y
i=1
k1−β <j≤ki
1 + e(jθi )
2
2Xj
(i)
Y
ki <j≤k
1 + e(jθi )
√
2
2Xj
.
To compute ER, we recall that if P is a Poisson random variable with mean λ, then
(i)
log EaP = eλ(a−1) . Since the variables Xj are independent Poisson variables with mean
(i)
EXj =
(64)
α
j,
it follows after some simplification that
m
X
X
X α cos(2πjθi )
α cos(2πjθi ) − 1
.
+
ER = exp
2j
j
1−β
i=1
k
ki <j≤k
<j≤ki
Next we apply the following bound from [EFG15a, Lemma 3.4]:
X cos(2πjθ)
(65)
= log min(k, kθk−1 ) + O(1).
j
j≤k
Substituting into (64) yields that
m
X
1
k β ∨ kk θi k 1
kβ
ki ∨ kk θi k
ER = exp α
log
− log
+ O(1) .
+ log
2
ki ∨ kk θi k
2
ki
1 ∨ kk θi k
i=1
By definition of ki , each summand is O(1). Therefore ER ε,m 1, as desired.
Finally we integrate the bound (59) with respect to θ to obtain a bound on the integral
appearing in (58).
Corollary 28. Let I = (k 1−β , k] and recall the function F (θ) = FI (θ) from (56). Then
Z
E1E |F (θ)|2 dθ ε,m k 1−m .
Tm
0
Proof. After the change of variables t = kθ, the bound reduces to verifying that
Z
E1E |F (θ)|2 dt ε,m 1.
k·Tm
0
Applying Lemma 27, we have
Z
Z
(66)
E1E |F (θ)|2 dt ε,m
k·Tm
0
Y
k·Tm
0 i : 1≤i≤m,
i6=j(t).
−(1+δ2 )
kti k ∨ 1
dt.
32
G. BRITO, C. FOWLER, M. JUNGE, AND A. LEVY
By symmetry, we may upper bound the integral (66) by
Z
(67)
m
m−1
Y
kti k ∨ 1
−(1+δ2 )
Z
!m−1
k/2
(|ti | ∨ 1)
dt = m
k·Tm
0
i=1
−(1+δ2 )
dti
.
−k/2
The rightmost integral in (67) is bounded above by
Z ∞
m−1
−(1+δ2 )
m 2
ti
dti + 2
.
1
The desired result now follows since the latter integral is finite and independent of k.
Substituting the result of Corollary 28 into (58) lets us show |S(I; X(1) , . . . , X(m) )| k m−1
with high probability.
Proposition 29. Let I = (k 1−β , k]. With probability greater than 1/2, S(I; X(1) , . . . , X(m) ) ⊂
[−ck, ck]m−1 for some constant c and |S(I; X(1) , . . . , X(m) )| k m−1 .
Proof.
L (I, X(i) ) ⊂
3mk By
Lemma 25 with ε = 1/3, with probability greater than 2/3 we have
(1)
⊂ [0, ck] for some constant c and all i. It follows that S(I; X , . . . , X(m) ) ⊂
0, α
m−1
[−ck, ck]
.
By Lemma 26 with ε = 1/20, the event E(ε, δ, m, α) holds with probably greater 19/20
and by Corollary 28,
Z
E1E |F (θ)|2 dθ ε,m k 1−m .
Tm
0
It follows by Markov’s inequality that for some sufficiently large L,
"(Z
)c #
" Z
#
2
1−m
c
2
1−m
P
|F (θ)| dθ k
≤ P [E ] + P 1E
|F (θ)| dθ > Lk
Tm
0
Tm
0
E1E Tm |F (θ)|2 dθ
1
0
+
20
Lk 1−m
1
≤
.
10
It then follows from (58) that with probability at least 9/10,
!−1
Z
R
≤
|S(I; X(1) , . . . X(m) )| ≥
|F (θ)|2 dθ
k m−1 .
Tm
0
Both events then occur simultaneously with probability at least 1/2, completing our
proof.
Proposition 30. Let D be sufficiently large constant
and I = (k 1−β ,Dk]. With probability
(i)
bounded away from zero there exists an array xj : 1 ≤ i ≤ m, j ∈ I for which
(i)
(i)
0 ≤ xj ≤ Xj
for all i and j ∈ I, and
X
j∈I
(1)
jxj
= ··· =
X
j∈I
(m)
jxj
.
EWENS SAMPLING AND INVARIABLE GENERATION
33
Proof. We define the interval I 0 = (k 1−β , k]. By Proposition 29 with probability greater
than 1/2
S(I 0 ; X(1) , . . . , X(m) ) ⊂ [−ck, ck]m−1
and |S(I 0 ; X(1) , . . . , X(m) )| k m−1 .
Let d = 2(5/2)1/α . Straightforward calculations tell us that independently with probability
(m)
greater than 1/2 there exists jm ∈ (2ck, cdk] such that Xjm > 0. Given this jm , there are
k m−1 collections of constants {j` }m−1
∈ (ck, (d + 1)ck]m−1 such that
1
(jm − j1 , . . . , jm − jm−1 ) ∈ S(I 0 ; X(1) , . . . , X(m) ).
Therefore, independently with probability bounded away from zero, there is one such
(`)
collection {j` }m−1
such that Xj` > 0 for all `.
1
m
By the definition of S(I 0 ; X(1) , . . . , X(m) ) we have a set of values ni ∈ L (I 0 , X(i) ) i=1
such that
m−1
(n1 − nm )1
= (jm − j` )m−1
.
1
Manipulating the terms coordinate-wise yields
j1 + n1 = · · · = jm + nm .
Our claim follows by setting D = (d + 1)c.
Tm
Corollary 31. i=1 L (X(i) ) is almost surely infinite.
Proof. We define k1 to be sufficiently large for Proposition 30 to hold, then inductively define
1/(1−β)
ki+1 = (Dki )
. The intervals Ii = (ki1−β , Dki ] are pairwise disjoint and the set
m
\
L (Ii , X(j) )
j=1
is non-empty with probability bounded away from zero. Since these events are independent
for distinct values of i, the claim now follows from the second Borel-Cantelli lemma.
Proof of Proposition 24. As we had fixed m = h(α), by Corollary 31 sα ≥ h(α) holds
for continuity points of h. For points of discontinuity, we fix α0 < α such that h(α0 ) =
(m)
h(α) = m. We couple two sets of independent realizations X(1) (α),
Tm. . . , X (i)(α)0 and
(1)
0
(m)
0
(i)
0
(i)
X (α ), . . . , X (α ) such that X (α ) ⊆ X (α) for all i. Now i=1 L (X (α )) is
almost surely finite by Corollary 31. Our coupling ensures L (X(i) (α0 ) ⊆ L (X(i) (α)) for all
Tm
Tm
0
i, so i=1 L (X(i)(α ) ) ⊆ i=1 L (X(i) (α)). Thus the second intersection is almost surely
infinite, and our claim follows.
Proof of Theorem 6. This now follows immediately from Proposition 23 and Proposition 24.
References
[ABT92]
[ABT00]
[ABT16]
[Apo76]
Richard Arratia, A. D. Barbour, and Simon Tavaré, Poisson process approximations for the
ewens sampling formula, Ann. Appl. Probab. 2 (1992), no. 3, 519–535.
R. Arratia, A. D. Barbour, and S. Tavaré, Limits of logarithmic combinatorial structures, Ann.
Probab. 28 (2000), no. 4, 1620–1644.
Richard Arratia, A. D. Barbour, and Simon Tavaré, Exploiting the feller coupling for the ewens
sampling formula, Statist. Sci. 31 (2016), no. 1, 27–29.
Tom M. Apostol, Introduction to analytic number theory, Undergraduate texts in mathematics,
Springer, New York, Berlin, Heidelberg, 1976, Autres tirages : 1984, 1986, 1995, 1998, 2010.
34
[AT92]
G. BRITO, C. FOWLER, M. JUNGE, AND A. LEVY
Richard Arratia and Simon Tavaré, The cycle structure of random permutations, Ann. Probab.
20 (1992), no. 3, 1567–1591.
[AT13]
, Independent process approximations for random combinatorial structures, arXiv preprint
arXiv:1308.3279 (2013).
[Bab81]
László Babai, On the order of uniprimitive permutation groups, Annals of Mathematics 113
(1981), no. 3, 553–568.
JD Bovey, The probability that some power of a permutation has small degree, Bulletin of the
[Bov80]
London Mathematical Society 12 (1980), no. 1, 47–51.
[BUV11]
Volker Betz, Daniel Ueltschi, and Yvan Velenik, Random permutations with cycle weights, Ann.
Appl. Probab. 21 (2011), no. 1, 312–331.
[Cra16a]
Harry Crane, Rejoinder: The ubiquitous ewens sampling formula, Statist. Sci. 31 (2016), no. 1,
37–39.
[Cra16b]
, The ubiquitous ewens sampling formula, Statist. Sci. 31 (2016), no. 1, 1–19.
[Dix92]
John D. Dixon, Random sets which invariably generate the symmetric group, Discrete Mathematics 105 (1992), no. 1, 25 – 39.
J.H. Davenport and G.C. Smith, Fast recognition of alternating and symmetric galois groups,
[DS00]
Journal of Pure and Applied Algebra 153 (2000), no. 1, 17 – 25.
[EFG15a] S. Eberhard, K. Ford, and B. Green, Invariable generation of the symmetric group, ArXiv
e-prints (2015).
[EFG15b] Sean Eberhard, Kevin Ford, and Ben Green, Permutations fixing a k-set, International Mathematics Research Notices (2015).
P Erdős and P Turán, On some problems of a statistical group-theory. ii, Acta Mathematica
[ET67]
Hungarica 18 (1967), no. 1-2, 151–163.
[EU14]
Nicholas M. Ercolani and Daniel Ueltschi, Cycle structure of random permutations with cycle
weights, Random Structures & Algorithms 44 (2014), no. 1, 109–133.
[Eul37]
Leonhard Euler, Observationes circa series, infinitas, Commentarii Academiae Scientarum
Petropolitanae 9 (1737), 160–188.
[Fel45]
William Feller, The fundamental limit theorems in probability, Bulletin of the American Mathematical Society 51 (1945), no. 11, 800–832.
[GM98]
Robert Guralnick and Kay Magaard, On the minimal degree of a primitive permutation group,
Journal of Algebra 207 (1998), no. 1, 127–145.
A. Gladkich and R. Peled, On the Cycle Structure of Mallows Permutations, ArXiv e-prints
[GP16]
(2016).
[Gra08]
Andrew Granville, The anatomy of integers and permutations, preprint.[Ack] (2008).
Joos Heintz, On polynomials with symmetric galois group which are easy to compute, Theoretical
[Hei86]
Computer Science 47 (1986), 99 – 105.
[KKRW15] R. Kenyon, D. Kral, C. Radin, and P. Winkler, Permutations with fixed pattern densities, ArXiv
e-prints (2015).
Tomasz Luczak and Lászlò Pyber, On random generation of the symmetric group, Combinatorics,
[LP93]
Probability and Computing 2 (1993), 505–512.
[Mal57]
C. L. Mallows, Non-null ranking models. i, Biometrika 44 (1957), no. 1/2, 114–130.
[MT84]
H. Maier and G. Tenenbaum, On the set of divisors of an integer, Inventiones mathematicae 76
(1984), no. 1, 121–128.
[Muk16]
Sumit Mukherjee, Fixed points and cycle structure of random permutations, Electron. J. Probab.
21 (2016), 18 pp.
[Mus78]
David R. Musser, On the efficiency of a polynomial irreducibility test, J. ACM 25 (1978), no. 2,
271–282.
[PPR15]
Robin Pemantle, Yuval Peres, and Igor Rivin, Four random permutations conjugated by an
adversary generate sn with high probability, Random Structures & Algorithms (2015), n/a–n/a.
[Wae34]
B.L. van der Waerden, Die seltenheit der gleichungen mit affekt, Mathematische Annalen 109
(1934), 13–16.
EWENS SAMPLING AND INVARIABLE GENERATION
Department of Mathematics, University of Washington
E-mail address: [email protected]
Department of Mathematics, University of Washington
E-mail address: [email protected]
Department of Mathematics, Duke University
E-mail address: [email protected]
Department of Mathematics, University of Washington
E-mail address: [email protected]
35
| 4math.GR
|
OVERCOMING THE LIMITATIONS OF PHASE
TRANSITION BY HIGHER ORDER ANALYSIS OF
REGULARIZATION TECHNIQUES
arXiv:1603.07377v3 [math.ST] 20 Oct 2017
By Haolei Weng∗ , Arian Maleki† and Le Zheng‡
Columbia University∗†‡
We study the problem of estimating a sparse vector β ∈ Rp from
2
the response variables y = Xβ + w, where w ∼ N (0, σw
In×n ), under the following high-dimensional asymptotic regime: given a fixed
number δ, p → ∞, while n/p → δ. We consider the popular class of
`q -regularized least squares (LQLS), a.k.a. bridge estimators, given
by the optimization problem:
1
β̂(λ, q) ∈ arg min ky − Xβk22 + λkβkqq ,
β 2
and characterize the almost sure limit of p1 kβ̂(λ, q) − βk22 , and call
it asymptotic mean square error (AMSE). The expression we derive
for this limit does not have explicit forms and hence is not useful in
comparing LQLS for different values of q, or providing information
in evaluating the effect of δ or sparsity level of β. To simplify the
expression, researchers have considered the ideal “error-free” regime,
i.e. w = 0, and have characterized the values of δ for which AMSE is
zero. This is known as the phase transition analysis.
In this paper, we first perform the phase transition analysis of
LQLS. Our results reveal some of the limitations and misleading features of the phase transition analysis. To overcome these limitations,
we propose the small error analysis of LQLS. Our new analysis framework not only sheds light on the results of the phase transition analysis, but also describes when phase transition analysis is reliable, and
presents a more accurate comparison among different regularizers.
1. Introduction.
1.1. Objective. Consider the linear regression problem where the goal is
to estimate the parameter vector β ∈ Rp from a set of n response variables
y ∈ Rn , under the model y = Xβ + w. This problem has been studied extensively in the last two centuries since Gauss and Legendre developed the least
squares estimate of β. The instability or high variance of the least squares
estimates led to the development of the regularized least squares. One of
MSC 2010 subject classifications: Primary 60K35, 60K35; secondary 60K35
Keywords and phrases: Bridge regression, phase transition, comparison of estimators,
small error regime, second order term, asymptotic mean square error, optimal tuning
1
2
HAOLEI WENG, ARIAN MALEKI AND LE ZHENG
the most popular regularization classes is the `q -regularized least squares
(LQLS), a.k.a. bridge regression [1, 2], given by the following optimization
problem:
(1.1)
1
β̂(λ, q) ∈ arg min ky − Xβk22 + λkβkqq ,
β 2
P
where kβkqq = pi=1 |βi |q and 1 ≤ q ≤ 2.1 LQLS has been extensively studied in the literature. In particular, one can prove the consistency of β̂(λ, q)
under the classical asymptotic analysis (p fixed while n → ∞) [4]. However,
this asymptotic regime becomes irrelevant for high-dimensional problems in
which n is not much larger than p. Under this high dimensional setting, if
β does not have any specific “structure”, we do not expect any estimator
to perform well. One of the structures that has attracted attention in the
last twenty years is the sparsity, that assumes only k of the elements of β
are non-zero and the rest are zero. To understand the behavior of the estimators under structured linear model in high dimension, a new asymptotic
iid
framework has been proposed in which it is assumed that Xij ∼ N (0, 1/n),
k, n, p → ∞, while n/p → δ and k/p → , where δ and are fixed numbers
[5, 6, 7, 8, 9].
One of the main notions that has been widely studied in this asymptotic
framework, is the phase transition [5, 6, 7, 10]. Intuitively speaking, phase
transition analysis assumes the error w equals zero and characterizes the
value of δ above which an estimator converges to the true β (in certain
sense that will be clarified later). While there is always error in the response
variables, it is believed that phase transition analysis provides reliable information when the errors are small. In this paper, we start by studying
the phase transition diagrams of LQLS for 1 ≤ q ≤ 2. Our analysis reveals
several limitations of the phase transition analysis. We will clarify these
limitations in the next section. We then propose a higher-order analysis of
LQLS in the small-error regime. As we will explain in the next section, our
new framework sheds light on the peculiar behavior of the phase transition
diagrams, and explains when we can rely on the results of phase transition
analysis in practice.
1.2. Limitations of the phase transition and our solution. In this section,
we intuitively describe the results of phase transition analysis, its limitations,
and our new framework. Consider the class of LQLS estimators and suppose
1
Bridge regression is a name used for LQLS with any q ≥ 0. In this paper we focus on
1 ≤ q ≤ 2. To analyze the case 0 ≤ q < 1, [3] has used the replica method from statistical
physics.
HIGHER ORDER ANALYSIS OF REGULARIZATION TECHNIQUES
3
that we would like to compare the performance of these estimators through
the phase transition diagrams. For the purpose of this section, we assume
that the vector β has only k non-zero elements, where k/p → with ∈
(0, 1). Since phase transition analysis is concerned with w = 0 setting, it
considers limλ→0 β̂(λ, q) which is equivalent to the following estimator:
arg min kβkqq ,
β
(1.2)
subject to y = Xβ.
Below we informally state the results of the phase transition analysis. We
will formalize the statement and describe in details the conditions under
which this result holds in Section 3.
Informal Result 1. For a given > 0 and q ∈ [1, 2], there exists a number
Mq () such that as p → ∞, if δ ≥ Mq () + γ (γ > 0 is an arbitrary number),
then (1.2) succeeds in recovering β, while if δ ≤ Mq () − γ, (1.2) fails.2
The curve δ = Mq () is called the phase transition curve of (1.2). We will
show that Mq () is given by the following formula:
1 if 2 ≥ q > 1,
(1.3) Mq () =
inf χ≥0 (1 − )Eη12 (Z; χ) + (1 + χ2 ) if q = 1,
where η1 (u; χ) = (|u|−χ)+ sign(u) denotes the soft thresholding function and
Z ∼ N (0, 1). While the above phase transition curves can be obtained with
different techniques, such as statistical dimension framework proposed in [7]
and Gordon’s lemma applied in [30, 21], we will derive them as a simple
byproduct of our main results in Section 3 under message passing framework. Also, the phase transition analysis of the regularized least squares
has already been performed in the literature [17, 6, 31]. Hence, we should
emphasize that the presentation of the phase transition results for bridge
regression is not our main contribution here. We rather use it to motivate
our second-order analysis of the asymptotic risk, which will appear later in
this section. Before we proceed further let us mention some of the properties
of M1 () that will be useful in our later discussions.
Lemma 1.
M1 () satisfies the following properties:
(i) M1 () is an increasing function of .
(ii) lim→0 M1 () = 0.
2
Different notions of success have been studied in the phase transition analysis. We will
mention one notion later in our paper.
4
HAOLEI WENG, ARIAN MALEKI AND LE ZHENG
(iii) lim→1 M1 () = 1.
(iv) M1 () > , for ∈ (0, 1).
Proof. Define F (χ, ) , (1−)Eη12 (Z; χ)+(1+χ2 ). It is straightforward
to verify that F (χ, ), as a function of χ over [0, ∞), is strongly convex and
has a unique minimizer. Let χ∗ () be the minimizer. We write it as χ∗ () to
emphasize its dependence on . By employing the chain rule we have
dM1 ()
d
∂F (χ∗ (), ) ∂F (χ∗ (), ) dχ∗ ()
∂F (χ∗ (), )
+
·
=
∂
∂χ
d
∂
∗
2
2
∗
∗
2
= 1 + (χ ()) − Eη1 (Z; χ ()) > 1 + (χ ()) − E|Z|2
=
= (χ∗ ())2 > 0,
which completes the proof of part (i). To prove (ii) note that
0 ≤ lim min(1 − )Eη12 (Z; χ) + (1 + χ2 )
→0 χ≥0
≤ lim (1 − )Eη12 (Z; log(1/)) + (1 + log2 (1/))
→0
Z ∞
= lim 2(1 − )
(z − log(1/))2 φ(z)dz
→0
log(1/)
Z ∞
= lim 2(1 − )
z 2 φ(z + log(1/))dz
→0
0
Z ∞
log2 (1/)
−
2
≤ lim 2(1 − )e
z 2 φ(z)dz = 0,
→0
0
where φ(·) is the density function of standard normal. Regarding the proof
of part (iii), first note that as → 1, χ∗ () → 0. Otherwise suppose χ∗ () →
χ0 > 0 (taking a convergent subsequence if necessary). Since Eη12 (Z; χ∗ ()) ≤
E|Z|2 = 1, we obtain
lim F (χ∗ (), ) = 1 + χ20 > 1.
→1
On the other hand, it is clear that
lim F (χ∗ (), ) ≤ lim F (0, ) = 1.
→1
→1
A contradiction arises. Hence the fact χ∗ () → 0 as → 1 leads directly to
lim M1 () = lim F (χ∗ (), ) = 1.
→1
→1
Part (iv) is clear from the definition of M1 ().
HIGHER ORDER ANALYSIS OF REGULARIZATION TECHNIQUES
5
1
0.8
q<1
q=1
q>1
ε
0.6
0.4
0.2
0
0
0.2
0.4
δ
0.6
0.8
1
Fig 1. Phase transition curves of LQLS for (i) q < 1: these results are derived in [3]
by the non-rigorous replica method from statistical physics. We have just included them
for comparison purposes. In this paper we have focused on q ≥ 1. (ii) q = 1: the blue
curve exhibits the phase transition of LASSO. Below this curve LASSO can “successfully”
recover β. (iii) q > 1: The magenta curve represents the phase transition of LQLS for any
q > 1. This figure is based on Informal Result 1 and will be carefully defined and derived
in Section 3.
Figure 1 shows Mq () for different values of q. We observe several peculiar
features: (i) As is clear from both Lemma 1 and Figure 1, q = 1 requires
much fewer observations than all the other values of q > 1 for successful
recovery of β. (ii) The values of the non-zero elements of β do not have any
effect on the phase transition curves. In fact, even the sparsity level does
not have any effect on the phase transition for q > 1. (iii) For every q > 1,
the phase transition of (1.2) happens at exactly the same value.
These features raise the following question: how much and to what extent
are these phase transition results useful in applications, where at least small
amount of error is present in the response variables? For instance, intuitively
speaking, we do not expect to see much difference between the performance
of LQLS for q = 1.01 and q = 1. However, according to the phase transition analysis, q = 1 outperforms q = 1.01 by a wide margin. In fact the
performance of LQLS for q = 1.01 seems to be closer to that of q = 2 than
q = 1. Also, in contrast to the phase transition implication, we may not
expect LQLS to perform the same for β with different values of non-zero
elements. The main goal of this paper is to present a new analysis that will
shed light on the misleading features of the phase transition analysis. It will
also clarify when and under what conditions the phase transition analysis is
reliable for practical guidance.
6
HAOLEI WENG, ARIAN MALEKI AND LE ZHENG
2 of the error w is assumed to be
In our new framework, the variance σw
small. We consider (1.1) with the optimal value of λ for which the asymptotic
kβ̂(λ,q)−βk2
2
mean square error, i.e., limp→∞
, is minimized. We first obtain
p
the formula for the asymptotic mean square error (AMSE) characterized
through a series of non-linear equations. Since σw is assumed small, we then
derive the asymptotic expansions for AMSE as σw → 0. As we will describe
later, the phase transition of LQLS for different values of q can be obtained
from the first dominant term in the expansion. More importantly, we will
show that the second dominant term is capable of evaluating the importance
of the phase transition analysis for practical situations and also provides a
much more accurate analysis of different bridge estimators. Here is one of
the results of our paper, presented informally to clarify our claims. All the
technical conditions will be determined in Sections 2 and 3.
Informal Result 2. If λ∗ denotes the optimal value of λ, then for any
q ∈ (1, 2), δ > 1, and < 1
2
q+1 (1 − )2 (E|Z|q )2
1
σw
2q δ
2q
kβ̂(λ∗ , q) − βk22 =
− σw
+ o(σw
),
p→∞ p
1 − 1/δ
(δ − 1)q+1 E|G|2q−2
lim
where Z ∼ N (0, 1) and G is a random variable whose distribution is specified
by the non-zero elements of β. We will clarify this in the next section. Finally,
the limit notation we have used above is the almost sure limit.
2
σw
As we will discuss in Section 3, the first term 1−1/δ
determines the phase
transition. Moreover, we have further derived the second dominant term in
the expansion of the asymptotic mean square error. This term enables us to
clarify some of the confusing features of the phase transitions. Here are some
important features of this term: (i) It is negative. Hence, the AMSE that is
predicted by the first term (and phase transition analysis) is overestimated
specially when q is close to 1. (ii) Fixing q, the magnitude of the second
dominant term grows as decreases. Hence, for small values of σw all values
of 1 < q < 2 benefit from the sparsity of β. Also, smaller values of q seem to
benefit more. (iii) Fixing and δ, the power of σw decreases as q decreases.
This makes the absolute value of the second dominant term bigger. As q
decreases to one, the order of the second dominant term gets closer to that
of the first dominant term and thus the predictions of phase transition analysis become less accurate. We will present a more detailed discussion of the
second order term in Section 3. To show some more interesting features of
our approach, we also informally state a result we prove for LASSO.
HIGHER ORDER ANALYSIS OF REGULARIZATION TECHNIQUES
7
Informal Result 3. Suppose that the non-zero elements of β are all larger
than a fixed number µ. If λ∗ denotes the value of λ that leads to the smallest
AMSE, and if δ > M1 (), then
(1.4)
2
1
δM1 ()σw
2
kβ̂(λ∗ , 1) − βk22 −
= O(exp(−µ̃/σw
)),
p→∞ p
δ − M1 ()
lim
where µ̃ is a constant that depends on µ.
As can be seen here, compared to the other values of q, q = 1 has smaller
first order term (according to Lemma 1), but much smaller (in magnitude)
second order term. The first implication of this result is that the first dominant term provides an accurate approximation of AMSE. Hence, phase transition analysis in this case is reliable even if small amount of noise is present;
that is one of the main reasons why the theoretically derived phase transition curve matches the empirical one for LASSO. Furthermore, note that in
order to obtain Informal Result 3, we have made certain assumption about
the non-zero components of β. As will be shown in Section 3, any violation
of this assumption has major impact on the second dominant term.
In the rest of the paper, we first state all the assumptions required for
our analysis. We then present the formal statements of aforementioned and
related results and provide a more comprehensive discussion.
1.3. Organization of the paper. The rest of the paper is organized as
follows: Section 2 presents the asymptotic framework of our analysis. Section
3 discusses the main contributions of our paper. Section 4 compares our
results with the related work. Section 5 shows some simulation results and
discusses some open problems that require further research. Section 6 and
Supplementary material are devoted to the proofs of all the results.
2. The asymptotic framework. The main goal of this section is to
formally introduce the asymptotic setting under which we study LQLS. In
the current and next sections only, we may write vectors and matrices as
β(p), X(p), w(p) to emphasize the dependence on the dimension of β. Similarly, we may use β̂(λ, q, p) as a substitute for β̂(λ, q). Note that since we
assume n/p → δ, we do not include n in our notations. Now we define a
specific type of a sequence known as a converging sequence. Our definition
is borrowed from other papers [11, 12, 13] with some minor modifications.
Definition 1. A sequence of instances {β(p), X(p), w(p)} is called a
converging sequence if the following conditions hold:
8
HAOLEI WENG, ARIAN MALEKI AND LE ZHENG
- The empirical distribution3 of β(p) ∈ Rp converges weakly to a probability measure pβ with bounded second moment. Further, p1 kβ(p)k22
converges to the second moment of pβ .
- The empirical distribution of w(p) ∈ Rn converges weakly to a zero
2 . And, 1 kw(p)k2 → σ 2 .
mean distribution with variance σw
w
2
n
- The elements of X(p) are iid with distribution N (0, 1/n).
For each of the problem instances in a converging sequence, we solve the
LQLS problem (1.1) and obtain β̂(λ, q, p) as the estimator. The interest is
to evaluate the accuracy of this estimator. Below we define the asymptotic
mean square error.
Definition 2. Let β̂(λ, q, p) be the sequence of solutions of LQLS for the
converging sequence of instances {β(p), X(p), w(p)}. The asymptotic mean
square error is defined as the almost sure limit of
p
1X
AMSE(λ, q, σw ) , lim
|β̂i (λ, q, p) − βi (p)|2 ,
p→∞ p
i=1
where the subscript i is used to denote the ith component of a vector.
Note that we have suppressed δ and pβ in the notation of AMSE for
simplicity, despite the fact that the asymptotic mean square error depends
on them as well. In the above definition, we have assumed that the almost
sure limit exists. Under the current asymptotic setting, the existence of
AMSE can be proved. In fact we are able to derive the asymptotic limit for
general loss functions as presented in the following theorem.
Theorem 2.1. Consider a converging sequence {β(p), X(p), w(p)}. For
any given q ∈ [1, 2], suppose that β̂(λ, q, p) is the solution of LQLS defined
in (1.1). Then for any pseudo-Lipschitz function4 ψ : R2 → R, almost surely
p
(2.1)
1X
lim
ψ β̂i (λ, q, p), βi (p) = EB,Z [ψ(ηq (B + σ̄Z; χ̄σ̄ 2−q ), B)],
p→∞ p
i=1
where B and Z are two independent random variables with distributions pβ
and N (0, 1), respectively; the expectation EB,Z (·) is taken with respect to
3
It is the distribution that puts a point mass 1/p at each of the p elements of the vector.
A function ψ : R2 → R is pseudo-Lipschitz of order k if there exists a constant L > 0
such that for all x, y ∈ R2 , we have |ψ(x) − ψ(y)| ≤ L(1 + kxkk−1
+ kykk−1
)kx − yk2 . We
2
2
consider pseudo-Lipschitz functions with order 2 in this paper.
4
HIGHER ORDER ANALYSIS OF REGULARIZATION TECHNIQUES
9
both B and Z; ηq (·; ·) is the proximal operator for the function k · kqq 5 ; and
σ̄ and χ̄ satisfy the following equations:
(2.2)
(2.3)
1
σ̄ 2 = σω2 + EB,Z [(ηq (B + σ̄Z; χ̄σ̄ 2−q ) − B)2 ],
δ
1
λ = χ̄σ̄ 2−q 1 − EB,Z [ηq0 (B + σ̄Z; χ̄σ̄ 2−q )] ,
δ
where ηq0 (·; ·) denotes the derivative of ηq with respect to its first argument.
The result for q = 1 has been proved in [13]. The key ideas of the proof
for generalizing to q ∈ (1, 2] are similar to those of [13]. We describe the
main proof steps in Appendix J of the supplementary material. According
to Theorem 2.1, in order to calculate the asymptotic mean square error (or
any other loss) of β̂(λ, q, p), we have to solve (2.2) and (2.3) for (σ̄, χ̄). The
following lemma shows these two nonlinear equations have a unique solution.
Lemma 2. For any positive values of λ, δ, σw > 0, any random variable
B with finite second moment, and any q ∈ [1, 2], there exists a unique pair
(σ̄, χ̄) that satisfies both (2.2) and (2.3).
The proof of this lemma can be found in Appendix D of the supplementary
material. Theorem 2.1 provides the first step in our analysis of LQLS. We
first calculate σ̄ and χ̄ from (2.2) and (2.3). Then, incorporating σ̄ and χ̄ in
(2.1) gives the following expression for the asymptotic mean square error:
(2.4)
AMSE(λ, q, σw ) = EB,Z (ηq (B + σ̄Z; χ̄σ̄ 2−q ) − B)2 .
Given the distribution of B (the sparsity level included), the variance of
2 , the number of response variables (normalized by the number
the error σw
of predictors) δ, and the regularization parameter λ, it is straightforward
to write a computer program to find the solution of (2.2) and (2.3) and
then compute the value of AMSE. However, it is needless to say that this
approach does not shed much light on the performance of bridge regression
estimates, since there are many factors involved in the computation and
each affects the result in a non-trivial fashion. In this paper, we would like
to perform an analytical study on the solution of (2.2) and (2.3) and obtain
an explicit characterization of AMSE in the small-error regime.
3. Our main contributions.
5
Proximal operator of k · kqq is defined as ηq (u; χ) , arg minz 21 (u − z)2 + χ|z|q . For
further information on these functions, please refer to the supplementary material.
10
HAOLEI WENG, ARIAN MALEKI AND LE ZHENG
3.1. Optimal tuning of λ. The performance of LQLS, as defined in (1.1),
depends on the tuning parameter λ. In this paper, we consider the value of λ
that gives the minimum AMSE. Let λ∗,q denote the value of λ that minimizes
AMSE given in (2.4). Then LQLS is solved with this specific value of λ, i.e.,
(3.1)
1
β̂(λ∗,q , q, p) ∈ arg min ky − Xβk22 + λ∗,q kβkqq .
β 2
Note that this is the best performance that LQLS can achieve in terms of
the AMSE. Theorem 2.1 enables us to evaluate this optimal AMSE of LQLS
for every q ∈ [1, 2]. The key step is to compute the solution of (2.2) and (2.3)
with λ = λ∗,q . Since λ∗,q has to be chosen optimally, it seemingly causes an
extra complication for our analysis. However, as we show in the following
corollary, the study of Equations (2.2) and (2.3) can be simplified to some
extent.
Corollary 1. Consider a converging sequence {β(p), X(p), w(p)}. Suppose that β̂(λ∗,q , q, p) is the solution of LQLS defined in (3.1). Then for any
q ∈ [1, 2]
(3.2)
AMSE(λ∗,q , q, σw ) = min EB,Z (ηq (B + σ̄Z; χ) − B)2 ,
χ≥0
where B and Z are two independent random variables with distributions
pβ and N (0, 1), respectively; and σ̄ is the unique solution of the following
equation:
(3.3)
σ̄ 2 = σω2 +
1
min EB,Z [(ηq (B + σ̄Z; χ) − B)2 ].
δ χ≥0
The proof of Corollary 1 is shown in Appendix E of the supplementary
material. Corollary 1 enables us to focus the analysis on a single equation
(3.3), rather than two equations (2.2) and (2.3). The results we will present
in the next section are mainly based on investigating the solution of (3.3).
3.2. Analysis of AMSE. In this paper since we are focused on the sparsity structure of β, from now on we assume that the distribution, to which
the empirical distribution of β ∈ Rp converges, has the form
pβ (b) = (1 − )δ0 (b) + g(b),
where δ0 (·) denotes a point mass at zero, and g(·) is a generic distribution
that does not have any point mass at 0. Here, the mixture proportion ∈
(0, 1) is a fixed number that represents the sparsity level of β. The smaller
HIGHER ORDER ANALYSIS OF REGULARIZATION TECHNIQUES
11
is, the sparser β will be. The distribution g(b) specifies the values of non-zero
components of β. We will use G to denote a random variable having such
a distribution. Since our results and proof techniques look very different for
the case q > 1 and q = 1, we study these cases separately.
3.2.1. Results for q > 1. Our first result is concerned with the optimal
AMSE of LQLS for 1 < q ≤ 2, when the number of response variables is
larger than the number of predictors p, i.e., δ > 1.
Theorem 3.1. Suppose P(|G| ≤ t) = O(t) (as t → 0) and E|G|2 < ∞,
then for 1 < q < 2, δ > 1 and ∈ (0, 1), we have
(3.4)
AMSE(λ∗,q , q, σw ) =
2
σw
δ q+1 (1 − )2 (E|Z|q )2 2q
2q
).
−
σ + o(σw
1 − 1/δ
(δ − 1)q+1 E|G|2q−2 w
For q = 2, δ > 1 and ∈ (0, 1), if E|G|2 < ∞, we have
AMSE(λ∗,q , q, σw ) =
4
2
δ 3 σw
σw
4
−
+ o(σw
).
1 − 1/δ (δ − 1)3 E|G|2
Note that Z ∼ N (0, 1) and G ∼ g(·) are independent.
The proof of the result is presented in Appendix F of the supplementary
material. There are several interesting features of this result that we would
like to discuss: (i) The second dominant term of AMSE is negative. This
means that the actual AMSE is smaller than the one predicted by the first
order term, especially for smaller values of q. (ii) Neither the sparsity level
nor the distribution of the non-zero components of β appear in the first
2
σw
dominant term, i.e. 1−1/δ
. As we will discuss later in this section, the first
dominant term is the one that specifies the phase transition curve. Hence,
these calculations show a peculiar feature of phase transition analysis we
discussed in Section 1.2, that the phase transition of q ∈ (1, 2] is neither
affected by non-zero components of β or the sparsity level. However, we see
that both factors come into play in the second dominant term. (iii) For the
fully dense vector, i.e. = 1, (3.4) may imply that for 1 < q < 2,
AMSE(λ∗,q , q, σw ) =
2
σw
2q
+ o(σw
).
1 − 1/δ
Hence, we require a different analysis to obtain the second dominant term
(with different orders). We refer the interested readers to [14] for further
information about this case. (iv) For < 1, the choice of q ∈ (1, 2] does
12
HAOLEI WENG, ARIAN MALEKI AND LE ZHENG
not affect the first dominant term. That is the reason why all the values of
q ∈ (1, 2] share the same phase transition curve. However, the value of q has a
major impact on the second dominant term. In particular, as q approaches
1, the order of the second dominant term in terms of σw gets closer to
that of the first dominant term. This means that in any practical setting,
phase transition analysis may lead to misleading conclusions. Specifically, in
contrast to the conclusion from phase transition analysis that q ∈ (1, 2] have
the same performance, the second order expansion enables us to conclude
that the closer to 1 the value of q is, the better its performance will be. Our
next theorem discusses the AMSE when δ < 1.
Theorem 3.2.
(3.5)
Suppose E|G|2 < ∞, then for 1 < q ≤ 2 and δ < 1,
lim AMSE(λ∗,q , q, σw ) > 0.
σw →0
The proof of this theorem is presented in Appendix G of the supplementary material. Theorems 3.1 and 3.2 together show a notion of phase
2 ), and hence it will go to
transition. For δ > 1, as σw → 0, AMSE = O(σw
zero, while AMSE 9 0 for δ < 1. In fact, the phase transition curve δ = 1
can be derived from the first dominant term in the expansion of AMSE. If
δ = 1, the first dominant term is infinity and there will be no successful
recovery, while it becomes zero when σw = 0 if δ > 1. A more rigorous
justification can be found in the proof of Theorems 3.1 and 3.2. Therefore,
we may conclude that the first order term contains the phase transition
information. Moreover, the derived second order term offers us additional
important information regarding the accuracy of the phase transition analysis. To provide a comprehensive understanding of these two terms, in Section
5 we will evaluate the accuracy of first and second order approximations to
AMSE through numerical studies.
3.2.2. Results for q = 1. So far we have studied the case 1 < q ≤ 2. In
this section, we study q = 1, a.k.a. LASSO. In Theorems 3.1 and 3.2, we
have characterized the behavior of LQLS with q ∈ (1, 2] for a general class
of G. It turns out that the distribution of G has a more serious impact on
the second dominant term of AMSE for LASSO. We thus analyze it in two
different settings. Our first theorem considers the distributions that do not
have any mass around zero.
Theorem 3.3.
Suppose P(|G| > µ) = 1 with µ being a positive constant
HIGHER ORDER ANALYSIS OF REGULARIZATION TECHNIQUES
13
and E|G|2 < ∞, then for δ > M1 ()6
(3.6)
δM1 () 2
AMSE(λ∗,1 , 1, σw ) =
σw + o φ
δ − M1 ()
r
δ − M1 () µ̃
,
δ
σw
where µ̃ is any positive constant smaller than µ and φ(·) is the density
function of standard normal.
The proof of Theorem 3.3 is given in Section 6. Different from LQLS with
q ∈ (1, 2], we have not derived the exact analytical expression of second
dominant term for LASSO. However, since it is exponentially small, the
first order term (or phase transition analysis) is sufficient for evaluating
the performance of LASSO in the small-error regime. This will be further
confirmed by the numerical studies in Section 5. Below is our result for the
distributions of G that have more mass around zero.
Theorem 3.4. Suppose that P(|G| ≤ t) = Θ(t` ) (as t → 0) with ` > 0
and E|G|2 < ∞, then for δ > M1 (),
δM1 () 2
σ
δ − M1 () w
`/2
1
`+2
,
& −Θ(σw
) · log log . . . log
|
{z
} σw
`+2
−Θ(σw
) & AMSE(λ∗,1 , 1, σw ) −
m times
where m is an arbitrary but finite natural number, and a & b means a ≥ b
holds for sufficiently small σw .
The proof of this theorem can be found in Appendix H of the supplementary material. It is important to notice the difference between Theorems 3.3
and 3.4. The first point we would like to emphasize is that the first dominant
terms are the same in both cases. The second dominant terms are different
though. As we will show in Section 6 and Appendix H, similar to LQLS
for 1 < q ≤ 2, the second dominant terms are in fact negative. Hence, the
actual AMSE will be smaller than the one predicted by the first dominant
term. Furthermore, note that the magnitude of the second dominant term
in Theorem 3.4 is much larger than that in Theorem 3.3. This seems intuitive. LASSO tends to shrink the parameter coefficients towards zero, and
hence, if the true β has more mass around zero, the AMSE will be smaller.
The more mass the distribution of G has around zero, the better the second
order term will be. Our next theorem discusses what happens if δ < M1 ().
6
Recall M1 () = inf χ≥0 (1 − )Eη12 (Z; χ) + (1 + χ2 ) with Z ∼ N (0, 1).
14
HAOLEI WENG, ARIAN MALEKI AND LE ZHENG
Theorem 3.5.
(3.7)
Suppose that E|G|2 < ∞. Then for δ < M1 (),
lim AMSE(λ∗,1 , 1, σw ) > 0.
σw →0
The proof is presented in Appendix I of the supplementary material. Similarly as we discussed in Section 3.2.1, Theorems 3.3, 3.4 and 3.5 imply the
phase transition curve of LASSO. Such information can be obtained from
the first dominant term in the expansion of AMSE as well.
4. Related work.
4.1. Other phase transition analyses and n/p → δ asymptotic results.
The asymptotic framework that we considered in this paper evolved in a
series of papers by Donoho and Tanner [15, 16, 5, 17]. This framework was
used before on similar problems in engineering and physics [18, 19, 20].
Donoho and Tanner characterized the phase transition curve for LASSO and
some of its variants. Inspired by this framework, many researchers started
exploring the performance of different algorithms or estimates under this
asymptotic settings [10, 7, 21, 8, 22, 23, 24, 25, 9, 26, 3, 27, 28, 12, 13].
Our paper performs the analysis of LQLS under such asymptotic framework. Also, we adopt the message passing analysis that was developed in a
series of papers [11, 6, 29, 12, 13]. The notion of phase transition we consider is similar to the one introduced in [11]. However, there are three major
differences: (i) The analysis of [11] is performed for LASSO, while we have
generalized the analysis to any LQLS with 1 < q ≤ 2. (ii) The analysis of
[11] is performed on the least favorable distribution for LASSO, while here
we characterize the effect of the distribution of G on the AMSE as well.
(iii) Finally, [11] is only concerned with the first dominant term in AMSE of
LASSO, while we derive the second dominant term whose importance has
been discussed in the last few sections.
Another line of research that has connections with our analysis is presented in a series of papers [30, 31, 21]. In [21] the authors have derived a
minimax formulation that (if it has a unique solution and is solved) can give
an accurate characterization of the asymptotic mean square error. Compared
with Theorem 2.1 in our paper, that result works for more general penalized M-estimators, while Theorem 2.1 holds for general pseudo-Lipschitz loss
functions. When applying the minimax formulation in [21] to bridge regression, the AMSE formula in (2.4) can be recovered. However, the derivation
of phase transition curves for bridge regression under optimal tuning is not
found in [21]. Furthermore, [30, 31] proposed a geometric approach to characterize the risk of penalized least square estimates with general convex
HIGHER ORDER ANALYSIS OF REGULARIZATION TECHNIQUES
15
penalties. In particular, both papers obtained phase transition results based
on a key convex geometry quantity called “Gaussian squared-distences”.
However, [30] only rigorously proved the negative results (equivalent to Theorems 3.2 and 3.5) and left the positive part as a conjecture. The phase
transition results in [31] are concerned with the prediction errors ky − X β̂k22
and kXβ − X β̂k22 , rather than the estimation error kβ̂ − βk22 . Also, neither
of the two papers went beyond the first-order or phase transition analysis of
the risk.
Several researchers have also worked on the analysis of LQLS for q < 1 [32,
3, 27]. These analyses are based on non-rigorous, but widely accepted replica
method from statistical physics. The current paper extends the analysis of
[3] to q ≥ 1 case, makes the analysis rigorous by using the message passing
framework rather than the replica method, and finally provides a higher
order analysis.
4.2. Other analysis frameworks. One of the first papers that compared
the performance of penalization techniques is [33] which showed that there
exists a value of λ with which Ridge regression, i.e. LQLS with q = 2,
outperforms the vanilla least squares estimator. Since then, many more regularizers have been introduced to the literature each with a certain purpose.
For instance, we can mention LASSO [34], elastic net [35], SCAD [36], bridge
regression [1], and more recently SLOPE [37]. There has been a large body
of work on studying all these regularization techniques. We partition all the
work into the following categories and explain what in each category has
been done about the bridge regression:
(i) Simulation results: One of the main motivations for our work comes
from the nice simulation study of the bridge regression presented in
[2]. This paper finds the optimal values of λ and q by generalized cross
validation and compares the performance of the resulting estimator
with both LASSO and ridge. The main conclusion is that the bridge
regression can outperform both LASSO and ridge. Given our results
we see that if sparsity is present in β, then smaller values of q perform
better than ridge (in their second dominant term).
(ii) Asymptotic study: Knight and Fu [4] studied the asymptotic properties of bridge regression under the setting where n → ∞, while p is
fixed. They established the consistency and asymptotic normality of
the estimates under quite general conditions. Huang et al. [38] studied LQLS for q < 1 under a high-dimensional asymptotic setting in
which p grows with n but is still assumed to be less than n. They
not only derived the asymptotic distribution of the estimators, but
16
HAOLEI WENG, ARIAN MALEKI AND LE ZHENG
also proved LQLS has oracle properties in the sense of Fan and Li
[36]. They have also considered the case p > n, and have shown that
under partial orthogonality assumption on X, bridge regression distinguishes correctly between covariates with zero and non-zero coefficients. Note that under the asymptotic regime of our paper, both
LASSO and the other bridge estimators have false discoveries [39] and
possibly non-zero AMSE. Hence, they may not provide consistent estimates. We should also mention that the analysis of bridge regression
with q ∈ [0, 1) under the asymptotic regime n/p → δ is presented in
[3]. Finally, the performance of LASSO under a variety of conditions
has been studied extensively. We refer the reader to [40] for the review
of those results.
(iii) Non-asymptotic bounds: One of the successful approaches that has
been employed for studying the performance of regularization techniques such as LASSO is the minimax analysis [41], [42]. We refer the
reader to [40] for a complete list of references on this direction. In this
minimax approach, a lower bound for the prediction error or mean
square error of any estimation technique is first derived. Then a specific estimate, like the one returned by LASSO, is considered and an
upper bound is derived assuming the design matrices satisfy certain
conditions such as restrictive eigenvalue assumption [41, 43], restricted
isometry condition [44], or coherence conditions [45]. These conditions
can be confirmed for matrices with iid subgaussian elements. Based
on these evaluations, if the order of the upper bound for the estimate
under study matches the order of the lower bound, we can claim that
the estimate (e.g. LASSO) is minimax rate-optimal. This approach
has some advantages and disadvantages compared to our asymptotic
approach: (i) It works under more general conditions. (ii) It provides
information for any sample size. The price paid in the minimax analysis is that the constants derived in the results are usually not sharp
and hence many schemes have similar guarantees and cannot be compared to each other. Our asymptotic framework looses the generality
and in return gives sharp constants that can then be used in evaluating
and comparing different schemes as we do in this paper. Along similar
directions, [43] has studied the penalized empirical risk minimization
with `q penalties for the values of q ∈ [1, 1 + log1 p ] and has found upper bounds on the excess risk of these estimators (oracle inequalities).
Similar to minimax analysis, although the results of this analysis enjoy
generality, they suffer from loose constants that impede an accurate
comparisons of different bridge estimators.
HIGHER ORDER ANALYSIS OF REGULARIZATION TECHNIQUES
0.45
17
0.08
true value
first order approximation
0.4
true value
first order approximation
0.07
0.35
0.06
0.3
MSE
MSE
0.05
0.25
0.2
0.04
0.03
0.15
0.02
0.1
0.01
0.05
0
0
0.05
0.1
σw
0.15
0.2
0
0.25
0
0.05
(a)
0.1
σw
0.15
0.2
0.25
0.2
0.25
0.2
0.25
(b)
0.16
0.06
true value
first order approximation
0.14
true value
first order approximation
0.05
0.12
0.04
MSE
MSE
0.1
0.08
0.03
0.06
0.02
0.04
0.01
0.02
0
0
0.05
0.1
σw
0.15
0.2
0
0.25
0
0.05
(c)
0.15
0.06
true value
first order approximation
0.1
true value
first order approximation
0.05
0.04
MSE
0.08
MSE
σw
(d)
0.12
0.06
0.03
0.04
0.02
0.02
0.01
0
0.1
0
0.05
0.1
σw
(e)
0.15
0.2
0.25
0
0
0.05
0.1
σw
0.15
(f)
Fig 2. Plots of actual AMSE and its approximations for (a) δ = 1.1 and = 0.7, (b)
δ = 1.1 and = 0.25, (c) δ = 1.5 and = 0.7, (d) δ = 1.5 and = 0.25, (e) δ = 2 and
= 0.7, (f ) δ = 2 and = 0.25.
18
HAOLEI WENG, ARIAN MALEKI AND LE ZHENG
5. Numerical results and discussions.
5.1. Summary. The analysis of AMSE we presented in Section 3.2 is performed as σw → 0. For such asymptotic analysis, it would be interesting to
check the approximation accuracy of the first and second order expansions
of AMSE over a reasonable range of σw . Towards this goal, this section performs several numerical studies to (i) evaluate the accuracy of the first and
second order expansions discussed in Section 3.2, (ii) discover situations in
which the first order approximation is not accurate (for reasonably small
noise levels) while the second order expansion is, and (iii) identify situations
where both first and second orders are inaccurate and propose methods for
improving the approximations. Sections 5.2 and 5.3 study the performance
of LASSO and other bridge regression estimators with q > 1 respectively.
Finally, we should also mention that all the results presented in this paper
are concerned with the asymptotic setting n, p → ∞ and n/p → δ. To evaluate the accuracy of these results for finite sample sizes, we have performed
additional simulations whose results are presented in Appendix B of the
supplementary material.
5.2. LASSO. One of the conclusions from Theorem 3.3 is that the first
dominant term provides a good approximation of AMSE for the LASSO
problem when the distribution of G does not have a large mass around 0.
To test this claim we conduct the following numerical experiment. We set
the parameters of our problem instances in the following way:
1.
2.
3.
4.
δ can take any value in {1.1, 1.5, 2}.
can take values in {0.25, 0.7}.
σw ranges within the interval [0, 0.25].
the distribution of G is specified as g(b) = 0.5δ1 (b) + 0.5δ−1 (b), where
δa (·) denotes a point mass at point a.
We then use the formula in Corollary 1 to calculate AMSE(λ∗,1 , 1, σw ). Finally, we compare AMSE(λ∗,1 , 1, σw ), computed numerically from (3.2) and
(3.3), with its first order approximation provided in Theorem 3.3. The results of this experiment are summarized in Figure 2. As is clear in this figure,
the first order expansion gives a very good approximation for AMSE over a
large range of σw .
5.3. Bridge regression estimators with q > 1. In this numerical experiment, we would like to vary σw and see under what conditions our first order
or second order expansions can lead to accurate approximation of AMSE for
a wide range of σw . Throughout this section, we set the distribution of G
19
HIGHER ORDER ANALYSIS OF REGULARIZATION TECHNIQUES
0.08
0.08
true value
first order approximation
second order approximation
0.08
true value
first order approximation
second order approximation
0.07
0.06
0.05
0.05
MSE
0.06
0.05
0.04
0.04
0.04
0.03
0.03
0.03
0.02
0.02
0.02
0.01
0.01
0
0
0.05
0.1
σw
0.15
0.2
0.25
(a)
0
true value
first order approximation
second order approximation
0.07
0.06
MSE
MSE
0.07
0.01
0
0.05
0.1
σw
(b)
0.15
0.2
0.25
0
0
0.05
0.1
σw
0.15
0.2
0.25
(c)
Fig 3. Plots of actual AMSE and its approximations for (a) δ = 5, = 0.7, q = 1.5, (b)
δ = 4, = 0.7, q = 1.6, (c) δ = 5, = 0.6, q = 1.8.
to g(b) = 0.5δ1 (b) + 0.5δ−1 (b), as we did in Section 5.2. We then investigate different conditions by specifying various values of other parameters in
our problem instances. The expansion of AMSE for q > 1 is presented in
Theorem 3.1. For q ∈ (1, 2), recall the two terms in the expansion below
(5.1) AMSE(λ∗,q , q, σw ) =
2
σw
δ q+1 (1 − )2 (E|Z|q )2 2q
2q
−
σ + o(σw
).
1 − 1/δ
(δ − 1)q+1 E|G|2q−2 w
We expect the first order term to present a good approximation over a
reasonably large range of σw , when the second order term is sufficiently
small. According to the analytical form of the second order term in (5.1),
it is small if the following three conditions hold simultaneously: (i) δ is
not close to 1, (ii) is not small, and (iii) q is not close to 1. Our first
numerical result shown in Figure 3 is in agreement with this claim. In this
simulation we have set three different cases for δ, and q so that they satisfy
the above three conditions. The non-zero elements of β are independently
drawn from 0.5δ1 (b) + 0.5δ−1 (b). As demonstrated in this figure, the first
order term approximates AMSE accurately. Another interesting finding is
that the second order expansion provides an even better approximation.
To understand the limitation of the first order approximation, we consider
the cases in which the second order term is large and suggests that at least
the first order approximation is not necessarily good. This happens when
either δ decreases to 1, decreases to 0 or q decreases to 1. The settings of
our experiments and the results are summarized below.
1. We keep q = 1.5 and = 0.7 fixed and study different values of
δ ∈ {5, 2, 1.5, 1.1}. Figure 4 summarizes the results of this simulation. As is clear in this figure (and is consistent with the message of
the second dominant term), as we decrease δ the first order approximation becomes less accurate. The second order approximation in these
20
HAOLEI WENG, ARIAN MALEKI AND LE ZHENG
0.08
0.14
true value
first order approximation
second order approximation
0.07
true value
first order approximation
second order approximation
0.12
0.06
0.1
MSE
MSE
0.05
0.04
0.08
0.06
0.03
0.04
0.02
0.02
0.01
0
0
0.05
0.1
σw
0.15
0.2
0
0.25
0
0.05
(a)
0.1
σw
0.15
0.2
0.25
0.2
0.25
(b)
0.2
true value
first order approximation
second order approximation
0.18
0.16
0.5
0.14
MSE
0.12
MSE
true value
first order approximation
second order approximation
0.6
0.1
0.08
0.06
0.4
0.3
0.2
0.04
0.1
0.02
0
0
0.05
0.1
σw
(c)
0.15
0.2
0.25
0
0
0.05
0.1
σw
0.15
(d)
Fig 4. Plots of actual AMSE and its approximations for q = 1.5, = 0.7 with (a) δ = 5,
(b) δ = 2, (c) δ = 1.5 and (d) δ = 1.1.
cases is more accurate than the first order approximation. However
interestingly, the second order approximation becomes less accurate
as δ decreases too. These observations suggest that to have a good
approximation for the values of δ that are very close to 1, although
the second order approximation outperforms the first order, it may
not be sufficient and higher order terms are required. Such terms can
be derived with strategies similar to the ones we used in the proof of
Theorem 3.1. Note that the insufficiency of the second order expansion
partially results from the wide range of σw ∈ [0, 0.25]. If we evaluate
the approximation when σw is small enough, we will expect the success
of the second order expansion.
2. In our second simulation, we fix δ = 5, = 0.4 and let q ∈ {1.8, 1.5, 1.1}.
All the simulation results are summarized in Figure 5. As we expected,
the first order approximation becomes less accurate when q decreases.
Furthermore, we notice that when q is very close to 1 (check q = 1.1
in the figure), even the second order approximation is not necessarily
good. This again calls for higher order approximation of the AMSE.
HIGHER ORDER ANALYSIS OF REGULARIZATION TECHNIQUES
0.03
0.03
true value
first order approximation
second order approximation
0.025
MSE
0.02
0.015
0.015
0.01
0.01
0.01
0.005
0.005
0.005
0
0
0.05
σw
(a)
0.1
0.15
0
true value
first order approximation
second order approximation
0.025
0.02
MSE
0.02
MSE
0.03
true value
first order approximation
second order approximation
0.025
0.015
21
0
0.05
σw
(b)
0.1
0.15
0
0
0.05
σw
0.1
0.15
(c)
Fig 5. Plots of actual AMSE and its approximations for δ = 5, = 0.4 with (a) q = 1.8,
(b) q = 1.5, and (c) q = 1.1.
3. For the last simulation, we fix δ = 5, q = 1.8, and let ∈ {0.7, 0.5, 0.3, 0.1}.
Our simulation results are presented in Figure 6. We see that as decreases the first order approximation becomes less accurate. The second order approximation is always better than the first one. Moreover,
we observe that when is very close to 0 (check = 0.1 in the figure),
even the second order approximation is not necessarily sufficient. As
we discussed in the previous two simulations, we might need higher
order approximation of the AMSE in such cases.
5.4. Discussion. Firstly, our numerical studies confirm that the first order term gives good approximations of AMSE for LASSO in the case where
the distribution of non-zero elements of β is bounded away from zero. Secondly, as the numerical results for q > 1 demonstrate, while the second order
approximation always improves over the first order term and works well in
many cases, in the following situations it may not provide very accurate
evaluation of AMSE: (i) when δ is close to 1, (ii) is close to zero, and (iii)
q is close to 1. In such cases, the value of the second order term becomes
large and hence the approximation is only accurate for very small value of
σw . The remedy that one can propose is to derive higher order expansions.
Such terms can be calculated with the same strategy that we used to obtain
the second dominant term.
6. Proof of Theorem 3.3. Due to the limited space in the main text,
we only present the proof of Theorem 3.3, one of our main results for LASSO,
in this section. The proofs of all the other results are deferred to the supplementary material. Although some techniques used in the proofs are quite
different for LASSO and LQLS with q ∈ (1, 2], the roadmap remains the
same. Hence we suggest readers to first read the proof in this section. Once
22
HAOLEI WENG, ARIAN MALEKI AND LE ZHENG
0.03
0.03
true value
first order approximation
second order approximation
0.025
0.02
MSE
MSE
0.02
0.015
0.015
0.01
0.01
0.005
0.005
0
true value
first order approximation
second order approximation
0.025
0
0.05
σw
0.1
0
0.15
0
0.05
(a)
0.15
0.03
true value
first order approximation
second order approximation
true value
first order approximation
second order approximation
0.025
0.02
MSE
0.02
MSE
0.1
(b)
0.03
0.025
0.015
0.015
0.01
0.01
0.005
0.005
0
σw
0
0.05
σw
(c)
0.1
0.15
0
0
0.05
σw
0.1
0.15
(d)
Fig 6. Plots of actual AMSE and its approximations for δ = 5 and q = 1.8 with (a)
= 0.7, (b) = 0.5, (c) = 0.3, and (d) = 0.1.
this relatively simple proof is clear, the other more complicated proofs will
be easier to read.
6.1. Roadmap of the proof. Since the proof of this result has several steps
and is long, we lay out the roadmap of the proof here to help readers navigate through the details. According to Corollary 1 (let us accept Corollary
1 for the moment; its proof will be fully presented in Appendix E of the
supplement), in order to evaluate AMSE(λ∗,1 , 1, σw ) as σw → 0, the crucial
step is to characterize σ̄ from the following equation
(6.1)
σ̄ 2 = σω2 +
1
min EB,Z [(η1 (B + σ̄Z; χ) − B)2 ].
δ χ≥0
To study (6.1), the key part is to analyze the term minχ≥0 EB,Z [(η1 (B +
σ̄Z; χ)−B)2 ]. A useful fact that we will prove in Section 6.4 can simplify the
analysis of (6.1): The condition δ > M1 () implies that σ̄ → 0, as σw → 0.
Hence one of the main steps of this proof is to derive the convergence rate of
minχ≥0 EB,Z [(η1 (B + σZ; χ) − B)2 ], as σ → 0. Once we obtain that rate, we
then characterize the convergence rate for σ̄ as σw → 0 from (6.1). Finally
we connect σ̄ to AMSE(λ∗,1 , 1, σw ) based on Corollary 1, and derive the
HIGHER ORDER ANALYSIS OF REGULARIZATION TECHNIQUES
23
expansion for AMSE(λ∗,1 , 1, σw ) as σw → 0. We introduce the following
notations:
R(χ, σ) = EB,Z [(η1 (B/σ + Z; χ) − B/σ)2 ],
χ∗ (σ) = arg min R(χ, σ),
χ≥0
where we have suppressed the subscript B, Z in E for notational simplicity.
According to [46], R(χ, σ) is a quasi-convex function of χ and has a unique
global minimizer. Hence χ∗ (σ) is well defined. It is straightforward to confirm
min EB,Z [(η1 (B + σZ; χ) − B)2 ] = σ 2 R(χ∗ (σ), σ).
χ≥0
Throughout the proof, we may write χ∗ for χ∗ (σ) when no confusion is
caused, and we use F (g) to denote the distribution function of |G|. The rest
of the proof of Theorem 3.3 is organized in the following way:
1. We first prove R(χ∗ (σ), σ) → M1 (), as σ → 0 in Section 6.2.
2. We further bound the convergence rate of R(χ∗ (σ), σ) in Section 6.3.
3. We finally utilize the convergence rate bound derived in Section 6.3 to
characterize the convergence rate of σ̄ and then derive the expansion
for AMSE(λ∗,1 , 1, σw ) in Section 6.4.
6.2. Proof of R(χ∗ (σ), σ) → M1 (), as σ → 0. Our goal in this section
is to prove the following lemma.
Lemma 3.
Suppose E|G|2 < ∞, then limσ→0 χ∗ (σ) = χ∗∗ and
lim R(χ∗ (σ), σ) = (1 − )E(η1 (Z; χ∗∗ ))2 + (1 + (χ∗∗ )2 ),
σ→0
where χ∗∗ is the unique minimizer of (1 − )E(η1 (Z; χ))2 + (1 + χ2 ) over
[0, ∞), and Z ∼ N (0, 1).
Proof. By taking derivatives, it is straightforward to verify that (1 −
)E(η1 (Z; χ))2 + (1 + χ2 ), as a function of χ over [0, ∞), is strongly convex
and has a unique minimizer. Hence χ∗∗ is well defined.
We first claim that χ∗ (σn ) is bounded for any given sequence σn → 0.
Otherwise there exists an unbounded subsequence χ∗ (σnk ) → +∞ with
σnk → 0. Since the distribution of G does not have point mass at zero and
η1 (G/σnk + Z; χ∗ (σnk )) = sign(G/σnk + Z)(|G/σnk + Z| − χ∗ (σnk ))+ ,
it is not hard to conclude that
|η1 (G/σnk + Z; χ∗ (σnk )) − G/σnk | → +∞, a.s.
24
HAOLEI WENG, ARIAN MALEKI AND LE ZHENG
By Fatou’s lemma, we then have
(6.2) R(χ∗ (σnk ), σnk ) ≥ E(η1 (G/σnk + Z; χ∗ (σnk )) − G/σnk )2 → +∞.
On the other hand, the optimality of χ∗ (σnk ) implies
R(χ∗ (σnk ), σnk ) ≤ R(0, σnk ) = 1,
contradicting the unboundedness in (6.2).
We next show the sequence χ∗ (σn ) converges to a finite constant, for any
σn → 0. Taking a convergent subsequence χ∗ (σnk ), due to the boundedness
of χ∗ (σn ), the limit of the subsequence is finite. Call it χ̃. Note that
E(η1 (G/σnk + Z; χ∗ (σnk )) − G/σnk )2
= 1 + E(η1 (G/σnk + Z; χ∗ (σnk )) − G/σnk − Z)2 +
2EZ(η1 (G/σnk + Z; χ∗ (σnk )) − G/σnk − Z).
Since η1 (u; χ) = sign(u)(|u| − χ)+ , we have the following three inequalities:
|η1 (Z; χ∗ (σnk ))|2 ≤ |Z|2 ,
(η1 (G/σnk + Z; χ∗ (σnk )) − G/σnk − Z)2 ≤ (χ∗ (σnk ))2 ,
|Z(η1 (G/σnk + Z; χ∗ (σnk )) − G/σnk − Z)| ≤ |Z|χ∗ (σnk ).
Furthermore, all the terms on the right hand side of the above inequalities
are integrable. Therefore we can apply the Dominated Convergence Theorem
(DCT) to obtain
lim R(χ∗ (σnk ), σnk )
nk →∞
= lim (1 − )E(η1 (Z; χ∗ (σnk )))2 + E(η1 (G/σnk + Z; χ∗ (σnk )) − G/σnk )2
nk →∞
= (1 − )E(η1 (Z; χ̃))2 + (1 + χ̃2 ).
Moreover, since χ∗ (σnk ) is the optimal threshold value for R(χ, σnk ),
lim R(χ∗ (σnk ), σnk ) ≤ lim R(χ∗∗ , σnk ) = (1−)E(η1 (Z; χ∗∗ ))2 +(1+(χ∗∗ )2 )
nk →∞
nk →∞
Combining the last two limiting results, we can conclude χ̃ = χ∗∗ . Since
χ∗ (σnk ) is an arbitrary convergent subsequence, this implies that the sequence χ∗ (σn ) converges to χ∗∗ as well. This is true for any σn → 0, hence
χ∗ (σ) → χ∗∗ , as σ → 0. limσ→0 R(χ∗ (σ), σ) can then be directly derived.
HIGHER ORDER ANALYSIS OF REGULARIZATION TECHNIQUES
25
6.3. Bounding the convergence rate of R(χ∗ (σ), σ). In Section 6.2 we
have shown R(χ∗ (σ), σ) → M1 () as σ → 0. Our goal in this section is to
bound the difference R(χ∗ (σ), σ) − M1 (). For that purpose, we first bound
the convergence rate of χ∗ (σ).
Lemma 4. Suppose P(|G| ≥ µ) = 1 with µ being a positive constant and
E|G|2 < ∞, then as σ → 0
|χ∗ (σ) − χ∗∗ | = O(φ(−µ/σ + χ∗∗ )),
where φ(·) is the density function of the standard normal.
Proof. Since χ∗ (σ) minimizes R(χ, σ), we have
gives the following expression for χ∗ (σ):
χ∗ (σ) =
∂R(χ∗ (σ),σ)
∂χ
= 0, which
2(1 − )φ(χ∗ ) + Eφ(χ∗ − G/σ) + Eφ(χ∗ + G/σ)
.
R −χ∗ −G/σ
R∞
R∞
φ(z)dz
2(1 − ) χ∗ φ(z)dz + E χ∗ −G/σ φ(z)dz + E −∞
Letting σ go to zero on both sides in the above equation, we then obtain
χ∗∗ =
2(1 − )φ(χ∗∗ )
R∞
,
2(1 − ) χ∗∗ φ(z)dz +
where we have applied Dominated Convergence Theorem (DCT). To bound
|χ∗ (σ) − χ∗∗ |, we first bound the convergence rate of the terms in the expression of χ∗ (σ). A direct application of the mean value theorem leads to
φ(χ∗ ) − φ(χ∗∗ ) = (χ∗∗ − χ∗ )χ̃φ(χ̃),
Z ∞
Z ∞
˜
φ(z)dz −
φ(z)dz = (χ∗∗ − χ∗ )φ(χ̃),
(6.3)
(6.4)
χ∗
χ∗∗
˜ being two numbers between χ∗ and χ∗∗ . We now consider the
with χ̃, χ̃
other four terms. By the condition P(|G| ≥ µ) = 1, we can conclude that for
sufficiently small σ
(6.5)
Eφ(χ∗ − G/σ) ≤ Eφ(χ∗ − |G|/σ) ≤ φ(µ/σ − χ∗ ),
(6.6)
Eφ(χ∗ + G/σ) ≤ Eφ(χ∗ − |G|/σ) ≤ φ(µ/σ − χ∗ ).
Moreover, it is not hard to derive
Z ∞
Z
(6.7) 1 − E
φ(z)dz − E
Z
∞Z
χ∗ −G/σ
χ∗ −g/σ
−χ∗ −G/σ
φ(z)dz
−∞
Z
χ∗ −µ/σ
φ(z)dzdF (g) ≤
=
0
−χ∗ −g/σ
−χ∗ −µ/σ
φ(z)dz ≤ 2χ∗ φ(µ/σ − χ∗ ),
26
HAOLEI WENG, ARIAN MALEKI AND LE ZHENG
where to obtain the last two inequalities we have used the condition P(|G| ≥
µ) = 1 and the fact χ∗ − µ/σ < 0 for σ small enough. We are now in the
position to bound |χ∗ (σ) − χ∗∗ |. Define the following notations:
Z
e1 , E
∞
Z
−χ∗ −G/σ
φ(z)dz − ,
φ(z)dz + E
χ∗ −G/σ
∗
−∞
∗
e2 , Eφ(χ − G/σ) + Eφ(χ + G/σ),
Z ∞
S , 2(1 − )φ(χ∗∗ ), T , 2(1 − )
φ(z)dz + .
χ∗∗
Using the new notations and Equations (6.3) and (6.4), we obtain
χ∗ (σ) =
S + 2(1 − )(χ∗∗ − χ∗ )χ̃φ(χ̃) + e2
,
˜ + e1
T + 2(1 − )(χ∗∗ − χ∗ )φ(χ̃)
χ∗∗ =
S
.
T
Hence we can do the following calculations:
χ∗ (σ) − χ∗∗ =
=
=
(6.8)
S + 2(1 − )(χ∗∗ − χ∗ )χ̃φ(χ̃) + e2
S
−
∗∗
∗
˜ + e1
T
T + 2(1 − )(χ − χ )φ(χ̃)
∗∗
∗
2(1 − )(χ − χ )χ̃φ(χ̃) + e2
−
˜ + e1
T + 2(1 − )(χ∗∗ − χ∗ )φ(χ̃)
˜ + e1 )
S(2(1 − )(χ∗∗ − χ∗ )φ(χ̃)
˜ + e1 )
T (T + 2(1 − )(χ∗∗ − χ∗ )φ(χ̃)
˜
2(1 − )(χ∗∗ − χ∗ )(χ̃φ(χ̃) − χ∗∗ φ(χ̃))
+
˜ + e1
T + 2(1 − )(χ∗∗ − χ∗ )φ(χ̃)
e2 − χ∗∗ e1
.
˜ + e1
T + 2(1 − )(χ∗∗ − χ∗ )φ(χ̃)
From (6.8) we obtain
˜
2(1 − )(χ̃φ(χ̃) − χ∗∗ φ(χ̃))
˜ + e1
T + 2(1 − )(χ∗∗ − χ∗ (σ))φ(χ̃)
∗∗
e2 − χ e1
.
˜ + e1
T + 2(1 − )(χ∗∗ − χ∗ (σ))φ(χ̃)
(χ (σ) − χ ) 1 +
∗
(6.9)
=
∗∗
˜ → χ∗∗ since
Note that in the above expression we have χ̃ → χ∗∗ and χ̃
∗
∗∗
∗∗
˜
χ (σ) → χ . Therefore, we conclude that χ̃φ(χ̃) − χ φ(χ̃) → 0 and (χ∗∗ −
˜ → 0. Moreover, since (6.5), (6.6) and (6.7) together show both
χ∗ (σ))φ(χ̃)
e1 and e2 go to 0 exponentially fast, we conclude from (6.9) that (χ∗ (σ) −
HIGHER ORDER ANALYSIS OF REGULARIZATION TECHNIQUES
27
χ∗∗ )/σ → 0. This enables us to proceed
|e2 − χ∗∗ e1 |
|χ∗ (σ) − χ∗∗ |
|χ∗ (σ) − χ∗∗ | (a)
=
lim
=
lim
σ→0 T φ(µ/σ − χ∗ )
σ→0 φ(µ/σ − χ∗∗ )
σ→0 φ(µ/σ − χ∗ )
lim
(b)
≤
2(1 + χ∗ (σ)χ∗∗ )φ(µ/σ − χ∗ )
2(1 + (χ∗∗ )2 )
=
.
σ→0
T φ(µ/σ − χ∗ )
T
lim
We have used (6.9) to obtain (a). We derived (b) by the following steps:
1. According to (6.7), |e1 | ≤ 2χ∗ φ(µ/σ − χ∗ ).
2. According to (6.5) and (6.6), |e2 | ≤ 2φ(µ/σ − χ∗ ).
This completes the proof of Lemma 4.
The next step is to bound the convergence rate of R(χ∗ (σ), σ) based on
the convergence rate of χ∗ (σ) we have derived in Lemma 4.
Lemma 5. Suppose P(|G| ≥ µ) = 1 with µ being a positive constant and
E|G|2 < ∞, then as σ → 0
|R(χ∗ (σ), σ) − M1 ()| = O(φ(µ/σ − χ∗∗ )),
where φ(·) is the density function of the standard normal.
Proof. We recall the two quantities:
M1 () = (1 − )E(η1 (Z; χ∗∗ ))2 + (1 + (χ∗∗ )2 ),
(6.10)
R(χ∗ (σ), σ) = (1 − )E(η1 (Z; χ∗ ))2 +
[1 + E(η1 (G/σ + Z; χ∗ ) − G/σ − Z)2 ]
+2EZ(η1 (G/σ + Z; χ∗ ) − G/σ − Z).
(6.11)
We bound |R(χ∗ (σ), σ) − M1 ()| by bounding the difference between the
corresponding terms in (6.11) and (6.10). From the proof of Lemma 4 we
know e1 < 0 and e2 > 0. Hence (6.9) implies χ∗ (σ) > χ∗∗ for small enough
σ. We start with
|E(η1 (Z; χ∗ ))2 − E(η1 (Z; χ∗∗ ))2 |
(6.12)
=
|E(η1 (Z; χ∗ ) − η1 (Z; χ∗∗ ))(η1 (Z; χ∗ ) + η1 (Z; χ∗∗ ))|
≤
E[|χ∗ − χ∗∗ + χ∗ I(|Z| ∈ (χ∗∗ , χ∗ ))| · |η1 (Z; χ∗ ) + η1 (Z; χ∗∗ )|]
(a)
≤
2(χ∗ − χ∗∗ ) · E|Z| + 2χ∗ E[I(|Z| ∈ (χ∗∗ , χ∗ ))|Z|]
≤
2(χ∗ − χ∗∗ ) · E|Z| + 4χ∗ (χ∗ − χ∗∗ )χ̃φ(χ̃) = O(φ(µ/σ − χ∗∗ )),
28
HAOLEI WENG, ARIAN MALEKI AND LE ZHENG
where we have used the fact |η1 (u; χ)| ≤ |u| to obtain (a); χ̃ is a number
between χ∗ (σ) and χ∗∗ ; and the last equality is due to Lemma 4. We next
bound the difference between E(η1 (G/σ + Z; χ∗ ) − G/σ − Z)2 and (χ∗∗ )2 :
(6.13)
|(χ∗∗ )2 − E(η1 (G/σ + Z; χ∗ ) − G/σ − Z)2 |
≤ |(χ∗ )2 − E(η1 (G/σ + Z; χ∗ ) − G/σ − Z)2 | + |(χ∗∗ )2 − (χ∗ )2 |.
To bound the two terms on the right hand side of (6.13), first note that
0 ≤ (χ∗ )2 − E(η1 (G/σ + Z; χ∗ ) − G/σ − Z)2
= E[I(|G/σ + Z| ≤ χ∗ ) · ((χ∗ )2 − (G/σ + Z)2 )]
Z ∞ Z −g/σ+χ∗
φ(z)dzdF (g)
≤ (χ∗ )2
0
(b)
≤ (χ∗ )2
(6.14)
Z
−g/σ−χ∗
−µ/σ+χ∗
φ(z)dz ≤ 2(χ∗ )3 φ(µ/σ − χ∗ )
−µ/σ−χ∗
∗∗
= O(φ(µ/σ − χ )),
where (b) is due to the condition P(|G| ≥ µ) = 1, and the last equality holds
since (χ∗ − χ∗∗ )/σ → 0 implied by Lemma 4. Furthermore, Lemma 4 yields
(6.15)
(χ∗ )2 − (χ∗∗ )2 = O(φ(µ/σ − χ∗∗ )).
Combining (6.13), (6.14), and (6.15), we obtain
(6.16)
|(χ∗∗ )2 − E(η1 (G/σ + Z; χ∗ ) − G/σ − Z)2 | = O(φ(µ/σ − χ∗∗ )).
Regarding the remaining term in R(χ∗ (σ), σ), we can derive
0 ≤ EZ(G/σ + Z − η1 (G/σ + Z; χ∗ ))
(c)
= E(1 − ∂1 η1 (G/σ + Z; χ∗ ))
(6.17)
(d)
= P(|G/σ + Z| ≤ χ∗ ) = O(φ(µ/σ − χ∗∗ )).
We have employed Stein’s lemma (see Lemma 11 in the supplement) to
obtain (c). Equality (d) holds due to (6.7). Putting the results (6.12), (6.16),
and (6.17) together finishes the proof.
6.4. Deriving the expansion of AMSE(λ∗,1 , 1, σw ). In this section we utilize the convergence rate result of R(χ∗ (σ), σ) from Section 6.3 to derive the
expansion of AMSE(λ∗,1 , 1, σw ) in (3.6), and thus finish the proof of Theorem 3.3. Towards that goal, we first prove a useful lemma.
HIGHER ORDER ANALYSIS OF REGULARIZATION TECHNIQUES
Lemma 6.
(6.18)
29
Let σ̄ be the solution to the following equation:
σ̄ 2 = σω2 +
1
min EB,Z [(η1 (B + σ̄Z; χ) − B)2 ].
δ χ≥0
Suppose δ > M1 (), then
2
σw
δ − M1 ()
=
.
2
σw →0 σ̄
δ
lim
Proof. We first claim that E(η1 (α + Z; χ) − α)2 is an increasing function
of α, because
d
E(η1 (α + Z; χ) − α)2 = 2E(αI(|α + Z| ≤ χ)) ≥ 0.
dα
Hence we obtain
(6.19) E(η1 (α + Z; χ) − α)2 ≤ lim E(η1 (α + Z; χ) − α)2 = 1 + χ2 .
α→∞
Inequality (6.19) then yields
R(χ, σ̄) = (1 − )E(η1 (Z; χ))2 + E(η1 (G/σ̄ + Z; χ) − G/σ̄)2
≤ (1 − )E(η1 (Z; χ))2 + (1 + χ2 ).
Taking minimum over χ on both sides above gives us
R(χ∗ (σ̄), σ̄) ≤ M1 ().
(6.20)
Moreover, since σ̄ is the solution of (6.18), it satisfies
2
σ̄ 2 = σw
+
(6.21)
σ̄ 2
R(χ∗ (σ̄), σ̄).
δ
Combining (6.20) and (6.21) with the condition δ > M1 (), we have
σ̄ 2 ≤
2
σw
,
1 − M1 ()/δ
which leads to σ̄ → 0, as σw → 0. Then applying Lemma 3 shows
lim R(χ∗ (σ̄), σ̄) = lim R(χ∗ (σ̄), σ̄) = M1 ().
σw →0
σ̄→0
Diving both sides of (6.21) by σ̄ 2 and letting σw → 0 finishes the proof.
30
HAOLEI WENG, ARIAN MALEKI AND LE ZHENG
To complete the proof of Theorem 3.3, first note that Corollary 1 tells us
AMSE(λ∗,1 , 1, σw ) = σ̄ 2 R(χ∗ (σ̄), σ̄),
2
σw
= σ̄ 2 −
σ̄ 2
R(χ∗ (σ̄), σ̄).
δ
We then have
(6.22)
δM1 () 2
σ
δ − M1 () w
δM1 ()
σ̄ 2
2
∗
2
∗
= σ̄ R(χ (σ̄), σ̄) −
· σ̄ − R(χ (σ̄), σ̄)
δ − M1 ()
δ
δ(R(χ∗ (σ̄), σ̄) − M1 ()) 2 (a)
σ̄ = O(σ̄ 2 φ(µ/σ̄ − χ∗∗ )),
=
δ − M1 ()
AMSE(λ∗,1 , 1, σw ) −
2
where (a) is due to Lemma 5. Finally, since limσw →0 σσ̄w2 = δ−Mδ 1 () according
to Lemma 6, it is not hard to see
r
δ − M () µ̃
1
(6.23) O(σ̄ 2 φ(µ/σ̄ − χ∗∗ )) = o(φ(µ̄/σ̄)) = o φ
,
δ
σw
where µ̄ and µ̃ are any constants satisfying 0 ≤ µ̃ < µ̄ < µ. Results (6.22)
and (6.23) together close the proof of Theorem 3.3.
Remark: (6.20) and (6.22) together imply that the second dominant term
of AMSE(λ∗,1 , 1, σw ) is in fact negative.
Acknowledgements. This work is supported by NSF grant CCF-1420328.
References.
[1] L. Frank and J. Friedman. A statistical view of some chemometrics regression tools.
Technometrics, 35(2):109–135, 1993.
[2] W. Fu. Penalized regressions: the bridge versus the lasso. J. Comp. Graph. Stat,
7(3):397–416, 1998.
[3] L. Zheng, A. Maleki, H. Weng, X. Wang, and T. Long. Does `p -minimization outperform `1 -minimization? Preprint. arXiv:1501.03704v2, 2016.
[4] K. Knight and W. Fu. Asymptotics for lasso-type estimators. Annals of statistics,
pages 1356–1378, 2000.
[5] D. Donoho and J. Tanner. Sparse nonnegative solution of underdetermined linear
equations by linear programming. Proc. Natl. Acad. Sci., 102(27):9446–9451, 2005.
[6] D. Donoho, A. Maleki, and A. Montanari. Message-passing algorithms for compressed
sensing. Proc. Natl Acad. Sci., 106(45):18914–18919, 2009.
[7] D. Amelunxen, M. Lotz, M. McCoy, and J. Tropp. Living on the edge: Phase transitions in convex programs with random data. Info. Infer., 2014.
[8] N. El Karoui, D. Bean, P. Bickel, C. Lim, and B. Yu. On robust regression with
high-dimensional predictors. Proc. Natl Acad. Sci., 110(36):14557–14562, 2013.
[9] J. Bradic and J. Chen. Robustness in sparse linear models: relative efficiency based
on robust approximate message passing. arXiv preprint arXiv:1507.08726, 2015.
HIGHER ORDER ANALYSIS OF REGULARIZATION TECHNIQUES
31
[10] M. Stojnic. Various thresholds for `1 -optimization in compressed sensing. arXiv
preprint arXiv:0907.3666, 2009.
[11] D. L. Donoho, A. Maleki, and A. Montanari. Noise sensitivity phase transition. IEEE
Trans. Inform. Theory, 57(10):6920–6941, Oct. 2011.
[12] M. Bayati and A. Montanri. The dynamics of message passing on dense graphs, with
applications to compressed sensing. IEEE Trans. Inform. Theory, 57(2), Feb. 2011.
[13] M. Bayati and A. Montanari. The LASSO risk for Gaussian matrices. IEEE Trans
Inform Theory, 58(4):1997–2017, 2012.
[14] H. Weng, A. Maleki, and L. Zheng. Overcoming The Limitations of Phase Transition
by Higher Order Analysis of Regularization Techniques. ArXiv e-prints, March 2016.
[15] D. Donoho. For most underdetermined systems of linear equations, the minimal `1 norm approximates the sparsest near-solution. Comm. Pure and Appl. Math, 2004.
[16] D. Donoho. High-dimensional centrally symmetric polytopes with neighborliness
proportional to dimension. Discrete & Computational Geometry, 35(4):617–652, 2006.
[17] D. Donoho and J. Tanner. Neighborliness of randomly projected simplices in high
dimensions. Proc. Natl Acad. Sci., 102(27):9452–9457, 2005.
[18] D. Guo and S. Verdú. Randomly spread cdma: Asymptotics via statistical physics.
IEEE Trans. Inform. Theory, 51(6):1983–2010, 2005.
[19] T. Tanaka. A statistical-mechanics approach to large-system analysis of cdma multiuser detectors. IEEE Trans. Info. Theory, 48(11):2888–2910, 2002.
[20] A. Coolen. The Mathematical Theory of Minority Games: Statistical Mechanics of
Interacting Agents (Oxford Finance Series). Oxford University Press, Inc., 2005.
[21] C. Thrampoulidis, E. Abbasi, and B. Hassibi. Precise error analysis of regularized
m-estimators in high-dimensions. arXiv preprint arXiv:1601.06233, 2016.
[22] N. Karoui. Asymptotic behavior of unregularized and ridge-regularized highdimensional robust regression estimators: rigorous results.
arXiv preprint
arXiv:1311.2445, 2013.
[23] D. Donoho and A. Montanari. High dimensional robust m-estimation: Asymptotic
variance via approximate message passing. Prob. Theo. Rel. Fields, pages 1–35, 2013.
[24] D. Donoho, M. Gavish, and A. Montanari. The phase transition of matrix recovery
from gaussian measurements matches the minimax mse of matrix denoising. Proceedings of the National Academy of Sciences, 110(21):8405–8410, 2013.
[25] D. Donoho and A. Montanari. Variance breakdown of huber (m)-estimators: n/p →
m. 2015.
[26] D. Donoho, A. Maleki, and A. Montanari. The noise-sensitivity phase transition in
compressed sensing. IEEE Trans. Info. Theory, 57(10), 2011.
[27] S. Rangan, V. Goyal, and A. Fletcher. Asymptotic analysis of map estimation via
the replica method and compressed sensing. In Advances in Neural Information
Processing Systems, pages 1545–1553, 2009.
[28] F. Krzakala, M. Mézard, F. Sausset, Y. Sun, and L. Zdeborová. Statistical-physicsbased reconstruction in compressed sensing. Physical Review X, 2(2):021005, 2012.
[29] A. Maleki. Approximate message passing algorithms for compressed sensing. Ph.D.
thesis, Stanford University, 2010.
[30] Samet Oymak, Christos Thrampoulidis, and Babak Hassibi. The squared-error of
generalized lasso: A precise analysis. In Communication, Control, and Computing
(Allerton), 2013 51st Annual Allerton Conference on, pages 1002–1009. IEEE, 2013.
[31] Samet Oymak and Babak Hassibi. Sharp mse bounds for proximal denoising. Foundations of Computational Mathematics, 16(4):965–1029, 2016.
[32] Y. Kabashima, T. Wadayama, and T. Tanaka. A typical reconstruction limit for
compressed sensing based on lp-norm minimization. Journal of Statistical Mechanics:
32
HAOLEI WENG, ARIAN MALEKI AND LE ZHENG
Theory and Experiment, 2009(09):L09003, 2009.
[33] A. Hoerl and R. Kennard. Ridge regression: Biased estimation for nonorthogonal
problems. Technometrics, 12(1):55–67, 1970.
[34] R. Tibshirani. Regression shrinkage and selection via the lasso. Journal of the Royal
Statistical Society. Series B, pages 267–288, 1996.
[35] H. Zou and T. Hastie. Regularization and variable selection via the elastic net.
Journal of the Royal Statistical Society: Series B, 67(2):301–320, 2005.
[36] J. Fan and R. Li. Variable selection via nonconcave penalized likelihood and its oracle
properties. Journal of the American statistical Association, 96(456):1348–1360, 2001.
[37] M. Bogdan, E. van den Berg, C. Sabatti, W. Su, and E. Candès. Slope:adaptive
variable selection via convex optimization. Annals App. Stat., 9(3):1103, 2015.
[38] J. Huang, J. Horowitz, and S. Ma. Asymptotic properties of bridge estimators in
sparse high-dimensional regression models. Annals Stat., pages 587–613, 2008.
[39] W. Su, M. Bogdan, and E. Candes. False discoveries occur early on the lasso path.
arXiv preprint arXiv:1511.01957, 2015.
[40] P. Bühlmann and S. Van De Geer. Statistics for high-dimensional data: methods,
theory and applications. Springer Science & Business Media, 2011.
[41] P. Bickel, Y. Ritov, and A. Tsybakov. Simultaneous analysis of lasso and dantzig
selector. Annals Stat., pages 1705–1732, 2009.
[42] G. Raskutti, M. Wainwright, and B. Yu. Minimax rates of estimation for highdimensional linear regression over-balls. IEEE Trans. Info. Theory, 57(10), 2011.
[43] V. Koltchinskii. Sparsity in penalized empirical risk minimization. In Annales de
l’IHP Probabilités et statistiques, volume 45, pages 7–57, 2009.
[44] E. Candès. The restricted isometry property and its implications for compressed
sensing. Comptes Rendus Mathematique, 346(9):589–592, 2008.
[45] F. Bunea, A. Tsybakov, and M. Wegkamp. Sparsity oracle inequalities for the lasso.
Electronic Journal of Statistics, 1:169–194, 2007.
[46] Ali Mousavi, Arian Maleki, and Richard G Baraniuk. Consistent parameter estimation for lasso and approximate message passing. Annals of Statistics. To appear.
[47] N. Parikh and S. Boyd. Proximal algorithms. Foundations and Trends in optimization, 1(3):127–239, 2014.
[48] C. Stein. Estimation of the mean of a multivariate normal distribution. The Annals
of Statistics, pages 1135–1151, 1981.
[49] ZD Bai and YQ Yin. Limit of the smallest eigenvalue of a large dimensional sample
covariance matrix. The annals of Probability, pages 1275–1294, 1993.
[50] Boris Sergeevich Kashin. Diameters of some finite-dimensional sets and classes of
smooth functions. Izvestiya Rossiiskoi Akademii Nauk. Seriya Matematicheskaya,
41(2):334–351, 1977.
HIGHER ORDER ANALYSIS OF REGULARIZATION TECHNIQUES
33
Supplementary material
APPENDIX A: ORGANIZATION
This supplement contains simulations and all the proof of lemmas and
theorems that have not been covered in the main text. We outline the structure of the supplement to help readers find the materials they are interested
in. The organization is as follows:
1. Appendix B contains additional simulation results. These numerical
experiments evaluate the accuracy of AMSE(λ∗,q , q, σw ) as an asymptotic prediction for finite-sample mean square error (MSE). Along this
line, the convergence rate of MSE is empirically verified.
2. Appendix C covers several important properties of the proximal operator function ηq (u; χ). These properties will later be used in our proofs.
It includes Lemmas 7, 8, 9, 10 and their proofs.
3. Appendix D proves Lemma 2 from the main text. This lemma is concerned with the uniquesness of the solution of (2.2) and (2.3). Lemmas
7, 8, 9, 10 from Appendix C are extensively used here.
4. Appendix E proves Corollary 1 in the main text. This corollary characterizes AMSE when the regularization parameter λ is optimally tuned.
This appendix uses Lemma 12 from Appendix D.
5. Appendix F includes the proof of Theorem 3.1, one of the main results in this paper. The theorem derives the second order expansion of
AMSE(λ∗,q , q, σw ) for q ∈ (1, 2]. We recommend interested readers to
read the proof in Section 6 before Appendix F.
6. Appendix G contains the proof of Theorem 3.2. This theorem identifies
the necessary condition for successful recovery with q ∈ (1, 2]. Phase
transition is implied by this theorem together with Theorem 3.1.
7. Appendix H proves Theorem 3.4. The proof of this theorem is along
the same lines as the proof of Theorem 3.3 presented in Section 6 of the
main text. We suggest the reader to study Section 6 before studying
this section.
8. Appendix I proves Theorem 3.5. The proof is essentially the same as
the proof of Theorem 3.2. Since we do not repeat the detailed arguments, readers may want to study Appendix G first.
34
HAOLEI WENG, ARIAN MALEKI AND LE ZHENG
9. Appendix J includes the proof of Theorem 2.1. Though Theorem 2.1
lays the first step towards analyzing the AMSE, we postpone the proof
to the end of the supplement. That is because the focus of this paper
is on the second order dominant term of AMSE and we do not want to
move the proof of this theorem ahead to potentially digress from the
highlights. Moreover, since the key ideas of this proof are similar to
those of [13], we will not detail out the entire proof and rather describe
several key steps. Note that in this proof we will use some results we
have developed in Appendices C and D.
APPENDIX B: ADDITIONAL NUMERICAL RESULTS
This section contains additional numerical studies to examine the relation
between AMSE and its finite-sample version. Throughout the simulations,
we fix the sparsity level = 0.4 and the distribution of G to g(b) = 0.5δ1 (b)+
0.5δ−1 (b). We also let the elements of X be i.i.d from N (0, 1/n), and the
i.i.d
2 ). The other parameters of our problem instances are
noise wi ∼ N (0, σw
set in the following way:
1.
2.
3.
4.
δ takes values in {1.5, 3, 5}
p takes any value in {20, 50, 200, 400}
q is chosen to be one of {1, 1.5, 1.8, 2}
σw ranges over [0, 0.15]
Given the values of , G, δ, σw and q, we compute AMSE(λ∗,q , q, σw ) according to the formulas in Corollary 1 from the main text. We also generate sami.i.d
ples from linear models by setting n = δp, {βi }pi=1 ∼ (1−)δ0 +G. We then
calculate the finite-sample MSE (under optimal tuning): inf λ≥0 kβ̂(λ, q) −
βk22 /p. The same experiment is repeated 200 times. For completeness, we
have included the first-order and second-order expansions of AMSE. The results are shown in Figures 7-15. We summarize our observations and findings
below.
• Figures 7-9 are about the case δ = 5. Figure 7 shows that the averaged
MSE is well matched with AMSE even when p is around 200 or 400.
This is true for different values of q as σw varies over [0, 0.15]. Even
for p = 20, the averaged MSE is already quite close to AMSE when
σw < 0.075. As confirmed by the simulations of Section 5 in the main
text, we see again that the first-order approximation for LASSO is very
accurate, and so is the second-order approximation for other values of
q.
HIGHER ORDER ANALYSIS OF REGULARIZATION TECHNIQUES
35
2 over the
• Figure 8 draws the standard deviation of MSE against σw
2
range [0, 0.15 ]. The plots with different values of q and p all indicate
that the standard deviation scales linearly with the noise variance
2 . Intuitively speaking, due to the quadratic form of MSE, its linear
σw
2 is somewhat expected.
scaling with σw
p
• In Figure 9, we multiply the standard deviation (SD) by p/20 and
2 . Interestingly, all the adjusted SD’s
plot the adjusted SD against σw
(with the same value of q) are aligned around the same straight line.
√
It implies that the scaling of SD on the dimension p is 1/ p, and the
√
convergence rate of MSE is n (recall n = δp). We leave a rigorous
proof of this empirical result for future research.
• Figures 10-12 are concerned with the case δ = 3. Compared with
Figures 7-9, it is seen that the observations we have made for δ = 5
still hold in δ = 3. We hence do not repeat the details.
• Figures 13-15 are for the scenario δ = 1.5. As the value of δ is decreased
to 1.5, the main messages we reveal in the case δ = 5 carry over. But
some details may change. For instance, only when p = 400 the difference between the averaged MSE and AMSE is negligible throughout
all the values of q and σw . Also, the linear pattern between SD (or
2 is less significant than when δ = 5. Given that
adjusted SD) and σw
for a fixed p smaller values of δ lead to smaller number of observations
(and hence less number of random elements in the design matrix),
these changes should not come as a surprise. In the language of statistical physics, since the the number of random elements in the system
decreases self averaging has not happened yet.
APPENDIX C: PRELIMINARIES ON ηq (u; χ)
This section is devoted to the properties of ηq (u; χ) defined as
1
ηq (u; χ) , arg min (u − z)2 + χ|z|q .
2
z
We start with some basic properties of these functions. Since the explicit
forms of ηq (u; χ) for q = 1 and 2 are known: η1 (u; χ) = (|u|−χ)sign(u)I(|u| >
u
χ), η2 (u; χ) = 1+2χ
, we focus our study on the case 1 < q < 2.
Lemma 7.
ηq (u; χ) satisfies the following properties:
(i) u − ηq (u; χ) = χqsign(u)|ηq (u; χ)|q−1 .
(ii) |ηq (u; χ)| ≤ |u|.
36
HAOLEI WENG, ARIAN MALEKI AND LE ZHENG
true value
first order approximation
second order approximation
empirical (p=20)
empirical (p=50)
empirical (p=200)
empirical (p=400)
MSE
0.020
0.020
0.015
0.010
0.010
0.000
0.000
0.005
MSE
q=1.5
0.030
q=1
true value
first order approximation
empirical (p=20)
empirical (p=50)
empirical (p=200)
empirical (p=400)
0.00
0.05
σw
0.10
0.15
0.00
0.05
q=1.8
0.15
0.10
0.15
0.000
0.010
MSE
0.020
true value
first order approximation
second order approximation
empirical (p=20)
empirical (p=50)
empirical (p=200)
empirical (p=400)
0.010
0.020
0.10
q=2
true value
first order approximation
second order approximation
empirical (p=20)
empirical (p=50)
empirical (p=200)
empirical (p=400)
0.000
MSE
σw
0.00
0.05
σw
0.10
0.15
0.00
0.05
σw
Fig 7. Plots of actual AMSE, its approximations, and finite-sample MSE. The MSE is
averaged over 200 times. The setup is δ = 5, = 0.4, g(b) = 0.5δ1 (b) + 0.5δ−1 (b).
(iii)
(iv)
(v)
(vi)
limχ→0 ηq (u; χ) = u and limχ→∞ ηq (u; χ) = 0.
ηq (−u; χ) = −ηq (u; χ).
For α > 0, we have ηq (αu; α2−q χ) = αηq (u; χ).
|ηq (u; χ) − ηq (ũ, χ)| ≤ |u − ũ|.
Proof. To prove (i), we should take the derivative of 21 (u − z)2 + χ|z|q
and set it to zero. Proofs of parts (ii), (iii) and (iv) are straightforward and
HIGHER ORDER ANALYSIS OF REGULARIZATION TECHNIQUES
q=1.5
0.008
p=20
p=50
p=200
p=400
SD
0.006
p=20
p=50
p=200
p=400
0.000
0.000
0.002
0.004
0.004
0.002
SD
0.006
0.008
q=1
0.000
0.005
0.010
σ2w
0.015
0.020
0.000
0.005
0.010
σ2w
0.015
0.020
0.015
0.020
q=2
p=20
p=50
p=200
p=400
SD
0.000
0.000
0.002
0.002
0.004
0.004
0.006
0.008
p=20
p=50
p=200
p=400
0.006
0.008
q=1.8
SD
37
0.000
0.005
0.010
σ2w
0.015
0.020
0.000
0.005
0.010
σ2w
Fig 8. Plots of the standard deviation of finite-sample MSE. The same experiment is
repeated 200 times. The setup is δ = 5, = 0.4, g(b) = 0.5δ1 (b) + 0.5δ−1 (b).
are hence skipped. To prove (v), note that
(C.1)
1
ηq (αu; α2−q χ) = arg min (αu − z)2 + χα2−q |z|q
z 2
α2
= arg min (u − z/α)2 + χα2 |z/α|q
z
2
1
= α arg min (u − z̃)2 + χ|z̃|q = αηq (u; χ).
z̃ 2
(vi) is a standard property of proximal operators of convex functions [47].
In many proofs, we will be dealing with derivatives of ηq (u; χ). For notational simplicity, we may use ∂1 ηq (u; χ), ∂12 ηq (u; χ), ∂2 ηq (u; χ), ∂22 ηq (u; χ)
38
HAOLEI WENG, ARIAN MALEKI AND LE ZHENG
q=1.5
0.008
0.006
p=20
p=50
p=200
p=400
0.000
0.000
0.002
0.004
Adjusted SD
0.006
0.004
0.002
Adjusted SD
0.008
q=1
p=20
p=50
p=200
p=400
0.000
0.005
0.010
σ2w
0.015
0.020
0.000
0.005
0.015
0.020
0.015
0.020
q=2
0.004
0.006
p=20
p=50
p=200
p=400
0.002
0.004
Adjusted SD
0.006
0.008
p=20
p=50
p=200
p=400
0.000
0.000
0.002
Adjusted SD
0.008
q=1.8
0.010
σ2w
0.000
0.005
0.010
σ2w
0.015
0.020
0.000
0.005
0.010
σ2w
Fig 9. Plots of the adjusted standard deviation of finite-sample MSE. The same experiment
is repeated 200 times. The setup is δ = 5, = 0.4, g(b) = 0.5δ1 (b) + 0.5δ−1 (b).
∂η (u,χ) ∂ 2 η (u,χ) ∂η (u,χ) ∂ 2 η (u,χ)
q
q
to represent q∂u , ∂u
, q∂χ , ∂χ
, respectively. Our next two
2
2
lemmas are concerned with differentiability of ηq (u; χ) and its derivatives.
Lemma 8. For every 1 < q < 2, ηq (u; χ) is a differentiable function of
(u, χ) for u ∈ R and χ > 0 with continuous partial derivatives. Moreover,
∂2 ηq (u; χ) is differentiable with respect to u, for any given χ > 0.
Proof. We start with the case u0 , χ0 > 0. The goal is to prove that
ηq (u; χ) is differentiable at (u0 , χ0 ). Since u0 > 0, ηq (u0 ; χ0 ) will be positive.
Then Lemma 7 part (i) shows ηq (u0 ; χ0 ) must satisfy
(C.2)
ηq (u0 ; χ0 ) + χ0 qηqq−1 (u0 ; χ0 ) = u0 .
39
HIGHER ORDER ANALYSIS OF REGULARIZATION TECHNIQUES
q=1
MSE
0.030
true value
first order approximation
second order approximation
empirical (p=20)
empirical (p=50)
empirical (p=200)
empirical (p=400)
0.010
0.020
0.020
0.015
0.010
0.000
0.000
0.005
MSE
q=1.5
true value
first order approximation
empirical (p=20)
empirical (p=50)
empirical (p=200)
empirical (p=400)
0.00
0.05
σw
0.10
0.15
0.00
0.05
q=1.8
0.15
0.10
0.15
0.000
0.000
0.010
MSE
0.030
true value
first order approximation
second order approximation
empirical (p=20)
empirical (p=50)
empirical (p=200)
empirical (p=400)
0.020
0.030
0.020
0.10
q=2
true value
first order approximation
second order approximation
empirical (p=20)
empirical (p=50)
empirical (p=200)
empirical (p=400)
0.010
MSE
σw
0.00
0.05
σw
0.10
0.15
0.00
0.05
σw
Fig 10. Plots of actual AMSE, its approximations, and finite-sample MSE. The MSE is
averaged over 200 times. The setup is δ = 3, = 0.4, g(b) = 0.5δ1 (b) + 0.5δ−1 (b).
Define the function F (u, χ, v) = u−v−χqv q−1 . Equation (C.2) says F (u, χ, v)
is equal to zero at (u0 , χ0 , ηq (u0 ; χ0 )). It is straightforward to confirm that
the derivative of F (u, χ, v) with respect to v is nonzero at (u0 , χ0 , ηq (u0 ; χ0 )).
By implicit function theorem, we can conclude ηq (u; χ) is differentiable at
(u0 , χ0 ). Lemma 7 part (iv) implies that the same result holds when u0 < 0.
We now focus on the point (0, χ0 ). Since ηq (0, χ0 ) = 0, we obtain
|ηq (u; χ0 )|
|u|1/(q−1)
≤ lim
= 0.
u→0
u→0 (χ0 q)1/(q−1) |u|
|u|
∂1 ηq (0; χ0 ) = lim
where the last inequality comes from (C.2). It is straightforward to see that
the partial derivative of ηq (u; χ) with respect to χ at (0, χ0 ) exists and
40
HAOLEI WENG, ARIAN MALEKI AND LE ZHENG
0.008
SD
0.004
0.000
0.005
0.010
σ2w
0.015
0.020
0.000
0.005
q=1.8
0.010
σ2w
0.015
0.020
0.015
0.020
q=2
p=20
p=50
p=200
p=400
0.008
p=20
p=50
p=200
p=400
SD
0.000
0.000
0.004
0.004
SD
0.008
0.012
q=1.5
p=20
p=50
p=200
p=400
0.000
SD
0.000 0.002 0.004 0.006 0.008 0.010
q=1
p=20
p=50
p=200
p=400
0.000
0.005
0.010
σ2w
0.015
0.020
0.000
0.005
0.010
σ2w
Fig 11. Plots of the standard deviation of finite-sample MSE. The same experiment is
repeated 200 times. The setup is δ = 3, = 0.4, g(b) = 0.5δ1 (b) + 0.5δ−1 (b).
is equal to zero as well. So far we have proved that ηq (u, χ) has partial
derivatives with respect to both u and χ for every u ∈ R, χ > 0. We next
show the partial derivatives are continuous. For u 6= 0, the result comes
directly from the implicit function theorem, because F (u, χ, v) is a smooth
function when v 6= 0. We now turn to the proof when u = 0. By taking
derivative with respect to u on both sides of (C.2), we obtain
(C.3)
∂1 ηq (u; χ) + χq(q − 1)ηqq−2 (u; χ)∂1 ηq (u; χ) = 1,
for any u, χ > 0. Moreover, it is clear from (C.2) that ηq (u; χ) → 0, as
HIGHER ORDER ANALYSIS OF REGULARIZATION TECHNIQUES
0.008
0.004
0.000
Adjusted SD
p=20
p=50
p=200
p=400
0.005
0.010
σ2w
0.015
0.020
0.000
0.005
0.010
σ2w
q=1.8
0.015
0.020
0.015
0.020
q=2
p=20
p=50
p=200
p=400
Adjusted SD
0.004
0.000
0.000
0.008
p=20
p=50
p=200
p=400
0.008
0.012
0.000
Adjusted SD
q=1.5
p=20
p=50
p=200
p=400
0.004
Adjusted SD
0.000 0.002 0.004 0.006 0.008 0.010
q=1
41
0.000
0.005
0.010
σ2w
0.015
0.020
0.000
0.005
0.010
σ2w
Fig 12. Plots of the adjusted standard deviation of finite-sample MSE. The same experiment is repeated 200 times. The setup is δ = 3, = 0.4, g(b) = 0.5δ1 (b) + 0.5δ−1 (b).
(u, χ) → (0+ , χ0 ). This fact combined with (C.3) yields
lim
(u,χ)→(0+ ,χ
0)
∂1 ηq (u; χ) =
1
lim
(u,χ)→(0+ ,χ
0)
1 + χq(q − 1)ηqq−2 (u; χ)
= 0.
Since ∂1 ηq (u; χ) = ∂1 ηq (−u; χ) implied by Lemma 7 part (iv), we conclude
lim
(u,χ)→(0,χ0 )
∂1 ηq (u; χ) = 0.
The same approach can prove that the partial derivative ∂2 ηq (u; χ) is continuous at (0, χ0 ). For simplicity we do not repeat the arguments.
42
HAOLEI WENG, ARIAN MALEKI AND LE ZHENG
q=1
0.06
true value
first order approximation
second order approximation
empirical (p=20)
empirical (p=50)
empirical (p=200)
empirical (p=400)
MSE
0.00
0.000
0.02
0.04
0.030
0.020
0.010
MSE
q=1.5
true value
first order approximation
empirical (p=20)
empirical (p=50)
empirical (p=200)
empirical (p=400)
0.00
0.05
σw
0.10
0.15
0.00
0.05
q=1.8
0.15
0.10
0.15
0.04
0.06
true value
first order approximation
second order approximation
empirical (p=20)
empirical (p=50)
empirical (p=200)
empirical (p=400)
0.00
0.00
0.02
MSE
0.04
0.06
0.10
q=2
true value
first order approximation
second order approximation
empirical (p=20)
empirical (p=50)
empirical (p=200)
empirical (p=400)
0.02
MSE
σw
0.00
0.05
σw
0.10
0.15
0.00
0.05
σw
Fig 13. Plots of actual AMSE, its approximations, and finite-sample MSE. The MSE is
averaged over 200 times. The setup is δ = 1.5, = 0.4, g(b) = 0.5δ1 (b) + 0.5δ−1 (b).
We now prove the second part of the lemma. Because F (u, χ, v) is infinitely many times differentiable in any open set with v 6= 0, implicit function theorem further implies ∂2 ηq (u; χ) is differentiable at any u 6= 0. The
rest of the proof is to show its differentiability at u = 0. This follows by
noting ∂2 ηq (0; χ) = 0, and
∂2 ηq (u; χ)
u→0
u
lim
(a)
=
=
−q|ηq (u; χ)|q−1
u→0 |u|(1 + χq(q − 1)|ηq (u; χ)|q−2 )
−(|u| − |ηq (u; χ)|)
= 0,
lim
u→0 χ|u|(1 + χq(q − 1)|ηq (u; χ)|q−2 )
lim
where (a) is by taking derivative with respect to χ on both sides of (C.2),
HIGHER ORDER ANALYSIS OF REGULARIZATION TECHNIQUES
q=1.5
SD
0.000
0.000
0.005
0.010
0.010
0.005
SD
p=20
p=50
p=200
p=400
0.015
0.020
p=20
p=50
p=200
p=400
0.015
0.020
q=1
0.000
0.005
0.010
σ2w
0.015
0.020
0.000
0.005
0.010
σ2w
0.015
0.020
0.015
0.020
q=2
p=20
p=50
p=200
p=400
SD
0.010
0.000
0.000
0.005
0.005
0.010
0.015
0.020
p=20
p=50
p=200
p=400
0.015
0.020
q=1.8
SD
43
0.000
0.005
0.010
σ2w
0.015
0.020
0.000
0.005
0.010
σ2w
Fig 14. Plots of the standard deviation of finite-sample MSE. The same experiment is
repeated 200 times. The setup is δ = 1.5, = 0.4, g(b) = 0.5δ1 (b) + 0.5δ−1 (b).
and the last two equalities above are due to Lemma 7 part (i) and (iii).
Lemma 9. Consider a given χ > 0, then for every 1 < q < 3/2,
∂1 ηq (u; χ) is a differentiable function of u for u ∈ R with continuous derivative; for q = 3/2, it is a weakly differentiable function of u; for 3/2 < q < 2,
∂1 ηq (u; χ) is differentiable at u 6= 0, but is not differentiable at zero.
Proof. As is clear from the proof of Lemma 8, the implicit function
theorem guarantees that ∂1 ηq (u; χ) is differentiable at u 6= 0 with continuous
derivative for 1 < q < 2. Hence we will be focused on u = 0. In the proof of
44
HAOLEI WENG, ARIAN MALEKI AND LE ZHENG
q=1.5
0.020
0.015
p=20
p=50
p=200
p=400
0.000
0.000
0.005
0.010
Adjusted SD
0.015
0.010
0.005
Adjusted SD
0.020
q=1
p=20
p=50
p=200
p=400
0.000
0.005
0.010
σ2w
0.015
0.020
0.000
0.005
0.015
0.020
0.015
0.020
q=2
0.015
p=20
p=50
p=200
p=400
0.000
0.000
0.005
0.010
0.010
Adjusted SD
0.015
0.020
p=20
p=50
p=200
p=400
0.005
Adjusted SD
0.020
q=1.8
0.010
σ2w
0.000
0.005
0.010
σ2w
0.015
0.020
0.000
0.005
0.010
σ2w
Fig 15. Plots of the adjusted standard deviation of finite-sample MSE. The same experiment is repeated 200 times. The setup is δ = 1.5, = 0.4, g(b) = 0.5δ1 (b) + 0.5δ−1 (b).
Lemma 8, we have derived
(C.4)
∂1 ηq (u; χ) =
1
,
1 + χq(q − 1)|ηq (u; χ)|q−2
for u 6= 0,
and ∂1 ηq (0; χ) = 0. We thus know
1
.
u→0 u + χq(q − 1)u|ηq (u; χ)|q−2
∂12 ηq (0; χ) = lim
(C.5)
Moreover, Lemma 7 part (i) implies
(C.6)
|ηq (u; χ)|2−q
u
= 1.
=
1
+
lim
u→0 χq|ηq (u; χ)|q−1 sign(u)
u→0
χq
lim
HIGHER ORDER ANALYSIS OF REGULARIZATION TECHNIQUES
45
For 1 < q < 3/2, (C.5) and (C.6) together give us
∂12 ηq (0; χ) = lim
u→0
1
q−2
= 0.
u + χq(q − 1)u(|u|/(χq)) q−1
We can also calculate the limit of ∂12 ηq (u; χ) (this second derivative can be
obtained from (C.4)) as follows.
−χq(q − 1)(q − 2)|ηq (u; χ)|q−3 sign(u)
= 0.
u→0
(1 + χq(q − 1)|ηq (u; χ)|q−2 )3
lim ∂12 ηq (u; χ) = lim
u→0
Therefore, ∂1 ηq (u; χ) is continuously differentiable on (−∞, +∞) for 1 <
q < 3/2. Regarding 3/2 < q < 2, Similar calculations yield
lim ∂12 ηq (0; χ) = +∞,
lim ∂12 ηq (0; χ) = −∞.
u→0−
u→0+
Finally to prove the weak differentiability for q = 3/2, we show ∂1 ηq (u; χ)
is a Lipschitz continuous function on (−∞, +∞). Note that for u 6= 0,
|∂12 ηq (u; χ)| =
χq(q − 1)(2 − q)|ηq (u; χ)|q−3
8
≤
,
(1 + χq(q − 1)|ηq (u; χ)|q−2 )3
9χ2
and ∂12 ηq (0+ ; χ) = −∂12 (0− ; χ) =
8
.
9χ2
Mean value theorem leads to
|∂1 ηq (u; χ) − ∂1 ηq (ũ; χ)| ≤
8
|u − ũ|,
9χ2
for uũ ≥ 0.
When uũ < 0, we can have
|∂1 ηq (u; χ) − ∂1 ηq (ũ; χ)| = |∂1 ηq (u; χ) − ∂1 ηq (−ũ; χ)|
8
8
≤
|u + ũ| ≤
|u − ũ|.
2
9χ
9χ2
This completes the proof of the lemma.
The last lemma in this section presents some additional properties regarding the derivatives of ηq (u; χ).
Lemma 10.
The derivatives of ηq (u; χ) satisfy the following properties:
(i) ∂1 ηq (u; χ) =
1
.
1+χq(q−1)|ηq (u;χ)|q−2
q−1
−q|ηq (u;χ)|
sign(u)
.
1+χq(q−1)|ηq (u;χ)|q−2
(ii) ∂2 ηq (u; χ) =
(iii) 0 ≤ ∂1 ηq (u; χ) ≤ 1.
46
HAOLEI WENG, ARIAN MALEKI AND LE ZHENG
(iv) For u > 0, ∂12 ηq (u; χ) > 0.
(v) |ηq (u; χ)| is a decreasing function of χ.
(vi) limχ→∞ ∂1 ηq (u; χ) = 0.
Proof. Parts (i) (ii) have been derived in the proof of Lemma 8. Part
(iii) is a simple conclusion of part (i). Part (iv) is clear from the proof of
Lemma 9. Part (v) is a simple application of part (ii). Finally, part (vi) is
an application of part(i) of Lemma 10 and part (iii) of Lemma 7.
APPENDIX D: PROOF OF LEMMA 2
D.1. Roadmap. The goal of this section is to show that Equations
(2.2) and (2.3) (rewritten below) have a unique solution (σ̄, χ̄):
1
σ̄ 2 = σω2 + EB,Z [(ηq (B + σ̄Z; χ̄σ̄ 2−q ) − B)2 ],
δ
1
0
2−q
2−q
λ = χ̄σ̄
1 − EB,Z [ηq (B + σ̄Z; χ̄σ̄ )] .
δ
Towards the goal we pursue the following two main steps:
1. We show the existence of the solution in Section D.2. In order to do
that, we first study the solution of Equation (2.2), and demonstrate
that for any χ ∈ (χmin , ∞) (χmin is a constant we will clarify later),
there exists a unique σχ such that (σχ , χ) satisfies (2.2). The proof will
be presented in Section D.2.1. We then show that by varying χ over
(χmin , ∞), the range of the value of the following term
1
χσχ 2−q 1 − EB,Z [ηq0 (B + σχ Z; χσχ2−q )]
δ
covers the number λ from Equation (2.3). That means Equations (2.2)
and (2.3) share at least one common solution pair (σχ , χ). We detail
out the proof in Section D.2.2.
2. We prove the uniqueness of the solution in Section D.3. The key idea
is to apply Theorem 2.1 to evaluate the asymptotic loss of the LQLS
estimates under two different pseudo-Lipschitz functions. These two
quantities determine the uniqueness of both σχ and χ in the common
solution pair (σχ , χ). Note that we have denoted this unique pair by
(σ̄, χ̄).
Before we start the details of the proof, we present Stein’s lemma [48]
that will be used several times in this paper.
HIGHER ORDER ANALYSIS OF REGULARIZATION TECHNIQUES
47
Lemma 11. Let g : R → R denote a weakly differentiable function. If
Z ∼ N (0, 1) and E|g 0 (Z)| < ∞, we have
E(Zg(Z)) = E(g 0 (Z)),
where g 0 denotes the weak-derivative of g.
D.2. Proving the existence of the solution of (2.2) and (2.3).
D.2.1. Studying the solution of (2.2). We first define a function that is
closely related to Equation (2.2):
Rq (χ, σ) , EB,Z [ηq (B/σ + Z; χ) − B/σ]2 .
(D.1)
Note that we have used the same definition for LASSO in Section 6. Here we
adopt a general notation Rq (χ, σ) to represent the function defined above
for any q ∈ [1, 2].
Lemma 12.
For 1 ≤ q ≤ 2, Rq (χ, σ) is a decreasing function of σ > 0.
Proof. We consider four different cases: (i) q = 1, (ii) q = 2, (iii) 1 <
q ≤ 3/2, (iv) 3/2 < q < 2.
(i) q = 1: Since R1 (χ, σ) is a differentiable function of σ, we will prove
(χ,σ)
this case by showing that ∂R1∂σ
< 0. We have
∂R1 (χ, σ)
∂σ
2
E [B(I(|B/σ + Z| > χ) − 1)(η1 (B/σ + Z; χ) − B/σ)]
σ2
2
= − 2 E I(|B/σ + Z| ≤ χ)(B 2 /σ) < 0.
σ
= −
The first equality above is due to Dominated Convergence Theorem.
u
(ii) q = 2: Since η2 (u; χ) = 1+2χ
, we have
∂R2 (χ, σ)
8χ2 E|B|2
=−
< 0.
∂σ
(1 + 2χ)2 σ 3
(iii) 1 < q ≤ 3/2: The strategy for this case is similar to that of the last
∂R (χ,σ)
two cases. We show that the derivative q∂σ
< 0.
∂Rq (χ, σ) (a)
= 2E (ηq (B/σ + Z; χ) − B/σ)(∂1 ηq (B/σ + Z; χ) − 1)(−B/σ 2 )
∂σ
= 2E (ηq (B/σ + Z; χ) − B/σ − Z)(∂1 ηq (B/σ + Z; χ) − 1)(−B/σ 2 )
(D.2) +2E Z(∂1 ηq (B/σ + Z; χ) − 1)(−B/σ 2 ) .
48
HAOLEI WENG, ARIAN MALEKI AND LE ZHENG
To obtain Equality (a), we have used Dominated Convergence Theorem (DCT); We employed Lemma 7 part (vi) to confirm the conditions
of DCT. Our goal is to show that the two terms in (D.2) are both negative. Regarding the first term, we first evaluate it by conditioning on
B = b for a given constant b > 0 (note that B and Z are independent):
EZ [(ηq (b/σ + Z; χ) − b/σ − Z)(∂1 ηq (b/σ + Z; χ) − 1)]
Z +∞
(b)
(ηq (z; χ) − z)(∂1 ηq (z; χ) − 1)φ(z − b/σ)dz
=
−∞
Z ∞
(ηq (z; χ) − z)(∂1 ηq (z; χ) − 1)φ(z − b/σ)dz +
=
0
Z 0
(ηq (z; χ) − z)(∂1 ηq (z; χ) − 1)φ(z − b/σ)dz
−∞
Z ∞
(d)
(c)
=
(ηq (z; χ) − z)(∂1 ηq (z; χ) − 1)(φ(z − b/σ) − φ(z + b/σ))dz > 0,
0
where φ(·) is the density function of standard normal; (b) is obtained
by a change of variables; (c) is due to the fact ∂1 ηq (−z; χ) = ∂1 ηq (z; χ)
implied by Lemma 7 part (iv); (d) is based on the following arguments:
According to Lemmas 7 part (ii) and Lemma 10 part (iii), ηq (z; χ) < z
and ∂1 ηq (z; χ) < 1 for z > 0. Moreover, φ(z − b/σ) − φ(z + b/σ) > 0
for z, b/σ > 0. Hence we have
EZ (ηq (b/σ + Z; χ) − b/σ − Z)(∂1 ηq (b/σ + Z; χ) − 1)(−b/σ 2 ) < 0.
Similarly we can show the above inequality holds for b < 0. It is clear
that the term on the left hand side equals zero when b = 0. Thus we
have proved the first term in (D.2) is negative. Now we should discuss
the second term. Again we condition on B = b for a given b > 0:
(e)
EZ [Z(∂1 ηq (b/σ + Z; χ) − 1)] = E(∂12 ηq (b/σ + Z; χ))
Z ∞
(D.3) =
[∂12 ηq (z; χ)(φ(z − b/σ) − φ(z + b/σ))]dz > 0.
0
Equality (e) is the result of Stein’s lemma, i.e. Lemma 11. Note that the
weak differentiability condition required in Stein’s lemma is guaranteed
by Lemma 9. To obtain the last inequality, we have used Lemma 10
part (iv) and the fact that φ(z − b/σ) − φ(z + b/σ) > 0 for z, b/σ > 0.
Hence we obtain that
EZ Z(∂1 ηq (b/σ + Z; τ ) − 1)(−b/σ 2 ) < 0.
HIGHER ORDER ANALYSIS OF REGULARIZATION TECHNIQUES
49
The same approach would work for b < 0, and clearly the left hand side
term of the above inequality equals zero for b = 0. We can therefore
conclude the second term in (D.2) is negative as well.
(iv) 3/2 < q < 2: The proof of this case is similar to the last one. The only
difference is that the proof steps we presented in (D.3) may not work,
due to the non-differentiability of ∂1 ηq (u; χ) for q > 3/2 as shown in
Lemma 9. Our goal here is to use an alternative approach to prove:
EZ [Z(∂1 ηq (b/σ + Z; χ) − 1)] > 0 for b > 0. We have
Z ∞
z(∂1 ηq (b/σ + z; χ) − 1)φ(z)dz
EZ [Z(∂1 ηq (b/σ + Z; χ) − 1)] =
−∞
Z ∞
z(∂1 ηq (b/σ + z; χ) − ∂1 ηq (b/σ − z; χ))φ(z)dz
=
Z0 ∞
(D.4) =
z(∂1 ηq (|b/σ + z|; χ) − ∂1 ηq (|b/σ − z|; χ))φ(z)dz,
0
where the last equality is due to the fact ∂1 ηq (u; χ) = ∂1 ηq (|u|; χ) for
any u ∈ R. Since |b/σ − z| < |b/σ + z| for z, b/σ > 0 and according to
Lemma 10 part (iv), we obtain
(D.5)
∂1 ηq (|b/σ + z|; χ) − ∂1 ηq (|b/σ − z|; χ) > 0.
Combining (D.4) and (D.5) completes the proof.
Lemma 12 paves our way in the study of the solution of (2.2). Define
(D.6)
1
χmin = inf χ ≥ 0 : E(ηq2 (Z; χ)) ≤ 1 ,
δ
where Z ∼ N (0, 1). The following corollary is a conclusion from Lemma 12.
Corollary 2.
For a given 1 ≤ q ≤ 2, Equation (2.2):
1
σ 2 = σω2 + EB,Z [(ηq (B + σZ; χσ 2−q ) − B)2 ],
δ
σw > 0
has a unique solution σ = σχ for any χ ∈ (χmin , ∞), and does not have any
solution if χ ∈ (0, χmin ).
Proof. First note that since σw > 0, σ = 0 is not a solution of (2.2).
Hence we can equivalently write Equation (2.2) in the following form:
(D.7)
1=
2
σw
1
+ Rq (χ, σ) , F (σ, χ).
σ2
δ
50
HAOLEI WENG, ARIAN MALEKI AND LE ZHENG
According to Lemma 12, F (σ, χ) is a strictly decreasing function of σ over
(0, ∞). We also know that F (σ, χ) is a continuous function of σ from the
proof of Lemma 12. Moreover, it is straightforward to confirm that
(D.8)
1
lim F (σ, χ) = E(ηq2 (Z; χ)).
σ→∞
δ
lim F (σ, χ) = ∞,
σ→0
Thus Equation (2.2) has a solution (the uniqueness is automatically guaranteed by the monotonicity of F (σ, χ)) if and only if 1δ E(ηq2 (Z; χ)) < 1. Recall
the definition of χmin given in (D.6). Since E(ηq2 (Z; χ)) is a strictly decreasing and continuous function of χ, 1δ E(ηq2 (Z; χ)) < 1 holds if χ ∈ (χmin , ∞)
and fails when χ ∈ (0, χmin ).
D.2.2. Studying the common solution of (2.2) and (2.3). Corollary 2
from Section D.2.1 characterizes the existence and uniqueness of solution
for Equation (2.2). Our next goal is to prove that (2.2) and (2.3) share at
least one common solution. Our strategy is: among all the pairs (σχ , χ) that
satisfy (2.2), we show that at least one of them satisfies (2.3). We do this in
the next few lemmas.
Lemma 13. Let δ < 1.
the value of σ that satisfies
2−q
lim χσχ
1−
(D.9)
χ→∞
2−q
lim χσχ
1−
χ→χ+
min
For each value of χ ∈ (χmin , ∞), define σχ as
(2.2). Then,
1
2−q
E[∂1 ηq (B + σχ Z; χσχ )]
= ∞,
δ
1
2−q
E[∂1 ηq (B + σχ Z; χσχ )]
= −∞.
δ
Proof. We first show that
2
lim σχ2 = σw
+
(D.10)
χ→∞
E|B|2
,
δ
lim σχ = ∞.
χ→χ+
min
2
2 + E|B| for any sequence
For the first part, we only need to show σχ2 k → σw
δ
χk → ∞. For that purpose, we first prove σχ = O(1). Otherwise, there exists
a sequence χn → ∞ such that σχn → ∞. Because
(ηq (B/σχn + Z; χn ) − B/σχn )2 ≤ 2(B/σχn + Z)2 + 2B 2 /σχ2 n ≤ 6B 2 + 4Z 2
for large enough n, we can apply Dominated Convergence Theorem (DCT)
to conclude
lim Rq (χn , σχn ) = E lim [ηq (B/σχn + Z; χn ) − B/σχn ]2 = 0.
n→∞
n→∞
HIGHER ORDER ANALYSIS OF REGULARIZATION TECHNIQUES
51
On the other hand, since the pair (σχn , χn ) satisfies (2.2) we obtain
lim Rq (χn , σχn ) = lim δ 1 −
n→∞
n→∞
2
σw
= δ.
σχ2 n
This is a contradiction. We next consider any convergent subsequence {σχkn }
of {σχk }. The facts σxk ≥ σw and σχk = O(1) imply σχkn → σ ∗ ∈ (0, ∞).
Moreover, since
) − B)2 ≤ 6B 2 + 5(σ ∗ )2 Z 2 ,
(ηq (B + σχkn Z; χkn σχ2−q
kn
when n is large enough. We can apply DCT to obtain,
2
2 1
2
2 EB
(σ ∗ )2 = lim σχ2 kn = σw
+ E lim (ηq (B+σχkn Z; χkn σχ2−q
)−B)
=
σ
+
.
w
kn
n→∞
δ n→∞
δ
Thus, we have showed any convergent subsequence of {σχ2 k } converges to
2 + EB 2 . Hence the sequence converges to that limit as well.
the same limit σw
δ
Regarding the second part in (D.10), if it is not the case, then there exists
a sequence χn → χ+
min such that σχn = O(1). Equation (D.7) shows,
1=
2
(a) σ 2
σw
1
1
σ2
1
+ Rq (χn , σχn ) ≥ 2w + Rq (χn , ∞) = 2w + E(ηq2 (Z; χn )),
2
σ χn
δ
σ χn
δ
σχn
δ
where (a) is due to Lemma 12. From the definition of χmin in (D.6), it is clear
that 1δ E(ηq2 (Z; χmin )) = 1 when δ < 1. Hence letting n → ∞ on the both
sides of the above inequaitiy leads to 1 ≥ Ω(1) + 1, which is a contradiction.
We are in position to derive the two limiting results in (D.9). To obtain
2
2 + E|B| , as χ → ∞. Therefore, Lemma 10
the first one, note that σχ2 → σw
δ
part (vi) combined with DCT gives us
lim E∂1 ηq (B + σχ Z; χσχ2−q ) = 0.
χ→∞
The first result of (D.9) can then be trivially derived. Regarding the second
result, we have showed that as χ → χ+
min , σχ → ∞. We also have
(b)
E∂1 ηq (B + σχ Z; χσχ2−q ) =
(c)
1
E(Zηq (B + σχ Z; χσχ2−q ))
σχ
= E(Zηq (B/σχ + Z; χ)),
52
HAOLEI WENG, ARIAN MALEKI AND LE ZHENG
where (b) holds by Lemma 11 and (c) is due to Lemma 7 part (v). Hence
lim E∂1 ηq (B + σχ Z; χσχ2−q )
χ→χ+
min
=
lim E(Zηq (B/σχ + Z; χ)) = E(Zηq (Z; χmin ))
χ→χ+
min
(d)
=
E(ηq2 (Z; χmin )) + χmin qE(|ηq (Z; χmin )|q )
=
δ + χmin qE(|ηq (Z; χmin )|q ),
where (d) is the result of Lemma 7 part (i). We thus obtain
1
1
2−q
(D.11) lim
1 − E∂1 ηq (B + σχ Z; χσχ ) = − χmin qE|ηq (Z; χmin )|q .
+
δ
δ
χ→χmin
Combining (D.11) and the fact that χmin > 0, σχ → ∞ finishes the proof.
Lemma 14. Let δ ≥ 1. For each value of χ ∈ (χmin , ∞), define σχ as
the value of σ that satisfies (2.2). Then,
1
lim χσχ2−q 1 − E[∂1 ηq (B + σχ Z; χσχ2−q )]
(D.12)
= ∞,
χ→∞
δ
1
2−q
2−q
= 0.
lim χσχ
1 − E[∂1 ηq (B + σχ Z; χσχ )]
δ
χ→χ+
min
Proof. The exactly same arguments presented in the proof of Lemma
13 can be applied to prove the first result in (D.12). We now focus on the
proof of the second one. Since E|∂1 ηq (B + σχ Z; χσχ2−q )| ≤ 1, our goal will
be to show χσχ2−q = o(1), as χ → 0+ (note that χmin = 0 when δ ≥ 1).
We first consider the case δ > 1. To prove χσχ2−q = o(1), it is sufficient
to show σχ = O(1). Suppose this is not true, then there exists a sequence
χn → 0 such that σχn → ∞. Recall that (χn , σχn ) satisfies (2.2):
(D.13)
1
2
σχ2 n = σw
+ E(ηq (B + σχn Z; χn σχ2−q
) − B)2 .
n
δ
Dividing both sides of the above equation by σχ2 n and letting n → ∞ yields
1 = 1δ < 1, which is a contradiction.
Regarding the case δ = 1, we first claim that σχ → ∞, as χ → 0. Otherwise, there exists a sequence χn → 0 such that σχn → σ ∗ ∈ (0, ∞). However,
2 + (σ ∗ )2
taking the limit n → ∞ on both sides of (D.13) gives us (σ ∗ )2 = σw
where contradiction arises. Hence, if we can show χσχ2 = O(1), then χσχ2−q =
HIGHER ORDER ANALYSIS OF REGULARIZATION TECHNIQUES
53
o(1) will be proved. Starting from (D.13) (replacing χn by χ) with δ = 1,
we can have for q ∈ (1, 2]
2
0 = σw
+ σχ2 E(ηq (B/σχ + Z; χ) − B/σχ − Z)2 +
2σχ2 EZ(ηq (B/σχ + Z; χ) − B/σχ − Z)
(a)
2
= σw
+ χσχ2 · E(χq 2 |ηq (B/σχ + Z; χ)|2q−2 ) +
|
{z
}
A
−2q(q − 1)|ηq (B/σχ + Z; χ)|q−2
,
χσχ2 · E
1 + χq(q − 1)|ηq (B/σχ + Z; χ)|q−2
|
{z
}
B
where to obtain (a) we have used Lemma 7 part (i), Lemma 10 part (i) and
Lemma 11. Therefore we obtain
2
χσχ2 = −σw
· (A + B)−1 .
(D.14)
Because σχ → ∞ as χ → 0, it is easily seen that
(D.15)
lim A = 0,
χ→0+
lim inf |B| ≥ 2q(q − 1)E|Z|q−2 .
χ→0+
Combining results (D.14) and (D.15) we can conclude that χσχ2 = O(1).
Finally for the case q = 1, we do similar calculations and have
1
E(ηq (B/σχ + Z; χ) − B/σχ − Z)2 ≤ χ → 0,
χ
2
4
|B| = P (|B/σχ + Z| ≤ χ) → √ .
χ
2π
|A| =
This completes the proof.
According to the results presented in
Lemmas 13 and 14, if the function
χσχ2−q 1 − 1δ E[∂1 ηq (B + σχ Z; χσχ2−q )] , is continuous with respect to χ ∈
(χmin , ∞), then we can conclude that Equations (2.2) and (2.3) share at least
one common solution pair. To confirm the continuity, it is straightforward
to employ implicit function theorem to show σχ is continuous about χ.
Moreover, According to Lemma 8, ∂1 ηq (u; χ) is also a continuous function
of its arguments.
D.3. Proving the uniqueness of the solution of (2.2) and (2.3).
The proof of uniqueness is motivated by the idea presented in [13]. Suppose there are two different solutions denoted by (σχ1 , χ1 ) and (σχ2 , χ2 ),
54
HAOLEI WENG, ARIAN MALEKI AND LE ZHENG
respectively. By applying Theorem 2.1 with ψ(a, b) = (a − b)2 , we have
AMSE(λ, q, σw )
=
) − B]2
E[ηq (B + σχ1 Z; χ1 σχ2−q
1
(a)
2
δ(σχ2 1 − σw
),
=
where (a) is due to (2.2). The same equations hold for the other solution pair
(σχ2 , χ2 ). Since they have the same AMSE, it follows that σχ1 = σχ2 . Next
we choose a different pseudo-Lipschitz function ψ(a, b) = |a| in Theorem 2.1
to obtain
p
1X
)|
|β̂i (λ, q, p)| = E|ηq (B + σχ1 Z; χ1 σχ2−q
1
p→∞ p
lim
i=1
= E|ηq (B + σχ2 Z; χ2 σχ2−q
)| = E|ηq (B + σχ1 Z; χ2 σχ2−q
)|
2
1
Since E|ηq (B + σχ1 Z; χ)|, as a function of χ ∈ (0, ∞), is strictly decreasing
based on Lemma 10 part (v), we conclude χ1 = χ2 .
APPENDIX E: PROOF OF COROLLARY 1
According to Theorem 2.1, the key of proving Corollary 1 is to analyze
the following equations:
(E.1)
(E.2)
1
σ 2 = σω2 + E[(ηq (B + σZ; χσ 2−q ) − B)2 ],
δ
1
λ∗,q = χσ 2−q 1 − E[ηq0 (B + σZ; χσ 2−q )] ,
δ
where λ∗,q = arg minλ≥0 AMSE(λ, q, σw ). We present the main result in the
following lemma.
Lemma 15. For every 1 ≤ q ≤ 2 and a given optimal tuning λ∗,q , there
exists a unique solution pair (σ̄, χ̄) that satisfies (E.1) and (E.2). Furthermore, σ̄ is the unique solution of
(E.3)
2
σ 2 = σw
+
1
min E(ηq (B + σZ; χ) − B)2 ,
δ χ≥0
and
(E.4)
χ̄ ∈ arg min E[(ηq (B + σ̄Z; χσ̄ 2−q ) − B)2 ].
χ≥0
HIGHER ORDER ANALYSIS OF REGULARIZATION TECHNIQUES
55
Proof. The first part of this lemma directly comes from Lemma 2. We
focus on the proof of the second part. Recall the definition of Rq (χ, σ) in
(D.1). Define
σ2
1
Gq (σ) , w2 + min Rq (χ, σ).
σ
δ χ≥0
Then (E.3) is equivalent to Gq (σ) = 1. We first show that Gq (σ) is a strictly
decreasing function of σ over (0, ∞). For any given σ1 > σ2 > 0, we can
choose χ1 , χ2 such that
χ1 = arg min E(ηq (B/σ1 + Z; χ) − B/σ1 )2 ,
χ≥0
χ2 = arg min E(ηq (B/σ2 + Z; χ) − B/σ2 )2 .
χ≥0
Applying Lemma 12 we have
Rq (χ1 , σ1 ) = min Rq (χ, σ1 ) ≤ Rq (χ2 , σ1 ) ≤ Rq (χ2 , σ2 ) = min Rq (χ, σ2 ).
χ≥0
χ≥0
Hence we obtain that Gq (σ1 ) < Gq (σ2 ). We next show
lim Gq (σ) = ∞,
(E.5)
σ→0
lim Gq (σ) < 1.
σ→∞
The first result in (E.5) is obvious. To prove the second one, we know
2
1
1
σw
+ Rq (χ, σ) = Eηq2 (Z; χ),
σ→∞ σ 2
δ
δ
lim Gq (σ) ≤ lim
σ→∞
for any given χ ≥ 0. Choosing a sufficiently large χ completes the proof for
the second inequality in (E.5). Based on (E.5) and the fact that Gq (σ) is a
strictly decreasing and continuous function of σ over (0, ∞), we can conclude
(E.3) has a unique solution. Call it σ ∗ , and denote
χ∗ ∈ arg min E[(ηq (B + σ ∗ Z; χ(σ ∗ )2−q ) − B)2 ].
χ≥0
Note that χ∗ may not be unique. Further define
1
λ∗ = χ∗ (σ ∗ )2−q 1 − E[∂1 ηq (B + σ ∗ Z; χ∗ (σ ∗ )2−q )] .
δ
It is straightforward to see that the pair (σ ∗ , χ∗ ) satisfies (2.2) and (2.3)
with λ = λ∗ . According to Theorem 2.1, we obtain
(E.6)
2
AMSE(λ∗ , q, σw ) = δ((σ ∗ )2 − σw
).
56
HAOLEI WENG, ARIAN MALEKI AND LE ZHENG
Also we already know for the optimal tuning λ∗,q
2
AMSE(λ∗,q , q, σw ) = δ(σ̄ 2 − σw
) ≤ AMSE(λ∗ , q, σw ).
Therefore σ̄ ≤ σ ∗ . On the other hand,
1
2
σ̄ 2 = σw
+ E(ηq (B + σ̄Z; χ̄σ̄ 2−q ) − B)2
δ
1
2
≥ σw + min E(ηq (B + σ̄Z; χ) − B)2 = σ̄ 2 · Gq (σ̄).
δ χ≥0
Thus Gq (σ̄) ≤ 1 = Gq (σ ∗ ). Since Gq (σ) is a strictly decreasing function, we
then obtain σ̄ ≥ σ ∗ . Consequently, we conclude σ̄ = σ ∗ . Finally we claim
(E.4) has to hold. Otherwise,
AMSE(λ∗,q , q, σw ) = E(ηq (B + σ̄Z; χ̄σ̄ 2−q ) − B)2
> min E(ηq (B + σ̄Z; χ) − B)2 = E(ηq (B + σ ∗ Z; χ∗ ) − B)2
χ≥0
= AMSE(λ∗ , q, σw ),
contradicts the fact that λ∗,q is the optimal tuning.
Remark: Lemma 15 leads directly to the result of Corollary 1. Furthermore, from the proof of Lemma 15, we see that if E[(ηq (B+σZ; χσ 2−q )−B)2 ],
as a function of χ, has a unique minimizer for any given σ > 0, then χ̄ = χ∗ .
That means the optimal tuning value λ∗,q is unique. [46] has proved that
E[(ηq (B + σZ; χσ 2−q ) − B)2 ] is quasi-convex and has a unique minimizer for
q = 1. We conjecture it is true for q ∈ (1, 2] as well and leave it for future
research.
APPENDIX F: PROOF OF THEOREM 3.1
F.1. Roadmap of the proof. Different from the result of Theorem
3.3 that bounds the second order term in AMSE(λ∗,1 , 1, σw ), Theorem 3.1
characterizes the precise analytical expression of the second dominant term
for AMSE(λ∗,q , q, σw ) with q ∈ (1, 2]. However, the idea of this proof is
similar to the one for Theorem 3.3 presented in Section 6 of the main text,
though the detailed proof steps are more involved here. We suggest interested
readers first going over the proof of Theorem 3.3 and then this section so
that both the proof idea and technical details are smoothly understood.
Recall the definition we introduced in (D.1):
Rq (χ, σ) = E(ηq (B/σ + Z; χ) − B/σ)2 ,
HIGHER ORDER ANALYSIS OF REGULARIZATION TECHNIQUES
57
where Z ∼ N (0, 1) and B with the distribution pβ (b) = (1 − )δ0 (b) + g(b)
are independent. Define
χ∗q (σ) = arg min Rq (χ, σ).
(F.1)
χ≥0
Based on Lemma 8 of Appendix C, it is straightforward to show Rq (χ, σ) is
a differentiable function of χ. It is also easily seen that
lim Rq (χ, σ) =
χ→∞
E|B|2
,
σ2
lim Rq (χ, σ) = 1.
χ→0
Therefore, the minimizer χ∗q (σ) exists at least for sufficiently small σ. If it
is not unique, we will consider the one having smallest value itself. As like
the proof of Theorem 3.3, the key is to characterize the convergence rate for
Rq (χ∗q (σ), σ) as σ → 0. After having that convergence rate result, we can
then obtain the convergence rate for σ̄ from Equation (3.3) and finally derive
the expansion of AMSE(λ∗,q , q, σw ) based on Corollary 1. We organize our
proof steps as follows:
1. We first characterize the convergence rate of χ∗q (σ) in Section F.2.
2. We then obtain the convergence rate of Rq (χ∗q (σ), σ) in Section F.3.
3. We finally derive the second order expansion for AMSE(λ∗,q , q, σw ) in
Section F.4.
As a final remark, once we show the proof for q ∈ (1, 2), since η2 (u; χ) =
has a nice explicit form, the proof for q = 2 can be easily derived. We
hence skip it for simplicity.
u
1+2χ
F.2. Characterizing the convergence rate of χ∗q (σ). The goal of
this section is to derive the convergence rate of χ∗q (σ) as σ → 0. We will
make use of the fact that χ∗q (σ) is the minimizer of Rq (χ, σ), to first show
χ∗q (σ) → 0 and then obtain the rate χ∗q (σ) ∝ σ 2q−2 . This is done in the
following three lemmas.
Lemma 16. Let χ∗q (σ) denote the minimizer of Rq (χ, σ) as defined in
(F.1). Then for every b 6= 0 and z ∈ R,
|ηq (b/σ + z; χ∗q (σ))| → ∞,
as σ → 0.
Proof. Suppose this is not the case. Then there exist a value of b 6=
0, z ∈ R and a sequence σk → 0, such that |ηq (b/σk + z; χ∗q (σk ))| is bounded.
Combined with Lemma 7 part (i) we obtain
1
|b/σk + z| − |ηq (b/σk + z; χ∗q (σk ))|
(F.2) χ∗q (σk ) =
=
Ω
.
q|ηq (b/σk + z; χ∗q (σk ))|q−1
σk
58
HAOLEI WENG, ARIAN MALEKI AND LE ZHENG
We next show that the result (F.2) implies for any other b̃ 6= 0 and z̃ ∈ R,
|ηq (b̃/σk + z̃; χ∗q (σk ))| is bounded as well. From Lemma 7 part (i) we know
(F.3) |b̃/σk + z̃| = |ηq (b̃/σk + z̃; χ∗q (σk ))| + χ∗q (σk )q|ηq (b̃/σk + z̃; χ∗q (σk ))|q−1 .
If |ηq (b̃/σk + z̃; χ∗q (σk ))| is unbounded, then the right hand side of equation
(F.3) (take a subsequence if necessary) has the order larger than 1/σk . Hence
(F.3) can not hold for all the values of k. We thus have reached the conclusion
that |ηq (b/σk + z; χ∗q (σk ))| is bounded for every b 6= 0 and z ∈ R. Therefore,
|ηq (G/σk + Z; χ∗q (σk )) − G/σk | → ∞ a.s., as k → ∞,
where G has the distribution g(·). We then use Fatou’s lemma to obtain
Rq (χ∗q (σk ), σk ) ≥ E|ηq (G/σk + Z; χ∗q (σk )) − G/σk |2 → ∞.
On the other hand, since χ∗q (σk ) minimizes Rq (χ, σk ),
Rq (χ∗q (σk ), σk ) ≤ Rq (0, σk ) = 1.
Such contradiction completes the proof.
Lemma 16 enables us to derive χ∗q (σ) → 0 as σ → 0. We present it in the
next lemma.
Lemma 17. Let χ∗q (σ) denote the minimizer of Rq (χ, σ) as defined in
(F.1). Then χ∗q (σ) → 0 as σ → 0.
Proof. First note that
Rq (χ, σ)
=
(1 − )E(ηq (Z; χ))2 + E(ηq (G/σ + Z; χ) − G/σ − Z)2 +
2EZ(ηq (G/σ + Z; χ) − G/σ − Z) +
(a)
=
(1 − )E(ηq (Z; χ))2 + χ2 q 2 E|ηq (G/σ + Z; χ)|2q−2 +
2E(∂1 ηq (G/σ + Z; χ) − 1) +
(b)
=
(F.4)
(1 − )E(ηq (Z; χ))2 + χ2 q 2 E|ηq (G/σ + Z; χ)|2q−2 +
1
2E
− 1 + .
1 + χq(q − 1)|ηq (G/σ + Z; χ)|q−2
We have employed Lemma 7 part (i) and Lemma 8 to obtain (a); (b) is due to
Lemma 10 part (i). According to Lemma 16, |ηq (G/σ + Z; χ∗q (σ))| → ∞ a.s.,
as σ → 0. Hence, if χ∗q (σ) 9 0, the second term in (F.4) (with χ = χ∗q (σ))
goes off to infinity, while the other terms remain finite, and consequently
Rq (χ∗q (σ), σ) → ∞. This is a contradiction with the fact Rq (χ∗q (σ), σ) ≤
Rq (0, σ) = 1.
HIGHER ORDER ANALYSIS OF REGULARIZATION TECHNIQUES
59
So far we have shown χ∗q (σ) → 0 as σ → 0. Our next lemma further
characterizes the convergence rate of χ∗q (σ).
Lemma 18. Suppose P(|G| ≤ t) = O(t) (as t → 0) and E|G|2 < ∞.
Then for q ∈ (1, 2) we have
χ∗q (σ)
(1 − )E|Z|q
lim 2q−2 =
.
σ→0 σ
qE|G|2q−2
Proof. We first claim that χ∗q (σ) = Ω(σ 2q−2 ). Otherwise there exists a
sequence σk → 0 such that χ∗q (σk ) = o(σk2q−2 ). According to Lemma 19 (we
postpone Lemma 19 since it deals with Rq (χ, σ)),
lim
k→∞
Rq (χ∗q (σk ), σk ) − 1
σk2q−2
= 0.
On the other hand, by choosing χ(σk ) = Cσk2q−2 with C =
19 implies that
Rq (χ(σk ), σk ) − 1
< 0,
lim
k→∞
σk2q−2
(1−)E|Z|q
,
qE|G|2q−2
Lemma
which contradicts with the fact that χ∗q (σk ) is the minimizer of Rq (χ, σk ).
Moreover, this choice of C shows that for sufficiently small σ there exists
χ(σ) such that
Rq (χ∗q (σ), σ) ≤ Rq (χ(σ), σ) < Rq (0, σ) = 1.
∂Rq (χ∗ (σ),σ)
q
That means χ∗q (σ) is a non-zero finite value. Hence it satisfies
=
∂χ
∗
∗
0. From now on we use χ to denote χq (σ) for notational simplicity. That
equation can be detailed out as follows:
0 = (1 − )Eηq (Z; χ∗ )∂2 ηq (Z; χ∗ ) + E(ηq (G/σ + Z; χ∗ ) − G/σ)∂2 ηq (G/σ + Z; χ∗ )
−q|ηq (Z; χ∗ )|q
(a)
= (1 − ) E
+ E(Z∂2 ηq (G/σ + Z; χ∗ ))
1 + χ∗ q(q − 1)|ηq (Z; χ∗ )|q−2
|
{z
}
|
{z
}
H2
H1
(F.5)
q 2 |ηq (G/σ + Z; χ∗ )|2q−2
+χ∗ E
,
1 + χ∗ q(q − 1)|ηq (G/σ + Z; χ∗ )|q−2
|
{z
}
H3
where we have used Lemma 7 part (i) and Lemma 10 part (ii) to obtain (a).
We now analyze the three terms H1 , H2 and H3 , respectively. According to
60
HAOLEI WENG, ARIAN MALEKI AND LE ZHENG
Lemma 17, we have that ηq (Z; χ∗ ) → Z as σ → 0. Lemma 7 part (ii) enables
us to bound the expression inside the expectation of H1 by q|Z|q . Hence we
can employ Dominated Convergence Theorem (DCT) to obtain
lim H1 = −qE|Z|q .
(F.6)
σ→0
For the term H3 , we first note that
lim ηq (G + σZ; σ 2−q χ∗ ) = G,
σ→0
|ηq (G + σZ; σ 2−q χ∗ )| ≤ |B| + σ|Z|.
We also know that |ηq (G/σ + Z; χ∗ )| → ∞, a.s. by Lemma 16. We therefore
can apply DCT to conclude
q 2 |ηq (G + σZ; σ 2−q χ∗ )|2q−2
H3
=
E
lim
= q 2 E|G|2q−2 .
σ→0 σ 2−2q
σ→0 1 + χ∗ q(q − 1)|ηq (G/σ + Z; χ∗ )|q−2
(F.7) lim
We now study the remaining term H2 . According to Lemma 8, ∂2 ηq (G/σ +
Z; χ∗ ) is differentiable with respect to its first argument. So we can apply
Lemma 11 to get
|ηq (G/σ + Z; χ∗ )|q−2
H2 = q(1 − q) E
(1 + χ∗ q(q − 1)|ηq (G/σ + Z; χ∗ )|q−2 )3
|
{z
}
+q 2 (1 − q) E
(1 +
|
J1
∗
χ |ηq (G/σ + Z; χ∗ )|2q−4
χ∗ q(q − 1)|ηq (G/σ + Z; χ∗ )|q−2 )3
{z
.
}
J2
It is straightforward to see that
J1 ≤ E
J2
1
,
+ χ∗ q(q − 1)
1/(q(q − 1))
≤ E
.
|ηq (G/σ + Z; χ∗ )|2−q + χ∗ q(q − 1)
|ηq (G/σ +
Z; χ∗ )|2−q
We would like to prove H2 → 0 by showing
(F.8)
lim E
σ→0
1
|ηq (G/σ +
Z; χ∗ )|2−q
+ χ∗ q(q − 1)
= 0.
Note that DCT may not be directly applied here, because the function inside the expectation cannot be easily bounded. Alternatively, we prove (F.8)
HIGHER ORDER ANALYSIS OF REGULARIZATION TECHNIQUES
61
by breaking the expectation into different parts and showing each part con1
verges to zero. Let α1 be a number that satisfies ηq (α1 ; χ∗ ) = (χ∗ ) 2−q , α2 =
1
(χ∗ ) 2 , and α3 a fixed positive constant that does not depend on σ. Denote
the distribution of |G| by F (g). Note the following simple fact about α1
according to Lemma 7 part (i):
1
α1 = ηq (α1 ; χ∗ ) + χ∗ qηqq−1 (α1 ; χ∗ ) = (q + 1)(χ∗ ) 2−q .
So α1 < α2 < α3 when σ is small. Define the following three nested intervals:
Ii (x) , [−x − αi , −x + αi ],
i = 1, 2, 3.
With these definitions, we start the proof of (F.8). We have
1
E
|ηq (G/σ + Z; χ∗ )|2−q + χ∗ q(q − 1)
Z ∞Z
1
φ(z)dzdF (g) +
=
∗
2−q + χ∗ q(q − 1)
0
z∈I1 (g/σ) |ηq (g/σ + z; χ )|
Z ∞Z
1
φ(z)dzdF (g) +
∗
2−q + χ∗ q(q − 1)
0
z∈I2 (g/σ)\I1 (g/σ) |ηq (g/σ + z; χ )|
Z ∞Z
1
φ(z)dzdF (g) +
∗ )|2−q + χ∗ q(q − 1)
|η
(g/σ
+
z;
χ
q
0
z∈I3 (b/σ)\I2 (g/σ)
Z ∞Z
1
φ(z)dzdF (g)
∗
2−q + χ∗ q(q − 1)
0
R\I3 (b/σ) |ηq (g/σ + z; χ )|
, G 1 + G2 + G3 + G4 ,
where φ(·) is the density function of standard normal. We now bound each
of the four integrals in (F.9) respectively.
Z ∞Z
1
G1 ≤
φ(z)dzdF (g)
∗ q(q − 1)
χ
0
z∈I1 (g/σ)
1
≤
2α1 φ(0)
2(q + 1)(χ∗ ) 2−q φ(0)
≤
→ 0, as σ → 0.
χ∗ q(q − 1)
χ∗ q(q − 1)
The last step is due to the fact that χ∗ → 0 by Lemma 17. For G2 we have
Z ∞Z
1
G2 ≤
φ(z)dzdF (g)
∗ 2−q
0
z∈I2 (g/σ)\I1 (g/σ) |ηq (g/σ + z; χ )|
Z ∞Z
(b)
1
≤
φ(z)dzdF (g)
∗
χ
0
z∈I2 (g/σ)\I1 (g/σ)
62
HAOLEI WENG, ARIAN MALEKI AND LE ZHENG
Z
≤
σ log(1/σ) Z
1
φ(z)dzdF (g) +
∗
χ
0
z∈I2 (g/σ)\I1 (g/σ)
Z ∞
Z
1
φ(z)dzdF (g)
∗
σ log(1/σ) z∈I2 (g/σ)\I1 (g/σ) χ
2φ(0)α2 2φ(log(1/σ) − α2 )α2
+
χ∗
χ∗
2φ(0)
2φ(log(1/σ) − α2 )
= P(|G| ≤ σ log(1/σ)) ∗ 1/2 +
(χ )
(χ∗ )1/2
≤ P(|G| ≤ σ log(1/σ))
(c)
≤ O(1) · σ 2−q log(1/σ) + O(1) ·
φ((log(1/σ))/2)
→ 0, as σ → 0.
σ q−1
In the above derivations, (b) is because
|ηq (g/σ + z; χ∗ )| ≥ ηq (α1 ; χ∗ ) = (χ∗ )1/(2−q) for z ∈
/ I1 (g/σ).
To obtain (c), we have used the condition P(|G| ≤ σ log(1/σ)) = O(σ log(1/σ))
and the result χ∗ = Ω(σ 2q−2 ) we proved at the beginning. Regarding G3 ,
Z ∞Z
1
φ(z)dzdF (g)
G3 ≤
∗ 2−q
0
z∈I3 (g/σ)\I2 (g/σ) |ηq (α2 ; χ )|
Z σ log 1/σ Z
1
≤
φ(z)dzdF (g)
∗ 2−q
|η
(α
;
q
2 χ )|
0
z∈I3 (g/σ)\I2 (g/σ)
Z ∞
Z
1
+
φ(z)dzdF (g)
∗ 2−q
|η
(α
;
q
2 χ )|
σ log 1/σ z∈I3 (g/σ)\I2 (g/σ)
≤
(d)
≤
P(|G| ≤ σ log(1/σ))
O(1) · σ q
2 −3q+3
2φ(0)α3
2φ(log(1/σ) − α3 )α3
+
∗
2−q
|ηq (α2 ; χ )|
|ηq (α2 ; χ∗ )|2−q
log(1/σ) + O(1) ·
φ((log(1/σ))/2)
→ 0, as σ → 0.
σ (q−1)(2−q)
The calculations above are similar to those for G2 . In (d) we have used the
following result:
ηq (α2 ; χ∗ )
ηq ((χ∗ )1/2 ; χ∗ )
=
lim
= lim ηq (1; (χ∗ )q/2 ) = 1.
σ→0 (χ∗ )1/2
σ→0
σ→0
(χ∗ )1/2
lim
Finally we can apply DCT to obtain
lim G4 = E lim
σ→0
σ→0
I(|G/σ + Z| > α3 )
= 0.
|ηq (G/σ + Z; χ∗ )|2−q + χ∗ q(q − 1)
HIGHER ORDER ANALYSIS OF REGULARIZATION TECHNIQUES
63
We have finished the proof of limσ→0 H2 = 0. This fact together with (F.5),
(F.6) and (F.7) gives us
χ∗
lim
σ→0 σ 2q−2
=
−(1 − )H1 − H2
(1 − )E|Z|q
=
.
2−2q
H3 /σ
qE|G|2q−2
F.3. Characterizing the convergence rate of Rq (χ∗q (σ), σ). Having derived the convergence rate of χ∗q (σ) in Section F.2, we aim to obtain
the convergence rate for Rq (χ∗q (σ), σ) in this section. Towards that goal, we
first present a useful lemma.
Lemma 19. Suppose P(|G| ≤ t) = O(t) (as t → 0) and E|G|2 < ∞. If
χ(σ) = Cσ 2q−2 with a fixed number C > 0, then for 1 < q < 2 we have
(F.9)
Rq (χ(σ), σ) − 1
= −2C(1 − )qE|Z|q + C 2 q 2 E|G|2q−2 .
σ→0
σ 2q−2
lim
Moreover, if χ(σ) = o(σ 2q−2 ) then
Rq (χ(σ), σ) − 1
= 0.
σ→0
σ 2q−2
(F.10)
lim
Proof. We first focus on the case χ(σ) = Cσ 2q−2 . According to (F.4),
Rq (χ, σ) − 1 = (1 − )E(ηq2 (Z; χ) − Z 2 ) + χ2 q 2 E|ηq (G/σ + Z; χ)|2q−2
{z
}
|
{z
} |
R2
R1
Z; χ)|q−2
|ηq (G/σ +
−2χq(q − 1)E
.
1 + χq(q − 1)|ηq (G/σ + Z; χ)|q−2
|
{z
}
R3
Now we calculate the limit of each of the terms individually. We have
R1
=
(1 − )E(ηq (Z; χ) + Z)(ηq (Z; χ) − Z)
(a)
=
−(1 − )E(ηq (Z; χ) + Z)(χq|ηq (Z; χ)|q−1 sign(Z))
=
−(1 − )χq(E|ηq (Z; χ)|q + E|Z||ηq (Z; χ)|q−1 ),
where (a) is due to Lemma 7 part (i). Hence we obtain
(F.11)
lim
σ→0
R1
σ 2q−2
= −C(1 − )q lim (E|ηq (Z; χ)|q + E|Z||ηq (Z; χ)|q−1 )
σ→0
= −2C(1 − )qE|Z|q .
64
HAOLEI WENG, ARIAN MALEKI AND LE ZHENG
The last equality is by Dominated Convergence Theorem (DCT). For R2 ,
lim
(F.12)
σ→0
R2
σ 2q−2
χ2 q 2 E|ηq (G/σ + Z; χ)|2q−2
σ→0
σ 2q−2
2 2
= C q lim E|ηq (G + σZ; χσ 2−q )|2q−2
=
lim
σ→0
= C 2 q 2 E|G|2q−2 .
Regarding the term R3 , we would like to show that if χ(σ) = O(σ 2q−2 ), then
(F.13)
lim
σ→0
χ
σ 2q−2
E
1
= 0.
|ηq (G/σ + Z; χ)|2−q + χq(q − 1)
Define α1 = 1 if 1 < q < 3/2 and α1 = σ 2q−3+c if 3/2 ≤ q < 2, where c > 0
is a sufficiently small constant that we will specify later. Let F (g) be the
distribution function of |G| and α2 > 1 a fixed constant. So α1 < α2 for
small σ. Define the following two nested intervals
Ii (x) , [−x − αi , −x + αi ],
i = 1, 2.
Then the expression in (F.13) can be written as
=
χ
σ 2q−2
χ
σ 2q−2
χ
2q−2
σ
χ
1
E
σ 2q−2 |ηq (G/σ + Z; χ)|2−q + χq(q − 1)
Z ∞Z
1
φ(z)dzdF (g) +
2−q + χq(q − 1)
0
z∈I1 (g/σ) |ηq (g/σ + z; χ)|
Z ∞Z
1
φ(z)dzdF (g) +
2−q + χq(q − 1)
|η
(g/σ
+
z;
χ)|
q
0
z∈I2 (g/σ)\I1 (g/σ)
Z ∞Z
1
φ(z)dzdF (g)
2−q + χq(q − 1)
0
R\I2 (g/σ) |ηq (g/σ + z; χ)|
, G 1 + G2 + G3 .
We will bound each of the three integrals above. The idea is similar as the
HIGHER ORDER ANALYSIS OF REGULARIZATION TECHNIQUES
65
one presented in the proof of Lemma 18. For the first integral,
Z ∞ Z −g/σ+α1
1
χ
G1 ≤
φ(z)dzdF (g)
σ 2q−2 0
χq(q
− 1)
−g/σ−α1
Z σ log 1/σ Z −g/σ+α1
χ
1
≤
φ(z)dzdF (g)
2q−2
σ
0
−g/σ−α1 χq(q − 1)
Z ∞
Z −g/σ+α1
1
χ
φ(z)dzdF (g)
+ 2q−2
σ
σ log 1/σ −g/σ−α1 χq(q − 1)
(b)
≤
≤
2α1 φ(0)
2α1 φ(log 1/σ − α1 )
+
q(q − 1)σ 2q−2
q(q − 1)σ 2q−2
α1
α1 φ((log 1/σ)/2)
O(1) · σ c log(1/σ) 2q−3+c + O(1) ·
σ
σ 2q−2
P(|G| ≤ σ log 1/σ)
(c)
→ 0, as σ → 0.
To obtain (b), we have used the following inequalities when σ is small:
Z
−g/σ+α1
φ(z)dz ≤ 2φ(0)α1 ,
for g ≤ σ log(1/σ),
−g/σ−α1
Z −g/σ+α1
φ(z)dz ≤ 2α1 φ(log(1/σ) − α1 ),
for g > σ log(1/σ).
−g/σ−α1
The limit (c) holds due to the choice of α1 . Regarding the second term G2 ,
Z ∞Z
χ
1
φ(z)dzdF (g)
G2 ≤
2q−2
2−q
σ
0
z∈I2 (g/σ)\I1 (g/σ) |ηq (g/σ + z; χ)|
Z σ log 1/σ Z
χ
1
=
φ(z)dzdF (g)
2q−2
2−q
σ
0
z∈I2 (g/σ)\I1 (g/σ) |ηq (g/σ + z; χ)|
Z ∞
Z
χ
1
+ 2q−2
φ(z)dzdF (g)
2−q
σ
σ log 1/σ z∈I2 (g/σ)\I1 (g/σ) |ηq (g/σ + z; χ)|
Z σ log 1/σ Z
(d)
χ
1
≤
φ(z)dzdF (g)
2q−2
2−q
σ
0
z∈I2 (g/σ)\I1 (g/σ) |ηq (α1 ; χ)|
Z ∞
Z
χ
1
+ 2q−2
φ(z)dzdF (g)
2−q
σ
σ log 1/σ z∈I2 (g/σ)\I1 (g/σ) |ηq (α1 ; χ)|
(e)
≤
(f )
≤
P(|G| ≤ σ log 1/σ)
O(1) · σ c log(1/σ)
2α2 φ(0)χ
2q−2
σ
|ηq (α1 ; χ)|2−q
1
α12−q σ c−1
+ O(1) ·
+
2α2 φ(log 1/σ − α2 )χ
σ 2q−2 |ηq (α1 ; χ)|2−q
φ((log 1/σ)/2)
α12−q
(g)
→ 0, as σ → 0,
66
HAOLEI WENG, ARIAN MALEKI AND LE ZHENG
where (d) is due to the fact |ηq (g/σ + z; χ)| ≥ ηq (α1 ; χ) for z ∈
/ I1 (g/σ); The
argument for (e) is similar to that for (b); (f ) holds based on two facts:
η (α ;χ)
1. limσ→0 q α11
= limσ→0 ηq (1; α1q−2 χ) = 1, since α1q−2 χ → 0. This is
obvious for the case α1 = 1. When α1 = σ 2q−3+c , we have α1q−2 χ =
2
O(1) · σ 2q +(c−5)q+4−2c and 2q 2 + (c − 5)q + 4 − 2c > 0 if c is chosen
small enough.
2. P(|G| ≤ σ log 1/σ) = O(σ log 1/σ). This is one of the conditions.
And finally (g) works as follows: it is clear that σ c log(1/σ)
1
σ c−1
α2−q
1
goes to
zero when α1 = 1; when α1 = σ 2q−3+c , we can sufficiently small c such that
2
α1q−2 σ 1−c = σ 2q +(c−7)q+7−3c = o(1). For the third integral G3 , we are able
to invoke DCT to obtain
I(|G/σ + Z| > α2 )
lim G3 = O(1) · lim E
= 0.
σ→0
σ→0 |ηq (G/σ + Z; χ)|2−q + χq(q − 1)
Note that DCT works because for small σ
I(|G/σ + Z| > α2 )
|ηq (G/σ + Z; χ)|2−q + χq(q − 1)
1
1
I(|G/σ + Z| > α2 )
≤
≤
.
≤
2−q
2−q
|ηq (G/σ + Z; χ)|
|ηq (α2 ; χ)|
|ηq (α2 ; 1)|2−q
We hence have showned that
R3
= 0.
σ 2q−2
Combining the results (F.11), (F.12), and (F.14) establishes (F.9).
To prove (F.10), first note that (F.14) has been derived in the general
setting χ(σ) = O(σ 2q−2 ). Moreover, we can use similar arguments to show
for χ(σ) = o(σ 2q−2 ),
R1
R2
lim 2q−2 = lim 2q−2 = 0.
σ→0 σ
σ→0 σ
This completes the proof.
(F.14)
lim
σ→0
We are in position to derive the convergence rate of Rq (χ∗q (σ), σ).
Lemma 20. Suppose P(|G| ≤ t) = O(t) (as t → 0) and E|G|2 < ∞.
Then for q ∈ (1, 2) we have
Rq (χ∗q (σ), σ) − 1
(1 − )2 (E|Z|q )2
=−
.
2q−2
σ→0
σ
E|G|2q−2
lim
Proof. According to Lemma 18, choosing C =
finishes the proof.
(1−)E|Z|q
qE|G|2q−2
in Lemma 19
HIGHER ORDER ANALYSIS OF REGULARIZATION TECHNIQUES
67
F.4. Deriving the expansion of AMSE(λ∗,q , q, σw ). According to
Corollary 1 we know
AMSE(λ∗,q , q, σw ) = σ̄ 2 · Rq (χ∗q (σ̄), σ̄),
(F.15)
where σ̄ satisfies the following equation:
2
σ̄ 2 = σw
+
(F.16)
σ̄ 2
Rq (χ∗q (σ̄), σ̄).
δ
Since χ∗q (σ̄) minimizes Rq (χ, σ̄) we obtain
Rq (χ∗q (σ̄), σ̄) ≤ Rq (0, σ̄) = 1.
Therefore, under the condition δ > 1,
δ
1
2
σ2 ,
σ̄ 2 ≤ σw
+ σ̄ 2 ⇒ σ̄ 2 ≤
δ
δ−1 w
(F.17)
which implies that σ̄ → 0 as σw → 0. Accordingly, we combine Equation
(F.16) with the fact Rq (χ∗q (σ̄), σ̄) → 1 as σ̄ → 0 from Lemma 20 to conclude
δ−1
2
2
σw
1
σw
∗
1
−
=
lim
=
lim
R
(χ
(σ̄),
σ̄)
=
.
q
q
σw →0 σ̄ 2
σ̄→0 σ̄ 2
σ̄→0
δ
δ
(F.18) lim
We now derive the expansion of AMSE(λ∗,q , q, σw ) presented in (3.4). From
(F.15) and (F.16) we can compute that
AMSE(λ∗,q , q, σw ) −
=
2
σw
1−1/δ
2q
σw
σ̄ 2 δ(Rq (χ∗q (σ̄), σ̄) − 1)
2q
(δ − 1)
σw
=
=
σ̄ 2 Rq (χ∗q (σ̄), σ̄) −
σ̄ 2q
2q
σw
1
1−1/δ
· (σ̄ 2 −
σ̄ 2
∗
δ Rq (χq (σ̄), σ̄))
2q
σw
Rq (χ∗q (σ̄), σ̄) − 1
δ
·
·
.
δ−1
σ̄ 2q−2
Letting σw → 0 on both sides of the above equation and using the results
from (F.18) and Lemma 20 completes the proof.
APPENDIX G: PROOF OF THEOREM 3.2
From (F.15) and (F.16) we see that
2
AMSE(λ∗,q , q, σw ) = δ(σ̄ 2 − σw
).
Hence Theorem 3.2 can be proved by showing
lim σ̄ > 0.
σw →0
68
HAOLEI WENG, ARIAN MALEKI AND LE ZHENG
For that purpose we first prove the following under the condition E|G|2 < ∞.
lim Rq (χ∗q (σ), σ) = 1.
(G.1)
σ→0
When q = 2, Rq (χ∗q (σ); σ) admits a nice explicit expression and can be
easily shown to converge to 1. For 1 < q < 2, since χ∗q (σ) is the minimizer
of Rq (χ, σ) we know
Rq (χ∗q (σ), σ) ≤ Rq (0, σ) = 1,
hence lim supσ→0 Rq (χ∗q (σ), σ) ≤ 1. On the other hand, (F.4) gives us
Rq (χ∗q (σ), σ) ≥ (1 − )E(ηq (Z; χ∗ ))2 −
1
+2E
1 + χ∗ q(q − 1)|ηq (G/σ + Z; χ∗ )|q−2
where we have used χ∗ to denote χ∗q (σ) for simplicity. Based on Lemmas
16 and 17, we can apply Fatou’s lemma to the above inequality to obtain
lim inf σ→0 Rq (χ∗q (σ), σ) ≥ 1.
Next it is clear that
(G.2)
lim Rq (χ∗q (σ), σ) ≤ lim lim Rq (χ, σ) = 0.
σ→∞
σ→∞ χ→∞
We now consider an arbitrary convergent sequence σ̄n → σ ∗ . We claim
σ ∗ 6= 0. Otherwise Equation (F.16) tells us
Rq (χ∗q (σ̄n ), σ̄n ) < δ < 1,
and letting n → ∞ above contradicts (G.1). Now that σ ∗ > 0 we can take
n → ∞ in (F.16) to obtain
Rq (χ∗q (σ ∗ ), σ ∗ ) = δ < 1.
According to Lemma 12, it is not hard to confirm Rq (χ∗q (σ), σ) is a strictly
decreasing and continuous function of σ. Results (G.1) and (G.2) then imply
that σ ∗ is the unique solution to Rq (χ∗q (σ), σ) = δ. Since this is true for any
sequence, we have proved limσw →0 σ̄ exists and larger than zero.
APPENDIX H: PROOF OF THEOREM 3.4
H.1. Roadmap. Theorem 3.4 differs from Theorem 3.3 in that the order of the second dominant term of AMSE(λ∗,1 , 1, σw ) becomes polynomial
(ignore the logarithm term) when the distribution of G has mass around zero.
HIGHER ORDER ANALYSIS OF REGULARIZATION TECHNIQUES
69
However, the proof outline remains the same. We hence stick to the same
notations used in the proof of Theorem 3.3. In particular, R(χ, σ), χ∗ (σ)
represent Rq (χ, σ), χ∗q (σ) with q = 1, respectively. We characterize the convergence rate of χ∗ (σ) in Section H.2, and bound the convergence rate of
R(χ∗ (σ), σ) in Section H.3. After we characterize R(χ∗ (σ), σ), the rest of
the proof is similar to that in Section 6.4 from the main text. We therefore
do not repeat it here.
H.2. Bounding the convergence rate of χ∗ (σ).
Lemma 21. Suppose P(|G| ≤ t) = Θ(t` ) with ` > 0 (as t → 0) and
E|G|2 < ∞, then for sufficiently small σ
`
v
u
1
u
αm σ ` ≤ χ∗ (σ) − χ∗∗ ≤ βm σ ` · tlog log . . . log
,
|
{z
} σ
m times
where m > 0 is an arbitrary integer number, αm , βm > 0 are two constants
depending on m, and χ∗∗ is the unique minimizer of (1 − )E(η1 (Z; χ))2 +
(1 + χ2 ) over [0, ∞).
Proof. According to Lemma 3, χ∗ (σ) → χ∗∗ as σ → 0. To characterize
the convergence rate, we follow the same line of proof that we presented for
Lemma 4 and adopt the same notations. For simplicity we do not detail out
the entire proof and instead highlight the differences. The key difference is
that neither e1 or e2 are exponentially small in the current setting. We now
start by bounding e2 . Let F (g) be the distribution function of |G| and define
logm (a) , log log . . . log(a).
|
{z
}
m times
Given an integer m > 0 and a constant c > 0, we then have
Eφ(χ∗ − |G|/σ)
m−1
X Z cσ(logm−i (1/σ))1/2
=
φ(χ∗ − g/σ)dF (g) +
i=1
Z
cσ(logm−i+1 (1/σ))1/2
cσ(logm (1/σ))1/2
∗
Z
∞
φ(χ − g/σ)dF (g) +
0
≤
m−1
X
φ(χ∗ − g/σ)dF (g)
cσ(log(1/σ))1/2
φ(c(logm−i+1 (1/σ))1/2 − χ∗ ) · P(|G| ≤ cσ(logm−i (1/σ))1/2 ) +
i=1
(H.1)
φ(0) · P(|G| ≤ cσ(logm (1/σ))1/2 ) + φ(c(log(1/σ))1/2 − χ∗ ).
70
HAOLEI WENG, ARIAN MALEKI AND LE ZHENG
The condition P(|G| ≤ t) = Θ(t` ) leads to
m−1
X
φ(c(logm−i+1 (1/σ))1/2 − χ∗ ) · P(|G| ≤ cσ(logm−i (1/σ))1/2 )
i=1
≤
∗ 2
m−1
e(χ ) /2 X −(c2 logm−i+1 (1/σ))/4
√
e
· Θ(σ ` (logm−i (1/σ))`/2 )
2π i=1
= O(1) ·
m−1
X
σ ` (logm−i (1/σ))`/2−c
2 /4
,
i=1
2 /2
where we have used the simple inequality e−(a−b)
also clear that
≤ e−b
2 /4
2 /2
· ea
. It is
1
∗ 2
2
φ(c(log(1/σ))1/2 − χ∗ ) ≤ √ e(χ ) /2 · σ c /4 .
2π
Therefore, by choosing a sufficiently large c we can conclude that the dominant term in (H.1) is φ(0)P(|G| ≤ cσ(logm (1/σ))1/2 ) = Θ(σ ` (logm (1/σ))`/2 ).
Furthermore, choosing a fixed constant C > 0 we have the following lower
bound
Z Cσ
∗
Eφ(χ +|G|/σ) ≥
φ(χ∗ +g/σ)dF (g) ≥ φ(C +χ∗ )·P(|G| ≤ Cσ) = Θ(σ ` ).
0
Because
Eφ(χ∗ + |G|/σ) ≤ Eφ(χ∗ ± G/σ) ≤ Eφ(χ∗ − |G|/σ),
We are able to derive
Θ(σ ` ) ≤ Eφ(χ∗ ± G/σ) ≤ Θ(σ ` (logm (1/σ))`/2 ).
As a result we obtain the bound for e2 :
(H.2)
Θ(σ ` ) ≤ e2 ≤ Θ(σ ` (logm (1/σ))`/2 )
To bound e1 , recall that
Z −G/σ+χ∗
e1 = −E
φ(z)dz = −2χ∗ Eφ(aχ∗ − G/σ),
−G/σ−χ∗
where |a| ≤ 1 depends on G. We can find two positive constants C1 , C2 > 0
such that for small σ
C1 Eφ(χ∗ + |G|/σ) ≤ Eφ(aχ∗ − G/σ) ≤ C2 Eφ(χ∗ − |G|/σ).
HIGHER ORDER ANALYSIS OF REGULARIZATION TECHNIQUES
71
Hence we have
Θ(σ ` ) ≤ −e1 ≤ Θ(σ ` (logm (1/σ))`/2 ).
(H.3)
Based on the results from (H.2) and (H.3) and Equation (6.9), we can use
similar arguments as in the proof of Lemma 4 to conclude
Θ(σ ` ) ≤ χ∗ (σ) − χ∗∗ ≤ Θ(σ ` (logm (1/σ))`/2 ).
H.3. Bounding the convergence rate of R(χ∗ (σ), σ).
Lemma 22. Suppose P(|G| ≤ t) = Θ(t` ) with ` > 0 (as t → 0) and
E|G|2 < ∞, then for sufficiently small σ
`
v
u
1
u
≤ R(χ∗ (σ), σ) − M1 () ≤ −αm σ ` ,
−βm σ ` · tlog log . . . log
|
{z
} σ
m times
where m > 0 is an arbitrary integer number and αm , βm > 0 are two constants depending on m.
Proof. We recall the two quantities:
M1 () = (1 − )E(η1 (Z; χ∗∗ ))2 + (1 + (χ∗∗ )2 )
R(χ∗ (σ), σ) = (1 − )E(η1 (Z; χ∗ ))2 + (1 + E(η1 (G/σ + Z; χ∗ ) − G/σ − Z)2 )
+2EZ(η1 (G/σ + Z; χ∗ ) − G/σ − Z).
Since χ∗ (σ) is the minimizer of R(χ, σ),
R(χ∗ (σ), σ) − M1 () ≤ R(χ∗∗ , σ) − M1 ()
=
[E(η1 (G/σ + Z; χ∗∗ ) − G/σ − Z)2 − (χ∗∗ )2 ] +
2EZ(η1 (G/σ + Z; χ∗∗ ) − G/σ − Z)
(a)
(H.4)
≤
(b)
−2EI(|G/σ + Z| ≤ χ∗∗ ) ≤ −Θ(σ ` ).
To obtain (a), we have used Lemma 11 and the fact |η1 (u; χ) − u| ≤ χ. (b)
is due to the similar arguments for bounding e1 in Lemma 21. To derive the
lower bound for R(χ∗ (σ), σ)−M1 (), we can follow the same reasoning steps
as in the proof of Lemma 5 and utilize the bound we derived for |χ∗ (σ)−χ∗∗ |
in Lemma 21. We will obtain
(H.5)
|R(χ∗ (σ), σ) − M1 ()| ≤ Θ(σ ` (logm (1/σ))`/2 ).
Putting (H.4) and (H.5) together completes the proof.
72
HAOLEI WENG, ARIAN MALEKI AND LE ZHENG
APPENDIX I: PROOF OF THEOREM 3.5
The proof of Theorem 3.5 is essentially the same as that of Theorem 3.2.
We do not repeat the details. Note the key argument limσ→0 R(χ∗ (σ), σ) =
M1 () has been shown in Lemma 3.
APPENDIX J: PROOF OF THEOREM 2.1
J.1. Roadmap of the proof. This appendix contains the proof of
Theorem 2.1. The proof for LASSO (q = 1) has been shown in [13]. We aim
to extend the results to 1 < q ≤ 2. We will follow similar proof strategy
as the one proposed in [13]. However, as will be described later some of the
steps are more challenging for 1 < q ≤ 2 (and some are easier). Motivated
by [13] we construct an approximate message passing (AMP) algorithm for
solving LQLS. We then establish an asymptotic equivalence between the
output of AMP and the bridge regression estimates. We finally utilize the
existing asymptotic results from AMP framework to prove Theorem 2.1. The
rest of the material is organized as follows. In Section J.2, we briefly review
approximate message passing algorithms and state some relevant results that
will be used later in our proof. Section J.3 collects two useful results to be
applied in the later proof. We describe the main proof steps in Section J.4.
J.2. Approximate message passing algorithms. For a function f :
R → R and a vector v ∈ Rm , we use f (v) ∈ Rm to denote the vector
(f (v1 ), . . . , f (vm )). Recall ηq (u; χ) is the proximal operator for the function
k·kqq . We are in the linear regression model setting: y = Xβ +w. To estimate
β, we adapt the AMP algorithm in [29] to generate a sequence of estimates
β t ∈ Rp , based on the following iterations (initialized at β 0 = 0, z 0 = y):
β t+1 = ηq (X T z t + β t ; θt ),
1
z t = y − Xβ t + z t−1 h∂1 ηq (X T z t−1 + β t−1 ; θt−1 )i,
(J.1)
δ
1 Pp
where hvi = p i=1 vi denotes the average of a vector’s components and
{θt } is a sequence of tuning parameters specified during the iterations. A
remarkable phenomenon about AMP is that the asymptotics of the sequence
{β t } can be characterized by one dimensional parameter, known as the state
of the system. The following theorem clarifies this claim.
Theorem J.1. Let {β(p), X(p), w(p)} be a converging sequence and ψ :
→ R be a pseudo-Lipschitz function. For any iteration number t > 0,
R2
p
1X
lim
ψ(βit+1 , βi ) = E[ψ(ηq (B + τt Z; θt ), B)],
p→∞ p
i=1
a.s.,
HIGHER ORDER ANALYSIS OF REGULARIZATION TECHNIQUES
73
where B ∼ pβ and Z ∼ N (0, 1) are independent and {τt }∞
t=0 can be tracked
through the following recursion (τ02 = σ 2 + 1δ E|B|2 ):
(J.2)
1
2
2
τt+1
= σw
+ E[ηq (B + τt Z; θt ) − B]2 , t ≥ 0.
δ
Proof. According to Lemma 7 part (vi), ηq (u; χ) is a Lipschitz continuous function of u. We can then directly apply Theorem 1 in [12] to complete
the proof.
Equation (J.2) is called state evolution. Theorem J.1 demonstrates that
the general asymptotic performance of {β t } is sharply predicted by the state
evolution. From now on, we will consider the AMP estimates {β t } with
θt = χτt2−q in (J.1). The positive constant χ is the solution of (2.2) and
(2.3). Note we have proved in Section D.2 that the solution exists. We next
present a useful lemma that characterizes the convergence of {τt }. Recall
the definition in (D.6):
1
χmin = inf χ ≥ 0 : E(ηq2 (Z; χ)) ≤ 1 .
δ
Lemma 23. For any given χ ∈ (χmin , ∞), the sequence {τt }∞
t=0 generated
from (J.2) with θt = χτt2−q converges to a finite number as t → ∞.
2 + 1 E[η (B + τ Z; χτ 2−q ) − B]2 . According to
Proof. Denote H(τ ) = σw
q
δ
Corollary 2, we know H(τ ) = τ 2 has a unique solution. Furthermore, since
H(0) > 0 and H(τ ) < τ 2 when τ is large enough, it is straightforward to
confirm the result stated in the above lemma.
Denote τt → τ∗ as t → ∞. Lemma 23 and (J.2) together yield
(J.3)
1
2
τ∗2 = σw
+ E[ηq (B + τ∗ Z; χτ∗2−q ) − B]2 .
δ
This is the same as Equation (2.2). We hence see the connection between
AMP estimates and bridge regression. The main part of the proof for Theorem 2.1 is to rigorously establish such connection. In particular we will
show the sequence {β t } converges (in certain asymptotic sense) to β̂(λ, q) as
t → ∞ in Section J.4. Towards that goal, we present the next theorem that
shows asymptotic characterization of other quantities in the AMP algorithm.
2−q
)i. Under the
Theorem J.2. Define wt , 1δ h∂1 ηq (X T z t−1 + β t−1 ; χτt−1
conditions of Theorem J.1, we have almost surely
74
HAOLEI WENG, ARIAN MALEKI AND LE ZHENG
kβ t+1 −β t k22
p
t→∞p→∞
= 0.
kz t+1 −z t k22
p
= 0.
(i) lim lim
(ii) lim lim
(iii)
t→∞p→∞
kz t k2
lim n 2
p→∞
= τt2 .
2−q
(iv) lim wt = 1δ E[∂1 ηq (B+τt−1 Z; χτt−1
)], where B, Z are the same random
p→∞
variables as in Theorem J.1.
Proof. All the results for q = 1 have been derived in [13]. We here
generalize them to the case 1 < q ≤ 2. Since the proof is mostly a direct
modification of that in [13], we only highlight the differences and refer the
reader to [13] for detailed arguments. According to the proof of Lemma 4.3
in [13], we have almost surely
kβ t+1 − β t k22
kz t+1 − z t k22
= lim
p→∞
p→∞
p
p
lim
2−q
χ)]2 ,
= E[ηq (B + Zt ; τt2−q χ) − ηq (B + Zt−1 ; τt−1
where (Zt , Zt−1 ) is jointly zero-mean gaussian, independent from B ∼ pβ ,
with covariance matrix defined by the recursion (4.13) in [13]. From Lemma
8, we know ηq (u; χ) is a differentiable function over (−∞, +∞) × (0, ∞).
Hence we can apply mean value theorem to obtain
2−q
E[ηq (B + Zt ; τt2−q χ) − ηq (B + Zt−1 ; τt−1
χ)]2
≤
2−q
)χ]2
E[∂1 ηq (a; b) · (Zt − Zt−1 ) + ∂2 ηq (a; b) · (τt2−q − τt−1
≤
2−q 2 2
) χ ]
2E[(∂1 ηq (a; b))2 · (Zt − Zt−1 )2 ] + 2E[(∂2 ηq (a; b))2 · (τt2−q − τt−1
(a)
≤
2−q 2 2 2
) χ q E|a|2q−2 ,
2E[(Zt − Zt−1 )2 ] + 2(τt2−q − τt−1
where (a, b) is a point on a line that connects the two points (B + Zt , τt2−q χ)
2−q
and (B + Zt−1 , τt−1
χ); we have used Lemma 7 part (ii) and Lemma 10 part
(i)(ii) to obtain (a). Note that Lemma 23 implies the second term on the
right hand side of the last inequality goes to zero, as t → ∞. Regarding the
first term, we can follow similar proof steps as for Lemma 5.7 in [13] to show
E(Zt − Zt−1 )2 → 0, as t → ∞.
The proof of part (iii) is the same as that of Lemma 4.1 in [13]. We
do not repeat the proof here. For (iv), Lemma F.3(b) in [13] implies the
empirical distribution of {((X T z t−1 +β t−1 )i , βi )}pi=1 converges weakly to the
2−q
distribution of (B + τt−1 Z, B). Since the function J(y, z) , ∂1 ηq (y; χτt−1
) is
bounded and continuous with respect to (y, z) according to Lemma 7 part
(i) and Lemma 8, (iv) follows directly from the Portmanteau theorem.
HIGHER ORDER ANALYSIS OF REGULARIZATION TECHNIQUES
75
J.3. Two useful theorems. In this section, we refer to two useful
theorems that have also been applied and cited in [13]. The first one is
regarding the limit of the singular values of random matrices taken from
[49].
Theorem J.3. (Bai and Yin, 1993). Let X ∈ Rn×p be a matrix having
i.i.d. entries with EXij = 0, EXij2 = 1/n. Denote by σmax (X), σmin (X) the
largest and smallest non-zero singular values of X, respectively. If n/p →
δ > 0, as p → ∞, then
1
lim σmax (X) = √ + 1, a.s.,
δ
1
lim σmin (X) = √ − 1 , a.s.
p→0
δ
p→0
The second theorem establishes the relation between `1 and `2 norm for
vectors from random subspace, showed in [50].
Theorem J.4. (Kashin, 1977). For a given constant 0 < v ≤ 1, there
exists a universal constant cv such that for any p ≥ 1 and a uniformly
random subspace V of dimension p(1 − v),
1
P ∀β ∈ V : cv kβk2 ≤ √ kβk1 ≥ 1 − 2−p .
p
J.4. The main proof steps. As mentioned before we will use similar
arguments as the ones shown in [13]. To avoid redundancy, we will not
present all the details and rather emphasize on the differences. We suggest
interested readers going over the proof in [13] before studying this section.
Similar to [13], we start with a lemma that summarizes several structural
properties of LQLS formulation. Define F(β) , 12 ky − Xβk22 + λkβkqq .
Suppose β, r ∈ Rp satisfy the following conditions:
√
krk2 ≤ c1 p
F(β + r) ≤ F(β)
√
k∇F (β)k2 P
≤ p
sup0≤µi ≤1 pi=1 |βi + µi ri |2−q ≤ pc2
0 < c3 ≤ σmin (X), where σmin (X) is defined in Theorem J.3
Lemma 24.
(i)
(ii)
(iii)
(iv)
(v)
(vi) krk k22 ≤ c4
7
krk k21
p .
The vector rk ∈ Rp is the projection of r onto ker(X)7
It is the nullspace of X defined as ker(X) = {β ∈ Rp | Xβ = 0}.
76
HAOLEI WENG, ARIAN MALEKI AND LE ZHENG
Then there exists a function f (, c1 , c2 , c3 , c3 , c4 , λ, q) such that
krk2 ≤
√
pf (, c1 , c2 , c3 , c4 , λ, q).
Moreover, f (, c1 , c2 , c3 , c4 , λ, q) → 0 as → 0.
Proof. First note that
∇F(β) = −X T (y − Xβ) + λq(|β1 |q−1 sign(β1 ), . . . , |βp |q−1 sign(βp ))T .
Combining it with Condition (ii) we have
0 ≥ F(β + r) − F(β)
1
1
= ky − Xβ − Xrk22 + λkβ + rkqq − ky − Xβk22 − λkβkqq
2
2
1
2
T T
= kXrk2 − r X (y − Xβ) + λ(kβ + rkqq − kβkqq )
2
p
X
1
2
T
= kXrk2 + r ∇F(β) + λ
(|βi + ri |q − |βi |q − qri |βi |q−1 sign(βi ))
2
i=1
p
(a)
(J.4)
λq(q − 1) X
1
≥ kXrk22 + rT ∇F(β) +
|βi + µi ri |q−2 ri2 ,
2
2
i=1
where (a) is obtained by Lemma 25 that we will prove shortly and {µi } are
numbers between 0 and 1. Note that we can decompose r as r = rk + r⊥
such that rk ∈ ker(X), r⊥ ∈ ker(X)⊥ . Accordingly Condition (v) yields
c23 kr⊥ k22 ≤ kXr⊥ k22 . This fact combined with Inequality (J.4) implies
c23 ⊥ 2 1
1
kr k2 ≤ kXr⊥ k22 = kXrk22 ≤ −rT ∇F(β) ≤ krk2 · k∇F(β)k2 ≤ c1 p,
2
2
2
where the last inequality is derived from Condition (i) and (iii). Hence we
can obtain
2c1 p
kr⊥ k22 ≤
.
c23
Our next step is to bound krk k22 . By Cauchy-Schwarz inequality we know
p
X
i=1
p
q
X
p
|ri | =
|βi + µi ri |2−q · ri2 |βi + µi ri |q−2 ,
i=1
v
v
u p
u p
uX
uX
t
2−q
≤
|βi + µi ri |
·t
ri2 |βi + µi ri |q−2 .
i=1
i=1
HIGHER ORDER ANALYSIS OF REGULARIZATION TECHNIQUES
77
So
p
X
(J.5)
krk21
.
2−q
i=1 |βi + µi ri |
ri2 |βi + µi ri |q−2 ≥ Pp
i=1
Combining Inequality (J.4) and (J.5) gives us
krk21 ≤
p
(b) 2c c
−2rT ∇F(β) X
1 2
·
|βi + µi ri |2−q ≤
p2 ,
λq(q − 1)
λq(q − 1)
i=1
where we have used Conditions (i), (iii), and (iv) to derive (b). Using the
upper bounds we obtained for krk21 and kr⊥ k22 , together with Condition (vi),
it is straightforward to verify the following chains of inequalities
krk k22 ≤
≤
c4 k 2 2c4
2c4
kr k1 ≤
(krk21 + kr⊥ k21 ) ≤
(krk21 + pkr⊥ k22 )
p
p
p
2c4 2c1 c2 2 2c1 2 4c1 c2 c4
4c1 c4
· p.
·
p + 2 p =
+ 2
p
λq(q − 1)
λq(q − 1)
c3
c3
We are finally able to derive
krk22 = krk k22 + kr⊥ k22 ≤
4c c c
4c1 c4 2c1
1 2 4
+ 2 + 2 · p.
λq(q − 1)
c3
c3
This completes the proof.
Note that Lemma 24 is a non-asymptotic and deterministic result. It
sheds light on the behavior of the cost function F(β) around its global
minimum. Suppose β + r is the global minimizer (a reasonable assumption
according to Condition (ii)), and if there is another point β having small
function value (indicated by its gradient from Condition (iii)), then the
distance krk2 between β and the optimal solution β + r should also be
small. This interpretation should not sound surprising, since we already
know F(β) is a strictly convex function. However, Lemma 24 enables us
to characterize this property in a precise way, which is crucial in the high
dimensional asymptotic analysis. Based on Lemma 24, we will set β + r =
β̂(λ, q), β = β t and then verify all the conditions in Lemma 24 to conclude
krk2 = kβ̂(λ, q) − β t k2 is small. In particular that small distance will vanish
as t → ∞, thus establishing the asymptotic equivalence between β̂(λ, q) and
β t . We perform the analysis in a sequel of lemmas and Proposition 1.
Lemma 25. Given a constant q satisfying 1 < q ≤ 2, for any x, r ∈ R,
there exists a number 0 ≤ µ ≤ 1 such that
(J.6)
|x + r|q − |x|q − rq|x|q−1 sign(x) ≥
q(q − 1)
|x + µr|q−2 r2 .
2
78
HAOLEI WENG, ARIAN MALEKI AND LE ZHENG
Proof. Denote fq (x) = |x|q . When q = 2, since f2 (x) is a smooth function over (−∞, +∞), we can apply Taylor’s theorem to obtain (J.6). For any
1 < q < 2, note that fq00 (0) = ∞, hence Taylor’s theorem is not applicable
to all the values of x ∈ R. We prove the inequality above in separate cases.
First observe that if (J.6) holds for any x > 0, r ∈ R, then it is true for any
x < 0, r ∈ R as well. It is also straightforward to confirm that when x = 0,
we can always choose µ = 1 to satisfy Inequality (J.6) for any r ∈ R. We
therefore focus on the case x > 0, r ∈ R.
a. When x + r > 0, since fq (x) is a smooth function over (0, ∞), we can
apply Taylor’s theorem to obtain (J.6).
b. If x+r = 0, choosing µ = 0, Inequality (J.6) is simplified to (q−1)xq ≥
q(q−1) q
2 x , which is clearly valid.
c. When x + r < 0, we consider two different scenarios.
i. First suppose −x − r ≥ x. We apply (J.6) to the pair −r − x and
x. Then we know there exists 0 ≤ µ̃ ≤ 1 such that
|x + r|q − |x|q ≥
q(q − 1)
|µ̃(−x − r) + (1 − µ̃)x|q−2 (2x + r)2
2
−(2x + r)q|x|q−1
It is also straightforward to verify that there is 0 ≤ µ ≤ 1 so that
µ(x + r) + (1 − µ)x = −µ̃(−x − r) − (1 − µ̃)x. Denote
g(y) =
q(q − 1)
|µ̃(−x − r) + (1 − µ̃)x|q−2 y 2 + q|x|q−1 y.
2
If we can show g(−2x − r) ≥ g(r), we can obtain the Inequality
(J.6). It is easily seen that the quadratic function g(y) achieves
global minimum at
y0 =
−1
−1
|x|q−1 |µ̃(−x − r) + (1 − µ̃)x|2−q ≤
|x| < −x.
q−1
q−1
Moreover, note that −2x − r ≥ 0, r < 0 and they are symmetric
around y = −x, hence g(−2x − r) ≥ g(r).
ii. Consider 0 < −x−r < x. We again use (J.6) for the pair −x−−r
and x to obtain
q(q − 1)
|µ̃(x + r) − (1 − µ̃)x|q−2 (2x + r)2
2
q(q − 1) q−2
−(2x + r)q|x|q−1 ≥ (−2x − r)q|x|q−1 +
|x| (2x + r)2 .
2
|x + r|q − |x|q ≥
HIGHER ORDER ANALYSIS OF REGULARIZATION TECHNIQUES
79
q−2 y 2 + q|x|q−1 y. If we can show h(−2x −
Denote h(y) = q(q−1)
2 |x|
r) ≥ h(r), Inequality (J.6) will be established with µ = 0. Since
−1
h(x) achieves global minimum at y0 = q−1
|x| < −x and −2x−r >
r, we can get h(−2x − r) ≥ h(r).
The next lemma is similar to Lemma 3.2 in [13]. The proof is adapted
from there.
Lemma 26. Let {β(p), X(p), w(p)} be a converging sequence. Denote the
solution of LQLS by β̂(λ, q), and let {β t }t≥0 be the sequence of estimates
generated from the AMP algorithm. There exists a positive constant C s.t.
kβ t k22
≤ C,
t→∞ p→∞
p
lim lim
lim sup
p→∞
a.s.,
kβ̂(λ, q)k2
≤ C,
p
a.s.
Proof. To show the first inequality, according to Theorem J.1 and Lemma
23, choosing a particular pseudo-Lipschitz function ψ(x, y) = x2 we obtain
kβ t k22
= EB,Z [ηq (B + τ∗ Z; χτ∗2−q )]2 < ∞,
t→∞ p→∞
p
lim lim
a.s.,
where B ∼ pβ and Z ∼ N (0, 1) are independent. For the second inequality,
first note that since β̂(λ, q) is the optimal solution we have
(J.7)
1
λkβ̂(λ, q)kqq ≤ F(β̂(λ, q)) ≤ F(0) = kyk22
2
1
2
2
2
2
=
kXβ + wk2 ≤ kXβk2 + kwk2 ≤ σmax (X)kβk22 + kwk22 .
2
We then consider the decomposition β̂(λ, q) = β̂(λ, q)⊥ + β̂(λ, q)k , where
β̂(λ, q)⊥ ∈ ker(X)⊥ and β̂(λ, q)k ∈ ker(X). Since ker(X) is a uniformly
random subspace with dimension p(1 − δ)+ , we can apply Theorem J.4 to
conclude that, there exists a constant c(δ) > 0 depending on δ such that the
following holds with high probability,
kβ̂(λ, q)k22 = kβ̂(λ, q)k k22 + kβ̂(λ, q)⊥ k22 ≤
(J.8)
kβ̂(λ, q)k k21
+ kβ̂(λ, q)⊥ k22
c(δ)p
≤
2kβ̂(λ, q)⊥ k21 + 2kβ̂(λ, q)k21
+ kβ̂(λ, q)⊥ k22
c(δ)p
≤
2kβ̂(λ, q)k21 2 + c(δ)
+
kβ̂(λ, q)⊥ k22 .
c(δ)p
c(δ)
80
HAOLEI WENG, ARIAN MALEKI AND LE ZHENG
Moreover, Hölder’s inequality combined with Inequality (J.7) yields
(J.9)
kβ̂(λ, q)k1
≤
p
kβ̂(λ, q)kqq
p
!1/q
≤
2 (X)kβk2 + kwk2
σmax
2
2
λp
!1/q
.
Using the results from (J.8) and (J.9), we can then upper bound kβ̂(λ, q)k22 :
kβ̂(λ, q)k22 (a) 2
≤
p
c(δ)
2 (X)kβk2 + kwk2
σmax
2
2
λp
!2/q
+
2 + c(δ)
⊥ 2
2 (X) kX β̂(λ, q) k2 .
pc(δ)σmin
2 (X)kβ̂(λ, q)⊥ k2 .
To obtain (a) we have used the fact kX β̂(λ, q)⊥ k22 ≥ σmin
2
We can further bound
(b)
(c)
(d)
2
kX β̂(λ, q)⊥ k22 ≤ 2ky − X β̂(λ, q)k22 + 2kyk22 ≤ 4kyk22 ≤ 8σmax
(X)kβk22 + 8kwk22 .
(b) is due to the simple fact X β̂(λ, q)⊥ = X β̂(λ, q); (c) and (d) hold since
ky − X β̂(λ, q)k22 ≤ 2F(β̂(λ, q)) and inequalities in (J.7). Combining the last
two chains of inequalities we obtain with probability larger than 1 − 2−p
kβ̂(λ, q)k22
2
≤
p
c(δ)
+
2 (X)kβk2 + kwk2
σmax
2
2
λp
!2/q
2 (X)kβk2 + kwk2
16 + 8c(δ) σmax
2
2
·
.
2 (X)
p
c(δ)σmin
Finally, because both σmin (X) and σmax (X) converge to non-zero constants by Theorem J.3 and (β, X, w) is a converging sequence, the right
hand side of the above inequality converges to a finite number.
Lemma 27. Let {β(p), X(p), w(p)} be a converging sequence. Denote the
solution of LQLS by β̂(λ, q), and let {β t }t≥0 be the sequence of estimates
generated from the AMP algorithm.There exists a positive constant C̃ s.t.
Pp
t 2−q
i=1 |µi β̂i (λ, q) + (1 − µi )βi |
lim sup lim sup sup
< C̃, a.s.
p
p→∞ 0≤µi ≤1
t→∞
Proof. For any given 0 ≤ µi ≤ 1, it is straightforward to see
|µi β̂i (λ, q)+(1−µi )βit |2−q ≤ max{|β̂i (λ, q)|2−q , |βit |2−q } ≤ |β̂i (λ, q)|2−q +|βit |2−q .
81
HIGHER ORDER ANALYSIS OF REGULARIZATION TECHNIQUES
Hence using Hölder’s inequality gives us
p
1X
|µi β̂i (λ, q) + (1 − µi )βit |2−q
p
0≤µi ≤1
sup
i=1
≤
1
p
p
X
p
p
p
2−q 1 X
2−q
1 X t 2−q 1 X
2
2
+
.
|βi |
≤
|β̂i (λ, q)|2
|βit |2
p
p
p
|β̂i (λ, q)|2−q +
i=1
i=1
i=1
i=1
Applying Lemma 26 to the above inequality finishes the proof.
The next lemma is similar to Lemma 3.3 in [13]. The proof is adapted
from there.
Lemma 28. Let {β(p), X(p), w(p)} be a converging sequence. Let {β t }t≥0
be the sequence of estimates generated from the AMP algorithm. We have
k∇F(β t )k22
= 0,
t→∞ p→∞
p
lim lim
a.s.
Proof. Recall the AMP updating rule (J.1):
2−q
β t = ηq (X T z t−1 + β t−1 ; τt−1
χ).
According to Lemma 7 part (i) we know β t satisfies
2−q
χq(|β1t |q−1 sign(β1t ), . . . , |βpt |q−1 sign(βpt ))T .
X T z t−1 + β t−1 = β t + τt−1
The rule (J.1) also tells us
z t = y − Xβ t + wt z t−1 ,
where wt is defined in Theorem J.2. Note
∇F(β t ) = −X T (y − Xβ t ) + λq(|β1t |q−1 sign(β1t ), . . . , |βpt |q−1 sign(βpt ))T .
We can then upper bound ∇F(β t ) in the following way:
1
1
2−q
√ k∇F(β t )k2 = √ k − X T (y − Xβ t ) + λ(τt−1 χ)−1 X T z t−1 + β t−1 − β t k2
p
p
1
2−q −1
= √ k − X T (z t − wt z t−1 ) + λ(τt−1
χ) (X T z t−1 + β t−1 − β t )k2
p
≤
2−q
λkβ t−1 − β t k2 kX T (z t−1 − z t )k2 |λ + τt−1 χ(wt − 1)| · kX T z t−1 k2
+
+
√
2−q √
2−q √
p
τt−1
χ p
τt−1
χ p
≤
2−q
λkβ t−1 − β t k2 σmax (X)kz t−1 − z t k2 σmax (X)|λ + τt−1 χ(wt − 1)| · kz t−1 k2
+
+
.
√
2−q √
2−q √
p
τt−1
χ p
τt−1
χ p
82
HAOLEI WENG, ARIAN MALEKI AND LE ZHENG
By Lemma 23, Theorem J.2 part (i)(ii) and Theorem J.3, it is straightforward to confirm that the first two terms on the right hand side of the
last inequality vanish almost surely, as p → ∞, t → ∞. For the third term,
Lemma 23 and Theorem J.2 part (iii)(iv) imply
2−q
|λ + τt−1
χ(wt − 1)| · kz t−1 k2
2−q √
t→∞ p→∞
τt−1
χ p
√
1 0
δτ∗
2−q
2−q
λ
−
τ
χ
1
−
= 0,
=
Eη
(B
+
τ
Z;
τ
χ)
∗
∗
∗
δ q
τ∗2−q χ
lim lim
a.s.
To obtain the last equality, we have used Equation (2.3).
We are in position to prove the asymptotic equivalence between AMP
estimates and bridge regression.
Proposition 1. Let {β(p), X(p), w(p)} be a converging sequence. Denote the solution of LQLS by β̂(λ, q), and let {β t }t≥0 be the sequence of
estimates generated from the AMP algorithm. We then have
(J.10)
lim lim
t→∞ p→∞
1
kβ̂(λ, q) − β t k22 = 0,
p
a.s.
Proof. We utilize Lemma 24. Let β +r = β̂(λ, q), β = β t . If this pair of β
krk2
t
k2
and r satisfies the conditions in Lemma 24, we will have p 2 = kβ̂(λ,q)−β
p
being very small. In the rest of the proof, we aim to verify that the conditions in Lemma 24 hold with high probability and establish the connection
between the iteration numbers t and in Lemma 24.
a. Condition (i) follows from Lemma 26:
√
krk2
kβ̂(λ, q)k2
kβ t k2
lim lim sup √ ≤ lim sup
+ lim lim √ ≤ 2 C,
√
t→∞ p→∞
t→∞p→∞
p
p
p
p→∞
b.
c.
d.
e.
f.
Condition
Condition
Condition
Condition
Condition
a.s.
(ii) holds since β̂(λ, q) is the optimal solution of F(β).
(iii) holds by Lemma 28. Note that → 0, as t → ∞.
(iv) is due to Lemma 27.
(v) is the result of Theorem J.3.
(vi) is a direct application of Theorem J.4.
Note all the claims made above hold almost surely as p → ∞; and → 0 as
t → ∞. Hence the result (J.10) follows.
HIGHER ORDER ANALYSIS OF REGULARIZATION TECHNIQUES
83
Based on the results from Theorem J.1, Lemma 23 and Proposition 1, we
can use exactly the same arguments as in the proof of Theorem 1.5 from [13]
to finish the proof of Theorem 2.1. Since the arguments are straightforward,
we do not repeat it here.
H. Weng
Department of Statistics
Columbia University
1255 Amsterdam Avenue
New York, NY, 10027
USA
E-mail: [email protected]
A. Maleki
Department of Statistics
Columbia University
1255 Amsterdam Avenue
New York, NY, 10027
USA
E-mail: [email protected]
L. Zheng
Department of Electrical Engineering
Columbia University
1300 S. W. Mudd Building, MC 4712
500 W. 120th Street
New York, NY 10027
USA
E-mail: [email protected]
| 7cs.IT
|
Quantum Artificial Life in an IBM Quantum Computer
U. Alvarez-Rodriguez,1, ∗ M. Sanz,1 L. Lamata,1 and E. Solano1, 2, 3
arXiv:1711.09442v1 [quant-ph] 26 Nov 2017
1
Department of Physical Chemistry, University of the Basque Country UPV/EHU, Apartado 644, 48080 Bilbao, Spain
2
IKERBASQUE, Basque Foundation for Science, Maria Diaz de Haro 3, 48013, Bilbao, Spain
3
Department of Physics, Shanghai University, 200444 Shanghai, People’s Republic of China
(Dated: November 28, 2017)
We present the first experimental realization of a quantum artificial life algorithm in a quantum computer. The quantum biomimetic protocol encodes tailored quantum behaviors belonging
to living systems, namely, self-replication, mutation, interaction between individuals, and death,
into the cloud quantum computer IBM ibmqx4. In this experiment, entanglement spreads throughout generations of individuals, where genuine quantum information features are inherited through
genealogical networks. As a pioneering proof-of-principle, experimental data fits the ideal model
with accuracy. Thereafter, these and other models of quantum artificial life, for which no classical
device may predict its quantum supremacy evolution, can be further explored in novel generations
of quantum computers. Quantum biomimetics, quantum machine learning, and quantum artificial
intelligence will move forward hand in hand through more elaborate levels of quantum complexity.
Introduction
As described by Deutsch, a quantum computer is a device that intends to fulfill the Deutsch-Church-Turing principle,
namely, to efficiently simulate a finitely realizable physical system in the framework of quantum mechanics [1]. In this
context, quantum supremacy would be reached when a quantum processor outperforms classical computers realizing
a given task. Along these lines, several proposals to achieve quantum supremacy for a variety of quantum algorithms
and quantum simulations have been proposed [2–7].
The keyword “quantum” has overflowed the limits to which was initially constrained and, currently, incessantly
spreads through the interdisciplinary scientific literature. Indeed, it is a source of inspiration for the breeding extensions of already existing models with their quantum counterparts [8–10]. Besides the appealing intellectual exercise,
this hybridization is often motivated by a plausible improvement in the conditions and enhancement in the efficiency
of the developed protocols. From all possible ramifications, including quantum machine learning and quantum artificial intelligence, our research in quantum biomimetics is concerned with the design of a framework for quantum
algorithms based on the imitation of biological processes, belonging to the macroscopic classical complexity, and
brought down by design to the microscopic quantum realm [11–17]. There may not always be a neat analogy between
the physical models underlying our protocols and those used to describe real biological systems, but our proposed
effective dynamics only partially aims at emulating core aspects of the mimicked process. From a wide perspective in
the history of arts and science, close imitation is a natural first layer and wish in the aesthetic process. In this sense,
plain simulation is a valid and fruitful engineering playground, where analogies abound and serve as communicating
vessels between unconnected fields. Our central goal in quantum simulations and quantum computing is to go beyond
it, through a higher creativity challenge, in the search of a second layer of a major art.
In the particular scenario of artificial life, simple models of organisms are able to undergo most common stages
of life in a controlled virtual environment [18–20]. When extending this to the quantum realm, particularities of
quantum physics, such as its limitation to linear dynamics, the no-cloning theorem, or the exponentially growing
dimensionality of Hilbert spaces, play a relevant role. The quantum artificial life protocol we have engineered and
implemented goes beyond the straightforward quantization of existing classical models. In this sense, and with similar
spirit of other contributions [21–24], it is noteworthy to mention that we leave open the question whether the origin
of life is genuinely quantum mechanical. What we prove here is that microscopic quantum systems can efficiently
encode quantum features and biological behaviours, usually associated with living systems and natural selection.
In this article, we report the first experimental implementation of a model for quantum artificial life [13] into a
quantum computer. To this end, we make use of the facilities provided by the IBM ibmqx4 quantum computing
chip [25]. This work should be aligned with the ramping developments in classical and quantum machine learning
and artificial intelligence: the development of algorithms and devices with the capacity to interpret and mimic human
behaviors in order to solve useful problems and improve the interaction with human beings. Along these lines, we may
∗ Electronic
address: [email protected]
2
foresee a future in which these idealized machines hybridize the knowledge in machine learning, artificial intelligence,
and artificial life, with an internal structure and dynamics following the laws of quantum physics, as is already
happening in the classical domain [26].
Results
We begin with a brief description of the model for quantum artificial life [13], whose most important elements are
the quantum living units or individuals. Each of them is expressed in terms of two qubits that we call genotype
and phenotype. The genotype contains the information describing the type of living unit, an information that
is transmitted from generation to generation. The state of the phenotype is determined by two factors: the genetic
information and the interaction between the individual and its environment. This state, together with the information
it encodes, is degraded during the lifetime of the individual.
The goal of the proposed model is to reproduce the characteristic processes of Darwinian evolution, adapted to the
language of quantum algorithms and quantum computing. The self-replication mechanism is based on two partial
quantum cloning events, an operation that entangles either the genotype or the phenotype with a blank state and
copies a certain expectation value of the original qubit in both of the outcome qubits. In this set of experiments,
the self-replication consists in duplicating the expectation value of σz in the genotype, in a blank state that will be
transformed in the genotype of the individual in the next generation [12]. The process is completed by copying again
σz of the new genotype in another blank state that will be transformed in the phenotype of the new individual. The
next subprotocol in the algorithm is the interaction between the individuals and the environment, which emulates the
aging of living units until an asymptotic state that represents its death. This evolution is encoded in a dissipative
dynamics that couples a bath with each of the phenotype qubits, with σ = |0ih1| as Lindblad operator. The effective
lifetime, i.e., the time the phenotype needs to arrive to the dark state of the Lindbladian up to a given error, depends
implicitly on the genotype. The protocol also accounts for mutations, performed via random single qubit rotations in
the genotype qubits or via errors in the self-replication process [13]. The final ingredient is the interaction between
individuals, which conditionally exchange the phenotypes depending on the genotypes [13]. This behavior is achieved
via a four-qubit unitary operation, where genotypes and phenotypes play the role of control and target qubits,
respectively. The conjunction of these components leads to a minimal but consistent Darwinian quantum scenario.
The protocol may be enriched when including spatial information, either quantum or classical, or increasing the model
complexity by considering a larger set of observables.
The first step for this implementation is to express each of the building blocks of the previous paragraph in terms
of the quantum gates available in the superconducting circuit architecture of IBM cloud quantum computer [25].
Since we have selected σz as the observable to clone, every partial quantum cloning event requires the realization of
a UCNOT gate, that can be directly performed in the experiment. Regarding the interaction with the environment,
we have adapted our protocol, because the experimental device does not allow to realize a conditional projection of
the quantum state to the |0i in the phenotypes. The alternative we propose is to implement the transition between
the basis states as a sequence of small rotations in σy for the phenotype qubits, e−iσy θ , with θ tuned according to the
duration of each simulated time step. This type of rotation can be realized in the experiment via the u3 (θ, φ, λ) gate,
cos θ2
−eiλ sin θ2
u3 (θ, φ, λ) =
,
eiφ sin θ2 ei(φ+λ) cos θ2
with u3 (θ, 0, 0). The same gate acting on genotype qubits can be used for the mutation events. Ideally, and in order
to emulate their randomness both in the phase θ and in the presence or absence of the event, we could design the
experimental runs following a classical program. For making the procedure tractable, we could discretize the range of
θ in n values, and divide the total experimental runs in n + 1 groups to account for each of the different possibilities.
The weight, or number of runs for each group, would depend on our selection for the mutation rate as well as on the
random parameters obtained with the external program. However, constrained by the flexibility of the experimental
device, we propose a less realistic but pragmatic procedure: assume that the mutations will only be of a specific θ,
and therefore eliminate a source of randomness and diversity in the protocol. The single-qubit gate accounting for
the mutations will be σx . Regarding the randomness in the presence or absence of mutation events, we will have to
adapt our algorithm to perform the mutations in groups of 1024 experimental runs, and achieve the mutation rate
accordingly. The last subprotocol, the inter-individual interactions, requires the implementation of the interaction
gate UI , whose effect is to exchange two pairs of quantum levels, while leaving the rest unaltered, as UI |xxyyi = |xyyxi
and UI |xyyxi = |xxyyi, for {x, y} ∈ {0, 1}. The challenge is to decompose UI in terms of the gates offered by the
†
experimental setup. Our solution is given by UI = S23 U12 (1 ⊗ F )U12 S23 , with F = U43 C34 C24 U23 C34
U43 U23 . Here,
the first and second subindices denote the control and target qubit respectively, U is the controlled-not gate, S is the
3
SWAP gate and C is the controlled square root of not gate. These can be rewritten in terms of the controlled-not √
gate
√
as Sij = Uij Uji Uij and C12 = (T ⊗ P u3 (−π/4, 0, 0))U12 (1 ⊗ u3 (π/4, 0, 0))U12 (1 ⊗ P † ), with P = σz and T = P .
An additional relation to point out is that the control target behavior in the controlled-not gate can be exchanged by
introducing Hadamard gates, U21 = (H ⊗ H)U12 (H ⊗ H). This is a useful formula for designing the quantum circuit
in an experimental platform that only allows a single direction for the implementation of the UCNOT .
A.
Experiments
Interaction between two individuals. We start with a quantum circuit designed for reproducing the dynamics
of two interacting individuals. Two precursor genotypes are initialized in |ψig1 = cos π8 |0i + sin π8 |1i and |ψig2 =
3π
cos 3π
8 |0i + sin 8 |1i with u3 . Afterwards, both individuals are completed by copying the genotype qubits in blank
3π
states via UCNOT gate, |ψi1 = cos π8 |00i + sin π8 |11i and |ψi2 = cos 3π
8 |00i + sin 8 |11i. In terms of θ1 = π/8 and
θ2 = 3π/8, the complete state, |ψ1 i ⊗ |ψ2 i, reads
|ψi = cos θ1 cos θ2 |0000i + cos θ1 sin θ2 |0011i + sin θ1 cos θ2 |1100i + sin θ1 sin θ2 |1111i.
We now apply the interaction gate UI to conclude this building block,
UI |ψi = cos θ1 cos θ2 |0000i + cos θ1 sin θ2 |0110i + sin θ1 cos θ2 |1001i + sin θ1 sin θ2 |1111i.
Notice that the interaction fully exchanges the phenotypes, hσz i2 and hσz i4 , that are now equal to the opposite
genotype, hσz i1 = cos2 θ1 − sin2 θ1 = hσz i4 and hσz i3 = cos2 θ2 − sin2 θ2 = hσz i2 .
The experiment is planned to reduce the total errors induced by the use of two-qubit gates. Consequently, we have
reordered the initial Hilbert space |g1 p1 g2 p2 i, where gi is genotype and pi is phenotype, as |p2 g2 p1 g1 i and assigned
each of these qubits to the experimental ones |Q0 Q1 Q2 Q3 i. See Fig. 1 for the remaining quantum circuit diagram.
|0i0
|0i1 3/4
|0i2
H
H
H
H
H
H S
H
H
1/4
1/4,0,1
1/4
-1/4
T+
T
1/4
-1/4 0,-1/2
H
0,3/4
H
H
H
H
H
|0i3 1/4
FIG. 1: Quantum circuit diagram for the protocol of two interacting individuals. The solid and dashed line in the gates denote
u2 (φ, λ) and u3 (θ, φ, λ), where the phase is indicated in units of π.
p√
√
The results, in Table I, agree with the ideal case with a 71.58% fidelity according to F (ρ1 , ρ2 ) = Tr
ρ1 ρ2 ρ1 [27],
computed with the data provided by the experiment when measuring in the computational basis. Therefore, this result
is valid, but not equivalent to the one that is expected when the complete wave function is considered, which is hindered
by the use of full tomography. The expectation values extracted from the data show a reasonable overlap between p1
and g2 , as expected, and a considerable distance between g1 and p2 .
Basis Element 0000 0001 0010 0011 0100 0101 0110 0111 1000 1001 1010 1011 1100 1101 1110 1111
Measured events 1104 338 647 542 693 355 2687 519 104 144 114 1
99 132 261 353
Predicted events 1012 0
0
0
0
0 5896 0
0 174 0
0
0
0
0 1012
TABLE I: Interaction between two individuals. Number of measurements for every element of the four-qubit basis. The
experimental values for hσz i are (0.70,-0.26,-0.27,0.41) while the ideal values are (0.71,-0.71,-0.71,0.71). Notice that the mapping
of reordering the qubits has been inverted to achieve the results in the |g1 p1 g2 p2 i basis.
Interaction with the environment. In this round of experiments we test the combination of partial quantum cloning
events and dissipation. A precursor genotype is initialized in |ψig1 = cos π3 |0i + sin π3 |1i, and the individual completed
with a first partial quantum cloning event via UCNOT and a blank state, |ψi1 = cos π3 |00i + sin π3 |11i. Then, a
single qubit rotation, u3 (π/8, 0, 0), is applied in the phenotype, that substitutes the dissipation in a discrete manner,
4
losing its exponential character. The course of time is simulated by this gate, by implementing one of them for every
simulated time step. Subsequently, a second individual is created in a complete self-replication event with two partial
quantum cloning operations. To conclude, u3 (π/8, 0, 0) is implemented again on both genotypes associated with a
next time step. We assign the Hilbert space of the simulating device as |Q0 Q1 Q2 Q4 i → |p2 g2 g1 p1 i to maximize the
efficiency of the protocol. See Fig. 2 for the quantum circuit diagram. The results, shown in Table II, account for
similar probability distributions between the ideal and the real data with a fidelity of 91.18%, as before computed
only for the computational basis.
|0i0
1/8
|0i1
|0i2
|0i4
2/3
1/4
FIG. 2: The initialization of a genotype before three partial quantum cloning events. The first of these will produce an initial
individual and the remaining two will replicate it into a second one. The protocol continues with single-qubit gates that emulate
the dissipation. The squares denote U3 (θ, φ, λ) gates where the number indicates the value of θ.
Basis element 0000 0001 0010 0011 0100 0101 0110 0111 1000 1001 1010 1011 1100 1101 1110 1111
Measured events 1491 103 42 224 387 46 31 249 67 91 108 916 149 354 439 3495
Predicted events 1682 66
0
0 288 11
0
0
0
0
34 866 0
0 200 5045
TABLE II: Self-replication and interaction with the environment in the σz basis. Number of measurements for every
element of the four-qubit basis. The experimental values for hσz i are (-0.37,-0.26,-0.34,-0.34) while the ideal values are (-0.5,0.35,-0.5,-0.46). Notice that the mapping of reordering the qubits has been inverted to achieve the results in the |g1 p1 g2 p2 i
basis.
For the self-replication instance, there is an additional property of the model that only arises when measuring some
purely quantum correlations of the system. The partial quantum cloning operation entangles the qubits which are
involved on it, transmitting hσx i of the original state into hσx ⊗ σx i. Note that this data can be extracted from the
experiment when measuring on the σx basis, which is done by introducing a Hadamard gate in every entry before
projecting. Therefore, one has to compute hσz ⊗ σz ⊗ σz ⊗ σz i in the new basis, to retrieve hσx ⊗ σx ⊗ σx ⊗ σx i. This
technique is based on the equality Tr[σx ρ] = Tr[σz HρH], since σx = Hσz H. Even if the calculation for the fidelity
yields a satisfactory 93.45%, the value of hσx ⊗ σx ⊗ σx ⊗ σx i still shows a sizable error with respect to the ideal one,
as we show in Table III.
Basis element 0000 0001 0010 0011 0100 0101 0110 0111 1000 1001 1010 1011 1100 1101 1110 1111
Measured events 753 246 277 52 448 747 569 513 343 493 616 177 717 679 345 749
Predicted events 624 1
77 547 157 1150 704 603 77 547 624 1 704 603 157 1150
TABLE III: Self-replication and interaction with the environment in the σx basis. Number of measurements for every
element of the four-qubit basis rotated to σx . The experimental value for hσx ⊗ σx ⊗ σx ⊗ σx i is 0.22 while the ideal value is
0.56. Again, the qubits have been reordered to coincide with the ideal results in the |g1 p1 g2 p2 i basis.
The implementation of mutations requires to combine the outcome of different designs of quantum circuit diagrams
and, therefore, experimental runs. In this case, we consider that a mutation event, which can affect both individuals,
is simulated with a σx . The complete result is achieved when gathering data from 4 different groups of experiments,
that correspond to the cases of mutation on the first genotype, mutation on the second genotype, mutation on both
genotypes, and no mutation. We have performed 1024 experimental runs for each of the three cases with mutations
and 8192 runs for the no-mutation rate. These results have been combined with the ones shown in Table II, that
coincide with the no-mutation case, with the goal of reducing the mutation rate for each individual, which takes a
5
final value of 2/19. See Table IV for the measured data with a fidelity of 94.86% with respect to the ideal case in the
σz basis.
Basis element 0000 0001 0010 0011 0100 0101 0110 0111 1000 1001 1010 1011 1100 1101 1110 1111
Measured events 3201 282 385 447 1344 217 541 494 235 608 297 2026 464 1509 946 6325
Predicted events 3466 137 309 12 1174 46 622 25 12 303 76 1930 26 659 401 10123
TABLE IV: Self-replication, interaction with the environment and mutations. Number of measurements for every
element of the four-qubit basis rotated to σx . The experimental values for hσz i are (-0.28,-0.23,-0.19,-0.23) while the ideal
values are (-0.40,-0.35,-0.40,-0.37). The experimental basis is also permuted to coincide with the ideal results in the |g1 p1 g2 p2 i
basis.
I
IV
a
b
c
-1
1
d
FIG. 3: Visualization of the ideal processes in experiments I and IV. We depict the individuals as combinations of two diamonds
that represent the genotype and phenotype qubits. The color in the genotype qubit, the upper diamond of each pair, depends
on the value of σz as indicated in the color bar. The color in the phenotype qubit is the same as in the genotype one, as the color
is meant to be showing the genetic information. Moreover, the opacity of this color is modified according to the expectation
value of σz being limited by the value of 1 that corresponds to the blank qubits. In both cases the right arrow separates
two consecutive time steps. Following these clarifications, we can see the exchange of phenotypes in I, and the self-replication
followed by different mutation possibilities in IV.
Realization of the complete model of quantum artificial life. The last round of experiments is devoted to the
reproduction of the aggregate of properties in the quantum artificial life algorithm. In order to maintain the fidelity
in values that allow us to claim that the experiment is indeed behaving according to the protocol, we restrict our
analysis to the case of two interacting individuals, which undergo mutations and dissipation. Then, the quantum
circuit diagram, shown in Fig. 4, is an upgraded version of the one shown in Fig. 1 that includes u3 (π/8, 0, 0) for
simulating the dissipation in the phenotypes. For the mutations, we follow the same strategy as in the previous
subsection, combining the data generated with different quantum circuit diagrams each of them emulating a specific
case of the presence or absence of mutation instances. In particular, 3 rounds of 8192 runs emulating the no-mutation
case and 1024 runs for each of the mutation cases determine a mutation rate of 2/27. The post-processing of the data,
in Table V, matches the ideal probability distribution in the computational basis with a fidelity of 93.94%.
6
|0i0
1/8
|0i1 3/4
|0i2
1/8
H
H
H
H
H
H S
H
H
1/4
1/4
1/4,0,1
-1/4
1/4
T+
-1/4 0,-1/2
H
0,3/4
H
T
1/8
H
H
H
H
1/8
|0i4 1/4
FIG. 4: Quantum circuit diagram for the complete quantum artificial life protocol. The squares with dashed and solid lines
represent u2 (φ, λ) and u3 (θ, φ, λ) single-qubit gates, respectively. The rotation angles are expressed in radians.
Basis element 0000 0001 0010 0011 0100 0101 0110 0111 1000 1001 1010 1011 1100 1101 1110 1111
Measured events 3449 1598 2656 2053 2361 298 6369 2166 521 500 629 529 594 692 656 1146
Predicted events 2221 924 2221 410 410 251 12401 2207 237 410 237 838 584 251 410 2207
TABLE V: Complete model. Number of measurements for every element of the four-qubit basis rotated to σx . The
experimental values for hσz i are (0.60,-0.09,-0.24,0.31) while the ideal values are (0.60,-0.43,-0.60,0.43). The experimental basis
is also permuted to coincide with the ideal results in the |g1 p1 g2 p2 i basis.
0.8
0.6
0.4
0.2
0
0.6
0.4
0.2
IV
0.5
0.05
0
Basis element
V
0.4
Probability
Probability
III
0.1
Basis element
0.5
0.4
0.3
0.2
0.3
0.2
0.1
0.1
0
0.15
0.6
0
Basis element
II
Probability
I
Probability
Probability
0.8
0
Basis element
Basis element
FIG. 5: Experimental and ideal probability distributions for all the cases analyzed. In each pair of columns the blue one in the
left denotes the experimental value while the yellow one in the right denotes the ideal estimation. The labels in the subplots
correspond to the tables for the different quantum artificial life instances.
Discussion
Regarding errors in the experimental protocol, even if the fidelities achieved are satisfactory, they do not correspond
to the fidelities of the complete quantum state. In this sense, the prediction for the number of events to measure is
done by simply multiplying the probability distribution by the number of events. These do not exactly match the
experimental data (see Table VI for all experimental outcomes). Indeed, the distances have to be properly weighed,
7
since the probability distribution is the key quantity to be extracted (see Fig. 5). Moreover, the overlap between
expectation values of observables in the measurement basis is lower than the fidelity of the probability distribution
for all cases analyzed here.
I
II
III
IVa
IVb
IVc
IVd
Va
Vb
Vc
Vd
Ve
Vf
0000
1104
1491
753
1511
136
39
24
1032
1087
1046
7
219
58
0001
338
103
246
121
14
28
16
422
414
445
6
269
42
0010
647
42
277
46
3
149
145
756
863
871
48
62
56
0011
542
224
52
201
14
7
1
594
696
631
45
63
24
0100
693
387
448
395
542
13
7
737
697
723
47
74
83
0101
355
46
747
86
32
34
19
51
44
46
5
87
65
0110
2687
31
569
27
9
46
428
2121
2076
2056
24
44
48
0111
519
249
513
213
6
3
23
731
692
634
61
27
21
1000
104
67
343
102
12
22
32
88
107
111
99
27
89
1001
144
91
493
192
29
137
159
109
107
114
89
15
66
1010
114
108
616
162
16
6
5
132
134
154
119
15
75
1011
1
916
177
984
122
2
2
113
97
148
122
7
42
1100
99
149
717
215
17
68
15
99
99
97
39
28
232
1101
132
354
679
635
12
444
64
179
191
162
48
57
55
1110
261
439
345
478
1
16
12
186
188
192
35
12
43
1111
353
3495
749
2824
5
1
0
383
304
348
68
18
25
TABLE VI: Experimental data. The rows denote the number of events in each of the elements of the computational basis for
all the instances considered. For the composite experiments, in order to retrieve the measurements shown in previous tables,
one has to add all the events in each of the individual runs. In the particular case of IV, the data in II also contributes to the
final result.
In parallel, the assignment between the simulated and the simulating Hilbert spaces is designed to maximize the
fidelity according to the calibration parameters provided by IBM. Nevertheless, the recalibration of the circuit changes
the gate and readout errors, so we reevaluate our circuit according to the new parameters and adapt it when the
fidelity outcome can be improved with a different labeling of qubits. Consequently, the performance of the different
experiments is not directly comparable, since they have been implemented under unequal conditions.
Despite the different factors degrading the implementation, the performed experiments reproduce the characteristic
properties of the sought quantum natural selection scenario. We have observed how the partial quantum cloning events
allow us to inherit the information of hσz i from qubit to qubit, and use this property to encode the self-replication
process. We have also seen how nonzero quantum correlations assure that both individuals have been part of a same
event in their timelines, in this case self-replication. Another relevant characteristic of the analysis is the inclusion
of mutations as a source of randomness that, counterintuitively, significantly improved the fidelity of the quantum
algorithm outcome. Our explanation is that mutations tend to homogenize the probability distribution, which is the
same effect as the one produced by the errors naturally present in the experimental platform. Surprisingly, in the
classical realm, mutations also help the species to adapt to changing environments.
This experimental realization of the proposed quantum algorithm represents the consolidation of the theoretical
framework of quantum artificial life. The improvement in scalable quantum computers will soon allow us for more
accurate quantum emulations with growing complexity towards quantum supremacy, even considering spatial variables
for the individuals and a mechanism for tracing out death living units. These future developments should lead towards
an autonomous character of the set of individuals, i.e., the evolution will be an intrinsic property of the system, and
the desired behavior will emerge without following the instructions of a previously designed quantum algorithm. In
this context, the system would be transformed into an intelligent source of quantum complexity whose evolutionary
plot for a large number of individuals may not be predicted classically and, consequently, has the capacity to produce
unexpected results when scaled up. An interesting question to address is to establish the relation existing between
the parameters defining the fundamental processes of the model, and the emergent multiqubit quantum state. Along
these lines, and following the frame of artificial life oriented genetic algorithms [28], we speculate about the idea of
channeling this complexity to encode optimization problems by tuning the self-replication, mutation and dissipation
rates that define the evolution. Furthermore, recent advances in quantum machine learning constitute a promising
material to work with in the study of algorithms combining the properties of both fields, pursuing the design of
intelligent and replicating quantum agents. Therefore, the creation of these quantum living units and their possible
applications are expected to have deep implications in the community of quantum simulation and quantum computing
in a variety of quantum platforms.
All in all, the experiments presented here entail the validation of quantum artificial life in the lab and, in particular,
in cloud quantum computers as that of IBM. Still another interesting step would be the development of autonomous
8
quantum devices following the theoretical and experimental results in quantum cellular automata [29, 30]. Our
quantum individuals are driven by an adaptation effort along the lines of a quantum Darwinian evolution, which
effectively transfer the quantum information through generations of larger multiqubit entangled states. We believe
that the presented results and vision, both in theory and experiments, should hoist this innovative research line as
one of the leading banners in the future of quantum technologies.
[1] D. Deutsch, Quantum theory, the Church-Turing principle and the universal quantum computer, Proc. Roy. Soc. London
A 400, 97 (1985).
[2] R. P. Feynman, Simulating physics with computers, Int. J. Theor. Phys. 21, 467 (1982).
[3] D. Deutsch and R. Josza, Rapid solutions of problems by quantum computation, Proc. Roy. Soc. London A 439, 553
(1992).
[4] L. K. Grover, A fast quantum mechanical algorithm for database search, Proceedings, 28th Annual ACM Symposium on
the Theory of Computing (STOC) 212 (1996).
[5] S. Lloyd, Universal quantum simulators, Science 273, 1073 (1996).
[6] P. W. Shor, Polynomial time algorithms for prime factorization and discrete logarithms on a quantum computer, SIAM J.
Sci. Statist. Comput. 26, 1484 (1997).
[7] S. Aaronson and A. Arkhipov, The Computational Complexity of Linear Optics, Proceedings, 43rd Annual AMC Symposium
on the Theory of Computing (STOC) 333 (2011).
[8] S. Wiesner, Conjugate coding, SIGACT News 15 78, (1983); C. H. Bennett and G. Brassard, Quantum cryptography:
Public key distribution and coin tossing. In Proceedings of IEEE International Conference on Computers, Systems and
Signal Processing, volume 175, page 8. New York, 1984; A. K. Ekert, Quantum cryptography based on Bells theorem,
Phys. Rev. Lett. 67, 661 (1991).
[9] Y. Aharonov, L. Davidovich, and N. Zagury, Quantum random walks, Phys. Rev. A 48, 1687 (1993); N. Shenvi, J. Kempe,
and K. B. Whaley, Quantum random-walk search algorithm, Phys. Rev. A 67, 052307 (2003).
[10] J. Biamonte, P. Wittek, N. Pancotti, P. Rebentrost, N. Wiebe, and S. Lloyd, Quantum machine learning, Nature 549, 195
(2017); V. Dunjko and H. J. Briegel, Machine learning & artificial intelligence in the quantum domain, arXiv:1709.02779
(2017); M. Schuld, I. Sinayskiy, and F. Petruccione, An introduction to quantum machine learning, Contemporary Phys. 56,
172 (2014); U. Alvarez-Rodriguez, L. Lamata, P. Escandell-Montero, J. D. Martı́n-Guerrero, and E. Solano, Supervised
Quantum Learning without Measurements, Scientific Reports 7, 13645 (2017); L. Lamata, Basic protocols in quantum
reinforcement learning with superconducting circuits, Scientific Reports 7, 1609 (2017).
[11] E. Schrödinger, What is life? (Cambridge University Press, Cambridge, UK, 1944).
[12] U. Alvarez-Rodriguez, M. Sanz, L. Lamata, and E. Solano, Biomimetic Cloning of Quantum Observables, Scientific Reports
4, 4910 (2014).
[13] U. Alvarez-Rodriguez, M. Sanz, L. Lamata, and E. Solano, Artificial Life in Quantum Technologies, Scientific Reports 6,
20956 (2016).
[14] U. Alvarez-Rodriguez, R. Di Candia, J. Casanova, M. Sanz, and E. Solano, Algorithmic quantum simulation of memory
effects, Phys. Rev. A 95, 020301 (2017) ; U. Alvarez-Rodriguez, A. Perez-Leija, I. L. Egusquiza, M. Gräfe, M. Sanz, L.
Lamata, A. Szameit, and E. Solano, Advanced-Retarded Differential Equations in Quantum Photonic Systems, Scientific
Reports 7, 42933 (2017).
[15] U. Las Heras, U. Alvarez-Rodriguez, E. Solano, and M. Sanz, Genetic Algorithms for Digital Quantum Simulations, Phys.
Rev. Lett. 116, 230504 (2016).
[16] P. Pfeiffer, I. L. Egusquiza, M. Di Ventra, M. Sanz, and E. Solano, Quantum Memristors, Scientific Reports 6, 29507
(2016).
[17] J. Salmilehto, F. Deppe, M. Di Ventra, M. Sanz, and E. Solano, Quantum Memristors with Superconducting Circuits,
Scientific Reports 7, 42044 (2017); M. Sanz, L. Lamata, and E. Solano, Quantum Memristors in Quantum Photonics.
arXiv:1709.07808 (2017).
[18] C. G. Langton, Artificial Life: An overview (MIT Press, Cambridge USA, 1997).
[19] M. Gardner, The fantastic combinations of John Conway’s new solitaire game life, Sci. Am. 223, 120 (1970).
[20] T. S. Ray, Ecology and Optimization of Digital Organisms. Santa Fe Institute (1992). Available at:
http://samoa.santafe.edu/media/workingpapers/92-08-042.pdf (Accessed: 10th October 2017).
[21] M. A. Martin-Delgado, On Quantum Effects in a Theory of Biological Evolution, Sci. Rep. 2, 302 (2012).
[22] D. Abbott, P. C. W. Davies, and A. K. Pati, Quantum Aspects of Life (Imperial College Press, London, 2008).
[23] P. Arrighi and J. Grattage, A Quantum Game of Life. Paper presented at Journées Automates Cellulaires, Turku, Finland
(2010, December 15).
[24] D. Bleh, T. Calarco, and S. Montangero, Quantum Game of Life. EPL 97, 20012 (2012).
[25] IBM Quantum Experience, https://www.research.ibm.com/ibm-q/ (last accessed 10 October 2017).
[26] L. Steels, The Artificial Life Roots of Artificial Intelligence, Artif. Life 1, 75 (1994).
[27] M. A. Nielsen and I. L. Chuang, Quantum Computation and Quantum Information (Cambridge University Press, Cambridge, UK, 2000).
[28] M. Mitchell and S. Forrest, Genetic Algorithms and Artificial Life, Artif. Life 1, 267 (1994).
9
[29] C. S. Lent, P. D. Tougaw, W. Porod, and G. H. Bernstein, Quantum cellular automata, Nanotechnology, 4, 1 (1993).
[30] A. O. Orlov, I. Amlani, G. H. Bernstein, C. S. Lent, and G. L. Snider, Realization of a Functional Cell for Quantum-Dot
Cellular Automata, Science 277, 928 (1997); I. Amlani, A. O. Orlov, G. Toth, G. H. Bernstein, C. S. Lent, and G. L.
Snider, Digital Logic Gate Using Quantum-Dot Cellular Automata, Science 284, 289 (1999); R. P. Cowburn and M. E.
Welland, Room Temperature Magnetic Quantum Cellular Automata, Science 287, 1466 (2000); A. Imre, G. Csaba, L. Ji,
A. Orlov, G. H. Bernstein, and W. Porod, Majority Logic Gate for Magnetic Quantum-Dot Cellular Automata, Science
311, 205 (2006).
Acknowledgements
We thank Armando Pérez-Leija and Alexander Szameit for enthusiastic and enlightening discussions. We acknowledge the use of IBM Quantum Experience for this work. The views expressed are those of the authors and do not reflect
the official policy or position of IBM or the IBM Quantum Experience team. We acknowledge support from Spanish
MINECO/FEDER FIS2015-69983-P, UPV/EHU new PhD program, Basque Government IT986-16, and Ramón y
Cajal Grant RYC-2012-11391.
Author Contributions
U. A.-R. performed the experiments on the IBM Quantum Experience. U. A.-R., M. S., L. L., and E. S. developed
the model and analyzed the data. All authors contributed to writing the manuscript.
Additional Information
Competing financial interests: The authors declare no competing financial interests.
| 2cs.AI
|
Towards Parallelizable Sampling–based
Nonlinear Model Predictive Control
R.V. Bobiti ∗ M. Lazar ∗
arXiv:1701.02660v2 [cs.SY] 12 Jan 2017
∗
Department of Electrical Engineering, Eindhoven University of Technology,
The Netherlands (e–mails: [email protected], [email protected])
Abstract: This paper proposes a new sampling–based nonlinear model predictive control (MPC)
algorithm, with a bound on complexity quadratic in the prediction horizon N and linear in the number of
samples. The idea of the proposed algorithm is to use the sequence of predicted inputs from the previous
time step as a warm start, and to iteratively update this sequence by changing its elements one by one,
starting from the last predicted input and ending with the first predicted input. This strategy, which
resembles the dynamic programming principle, allows for parallelization up to a certain level and yields
a suboptimal nonlinear MPC algorithm with guaranteed recursive feasibility, stability and improved
cost function at every iteration, which is suitable for real–time implementation. The complexity of the
algorithm per each time step in the prediction horizon depends only on the horizon, the number of
samples and parallel threads, and it is independent of the measured system state. Comparisons with
the fmincon nonlinear optimization solver on benchmark examples indicate that as the simulation time
progresses, the proposed algorithm converges rapidly to the “optimal” solution, even when using a small
number of samples.
Keywords: suboptimal nonlinear predictive control, sampling–based optimization, embedded model
predictive control, dynamic programming, constrained control
1. INTRODUCTION
A vast literature on nonlinear model predictive control (NMPC)
has proven both the theoretical (Grüne and Pannek, 2011), as
well as the practical advantage (Magni et al., 2009) of this
method in treating optimally the control of multi–variable nonlinear systems subject to constraints on state and inputs. Common research interests include methods for reducing the complexity of the NMPC algorithms, to make them applicable on
devices of low memory (ASIC, FPGA), such as explicit NMPC
(ENMPC), see, e.g., (Johansen, 2004). Much work has been
done on treating other limiting factors, such as the requirement
of NMPC of solving an optimization problem online. This is
not well achieved by common optimization tools, which have
no specific termination time, especially due to non–convexity
which involves multiple local–minima. Therefore, real time
requirements are not met, which limits the industrial impact of
NMPC. To alleviate these concerns, the literature has proposed
multiple solutions, such as approximate dynamic programming
(DP) (Bertsekas, 2005a), suboptimal MPC (Scokaert et al.,
1999), approximative DP and suboptimal control via rollout
algorithms (Bertsekas, 2005b), NMPC based on system approximation by neural models (Ławryńczuk, 2009).
An alternative strategy in NMPC is to draw samples from either
the state or input space, to design computationally feasible
NMPC methods, see for example, (Piovesan and Tanner, 2009)
which proposes a randomized approach to sampling the space
of predicted input sequences. More recently, in (Chakrabarty
et al., 2016), an ENMPC method was proposed based on
sampling of the state space for continuously differentiable
nonlinear systems. The method therein solves optimization
problems offline to find optimal control sequences, which are
used to construct the ENMPC strategy. While there are still
concerns in ENMPC related to robustness, feasibility of the
offline optimization and finding the neighbors in the sampled
grid for the off–grid states, ENMPC, when successful, reduces
significantly the computational load of MPC at the expense of
an acceptable cost degradation.
Input and state space sampling–methods for solving NMPC
via approximate DP have also been proposed, see (Bertsekas,
2005a), though they inherit the dimensionality issues of DP
(Lee and Lee, 2004).
Another relevant sampling–based strategy, the so–called sampling based MPC (SBMPC), was proposed in (Dunlap et al.,
2010). The method therein is applicable to nonlinear systems
in general, though, its performance is dependent on a user–
specified heuristic. A more ample discussion on sampling–
based DP and SBMPC, in the light of the method proposed in
this paper is reported in Section 2.3.
A common problem of sampling–guided methods for NMPC is
the sampling strategy. For example, with each input sample, a
tree is expanded. After the tree is built, the path of least cost in
the tree is selected from the initial state to the desired state.
If the sampling is performed over the input space, and each
sample is connected to all the samples in the input space for
the next time step in the control horizon, then the tree growth
is exponential with the horizon. Alternatively, as in randomized
MPC (Piovesan and Tanner, 2009), sampling randomly in the
input space, of dimension m, augmented to the horizon of
dimension N requires a large number of samples, in an mN
dimensional space, to achieve a significant accuracy.
In this paper we adopt a suboptimal formulation of NMPC,
as originally proposed in (Scokaert et al., 1999), where it
was shown that feasibility of a solution implies stability under
suitable conditions. This, together with the fact that suboptimal
NMPC has the same inherent robustness properties as optimal
NMPC, see (Pannocchia et al., 2011) and (Lazar and Heemels,
2009), suggest that suboptimal NMPC is a viable and in fact the
best one can hope for when a sampling–guided MPC strategy is
undertaken for the control of nonlinear systems. Furthermore,
we aim at a sampling method which provides a suboptimal
solution that yields good control performance, has a reasonable
computational complexity increase with the prediction horizon
and allows for parallel implementation up to some level.
In this paper, the main idea for achieving this goal is to use
the shifted sequence of predicted inputs from the previous time
step as a warm start, and to iteratively update this sequence
by changing its elements one by one, starting from the last
predicted input and ending with the first predicted input. This
strategy resembles the dynamic programming principle, especially the rollout algorithms, see (Bertsekas, 2005b), which improve a heuristic base policy for optimal control. Additionally,
in this paper, we sample the original input space, which is
typically represented by a proper set U ⊂ Rm . Sampling allows
for parallelization of the calculations performed for updating
each of the elements of the predicted sequence of inputs and
it enables limiting the computational time according to the
requirements of the considered application. An upper–bound
on the complexity of the overall algorithm is quadratic with the
prediction horizon N and linear with the number of samples in
U. This enables the usage of long prediction horizons or real–
time implementation on inexpensive computing devices such as
ASIC and FPGA. The suitability for real–time implementation
is also enhanced by the fact that the algorithm can be stopped
at any iteration performed within a sampling period, while the
complexity of the calculations per iteration depends only on
the horizon N , the number of samples and parallel threads,
and it does not depend on the measured state of the system.
Moreover, the updated predicted sequence of inputs obtained
at any iteration will guarantee recursive feasibility, stability
and an improved cost function under the same conditions as
in suboptimal NMPC (Scokaert et al., 1999).
The remainder of this paper is organized as follows. In Section 2, basic notation is introduced, together with the problem
formulation and a discussion on the relation with the existing
methods. Section 3 presents the main result as a prototype algorithm and its complexity analysis. In Section 4, three examples
illustrate the potential of the developed method, and Section 5
concludes the paper.
2. PRELIMINARIES
2.1 Notation
Let R, R+ , Z and Z+ denote the field of real numbers, the set
of non–negative reals, the set of integers and the set of non–
negative integers, respectively. For every c ∈ R and Π ⊆ R,
define Π≥c := {k ∈ Π | k ≥ c} and similarly Π≤c . Let int(S)
denote the interior of a set S. Let Sh := S × . . . × S for any
h ∈ Z≥1 denote the h–times Cartesian–product of S ⊆ Rn . A
set S ⊂ Rn is called proper if it is non–empty, compact and
0 ∈ int(S).
For a vector x ∈ Rn , the symbol kxk is used to denote
an arbitrary p–norm; it will be made clear when a specific
norm is considered. For a vector x ∈ Rn , define |x| :=
[|x1 | . . . |xn |]T . Also for a vector x ∈ Rn , by max |x| we denote
max{|x1 |, . . . , |xn |}. For a scalar x ∈ R, denote by dxe the
smallest integer number larger than x.
A function α : R+ → R+ is said to belong to class K, i.e.,
α ∈ K, if it is continuous, strictly increasing and α(0) = 0.
Furthermore, α ∈ K∞ if α ∈ K and lims→∞ α(s) = ∞.
2.2 Suboptimal MPC problem formulation
Let us consider the discrete–time system described by
xk+1 = f (xk , uk ),
n
(1)
m
where xk ∈ R is the state and uk ∈ R is the control vector
at discrete–time k ∈ Z+ . We assume that the map f : Rn ×
Rm → Rn satisfies f (0, 0) = 0, which is, the origin is an
equilibrium point for system (1).
The goal of MPC is to regulate the state to the origin while
satisfying control and state constraints, i.e., uk ∈ U ⊂ Rm and
xk ∈ X ⊂ Rn for all k ∈ Z+ , where U and X are proper
sets. MPC relies on a receding–horizon control law in order to
determine, for each k, a finite–sequence of control inputs
U (k) = {uk|k , uk+1|k , . . . , uk+N −1|k },
where N is the control and prediction horizon, which are
considered equal in this paper, for simplicity of exposition. If
the initial state is xk|k = xk and the control sequence is U (k),
the solution of system (1) in closed-loop with U (k) at time k +i
is denoted by φ(xk|k , U (k), i). The current control action uk , is
selected as the first control action in U (k), i.e., uk = uk|k .
To achieve this, optimal MPC minimizes, at each discrete–time
k, a cost function of the type
J(xk|k , U (k)) = Vf (xk+N |k ) +
N
−1
X
L(xk+i|k , uk+i|k ), (2)
i=0
where J : Rn × Rm → R+ is the total cost function, Vf :
Rn → R+ is a terminal cost and L : Rn × Rm → R+ is a stage
cost. The minimization is performed with respect to U (k) and
it is subject to
xk+j|k = f (xk+j−1|k , uk+j−1|k ),
∀j ∈ Z[1,N ] ,
(3)
and to the state and input constraints.
When the dynamics f is a nonlinear, possibly non-convex
function, the optimization of the cost (2) cannot be guaranteed
to converge to a global optimum, in general. Alternatively,
suboptimal MPC was proposed as a viable alternative, see, e.g.,
(Scokaert et al., 1999), to deal with this inherent shortcoming
of nonlinear global optimization.
Assume that the following constraints are required to hold at
each iteration of the MPC problem:
xk+i|k ∈ X, uk+i|k ∈ U,
∀i ∈ Z[0,N −1] ,
(4)
and
xk+N |k ∈ XT ,
(5)
where XT ⊆ X is a proper set which represents a terminal
constraint. Moreover, define by U(xk|k ) the set of control
sequences U (k) which, applied on xk|k , satisfy (3), (4) and (5).
Suboptimal MPC relies on an initial feasible solution, a warm
start sequence Uwarm (k) ∈ U(xk|k ) at each step k, which is
improved iteratively. The suboptimal MPC problem considered
in this paper is formulated as follows:
Problem 2.1. For each k ∈ Z≥1 , given Uwarm (k) find a
sequence U (k) ∈ U(xk|k ) such that
J(xk|k , Uwarm (k)) > J(xk|k , U (k)),
and the constraints (3), (4) and (5) are satisfied.
(6)
Consider now a locally stabilizing control law kf : XT → U.
Assume that XT is a sublevel set of Vf . For stability of the MPC
closed–loop system it is also required that:
• Vf (f (x, kf (x))) + L(x, kf (x)) ≤ Vf (x) for all x ∈ XT ;
• there exist α1 , α2 ∈ K∞ such that α1 (|x|) ≤ Vf (x) ≤
α2 (|x|) for all x ∈ XT ;
• there exist α3 ∈ K∞ such that L(x, u) ≥ α3 (|x|) for all
(x, u) ∈ X × U;
Remark 2.2. The first property above listed implies that XT is
positively invariant for the system xk+1 = f (xk , kf (xk )). The
second and third properties can be satisfied if, for example Vf
and L are positive definite quadratic functions.
These properties imply that the cost function J(·, ·) is a Lyapunov function, see (Mayne and Rawlings, 2009, Lemma 2.14).
As such, if F is the set of states in X for which there exists a control sequence U (k) which satisfies the constraints (3), (4) and
(5), then the solution to Problem 2.1 provides an asymptotically
stabilizing controller with a region of attraction F.
In (Mayne and Rawlings, 2009), an algorithm was proposed for
suboptimal MPC with stability guarantees. Given the current
state xk|k and the previous control sequence U (k − 1) as an
input, the steps of the algorithm therein can be summarized as
follows:
• If xk|k ∈
/ XT , use the warm start sequence:
Uwarm (k) ={uk|k−1 , uk+1|k−1 , . . . ,
uk+N −2|k−1 , kf (xk+N −1|k−1 )}.
(7)
Solve iteratively Problem 2.1 via optimization to improve
Uwarm (k) with a U (k) ∈ U(xk|k ). Apply control uk =
uk|k .
• If xk|k ∈ XT , set uk|k = kf (xk|k ), or, similarly to the previous case, use the warm start and solve an optimization
algorithm iteratively to find an improved control sequence
U (k) ∈ U(xk|k ).
Optimization solvers, in both optimal and suboptimal MPC,
present difficulties in terms of parallelization and a priori
known execution time independently of the current state xk .
To circumvent these drawbacks and enable a computationally
efficient and parallelizable nonlinear MPC algorithm, we propose a sampling based approach to solving Problem 2.1, as
illustrated in Section 3. To this end, we first review existing,
similar approaches within nonlinear MPC.
2.3 Existing approaches based on sampling
This section provides a brief in depth review of two main
existing approaches for NMPC based on sampling, namely
approximate DP and sampling–based MPC, which were also
mentioned in the Introduction.
An approximate version of DP, as a tool for solving optimization problems, was proposed in (Bertsekas, 2005a, Section
6.6.1). DP has been successfully applied for determining explicit solutions for linear MPC controllers, see, e.g., (Muñoz
de la Peña et al., 2004). In nonlinear MPC, the state and input
states are typically discretized to apply DP algorithms. The
main idea is to discretize the state space with a finite grid and
to express each state outside of the grid as an interpolation of
nearby grid elements. The same interpolation law is applied to
compute the cost of the current nongrid state as a function of the
costs of the nearby grid states. As such, by discretizing both the
state space for each time in the control horizon and the input
space, a transition diagram is obtained which approximates
the dynamics of the system in the continuous space. On this
discrete transition system, DP is applied to determine the path
with the smallest cost, which, for a given initial state xk|k ,
provides the control sequence U (k).
Solving MPC with DP via discretization suffers from the “curse
of dimensionality”, due to sampling of both state and input
spaces and the requirement for constructing the complete transition diagram, by evaluating the subsequent state and cost for
each sampled state and all the samples in the input space.
An alternative to approximate DP, namely sampling based
MPC, was developed within the area of Robotics, where typically optimization problems arising in control are non–convex,
due to either kinematic constraints or constraints posed by
obstacle avoidance. Sampling–based motion planning such as
rapidly-exploring random trees (RRTs) (LaValle, 1998) or randomized A* algorithms (Likhachev and Stentz, 2008), have
been extensively used to construct trees which connect an initial
state to a final state based on sampling states in the search
space and searching for feasible inputs to connect these states.
To approach issues related to the, possibly unfeasible, search
for an input after sampling only in the state space, a method
(SBMPC) was proposed in (Dunlap et al., 2010), which samples
the input space at each sampling period and creates trees that
contain feasible state trajectories. The optimal path to a goal in
the state–space is then searched for within the tree using goal–
directed search algorithms, such as LPA*. Such algorithms rely
on computing a heuristic measure of the distance from the
current sample to the goal. Selecting the heuristic is however,
not an obvious task for general nonlinear systems.
Therefore, a desirable feature of a sampling–based suboptimal
NMPC algorithm is a non–exponential growth in the tree generated through sampling of the state or input space at each
step in the horizon. Furthermore, it is also desirable to reduce
the dependency of the algorithm on the non–obvious selection
of a heuristic, which significantly impacts the performance of
the sampling–based strategy. To circumvent these issues, an
alternative suboptimal strategy for sampling the input space is
proposed in the next section, based on sequentially updating a
warm start feasible sequence of predicted inputs.
3. MAIN RESULT
In this section we present the proposed sampling–based algorithm for solving suboptimal NMPC problems and we provide
a detailed complexity analysis.
3.1 Prototype algorithm
Inspired by the suboptimal MPC and DP principles, we propose a sampling–based solution to Problem 2.1, which, by the
mechanism involved in the iterative improvement of the initial
feasible control sequence Uwarm (k), has a low increase with
the control horizon of the computational complexity.
j=0
j=1
j=2
uw
2|0
uw
0|0
u12|0
u22|0
uw
1|0
u32|0
j=1
j=0
u41|0
uw
0|0
u11|0
j=2
u31|0
uw
1|0
uw
2|0
=
u32|0
u21|0
j=1
j=0
u10|0
uw
1|0
j=2
= u41|0
uw
0|0 u2
0|0
u30|0
u40|0
3
uw
2|0 = u2|0
Fig. 1. Sampling principle.
Algorithm 1 Sampling–based suboptimal NMPC algorithm.
Input: N , {nj }j∈Z[0,N −1] , xk|k , X, U, XT , J(·, ·)
w
Uwarm (k) = {uw
k|k , . . . , uk+N −1|k }
Output: U (k), uk
1:
2:
3:
4:
5:
6:
7:
8:
9:
10:
11:
12:
13:
14:
Jsub ← J(xk|k , Uwarm (k));
for all j = N − 1 : −1 : 0 do
Select nj samples uqj+k|k ∈ U, q ∈ Z[1,...,nj ] ;
for all q = 1 : 1 : nj do
if j ≥ 1 then
q
w
U (k)jq = {uw
k|k , . . . , uk+j−1|k , uk+j|k ,
w
uw
k+j+1|k , . . . , uk+N −1|k };
else
w
U (k)jq = {uqk|k , uw
k+1|k , . . . , uk+N −1|k };
if φ(xk|k , U (k)jq , i) ∈ X, ∀i ∈ Z[j+1,N −1] and
φ(xk|k , U (k)jq , N ) ∈ XT then
Jnew ← J(xk|k , U (k)jq );
if Jnew < Jsub then
Jsub ← Jnew ;
Uwarm (k) ← U (k)jq ;
U (k) = Uwarm (k), uk = uw
k|k ;
The principle behind the proposed sampling–based approach is
illustrated in Fig. 1 and formalized in Algorithm 1. In Fig. 1, the
iterative improvement of an initial cost provided by an initial
feasible control sequence Uwarm (k) is illustrated for the case
when N = 3. The algorithm keeps always Uwarm (k) as a
reference sequence, and it covers the horizon in a backward
fashion, in N iterations. Starting with j = N − 1, at each
iteration step, nj samples {uqk+j|k }q∈Z[1,nj ] are drawn from the
input constraint set U. For each sample, the reference sequence
Uwarm (k) is modified in the j th location, and a new sequence,
U (k)jq is obtained. If the following properties hold:
• U (k)jq is a feasible sequence, i.e., the constraints (3), (4)
and (5) hold,
• the new cost, Jnew = J(xk|k , U (k)jq ) decreases with
respect to J(xk|k , Uwarm (k)),
then Uwarm (k) is replaced by U (k)jq and the algorithm continues backwards with respect to the prediction time j, in a
similar manner. With this approach, at any point in time, if the
maximally allowed computational time is exceeded, a feasible,
improved control sequence exists and it can be utilized as a
suboptimal MPC solution.
By choosing to cover the control horizon backwards, we can
reuse at each step j, the states φ(xk|k , Uwarm (k), i) for all
i ∈ Z[1,j] , which, by the feasibility of Uwarm (k), already
satisfy the state and input constraints. This holds not only for
the original Uwarm (k), but for any subsequent improvement
of Uwarm (k). As such, also the stage costs up to the j th state
can be recovered from previous computations. The reusability
of previously computed costs and states by navigating the
control horizon in a backward manner resembles the working
principle of DP. This suggests intuitively that the proposed cost
improvement method could deliver good performance, which
is supported by results obtained in non–trivial case studies, see
Section 4. A formal analysis of convergence towards the DP
solution, as the discrete–time k increases, makes the object of
future research.
When k = 0, we can select an initial sequence Uwarm (0)
as the solution of the optimal MPC problem. In this case, we
can proceed with k = 1. Alternatively, we can select a feasible sequence Uwarm (0), an “oracle”, by randomly selecting
sequences of inputs in U until a feasible Uwarm (0) is found.
In this case, if it is feasible to afford the computational time,
we can proceed with Algorithm 1 in an attempt of obtaining an
improved sequence.
For k ∈ Z≥1 , to choose the input Uwarm (k) for Algorithm 1,
one can use the receding horizon principle of MPC. As such,
the input sequence Uwarm (k) = {uk|k−1 , . . . , uk+N −2|k−1 , u}
is a warm start at time k. If φ(xk−1|k−1 , U (k − 1), N ) ∈
XT and XT is positively invariant for the system xk+1 =
f (xk , kf (xk )), one can select u = kf (xk+N −1|k−1 ). In this
case, Uwarm (k) is a feasible solution, and therefore a candidate
warm start for every k ∈ Z≥1 . If there exists no terminal set XT
and no kf (·), then one can select u ∈ U such that Uwarm (k)
remains feasible, i.e., φ(xk−1|k−1 , Uwarm (k), N ) ∈ X. In these
circumstances, however, stability of the closed loop is not
guaranteed. Such an example is illustrated in Section 4.3.
Remark 3.1. Common sampling schemes are employed for
sampling of the input space U at each iteration, among which
we consider deterministic uniform sampling, which places each
sample at equal distance from each other, to cover uniformly the
space U. Alternatively, “true” random samples can be selected,
which are simpler to draw in higher dimensional spaces. Quasi–
random low–discrepancy sequences, see (Chakrabarty et al.,
2016), may be used as well, considering the fact that they
appear to be random for multiple purposes, such as Monte Carlo
simulations. Such sampling methods, e.g., Sobol or Halton sequences, have been shown, see, e.g., (de Dios Ortúzar et al.,
1994), to better cover the space than random sequences.
Remark 3.2. Assume that kf is a locally stabilizing control law
on XT , a sublevel set of Vf , which is positively invariant for the
system xk+1 = f (xk , kf (xk )) and Vf and L are, e.g., positive
definite quadratic functions. Considering that the sequence
U (k) provided by Algorithm 1 is a suboptimal solution solving
Problem 2.1, then, by the arguments in Section 2.2, Problem 2.1
is recursively feasible and it ensures stability of system (1) in
closed loop with uk = uw
k|k .
Remark 3.3. The working mechanism of Algorithm 1 resembles the working principle of the rollout algorithm described
in the survey paper (Bertsekas, 2005b). Therein, a rollout algorithm improves iteratively a base policy (here, the feasible
control sequence Uwarm (k)) to provide a suboptimal solution
to an optimal control problem via approximative DP and suboptimal control. The working principle of the rollout algorithm
with sampling based MPC, i.e., choosing at time k the iterated
solution of the previous time instance, k − 1, as a warm start,
and a sampling strategy, is not provided therein.
Remark 3.4. Due to the sampling procedure having a discontinuous behaviour, the inputs might be varying more than it is
safe for some applications. This problem might be alleviated
by either smoothening the inputs via interpolation of the obtained input sequence with the initial sequence, or by penalizing
∆uk = uk − uk−1 via constraints or via the cost function, such
that the input variability is limited to acceptable bounds.
3.2 Complexity analysis
In general, if we have p ∈ Z[2,n) processors, then the complexity of Algorithm 1 is
N (N + 1)
C = dn/pe c1
+ c2 N ,
(11)
2
where the term dn/pe appears due to the fact that a thread
can not engage in computations related to a subsequent time
horizon until all the threads have finalized the computations
related to the current time horizon j.
4. ILLUSTRATIVE EXAMPLES
The sampling–based suboptimal NMPC strategy proposed in
Section 3 is illustrated on three nonlinear systems, to highlight
various features of this method. All tests have been performed
on a system with the following specifications: Intel Core i73770 CPU 3.4GHz, 16GB RAM, 64-bit OS.
4.1 Cart–spring system
In order to analyze the complexity of Algorithm 1, the following assumptions are undertaken, for a given state x, input u and
input sequences U , U1 , U2 :
1) The cost of evaluating f (x, u) and performing a feasibility
test f (x, u) ∈ X is c1 ;
2) The cost of evaluating J(x, U ) is c2 ;
3) The cost of comparing J(x, U1 ) < J(x, U2 ) and changing
Jsub and Uwarm (k) if necessary, i.e., steps 11-13 in Algorithm 1, is negligible.
4) The current cost Jsub is instantaneously available for comparison with each of the nj samples according to step 11
in Algorithm 1 and each of the new sequences U (k)jq may
modify Jsub if the new cost Jnew is smaller than Jsub .
The complexity of Algorithm 1 for a given xk|k is the following:
N
−1
N
−1
X
X
C = c1
(N − j)nj + c2
nj .
(8)
j=0
N (N + 1)
+ c2 N.
(10)
2
The complexity bound given in (10), though quadratic in the
prediction horizon, yields a reasonable complexity, considering
that, in NMPC, a horizon N = 10 is considered a reasonably
large horizon.
C = c1
j=0
The method developed in this paper will first be applied to
a system incorporating an exponential nonlinearity, i.e., the
model of a cart with mass M , which is moving on a plane,
see (Raimondo et al., 2009). This cart is attached to a wall via
a spring with elastic constant k varying with the first state k =
k0 e−x1 , where x1 stands for the displacement of the carriage
from the equilibrium position. A damper acts as a resistor in
the system, with damping hd . The discretized nonlinear model
of the cart and spring system is the following:
x1 (k + 1)
x(k + 1) =
= f1 (x(k)) + F2 u(k),
(12)
x2 (k + 1)
where
x1 (k) + Ts x2 (k)
f1 (x(k)) =
,
ρ0 −x1
x2 (k) − Ts M
e
x1 (k) − Ts hMd x2 (k)
T T
F2 = 0 Ms
,
(13)
where x2 is the velocity of the cart and u is an external force
which acts as an input to the system. The parameter values are
Ts = 0.4s, ρ0 = 0.33, M = 1, hd = 1.1.
If we assume nj ≤ n for all j ∈ Z[0,N −1] , then the complexity
(8) can be upper bounded by
N (N + 1)
C = nc1
+ c2 N n.
(9)
2
A possible reduction of the bound (9) might be attained, considering the fact that all the states subsequent to a non–feasible
state are no longer evaluated and checked for feasibility. This
means that, in step 9 of Algorithm 1, if φ(xk|k , U (k)jq , i) ∈
/X
for a specific i ∈ Z[j+1,N −1] , then φ(xk|k , U (k)jq , r) for all
r ∈ Z[i+1,N ] , are no longer evaluated, in which case steps 2)
and 3) are skipped all together.
The MPC controller has to steer the cart to the origin from
a non–zero initial state, while satisfying the input and state
constraints, which are
|u| ≤ 4.5, |x1 | ≤ 2.65,
(14)
and reducing the cost (2), where the stage cost and terminal
cost are quadratic functions, i.e., L(x, u) = xT Qx + uT Ru
and Vf (x) = xT P x. Choose the following parameters for the
MPC problem:
7.0814 3.3708
M = 4, P =
, Q = diag(1, 1), R = 1.
3.3708 4.2998
Consider now that many threads are available, from multiple
processors. Notice also that at each time in the horizon, all nj
computations can be performed separately. In these conditions,
assuming we have n threads available, then the complexity of
Algorithm 1 is upper bounded by
In (Raimondo et al., 2009) it is shown that the control law
u = kf (x) = − [ 0.8783 1.1204 ] f1 (x),
is locally stabilizing in the set
XT = {x|Vf (x) ≤ 4.7},
State trajectory
evaluation time[s]
0.16
6
terminal set
fmincon
nj=0
5
0.14
0.12
nj=5
nj=10
time [s]
4
nj=30
0.1
0.08
3
x
2
0.06
2
0.04
1
0.02
1
2
3
4
5
6
7
8
9
10
11
12
k
0
Fig. 4. Computation time.
−1
−2
−2
−1
0
x1
1
2
Fig. 2. State trajectory.
inputs
4
2
u
0
−2
−4
−6
2
4
6
8
10
12
14
16
18
20
k
Fig. 3. Inputs applied to the system.
which is a terminal set where the conditions for stability of
MPC mentioned in Section 2.2 are satisfied.
For the first experiment, fix N = 10. An initial Uwarm (0)
is provided by a random “oracle”. Consider nj = n, for all
j ∈ Z[0,N −1] , taking various values in the set {0, 5, 10, 30}.
For n = 0, Uwarm (0) is propagated though iterations without
any intervention or change from the sampling mechanism. This
serves as a reference, to notice the improvements brought in by
Algorithm 1. The results are illustrated in Fig. 2–5, where the
legend from Fig. 2 holds until Fig. 5. Iterations are considered
from k = 1 until k = 20. Though the different sampling
options proposed in Remark 3.1 provide in this example similar
outcomes, Halton points have been used here for illustration.
This choice is motivated by the practical feature that, adding
extra points only when they are required does not have impact
on the coverage of the set U.
In Fig. 2 and Fig. 3 the state trajectory and the inputs u(k)
applied to the system are illustrated. The constraint specifications (14) are satisfied for all cases, at all times. Notice the
Fig. 5. The cost Jsub .
Computational performance.
fmincon
sampling
bound (9)
bound (10)
20
15
Computational details for N<=20.
0.3
time [s]
25
time [s]
Algorithm 1 is applied for the MPC control of system (12). We
compare the results of this method with the results provided by
fmincon in Matlab, even though, as it will be seen later, the
optimization tool does not always provide the optimum, due
to local minima. The scalability of the algorithm is tested by
varying both the number of samples n and the control horizon
N . The tests illustrated in this paper are performed on a feasible
initial state x0|0 = [−2.5, 3]. The choice of the initial condition
does not influence greatly the results, which are similar for
other feasible initial states.
10
0.2
0.1
5
0
0
20
40
60
N
80
100
5
10
15
20
N
Fig. 6. Computational performance versus proposed bounds.
immediate smoothening of the trajectories and input sequence
even for a small n. In Fig. 4, the computational time, without
parallelization, is illustrated. At k = 1, the computational cost
of finding an oracle is included. Fig. 5 illustrates the values of
Jsub for each iteration. Notice, overall, that even for a small
number of samples, the performance of the closed loop system is significantly improved and the computational time is
promising, even for a non–parallel implementation. Also, as the
iteration k advances, the initially modest performance increases
significantly, due to the continuous improvement of Uwarm and
the receding horizon principle. Interestingly, even for n = 5,
at iteration k = 4, the cost Jsub (4) is smaller than the cost
computed via fmincon, which, due to local minima, provided
a feasible but not optimal solution.
Cost function
250
fmincon
nj=0
Jsub
200
nj=5
150
nj=20
nj=50
100
50
0
0
5
10
15
20
25
k
30
35
40
45
50
35
40
45
50
evaluation time[s]
1.4
1.2
1
t[s]
For the second experiment we aim to test the computational
complexity of Algorithm 1 in terms of the control horizon.
The implementation used here does not use parallelization. Fix
n = 10 and N takes values in the set {3, 10, 20, 50, 100}. The
results are illustrated in Fig. 6. With red, the computational
complexity for Algorithm 1 is depicted, and it is always smaller
than the bound (9), which is followed closely. It is expected that
on dedicated devices, the complexity of Algorithm 1 is smaller,
due to the processors not running in parallel threads related
to other system applications. With parallelization, further reduction in complexity is expected, as described in Section 3.2
and illustrated in Fig. 6. Notice that, for smaller horizons,
the complexity of Algorithm 1 is smaller than the complexity
of fmincon, even without parallelization. For N = 100,
fmincon provides solutions which are not feasible, while the
proposed Algorithm 1 still terminates in 21.6 seconds.
0.8
0.6
0.4
4.2 Buck–Boost power converter
0.2
Next, the bilinear model of a Buck–Boost power converter is
considered, as in (Spinu et al., 2011):
x(k)T C1
x(k + 1) = Ax(k) + Bu(k) +
u(k),
(15)
x(k)T C2
T
where x := [ vC iL ] ∈ X ⊂ R2 is the state vector consisting
of the voltage across the output capacitor and the current
T
through the filter inductor. The input u := [ d1 d2 ] ∈ U ⊂ R2
stands for the duty–cycle ratio of the control signal applied to
the switching node. The parameters are
− RH1 C 0
0 0
A = I2 + Ts
,
B
=
Ts ,
v
s
0
− RLL
L 0
0 0
0 − L1
C1 =
T
,
C
=
Ts ,
s
2
0 C1
0 0
with the values RL = 0.2Ω, C = 22µF , L = 220µH,
Ts = 10µs.
The aim of the control loop is to stabilize the system to the
T
T
equilibrium point xe := [ 20 0.5 ] , ue := [ 0.81 0.4 ] , under
2
the constraints iL ∈ R[0,3] , vC ∈ R[−0.1,22.5] , u ∈ R[0,1] . The
terminal controller
−0.0014 −0.3246
u = ue + K(x − xe ), K =
,
0.0001 −0.0055
stabilizing the system in the terminal set XT given in (Spinu
et al., 2011, set P in Section IV.C), and the quadratic cost with
the matrices
46.6617 42.8039
Q = diag(1, 2), R = diag(1, 1), P =
,
42.8039 69.4392
satisfies all the conditions for stability formulated in Section 2.2.
Similarly to the previous example, we apply Algorithm 1 for
the MPC control of system (15). Fix the initial state x0|0 =
T
[ 1 2 ] + xe , which is outside of the terminal set XT , and
N = 10. Uwarm (0) is given by an oracle and we use random
sampling of U. See in Fig. 7 the effect of varying n on the
cost function Jsub and the time necessary, per iteration, to
compute the corresponding control law. Notice the effect of
the unknown termination time on the evaluation time for the
optimization performed through fmincon and the relatively
equal computational time of Algorithm 1 over the iterations k.
0
0
5
10
15
20
25
k
30
Fig. 7. Performance of Algorithm 1 on the power converter.
All the constraints were satisfied for all the presented situations,
and it is expected that, through implementation on dedicated
multi–thread systems, the computational time for the control
law decreases to the extent of fitting the tight sampling period
of the power converter.
4.3 Wheeled mobile robot
The last example illustrates the methodology developed in this
paper for an obstacle avoidance task by a nonholonomic system
with trigonometric nonlinearities, due to kinematics, i.e., a
model of the wheeled mobile robot (WMR), as described in
(Kuhne et al., 2005):
"
#
x1 (k) + u1 (k) cos x3 (k)Ts
x(k + 1) = x2 (k) + u1 (k) sin x3 (k)Ts .
(16)
x3 (k) + u2 (k)Ts
In (16), the state x ∈ R3 describes the position and the orientation of the robot with respect to a global inertial frame
{O, X, Y }, and the input u ∈ R2 gives the linear and angular velocity, respectively. The parameter Ts = 0.1s is the
discretization period of system (16).
The MPC strategy aims at driving the WMR from an initial
T
state x0|0 = [ 0 6 0 ] to the origin of the inertial frame, i.e.,
T
xg = [ 0 0 0 ] , while satisfying the input constraints u1 ∈
R[−0.47,0.47] , u2 ∈ R[−3.77,3.77] . A quadratic cost function of
the form
N
−1
X
T
J(xk|k , U (k)) =xk+N |k P xk+N |k +
xTk+j|k Q(j)xk+j|k +
j=1
+
N
−1
X
uTk+j|k Ruk+j|k
j=0
is considered, with the parameters: Q(j) = 2j−1 Q, P =
50Q(N ), Q = diag(1, 1, 0.5), R = diag(0.1, 0.1), N = 5.
In this example we illustrate Algorithm 1 with the above parameters. Consider Uwarm (0) generated by an “oracle”, and
sampling of U based on random sequences. The result for
n = 30 is illustrated in Fig 8. fmincon could not be applied
State trajectory around obstacle.
6
Algorithm 1
SBMPC
obstacle
5
x2
4
improves in terms of cost an initially feasible control sequence.
This suboptimal NMPC strategy provides a promising computational complexity even for large control horizons, with good
perspectives for parallel implementation. Future work aims at
investigating the convergence of the proposed algorithm to the
optimal solution, the scalability in terms of system dimension
and the effect of parallelization on computational time for real–
life applications.
3
REFERENCES
2
1
0
-1.6
-1.4
-1.2
-1
-0.8
-0.6
-0.4
-0.2
0
0.2
0.4
x1
Fig. 8. Trajectories of WMR.
total verification time[s]
0.8
0.7
time [s]
0.6
0.5
0.4
0.3
0.2
0.1
0
50
100
150
200
250
300
350
400
k
Fig. 9. Time performance of WMR.
due to feasibility issues related to non–convexity caused by the
presence of the obstacle. Notice the avoidance of the obstacle
of the WMR under the control law generated by Algorithm 1.
Due to the sampling mechanism, however, the inputs are not
smooth, which could be alleviated, e.g., by an interpolation
mechanism. For comparison we illustrate also the result using
SBMPC (Dunlap et al., 2010), with the same horizon N and
n = 30. Notice, in the state trajectory plot in Fig 8 the fact
that, after 400 iterations, the WMR did not reach xg yet, and
the input values are not yet 0, which means that the sampling
mechanism did not consider that the WMR is close to the goal.
In this case, it is recommended to sample more densely around
0 in the set U, rather than uniformly covering U with samples.
Notice in Fig 9 that, due to the building of a tree and the search
in a tree for the best path, SBMPC is more computationally
demanding. The complexity of Algorithm 1 with the given n
fits the sampling period Ts = 0.1s of the WMR, which makes
the method applicable for real–time control of this system.
5. CONCLUSION
In this paper, an algorithm based on sampling of the input space
at each time in the horizon was proposed, which iteratively
Bertsekas, D.P. (2005a). Dynamic programming and optimal
control, volume 1. Athena Scientific Belmont, MA.
Bertsekas, D.P. (2005b). Dynamic programming and suboptimal control: A survey from ADP to MPC. European Journal
of Control, 11(4), 310–334.
Chakrabarty, A., Dinh, V., Corless, M., Rundell, A., Zak,
S., and Buzzard, G. (2016).
Support vector machine informed explicit nonlinear model predictive control using low-discrepancy sequences. IEEE Transactions
on Automatic Control, preprint published on–line, DOI:
10.1109/TAC.2016.2539222, PP(99).
de Dios Ortúzar, J., Willumsen, L.G., et al. (1994). Modelling
transport. Wiley New Jersey.
Dunlap, D.D., Caldwell, C.V., and Collins, E.G. (2010). Nonlinear model predictive control using sampling and goaldirected optimization. In 2010 IEEE International Conference on Control Applications, 1349–1356. IEEE.
Grüne, L. and Pannek, J. (2011). Nonlinear model predictive
control. Theory and algorithms. London: Springer–Verlag.
Johansen, T.A. (2004). Approximate explicit receding horizon
control of constrained nonlinear systems. Automatica, 40(2),
293–300.
Kuhne, F., Lages, W.F., and Da Silva, J.G. (2005). Point stabilization of mobile robots with nonlinear model predictive
control. In IEEE International Conference Mechatronics and
Automation, 2005, volume 3, 1163–1168. IEEE.
LaValle, S.M. (1998). Rapidly-exploring random trees: A new
tool for path planning. TR 98-11, Computer Science Dept.,
Iowa State University.
Ławryńczuk, M. (2009). Computationally efficient nonlinear
predictive control based on state-space neural models. In International Conference on Parallel Processing and Applied
Mathematics, 350–359. Springer.
Lazar, M. and Heemels, W. (2009). Predictive control of hybrid systems: Input-to-state stability results for sub-optimal
solutions. Automatica, 45(1), 180–185.
Lee, J.M. and Lee, J.H. (2004). Approximate dynamic programming strategies and their applicability for process control: A review and future directions. International Journal of
Control Automation and Systems, 2, 263–278.
Likhachev, M. and Stentz, A. (2008). R* search. Lab Papers
(GRASP), 23.
Magni, L., Raimondo, D.M., and Allgöwer, F. (2009). Nonlinear model predictive control: Towards new challenging
applications, volume 384. Springer, Lecture notes in Control
and Information Sciences.
Mayne, D. and Rawlings, J. (2009). Model predictive control:
theory and design. Madison, WI: Nob Hill Publishing, LCC.
Muñoz de la Peña, D., Alamo, T., Bemporad, A., and Camacho,
E.F. (2004). A dynamic programming approach for determining the explicit solution of linear MPC controllers. In
IEEE Conference on Decision and Control.
Pannocchia, G., Rawlings, J.B., and Wright, S.J. (2011). Conditions under which suboptimal nonlinear MPC is inherently
robust. Systems & Control Letters, 60(9), 747–755.
Piovesan, J.L. and Tanner, H.G. (2009). Randomized model
predictive control for robot navigation. In IEEE International Conference on Robotics and Automation, 2009, 94–99.
Kobe, Japan.
Raimondo, D.M., Limon, D., Lazar, M., Magni, L., and Camacho, E.F. (2009). Min-max model predictive control of nonlinear systems: A unifying overview on stability. European
Journal of Control, 15(1), 5–21.
Scokaert, P.O., Mayne, D.Q., and Rawlings, J.B. (1999). Suboptimal model predictive control (feasibility implies stability). IEEE Transactions on Automatic Control, 44(3), 648–
654.
Spinu, V., Lazar, M., and van den Bosch, P. (2011). An explicit
state-feedback solution to constrained stabilization of dc-dc
power converters. In 2011 IEEE International Conference
on Control Applications (CCA), 1112–1118. IEEE.
| 3cs.SY
|
NONCOMMUTATIVE RESOLUTIONS OF DISCRIMINANTS
RAGNAR-OLAF BUCHWEITZ, ELEONORE FABER, AND COLIN INGALLS
arXiv:1702.00791v1 [math.AG] 2 Feb 2017
A BSTRACT. We give an introduction to the McKay correspondence and its connection to
quotients of Cn by finite reflection groups. This yields a natural construction of noncommutative resolutions of the discriminants of these reflection groups. This paper is an extended version of E. F.’s talk with the same title delivered at the ICRA.
1. I NTRODUCTION
This article has two objectives: first, we want to present the components of the classical
McKay correspondence, which relate algebra, geometry and representation theory. On
the other hand, we give a short introduction to discriminants of finite reflection groups
and our construction of their noncommutative desingularizations. The details of our new
results will be published elsewhere [BFI16].
This project grew out of the following (which is NOT how we will present our story!):
Start with a commutative ring R. Then a noncommutative resolution of singularities
(=NCR) of R (or of Spec( R)) is an endomorphism ring EndR M of a faithful module M
such that the global dimension of EndR M is finite. Endomorphism rings of finite global
dimension have got a lot of attention lately because of their connections to various parts
of mathematics, such as commutative algebra, noncommutative algebraic geometry and
representation theory. The problem of constructing explicit NCRs is difficult in general
and one only knows some scattered examples. In particular, NCRs for non-normal rings
have not been considered much in the literature, mostly only in examples where R has
Krull-dimension ≤ 2 (geometrically: R is the coordinate ring of a collection of points,
curves or surfaces).
In [DFI15] we were interested in noncommutative resolutions (=NCRs) of certain nonnormal hypersurfaces, so-called free divisors, and explicitly constructed a NCR of a normal crossings divisor. After some extensive calculations we realized that there was a more
conceptual way to understand this particular NCR as a quotient of the skew group ring
of the polynomial ring in n variables and the reflection group (µ2 )n by a two-sided ideal
generated by an idempotent of the skew group ring. And since the normal crossings divisor is the discriminant of a real reflection group, we had the idea to carry out the same
quotient procedure for other discriminants of reflection groups. It was readily verified
that indeed this construction works for any discriminant of a finite reflection group.
Date: February 6, 2017.
2010 Mathematics Subject Classification. 14E16, 13C14, 14E15, 14A22 .
Key words and phrases. Reflection groups, hyperplane arrangements, maximal Cohen–Macaulay modules,
matrix factorizations, noncommutative desingularization.
R.-O.B. was partially supported by an NSERC Discovery grant, E.F. was partially supported by an Oberwolfach Leibniz Fellowship, C.I. was partially supported by an NSERC Discovery grant.
1
2
RAGNAR-OLAF BUCHWEITZ, ELEONORE FABER, AND COLIN INGALLS
After some more painful calculations and serious doubts, that these quotients are endomorphism rings in general, we were finally able to also prove this for reflection groups
generated by reflections of order 2, and discovered the surprising fact that the corresponding reflection arrangement is in some sense the noncommutative resolution of the
discriminant of a reflection group. Moreover, our work generalizes Auslander’s theorem
about skew group rings of small subgroups G ⊆ GL(n, C) and thus our NCRs give rise
to a McKay correspondence for reflection groups. For reflection groups G ⊆ GL(2, C)
our results nicely fit in the picture of the classical McKay correspondence: we see that the
reflection arrangement provides a representation generator for the torsion-free modules
over the discriminant curves.
This paper is organized as follows:
C ONTENTS
1.
Introduction
1
2.
Reflection groups
2
2.1.
Basics
3
2.2.
Discriminant and hyperplane arrangement
3
3.
(Noncommutative) resolutions of singularities
5
4.
The classical McKay correspondence
6
4.1.
Dual resolution graphs
6
4.2.
McKay quiver
8
4.3.
AR-quiver
9
4.4.
Auslander’s theorem and reflections
10
5.
NCRs of discriminants
11
6.
Further questions
12
7.
Acknowledgements
13
References
13
2. R EFLECTION GROUPS
Throughout the paper, we assume that k = C and that G ⊆ GL(V ) is a finite subgroup
acting on V ∼
= Cn . Thus, G ⊆ GL(V ) means G ⊆ GL(n, C). Then Symk (V ) ∼
= k [ x1 , . . . , x n ]
is a (graded) polynomial ring, with the standard grading | xi | = 1. Assume that G acts
on V, then any g ∈ G acts also on S via g( f ( x1 , . . . , xn )) = f ( g( x1 , . . . , xn )). Note that in
most of the literature (about the McKay correspondence), one considers the completion
k [[ x1 , . . . , xn ]] of S, but in fact, all the results we need work for S the graded polynomial
ring.
Here we gather some well-known results about complex reflection groups, see e.g. [Bou81,
LT09, OT92], we follow the notation of [OT92].
NONCOMMUTATIVE RESOLUTIONS
3
2.1. Basics. An element g ∈ GL(V ) is called a pseudo-reflection if it is conjugate to a diagonal matrix diag(ζ, 1, . . . , 1) with ζ 6= 1 a root of unity. If ζ = −1, we call g a (true)
reflection. A pseudo-reflection g fixes a unique hyperplane H pointwise. This hyperplane
will be called the mirror of g. If G is generated by pseudo-reflections, it is called a complex
reflection group (or pseudo-reflection group). If G is generated by true reflections, then we
will call G a true reflection group. True reflection groups are generated by reflections of
order 2 and thus are precisely the finite Coxeter groups. A finite subgroup G ⊆ GL(V ) is
called small, if G does not contain any pseudo-reflections.
Let now G ⊆ GL(V ) be finite. The invariant ring of the action of G on V is R = SG = { f ∈
S : g( f ) = f for all g ∈ G }. By the theorem of Hochster–Roberts [HR74], R is Cohen–
Macaulay and it is Gorenstein if G ⊆ SL(V ), by [Wat74]. If G is a pseudo-reflection group,
we are even in a better situation, cf. [Che55], [OT92, Thm. 6.19]:
Theorem 2.1 (Chevalley–Shephard–Todd). (a) Let G ⊆ GL(V ) be a finite group acting on
V. Then G is a pseudo-reflection group if and only if there exist homogeneous polynomials
f 1 , . . . , f n ∈ R such that R = k [ f 1 , . . . , f n ].
(b) Moreover, S is free as an R-module, more precisely, S ∼
= R ⊗k kG as R-modules, where kG
denotes the group ring of G.
The polynomials f i in the theorem are called basic invariants of G. They are not unique,
but their degrees deg f i = di are uniquely determined by the group G.
2.2. Discriminant and hyperplane arrangement. Let G ⊆ GL(V ) be any finite subgroup.
Then we denote by V/G the quotient by this action. It is an affine variety given by V/G =
Spec(SG ), see e.g. [CLO07, Chapter 7]. In general V/G is singular, in particular, if G is
a small subgroup of GL(V ). If G is a pseudo-reflection group, then by Thm. 2.1, V/G is
smooth and isomorphic to V ∼
= kn . We have a natural projection:
π:V∼
= Spec(S) −→ V/G ∼
= Spec(SG ) .
On the ring level, this corresponds to the natural inclusion of rings
SG ∼
= k[ f 1 , . . . , f n ] ,→ S = k[ x1 , . . . , xn ] .
Now let us define two of the central characters in our story: Let G be a finite complex
reflection group. The set of mirrors of G is called the reflection arrangement of G, denoted
by A( G ). Note that A( G ) ⊆ V. Let H ⊆ V be one of these mirrors, that is, H is fixed
by a cyclic subgroup generated by a pseudo-reflection g H ∈ G of order ρ H > 1, and let
{l H = 0} be the defining linear equation of H. Then the reflection arrangement is given
by the (reduced) equation z := ∏ H ⊆A(G) l H = 0. The image of the reflection arrangement
A( G ) under the projection π : V → V/G is called the discriminant of G and given by a
reduced polynomial ∆. Note that ∆ ∈ SG , whereas z 6∈ SG .
One can express the discriminant in terms of the basic invariants of G: first, one can show
that
∂ fi
ρ −1
= u ∏ l HH ,
J = Jac( f 1 , . . . , f n ) = det
∂x j i,j=1,...,n
H ⊆A( G )
where u ∈ k∗ . The Jacobian J is an anti-invariant for G, that is, gJ = (det g)−1 J for any
g ∈ G. The discriminant ∆ is then given by
∆ = Jz =
∏
H ⊆A( G )
ρ
l HH ,
4
RAGNAR-OLAF BUCHWEITZ, ELEONORE FABER, AND COLIN INGALLS
which is an element of SG .
In the case when G is a true reflection group, we have ρ H = 2 for all H and thus z = J,
and the computation of ∆ simplifies to ∆ = z2 .
Let us conclude this section with some (non-)examples:
ζ3 0
,
Example 2.2. Let G ⊆ SL(V ), dim V = 2, be the cyclic group C3 generated by
0 ζ 3−1
where ζ 3 is a third root of unity acting on S = k [ x, y]. Then V/G is a Kleinian surface singularity, denoted by A2 . The invariants of G are u = x3 + y3 , v = xy, w = x3 − y3 and satisfy the relation w2 − u2 + 4v3 . Thus (after a change of coordinates) SG ∼
= k[u, v, w]/(w2 +
u2 + v3 ). A real picture of V/G can be seen in Fig. 1.
Example 2.3. Let G = (µ2 )n be acting on kn by reflecting about the coordinate hyperplanes. Thus the generators of G are the diagonal matrices
1 0 0 0 0
..
0 . 0 0 0
Si =
0 0 −1 0 0 .
0 0 0 . . . 0
0 0 0 0 1
The hyperplane arrangement A( G ) is the union of the coordinate hyperplanes. This can
easily be seen computationally: The basic invariants in this example are f i = xi2 , that is,
SG = k [ x12 , . . . , xn2 ]. Thus Jac( f 1 , . . . , f n ) = 2n x1 · · · xn and we can take z = J = x1 · · · xn .
Since G is a true reflection group, ∆ = z2 = x12 · · · xn2 = f 1 · · · f n . The coordinate ring of
the discriminant is SG /(∆) = k [ f 1 , . . . , f n ]/( f 1 · · · f n ). In this example, the discriminant is
also the union of coordinate hyperplanes, but now seen in the quotient V/G. See Fig. 1
for a picture in in the 3-dimensional case.
F IGURE 1. The A2 -singularity (left) and the discriminant of µ32 . (right).
Example 2.4. Let G = G (2, 1, 3), that is, the Weyl group B3 = S4 × µ2 acting on k3 by
permutation and sign changes. The invariants of S = k [ x1 , x2 , x3 ] under this group action
are u = x12 + x22 + x32 , v = x14 + x24 + x34 and w = x16 + x26 + x36 . Since G (2, 1, 3) is a Coxeter
group, it is generated by true reflections, and thus z = J. A calculation of the Jacobian
determinant of (u, v, w) shows that
z = J = x1 x2 x3 ( x12 − x22 )( x12 − x32 )( x22 − x32 ).
NONCOMMUTATIVE RESOLUTIONS
5
The invariant ring is SG = k [u, v, w] and an investigation of z2 shows that it is equal to
the product of w with the discriminant of the polynomial P(t) = t3 + ut2 + vt + w, see
[YS81]. The equation for the discriminant is
∆ = w(u2 v2 − 4v3 − 4u3 w + 18uvw − 27w2 ).
See Fig. 2 below for pictures of the hyperplane arrangement, given by z and the discriminant ∆ in R3 .
F IGURE 2. The B3 -reflection arrangement (left) and the corresponding discriminant with equation w(u2 v2 − 4v3 − 4u3 w + 18uvw − 27w2 ) (right).
3. (N ONCOMMUTATIVE ) RESOLUTIONS OF SINGULARITIES
Here we come to the geometric part. Let X be an affine algebraic variety, that is X =
Spec( R), where R is a commutative finitely generated k-algebra. If R is not a regular ring,
then we say that X is a singular algebraic variety and denote by Sing X its singular locus. A
e → X from a smooth scheme
resolution of singularities of X is a proper birational map π : X
e
X to X such that π is an isomorphism on the smooth points of X.
The central result regarding resolution of singularities is
Theorem 3.1 (Hironaka [Hir64]). Let X be a scheme over a field of characteristic 0 (e.g., X an
e → X.
affine algebraic variety defined over C). Then there exists a resolution of singularities π : X
One can impose some extra conditions on π, for example that π −1 (Sing( X )) is a normal
e and π is a composition of blowups in smooth centers, or (in the
crossings divisor in X
case X an irreducible reduced variety over k), that X has a so-called embedded resolution, see [Hau03] for more on this. So existence of resolutions of singularities is (with
our assumption that k = C, that is, k of characteristic 0) not an issue. However, from
e If X is of
an algebraic point of view, it would be desirable to have some control over X.
Krull-dimension 2, then there exists a unique minimal resolution of singularities, that is, a
e of X factors through Y. This is false in
resolution Y of X such that any other resolution X
higher dimensions, in general there does not even exist a minimal resolution.
For some Cohen–Macaulay varieties X (that is: if X = Spec( R), then R is a Cohen–
Macaulay ring) there still exist preferred resolutions of singularities: namely those, that
e → X is called
do not affect the canonical class of X. A resolution of singularities π : X
∗
crepant if π ωX = ωXe . Crepant resolutions first appeared in the minimal model program,
where the term “crepant” was coined by M. Reid to reflect the fact that the pull-back of
6
RAGNAR-OLAF BUCHWEITZ, ELEONORE FABER, AND COLIN INGALLS
the canonical sheaf of X does not pick up discrepancy, see e.g. [Rei02]. Crepant resolutions still do not always exist and are usually not isomorphic to each other, but at least
their cohomologies should be the same, see [BO02]:
Conjecture 3.2 (Bondal–Orlov). Let Y and Y + be two crepant resolutions of an algebraic variety
X. Then there is a derived equivalence of categories
D b (Coh Y ) ' D b (Coh Y + ) ,
where Coh(−) denotes the category of coherent sheaves.
Building on work by Bridgeland–King–Reid for 3-dimensional quotient singularities [BKR01],
Bridgeland proved [Bri02] the conjecture for complex 3-dimensional varieties with only
terminal singularities. Inspired by their work, Van den Bergh defined a noncommutative
crepant resolution (=NCCR) [VdB04a], also see [VdB04b]: if X = Spec( R), then the main
point in their approach is to construct a (non-commutative) R-algebra such that for any
crepant resolution π : Y → X one has D b (Coh Y ) ' D b (mod A). Here mod A denotes
the category of finitely generated A-modules. The algebra A is thus a noncommutative
analogue of a crepant resolution of singularities, a NCCR. Later, NCCRs were studied
and constructed in various instances, see e.g., [BLvdB10, IW14]. The more general concept of noncommutative resolution (=NCR) was defined by Dao–Iyama–Takahashi–Vial
[DITV15] in 2015. Unfortunately there is (so far) not a good theory about general existence and properties of NCRs, only mostly examples, see e.g. [BIKR08, DFI15, DII+ 16,
DFI16, Leu07]. A good introduction to NCCRs and categorical geometry with many references can be found in [Leu12].
But let us state the definitions:
Definition 3.3. Let R be a commutative reduced noetherian ring. An R-algebra A is called
a noncommutative resolution (NCR) of R (or of Spec( R)) if A is of the form EndR M for M
a finitely generated, faithful R-module and gldim( A) < ∞. The algebra A is called a
noncommutative crepant resolution (NCCR) of R (or of Spec( R)) if A is a nonsingular order,
that is, gldim Ap = gldim Rp for all p ∈ Spec( R) and A is a maximal Cohen–Macaulay
module over R.
Remark 3.4. In Van den Bergh’s original definition, R was assumed to be a commutative
normal Gorenstein domain. The reasoning behind this more general definition can be
found in [DFI15].
In general it is not clear how to construct NCRs and whether they exist. However, there
is one beautiful example, which will lead to our construction of NCRs for discriminants
of reflection groups.
4. T HE CLASSICAL M C K AY CORRESPONDENCE
Here we give brief sketches of the components of the classical McKay correspondence.
Up to this date, the correspondence has been generalized in various directions, for an
account and more detailed references see [Buc12].
4.1. Dual resolution graphs. Let Γ ⊆ SL(V ) be a finite subgroup acting on V ∼
= k2 .
Γ
Denote as above by S = k [ x, y] the symmetric algebra of V and the invariant ring S = R.
Then X := V/Γ = Spec( R) is a quotient singularity. More precisely, these quotient
singularities (aka Kleinian singularities) have been classified by Felix Klein [Kle93]:
NONCOMMUTATIVE RESOLUTIONS
7
Theorem 4.1 (Klein 1884). Let Γ and X be as above. Then X is isomorphic to an isolated surface
singularity of type ADE, that is, R ∼
= k[ x, y, z]/( f ):
•
•
•
•
•
An :
Dn :
E6 :
E7 :
E8 :
f = z2 + y2 + x n+1 , if Γ is a cyclic group,
f = z2 + x (y2 + x n−2 ) for n ≥ 4, if Γ is a binary dihedral group,
f = z2 + x3 + y4 , if Γ is the binary tetrahedral group,
f = z2 + x ( x2 + y3 ), if Γ is the binary octahedral group,
f = z2 + x3 + y5 , if Γ is the binary icosahedral group.
F IGURE 3. The D4 -singularity with equation z2 + x (y2 − x2 ) = 0 (left) and
the E8 -singularity with equation z2 + x3 + y5 = 0 (right).
The Kleinian singularities X are also classified by their so-called dual resolution graphs,
which happen to be the corresponding ADE-Coxeter-Dynkin-diagrams (for a more detailed account of resolution of surface singularities and especially the rational double
e → X of a Kleinian
points, see [Gre92], [Dur79]): consider the minimal resolution π : X
singularity X. The minimal resolution exists, since X is a normal surface singularity and
e is moreover unique, see [Lau71]. One can for example obtain X
e by successively blowX
ing up singular points. This process can be quite intricate, see for example the dessert in
e will be denoted by E = Sin=1 Ei
this menu [FH10]. The preimage of {0} = Sing( X ) in X
e (this holds for example because one can
and is a union of irreducible curves Ei on X
e
construct X as an embedded resolution of X ⊆ k3 , which implies that the preimage of
Sing X must be a normal crossing divisor). Since X is a so-called rational singularity, the
irreducible components Ei of E are isomorphic to P1 ’s and intersect each other at most
e Now one can
pairwise transversally (this is to say: E is a normal crossings divisor on X).
form the dual resolution graph of X: the vertices are indexed by the Ei , for i = 1, . . . , n and
there is an edge between Ei and E j if and only if Ei ∩ E j 6= ∅.
Moreover, one defines the intersection matrix E := ((Ei · E j ))i,j=1,...,n . Computation
shows that (Ei · Ei ) = −2 and (Ei · E j ) ∈ {0, 1} for i 6= j and that E is symmetric and
e of the form Z 0 = ∑in=1 mi Ei
negative-definite. Moreover, there exist divisors Z 0 on X
0
supported on E with mi > 0, such that Z · Ei ≤ 0 for all i. The smallest such divisor
is denoted by Z and called the fundamental divisor of X. One decorates the vertices of
the dual resolution graph with the multiplicities mi of Z. For an illustration of the D4 singularity see Fig. 4.1, and the resolution with the resolution graph are below in Fig. 4.
8
RAGNAR-OLAF BUCHWEITZ, ELEONORE FABER, AND COLIN INGALLS
1 E3
E4
1
E1
2
E2
1
F IGURE 4. The minimal resolution of the D4 -singularity (with the exceptional curves sketched in red) and the dual resolution graph of type D4
with the multiplicities of the fundamental divisor in red.
4.2. McKay quiver. In 1979, John McKay built the following graph out of a finite subgroup Γ ⊆ SL(V ), dim V = 2: the embedding of Γ in SL(V ) defines the so-called canonical representation c : Γ ,→ GL(V ) of Γ. Any finite group only has finitely many isomorphism classes of finite dimensional irreducible k-representations, given by group homomorphisms ρi : Γ → GL(Vi ), i = 0, . . . , n for vector spaces Vi . The trivial representation
will be denoted by ρ0 : Γ → k∗ , sending any γ to 1. The McKay quiver of Γ consists of
the vertices indexed by the Vi and there are mij arrows from Vi to Vj if and only if Vi is
contained with multiplicity mij in Vj ⊗ V.
Example 4.2. Consider the group Γ generated by
10
i 0
0 1
0 i
±
,±
,±
,±
.
01
0 −i
−1 0
i 0
This is the binary dihedral group D4 . It has five irreducible representations ρi , four of
which are one-dimensional and one two-dimensional ρ1 , which is the canonical representation c. Using character theory, one obtains the McKay quiver of Γ (the dimensions of
the irreducible representations in red):
ρ4 1
2
1 ρ0
ρ1
ρ3 1
ρ2 1
The following is due to McKay [McK80]:
Observation: Let Γ be as above. Then the McKay quiver of Γ is connected and contains no
loops and mij = m ji ∈ {0, 1}. By calculation, one sees that the McKay quiver of Γ is the
extended Coxeter–Dynkin diagram for Γ, with arrows in both directions. If one deletes
the vertex corresponding to the trivial representation and collapses all arrows Vi Vj to
an edge, then one obtains the Coxeter–Dynkin diagram associated to Γ. If one decorates
NONCOMMUTATIVE RESOLUTIONS
9
the vertices Vi with the dimensions dim Vi , then one gets back the dual resolution graph
of V/Γ.
Thus one obtains a 1 − 1-correspondence between the irreducible components of the exceptional divisor of the minimal resolution of the quotient singularity V/Γ and the nontrivial irreducible representations of Γ! This was made precise later by [GSV83].
4.3. AR-quiver. But there is also an algebraic part of the correspondence: Let again R =
SΓ for Γ ⊆ SL(V ), where V ∼
= k2 . We consider reflexive modules over R. Since R is
normal, M is reflexive if and only if it is maximal Cohen–Macaulay (=CM), and we write
CM( R) for the category of CM-modules over R. By Herzog’s theorem [Her78], there is
an isomorphism of R-modules:
S∼
=
M
M a M , where the integer a M = rankR M.
M ∈CM( R)
Differently phrased: S is a representation generator for CM( R). In particular, there are only
finitely many indecomposable objects in CM( R). One then says that R is of finite CMtype. By [BGS87, Her78, Knö87], a Gorenstein ring R is of finite CM-type if and only if it
is isomorphic to a simple hypersurface singularity (if dim R = 2, these are precisely the
ADE-surface singularities).
Starting from the indecomposable CM-modules over R, one obtains a third quiver, the
Auslander–Reiten-quiver (short: AR-quiver) of R: its vertices are given by the indecomposable CM-modules Mi , i = 1, . . . , n of R and there are mij arrows from Mi to M j if and only
if in the almost split sequence 0 → τM j → E → M j → 0 ending in M j , Mi appears with
multiplicity mij in E. Here one can also see that τMi = Mi for all Mi 6∼
= R and that the
AR-quiver of R is precisely the McKay quiver of Γ. Moreover, the ranks of the indecomposable modules Mi correspond to the dimensions of the irreducible representations of Γ.
This astounding correspondence was proven by Maurice Auslander [Aus86] and looking
at his proof more carefully, we will obtain the link to NCCRs. Therefore we need some
general notation first:
Let G ⊆ GL(V ) finite act on S = Symk (V ) = k [ x1 , . . . , xn ] for some vector space V ∼
= kn .
Then define A = S ∗ G to be the skew group ring (or twisted group ring) of G. As an Smodule, A is just S ⊗k kG, but the multiplication on A is twisted by the action of G: for
elements s, s0 ∈ S and g, g0 ∈ G define
sg · s0 g0 := (sg(s0 ))( gg0 ) ,
and extend by linearity. The following result is the key theorem in proving the correspondence between indecomposable CM-modules and irreducible representations. However
note that the assumptions on G are more general than before: we may take any G which
does not contain any pseudo-reflections and the dimension of V may be greater than 2:
Theorem 4.3 (Auslander). Let S be as above and assume that G ⊆ GL(V ), dim V = n, is
small and set R = SG . Then we have an isomorphism of algebras:
∼
=
A = S∗G −
→ EndR (S) , sg 7→ ( x 7→ sg( x )) .
Moreover, S ∗ G is a CM-module over R and gldim(S ∗ G ) = n, and Z ( A) = R.
10
RAGNAR-OLAF BUCHWEITZ, ELEONORE FABER, AND COLIN INGALLS
This result can be phrased in terms of NCCRs:
Corollary 4.4. If G ⊆ SL(V ), that is, R is a Gorenstein singularity, then A is an NCCR of R. If
G ⊆ GL(V ) is small and not in SL(V ), then A is an NCR of R.
Proof. The twisted group ring A ∼
= EndR S is an endomorphism ring of a faithful CMmodule over R. Since gldim A = n < ∞, A is an NCR. For the crepancy in the SL(V )
case first note that R is a normal Gorenstein domain by the theorems of Hochster–Roberts
and Watanabe. Then the fact that S ∗ G ∼
= EndR S is CM over R implies that EndR S is a
nonsingular order, see [VdB04a, Lemma 4.2] and thus an NCCR.
For G = Γ ⊆ SL(V ) and dim V = 2, Auslander’s theorem implies a 1-1-correspondence
between the indecomposable CM-modules over R, the indecomposable summands of S,
the indecomposable projective modules over S ∗ G, and the irreducible Γ-representations.
Then McKay’s observation gives the correspondence to the geometry, namely, the bijection to the irreducible components of the irreducible divisor on the minimal resolution of
R.
We shortly sketch the correspondences: first, by Herzog’s theorem the indecomposable
objects in CM( R) are in bijection to the indecomposable R-summands of S. For the second
part, send the indecomposable summand M of S to HomR (S, M), which is an indecomposable projective EndR (S)-module. By Auslander’s theorem, it is an indecomposable
projective S ∗ Γ-module. For the last correspondence, send a indecomposable projective
S ∗ Γ-module to P/( x, y) P, which is a simple module over kΓ and hence an irreducible
Γ-representation. For details about these constructions, see [Aus86, Yos90, LW12].
4.4. Auslander’s theorem and reflections. Looking at Auslander’s theorem, it is natural
to ask:
Question 4.5. Let G ⊆ GL(V ), dim V = n, be generated by pseudo-reflections. Is there
an analogue of Auslander’s theorem? In particular, is there also a noncommutative resolution hidden somewhere?
The first problem in generalizing Auslander’s theorem is that if G is a pseudo-reflection
group, then the map
A = S ∗ G ,→ EndR (S)
is no longer surjective. One can see this for example for G = Sn . Then EndR (S) contains
the so-called Demazure operators, see e.g. [LS92], which are not contained in the skew
group ring. However, one can show in this case that A is the intersection of EndR (S) ∩
(EndR (S))τ in Q( A), the quotient ring of A, where τ is a particular anti-involution on
EndR (S), for details see [KK86].
Another “problem” is that R = SG is no longer singular if G is generated by pseudoreflections, by Thm. 2.1. This, however, can be remedied by looking at the discriminant
∆ ∈ R: the discriminant defines a singular hypersurface, whose singularities occur in
codimension 1. More precisely: it was shown by Saito and Terao that the discriminant of
a reflection arrangement A( G ) is a so-called free divisor in R, that is, the Jacobian ideal of
∆ is a CM-module over R/(∆). One can also show that the reflection arrangement itself
defines a free divisor in S. For more details, see [OT92].
NONCOMMUTATIVE RESOLUTIONS
11
The above facts already suggest our strategy to attack the problem of finding an analogue
of Auslander’s theorem and a noncommutative resolution for the discriminant: we will
start with the twisted group ring S ∗ G and using the small group G ∩ SL(V ) we will cook
up a NCR of ∆, which will also yield a 1−1-correspondence to irreducible representations
of G.
5. NCR S OF DISCRIMINANTS
For this section let us change the notation slightly: first, we will assume that G is a true
reflection group, i.e., G ⊆ GL(V ), dim V = n, is generated by reflections of order 2 (some
of our results also hold for pseudo-reflection groups but the main correspondence only
works in this case so far). We will denote the invariant ring SG by T. By Theorem 2.1
T ∼
= k[ f 1 , . . . , f n ] for the basic invariants f i ∈ S. Since G is a true reflection group J = z
∂f
and the hyperplane arrangement A( G ) is defined by the polynomial J = det( ∂xij ). The
discriminant ∆ is then defined by the equation J 2 = ∆, where ∆ is in T. Its coordinate
ring will be denoted by T/(∆). For our construction of the NCR of ∆, we consider the
small group Γ = G ∩ SL(V ), with invariant ring R = SΓ . We have an exact sequence of
groups 1 → Γ → G → H → 1, where the quotient is H ∼
= µ2 = hσi. Moreover, as above,
we denote by A = S ∗ G the twisted group ring of G.
The first result deals with the problem that A is not isomorphic to EndT (S). We can show
that A is still isomorphic to an endomorphism ring - yet over another twisted group ring:
Theorem 5.1. With notation introduced in the paragraph above, there is an isomorphism of rings
EndR∗ H (S ∗ Γ) ∼
= S ∗ G.
This result follows from generalizing the approach of H. Knörrer for curves [Knö84] (also
using ideas from J. Bilodeau [Bil05]).
For simplifying notation set B := R ∗ H. Now one can interpret B as the path algebra over
T of a quiver Q modulo some relations. Then B has idempotents e± = 12 (1 ± σ), where
σ is the generator of H, as defined above. One crucial observation here is that B/Be± B is
isomorphic to T/(∆) as rings. Thus, using the functor
i∗ (−)=−⊗ B B/Be− B
mod( B/Be− B) ←−−−−−−−−−−− mod( B)
from the standard recollement we obtain the following result:
Theorem 5.2. The functor i∗ induces an equivalence of categories
CM( B)/he− Bi ' CM( T/(∆)),
where he− Bi denotes the ideal generated by the object e− B in the category CM( B).
Theorem 5.1 shows that A is in some sense too large for being a NCR of ∆. But theorem
5.2 suggests that we have to quotient A by an ideal: since A ∈ CM( B), we calculate its
image in CM( B)/he− Bi and find that it is isomorphic to the quotient Ā = A/AeA, where
e = |G1 | ∑ g∈G g ∈ A is the idempotent corresponding to the trivial representation of G. On
the other hand, calculate the image of A under the recollement functor i∗ : one gets that
i∗ ( A) = EndT/(∆) (i∗ (S ∗ H )) .
12
RAGNAR-OLAF BUCHWEITZ, ELEONORE FABER, AND COLIN INGALLS
Exploiting results from Stanley about the structure of R as T-module [Sta77], we obtain
that i∗ (S ∗ H ) ∼
= S/( J ), as T/(∆)-module. As a ring, S/( J ) is precisely the coordinate
ring of the reflection arrangement A( G )! In the last step, a result of Auslander–Platzeck–
Todorov [APT92] about global dimension of quotients yields the following
Theorem 5.3. With notation as above,
Ā := A/AeA ∼
= EndT/(∆) (S/( J )),
and gldim Ā < ∞. Thus Ā yields a NCR of the free divisor T/∆.
Moreover, the indecomposable projective modules over Ā are in bijection with the non-trivial representations of G and also with certain CM-modules over the discriminant, namely the T/(∆)-direct
summands of S/( J ).
Thus we not only obtain a NCR of discriminants of reflection arrangements but also a
McKay correspondence for reflection groups G.
Example 5.4. Let dim V = 2 and let G ⊆ GL(V ) be a true reflection group. Then the
discriminant ∆ is an ADE-curve singularity. We obtain that S/( J ) is a generator of the category of CM-modules over the T/(∆), that is, add(S/( J )) = CM( T/(∆)). Moreover, the
multiplicities, with which the indecomposables appear in S/( J ) are precisely the ranks of
the corresponding irreducible representations.
As a particular example, take G = S3 , whose discriminant ∆ is a cusp (aka A2 -singularity).
For the cusp, one can show that S/( J ) ∼
= T/(∆) ⊕ m⊕2 as T/(∆)-modules, where m denotes the maximal ideal (u, v) in T/(∆) = k [u, v]/(u3 − v2 ). Here one sees that T/(∆)
corresponds to the determinantal representation and m to the canonical representation
given by G ,→ GL(2, k ).
6. F URTHER QUESTIONS
In general we are still missing a conceptual explanation for the T/(∆)-direct summands
of S/( J ): in dimension 2, see example 5.4, any indecomposable CM-module over T/(∆)
can be obtained from S/( J ) (via matrix factorizations). But if dim V ≥ 3, that is, the
Krull-dimension of T/(∆) is greater than or equal to 2, T/(∆) is not of finite CM-type.
However, the factors of S/( J ) are still CM-modules and can be calculated as matrix factorizations. So far we only have few examples (the normal crossing divisor, as studied in
Dao–Faber–Ingalls [DFI15] or the swallowtail, that is, the discriminant of S4 , where we
can describe the T/(∆)-direct summands using Hovinen’s classification of graded rank 1
MCM-modules [Hov09]).
Thus, for a geometric interpretation of the direct summands of S/( J ) we want to establish
a similar correspondence as in [GSV83]. A next step would be to realize geometric resolutions as moduli spaces of isomorphism classes of representations of certain algebras as
in [CB00].
But, in a different vein, the structure of Ā as an S/( J )-module seems to be easier to understand: Ā is isomorphic to the cokernel of the map ϕ given by left multiplication on G, i.e.,
the matrix of ϕ corresponds to the multiplication table of the group G. Since Frobenius
and Dedekind, the block decomposition of this matrix is well-known. Our goal is to use
this surprising discovery to learn more about the decomposition of Ā and hence of S/( J )
over the discriminant.
NONCOMMUTATIVE RESOLUTIONS
13
Moreover, looking at the quiver of Ā, the exact form of the generating (necessarily quadratic) relations remains mysterious and will be the subject of further research.
7. A CKNOWLEDGEMENTS
We want to thank the organizers of the ICRA for making it such a pleasant and stimulating conference. In particular, we thank Graham Leuschke for his encouragement
to write this proceedings article. We acknowledge and highly appreciate that the research presented in this paper was supported through the program “Research in Pairs” in
2015 and the program “Oberwolfach Leibniz Fellowships” in 2016 by the Mathematisches
Forschungsinstitut Oberwolfach and by the Institut Mittag-Leffler (Djursholm, Sweden)
in frame of the program “Representation theory” in 2015.
R EFERENCES
M. Auslander, M. I. Platzeck, and G. Todorov. Homological theory of idempotent ideals. Trans.
Amer. Math. Soc., 332(2):667–692, 1992. 12
[Aus86]
M. Auslander. Rational singularities and almost split sequences. Transactions of the AMS,
293(2):511–531, 1986. 9, 10
[BFI16]
R.-O. Buchweitz, E. Faber, and C. Ingalls. A McKay correspondence for reflection groups. 2016.
in preparation. 1
[BGS87]
R.-O. Buchweitz, G.-M. Greuel, and F.-O. Schreyer. Cohen-Macaulay modules on hypersurface
singularities. II. Invent. Math., 88(1):165–182, 1987. 9
[BIKR08] I. Burban, O. Iyama, B. Keller, and I. Reiten. Cluster tilting for one-dimensional hypersurface
singularities. Adv. Math., 217(6):2443–2484, 2008. 6
[Bil05]
J. Bilodeau. Auslander algebras and simple plane curve singularities. In Representations of algebras
and related topics, volume 45 of Fields Inst. Commun., pages 99–107. Amer. Math. Soc., Providence,
RI, 2005. 11
[BKR01]
T. Bridgeland, A. King, and M. Reid. The MacKay correspondence as an equivalence of derived
categories. J. Amer. Math. Soc., 14(3):535–554, 2001. 6
[BLvdB10] R.-O. Buchweitz, G. J. Leuschke, and M. van den Bergh. Non-commutative desingularization of
determinantal varieties I. Invent. Math., 182(1):47–115, 2010. 6
[BO02]
A. Bondal and D. Orlov. Derived categories of coherent sheaves. In Proceedings of the International
Congress of Mathematicians, Vol. II (Beijing, 2002), pages 47–56. Higher Ed. Press, Beijing, 2002. 6
[Bou81]
Nicolas Bourbaki. Éléments de mathématique. Masson, Paris, 1981. Groupes et algèbres de Lie.
Chapitres 4, 5 et 6. [Lie groups and Lie algebras. Chapters 4, 5 and 6]. 2
[Bri02]
T. Bridgeland. Flops and derived categories. Invent. Math., 147(3):613–632, 2002. 6
[Buc12]
R.-O. Buchweitz. From platonic solids to preprojective algebras via the McKay correspondence.
Oberwolfach Jahresbericht, pages 18–28, 2012. 6
[CB00]
W. Crawley-Boevey. On the exceptional fibres of Kleinian singularities. Amer. J. Math.,
122(5):1027–1037, 2000. 12
[Che55]
C. Chevalley. Invariants of finite groups generated by reflections. Amer. J. Math., 77:778–782, 1955.
3
[CLO07] D. Cox, J. Little, and D. O’Shea. Ideals, Varieties, and Algorithms. Springer, New York, 3 edition,
2007. 3
[DFI15]
H. Dao, E. Faber, and C. Ingalls. Noncommutative (crepant) desingularizations and the global
spectrum of commutative rings. Algebr. Represent. Theory, 18(3):633–664, 2015. 1, 6, 12
[DFI16]
B. Doherty, E. Faber, and C. Ingalls. Computing global dimension of endomorphism rings via
ladders. J. Algebra, 458:307–350, 2016. 6
[DII+ 16] H. Dao, O. Iyama, S. Iyengar, R. Takahashi, M. Wemyss, and Y. Yoshino. Noncommutative resolutions using syzygies. 2016. arXiv:1609.04842. 6
[DITV15] H. Dao, O. Iyama, R. Takahashi, and C. Vial. Non-commutative resolutions and Grothendieck
groups. J. Noncommut. Geom., 9(1):21–34, 2015. 6
[Dur79]
A.-H. Durfee. Fifteen characterizations of rational double points. L’Enseignement Mathématique,
25:131–163, 1979. 7
[APT92]
14
[FH10]
[Gre92]
[GSV83]
[Hau03]
[Her78]
[Hir64]
[Hov09]
[HR74]
[IW14]
[KK86]
[Kle93]
[Knö84]
[Knö87]
[Lau71]
[Leu07]
[Leu12]
[LS92]
[LT09]
[LW12]
[McK80]
[OT92]
[Rei02]
[Sta77]
[VdB04a]
[VdB04b]
[Wat74]
[Yos90]
RAGNAR-OLAF BUCHWEITZ, ELEONORE FABER, AND COLIN INGALLS
E. Faber and H. Hauser. Today’s menu: geometry and resolution of singular algebraic surfaces.
Bull. Amer. Math. Soc., 47(3):373–417, 2010. 7
G.-M. Greuel. Deformation und Klassifikation von Singularitäten und Moduln. Deutsch. Math.Verein. Jahresber. Jubiläumstagung 1990, pages 177–238, 1992. 7
G. Gonzalez-Sprinberg and J.-L. Verdier. Construction géométrique de la correspondance de
McKay. Ann. Sci. École Norm. Sup. (4), 16(3):409–449 (1984), 1983. 9, 12
H. Hauser. The Hironaka Theorem on resolution of singularities. Bull. Amer. Math. Soc., 40:323–
403, 2003. 5
Jürgen Herzog. Ringe mit nur endlich vielen Isomorphieklassen von maximalen, unzerlegbaren
Cohen-Macaulay-Moduln. Math. Ann., 233(1):21–34, 1978. 9
H. Hironaka. Resolution of singularities of an algebraic variety over a field of characteristic zero.
Ann. Math., 79:109–326, 1964. 5
B. Hovinen. Matrix factorizations of the classical determinant. Phd thesis, University of Toronto,
2009. 12
Melvin Hochster and Joel L. Roberts. Rings of invariants of reductive groups acting on regular
rings are Cohen-Macaulay. Advances in Math., 13:115–175, 1974. 3
O. Iyama and M. Wemyss. Maximal modifications and Auslander-Reiten duality for non-isolated
singularities. Invent. Math., 197(3):521–586, 2014. 6
Bertram Kostant and Shrawan Kumar. The nil Hecke ring and cohomology of G/P for a KacMoody group G. Proc. Nat. Acad. Sci. U.S.A., 83(6):1543–1545, 1986. 10
F. Klein. Vorlesungen über das Ikosaeder und die Auflösung der Gleichungen vom fünften Grade.
Birkhauser Verlag, Berlin, 1993. Reprint of the 1884 original, edited, with an introduction and
commentary by Peter Slodowy. 6
H. Knörrer. The Cohen–Macaulay modules over simple hypersurface singularities. 1984. MaxPlanck Institut für Mathematik und Sonderforschungsbereich theoretische Physik MPI/SFB 8451. 11
Horst Knörrer. Cohen-Macaulay modules on hypersurface singularities. I. Invent. Math.,
88(1):153–164, 1987. 9
H. Laufer. Normal Two-dimensional Singularities, volume 71 of Annals of Math. Studies. Princeton
Univ. Press, 1971. 7
G. J. Leuschke. Endomorphism rings of finite global dimension. Canad. J. Math., 59(2):332–342,
2007. 6
G. J. Leuschke. Non-commutative crepant resolutions: scenes from categorical geometry. In
Progress in commutative algebra 1, pages 293–361. de Gruyter, Berlin, 2012. 6
Alain Lascoux and Marcel-Paul Schützenberger. Décompositions dans l’algèbre des différences
divisées. Discrete Math., 99(1-3):165–179, 1992. 10
G. I. Lehrer and D. E. Taylor. Unitary reflection groups, volume 20 of Australian Mathematical Society
Lecture Series. Cambridge University Press, Cambridge, 2009. 2
G. J. Leuschke and R. Wiegand. Cohen-Macaulay representations, volume 181 of Mathematical Surveys and Monographs. American Mathematical Society, Providence, RI, 2012. 10
John McKay. Graphs, singularities, and finite groups. In The Santa Cruz Conference on Finite Groups
(Univ. California, Santa Cruz, Calif., 1979), volume 37 of Proc. Sympos. Pure Math., pages 183–186.
Amer. Math. Soc., Providence, R.I., 1980. 8
P. Orlik and H. Terao. Arrangements of hyperplanes, volume 300 of Grundlehren der Mathematischen
Wissenschaften. Springer-Verlag, Berlin, 1992. 2, 3, 10
M. Reid. La correspondance de McKay. Astérisque, (276):53–72, 2002. Séminaire Bourbaki, Vol.
1999/2000. 6
R. P. Stanley. Relative invariants of finite groups generated by pseudoreflections. J. Algebra,
49(1):134–148, 1977. 12
M. Van den Bergh. Non-commutative crepant resolutions. In The legacy of Niels Henrik Abel, pages
749–770. Springer, Berlin, 2004. 6, 10
M. Van den Bergh. Three-dimensional flops and noncommutative rings. Duke Math. J., 122(3):423–
455, 2004. 6
Keiichi Watanabe. Certain invariant subrings are Gorenstein. I, II. Osaka J. Math., 11:1–8; ibid. 11
(1974), 379–388, 1974. 3
Y. Yoshino. Cohen–Macaulay modules over Cohen–Macaulay rings, volume 146 of London Mathematical Society Lecture Note Series. Cambridge University Press, Cambridge, 1990. 10
NONCOMMUTATIVE RESOLUTIONS
[YS81]
15
Tamaki Yano and Jirō Sekiguchi. The microlocal structure of weighted homogeneous polynomials
associated with Coxeter systems. II. Tokyo J. Math., 4(1):1–34, 1981. 5
D EPT. OF C OMPUTER AND M ATHEMATICAL S CIENCES , U NIVERSITY OF T ORONTO AT S CARBOROUGH , 1265
M ILITARY T RAIL , T ORONTO , ON M1C 1A4, C ANADA
E-mail address: [email protected]
D EPARTMENT OF M ATHEMATICS , U NIVERSITY OF M ICHIGAN , A NN A RBOR MI, 48109, USA
E-mail address: [email protected]
D EPARTMENT OF M ATHEMATICS
E3B 5A3, C ANADA
E-mail address: [email protected]
AND
S TATISTICS , U NIVERSITY
OF
N EW B RUNSWICK , F REDERICTON , NB.
| 0math.AC
|
Implementation of a “language driven” Backpropagation algorithm
I.V. Grossua,*, C.I. Ciuluvica (Neagu)b,*
a
University of Bucharest, Faculty of Physics, Bucharest-Magurele, P.O. Box MG 11, 077125,
Romania
b
University of Bucharest, Faculty of Psychology and Education Sciences, 90 Panduri Street,
District 5 Bucharest, P.O. 050663, Romania
ABSTRACT
Inspired by the importance of both communication and feedback on errors in human learning, our main goal was
to implement a similar mechanism in supervised learning of artificial neural networks. The starting point in our
study was the observation that words should accompany the input vectors included in the training set, thus
extending the ANN input space. This had as consequence the necessity to take into consideration a modified
sigmoid activation function for neurons in the first hidden layer (in agreement with a specific MLP apartment
structure), and also a modified version of the Backpropagation algorithm, which allows using of unspecified
(null) desired output components. Following the belief that basic concepts should be tested on simple examples,
the previous mentioned mechanism was applied on both the XOR problem and a didactic color case study. In
this context, we noticed the interesting fact that the ANN was capable to categorize all desired input vectors in
the absence of their corresponding words, even though the training set included only word accompanied inputs,
in both positive and negative examples. Further analysis along applying this approach to more complex
scenarios is currently in progress, as we consider the proposed “language-driven” algorithm might contribute to
a better understanding of learning in humans, opening as well the possibility to create a specific category of
artificial neural networks, with abstraction capabilities.
1. Introduction
Given the importance of communication in human learning [1-4], in this work we present
an attempt of implementing a “communication language” in supervised learning of artificial
neural networks. In this context, the significant role played by error feedbacks was noticed
also. Thus, as Blair [1] suggests, nearly all categorization models take errors to be the
essential ingredient in learning. According to Van Dyck et al. [5], the evaluation of past
behaviors and acting upon the awareness that errors hold useful information are indeed
considered as important practices to learn from errors. In literature, several authors:
Edmonson [6], Van Dyck et al. [5], Rochlin [7], argue that communication is one of the most
important conditions so that learning from errors should manifest.
In the field of artificial intelligence, communication in multi-agent systems [8] using
neural networks is treated, for example, in the research of Steels & Belpaeme [9]. In the
present work, we analyze a more restrictive, unidirectional (“trainer” - ANN) communication
type, emphasizing the idea that words injected from the environment, in both positive and
negative examples, should act as “learning triggers”. Although we believe some of the
discussed concepts could be used in the frame of other paradigms, for simplicity, this study
was limited to the case of supervised learning.
*Corresponding authors. E-mail addresses:
[email protected] (I.V.Grossu), [email protected] (C.I.Ciuluvica)
2. Implementation of a “language-supervised” training algorithm
Starting from the idea that both communication and feedback to errors should play an
important role in supervised learning, our main purpose was to create an ANN with the
following capabilities:
•
•
•
•
Communicability. It is important to notice that we are treating only the particular case
of unidirectional “trainer” - ANN communication, in which the language is used as a
training instrument;
Abstraction. Capability of recognizing words in the absence of any input vector;
Pattern recognition. Capability of categorizing input vectors in the absence of
corresponding language information;
Sensitivity to errors. Capability of recognizing incorrect (word, input vector)
associations.
Based on the observation that, in the training process, words should extend the ANN
input space, acting as “learning triggers”, the proposed model is constructed by analogy with
the human brain, which neural activity is based on two fundamental inseparable events:
excitation and inhibition. In the human brain there are two simple inhibitory circuits
(feedback – for local network activity), and feed-forward (generated by the activity of others
cortical regions), representing the fundamental building blocks of the cortical architecture
[10]. According to the literature, the relationship between excitation and inhibition
may be involved in various physiological functions, including increasing stability of cortical
activity, preventing runaway excitation [11], regulating the neuronal output [12], and also the
supervised learning of human behavior (the Pavlov’s Conceptual Nervous System).
In this context, we considered a specific ANN modular architecture, comprising one word
recognition component (which might be implemented using various specific systems [1315]), and a Multilayer Perceptron, responsible for learning the specific function of interest.
The MLP main module has an “apartment” structure (interconnected regions of adjacent
hidden and output neurons), and receives as input both regular data and the output of the
word recognition component. We considered a specific relation between words and neural
apartments. Thus, presenting a word to the ANN will result in inhibiting (infinite dendrite
weights) all apartments, excepting a corresponding one. In a dynamically approach, each time
a new word is presented to the ANN input, a new empty apartment will be allocated
accordingly.
For avoiding the complexity involved by lexical ambiguity [13], one can restrict the
communication to a basic dictionary, containing only unambiguous words. In this case, a
unique positive integer could be associated to each word (zero indicating the absence of
language information). For simplicity, we will further use the term “word” for both words
and for their corresponding unique identifiers. In this context, the previous described ANN
could be replaced with an equivalent, simplified system, which includes only the apartment
MLP module, and which input space is extended with the language component. The relation
between words and apartments is resolved in the first hidden layer, by considering constant
language dendrite weights (not modified during the training process), equaling their
corresponding words. The word dendrite contributes also to the neuron activation:
2
where y is the neuron output, f the activation function, t the neuron threshold, wword is the
weight of the language dendrite, n the dimension of the input data, w the dendrites weights
vector, and x is the input vector.
The apartment inhibition is achieved by implementing a modified sigmoid activation
function for neurons in the first hidden layer:
where xword is the word input, and w the weight of the word dendrite
Inspired by the importance of error feed-backs in human learning [1-7], we considered
that the training set should include both positive and negative examples. For error
recognition, one can envisage various approaches:
•
•
•
Extending the output space by adding one neuron for each error category;
Using bipolar neurons (1 will indicate an affirmation, while -1 a negation);
Associate errors with the absence of any output signal (the “zero output” criteria).
As training technique we used a modified version of the Backpropagation [16-18]
algorithm, capable of ignoring error signals coming from a specified set of output neurons.
The neurons that must be ignored are identified, during the training process, by their
corresponding null (unspecified) desired outputs. The main role of this mechanism is related
to the creation of some privileged training neural channels (apartments) inside the ANN.
3. Application to the XOR problem
For testing the proposed approach, an object oriented three layers perceptron was
implemented in Microsoft C# 2010. The input space was extended with the language
dimension (which accepts only integer values). The activation functions for neurons in the
first hidden layer were defined according to Eq.2, while all other neurons are using the
classical sigmoid function. The previously discussed version of the Backpropagation
algorithm ignores the error signals coming from output neurons with unspecified (null)
expected values. Thus, the desired output was designed as a vector of nullable values [19]. A
set of specific techniques [20] have been also considered for improving the efficiency of the
training algorithm: the neural weights are initialized using random values in the [-2/nd, 2/nd]
interval, nd representing the dendrites number for the corresponding neuron, the learning rate
c is decreased in time following a c/n law (where n is a strictly positive integer, incremented
after a specified number of iterations), and the elements of the training set are permuted after
each iteration. A “shaking ANN weights” mechanism was also implemented for avoiding
local minimum attractors.
The training set must contain both positive and negative examples. In this context, the
previously discussed “zero output” error recognition criterion was empirically chosen. As a
training success condition, a maximum target error, specified as parameter, was required to
be satisfied by each input vector from the training set. The classification error is defined
based on the Euclidian distance between the ANN response and the desired output:
3
where d is the desired output, nulls being ignored, and o the corresponding output.
Following the belief that basic concepts should be tested on simple examples, the XOR
problem has been chosen as a first concrete application to be discussed. A two-apartment
MLP was used following this purpose (Fig.1.). As the number of effective patterns (word –
input vector combinations) is increased with respect to the classical XOR problem, a higher
number of neurons should be allocated in the hidden layer [21]. The training set, together
with one approximate solution (c = 0.5, and target error = 0.001) is presented, for
exemplification, in Tab.1.
x1
x2
xword
0
1
0
1
0
1
1
0
0
1
0
1
0
1
1
0
0
1
0
1
0
1
1
0
d1
o1
d2
positive examples
1
1
0.96
null
1
1
0.97
null
2
null
0.09
1
2
null
0.14
1
negative examples
2
null
0.13
0
2
null
0.11
0
1
0
0.04
null
1
0
0.04
null
result (not included in the training set)
0
0.96
0
0.96
0
0.04
0
0.03
-
o2
Dε
0.45
0.47
0.96
0.96
0.0009
0.0004
0.0005
0.0007
0.02
0.04
0.49
0.39
0.0002
0.001
0.0007
0.0008
0.02
0.05
0.95
0.97
-
Tab.1. An approximate solution for the XOR problem.
One can notice that the ANN is able to learn all examples from the training set, but it is
also capable of categorizing all input vectors in the absence of their corresponding words,
even though this kind of information was not explicitly included in the training set. The
success rate, measured for 1000 distinct training processes, is 98%. For the example
presented in Tab.2, it is also interesting to remark that, for all null desired values, the
corresponding outputs are close to zero. However, this effect is obtained with a low
frequency (19%). For avoiding this uncertainty, the algorithm was tested also with the
complete training set (which includes input without word examples, and in which all nulls are
replaced with zero). The same success rate was obtained in this case.
4
Fig.1. The ANN topology used for the XOR problem.
4. A didactic color case study
As a more complex example, we considered the didactic problem of categorizing the
colors corresponding to the RGB cube vertices. An eight-apartment structure was used in this
case. The number of hidden neurons per apartment (12) was empirically established. For each
class, a number of 29 input vectors were symmetrically chosen around each vertex (Fig.2.).
The training set is constructed by analogy with the data presented in Tab.1. It contains both
positive and negative examples, summarizing 1856 elements. For avoiding high execution
times, the target error was set to 0.1, while the learning rate was empirically chosen (c =
0.005).
Fig.2. The RGB vectors used in the training set.
5
Similarly with the, previously discussed, XOR problem, the ANN is capable of
categorizing all inputs without word vectors (Tab.2), even though this kind of information is
missing from the training set.
xRed
0
255
0
0
255
0
255
255
xGreen xBlue
oBlack oBlue
oMagenta oRed
oYellow oGreen oCyan
oWhite
0
0
0.91
0.06
0
0.04
0.08
0.05
0
0.05
0
0
0.1
0
0.06
0.94
0.07
0.02
0.01
0.03
255
0
0.1
0.01
0
0.01
0.06
0.94
0.04
0.04
0
255
0.05
0.88
0.05
0.02
0
0.02
0.13
0.04
255
0
0.03
0
0.01
0.01
0.92
0.03
0
0.04
255
255
0.04
0.01
0
0.01
0
0.02
0.87
0.05
0
255
0.04
0.01
0.9
0.02
0.02
0.01
0.01
0.04
255
255
0.05
0.01
0
0.02
0.01
0.01
0.01
0.89
Tab.2. An approximate solution for the discussed RGB color categorization problem.
The algorithm success rate, measured for 300 distinct training processes, is 99.6%. We
noticed also that, the output values for neurons without expectations (null desired value) have
an accumulation to zero tendency (the absolute maximum value is 0.89, while, in average,
71% of all outputs are below 0.5, which is convenient for the target error we chose). The
probability distribution of the maximum output for neurons with unspecified desired values is
presented in Fig.3. Thus, the zero output error recognition condition could be extended to a
more practical, “noisy output criterion” (high entropy output, for which a maximum value
cannot be clearly distinguished). As opposed to the XOR problem, the algorithm did not
prove its efficiency for the complete training set. One possible amelioration could come from
using a robust Backpropagation algorithm [22], together with a more advanced local
minimum avoiding technique (e.g. the simulated annealing [23]).
Fig.3. The probability distribution of the maximum output for neurons with unspecified desired values
6
5. Conclusions
Starting from the importance of both communication and feedback on errors in human
learning, we tried to implement similar concepts in supervised learning of artificial neural
networks. Following this purpose, the input space was extended with the language dimension,
a modified sigmoid activation function was considered for neurons in the first hidden layer
(in agreement with a specific MLP apartment structure), a modified Backpropagation
algorithm was implemented, in order to allow using unspecified (null) desired output
components, and both positive and negative examples were included in the training set. For a
better understanding of basic concepts involved, the previous discussed mechanism have
been tested on both the XOR problem and a simple RGB color case study. As an interesting
result, one can notice that the ANN is capable of recognizing input vectors in the absence of
their corresponding words, even when this kind of information is not explicitly included in
the training set. This fact could be connected with the interference between communication
and feedback on errors mechanisms, whose importance is discussed also in psychology [5-7].
Compared to the classical Backpropagation algorithm, the proposed approach is much
more resource intensive, especially because both positive and negative examples has been
taken into consideration. However, the amount of data stored into the ANN weights is
augmented too, as it includes also language information. The “language-driven”
Backpropagation algorithm could have also a cultural role, as words injected from the
environment have an effect of modifying the distance between input vectors, by dividing the
extended input space in sub-spaces sharing the same language component.
Further analysis along applying this approach to more complex scenarios is currently in
progress, as we consider the proposed “language-driven” algorithm might contribute to a
better understanding of learning in humans, opening as well the possibility of creating a
specific category of artificial neural networks, with abstraction capabilities.
References
[1]
[2]
[3]
[4]
[5]
[6]
[7]
[8]
[9]
R.M. Blair, R.M. Watson, M.M. Kimberly. Errors, efficiency, and the interplay between
attention and category learning, Cognition, 2009, 112, 330-336
G.L. Murphy, The Big Book of Concepts, MIT Press, 2002
B.H. Ross, E.G. Taylor, and E.L. Middleton, Concept and Category Learning in Humans,
Learning and Memory, A Comprehensive Reference, 2008, 2, 535-556
J.R. Anderson, The adaptive nature of human categorization, Psychological Review,
1991, 98, 409–429
C. Van Dyck, M. Frese, M. Baer, & S. Sonnentag, Organizational error management
culture and its impact on performance: A two-study replication, Journal of Applied
Psychology, 2005, 90, 1228–1240
A.C. Edmondson, Learning from mistakes is easier said than done: Group and
organizational influences on the detection and correction of human error. Journal of
Applied Behavioral Science, 1996, 32, 5–28
G.I. Rochlin, Safe operation as a social construct, Ergonomics, 1999,42(11),1549–60
Ferber, J. (1998) Multi-agent systems: An introduction to distributed artificial
intelligence. Addison-Wesley. [aLS, TS]
Steels L, Belpaeme T., Coordinating perceptually grounded categories through language:
a case study for colour, Behav Brain Sci. 2005 Aug;28(4):469-89
7
[10]
[11]
[12]
[13]
[14]
[15]
[16]
[17]
[18]
[19]
[20]
[21]
[22]
[23]
Van Vreeswijk, C., and Sompolinsky, H., (1996). Chaos in neuronal networks with
balanced excitatory and inhibitory activity. Science 274, 1724-1726;
Tsodyks, M.V., Skaggs, W.E., Sejnowssky, T.J., McNaughton, B.L., (1997). Paradoxical
effects of external modulation of inhibitory interneurons. Neurosciences Journal 17:43824388;
Isaacson, S.I., and Scanziani, M., (2011). How inhibition shapes cortical activity. Neuron
72: 231-243
A.H. Kawamoto, Journal of Memory and Language, 08/1993; 32(4):474–516.
S. Pirmoradian, A. Treves, Cognitive Computation, December 2011, Volume 3, Issue 4,
pp 539-553
J.Cadore, F.J.Valverde-Albacete, A. Gallardo-Antolín, C. Peláez-Moreno, Cognitive
Computation, December 2013, Volume 5, Issue 4, pp 426-441
P. Antognetti, V. Milutinovic, Neural Networks. Concepts. Applications and
Implementations, vol 1…4, Prentice Hall, Englewood Cliffs, New Jersey, 1991
J.A. Freeman, D.M. Skapura, Neural Networks. Algorithms, Applications, and
Programming techniques, Addison-Wesley Publ. Co., 1991
R.H. Silverman, A.S. Noetzel, Image processing and Pattern Recognition in
Ultrasonograms by Backpropagation, Neural Networks, vol. 3 (1990), 593-603
Joseph Albahari, Ben Albahari, C# 4.0 in a Nutshell, O’Reilly, US, 2010, 237-240, 681733, 789-941
D. Dumitrescu, H. Costin, Retele Neuronale Teorie si Aplicatii, Teora, Romania,
1996,192-227
A. Dumitras, Proiectarea retelelor neurale artificiale, Odeon, Bucuresti, 1997, 81-202
Huber, P.J., Robust Statistics, Wiley, New York, 1981
Szu, H., Hartley, R., Fast Simulated Annealing, Physics Letters, 1222 (1987), 157-162
8
| 9cs.NE
|
1
On the Properties of the Compound Nodal
Admittance Matrix of Polyphase Power Systems
arXiv:1712.08764v1 [cs.SY] 23 Dec 2017
Andreas Martin Kettner, Member, IEEE, and Mario Paolone, Senior Member, IEEE
Abstract—Most techniques for power system analysis model
the grid by exact electrical circuits. For instance, in power flow
study, state estimation, and voltage stability assessment, the use
of admittance parameters (i.e., the nodal admittance matrix)
and hybrid parameters is common. Moreover, network reduction
techniques (e.g., Kron reduction) are often applied to decrease the
size of large grid models (i.e., with hundreds or thousands of state
variables), thereby alleviating the computational burden. However, researchers normally disregard the fact that the applicability
of these methods is not generally guaranteed. In reality, the nodal
admittance must satisfy certain properties in order for hybrid
parameters to exist and Kron reduction to be feasible. Recently,
this problem was solved for the particular cases of monophase and
balanced triphase grids. This paper investigates the general case
of unbalanced polyphase grids. Firstly, conditions determining
the rank of the so-called compound nodal admittance matrix
and its diagonal subblocks are deduced from the characteristics
of the electrical components and the network graph. Secondly, the
implications of these findings concerning the feasibility of Kron
reduction and the existence of hybrid parameters are discussed. In
this regard, this paper provides a rigorous theoretical foundation
for various applications in power system analysis.
Index Terms—Admittance parameters, hybrid parameters,
Kron reduction, multiport network, nodal admittance matrix,
polyphase power systems, unbalanced power grids
I. I NTRODUCTION
I
NHERENTLY, techniques for power system analysis need
an exact analytical description of the grid. This description
is normally deduced from an equivalent electrical circuit. For
instance, in Power Flow Study (PFS, [1]–[3]), State Estimation
(SE, [4]–[6]), and Voltage Stability Assessment (VSA, [7]–[9]),
the use of admittance parameters (i.e., the nodal admittance
matrix) or hybrid parameters (i.e., hybrid parameters matrices)
is a common practice. Since the solution methods employed for
these applications are computationally heavy (e.g., [10]–[14]),
network reduction techniques like Kron reduction (see [15]) are
often applied in order to reduce the problem size. Thereby, the
computational burden is decreased, and the execution speed is
increased without the use of high-performance computers (e.g.,
[16]). However, neither the reducibility of the nodal admittance
matrix nor the existence of hybrid parameters are guaranteed
a priori. In order for this to be the case, the nodal admittance
matrix has to satisfy certain properties (i.e., the corresponding
diagonal subblocks have to be invertible).
Interestingly, most researchers and practitioners apparently
ignore this fact. There exist a handful of publications which
investigate the feasibility of Kron reduction (e.g., [17]) and
the existence of hybrid parameters matrices (e.g., [18]–[20]),
but their validity is limited. The former base their reasoning
upon trivial cases (i.e., purely resistive or inductive monophase
grids), and the latter establish feeble guarantees (i.e., hybrid
This work was supported by the Swiss National Science Foundation through
the National Research Programme NRP 70 “Energy Turnaround”.
The authors are with the Faculté Sciences et Techniques de l’Ingénieur
(STI) of the École Polytechnique Fédérale de Lausanne (EPFL) in Lausanne,
Vaud, Switzerland (e-mail: [email protected], [email protected]).
parameters may only exist for one partition of the nodes). Recently, the authors proved that the feasibility of Kron reduction
and the existence of hybrid parameters matrices (for arbitrary
partitions of the nodes) are ensured for electrical circuits which
are connected and have lossy branch impedances (see [21]). As
the development is based on two-port equivalents, the findings
hold for monophase grids or balanced triphase grids, which can
be equivalently transformed into positive-sequence networks
using the method of symmetrical components (see [22]).
This paper investigates the general case of unbalanced polyphase power systems (i.e., grids which cannot be decomposed
via the method of symmetrical components). As the method of
symmetrical components is not applicable, the polyphase grid
cannot be equivalently represented by decoupled monophase
grids. Hence, the electromagnetic coupling between the phases
has to be explicitly accounted for. For this purpose, the concept
of compound electrical parameters (see [23]) is adopted. The
authors of this paper argue that physical electrical components
are represented by equivalent circuits whose compound electrical parameters are symmetric, invertible, and passive. Using
mathematical derivation and physical reasoning, it is proven
that any diagonal subblock of the compound nodal admittance
matrix has full rank if the network graph is connected. Based
on this, it is demonstrated that the feasibility of Kron reduction
and the existence of hybrid parameters are guaranteed under
practical conditions. In that sense, this paper provides – for the
first time in the literature – a rigorous theoretical foundation
for the analysis of polyphase power systems, in particular for
PFS, SE, and VSA (cf. [24]–[26]).
The remainder of this paper is structured as follows: First,
the basic theoretical foundations are laid in Sec. II. Thereupon,
the properties of the compound nodal admittance matrix are
developed in Sec. III. Afterwards, the implications with respect
to Kron reduction and hybrid parameters matrices are deduced
in Sec. IV. Finally, the conclusions are drawn in Sec. V.
II. F OUNDATIONS
A. Numbers
Scalars are denoted by ordinary letters. The real and imaginary part of a complex scalar z ∈ C are denoted by ℜ{z} and
ℑ{z}, respectively. Thus, z can be expressed in rectangular
coordinates as z = ℜ{z} + jℑ{z}. The complex conjugate of
z is denoted by z ∗ . The absolute value and the argument of
z are denoted by |z| and arg (z), respectively. Thus, z can be
expressed in polar coordinates as z = |z| ∠ arg (z).
B. Set Theory
Sets are denoted by calligraphic letters. The cardinality of
a set A is denoted by |A|. The (set-theoretic) difference A \ B
of two sets A and B is defined as
A \ B := {x | x ∈ A, x ∈
/ B}
(1)
The Cartesian product A × B of A and B is defined as
A × B := {(a, b) | a ∈ A, b ∈ B}
(2)
2
(5)
Let MI′ ×J′ (I′ ⊆ I, J′ ⊆ J) be the submatrix consisting of
the blocks Mij (i ∈ I′ , j ∈ J′ ). Now, consider the particular
case I′ := I \ {|I|} and J′ := J \ {|J|}, and define
MI′ ×J′
MI′ ×{|J|}
A B
:=
M=
(12)
C D
M{|I|}×J′ M|I||J|
(6)
If D is invertible, the Schur complement M/D of D in M is
A partition of a set A is a family of sets {Ak | k ∈ K}, where
K := {1, · · · , |K|} is an integer interval, for which
Ak ⊆ A ∀k ∈ K
(3)
Ak 6= ∅ ∀k ∈ K
(4)
Ak ∩ Al = ∅ ∀k, l ∈ K, k 6= l
[
Ak = A
k∈K
M/D := A − BD−1 C
That is, the parts Ak are non-empty and disjoint subsets of A,
whose union is exhaustive. If all parts are proper subsets of A
Ak ( A (∀k ∈ K)
(7)
The following properties hold.
Lemma 6. det(M) = det(D) det(M/D) (see [30]).
Lemma 7. The Schur complement can be computed blockwise
Aij Bi
−1
(M/D)ij = Aij − Bi D Cj =
/ D (14)
Cj D
and {Ak | k ∈ K} 6= {A} (i.e., the partition is non-trivial).
C. Linear Algebra
Matrices and vectors are denoted by bold letters. Consider a
matrix M = (Mrc ) (r ∈ R, c ∈ C), where R := {1, · · · , |R|}
and C := {1, · · · , |C|} are the sets of row and column indices.
The field of matrices of size |R| × |C| over a field of numbers
F is denoted by M|R|×|C| (F). The transpose MT of M is
(M T )cr := Mrc (∀r ∈ R, ∀c ∈ C)
(13)
(8)
If M = MT , it is called symmetric. The rank of a matrix is
the dimensionality of the vector spaces spanned by its row or
column vectors. The following properties hold
where i ∈ I′ and j ∈ J′ .
Proof. Writing (13) in block form yields
A11 · · · A1|J′ |
..
..
..
M/D =
.
.
.
′
′
′
A|I |1 · · · A|I ||J |
B1
.
− .. D−1 C1 · · · C|J′ |
B|I′ |
Lemma 1. rank(MT M) = rank(M) ∀M ∈ M|R|×|C| (C).
The claim follows straightforward.
Lemma 2. Let A ∈ M|R|×|R| (C) and B ∈ M|C|×|C| (C) be
non-singular. Then, for any matrix M ∈ M|R|×|C| (C), it holds
that rank(AM) = rank(M) = rank(MB).
D. Graph Theory
A non-singular complex matrix is called unitary if its inverse
equals its conjugate transpose, that is M−1 = (M∗ )T .
T
(15)
A directed graph G = (V, E) consists of a set of vertices
V ( G) := V and a set of directed edges E ( G) := E, where
E ( G) ⊆ {(v, w) ∈ V ( G) × V ( G) | v 6= w}
(16)
Lemma 3. A matrix M ∈ M|R|×|R| (C), for which M = M ,
can be factorized as M = UT DU, where U ∈ M|R|×|R| (C)
is unitary, and D ∈ M|R|×|R| (R) is non-negative diagonal
(Autonne-Takagi factorization, see [27]). If M is non-singular,
then D is positive diagonal.
The sets of outgoing and incoming edges of a vertex v ∈ V ( G)
are denoted by Eout ( G, v) and Ein ( G, v). Formally
A real symmetric matrix is called positive definite (M ≻ 0) or
negative definite (M ≺ 0), respectively, when
A set of internal edges Eint ( G, W) with respect to W ⊆ V ( G)
contains all directed edges that start and end in W. So
M ≻ 0 : xT Mx > 0 ∀x 6= 0
T
M ≺ 0 : x Mx < 0 ∀x 6= 0
Eint ( G, W) := {(u, v) ∈ E ( G) | u, v ∈ W ⊆ V ( G)} (19)
(10)
A cut-set Ecut ( G, W) with respect to W ( V ( G) contains all
directed edges that start in W and end in V ( G) \ W. That is
(
)
u ∈ W ( V ( G) ,
(20)
Ecut ( G, W) := (u, v) ∈ V ( G)
v ∈ V ( G) \ W
Lemma 4. Let M ∈ M|R|×|R| (C) and ℜ{M} ≻ 0. Then, M
is non-singular and ℜ{M−1 } ≻ 0 (see [28]).
Lemma 5. Let M ∈ M|R|×|R| (C) and ℑ{M} ≻ 0. Then, M
is non-singular and ℑ{M−1 } ≺ 0 (see [29]).
Let {Ri | i ∈ I} and Cj | j ∈ J be partitions of R and C,
where I := {1, · · · , |I|} and J := {1, · · · , |J|}. The block
formed by the intersection of the rows Ri (i ∈ I) with the
columns Cj (j ∈ J) is denoted by Mij . Accordingly
M11
· · · M1|J|
..
..
..
M = Mij =
(11)
.
.
.
· · · M|I||J|
(17)
(18)
(9)
If the inequality includes 0, M is called positive semi-definite
(M 0) or negative semi-definite (M 0).
M|I|1
Eout ( G, v) := {e ∈ E ( G) | e = (v, w), w ∈ V ( G)}
Ein ( G, v) := {e ∈ E ( G) | e = (w, v), w ∈ V ( G)}
The connectivity of the graph G is given by its (edge-to-vertex)
incidence matrix A G , whose elements are given by (see [31])
+1 if ei ∈ Eout ( G, v)
−1 if ei ∈ Ein ( G, v)
A G,iv :=
(21)
0 otherwise
A directed graph G is said to be weakly connected if there
exists a connecting path (which need not respect the directivity
of the edges) between any pair of vertices.
Lemma 8. If the directed graph G is weakly connected, then
rank(A G ) = | V ( G) | − 1 (see [32]).
3
m∈N
n∈N
polyphase nodes (i.e., physical and virtual). The vertices are
V := N ∪ G
P
ZΠ,(m,n)
P
The edges fall into two categories, namely polyphase branches
and polyphase shunts. The former connect a pair of polyphase
nodes, the latter a polyphase node and ground. Let LΠ ⊆ N×N
and LT ⊆ N × NT be the polyphase branches associated with
the Π-section and T-section equivalent circuits, respectively.
The set of all polyphase branches is L := LΠ ∪ LT . Similarly,
let TE := NE × G and TT := NT × G be the polyphase shunts
associated with NE and NT , respectively. Thus, the set of all
polyphase shunts is T := TE ∪ TT . The edges are obtained as
P
YΠ,m|(m,n)
YΠ,m|(m,n)
g∈G
(a) Π-section equivalent circuit.
x ∈ NT
m∈N
P
E := L ∪ T
n∈N
P
ZT,(m,x)
(22)
ZT,(x,n)
P
YT,x
g∈G
(b) T-section equivalent circuit.
Fig. 1. Multiport equivalents of electrical components.
E. Power System Analysis
Consider the case of an unbalanced polyphase power system,
which is equipped with a neutral conductor. With regard to the
wiring, the following is assumed:
Hypothesis 1. The reference point of every voltage or current
source is connected with the neutral conductor. Furthermore,
the neutral conductor is grounded using an effective earthing
system, which is capable of establishing a null voltage between
the neutral conductor and the physical ground (see [33], [34]).
Therefore, the phase-to-neutral voltages effectively correspond
to phase-to-ground voltages. Let G := {0} be the ground node,
and P := {1, . . . , |P|} the phases. An array of terminals which
belong together (one for every phase p ∈ P) form a polyphase
node. Define NE := {1, . . . , |NE |} as the set of the physically
existent polyphase nodes of the grid (i.e., where actual voltages
and currents could be measured). The grid consists of electrical
components that link the polyphase nodes with each other and
the ground. As to the grid, the following is presumed:
Hypothesis 2. The grid consists of electrical components that
are linear and passive. Further, electromagnetic coupling only
matters inside of electrical components, but not between them.
Hence, they are represented either by Π-section or T-section
equivalent circuits without mutual coupling as shown in Fig. 1
(see [23]).
Every T-section equivalent circuit comes with an additional
polyphase node (see Fig. 1b). These nodes are purely virtual.
That is, they are part of the model, but do not exist in reality.
Let NT encompass all virtual polyphase nodes originating from
the T-section equivalent circuits. The topology of the grid
model is described by the directed graph G = (V, E), which is
constructed as follows. Define N := NE ∪ NT as the set of all
(23)
The polyphase branches are related to the longitudinal electrical parameters of the multiport equivalents. More precisely,
every polyphase branch ℓi ∈ L is associated with a compound
branch impedance ZP
L,i , which is defined as follows
(
ZP
if ℓi ∈ LΠ
P
Π,ℓi
(24)
ZL,i :=
P
ZT,ℓi if ℓi ∈ LT
Similarly, the polyphase shunts relate to the transversal electrical parameters of the multiport equivalents. The aggregated
P
shunt admittance YΠ,n
at the polyphase node n ∈ N resulting
from the connected Π-section equivalent circuits is given by
X
X
P
P
P
:=
YΠ,n|(n,x)
YΠ,n|(x,n)
YΠ,n
+
(25)
(n,x)∈LΠ
(x,n)∈LΠ
P
The compound shunt admittance YT,n
at the polyphase node
n ∈ N is therefore given by
(
P
YΠ,n
if n ∈ NE
P
YT,n :=
(26)
P
YT,n if n ∈ NT
With respect to the compound electrical parameters of the grid
model, the following assumption is made:
Hypothesis 3. The compound branch impedances ZP
L,k given
in (24) are symmetric, invertible, and passive. That is
P T
ZP
L,i = (ZL,i )
∀ℓi ∈ L :
P
−1
∃YL,i
= (ZP
L,i )
ℜ{ZP
L,i }
(27)
0
In particular, this implies ZP
L,k 6= 0. Conversely, the compound
P
shunt admittances YT,n
given in (26) may be zero. If not, then
they are also symmetric, invertible, and passive. That is
P
P
YT,n
= (YT,n
)T
P
if YT,n
6= 0 :
P
−1
∃ZP
T,n = (YT,n )
P
ℜ{YT,n
}
(28)
0
In practice, the above-stated assumptions are valid for a broad
variety of power system components, like transmission lines,
transformers, and various FACTS devices. Further information
about this subject is given in App. A.
Let Vn,p and In,p be the phasors of the nodal voltage and the
nodal current in phase p ∈ P of the polyphase node n ∈ N.
By definition, the nodal voltage is referenced to the ground
node, and the nodal current flows from the ground node into
the corresponding terminal (see Fig. 2). Analogous quantities
are defined for a polyphase node as a whole
P
:= col (Vn,p ), p ∈ P
VN,n
IP
N,n
:= col (In,p ) , p ∈ P
(29)
(30)
4
P
P
P
IN,|N|
IN,1
n
IN,|N|
IN,1
P
P
ZL,j
P
bN
V
′
...
...
G
B
P
ZL,i
+
′
B
P
P
VN,|N|
VN,1
P
bG
V
G
In,p
P
YT,n
Vn,p
(a)
P
YT,n
P
(b) ∃n ∈ N s.t. YT,n 6= 0.
= 0 ∀n ∈ N.
Fig. 3. Proof of Theorem 1 (surfaces indicate weakly connected graphs).
−
P
In other words, YN
is non-singular if there is at least one nonzero shunt admittance in the circuit, and singular otherwise.
Fig. 2. Definition of the nodal quantities.
The proof is conducted separately for the two cases.
and for the grid as a whole
P
P
:= col (VN,n
), n ∈ N
VN
IP
N
:=
col (IP
N,n )
, n∈N
(31)
P
P
:= diag(YL,i
YL
) , ℓi ∈ L
(33)
YTP
(34)
:=
n∈N
P
P T P P
) YL AB
YN
= (AB
(32)
P
The primitive compound branch admittance matrix YL
and
P
the primitive compound shunt admittance matrix YT are built
in the following way
P
diag(YT,n
),
P
Proof (Case I: YT,n
= 0 ∀n ∈ N, see Fig. 3a.). From (34),
P
it follows that YT = 0. Therefore, (37) simplifies to
(39)
P
YL
According to (33),
is block diagonal. By Hypothesis 3, its
P
blocks YL,i
(ℓi ∈ L) are symmetric and invertible. Therefore,
P
YL
is also symmetric and invertible. According to Lemma 3,
there exists a factorization of the form
P
T P P
YL
= (UP
L ) DL UL
UP
L
(40)
DP
L
Let B := (N, L) be the subgraph of G comprising the
branches only. Define its polyphase incidence matrix AP
B as
where
is unitary and
is positive diagonal. Therefore,
P T P
there exists a matrix EP
such
that DP
L
L = (EL ) EL , which
P
is itself positive diagonal. Hence, YN can be written as
P
AB
: AP
B,ij := AB,ij I|P|
P
T
P
P
P P P
YN
= (MP
L ) ML , ML = EL UL AB
(35)
(41)
where I|P| is the identy matrix of size |P|×|P|. The compound
P
P
nodal admittance matrix YN
, which relates IP
N with VN as
P
Note that the term EP
L UL is non-singular. Therefore, it follows
from Lemmata 1 & 2 that
P P
IP
N = YN VN
P
P
rank(YL
) = rank(MP
L ) = rank(AB )
(36)
is given by (see [23], [32])
P
YN
P T
= (AB )
Recall the definition of AB (35). It is evident that
P P
YL
AB
+
YTP
(37)
P
By definition (31)–(32), VN
and IP
N are block vectors, whose
blocks correspond to the polyphase nodes of the grid. ThereP
P
P
fore, YN
can be written in a block form as YN
= (YN,mn
),
P
P
P
where the block YN,mn relates IN,m and VN,n (m, n ∈ N).
As known from circuit theory
P P
P P
P
Lemma 9.
YN,nx =
YN,yn = YT,n
(see [32]).
x∈N
(42)
P
y∈N
III. P ROPERTIES
The properties of the compound nodal admittance matrix are
developed in two steps. First, it is proven that the compound
shunt admittances determine the rank of the whole matrix in
Sec. III-A. Afterwards, it is shown that the compound branch
impedances determine the rank of its diagonal blocks in III-B).
A. Rank
Theorem 1. Suppose that Hypotheses 1–3 hold. If the branch
graph B = (N, L) is weakly connected, it follows that
P
(|N| − 1)|P| if YT,n
= 0 ∀n ∈ N
P
rank(YN
)=
(38)
|N||P|
otherwise
rank(AP
B ) = |P| · rank(AB )
(43)
Since B = (N, L) is weakly connected, rank(AB ) = |N| − 1
according to Lemma 8. This proves the claim.
P
Proof (Case II: ∃n ∈ N s.t. YT,n
6= 0, see Fig. 3b.). Introduce
′
a virtual ground node G , and build a modified grid by turning
the original ground node into another polyphase node. That is
N′ := N ∪ G
(44)
Accordingly, polyphase shunts with non-zero admittance become polyphase branches in the modified grid. That is
n
o
P
L′ = L ∪ (n, g) ∈ T | YT,n
6= 0
(45)
Hypothesis 3 implicates that the compound branch impedances
′
′
ZP
L ,i (ℓi ∈ L ) are symmetric and invertible. By construction
(
ZP
if ℓi ∈ L
P
L,i
(46)
ZL′ ,i =
P
(YT,n
)−1 if ℓi = (n, g) ∈ L′ \ L
The claimed properties follow from (27) and (28), respectively.
Let B′ := (N′ , L′ ) be the analogon of B for the modified grid.
Obviously, if B is weakly connected, B′ is weakly connected,
too. Clearly, the modified grid satisfies the conditions required
5
for the application of Theorem 1. Observe that the compound
shunt admittances from the polyphase nodes N′ to the virtual
ground node G′ are zero by construction. That is
YTP′ ,n = 0 ∀n ∈ N′
B
(47)
Ecut (B, Mk )
Therefore, the first part of (38), which has already been proven,
can be applied. Accordingly
P
′
′ ) = (|N | − 1)|P| = |N||P|
rank(YN
C
Ck
Ecut (B, N \ Mk )
(48)
Now, define yTP as
P
), n ∈ N
yTP := col(YT,n
(49)
(a) Branch graphs (surfaces indicate weakly connected graphs).
Given that the voltages of the modified grid are referenced to
the virtual ground node, Ohm’s law (36) reads as follows
P
P
P
P
−yT
YN
bN
V
IN
=
(50)
P P
P T
P
P
Y
−(y
)
bG
T,n
T
IG
V
m ∈ Mk
ℓj ∈ L k
n ∈ Mk
eP
Z
L,j
ℓi ∈ X k
x ∈ N \ Mk
P
ZL,i
n∈N
P
′ is given in block form as
That is, YN
P
−yTP
YN
P
′ =
YN
P P
YT,n
−(yTP )T
P
YT,n
P
e T,n
Y
(51)
n∈N
It is known that elementary operations on the rows and column
of a matrix do not change the rank. Thus, one can add the first
P
′ to the last block row / column
|N| block rows / columns of YN
without affecting its rank. Using Lemma 9, one finds
P
YN
−yTP
X
rown (52)
P P row|N|+1 +=
P T
−(yT )
YT,n
n∈N
n∈N
P
−yTP
YN
0
0
P
YN 0
0
0
col|N|+1 +=
X
coln
(53)
n∈N
(54)
It follows straightforward that
P
P
′ ) = |N||P|
rank(YN
) = rank(YN
(55)
g∈G
(b) Grounding of branches in the fictional grid.
Fig. 4. Proof of Theorem 2.
(i.e., Eint ( B, M)) persist, whereas those interconnecting M
and N \ M (i.e., Ecut (B, M) and Ecut (B, N \ M)) become
polyphase shunts in the modified grid. The remaining branches
are described by the graph
C := (M, Eint (B, M))
(58)
In general, C is disconnected. However, since B is connected
as a whole, there exists a partition {Mk | k ∈ K} of M so that
the subgraphs Ck associated with Mk
Ck := (Mk , Eint (B, Mk )), k ∈ K
(59)
are weakly connected, and mutually disconnected (see Fig. 4a).
P
Therefore, YN,M×M
is block diagonal. Namely
which proves the claim.
B. Block Rank
P
P
YN,M×M
= diag(YN,M
), k ∈ K
k ×Mk
Theorem 2. Suppose that Hypotheses 1–3 hold. If the branch
graph B = (N, L) is weakly connected, and all the compound
branch impedances ZP
L,i (ℓi ∈ L) are strictly passive
ℜ{ZP
L,i } ≻ 0 ∀ℓi ∈ L
(56)
P
rank(YN,M×M
) = |M||P| ∀M ( N
(57)
then it follows that
That is, every proper diagonal subblock of the compound nodal
P
admittance matrix YN
has full rank.
P
Proof. As known from circuit theory, YN,M×M
relates IP
N,M
P
P
and VN,M supposing that VN,N\M = 0 (i.e., the polyphase
nodes N\M are short-circuited). This is due to the assumption
that the state of each electrical component is composed solely
P
of ground-referenced nodal voltages. In that sense, YN,M×M
can be regarded as the compound nodal admittance matrix of
a fictional grid obtained by grounding the polyphase nodes
N \ M. Therefore, the polyphase branches internal to M
(60)
P
In turn, YN,M
can be interpreted as the compound nodal
k ×Mk
admittance matrix of a fictional grid, which is constructed by
grounding the polyphase nodes N \ Mk .
eP
Now, it is shown that the compound branch impedances Z
L,i
P
e
and compound shunt admittances YT,n of this fictional grid
satisfy Hypothesis 3. To keep the notation concise, define
Lk := E ( Ck ) = Eint (B, Mk ) ⊂ L
(61)
The grounding process does not modify the compound branch
impedances. Thus, there is a simple one-to-one correspondence
P
eP
Z
L,i = ZL,i , ℓi ∈ Lk
(62)
Xk := Ecut ( B, Mk ) ∪ Ecut (B, N \ Mk )
(63)
eP
Therefore, (27) of Hypothesis 3 obviously applies to Z
L,i , too.
In constrast, the compound shunt admittances are changed, as
some of the polyphase branches become polyphase shunts. The
set Xk containing these polyphase branches is given by
6
From IP
N,Z = 0, it follows straightforward that
Since B is weakly connected, it holds that
Xk 6= ∅ ∀k ∈ K
This implies that, for every Mk , at least one polyphase branch
of the original grid becomes a polyphase shunt in the modified
grid (see Fig. 4b). It follows that
X
X
P
P
P
P
e T,n
Y
= YT,n
+
YL,i
+
YL,i
, n ∈ Mk (65)
ℓi ∈Xk
ℓi =(n,x)
ℓi ∈Xk
ℓi =(x,n)
If the sums are empty, (28) of Hypothesis 3 obviously applies.
Otherwise, it follows from Hypothesis 3, (56), and Lemma 4
P
e T,n
that Y
is symmetric and invertible, and has positive definite
real part. Hence, (28) of Hypothesis 3 applies in this case, too.
As Hypothesis 3 holds, Theorem 1 can be applied. Due to
the fact that Xk 6= ∅ ∀k ∈ K, it follows that ∃m ∈ Mk ∀k ∈ K
P
P
e T,m
for which Y
6= 0 even if YT,n
= 0 ∀n ∈ Mk . Therefore
P
rank(YN,M
) = |Mk ||P| ∀k ∈ K
k ×Mk
(66)
P
P
Since YN,M×M
is block diagonal with blocks YN,M
k ×Mk
X
P
P
rank(YN,M×M ) =
rank(YN,Mk ×Mk )
(67)
k∈K
= |P|
X
|Mk |
(68)
k∈K
= |P||M|
(69)
This proves the claim.
IV. I MPLICATIONS
Using the properties developed in the pevious section, the
findings of [21] can be extended to polyphase power systems.
First, is is proven that Kron reduction is feasible in Sec. IV-A.
Then, it is demonstrated that hybrid parameters matrices exist
for arbitrary partitions of the polyphase nodes in Sec. IV-B.
A. Kron Reduction
Corollary 1. Suppose that Theorem 2 applies. Let Z ( N be
a non-empty subset of N with zero current injection. That is
IP
N,Z
=0
(70)
Then, the voltages at the zero-injection buses Z linearly depend
on those at the other polyphase buses Z∁ := N \ Z. Therefore,
Ohm’s law (36) reduces to
P
P
bP P
bP
IP
N,Z∁ = YZ∁ VN,Z∁ , YZ∁ := YN /YN,Z×Z
(71)
P
bZ
The reduced compound nodal admittance matrix Y
satisfies
∁
P
bZ
rank Y
= |M||P| ∀M ( Z∁
(72)
∁ ,M×M
P
P
bZ
That is, Y
has the same properties as YZ
with
∁ ,M×M
∁ ,M×M
respect to the rank of its proper diagonal subblocks.
Concerning (70), observe that a zero-injection node has zero
current injection in every single of its phases.
Proof. Observe that Z and Z∁ form a partition of N. Therefore,
Ohm’s law (36) can be written in block form as
# "
"
#"
#
P
P
P
YN,Z
Y
IP
V
N,Z∁
N,Z
×Z
N,Z
×Z
∁
∁
∁
∁
=
P
P
P
IP
VN,Z
YN,Z×Z
YN,Z×Z
N,Z
∁
(73)
P
has full rank. Therefore
According to Theorem 2, YN,Z×Z
P
−1
∃ZP
N,Z×Z := (YN,Z×Z )
P
P
P
= −ZP
VN,Z
N,Z×Z YN,Z×Z∁ VN,Z∁
(64)
(74)
(75)
P
P
is a linear function of VN,Z
So, VN,Z
as claimed. Substitute
∁
this formula back into Ohm’s law, and obtain
P
P
P
P
IP
N,Z∁ = YN,Z∁ ×Z∁ VN,Z∁ + YN,Z∁ ×Z VN,Z
"
P
P
YN,Z
VN,Z
∁ ×Z∁
∁
=
P
P
P
P
−YN,Z
Z
N,Z×Z YN,Z×Z∁ VN,Z∁
∁ ×Z
P
P
P
VN,Z
= YN
/YN,Z×Z
∁
=
P
P
bZ
Y
VN,Z
∁
∁
(76)
(77)
(78)
(79)
This proves the first claim (71). According to Lemma 7, the
P
bZ
diagonal block of Y
associated with M ( Z∁ is given by
∁
P
P
P
bZ
Y
(80)
=
Y
/Y
,M×M
N
N,Z×Z
∁
M×M
"
#
P
P
YN,M×M
YN,M×Z
P
=
/ YN,Z×Z
(81)
P
P
YN,Z×M YN,Z×Z
P
P
= YN,(M∪Z)×(M∪Z)
/YN,Z×Z
(82)
From M ( Z∁ = N\Z and Z ( N, it follows that M∪Z ( N.
That is, both M ∪ Z and Z are proper subsets of N. According
P
to Theorem 2, the diagonal subblocks of YN
associated with
M ∪ Z and Z therefore have full rank. By consequence
P
6= 0
(83)
DM∪Z := det YN,(M∪Z)×(M∪Z)
P
DZ := det YN,Z×Z
6= 0
(84)
From Lemma 6, it follows that
P
bZ
= DM∪Z · DZ 6= 0
det Y
,M×M
∁
P
bZ
Therefore, Y
has full rank. Namely
∁ ,M×M
P
bZ
rank Y
= |M||P|
,M×M
∁
(85)
(86)
As M ( Z∁ is abitrary, this proves the second claim (72).
This reduction technique is widely known as Kron reduction.
P
Obviously, eliminating the redundant variables VN,Z
reduces
the size of the model. Thereby, computationally heavy tasks
such as PFS, SE, and VSA can be accelerated without having
to use high-performance computers (see [10]–[14] and [16]).
Interestingly, researchers and practicioners hardly ever verify
whether Kron reduction is truly feasible, although the method
is frequently applied. In fact, not even the inventor (i.e., [15])
considered this issue. Lately, some researchers (i.e., [17], [21])
investigated the problem, but neither of them covered the case
of generic polyphase grids. This gap in the literature is closed
by Corollary 1, which provides a guarantee for the feasibility
of Kron reduction in polyphase grids.
Finally, it is worthwhile to note that Corollary 1 has another
fundamental implication.
Observation 1. Kron reduction preserves the property which
enables its applicability in the first place (i.e., that every proper
diagonal subblock of a compound nodal admittance matrix has
full rank). Therefore, if Z is partitioned as {Zk | k ∈ K}, the
parts Zk can be reduced one after another, and the (partially or
fully) reduced compound nodal admittance matrices obtained
after every step of the reduction also exhibit the rank property.
Performing a reduction partially, or sequentially rather than
“en bloc”, can be beneficial in terms of computational burden,
7
as the Schur complement (13) involves a matrix inversion. This
operation is computationally expensive, and scales poorly with
problem size (even if the inverse is not computed explicitly).
B. Hybrid Parameters
Corollary 2. Suppose that Theorem 2 applies. Let M ( N be
non-empty, that is M and M∁ := N \ M form a partition of N.
Then, there exists a compound hybrid parameters matrix HP
N,
and the grid is described by the hybrid multiport equations
#
" P
#
" P
VN,M∁
IN,M∁
P
(87)
= HN
IP
VP
N,M
N,M
The blocks of HP
N are given as follows
HP
N,M×M
HP
N,M×M∁
HP
N,M∁ ×M
HP
N,M∁ ×M∁
P
= (YN,M×M
)−1
=
=
=
P
P
)−1 YN,M×M
−(YN,M×M
∁
P
−1
P
YN,M
(Y
)
N,M×M
∁ ×M
P
P
YN
/YN,M×M
(88)
(89)
(90)
(91)
Proof. Write Ohm’s law (36) for IP
N,M
P
P
P
P
IP
N,M = YN,M×M∁ VN,M∁ + YN,M×M VN,M
(92)
P
According to Theorem 2, YN,M×M
has full rank, and is hence
P
P
invertible. Define ZN,M×M := (YN,M×M
)−1 , and solve the
P
aforementioned equation for VN,M . This yields
P
P
P
P
(93)
= ZP
VN,M
N,M×M IN,M − YN,M×M∁ VN,M∁
P
P
P
= HP
N,M×M IN,M + HN,M×M∁ VN,M∁
(94)
as claimed in (88)–(89). Write Ohm’s law (36) for IP
N,M∁ , and
P
substitute the above formula for VN,M
. This yields
P
P
P
P
IP
N,M∁ = YN,M∁ ×M∁ VN,M∁ − YN,M∁ ×M VN,M
P
IP
ZP
YN,M
∁ ×M N,M×M N,M
P
P
=
+YN,M∁ ×M∁ VN,M∁
P
P
P
−YN,M
ZP
YN,M×M
VN,M
∁ ×M N,M×M
∁
∁
P
P
P
= HP
N,M∁ ×M IN,M + HM∁ ×M∁ VN,M∁
(95)
(96)
(97)
as claimed in (90)–(91).
Various works have studied hybrid descriptions of multiport
circuits, but they only treat the single-phase case. Most of these
works exhibit further theoretical shortcomings or limitations.
Some authors plainly describe how the hybrid representation is
obtained given that it exists at all (e.g., [35]–[37]). Others do
provide criteria for the existence of the hybrid representation,
but only for some (i.e., perhaps one sole) partition of the nodes
(e.g., [18]–[20]). Moreover, the existing works apply to special
cases of polyphase grids at best, such as balanced triphase grids
(e.g., [21]). In contrast, Corollary 2 guarantees the existence of
hybrid parameters for arbitrary numbers of phases and arbitrary
partitions of the nodes.
Finally, it is worth noting that compound hybrid parameters
matrices also exist for Kron-reduced grids. Namely
Observation 2. The existence of compound hybrid parameters
matrices is based on the same rank property as the feasibility
of Kron reduction. Since Kron reduction preserves this property
(recall Observation 1), compound hybrid parameters matrices
can be obtained from unreduced, partially reduced, and fully
reduced compound nodal admittance matrices. In this regard,
P
P
bZ
YN
can be replaced by Y
in Corollary 2.
∁
V. C ONCLUSIONS
This paper examined the properties of the compound nodal
admittance matrix of polyphase power systems, and illustrated
their implications for power system analysis. Using the concept
of compound electrical parameters, and exploiting the physical
characteristics of electrical components, rank properties for the
compound nodal admittance matrix and its diagonal subblocks
were deduced. Notably, it was proven that the diagonal blocks
have full rank if the grid is connected and lossy. Based on these
findings, it was shown that the feasibility of Kron reduction and
the existence of hybrid parameters are guaranteed in practice.
Thus, this paper provided a rigorous theoretical foundation for
the analysis of generic polyphase power systems.
A PPENDIX A
P OWER S YSTEM C OMPONENTS
A. Transmission Lines
Consider a multiconductor transmission line with |P| phase
conductors plus one neutral conductor. Let v(z, t) and i(z, t)
describe the phase-to-neutral voltages and the phase conductor
currents, respectively. That is
v(z, t) = col vp (z, t) , p ∈ P
(98)
i(z, t) = col ip (z, t) , p ∈ P
(99)
where z is the position along the line. If (i) the electromagnetic
parameters of the line are state-independent, (ii) the conductors
are parallel, and the perpendicular distance between any two of
them is much shorter than the wavelength, (iii) the conductors
have finite conductance, (iv) the electromagnetic field outside
of the conductors produced by the charges and currents inside
of them is purely transversal, and (v) the sum of the conductor
currents is zero, then Maxwell’s equations simplify to the socalled telegrapher’s equations (see [38])
∂
′
′ ∂
i(z, t)
(100)
v(z, t) = − R + L
∂z
∂t
∂
∂
v(z, t)
(101)
i(z, t) = − G′ + C′
∂z
∂t
R′ and L′ are the resistance and inductance per unit length of
the conductors, G′ & C′ are the conductance and capacitance
per unit length of the dielectric. These matrices are symmetric.
Consider a segment of infinitesimal length ∆z at position z.
Let Ee (z, t) and Em (z, t) be the energy stored in the electric
and magnetic field, respectively. They are given by (see [38])
∆z
(v(z, t))T C′ v(z, t)
2
∆z
(i(z, t))T L′ i(z, t)
Em (z, t) =
2
Ee (z, t) =
(102)
(103)
As Ee (z, t) > 0 ∀v(z, t) 6= 0 and Em (z, t) > 0 ∀i(z, t) 6= 0,
it follows straightforward that C′ ≻ 0 and L′ ≻ 0. Similarly,
let Pd (z, t) and Pc (z, t) be the power dissipated in the conductors and the ambient dielectric. They are given by (see [38])
Pd (z, t) = ∆z · (v(z, t))T G′ v(z, t)
T
′
Pc (z, t) = ∆z · (i(z, t)) R i(z, t)
(104)
(105)
Since real systems are lossy, Pd (z, t) > 0 ∀v(z, t) 6= 0 and
Pc (z, t) > 0 ∀i(z, t) 6= 0. Therefore, G′ ≻ 0 and R′ ≻ 0. If a
transmission line is electrically short (i.e., its length |zm −zn | is
8
significantly shorter than the wavelength), it can be represented
by a Π-section equivalent circuit with
′
G + jωC′
P
(106)
YΠ,m|(m,n)
= |zm − zn |
2
P
P
YΠ,n|(m,n)
= YΠ,m|(m,n)
ZP
Π,(m,n)
′
′
′
m
n
P
ZΛ,(m,n)
(107)
′
′
= |zm − zn |(R + jωL )
′
compensator
(108)
P
B. Transformers
When analyzing power systems operating at rated frequency,
transformers are represented by a T-section equivalent circuits,
which correspond one-to-one to the composition of the devices
(see [39]). Consider a polyphase transformer connecting two
polyphase nodes m, n ∈ N. Let m be its primary side, and n
its secondary side. In this case, the parameters of the T-section
equivalent circuit are given as
P
YΛ,m|(m,n)
P
YΠ,m|(m,n)
,
As G , C , R , and L are real positive definite,
P
YΠ,n|(m,n)
, and ZP
Π,(m,n) are symmetric with positive definite
real part. So, according to Lemmata 4 & 5, they are invertible.
This is in accordance with Hypothesis 3.
P
ZΓ,(m,n)
YΛ,n|(m,n)
g
(a) Series compensation.
m
n
P
ZΛ,(m,n)
P
YΛ,m|(m,n)
compensator
P
YΓ,n
P
YΛ,n|(m,n)
ZP
T,(m,x) = Rw,1 + jωLℓ,1
(109)
ZP
T,(x,n)
P
YT,x
= Rw,2 + jωLℓ,2
(110)
g
= Gh + jωBm
(111)
(b) Shunt compensation (for simplicity, only one compensator is shown).
The resistance matrices Rw,1 and Rw,2 represent the winding
resistances, and the inductance matrices Lℓ,1 and Lℓ,2 the
leakage inductances of the coils on the primary and secondary
side, respectively. The former are positive diagonal, the latter
are positive definite. The conductance matrix Gh represents
to the hysteresis losses, and the susceptance matrix Bm to the
magnetization of the transformer’s core. The former is positive
diagonal, the latter is positive definite. Therefore, according to
P
P
Lemmata 4 & 5, ZP
T,(m,x) , ZT,(x,n) , and YT,x are invertible.
This applies irrespective of whether the polyphase transformer
is implemented using a single core with multiple legs or several
separate cores. In the latter case, Lℓ,1 , Lℓ,2 , and Bm simply
become diagonal, since the separate cores are magnetically
decoupled. In either case, the compound electrical parameters
satisfy the assumptions of Hypothesis 3.
C. FACTS Devices
There exist three families of FACTS devices, namely series
compensators, shunt compensators, and combined series-andshunt compensators (see [40]). Here, the first two cases are
P
P
discussed. Let ZP
Λ,(m,n) , YΛ,m|(m,n) , and YΛ,n|(m,n) be the
compound electrical parameters describing a transmission line
without compensation (see App. A-A). If a series compensator
is installed, the compound branch impedance of the respective
transmission line is altered. In the Π-section equivalent circuit,
this is reflected by adding the compound impedance ZΓ,(m,n)
of the compensator to the compound branch impedance of the
transmission line (see Fig. 5a). Namely
P
ZP
Π,(m,n) = ZΛ,(m,n) + ZΓ,(m,n)
(112)
Conversely, if shunt compensators are installed, the compound
shunt admittances are altered. This is accounted for by adding
the compound admittances YΓ,m and YΓ,n of the compensators to the compound shunt admittances of the transmission
line (see Fig. 5b). That is
P
P
YΠ,m|(m,n)
= YΛ,m|(m,n)
+ YΓ,m
P
YΠ,n|(m,n)
=
P
YΛ,n|(m,n)
+ YΓ,n
(113)
(114)
Fig. 5. Incorporation of FACTS devices into the transmission line model.
Typically, this kind of compensators are implemented as banks
of capacitors or reactors, which can be stepwise connected and
disconnected. Such devices are symmetrical with respect to the
phases, and lossy. By consequence
ZΓ,(m,n) = ZTΓ,(m,n) , ℜ{ZΓ,(m,n) } ≻ 0
YΓ,m/n =
T
YΓ,m/n
,
ℜ{YΓ,m/n } ≻ 0
(115)
(116)
Recall that the compound electrical parameters of transmission
lines comply with Hypothesis 3 (see App. A-A). Accordingly,
the results of (112) & (113)–(114) are symmetric, have positive
definite real part, and are thus invertible (see Lemma 4). Hence,
it can be concluded that transmission lines which are equipped
with series or shunt compensators satisfy Hypothesis 3.
R EFERENCES
[1] B. Stott, “Review of Load-Flow Calculation Methods,” Proc. IEEE,
vol. 62, no. 7, pp. 916–929, Jul. 1974.
[2] F. Milano, Power System Modelling and Scripting. Berlin, BER, DE:
Springer Science & Business Media, 2010.
[3] A. J. Monticelli, A. Garcia, and O. Saavedra, “Fast Decoupled Load
Flow: Hypothesis, Derivations, and Testing,” IEEE Trans. Power Syst.,
vol. 5, no. 4, pp. 1425–1431, Nov. 1990.
[4] F. F. Wu, “Power System State Estimation: A Survey,” Int. J. Elect.
Power Energy Syst., vol. 12, no. 2, pp. 80–87, Apr. 1990.
[5] A. J. Monticelli, State Estimation in Electric Power Systems: A Generalized Approach. Berlin, BER, DE: Springer Science & Business Media,
1999.
[6] A. Abur and A. Gómez Expósito, Power System State Estimation: Theory
and Implementation. Boca Ranton, FL, USA: CRC Press, 2004.
[7] P. Kundur, Power System Stability and Control, N. J. Balu and M. G.
Lauby, Eds. New York City, NY, USA: McGraw-Hill Education, 1994.
[8] E. W. Kimbark, Power System Stability. Hoboken, NJ, USA: John
Wiley & Sons, 1995.
[9] T. Van Cutsem and C. D. Vournas, Voltage Stability of Electric Power
Systems. Berlin, BER, DE: Springer Science & Business Media, 1998.
[10] V. Ajjarapu and C. Christy, “The Continuation Power Flow: A Tool for
Steady State Voltage Stability Analysis,” IEEE Trans. Power Syst., vol. 7,
no. 1, pp. 416–423, Feb. 1992.
[11] J. N. Mendel, “Computational Requirements for a Discrete Kalman
Filter,” IEEE Trans. Autom. Control, vol. 16, no. 6, pp. 748–758, Dec.
1971.
9
[12] A. M. Kettner and M. Paolone, “Sequential Discrete Kalman Filter for
Real-Time State Estimation in Power Distribution Systems: Theory and
Implementation,” IEEE Trans. Instrum. Meas., vol. 66, no. 9, pp. 2358–
2370, Sep. 2017.
[13] P.-A. Löf, G. Andersson, and D. J. Hill, “Voltage Stability Indices for
Stressed Power Systems,” IEEE Trans. Power Syst., vol. 8, no. 1, pp.
326–335, Feb. 1993.
[14] G. Irisarri, X. Wang, J. Tong, and S. Mokhtari, “Maximum Loadability
of Power Systems using Interior Point Nonlinear Optimization Method,”
IEEE Trans. Power Syst., vol. 12, no. 1, pp. 162–172, Feb. 1997.
[15] G. Kron, Tensors for Circuits, 2nd ed. Mineola, NY, USA: Dover
Publications, 1959.
[16] S. K. Khaitan and A. Gupta, Eds., High-Performance Computing in
Power and Energy Systems. Berlin, BER, DE: Springer Science &
Business Media, 2013.
[17] F. Dörfler and F. Bullo, “Kron Reduction of Graphs with Applications to
Electrical Networks,” IEEE Trans. Circuits Syst. I: Reg. Papers, vol. 60,
no. 1, pp. 150–163, Jan. 2013.
[18] H. C. So, “On the Hybrid Description of a Linear n-Port Resulting from
the Extraction of Arbitrarily Specified Elements,” IEEE Trans. Circuit
Theory, vol. 12, no. 3, pp. 381–387, Sep. 1965.
[19] J. Zuidweg, “Every Passive Time-Invariant Linear n-Port has at least
one H Matrix,” IEEE Trans. Circuit Theory, vol. 12, no. 1, pp. 131–
132, Mar. 1965.
[20] B. Anderson, R. Newcomb, and J. Zuidweg, “On the Existence of H
Matrices,” IEEE Trans. Circuit Theory, vol. 13, no. 1, pp. 109–110, Mar.
1966.
[21] A. M. Kettner and M. Paolone, “On the Properties of the Power Systems
Nodal Admittance Matrix,” IEEE Trans. Power Syst., vol. 33, no. 1, pp.
1130–1131, Jan. 2018.
[22] C. L. Fortescue, “Method of Symmetrical Co-Ordinates Applied to the
Solution of Polyphase Networks,” Trans. AIEE, vol. 37, no. 2, pp. 1027–
1140, Jun. 1918.
[23] J. Arrillaga and C. P. Arnold, Computer Analysis of Power Systems.
Hoboken, NJ, USA: John Wiley & Sons, 1990.
[24] S. N. Tiwari and L. P. Singh, “Mathematical Modelling and Analysis of
Multiphase Systems,” IEEE Trans. Power App. Syst., no. 6, pp. 1784–
1793, Jun. 1982.
[25] ——, “Six-Phase (Multiphase) Power Transmission Systems: A Generalized Investigation of the Load Flow Problem,” Elect. Power Syst. Res.,
vol. 5, no. 4, pp. 285–297, 1982.
[26] A. S. Meliopoulos and F. Zhang, “Multiphase Power Flow and State
Estimation for Power Distribution Systems,” IEEE Trans. Power Syst.,
vol. 11, no. 2, pp. 939–946, May 1996.
[27] R. A. Horn and C. R. Johnson, Matrix Analysis, 2nd ed. Cambridge,
CAM, UK: Cambridge University Press, 2012.
[28] D. London, “A Note on Matrices with Positive Definite Real Part,” Proc.
AMS, vol. 82, pp. 322–324, Jul. 1981.
[29] K. Fan, “Generalized Cayley Transforms and Strictly Dissipative Matrices,” Lin. Alg. Applicat., vol. 5, no. 2, pp. 155–172, 1972.
[30] F. Zhang, The Schur Complement and its Applications. Berlin, BER,
DE: Springer Science & Business Media, 2006.
[31] J. L. Gross, J. Yellen, and P. Zhang, Handbook of Graph Theory, 2nd ed.
Boca Ranton, FLA, USA: CRC Press, 2013.
[32] C. Desoer and E. Kuh, Basic Circuit Theory. New York City, NY, USA:
McGraw-Hill Education, 1969.
[33] J. C. Das and R. H. Osman, “Grounding of AC and DC Low-Voltage
and Medium-Voltage Drive Systems,” IEEE Trans. Ind. Appl., vol. 34,
no. 1, pp. 205–216, Jan. 1998.
[34] T.-H. Chen and W.-C. Yang, “Analysis of Multi-Grounded Four-Wire
Distribution Systems Considering the Neutral Grounding,” IEEE Trans.
Power Del., vol. 16, no. 4, pp. 710–717, Oct. 2001.
[35] I. N. Hajj, “Computation of Hybrid Equations of Linear Multiports,”
IEEE Trans. Circuits Syst., vol. 24, no. 11, pp. 655–656, Nov. 1977.
[36] S. Sun and L. Qian, “Formulation of Generalised State Equations and
Multiport Equations: A Novel Approach,” IEE Proc. G - Circuits Devices
Syst., vol. 137, no. 1, pp. 49–52, Feb. 1990.
[37] J. Augusto and C. Almeida, “Use of Modified Nodal Analysis to Write
Multiport Hybrid Matrix,” IET Electron. Lett., vol. 27, no. 19, pp. 1750–
1752, Sep. 1991.
[38] C. R. Paul, Analysis of Multiconductor Transmission Lines. Hoboken,
NJ, USA: John Wiley & Sons, 2008.
[39] M. J. Heathcote, The Johnson & Phillips Transformer Book, 12th ed.
Oxford, OXF, UK: Newnes, 2011.
[40] S. Gerbex, R. Cherkaoui, and A. J. Germond, “Optimal Location of
Multi-Type FACTS Devices in a Power System by Means of Genetic
Algorithms,” IEEE Trans. Power Syst., vol. 16, no. 3, pp. 537–544, Aug.
2001.
| 3cs.SY
|
Entropy 2017, , 1-; doi: To be determined
OPEN ACCESS
entropy
ISSN 1099-4300
www.mdpi.com/journal/entropy
Article
Computing Entropies With Nested Sampling
Brendon J. Brewer
arXiv:1707.03543v2 [stat.CO] 16 Aug 2017
Department of Statistics, The University of Auckland, Private Bag 92019, Auckland 1142, New Zealand
Accepted 08/2017
Abstract: The Shannon entropy, and related quantities such as mutual information, can
be used to quantify uncertainty and relevance. However, in practice, it can be difficult to
compute these quantities for arbitrary probability distributions, particularly if the probability
mass functions or densities cannot be evaluated. This paper introduces a computational
approach, based on Nested Sampling, to evaluate entropies of probability distributions that
can only be sampled. I demonstrate the method on three examples: a simple gaussian
example where the key quantities are available analytically; (ii) an experimental design
example about scheduling observations in order to measure the period of an oscillating
signal; and (iii) predicting the future from the past in a heavy-tailed scenario.
Keywords: information theory; entropy; mutual information; monte carlo; nested sampling;
bayesian inference
1. Introduction
If an unknown quantity x has a discrete probability distribution p(x), the Shannon entropy [1,2] is
defined as
X
H(x) = −
p(x) log p(x)
(1)
x
where the sum is over all of the possible values of x under consideration. The entropy quantifies the
degree to which the issue “what is the value of x, precisely?” remains unresolved [3]. More generally, the
‘Kullback-Leibler divergence’ quantifies the degree to which a probability distribution p(x) departs from
a ‘base’ or ‘prior’ probability distribution q(x):
X
p(x)
DKL p || q =
p(x) log
(2)
q(x)
x
Entropy 2017,
2
[4,5]. If q(x) is taken to be uniform, this is equivalent (up to an additive constant and a change of sign) to
the Shannon entropy H(x). With the negative sign reinstated, Kullback-Leibler divergence is sometimes
called relative entropy, and is more fundamental than the Shannon entropy [4]. However, this paper
focuses on the Shannon entropy for simplicity, and because the resulting algorithm calculates the expected
value of a log-probability.
The Shannon entropy can be interpreted straightforwardly as a measure of uncertainty. For example, if
there is only one possible value which has probability one, H(x) is zero. Conventionally, 0 × log 0 is
defined to be equal to limx→0+ (x log x), which is zero; i.e., ‘possibilities’ with probability zero do not
contribute to the sum. If there are N possibilities with equal probabilities 1/N each, then the entropy
is H(x) = log(N ). If the space of possible x-values is continuous so that p(x) is a probability density
function, the differential entropy
Z
H(x) = − p(x) log p(x) dx
(3)
quantifies uncertainty by generalising the log-volume of the plausible region, where volume is defined
with respect to dx. Crucially, in continuous spaces, the differential entropy is not invariant under changes
of coordinates, so any value given for a differential entropy must be understood with respect to the
coordinate system that was used. Even in discrete cases, using the Shannon entropy assumes that each
possibility contributes equally when counting possibilities — if this is inappropriate, the relative entropy
should be used.
Entropies, relative entropies, and related quantities, tend to be analytically available only for a few
families of probability distributions. On the numerical side, if log p(x) can be evaluated for any x, then
simple Monte Carlo will suffice for approximating H(x). On the other hand, if p(x) can only be sampled
but not evaluated (for example, if p(x) is a marginal or conditional distribution, or the distribution of a
quantity derived from x), then this will not work. Kernel density estimation or the like [e.g. 6] may be
effective in this case, but is unlikely to generalise well to high dimensions.
This paper introduces a computational approach to evaluating the Shannon entropy H(x) of probability
distributions that can only be sampled using Markov Chain Monte Carlo (MCMC). The key idea is that
Nested Sampling can create an unbiased estimate of a log-probability, and entropies are averages of
log-probabilities.
1.1. Notation and conventions
Throughout this paper, I use the compact ‘overloaded’ notation for probability distributions favoured
by many Bayesian writers [7,8], writing p(x) for either a probability mass function or a probability
density function, instead of the more cumbersome P (X = x) or fX (x). In the compact notation, there
is no distinction between the ‘random variable’ itself (X) and a similarly-named dummy variable (x).
Probability distributions are implicitly conditional on some prior information, which is omitted from the
notation unless necessary. All logarithms are written log and any base can be used, unless otherwise
specified (for example by writing ln or log10 ). Any numerical values given for the value of specific
entropies are in nats, i.e., the natural logarithm was used.
Even though the entropy is written as H(x), it is imperative that we remember it is not a property of
the value of x itself, but a property of the probability distribution used to describe a state of ignorance
Entropy 2017,
3
about x. Throughout this paper, H(x) is used as notation for both Shannon entropies (in discrete cases)
and differential entropies (in continuous cases). Which one it is should be clear from the context of the
problem at hand. I sometimes write general formulae in terms of sums (i.e., as they would appear in a
discrete problem), and sometimes as integrals, as they would appear in a continuous problem.
2. Entropies in Bayesian inference
Bayesian inference is the use of probability theory to describe uncertainty, often about unknown
quantities (‘parameters’) θ. Some data d, initially unknown but thought to be relevant to θ, is obtained.
The prior information leads the user to specify a prior distribution p(θ) for the unknown parameters,
along with a conditional distribution p(d|θ) describing prior knowledge of how the data is related to the
parameters (i.e., if the parameters were known, what data would likely be observed?). By the product
rule, this yields a joint prior
p(θ, d) = p(θ)p(d|θ)
(4)
which is the starting point for Bayesian inference [5,9]. In practical applications, the following operations
are usually feasible and have a low computational cost:
1. Samples can be generated from the prior p(θ);
2. Simulated datasets can be generated from p(d|θ) for any given value of θ;
3. The likelihood, p(d|θ), can be evaluated cheaply for any d and θ. Usually it is the log-likelihood
that is actually implemented, for numerical reasons.
Throughout this paper I assume these operations are available and inexpensive.
2.1. The relevance of data
The entropy of the prior p(θ) describes the degree to which the question “what is the value of θ,
precisely?” remains unanswered, while the entropy of the joint prior p(θ, d) describes the degree to
which the question “what is the value of the pair (θ, d), precisely?” remains unanswered. The degree to
which the question “what is the value of θ?” would remain unresolved if d were resolved is given by the
conditional entropy
X
X
p(d)
p(θ|d) log p(θ|d)
(5)
H(θ|d) = −
d
θ
which is the expected value of the entropy of the posterior, averaged over all possible datasets which
might be observed.
One might wish to compute H(θ|d), perhaps to compare it to H(θ) and quantify how much might be
learned about θ. This would be difficult because the expression for the posterior distribution
p(θ|d) =
p(θ)p(d|θ)
p(d)
(6)
Entropy 2017,
4
contains the marginal likelihood:
p(d) =
X
p(θ)p(d|θ)
(7)
θ
also known as the ‘evidence’, which tends to be computable but costly, especially when the summation is
replaced by an integration over a high-dimensional parameter space.
It is important to distinguish between H(θ|d) and the entropy of θ given a particular value of d, which
might be written H(θ|d = dobs ) (‘obs’ for observed). The former measures the degree to which one
question answers another ex ante, and is a function of two questions. The latter measures the degree to
which a question remains unresolved after conditioning on a specific statement (ex post), and is a function
of a question and a statement.
2.2. Mutual information
The mutual information is another way of describing the relevance of the data to the parameters. Its
definition, and relation to other quantities, is
XX
p(θ, d)
I(θ; d) =
p(θ, d) log
(8)
p(θ)p(d)
θ
d
= H(θ) + H(d) − H(θ, d)
= H(d) − H(d|θ)
= H(θ) − H(θ|d).
(9)
(10)
(11)
The mutual information can also be written as the expected value (with respect to the prior over datasets
p(d)) of the Kullback-Leibler divergence from prior to posterior:
X
I(θ; d) =
p(d)DKL p(θ|d) || p(θ) .
(12)
d
In terms of the prior, likelihood, and evidence, it is
XX
I(θ; d) =
p(θ)p(d|θ) [log p(d|θ) − log p(d)] ,
θ
(13)
d
i.e., the mutual information is the prior expected value of the log likelihood minus the log evidence.
As with the conditional entropy, the computational difficulty appears in the form of the log evidence,
P
log p(d) = log ( θ p(θ)p(d|θ)), which must be evaluated or estimated for many possible datasets.
For experimental design purposes, maximising the expected amount of information obtained from
the data is a sensible goal. Formally, either maximising I(θ; d) or minimising H(θ|d) will produce the
same result because the prior p(θ) does not vary with the experimental design. Reference priors [10] also
maximise I(θ; d) but vary the prior p(θ) in the maximisation process while keeping the experimental
design fixed.
Entropy 2017,
5
3. Nested Sampling
Nested Sampling (NS), introduced by Skilling [11], is an algorithm whose aim is to calculate the
evidence
Z
p(θ) = p(θ)p(d|θ) dθ,
(14)
or, in the simplified notation common when discussing computational matters,
Z
Z = π(θ)L(θ) dθ,
(15)
where θ is the unknown parameter(s), π is the prior distribution, and L is the likelihood function. The
original NS algorithm, and variations within the same family [12–14], have become popular tools in
Bayesian data analysis [15–17] and have also found use in statistical physics [18–20], which was another
of its aims. NS also estimates the Kullback-Leibler divergence from the prior π(θ) to the posterior
π(x)L(x)/Z, which Skilling calls the information. This is a measure of how much was learned about θ
from the specific dataset, and is also useful for defining a termination criterion for NS.
While Z is a simple expectation with respect to π, the distribution of L-values implied by π tends to
be very heavy-tailed, which is why simple Monte Carlo does not work. Equivalently, the integral over θ
is dominated by very small regions where L(θ) is high.
To overcome this, NS evolves a population of N particles in the parameter space towards higher
likelihood regions. The particles are initially initialised from the prior π(θ), and the particle with the
lowest likelihood, L∗1 , is found and its details are recorded as output. This worst particle is then discarded
and replaced by a new particle drawn from the prior π but subject to the constraint that its likelihood
must be above L∗1 . This is usually achieved by cloning a surviving particle and evolving it with MCMC
to explore the prior distribution but with a constraint on likelihood value, L(θ) > L∗1 . This ‘constrained
prior’ distribution has probability density proportional to
π(θ)1 (L(θ) > L∗1 )
(16)
where 1() is an indicator function which returns one if the argument is true and zero if it is false.
This process is repeated, resulting in an increasing sequence of likelihoods
L∗1 , L∗2 , L∗3 , ...
Defining X(`) as the amount of prior mass with likelihood greater than some threshold `,
Z
X(`) = π(θ)1 (L(θ) > `) dθ,
(17)
(18)
each discarded point in the sequence can be assigned an X-value, which transforms Z to a one-dimensional
integral. Skilling [11]’s key insight was that the X-values of the discarded points are approximately known
from the nature of the algorithm. Specifically, each iteration reduces the prior mass by approximately a
factor e−1/N . More accurately, the conditional distribution of Xi given Xi−1 (and the definition of the
algorithm) is obtainable by letting ti ∼ Beta(N, 1) and setting Xi+1 := ti Xi−1 (where X1 = t1 ). The ti
quantities are the proportion of the remaining prior mass that is retained at each iteration, after imposing
the constraint of the latest L∗ value.
Entropy 2017,
6
3.1. The sequence of X values
The sequence of X values is defined by
X1 = t1
(19)
X2 = t2 X1
(20)
X3 = t3 X2
(21)
and so on, i.e. Xi can be written as
Xi =
i
Y
tk .
(22)
k=1
Taking the log and multiplying both sides by -1, we get
− ln Xi =
i
X
(− ln tk ) .
(23)
k=1
Consider ti , the fraction of the remaining prior mass that is retained in each NS iteration. By a change
of variables, the distribution for these compression factors ti ∼Beta(N, 1) corresponds to an exponential
distribution for − ln(ti ):
− ln ti ∼ Exponential(N ).
(24)
From Equations 23 and 24, we see that the sequence of − ln X values produced by a Nested Sampling
run can be considered as a Poisson process with rate equal to N , the number of particles. This is why
separate NS runs can be simply merged [11,21]. Walter [22] showed how this view of the NS sequence of
points can be used to construct a version of NS that produces unbiased (in the frequentist sense) estimates
of the evidence. However, it can also be used to construct an unbiased estimator of a log-probability,
which is more relevant to information theoretic quantities discussed in the present paper.
Consider a particular likelihood value ` whose corresponding X-value we would like to know. Since
the − ln Xi values have the same distribution as the arrival times of a Poisson process with rate N , the
probability distribution for the number of points in an interval of length w is Poisson with expected value
N w. Therefore the expected number of NS discarded points with likelihood below ` is −N ln X(`):
n(L(θ i ) ≤ `) = −N ln X(`).
(25)
We can therefore take n(L(θ i ) ≤ `)/N , the number of points in the sequence with likelihood below `
divided by the number of NS particles, as an unbiased estimator of − ln X(`). The standard deviation of
p
the estimator is − ln X(`)/N .
4. The algorithm
The above insight, that the number of NS discarded points with likelihood below ` has expected value
−N ln X(`), is the basis of the algorithm. If we wish to measure a log-probability − ln X(`), we can use
Entropy 2017,
7
NS to do it. π(θ) and L(θ) need not be the prior and likelihood respectively, but can be any probability
distribution (over any space) and any function whose expected value is needed for a particular application.
See Algorithm 1 for a step-by-step description of the algorithm. The algorithm is written in terms of
an unknown quantity θ whose entropy is required. In applications, θ may be parameters, a subset of the
parameters, data, some function of the data, parameters and data jointly, or whatever.
The idea is to generate a reference particle θ ref from the distribution p(θ) whose entropy is required.
Then, a Nested Sampling run evolves a set of particles {θ}N
i=1 , initially representing p(θ), towards θ ref in
order to measure the log-probability near θ ref (see Figure 1). Nearness is defined using a distance function
d(θ; θ ref ), and the number of NS iterations taken to reduce the distance to below some threshold r (for
‘radius’) provides an unbiased estimate of
(depth) = − log [P (d(θ; θ ref ) < r)]
(26)
which I call the ‘depth’. E.g., if N = 10 Nested Sampling particles are used, and it takes 100 iterations to
reduce the distance to below r, then the depth is 10 nats.
If we actually want the differential entropy
Z
H(θ) = − p(θ) log p(θ) dθ
(27)
we can use the fact that density equals mass divided by volume. Assume r is small, so that
Z
P (d(θ; θ ref ) < r | θ ref ) ≈ p(θ ref )
dθ
(28)
d(θ;θ ref )<r
= p(θ ref ) V
(29)
where V is the volume of the region where d(θ; θ ref ) < r. Then H(θ) = h(depth)ip(θ) + log V (r). It
is important to remember that the definition of volume here is in terms of the coordinates θ, and that a
change of coordinates (or a change of the definition of the distance function) would result in a different
value of the differential entropy. This is equivalent to counting possibilities or volumes with a different
choice of base measure q(θ).
Reducing r towards zero ought to lead to a more accurate result because the probability density at
a point will be more accurately approximated by the probability of a small region containing the point,
divided by the volume of that region. On the other hand, smaller values of r will lead to more Nested
Sampling iterations required, and hence less accuracy in the depth estimates. Thus, there is a an awkward
trade-off involved in choosing the value of r, which will need to be investigated more thoroughly in the
future. However, the algorithm as it currently stands ought to be useful for calculating the entropies of
any low-to-moderate dimensional functions of an underlying high-dimensional space.
If the distance function d(θ ref , θ) is chosen to be Euclidean, the constraint d(θ ref , θ) < r corresponds
to a ball in the space of possible θ ref values. The volume of such a ball, of radius r in n dimensions, is
V =
(πr2 )n/2
.
( n2 )!
(30)
In one dimension, if the threshold r is not small, we might think we are calculating the entropy associated
with the question ‘what is θ, to within a tolerance of ±r?’. This is not quite correct. See Appendix A for
a discussion of this type of question.
Entropy 2017,
8
Algorithm 1 The algorithm which estimates the expected value of the depth:
R
R
− p(θ ref ) p(θ) log [P (d(θ; θ ref ) < r | θ ref )] dθ dθ ref ,
that is, minus the expected value of the log-probability of a small region near θ ref , which can be converted
to an estimate of an entropy or differential entropy. The part highlighted in green is standard Nested
Sampling with quasi-prior p(θ) and quasi-likelihood given by minus a distance function d(θ; θ ref ). The
final result, hd
final , is an estimate of the expected depth.
Set the numerical parameters:
N ∈ {1, 2, 3, ...}
. the number of Nested Sampling particles to use
r≥0
. the tolerance
b ← {}
h
. Initialise an empty list of results
while more iterations desired do
k←0
. Initialise counter
Generate θ ref from p(θ)
. Generate a reference point
N
Generate {θ i }i=1 from p(θ)
. Generate initial NS particles
Calculate di ← d(θ i ; θ ref ) for all i . Calculate distance of each particle from the reference point
i∗ ← argmin(λi → di )
. Find the worst particle (greatest distance)
dmax ← di∗
. Find the greatest distance
while dmax > r do
Replace θ i∗ with θ new from p (θ | d(θ) < dmax )
. Replace worst particle
Calculate di∗ ← d(θ new ; θ ref )
. Calculate distance of new particle from reference point
i∗ ← argmin(λi → di )
. Find the worst particle (greatest distance)
dmax ← di∗
. Find the greatest distance
k ←k+1
. Increment counter k
b ← append h,
b k/N
h
. Append latest estimate to results
P
1
b
hd
h
. Average results
final ← num_iterations
Entropy 2017,
9
10.0
7.5
5.0
y
2.5
0.0
−2.5
−5.0
−7.5
−10.0
−10
−5
0
5
10
x
Figure 1. To evaluate the log-probability (or density) of the blue probability distribution near
the red point, Nested Sampling can be used, with the blue distribution playing the role of the
“prior” in NS, and the Euclidean distance from the red point (illustrated with red contours)
playing the role of the negative log-likelihood. Averaging over selections of the red point
gives an estimate of the entropy of the blue distribution.
In the following sections I demonstrate the algorithm on three example problems. For each of the
examples, I used 1000 repetitions of NS (i.e., 1000 reference particles), and used 10 NS particles and
10,000 MCMC steps per NS iteration.
5. Example 1: Entropy of a prior for the data
This example demonstrates how the algorithm may be used to determine the entropy for a quantity
whose distribution can only be sampled, in this case, a prior distribution over datasets.
Consider the basic statistics problem of inferring a quantity µ from 100 observations x = {x1 , ..., x100 }
whose probability distribution (conditional on µ) is
p(xi |µ) ∼ Normal(µ, 1).
(31)
I assumed a Normal(0, 102 ) prior for µ. Before getting the data, the prior information is described by the
joint distribution
p(µ, x) = p(µ)p(x|µ).
(32)
The prior over datasets is given by the marginal distribution
Z
p(x) = p(µ, x) dµ.
(33)
Entropy 2017,
10
This is the distribution whose differential entropy I seek in this section. Its true value is available
analytically in this case, by noting that the posterior is available in closed form for any dataset:
2 !
100
1 X
10
.
(34)
µ|x ∼ Normal
xi , √
100.01 i=1
10001
which enables the calculation of the mutual information and hence H(x). The true value of H(x) is
146.499 nats. However, in most situations, the entropy of a marginal distribution such as this is not
available in closed form.
To use the algorithm, a distance function d(; ) must be defined, to quantify how far the NS particles are
from the reference particle. To calculate the entropy of the data, I applied the algorithm in the joint space
of possible parameters and datasets, i.e., distribution of reference particles followed Equation 32. For the
distance function, which helps the NS particles approach the reference particle, I used a simple Euclidean
metric in the space of datasets:
v
u 100
uX
d (θ; θ ref ) = t
(xi − xref,i )2 .
(35)
i=1
Since the distance function refers only to the data x and not the parameter µ, the sampling is effectively
done in the space of possible datasets only — the parameter µ can be thought of as merely a latent variable
allowing p(x) to be explored conveniently.
√
I ran the algorithm and computed the average depth using a tolerance of r = 10−3 100, so that the
RMS difference between points in the two datasets was about 10−3 . From the algorithm, the estimate of
the expected depth was
h(depth)x i = 698.25 ± 0.34.
(36)
The uncertainty here is the standard error of the mean, i.e., the standard deviation of the depth estimates
divided by square root of the number of NS repetitions (the number of reference particles whose depth
was evaluated).
√
The log-volume of a 100-dimensional ball of radius r = 10−3 100 is −551.76. Therefore, the
differential entropy is estimated to be
H(x) ≈ 146.49 ± 0.34.
(37)
which is very close to the true value (perhaps suspiciously close, but this was just a fluke).
To carry out the sampling for this problem, Metropolis-Hastings moves that explore the joint prior
distribution p(µ, x) had to be implemented. Four basic kinds of proposals were used: i) those which
change µ while leaving x fixed; ii) those which change µ and shift x correspondingly; iii) those which
resample a subset of the xs from p(x|µ); and iv) those which move a single x slightly.
Entropy 2017,
11
6. Example 2: Measuring the period of an oscillating signal
In physics and astronomy, it is common to measure an oscillating signal at a set of times {t1 , ..., tn }, in
order to infer the amplitude, period, and phase of the signal. Here, I demonstrate the algorithm on a toy
version of Bayesian experimental design: at what times should the signal be observed in order to obtain
as much information as possible about the period? To answer this, we need to be able to calculate the
mutual information between the unknown period and the data.
As the algorithm lets us calculate the entropy of any distribution which can be sampled by MCMC,
there are several options. The mutual information can be written in various ways, such as the following.
I(θ; d) = H(θ) + H(d) − H(θ, d)
I(θ; d) = H(d) − H(d|θ)
(38)
(39)
I(θ; d) = H(θ) − H(θ|d)
Z
I(θ; d) = p(d) DKL (p(θ|d) || p(θ)) dd
(40)
I(θ; d) = DKL (p(θ, d) || p(θ)p(d)) .
(42)
(41)
In a Bayesian problem where θ is an unknown parameter and d is data, the first and second formulas would
be costly to compute, because the high dimensional probability distribution for the dataset would require
a large number of NS iterations to compute H(d) and H(θ, d). The fourth involves the Kullback-Leibler
divergence from the prior to the posterior, averaged over all possible datasets. This is straightforward
to approximate with standard Nested Sampling, but the estimate for a given dataset may be biased. It is
quite plausible that the bias is negligible compared to the variance from averaging over many datasets.
This is probably true in many situations.
However, more significantly, the above method would not work if we want I(θ; d) for a single
parameter or a subset of them, rather than for all of the parameters. The prior-to-posterior KL divergence
measured by standard NS is the amount learned about all parameters, including nuisance parameters.
Therefore, I adopted the third formula as the best way to compute I(θ; d). In the sinusoidal example, the
parameter of interest is the period, whereas the model will contain other parameters as well. Since the
period will have its prior specified explicitly, H(θ) will be available analytically. I then use the algorithm
to obtain H(θ|d) as follows:
Z
H(θ|d) = p(dref )H(θ|d = dref ) ddref .
(43)
This is an expected value over datasets, of the entropy of the posterior given each dataset. I ran the
algorithm by generating reference particles from p(θ, d), i.e., generating parameter values from the prior,
along with corresponding simulated datasets. Given each simulated dataset, I used NS to compute an
unbiased estimate of H(θ|d = dref ), by evolving a set of particles (initially from the posterior) towards the
reference particle. The role of the “prior” in the NS is actually taken by the posterior, and the “likelihood”
was minus a distance function defined in terms of the parameter(s) of interest.
To generate the initial NS particles from the posterior, I used standard MCMC (targeting the posterior),
initialised at the θ-value that produced the simulated dataset dref , which is a perfect sample from the
posterior.
Entropy 2017,
12
6.1. Assumptions
I consider two possible observing strategies, both involving n = 101 measurements. The ‘even’
strategy has observations at times
ti =
i−1
n−1
(44)
for i ∈ {1, 2, ..., 101}, that is, the observation times are 0, 0.01, 0.02, ..., 0.99, 1 — evenly spaced from
t = 0 to t = 1, including the endpoints. The second, ‘uneven’ observing strategy schedules the
observations according to
3
i − 12
ti =
(45)
n
which schedules observations close together initially, and further apart as time goes on.
A purely sinusoidal signal has the form
2πt
+φ
y(t) = A sin
T
(46)
where A is the amplitude, T is the period, and φ is the phase. Throughout this section, I parameterise the
period by its logarithm, τ = log10 T . I assumed the following priors:
ln A ∼ Normal 0, 0.12
(47)
τ ∼ Uniform(−1, 0)
φ ∼ Uniform(0, 2π)
(48)
(49)
and the following conditional prior for the data:
Yi |A, T, φ ∼ Normal y(ti ), 0.12 .
(50)
That is, the data is just the signal y(t) observed at particular times {t1 , ..., tn }, with gaussian noise of
standard deviation 0.1. The amplitude of the sinusoid is very likely to be around 10 times the noise level,
and the period is between 0.1 and 1 times the duration of the data. An example signal observed with the
even and uneven observing schedules is shown in Figure 2.
To carry out the sampling for this problem, Metropolis-Hastings moves that explore the posterior
distribution p(A, τ, φ|Y ) had to be implemented. I used heavy-tailed proposals (as in Section 8.3 of
Brewer and Foreman-Mackey [23]) which modified one parameter at a time.
6.2. Results
Letting τ = log10 T , and treating A and φ are nuisance parameters, I calculated the conditional entropy
H(τ |Y ) using the algorithm with a tolerance of r = 10−5 . The distance function was just the absolute
difference between τ and τref .
The result was
h(depth)i = 5.379 ± 0.038
(51)
Entropy 2017,
13
True signal
Even data
Uneven data
2.0
1.5
y
1.0
0.5
0.0
−0.5
−1.0
0.0
0.2
0.4
0.6
0.8
1.0
t
Figure 2. A signal with true parameters A = 1, τ = −0.5, and φ = 0, observed with
noise standard deviation 0.1 with the even (gold points) and uneven (green points) observing
strategies.
Converting to a differential entropy by adding ln(2r) = −10.820, the conditional entropy is
H(τ |Y ) = −5.441 ± 0.038 nats.
(52)
Since the Uniform(0,1) prior for τ has a differential entropy of zero, the mutual information is I(τ ; Y ) =
H(τ ) − H(τ |Y ) = 5.441 ± 0.038 nats.
The results for the uneven observing schedule were
h(depth)i = 5.422 ± 0.038
H(τ |Y ) = −5.398 ± 0.038
I(τ ; Y ) = 5.398 ± 0.038 nats.
(53)
(54)
(55)
The difference in mutual informations between the two observation schedules is trivial. The situation
may be different if we had allowed for shorter periods to be possible, as irregular observing schedules
are known to reduce aliasing and ambiguity in the inferred period [24]. When inferring the period
of an oscillating signal, multimodal posterior distributions for the period are common [17,25]. The
multimodality of the posterior here raises an interesting issue. Is the question we really want answered
“what is the value of T precisely?”, to which the mutual information relates? Most practicing scientists
would not feel particularly informed to learn that the vast majority of possibilities had been ruled out, if
the posterior still consisted of several widely separated modes! Perhaps, in some applications, a more
appropriate question is “what is the value of T to within ± 10%”, or something along these lines. See
Appendix A for more about this issue.
Entropy 2017,
14
In a serious experimental design situation, the relevance is not the only consideration (or the answer to
every experimental design problem would be to obtain more data without limit), but it is a very important
one.
7. Example 3: Data with Pareto distribution
Consider a situation where the probability distribution for some data values x = {x1 , x2 , ..., xn } is a
Pareto distribution, conditional on a minimum value xmin = 1 and a slope α:
n
Y
α
p(x|α) =
α+1
x
i=1 i
(56)
where all of the xi are greater than xmin = 1. If α is relatively low (close to zero) then this distribution is
very heavy-tailed, and as α increases the probability density for the xs concentrates just above x = 1.
With heavy-tailed distributions, it can be difficult to predict future observations from past ones, a notion
popularised by Taleb [26]. This occurs for three main reasons. First, the observations may not provide
much information about the parameters (in this case, α). Second, even conditional on the parameters, the
predictive distribution for future data can be heavy-tailed, implying a lot of uncertainty. Finally, one may
always doubt the probability assignment on which such a calculation is based. However, the data clearly
provides some nonzero amount of information about the future.
Consider n = 100 data values, partitioned into two halves, so y = {x1 , ..., x50 } = {y1 , ..., y50 } is the
first half of the data and z = {x51 , ..., x100 } = {z1 , ..., z50 } is the second half. Let ytot and ztot be the
totals of the two subsets of the data:
ytot =
50
X
yi
(57)
zi .
(58)
i=1
ztot =
50
X
i=1
In this section I use the algorithm to answer the questions: i) How uncertain is ytot ?; ii) How uncertain is
ztot ?; and iii) How informative is ytot about ztot ? These are quantified by H(ytot ), H(ztot ), and I(ytot ; ztot )
respectively. This can be done despite the fact that α is unknown.
The prior for α was chosen to be lognormal such that the prior median of α is 1 and the prior standard
deviation of ln α is also 1. Due to the heavy-tailed nature of the Pareto distribution, I ran the algorithm in
terms of ln ytot and ln ztot . This affects the differential entropies but not the mutual information (which is
invariant under changes of parameterisation). The joint distribution for ln ytot and ln ztot , whose properties
I aimed to obtain, is shown in Figure 3.
Three types of Metropolis-Hastings proposals were used for this example: i) A proposal which changes
α and modifies all of the xs accordingly; ii) A proposal which resamples a subset of the xs from the
Pareto distribution; and iii) A proposal which moves a single x-value slightly. A fourth proposal which
moves α while keeping x fixed, i.e., a posterior-sampling move, was not implemented.
Entropy 2017,
15
10
8
ln(ztot)
6
4
2
0
0
2
4
ln(ytot)
6
8
10
Figure 3. The joint distribution for the log-sum of the first half of the ‘Pareto data’ and the
second half. The goal is to calculate the entropy of the marginal distributions, the entropy of
the joint distribution, and hence the mutual information. The distribution is quite heavy tailed
despite the logarithms, and extends beyond the domain shown in this plot.
To compute the entropy of ln ytot , the distance function was just the absolute value of the distance
between the NS particles and the reference particle in terms of ln ytot . The result, using a tolerance of
r = 10−3 , was:
h(depth)ln ytot i = 8.851 ± 0.057.
(59)
Correcting for the volume factor yields a differential entropy estimate of H(ln ytot ) = 2.366 ± 0.057 nats.
By the symmetry of the problem specification, H(ln ztot ) = H(ln ytot ).
The joint entropy H(ln ytot , ln ztot ) was also estimated by using a Euclidean distance function and a
tolerance of 10−3 . The result was
(depth)(ln ytot ,ln ztot ) = 16.03 ± 0.11.
(60)
Applying the log-volume correction yields a differential entropy estimate of H(ln ytot , ln ztot ) = 3.40 ±
0.11 nats. Therefore, the mutual information is
I(ytot ; ztot ) = H(ln ytot ) + H(ln ztot ) − H(ln ytot , ln ztot )
= 1.130 ± 0.082 nats.
(61)
(62)
The variance of this mutual information estimate was reduced by using a common sequence of reference
points for the marginal and joint entropy runs. This is an instance of the ‘common random numbers’
Entropy 2017,
16
technique1 . In short, for those reference particles with a large depth in terms of the ln ytot distance function,
the depth in terms of the (ln ytot , ln ztot ) distance function was also large. Since the mutual information
involves a difference of the averages of these depths, it is more accurate to compute the average difference
(i.e. compute the two depths using the same reference particle, to compute the difference between the two
depths) than the difference of the averages.
8. Computational cost, limitations, and conclusion
The computational resources needed to compute these quantities was quite large, as Nested Sampling
appears in the inner loop of the algorithm. However, it is worth reflecting on the complexity of the
calculation that has been done.
In a Bayesian inference problem with parameters θ and data d, the mutual information is
ZZ
p(θ, d)
dθ dd
(63)
I(θ; d) =
p(θ, d) ln
p(θ)p(d)
and this measures the dependence between θ and d. A Monte Carlo strategy to evaluate this is to sample
from p(θ, d) and average the value of the logarithm, which involves computing the log-evidence
Z
ln p(d) = ln
p(θ)p(d|θ) dθ
(64)
for each possible data set. Computing p(d) for even a single data set has historically been considered
difficult.
However, if we want the mutual information between the data and a subset of the parameters (i.e.,
there are nuisance parameters we don’t care about), things become even more tricky. Suppose we partition
the parameters θ into important parameters φ and nuisance parameters η, such that we want to calculate
the mutual information I(φ; d). This is given by
ZZ
p(φ, d)
I(φ; d) =
p(φ, d) ln
dφ dd
(65)
p(φ)p(d)
which is equivalent to Equation 63 but with θ replaced throughout by φ. However, this is an even
more difficult calculation than before, as p(d|φ), the marginal likelihood function with the nuisance
parameters integrated out, is typically unavailable in closed form. If we were to marginalise out the
nuisance parameters η explicitly, this would give us Equation 65 with every probability distribution
written as an explicit integral over η:
Z Z Z
I(φ; d) =
R
p(φ, η)p(d|φ, η) dη
R
p(φ, η)p(d|φ, η) dη ln
dφ dd
p(φ) p(φ, η)p(d|φ, η) dφ dη
(66)
It should not be surprising that this is costly. Indeed, it seems possible to me that any general algorithm
for computing entropies will have the property that NS (or a similar algorithm capable of measuring small
1
http://demonstrations.wolfram.com/TheMethodOfCommonRandomNumbersAnExample/
Entropy 2017,
17
volumes) appears in an inner loop, since entropies are sums of terms, each of which relates to the downset
of a given statement — the downset of a statement S being the set of all statements that imply S [3].
The current implementation of the algorithm (see Appendix B) is in C++ and has not been parallelised.
Given the high computational costs, parallelisation should be attempted. This could follow the method of
[21], or alternatively, different threads could work on different reference particles simultaneously.
The algorithm as presented will tend to become less accurate on high dimensional problems, as the
required depth of the runs will increase correspondingly, since many NS iterations will be required to get
a high dimensional vector of quantities to be close to the reference vector. The accuracy of the ln(depth)
p
estimates should degrade in proportion to (depth), due to the Poisson-process nature of the − ln X
sequence.
Throughout this paper, the question of how to sample from the constrained distributions required by
NS has not been discussed in detail. For the examples, I used the original Skilling [11] suggestion of
using a short MCMC run initialised at a surviving point. In some problems, this will not be effective,
and this will cause the resulting estimates to be incorrect, as in standard Nested Sampling applications.
Testing the sensitivity of the results to changes in the number of MCMC steps used is a simple way to
increase one’s confidence in the results, but is not a guarantee (no general guarantee can exist, because
any algorithm based on sampling a subset of a space might miss a crucial but hard-to-locate dominant
spike). Despite this weakness, I hope the algorithm will be useful in a wide range of applications.
Acknowledgements
I would like to thank John Skilling (Maximum Entropy Data Consultants) and the anonymous referee
for their comments. It is a pleasure to thank the following people for interesting and helpful conversations
about this topic, and/or comments on a draft: Ruth Angus (Flatiron Institute), Ewan Cameron (Oxford),
James Curran (Auckland), Tom Elliott (Auckland), David Hogg (NYU), Kevin Knuth (SUNY Albany),
Thomas Lumley (Auckland), Iain Murray (Edinburgh), Jared Tobin (jtobin.io). This work was
supported by Centre for eResearch at the University of Auckland.
Conflicts of Interest
The authors declare no conflicts of interest.
References
1. Shannon, C.E. A mathematical theory of communication. The Bell System Technical Journal
1948, 27, 379–423.
2. Cover, T.M.; Thomas, J.A. Elements of information theory; John Wiley & Sons, 2012.
3. Knuth, K.H. Toward question-asking machines: the logic of questions and the inquiry calculus
2005.
4. Knuth, K.H.; Skilling, J. Foundations of inference. Axioms 2012, 1, 38–73.
5. Caticha, A.; Giffin, A. Updating Probabilities. Bayesian Inference and Maximum Entropy
Methods In Science and Engineering. AIP Publishing, 2006, Vol. 872, pp. 31–42.
Entropy 2017,
18
6. Szabó, Z. Information Theoretical Estimators Toolbox. Journal of Machine Learning Research
2014, 15, 283–287.
7. Jaynes, E.T. Probability theory: The logic of science; Cambridge university press, 2003.
8. MacKay, D.J. Information theory, inference and learning algorithms; Cambridge university press,
2003.
9. Caticha, A. Lectures on probability, entropy, and statistical physics.
arXiv preprint
arXiv:0808.0012 2008.
10. Bernardo, J.M. Reference analysis. Handbook of statistics 2005, 25, 17–90.
11. Skilling, J. Nested sampling for general Bayesian computation. Bayesian analysis 2006,
1, 833–859.
12. Feroz, F.; Hobson, M.; Bridges, M. MultiNest: an efficient and robust Bayesian inference tool
for cosmology and particle physics. Monthly Notices of the Royal Astronomical Society 2009,
398, 1601–1614.
13. Brewer, B.J.; Pártay, L.B.; Csányi, G. Diffusive nested sampling. Statistics and Computing 2011,
21, 649–656.
14. Handley, W.; Hobson, M.; Lasenby, A. POLYCHORD: next-generation nested sampling. Monthly
Notices of the Royal Astronomical Society 2015, 453, 4384–4398.
15. Knuth, K.H.; Habeck, M.; Malakar, N.K.; Mubeen, A.M.; Placek, B. Bayesian evidence and
model selection. Digital Signal Processing 2015, 47, 50–67.
16. Pullen, N.; Morris, R.J. Bayesian model comparison and parameter inference in systems biology
using nested sampling. PloS one 2014, 9, e88419.
17. Brewer, B.J.; Donovan, C.P. Fast Bayesian inference for exoplanet discovery in radial velocity
data. Monthly Notices of the Royal Astronomical Society 2015, 448, 3206–3214.
18. Pártay, L.B.; Bartók, A.P.; Csányi, G. Efficient sampling of atomic configurational spaces. The
Journal of Physical Chemistry B 2010, 114, 10502–10512.
19. Baldock, R.J.; Pártay, L.B.; Bartók, A.P.; Payne, M.C.; Csányi, G.
Determining
pressure-temperature phase diagrams of materials. Physical Review B 2016, 93, 174108.
20. Martiniani, S.; Stevenson, J.D.; Wales, D.J.; Frenkel, D. Superposition enhanced nested sampling.
Physical Review X 2014, 4, 031034.
21. Henderson, R.W.; Goggans, P.M.; Cao, L. Combined-chain nested sampling for efficient Bayesian
model comparison. Digital Signal Processing 2017, 70, 84 – 93.
22. Walter, C. Point process-based Monte Carlo estimation. Statistics and Computing 2017,
27, 219–236.
23. Brewer, B.J.; Foreman-Mackey, D. DNest4: Diffusive Nested Sampling in C++ and Python.
Journal of Statistical Software, accepted. arxiv: 1606.03757 2016, [arXiv:stat.CO/1606.03757].
24. Bretthorst, G.L. Nonuniform sampling: Bandwidth and aliasing. AIP conference proceedings.
AIP, 2001, Vol. 567, pp. 1–28.
25. Gregory, P.C. A Bayesian Analysis of Extrasolar Planet Data for HD 73526. The Astrophysical
Journal 2005, 631, 1198–1214.
26. Taleb, N.N. The Black Swan: The Impact of the Highly Improbable. ISBN: 978-1400063512
2007.
Entropy 2017,
19
A. Precisional questions
The central issue about the value of a parameter θ asks “what is the value of θ, precisely?”. However,
in practice we often don’t need or want to know θ to arbitrary precision. Using the central issue can lead
to counterintuitive results if you don’t keep its specific definition in mind. For example, suppose x could
take any integer value from 1 to 1 billion. If we learned the last digit of x, we will have ruled out nine
tenths of the possibilities, and therefore obtained a lot of information about the central issue. However,
this information might be useless for practical purposes. If we learned the final digit was a 9, x could still
be 9, or 19, or 29, any number ending in 9 up to 999,999,999.
In practice, we may want to ask a question that is different from the central issue. For example, suppose
x ∈ {1, ..., 10} and we want to know the value of x to within a tolerance of ±1. Any of the following
statements would resolve the issue:
• x ∈ {1, 2, 3} and anything that implies it
• x ∈ {2, 3, 4} and anything that implies it
• x ∈ {3, 4, 5} and anything that implies it
• and so on.
According to Knuth [3], the entropy of a question is computed by applying the sum rule over all statements
that would answer the question.
We can write the question of interest, Q, as a union of ideal questions (the ideal questions are downsets
of statements: the set of all statements that imply S is the downset of S and is denoted ↓ S):
Q = [↓ (x ∈ {1, 2, 3})] ∪ [↓ (x ∈ {2, 3, 4})] ∪ [↓ (x ∈ {3, 4, 5})] ∪ ...
(67)
The entropy of this question is
H(Q) = h1,2,3
+ (h2,3,4 − h2,3 )
+ (h3,4,5 − h3,4 )
(68)
(69)
(70)
+ ...
(71)
+ (h8,9,10 − h8,9 ) .
(72)
where hx = −P (x) log P (x).
Continuous case
Consider a probability density function f (x), defined on the real line. The probability contained in an
interval [x0 , x0 + r], which has length r, is
Z x0 +r
f (x) dx = F (x0 + r) − F (x0 )
(73)
x0
Entropy 2017,
20
where F (x) is the cumulative distribution function (CDF). This probability can be considered as a function
of two variables, x0 and r, which I will denote by P (, ):
P (x0 , r) = F (x0 + r) − F (x0 ).
(74)
The contribution of such an interval to an entropy expression is −P log P , i.e.,
Q(x0 , r) = −P (x0 , r) log P (x0 , r).
(75)
Consider the rate of change of Q as the interval is shifted to the right, but with its width r held constant:
∂Q
∂
=
[−P log P ]
∂x0
∂x0
∂P
∂x0
= − [1 + log P (x0 , r)] [f (x0 + r) − f (x0 )] .
= − [1 + log P (x0 , r)]
(76)
(77)
(78)
Consider also the rate of change of Q as the interval is expanded to the right, while keeping its left edge
fixed:
∂
∂Q
=
[−P (x0 , r) log P (x0 , r)]
(79)
∂r
∂r
∂P
= − [1 + log P (x0 , r)]
(80)
∂r
= − [1 + log P (x0 , r)] f (x0 + r).
(81)
The entropy of the precisional question is built up from Q terms. The extra entropy from adding the
interval [x0 , x0 + r] and removing the overlap [x0 , x0 + r − h], for small h, is
δH = Q(x0 , r) − Q(x0 , r − h)
∂Q
=h
∂r
= −h [1 + log P (x0 , r)] f (x0 + r).
(82)
(83)
(84)
Therefore, the overall entropy is
H=−
=−
Z
∞
Z−∞
∞
−∞
[1 + log P (x − r, r)] f (x) dx
(85)
(1 + log [F (x) − F (x − r)]) f (x) dx.
(86)
Interestingly, calculating the log-probability of an interval to the right of x gives an equivalent result:
Z ∞
0
H =−
(1 + log [F (x + r) − F (x)]) f (x) dx
(87)
−∞
= H.
(88)
This is not quite the same as the expected log-probability calculated by the version of the algorithm
proposed in this paper (when the tolerance is not small). However, the algorithm can be made to estimate
the entropy of the precisional question, by redefining the distance function. For a one-dimensional quantity
of interest, the distance function can be defined such that d(x; xref ) can only ever be below r if x < xref
(or alternatively, if x > xref ).
Entropy 2017,
21
B. Software
A C++ implementation of the algorithm is available in a git repository located at
https://github.com/eggplantbren/InfoNest
and can be obtained using the following git command, executed in a terminal:
git clone https://github.com/eggplantbren/InfoNest
The following will compile the code and execute the first example from the paper:
cd InfoNest/cpp
make
./main
The algorithm will run for 1000 ‘reps’, i.e., 1000 samples of θ ref , which is time consuming. Output is
saved to output.txt. At any time, you can execute the Python script postprocess.py to get an
estimate of the depth:
python postprocess.py
By default, postprocess.py estimates the depth with a tolerance of r = 10−3 . This value
can be changed by calling the postprocess function with a different value of its argument tol.
E.g., postprocess.py can be edited so that its last line is postprocess(tol=0.01) instead of
postprocess().
The numerical parameters, and the choice of which problem is being solved, are specified in
main.cpp. The default problem is the first example from the paper. For Example 2, since it is a
conditional entropy which is being estimated (which requires a slight modification to the algorithm),
an additional argument InfoNest::Mode::conditional_entropy must be passed to the
InfoNest::execute function.
c 2017 by the author; licensee MDPI, Basel, Switzerland. This article is an open access article
distributed under the terms and conditions of the Creative Commons Attribution license
(http://creativecommons.org/licenses/by/4.0/).
| 7cs.IT
|
Promises and Caveats of Uplink
IoT Ultra-Dense Networks
Ming Ding‡ , David López Pérez†
‡
arXiv:1801.06623v1 [cs.NI] 20 Jan 2018
†
Data61, Australia {[email protected]}
Nokia Bell Labs, Ireland {[email protected]}
Abstract—In this paper, by means of simulations, we evaluate
the uplink (UL) performance of an Internet of Things (IoT)
capable ultra-dense network (UDN) in terms of the coverage
probability and the density of reliably working user equipments
(UEs). From our study, we show the benefits and challenges
that UL IoT UDNs will bring about in the future. In more
detail, for a low-reliability criterion, such as achieving a UL
signal-to-interference-plus-noise ratio (SINR) above 0 dB, the
density of reliably working UEs grows quickly with the network
densification, showing the potential of UL IoT UDNs. In contrast,
for a high-reliability criterion, such as achieving a UL SINR
above 10 dB, the density of reliably working UEs remains to
be low in UDNs due to excessive inter-cell interference, which
should be considered when operating UL IoT UDNs. Moreover,
considering the existence of a non-zero antenna height difference
between base stations (BSs) and UEs, the density of reliably
working UEs could even decrease as we deploy more BSs. This
calls for the usage of sophisticated interference management
schemes and/or beam steering/shaping technologies in UL IoT
UDNs.
I. I NTRODUCTION
Recent years have seen rapid advancement in the development and deployment of Internet of Things (IoT) networks,
which can be attributed to the increasing communication and
sensing capabilities combined with the falling prices of IoT
devices [1]. For example, base stations (BSs) can be equipped
with the latest IoT technologies, such as the new generation
of machine type communications [2], to collect data from gas,
water, and power meters via uplink (UL) transmissions. In
practice, such BSs can take the form of both terrestrial and
aerial ones [3].
This poses, however, a challenge to the wireless industry,
which must offer an increasing volume of reliable traffic in a
profitable and energy efficient manner, especially for the UL
communications. In this context, the orthogonal deployment
of ultra-dense (UD) small cell networks (SCNs), or simply
ultra-dense networks (UDNs), have been selected as one of the
workhorse for network enhancement in the fourth-generation
(4G) and fifth-generation (5G) networks developed by the 3rd
Generation Partnership Project (3GPP) [2]. Here, orthogonal
deployment means that UDNs and macrocell networks operate
on different frequency spectrum, which simplifies network
management due to no inter-tier interference.
The performance analysis of IoT UDNs is, however, particularly challenging for the UL because i) UDNs are fundamentally different from the current sparse/dense networks [4]
and ii) the UL power control mechanism operates according
to the random user equipment (UE) positions in the network,
which is quite different from the constant power setting in the
downlink (DL) [5].
In this paper, by means of simulations, we evaluate the
network performance of UL IoT UDNs in terms of the
coverage probability and the density of reliably working UEs.
The main findings of this paper are as follows:
• We find that for a low-reliability criterion, such as achieving a UL signal-to-interference-plus-noise ratio (SINR)
above γ = 0 dB, the density of reliably working UEs
quickly grows with the network densification, showing
the benefits of UL IoT UDNs. In contrast, for a highreliability criterion, such as achieving a UL SINR above
γ = 10 dB, the density of reliably working UEs remains
low in UDNs due to excessive inter-cell interference,
which should be considered when operating UL IoT
UDNs.
• We find that due to the existence of a non-zero antenna
height difference between BSs and UEs, the density of
reliably working UEs could even decrease as we deploy
more BSs in a UL IoT UDN. This calls for the usage
of sophisticated interference management schemes and/or
beam steering/shaping technologies in UL IoT UDNs.
• We find that the correlated shadow fading allows a BS
with a lower environmental fading factor to provide connection to a larger number of UEs. Thus, its theoretical
analysis is an open problem for further study.
• We find that the optimized hexagonal-like BS deployment
can improve network performance for relatively sparse
networks, but not for UDNs. Thus, its theoretical study
is not urgent.
II. D ISCUSSION
ON THE
A SSUMPTIONS
OF
UL I OT UDN S
In this section, we discuss several important assumptions in
UL IoT UDNs.
A. BS Deployments
In general, to study the performance of a UL IoT network,
two types of BS deployments can be found in the literature,
i.e., the hexagonal and the random deployments as shown in
Fig. 1.
In Fig. 1, BSs are represented by markers “x” and cell
coverage areas are outlined by solid lines. Note that the hexagonal BS deployment leads to an upper-bound performance
because BSs are evenly distributed in the network scenario,
(a) The hexagonal BS deployment.
(b) The random BS deployment.
Fig. 1: Illustration of two widely accepted types of BS deployments. Here, BSs are represented by markers “x” and cell
coverage areas for user equipment (UE) distribution are outlined by solid lines.
and thus very strong interference due to close BS proximity is
precluded [6]. In contrast, the random BS deployment reflects
a more realistic network deployment with more challenging
interference conditions [2, 7]. For completeness, in our following performance evaluation, we will consider both BS
deployments.
B. Antenna Heights
In the performance analysis of the conventional sparse or
dense cellular networks, the antenna height difference between
BSs and UEs is usually ignored due to the dominance of
the horizontal distance. However, with a much shorter distance between a BS and its served UEs in an UDN, such
antenna height difference becomes non-negligible [7]. The
performance impact of such antenna height difference between
BSs and UEs on the DL UDNs has been investigated in [8].
More specifically, the existence of a non-zero antenna height
difference between BSs and UEs gives rise to a non-zero
cap on the minimum distance between a BS and its served
UEs, and thus a cap on the received signal power strength.
Thus, and although each inter-cell interference power strength
is subject to the same cap, the aggregate inter-cell interference
power will overwhelm the signal power in an UDN due to the
sheer number of strong interferers. Consequently, the antenna
height difference between BSs and UEs should be considered
in the performance evaluation of UL IoT UDNs.
C. Line-of-Sight Transmissions
A much shorter distance between a BS and its served
UEs in an UDN implies a higher probability of line-of-sight
(LoS) transmissions. The performance impact of such LoS
transmissions on the DL of UDNs has been shown to be
significant in [9, 10]. Generally speaking, LoS transmissions
are more helpful to enhance the received signal strength
than non-line-of-sight (NLoS) transmissions. However, after
a certain level of BS densification, not only the signal power,
but also the inter-cell interference power will significantly
grow due to the emergence of LoS interfering paths. Thus,
the transition of a large number of interfering paths from
NLoS to LoS will overwhelm the signal power in an UDN.
Consequently, the probabilistic LoS transmissions should also
be considered in the performance evaluation of UL IoT UDNs.
D. BS Idle Mode
Considering the surplus of BSs in UDNs, a BS can be put
to sleep when there is no active UE connected to it, which is
referred as the BS idle mode capability (IMC) [11, 12]. As a
result, the active UEs’ SINR can benefit from i) a BS selection
diversity gain, i.e., each UE has a plurality of BSs to select its
server from, and ii) a tight control of inter-cell interference,
as effective UL inter-cell interference only comes from active
UEs served by active neighboring BSs. The surplus of BSs
together with the IMC can be seen as a powerful tool that can
mitigate the interference problems presented in the previous
subsections. However, it should be noted that switching off
BSs has a negative impact on the number of IoT UEs that can
concurrently transmit.
III. S YSTEM M ODEL
For a certain time-frequency resource block (RB), we consider a UL IoT network with BSs deployed on a plane according to the hexagonal deployment or the random deployment,
as shown in Fig. 1. For both BS deployments, the density of
BSs is denoted by λ BSs/km2 . Furthermore, we consider a
homogeneous Poisson point process (HPPP) Φ to characterize
the random deployment.
Active UEs are assumed to be distributed following an
HPPP with a density of ρ UEs/km2 . Here, we only consider
active UEs in the network because non-active UEs do not
trigger data transmission. Note that the total number of UEs,
e.g., phones, gateways, sensors, tags, etc., in a UL IoT network
should be much higher than the number of the active UEs.
However, we believe that in a certain time-frequency RB, the
active UEs with non-zero data traffic demands should still be
not many. For example, a typical density of active UEs is
around 300 UEs/km2 in 5G [2].
NL
A. Channel Model
The two-dimensional (2D) distance between a BS and a
UE is denoted by r. Moreover, the absolute antenna height
difference between a BS and a UE is denoted by L. Thus, the
3D distance between a BS and a UE can be expressed as
p
w = r 2 + L2 .
(1)
Note that the value of L is in the order of several meters [13].
Following [10], we adopt a general path loss model, where
the path loss ζ (w) is a multi-piece function of w written as
ζ1 (w) , when L ≤ w ≤ d1
ζ2 (w) , when d1 < w ≤ d2
ζ (w) = .
,
(2)
..
..
.
ζN (w) , when w > dN −1
where each piece ζn (w) , n ∈ {1, 2, . . . , N } is modeled as
(
L
ζ L (w) = ALn w−αn ,
LoS: PrLn (w)
ζn (w)= nNL
, (3)
NL
−αn
ζn (w) = ANL
, NLoS: 1 − PrLn (w)
n w
where
L
NL
• ζn (w) and ζn (w) , n ∈ {1, 2, . . . , N } are the n-th piece
path loss functions for the LoS and the NLoS cases,
respectively,
L
NL
• An and An are the path losses at a reference 3D distance
w = 1 for the LoS and the NLoS cases, respectively,
L
NL
• αn and αn are the path loss exponents for the LoS and
the NLoS cases, respectively, and
L
• Prn (w) is the n-th piece LoS probability function that a
transmitter and a receiver separated by a 3D distance w
has an LoS path, which is assumed to be a monotonically
decreasing function with respect to w. Existing measurement studies have confirmed this assumption [13].
Moreover, we assume that each BS/UE is equipped with an
isotropic antenna, and that the multi-path fading between a BS
and a UE is independently identical distributed (i.i.d.) Rayleigh
distributed [9, 10, 14].
B. User Association Strategy
We assume a practical user association strategy (UAS), in
which each UE is connected to the BS giving the maximum
average received signal strength [10, 14]. Such UAS can be
formulated by
bo = arg max R̄b (w) ,
(4)
b
where R̄b (w) denotes the average received signal strength
from BS b and the UE of interest, separated by a distance
of w. Assuming a constant BS transmission power, R̄b (w)
can be equivalently evaluated by ζ (w) defined in (2).
As a special case to show our numerical results in the
simulation section, we consider a practical two-piece path
loss function and a two-piece exponential LoS probability
function, defined by the 3GPP [13]. More specifically, in (2)
L
we use N = 2, ζ1L (w) = ζ2L (w) = AL w−α , ζ1NL (w) =
ζ2NL (w) = ANL w−α , PrL1 (w) = 1 − 5 exp (−R1 /w), and
PrL2 (w) = 5 exp (−w/R2 ), where R1 = 156 m, R2 = 30 m,
1
= 67.75 m [13]. For clarity, this path loss case
and d1 = lnR10
is referred to as the 3GPP Case hereafter.
C. BS Activation Model
As discussed in Subsection II-D, a BS will enter an idle
mode if there is no UE connected to it. Thus, the set of active
BSs is determined by the UAS. Since UEs are randomly and
uniformly distributed in the network and given the adopted
UAS strategy, we can assume that the active BSs also follow
an HPPP distribution Φ̃ [11], the density of which is denoted
by λ̃ BSs/km2 , where λ̃ ≤ λ and λ̃ ≤ ρ. Note that λ̃ also
characterizes the density of active UEs because no collision
exists in the centralized cellular IoT UDNs.
For illustration purposes, considering a single-slope path
loss model and a nearest-BS UAS, λ̃ can be calculated as [11]
1
q ,
λ̃ = λ 1 −
(5)
ρ
1 + qλ
where an empirical value of 3.5 was suggested for q in [11]1 .
D. UL Power Control Model
The UE power, denoted by P UE , is subject to semi-static
power control (PC) in practice. In this paper, we adopt the
fractional path loss compensation (FPC) scheme standardized
in 4G [13], which can be modeled as
P0
P UE = 10 10 [ζ (w)]
−η
N RB ,
(6)
where P0 is the target received power in dBm on each RB at
the BS, η ∈ (0, 1] is the FPC compensation factor, and N RB
is the number of RBs in the frequency domain.
E. The Coverage Probability
Based on this system model, we can define the coverage
probability that the typical UE’s UL SINR is above a designated threshold γ as
pcov (λ, γ) = Pr SINRU > γ ,
(7)
where the UL SINR is calculated by
SINRU =
PbUE
ζ (wbo ) h
o
,
U + PU
Iagg
N
(8)
where bo denotes the serving BS of the typical UE, PbUE
is
o
the UE transmission power given by (6), wbo is the distance
from the typical UE to its serving BS bo , h is the Rayleigh
channel gain modeled as an exponentially distributed random
variable (RV) with a mean of one as mentioned above, PNU
is the additive white Gaussian noise (AWGN) power at the
U
serving BS bo , and Iagg
is the UL aggregate interference.
1 Note that according to [12], q should also depend on the path loss model
with LoS/NLoS transmissions. Having said that, [12] also showed that (5) is
generally very accurate to characterize λ̃ for the 3GPP Case with LoS/NLoS
transmissions.
F. The Density of Reliably Working UEs
Based on the definitions of the active BS density in Subsection III-C and the coverage probability in Subsection III-E,
we can further define a density of reliably working UEs that
can operate above a target UL SINR threshold γ as
ρ̃ = λ̃pcov (λ, γ) ,
(9)
where the active BS density λ̃ measures the maximum density
of UEs that can simultaneously transmit, and the coverage
probability pcov (λ, γ) scales down λ̃, giving the density of
reliably working UEs. The larger the UL SINR threshold γ, the
higher the reliability of the IoT communications, and thus the
less the UEs that can simultaneously achieve such reliability.
G. More Refined Assumptions
In performance analysis, the multi-path fading is usually
modeled as Rayleigh fading for simplicity. However, in the
3GPP, a more practical model based on generalized Rician
fading is widely adopted for LoS transmissions [15]. In such
model, the K factor in dB scale (the ratio between the power
in the direct path and the power in the other scattered paths)
is modeled as K[dB] = 13 − 0.03w, where w is defined in (1).
More specifically, let hL denote the multi-path fading power
for LoS transmissions. Then, for the 3GPP model of Rician
fading, hL follows a non-central chi-squared distribution with
its PDF given by [16]
= (K + 1) exp −K − (K + 1) hL
f hL
q
L
×I0 2 K (K + 1) h ,
(10)
where K is the distance-dependent value discussed above and
I0 (·) is the 0-th order modified Bessel function of the first
kind [16].
Moreover, the shadow fading is also usually not considered
or simply modeled as i.i.d. RVs in performance analysis.
However, in the 3GPP, a more practical correlated shadow
fading is often used [13, 15, 17], where the shadow fading in
dB unit is modeled as zero-mean Gaussian RV [13]. More
specifically, the shadow fading coefficient in dB unit between
BS b and UE u is formulated as [13]
√
√
(11)
Sbu = τ SuUE + 1 − τ SbBS ,
where τ is the correlation coefficient of shadow fading, SuUE
and SbBS are i.i.d. zero-mean Gaussian RVs attributable to UE
u and BS b, respectively. The variance of SuUE and SbBS is
2
denoted by σShad
. In [13], it is suggested that τ = 0.5 and
σShad = 10 dB.
Considering the distance-dependent Rician fading for LoS
transmissions and the correlated shadow fading, we can upgrade the 3GPP Case to an advanced 3GPP Case. In the next
section, we will present simulation results of UL IoT UDNs
for both the 3GPP Case and the Advanced 3GPP Case. It
should be noted that for the Advanced 3GPP Case, shadow
fading should be considered in the computation of R̄b (w),
Sbu
i.e., R̄b (w) should be evaluated by ζ (w) × 10 10 in (4).
IV. S IMULATION AND D ISCUSSION
In this section, we present numerical results to validate the
accuracy of our analysis. According to Tables A.1-3~A.1-7
of [13], we adopt the following parameters for the 3GPP Case:
αL = 2.09, αNL = 3.75, AL = 10−10.38 , ANL = 10−14.54 ,
P0 = −76 dBm, η = 0.8, N RB = 55, PN = −91 dBm (with
a noise figure of 13 dB), τ = 0.5 and σShad = 10 dB.
A. Performance Results of the 3GPP Case
In Figs. 2 and 3, we plot the performance results of the
3GPP Case for γ = 0 dB and γ = 10 dB, respectively.
Here, we only consider realistic networks with the random
deployment of BSs. From these two figures, we can draw
the following observations:
• Figs. 2a and 3a show that the active BS density λ̃
monotonically increases with the network densification,
and it is bounded by ρ = 300 UEs/km2 . Such results are
in line with the analytical results in (5) [11, 12]. However,
the density of reliably working UEs ρ̃, i.e., λ̃pcov (λ, γ)
defined in (9), does not necessarily grow as the BS density
λ increases. This is because pcov (λ, γ) is a non-monotone
function with respect to λ, which will be explained in the
following.
• When
BS density λ is around λ
∈
−1 the
10 , 70 BSs/km2 , the network is noise-limited,
and thus pcov (λ, γ) increases with λ as the network is
lightened up with coverage and the signal power strength
benefits form LoS transmissions.
2
• When the BS density λ is around λ ∈ [70, 400] BSs/km ,
cov
p (λ, γ) decreases with λ. This is due to the transition
of a large number of interfering paths from NLoS to
LoS, which accelerates the growth of the aggregate intercell interference. Such performance behavior has been
reported in [10] for the
[18] for the UL.
DL and
4
• When λ ∈ 400, 10
BSs/km2 , pcov (λ, γ) continuously
increases thanks to the BS IMC [12], i.e., the signal power
continues increasing with the network densification due
to the BS diversity gain, while the aggregate interference
power becomes constant, as λ̃ is bounded by ρ. However,
as shown in Figs. 2b and 3b, the antenna height difference
L between BSs and UEs has a large impact on pcov (λ, γ)
because a non-zero L places a bound on the signal power
strength, which degrades the coverage performance. In
more detail, when λ = 104 BSs/km2 and γ = 0 dB,
the coverage probability with L = 8.5 m loses 13 %
compared to that with L = 0 m. Such performance degradation further enlarges to 32 % when λ = 104 BSs/km2
and γ = 10 dB, showing a much less chance of UE
working reliably above 10 dB.
• Due to the complicated performance behavior of
pcov (λ, γ), the density of reliably working UEs ρ̃ displayed in Figs. 2c and 3c depends on the following
factors:
– For a low-reliability criterion, such as surpassing a
UL SINR threshold of γ = 0 dB, the density of reliably working UEs ρ̃ grows quickly with the network
(a) The active BS density λ̃.
(b) The coverage probability pcov (λ, γ).
(c) The density of reliably working UEs ρ̃.
Fig. 2: Performance results of the 3GPP Case (Rayleigh fading, no shadow fading) with γ = 0 dB.
(a) The active BS density λ̃.
(b) The coverage probability pcov (λ, γ).
(c) The density of reliably working UEs ρ̃.
Fig. 3: Performance results of the 3GPP Case (Rayleigh fading, no shadow fading) with γ = 10 dB.
(a) The active BS density λ̃.
(b) The coverage probability pcov (λ, γ).
(c) The density of reliably working UEs ρ̃.
Fig. 4: Performance results of the Advanced 3GPP Case (Rician fading for LoS, correlated shadow fading) with γ = 0 dB.
(a) The active BS density λ̃.
(b) The coverage probability pcov (λ, γ).
(c) The density of reliably working UEs ρ̃.
Fig. 5: Performance results of the Advanced 3GPP Case (Rician fading for LoS, correlated shadow fading) with γ = 10 dB.
densification, showing the benefits of UL IoT UDNs.
In contrast, for a high-reliability criterion, such as
surpassing a UL SINR threshold of γ = 10 dB, the
density of reliably working UEs ρ̃ does not exhibit
a satisfactory performance even in UDNs, e.g., we
merely get ρ̃ < 50 UEs/km2 when λ = 103 BSs/km2 .
The situation only improves when the BS IMC fully
kicks in, e.g., λ > 103 BSs/km2 .
– Considering the existence of a non-zero L, the
density of reliably working UEs ρ̃ could even decrease as we deploy more BSs (see Fig. 3c, λ ∈
[200, 600] BSs/km2 ). This calls for the usage of
sophisticated interference management schemes [19]
in UL IoT UDNs. Another solution to mitigate
such strong inter-cell interference is beam steering/shaping using multi-antenna technologies [2].
B. Performance Results of the Advanced 3GPP Case
In Figs. 4 and 5, we plot the performance results of the
Advanced 3GPP Case for γ = 0 dB and γ = 10 dB, respectively. To make our study more complete, here we consider
both the random and the hexagonal deployments of BSs. From
these two figures, we can see that the previous conclusions are
qualitatively correct, which indicates that it is not urgent to
investigate Rician fading and/or correlated shadow fading in
the context of UDNs. However, there are two new observations
that are worth mentioning:
• From Figs. 2a/3a and Figs 4a/5a, we can see that the
active BS density λ̃ of the Advanced 3GPP Case is
smaller than that of the 3GPP Case. This means that (5)
is no longer accurate to characterize λ̃ for the Advanced
3GPP Case. This is because the correlated shadow fading
allows a BS with a lower environmental fading factor,
i.e., SbBS , to attract more UEs than other BSs with higher
values of SbBS . Its theoretical analysis is an open problem
for further study.
• The hexagonal deployment of BSs can improve network performance for relatively sparse networks (e.g.,
λ < 102 BSs/km2 ), but not for UDNs (e.g., λ >
103 BSs/km2 ). Such conclusion indicates that it is not
urgent to investigate the performance of UDNs with the
hexagonal deployment of BSs, which has been a longstanding open problem for decades [5].
V. C ONCLUSION
We presented simulation results to evaluate the network
performance of UL IoT UDNs. From our study, we can see that
for a low-reliability criterion, the density of reliably working
UEs grows quickly with the network densification. However,
in our journey to realize a more reliable UL IoT UDNs, we
should be aware of several caveats:
• First, for a high-reliability criterion, the density of reliably
working UEs remains low in UDNs due to excessive
inter-cell interference, which should be considered when
operating UL IoT UDNs.
•
•
Second, due to the existence of a non-zero antenna height
difference between BSs and UEs, the density of reliably
working UEs could even decrease as we deploy more
BSs. This calls for further study of UL IoT UDNs.
Third, well-planned hexagonal-like BS deployments can
improve network performance for relatively sparse networks, but not for UDNs, showing that alternative solutions other than BS position optimization should be
considered in the future UL IoT UDNs.
R EFERENCES
[1] A. Al-Fuqaha, M. Guizani, M. Mohammadi, M. Aledhari, and
M. Ayyash, “Internet of things: A survey on enabling technologies,
protocols, and applications,” IEEE Communications Surveys Tutorials,
vol. 17, no. 4, pp. 2347–2376, Fourthquarter 2015.
[2] D. López-Pérez, M. Ding, H. Claussen, and A. Jafari, “Towards 1
Gbps/UE in cellular systems: Understanding ultra-dense small cell
deployments,” IEEE Communications Surveys Tutorials, vol. 17, no. 4,
pp. 2078–2101, Jun. 2015.
[3] A. Fotouhi, M. Ding, and M. Hassan, “Understanding autonomous drone
maneuverability for Internet of Things applications,” pp. 1–6, Jun. 2017.
[4] M. Ding, D. López-Pérez, H. Claussen, and M. A. Kâafar, “On the
fundamental characteristics of ultra-dense small cell networks,” to
appear in IEEE Network Magazine, arXiv:1710.05297 [cs.IT], Oct.
2017. [Online]. Available: http://arxiv.org/abs/1710.05297
[5] M. Ding, D. López-Pérez, G. Mao, and Z. Lin, “Microscopic analysis
of the uplink interference in FDMA small cell networks,” IEEE Trans.
on Wireless Communications, vol. 15, no. 6, pp. 4277–4291, Jun. 2016.
[6] J. Andrews, F. Baccelli, and R. Ganti, “A tractable approach to coverage
and rate in cellular networks,” IEEE Transactions on Communications,
vol. 59, no. 11, pp. 3122–3134, Nov. 2011.
[7] 3GPP, “TR 36.872: Small cell enhancements for E-UTRA and EUTRAN - Physical layer aspects,” Dec. 2013.
[8] M. Ding and López-Pérez, “Performance impact of base station antenna
heights in dense cellular networks,” to appear in IEEE Transactions
on Wireless Communications, arXiv:1704.05125 [cs.NI], Sep. 2017.
[Online]. Available: https://arxiv.org/abs/1704.05125
[9] X. Zhang and J. Andrews, “Downlink cellular network analysis with
multi-slope path loss models,” IEEE Transactions on Communications,
vol. 63, no. 5, pp. 1881–1894, May 2015.
[10] M. Ding, P. Wang, D. López-Pérez, G. Mao, and Z. Lin, “Performance
impact of LoS and NLoS transmissions in dense cellular networks,”
IEEE Transactions on Wireless Communications, vol. 15, no. 3, pp.
2365–2380, Mar. 2016.
[11] S. Lee and K. Huang, “Coverage and economy of cellular networks with
many base stations,” IEEE Communications Letters, vol. 16, no. 7, pp.
1038–1040, Jul. 2012.
[12] M. Ding, D. López-Pérez, G. Mao, and Z. Lin, “Performance impact of
idle mode capability on dense small cell networks with LoS and NLoS
transmissions,” to appear in IEEE Transactions on Vehicular Technology,
arXiv:1609.07710 [cs.NI], Sep. 2017.
[13] 3GPP, “TR 36.828: Further enhancements to LTE Time Division Duplex
for Downlink-Uplink interference management and traffic adaptation,”
Jun. 2012.
[14] T. Bai and R. Heath, “Coverage and rate analysis for millimeter-wave
cellular networks,” IEEE Transactions on Wireless Communications,
vol. 14, no. 2, pp. 1100–1114, Feb. 2015.
[15] Spatial Channel Model AHG, “Subsection 3.5.3, Spatial Channel Model
Text Description V6.0,” Apr. 2003.
[16] I. Gradshteyn and I. Ryzhik, Table of Integrals, Series, and Products
(7th Ed.). Academic Press, 2007.
[17] M. Ding, M. Zhang, D. López-Pérez, and H. Claussen, “Correlated
shadow fading for cellular network system-level simulations with wraparound,” 2015 IEEE International Conference on Communications
(ICC), pp. 2245–2250, Jun. 2015.
[18] T. Ding, M. Ding, G. Mao, Z. Lin, D. López-Pérez, and A. Y. Zomaya,
“Uplink performance analysis of dense cellular networks with LoS and
NLoS transmissions,” IEEE Transactions on Wireless Communications,
vol. 16, no. 4, pp. 2601–2613, Apr. 2017.
[19] M. Ding and H. Luo, Multi-point Cooperative Communication Systems:
Theory and Applications. Springer, 2013.
| 7cs.IT
|
1
On the Support Recovery of Jointly Sparse
Gaussian Sources using Sparse Bayesian Learning
Saurabh Khanna, Student Member, IEEE and Chandra R. Murthy, Senior Member, IEEE
arXiv:1703.04930v3 [cs.IT] 26 Oct 2017
Dept. of ECE, Indian Institute of Science, Bangalore, India
{sakhanna, cmurthy}@iisc.ac.in
Abstract—In this work, we provide non-asymptotic, probabilistic guarantees for successful sparse support recovery by
the multiple sparse Bayesian learning (M-SBL) algorithm in
the multiple measurement vector (MMV) framework. For joint
sparse Gaussian sources, we show that M-SBL perfectly recovers
their common nonzero support with arbitrarily high probability
using only finitely many MMVs. In fact, the support error
probability decays exponentially fast with the number of MMVs,
with the decay rate depending on the restricted isometry property
of the self Khatri-Rao product of the measurement matrix.
Our analysis theoretically confirms that M-SBL is capable of
recovering supports of size as high as O(m2 ), where m is the
number of measurements per sparse vector. In contrast, popular
MMV algorithms in compressed sensing such as simultaneous
orthogonal matching pursuit and row-LASSO can recover only
O(m) sized supports. In the special case of noiseless measurements, we show that a single MMV suffices for perfect recovery
of the k-sparse support in M-SBL, provided any k + 1 columns
of the measurement matrix are linearly independent. Unlike
existing support recovery guarantees for M-SBL, our sufficient
conditions are non-asymptotic in nature, and do not require the
orthogonality of the nonzero rows of the joint sparse signals.
Index Terms—Sparse Signal Processing, Sparse Bayesian
Learning, Joint Sparsity, Restricted Isometry Property, KhatriRao product, Statistical Learning Theory.
I. I NTRODUCTION
In recent years, joint sparsity has emerged as one of the most
important and versatile signal structures in the field of sparse
signal processing. Two or more vectors are said to be jointly
sparse if their nonzero coefficients belong to the same index
set, i.e., they share a common nonzero support. Joint sparsity
arises naturally in multi-modal or multi-channel analysis of
signals residing in low dimensional signal subspaces. The underlying joint sparsity can be exploited to resolve ambiguities
which may arise due to erroneous estimation of the support of
the individual sparse vectors from noisy measurements. This
idea has been exploited in several practical scenarios such as
MIMO channel estimation [1]–[3], distributed source coding
[4], [5], multi-task compressive sensing [6], distributed event
localization [7], array signal processing [8], and cooperative
spectrum sensing [9]–[11].
In the sparse signal recovery literature, the estimation of
jointly sparse signals is referred to as the multiple measurement
vector (MMV) problem [12] where the signal of interest is a
matrix X ∈ Rn×L whose columns are jointly sparse vectors
in Rn . As a result, X is a row sparse matrix with only a
fraction of its rows containing nonzero elements and the rest
of the rows made up entirely of zeros. In the MMV problem,
the goal is to recover X from its noisy, linear measurements
Y ∈ Rm×L . The measurement matrix Y (each column is
called a single measurement vector (SMV)) is generated as
Y = AX + W,
(1)
where A ∈ Rm×n is a known measurement matrix and W ∈
Rm×L is the unknown noise matrix. For m < n, the above
linear system is under-determined and therefore has infinitely
many solutions for X. However, if A satisfies certain restricted
isometry properties, a unique row-sparse solution can still be
guaranteed [12]–[15].
In many applications, the performance of a joint sparse
signal recovery algorithm is judged on the basis of how
accurately it can identify the true support or the locations of
the nonzero rows of X. This gives rise to the joint sparse
support recovery (JSSR) problem where the goal is to recover
the row support of X given Y and A. Interestingly, unlike the
nonzero coefficients in a k-sparse X which can be recovered
only if m ≥ k, the nonzero support can be recovered even
from m < k measurements. In fact, for i.i.d. Gaussian entries
in both A and X, a non-iterative, correlation based algorithm
called One Step Greedy Algorithm (OSGA) [16] is capable of
recovering the true support using only m ≥ 1 measurement
per signal with probability approaching one as L → ∞.
A majority of the existing MMV algorithms [12], [14], [17]–
[21] implicitly assume that the number of nonzero rows in
X is less than the number of measurements per signal, i.e,
k < m. Currently, M-SBL [22], Co-LASSO [23] and RDCMP [24] are the only MMV algorithms which have been
theoretically or empirically shown to be capable of recovering
a k-sparse support from fewer than k measurements per signal.
However, a theoretical understanding of when these algorithms
are guaranteed to be successful is far from complete. In
this work, we derive non-asymptotic guarantees for M-SBL’s
performance in the JSSR problem. Specifically, we derive the
sufficient conditions for exact support recovery in terms of the
number of MMVs and properties of A.
M-SBL’s support recovery performance has been investigated in [25] for k < m, and in [22] for k ≥ m. In both studies, the orthogonality of the nonzero rows of X is identified
as one of the sufficient conditions for exact support recovery.
For finite L, the row-orthogonality condition is too restrictive
for a deterministic X and almost never true for a continuously
distributed random source. Thus, the existing support recovery
guarantees for M-SBL are in reality only applicable in the
asymptotic sense when L → ∞. Furthermore, the earlier analysis is restricted only to noiseless measurements. In contrast,
the new sufficient conditions derived in this work are nonasymptotic in nature and do not assume row orthogonality
2
in the signal matrix X. Our analysis also accounts for the
presence of measurement noise. While our focus here is on
sufficient conditions for exact support recovery, the necessary
conditions have been partly covered in [26] and [27].
A. Existing Theoretical Guarantees for Support Recovery
The earliest theoretical work focused on seeking guarantees
for a unique joint sparse solution to the canonical ℓ0 norm
minimization problem:
L0 :
min
X∈Rm×L
||X||0
s.t. AX = Y,
(2)
i.i.d. N (0, diag(1S )), where S denotes the unknown
support
set. For m = Ω k log nk , it is shown that L ≥ O logloglogn n
suffices for diminishing support error probability with increasing L. One of our contributions in this paper is to extend this
result to a more general signal prior on X and show that the
support error probability vanishes even if m scales sublinearly
in the support size k.
In [32], the support recovery problem is analyzed as a
single-input-multi-output MAC communication problem. For
n
number of nonzero rows fixed to k, m = log
c(X) is shown to be
both necessary and sufficient for successful support recovery
as the problem size tends to infinity. The quantity c(X) is
a capacity like term which depends on the elements of the
nonzero rows in X and the noise power. Even fewer measurements m = Ω( Lk log n) suffices when each measurement
vector is generated using a different measurement matrix [33].
In the above discussion, in particular about algorithmspecific support recovery guarantees, the number of nonzero
rows in X is typically assumed to be less than m, the number
of measurements per MMV. In the following subsection, we
review the covariance matching framework, which is capable
of recovering supports of size k greater than m.
where ||X||0 denotes the number of nonzero rows in X.
In [12], [13], the authors showed that for spark1 (A) = m + 1
and rank(Y) ≤ m, the L0 problem admits a unique k-sparse
solution if k < ⌈(spark(A) − 1 + rank(Y)) /2⌉. This result
confirmed that the SMV bottleneck of k < m/2 for ℓ0 norm
based support recovery is not applicable when multiple measurement vectors are used. Furthermore, the sparsity bound
suggests that supports of size k < m are potentially uniquely
recoverable. Since this result, numerous JSSR algorithms have
been proposed in the quest to meet the k < m sparsity bound.
To circumvent the combinatorial hardness of the L0 problem, [17] proposed minimizing the mixed-norm ℓp,q norm
of X instead of the ℓ0 norm. Variants of the ℓp,q norm
minimization problem with different combinations of p and
q have been investigated independently in several works [12],
[13], [28], [29]. For p ≥ 1, q = 1, [13] has shown that ℓp,q
norm minimization problem has a unique k-sparse solution,
< 1, for all j ∈
/ S and for
provided A satisfies A†S aj
1
−1 T
†
T
all S ⊂ [n], |S| ≤ k, where AS = AS AS
AS . This
also serves as a sufficient condition for exact support recovery
in Simultaneous Orthogonal Matching Pursuit (SOMP) [14],
a greedy support reconstruction algorithm. In [18], support
recovery performance of various correlation based greedy and
iterative hard-thresholding type algorithms is studied in the
noiseless MMV setup. The sufficient conditions for exact
support recovery are specified in terms of the asymmetric
restricted isometry constants of the measurement matrix A.
A limitation of the aforementioned support recovery techniques and associated guarantees is that only supports of size
up to k < m/2 are uniquely recoverable. In [30], rank aware
OMP and rank aware order recursive matching pursuit are
proposed and analyzed, and it is shown that both these algorithms perfectly recover any k-sparse support from noiseless
measurements as long as k < spark(A) − 1 and rank(X) = k.
For the rank defective case, i.e., rank(X) < k, compressed
sensing MUSIC [20] and subspace-augmented MUSIC [19]
are still capable of recovering any k < spark(A) − 1 sized
support as long as k−rank(X) partial support can be estimated
by another sparse signal recovery algorithm.
In [31], the support recovery problem is formulated as a
multiple hypothesis testing problem. Necessary and sufficient
conditions for perfect support recovery with high probability
are derived under the assumption that the columns of X are
In (3), the linear constraints are the vector form of the
second order moment constraints, i.e., the covariance matching
equation: L1 YYT = Adiag(γ)AT + σ 2 Im , where σ 2 Im
denotes the noise covariance matrix. Since the constraints in
(3) comprise up to (m2 +m)/2 linearly independent equations,
sparsity levels as high as O(m2 ) are potentially recoverable.
To recover the maximum level of sparsity, k = (m2 + m)/2,
a necessary condition derived in [23], [27] is that the columnwise self Khatri-Rao product matrix A ⊙ A must have full
Kruskal rank,2 i.e., Krank(A ⊙ A) = (m2 + m)/2.
The M-SBL algorithm [25], our focus in this paper, also
imposes a common Gaussian prior N (0, diag(γ)) on the
columns of X and hence implicitly exploits the latent uncorrelatedness of the nonzero entries in X. Interestingly, similar
to Co-LASSO, the support recovery performance of M-SBL is
closely related to the properties of the self Khatri-Rao product
1 Spark of a matrix A is the smallest integer p such that there exist p
linearly dependent columns in A.
2 The Kruskal rank of an m × n matrix A is the largest integer k such that
any k columns of A are linearly independent.
B. Covariance Matching Based Support Recovery
A key insight was propounded in [23], that there often exists
a latent structure in the MMV problem: the nonzero entries
of X are uncorrelated. This signal structure can be enforced
by assuming each column of X to be i.i.d. N (0, diag(γ)),
where γ ∈ Rn+ is a non-negative vector of variance parameters.
Under this source model, identifying the nonzero rows of X
is tantamount to estimating the support of γ. In [23], the
Co-LASSO algorithm was proposed for the recovery of γ.
Instead of working directly with the linear observations, CoLASSO uses their covariance form, L1 YYT , as input, and γ
is recovered by solving the following non-negative ℓ1 norm
minimization problem:
1
YYT − σ 2 Im . (3)
minn ||γ||1 s.t. (A ⊙ A)γ = vec
γ∈R+
L
3
A ⊙ A. Making this relation explicit is one of the main
contributions of this paper.
C. Contributions
Following are the main contributions of this work.
1) A new interpretation of the M-SBL algorithm as a
Bregman matrix divergence minimization problem, which
opens up new avenues to exploit the vast literature on
Bregman divergence minimization for devising faster,
more robust algorithms towards minimizing the M-SBL
cost function.
2) New sufficient conditions under which M-SBL exactly
recovers the true support of the joint sparse vectors
with high probability in the noisy MMV problem. The
conditions are specified in terms of the number of MMVs
and the properties of the measurement matrix.
3) The support error probability in M-SBL is theoretically
shown to decay exponentially with the number of MMVs.
The error exponent is related to the restricted isometry
property of A ⊙ A, the self Khatri-Rao product of A.
Explicit upper bounds on the number of MMVs sufficient
for vanishing support error probability in M-SBL for both
noisy as well as noiseless measurements are derived.
A key feature of our analysis is that our sufficient conditions
are expressed in terms of the restricted isometry property (RIP)
constants of the measurement matrix A and that of its selfKhatri-Rao product A ⊙ A. This makes our results applicable
to both random as well as deterministic constructions of A.
The remainder of the paper is organized as follows. In
section II, we formulate the JSSR problem and introduce our
signal model for X. We also review the M-SBL algorithm
[25] and interpret the M-SBL cost function as a Bregman
matrix divergence. In section III, we cover some preliminary
concepts which are used while analyzing the support recovery
performance of M-SBL. In section IV, we derive an abstract
upper bound for the support error probability, which is used
in section V to derive our main result, namely, the sufficient
conditions for vanishing support error probability in M-SBL.
In section VI, we discuss the implications of the new results
in the context of several interesting special cases. Our final
conclusions are presented in section VII.
D. Notation
Throughout this paper, scalar variables are denoted by
lowercase alphabets and vectors are denoted by boldface lowercase alphabets. Matrices are denoted by uppercase boldface
alphabets and uppercase calligraphic alphabets denote sets.
Given a vector x, x(i) represents its ith entry. supp(x)
denotes the support of x, the set of indices corresponding
to nonzero entries in x. Likewise, R(X) denotes the set of
indices of all nonzero rows in X and is called the rowsupport of X. For any n ∈ N, [n] , {1, 2, . . . , N }. For any
n dimensional vector x and index set S ⊆ [n], the vector
xS is an |S| × 1 sized vector retaining only those entries
of x which are indexed by elements of S. Likewise, AS
is a submatrix comprising the columns of A indexed by S.
Null(A) and Col(A) denote the null space and column space
of the matrix A, respectively. The spectral, Frobenius and
maximum absolute row sum matrix norms of A are denoted
by |||A|||2 , ||A||F , and |||A|||∞ , respectively. P(E) denotes
the probability of event E. N (µ, Σ) denotes the Gaussian
probability density with mean µ and covariance matrix Σ.
For any square matrix C, tr(C) and |C| denote its trace and
n
determinant, respectively. Lastly, S++
denotes the set of all
n × n positive definite matrices.
II. S YSTEM M ODEL AND M-SBL A LGORITHM
A. Joint Sparse Support Recovery (JSSR)
Suppose x1 , x2 , . . . , xL are L distinct joint-sparse vectors
in Rn with a common nonzero support denoted by the index
set S ∗ ⊆ [n]. Let K be the maximum size of the common
support, i.e., |S ∗ | ≤ K. In JSSR, we are interested in recovering S ∗ from noisy underdetermined linear measurements
y1 , y2 , . . . , yL generated as
yj = Axj + wj , 1 ≤ j ≤ L.
(4)
The measurement matrix A ∈ Rm×n is assumed to be a
non-degenerate matrix, with m ≤ n. By non-degeneracy
of A, it is implied that any m columns of A are linearly
independent, or spark(A) = m + 1. The noise vector w ∈ Rm
is assumed to be zero mean Gaussian distributed with diagonal
covariance matrix σ 2 Im . The measurement matrix A and
the noise variance σ 2 are assumed to be known. The linear
measurement model in (4) can be rewritten in a compact
MMV form as Y = AX + W, where Y = [y1 , y2 , . . . , yL ],
X = [x1 , x2 , . . . , xL ] and W = [w1 , w2 , . . . , wL ] are the
observation, signal and noise matrices, respectively. Since the
columns of X are jointly sparse, with common support S ∗ , X
is a row sparse matrix with row support R(X) = S ∗ .
B. Source Signal Model
We assume that if the ith row of the unknown signal matrix
X nonzero, then it is a Gaussian ensemble of L i.i.d. zero mean
random variables with a common variance γ ∗ (i) belonging
to the interval [γ min , γ max ]. In the sequel, we refer to this as
Assumption (A1). An immediate consequence of (A1) is that
there exists a bounded, nonnegative, and at most K sparse
vector, γ ∗ ∈ Rn+ , such that the columns xj are i.i.d. N (0, Γ∗ )
with Γ∗ , diag(γ ∗ ). Furthermore, R(X) and supp(γ ∗ ) are
the same and equal to S ∗ .
C. M-SBL Algorithm
In this section, we review the M-SBL algorithm [25], a
type-II maximum likelihood (ML) procedure for estimation
of joint sparse signals from compressive linear measurements. In M-SBL, the columns of X are assumed to be
i.i.d. N (0, Γ) distributed, where Γ = diag(γ), and γ =
T
[γ(1), γ(2), . . . , γ(n)] is an n × 1 vector of non-negative
variance parameters. The elements of γ are collectively called
hyperparameters as they represent the parameters of the signal
prior. Since the hyperparameter γ(i) models the common
variance of the ith row of X, if γ(i) = 0, it drives the posterior
4
estimate of xj (i) to zero for 1 ≤ j ≤ L. Consequently, if γ
is a sparse vector, it induces joint sparsity in X.
In M-SBL, the hyperparameter vector γ is chosen to maximize the Bayesian evidence p(Y; γ), which is tantamount to
finding the maximum likelihood (ML) estimate of γ. Let γ̂ ML
denote the ML estimate of γ, i.e.,
γ̂ ML = arg max log p(Y; γ).
(5)
γ∈Rn
+
The Gaussian prior on xj combined with the linear measurement model induces Gaussian observations, i.e., p(yj ; γ) =
N (0, σ 2 Im + AΓAT ). For a fixed γ, the MMVs yj are
mutually independent. Hence, it follows that
log p(Y; γ) =
L
X
log p(yj ; γ)
j=1
2
T
,
∝ −L log |Σγ | − Tr Σ−1
γ YY
(6)
T
where Σγ = σ Im + AΓA . The log likelihood log p(Y; γ)
in (6) is a non convex function of γ and its global maximizer
γ̂ ML is not available in closed form. However, its local
maximizer can still be found via fixed point iterations or the
Expectation-Maximization (EM) procedure.
In [25], it is empirically shown that the EM procedure
faithfully recovers the true support S ∗ , provided m and L
are sufficiently large. In this paper, we derive new sufficient
conditions for supp (γ̂ ML ) = S ∗ , i.e., for perfect support
recovery via maximization of M-SBL’s log-likelihood cost.
Σγ , with point wise distances measured using the Bregman
LogDet divergence. Thus, the M-SBL algorithm, at its core,
is essentially a second order moment matching or covariance
matching procedure which selects γ such that the associated
covariance matrix Σγ is closest to the sample covariance
matrix, in the Bregman LogDet divergence sense. In a later
section, we theoretically show that if the second moment
matching equations are too ill-conditioned, then M-SBL fails
to recover the true support of the joint sparse columns of X.
This new interpretation of the M-SBL cost as a Bregman
matrix divergence beckons two interesting questions:
i Are there other matrix divergences besides LogDet Bregman matrix divergence which are better suited for covariance matching?
ii How to exploit the structural similarities between the MSBL cost and the Bregman (LogDet) matrix divergence
to devise faster and more robust techniques for the type-2
maximum likelihood procedure?
We believe that evaluating the performance of other matrix
divergences for covariance matching is a worthwhile exercise
to pursue and can lead to development of new, improved
algorithms for the JSSR problem. Our preliminary results in
this direction can be found in [24].
III. P RELIMINARY C ONCEPTS
In this section, we introduce a few key definitions and
results which are used in the later sections.
D. The M-SBL cost is a Bregman Matrix Divergence
We now present an interesting interpretation of M-SBL’s
log-marginalized likelihood cost in (6) which facilitates a
deeper understanding of what is accomplished by its maximization. We begin by introducing the Bregman matrix divergence Dϕ (X, Y) between any two n × n positive definite
matrices X and Y as
Dϕ (X, Y) , ϕ(X) − ϕ(Y) − h∇ϕ(Y), X − Yi,
(7)
n
where ϕ : S++
→ R is a convex function with ∇ϕ(Y)
as its first order derivative evaluated at point Y. In (7), the
matrix inner product hX, Yi is evaluated as tr XYT . For the
specific case of ϕ(·) = − log | · |, a strongly convex function,
we obtain the Bregman LogDet matrix divergence given by
(8)
Dlogdet (X, Y) = tr XY−1 − log XY−1 − n.
By termwise comparison of (6) and (8), we observe that the
negative log likelihood − log p(Y; γ) and Dlogdet (RY , Σγ ) are
the same up to a constant. In fact, it is shown in [34], [35] that
every regular exponential family of probability distributions is
associated with a unique Bregman divergence.
In the divergence term Dlogdet (RY , Σγ ), the first argument
RY , L1 YYT is the sample covariance matrix of the
observations Y and the second argument Σγ = σ 2 I + AΓAT
is the parameterized covariance matrix of Y. This connection between M-SBL’s log likelihood cost and the LogDet
divergence reveals that by maximizing the M-SBL cost, we
seek a γ which minimizes the distance between RY and
A. Restricted Isometry Property
A matrix A ∈ Rm×n is said to satisfy the restricted
isometry property (RIP) of order k if there exists a constant
δkA ∈ (0, 1) such that
2
2
2
(1 − δkA ) ||x||2 ≤ ||Ax||2 ≤ (1 + δkA ) ||x||2
(9)
holds for any k-sparse vector x ∈ Rn . The smallest such δkA
is called the k th order restricted isometry constant (RIC) of A.
B. ǫ-Cover, ǫ-Net and Covering Number
Suppose T is a set equipped with a pseudo-metric d. For
any set A ⊆ T , its ǫ-cover is defined as the coverage of A with
open balls of radius ǫ and centers in T . The set Aǫ comprising
the centers of these covering balls is called an ǫ-net of A. The
minimum number of ǫ-balls which can cover A is called the
ǫ-covering number of A, and is given by
ǫ
Ncov
(A, d) = min {|Aǫ | : Aǫ is an ǫ-net of A}
(10)
In computational theory of learning, ǫ-net constructs are often
useful in converting a union over the elements of a continuous
set to a finite sized union.
Proposition 1 ( [36]). Let B(0, 1) be a unit ball in Rn centered
at 0. Then, its ǫ-covering number with respect to the standard
n
ǫ
Euclidean metric is bounded as Ncov
(B(0, 1), ||·||2 ) ≤ (3/ǫ) .
5
C. Strong Convexity of − log det
A differentiable function f is called strongly convex with
strong convexity constant mf > 0 if the following inequality
holds for all points x, y in the domain.
mf
2
f (y) ≥ f (x) + h∇f (x), y − xi +
||y − x||2 ,
(11)
2
where ||·||2 is the Frobenius norm.
m
Proposition 2. Let MK ⊆ S++
be the collection of ma2
trices of form Σγ = σ I + AΓAT , where Γ = diag(γ),
γ ∈ Rn+ , ||γ||0 ≤ K, and A ∈ Rm×n . The matrix function ψ(·) = − log det(·) is strongly convex in the domain
MK with strong convexity constant mψ given by mψ =
−1
supγ∈Rn+ ,||γ||0 ≤2K |||Σγ ◦ Σγ |||2
, where ◦ denotes the
Hadamard product.
Proof. See Appendix A.
D. Rényi Divergence
Let (X , F ) be a measurable space and P and Q be two
probability measures on F with densities p and q, respectively,
with respect to the dominating Lebesgue measure µ on F .
For α ∈ R+ \1, the Rényi divergence of order α between P
and Q, denoted Dα (p||q), is defined as
Z
1
p(x)α q(x)1−α µ(dx).
(12)
log
Dα (p||q) =
α−1
X
For p = N (0, Σ1 ) and q = N (0, Σ2 ), the α-Rényi
divergence Dα (p||q) is available in closed form [37],
Dα (p||q) =
1
|(1 − α)Σ1 + αΣ2 |
.
log
1−α
α
2(1 − α)
|Σ1 |
|Σ2 |
(13)
E. A Lower Bound on the α-Rényi Divergence Between Multivariate Gaussian Distributions
The strong convexity of − log det(·) can be used to derive a
lower bound on the Rényi divergence between two multivariate
Gaussian distributions, stated below as Theorem 1.
Theorem 1. Let p1 and p2 be two multivariate Gaussian
distributions with zero mean and positive definite covariance
matrices Σ1 and Σ2 , respectively. Then, for α ∈ [0, 1), the
α-Rényi divergence between p1 and p2 satisfies
1
2
(14)
Dα (p1 , p2 ) ≥ αm∗ ||Σ2 − Σ1 ||F ,
4
where m∗ is the strong convexity constant of the matrix
function ψ(Σ) = − log |Σ| over the domain Θ(Σ1 , Σ2 ), the
collection of all convex combinations of Σ1 and Σ2 .
Proof. See Appendix B.
We now proceed with the analysis of support recovery performance of M-SBL algorithm. We derive an M-SBL specific
Chernoff bound for the support error probability in the JSSR
problem under assumption (A1). We begin by introducing
some of the frequently used notation in the following table.
K
Maximum number of nonzero rows in X
Sk
Collection of all support sets of k or lesser
size, i.e., Sk = {S ⊆ [n], |S| ≤ k}
Θ(S) A bounded hyperparameter set associated with
the
set S, formally defined as Θ(S) ,
support
γ ∈ Rn+ : supp(γ) = S, γ min γ S γ max .
ΘK
Collection of all nonnegative K or less sparse
vectors in Rn+ whose nonzeroSelements belong
to [γ min , γ max ]. Also, ΘK = S∈SK Θ(S).
S∗
True row support of X.
γ∗
Principal diagonal of the common covariance
matrix Γ∗ of the i.i.d. columns in X.
Also, supp(γ ∗ ) = S ∗ .
By assumption (A1) on X, we have γ ∗ ∈ ΘK . Thus, in
order to recover γ ∗ from Y, we consider solving a constrained
version of the M-SBL optimization problem in (5):
C
ǫ2
log 2δ .
γ̂ = arg max L(Y; γ).
cM-SBL(K):
(16)
γ∈ΘK
In (16), the objective L(Y; γ) , log p(Y; γ) is the same as
the M-SBL log-likelihood cost in (6). The row support of X
is then estimated as supp(γ̂), where γ̂ is the solution of (16).
We define the set of bad MMVs,
(17)
ES ∗ , Y ∈ Rm×L : supp (γ̂) 6= S ∗ ,
which result in erroneous estimation of the true support. We
are interested in identifying the conditions under which P(ES ∗ )
can be made arbitrarily small.
Note that ES ∗ is the collection of undesired MMVs for
which the log-likelihood objective is globally maximized by
γ ∈ Θ(S), for S 6= S ∗ . Based on this interpretation, ES ∗ can
be rewritten as
[
Y : max L(Y; γ) ≥ ′ max ∗ L(Y; γ ′ ) .
ES ∗ =
S∈SK \{S ∗ }
Since
γ∈Θ(S)
γ ∈Θ(S )
(18)
max ∗ L(Y; γ ′ ) ≥ L(Y; γ ∗ ), we have
′
γ ∈Θ(S )
ES ∗ ⊆
[
S∈SK \{S ∗ }
=
F. Concentration of Sample Covariance Matrix
Proposition 3 (Vershynin [38]). Let y1 , y2 , . . . , yL ∈ Rm
beP
L independent samples from N (0, Σ), and let ΣL =
L
1
T
j=1 yj yj denote the sample covariance matrix. Then, for
L
any ǫ > 0,
|||ΣL − Σ|||2 ≤ ǫ |||Σ|||2 ,
(15)
holds with probability exceeding 1 − δ provided L ≥
Here, C is an absolute constant.
IV. S UPPORT E RROR P ROBABILITY A NALYSIS
[
S∈SK
\{S ∗ }
Y : max L(Y; γ) ≥ L(Y; γ ∗ )
γ∈Θ(S)
[
{Y : L(Y; γ) − L(Y; γ ∗ ) ≥ 0} .(19)
γ∈Θ(S)
The continuous union over infinitely many elements of Θ(S)
in (19) can be relaxed to a finite sized union by using the
following ǫ-net argument.
Consider Θǫ (S), a finite sized ǫ-net of the hyperparameter
set Θ(S), such that for any γ ∈ Θ(S), there exists an element
γ ′ ∈ Θǫ (S) so that |L(Y; γ) − L(Y; γ ′ )| ≤ ǫ. Proposition 4
gives an upper bound on the size of such an ǫ-net.
6
Proposition 4. Given a support set S ⊆ [n], there exists a
finite set Θǫ (S) ⊂ Θ(S) such that it simultaneously satisfies
(i) For any γ ∈ Θ(S), there exists a γ ′ ∈ Θǫ (S) such that
|L(Y; γ) − L(Y;γ ′ )| ≤ ǫ.
|S|
p
ǫ
(ii) |Θ (S)| ≤ max 1, 3CL,S (γ max − γ min ) |S|/ǫ
,
where CL,S is the Lipschitz constant of L(Y; γ) with
respect to γ in the bounded domain Θ(S).
The set Θǫ (S) is an ǫ-net of Θ(S).
Proof. See Appendix C.
From Proposition 4-(ii), we observe that the construction
of Θǫ (S) depends on the Lipschitz continuity of the loglikelihood L(Y; γ) with respect to γ over the domain Θ(S).
By virtue of the data-dependent nature of L(Y; γ), its Lipschitz constant CL,S depends on the instantaneous value of
Y. To make the rest of the analysis independent of Y, we
introduce a new MMV set G such that CL,S can be uniformly
bounded solely in terms of the statistics of Y. A possible
choice of G could be
1
T
T
m×L
≤ 2E y1 y1
. (20)
YY
G, Y⊂R
:
L
2
For Y ∈ G, L(Y; γ) is uniformly Lipschitz continuous
irrespective of Y, and hence the ǫ-net Θǫ (S) can now be
constructed entirely independent of Y.
Since for arbitrary sets A and B, A ⊆ (A ∩ B) ∪ B c , the
RHS in (19) relaxes as
[
[
{L(Y; γ) − L(Y; γ ∗ ) ≥ 0} ∩ G ∪ G c .
ES ∗ ⊆
∗
S∈SK \S γ∈Θ(S)
(21)
Let Θǫ (S)|G denote an ǫ-net of Θ(S) constructed under the
assumption that Y ∈ G. Then, the continuous union over Θ(S)
relaxes to a finite sized union over Θǫ (S)|G as shown below.
[
[
ES ∗ ⊆
{L(Y; γ)−L(Y; γ ∗ ) ≥ −ǫ} ∩ G ∪ G c
S∈SK \S ∗ γ∈Θǫ (S)|G
[
[
{L(Y; γ) − L(Y; γ ∗ ) ≥ −ǫ} ∪ G c .
⊆
ǫ
∗
S∈SK \S
γ∈Θ (S)|G
By applying the union bound, we obtain
X
X
P (L(Y; γ)−L(Y; γ ∗ ) ≥ −ǫ)
P (ES ∗ ) ≤ P (G c )+
S∈SK \S ∗ γ∈Θǫ (S)|G
(22)
From (22), P(ES ∗ ) will be small if each of the constituent
probabilities P (L(Y; γ) − L(Y; γ ∗ ) ≥ −ǫ), γ ∈ Θǫ (S)|G are
sufficiently small so that their collective contribution to the
double summation in (22) remains small. In Theorem 2, we
show that each event within the summation is a large deviation
event which occurs with an exponentially decaying probability.
Theorem 2. Let pγ (y) denote the marginal probability density
of the columns of Y induced by the joint sparse X with
columns drawn independently from N (0, diag(γ)). Then, the
P
log-likelihood L(Y; γ) = L
j=1 log pγ (yj ) satisfies the following large deviation property.
P (L(Y; γ) − L(Y; γ ∗ ) ≥ −ǫ) ≤ exp (−Lψ ∗ (−ǫ/L)),
(23)
where ψ ∗ (·) is the Legendre transform3 of ψ(t) , (t −
1)Dt (pγ , pγ ∗ ), and Dt is the t-Rényi divergence (of order
t > 0) between the probability densities pγ and pγ ∗ .
Proof. See Appendix D.
Note that, when the measurement noise is Gaussian, the
marginal density pγ (yj ) of the individual observations is
also Gaussian with zero mean and covariance matrix Σγ =
σ 2 Im +AΓAT . If σ 2 > 0, both marginals pγ and pγ ∗ are nondegenerate and hence the Rényi divergence term Dt (pγ , pγ ∗ )
in Theorem 2 is well defined. We now restate Theorem 2 as
Corollary 1, which is the final form of the large deviation
result for L(Y; γ) that will be used in bounding P(ES ∗ ).
Corollary 1. For an arbitrary γ ∈ Rn+ , and the true variance
parameters γ ∗ , let the associated marginal densities pγ and
pγ ∗ be as defined in Theorem 2, and suppose σ 2 > 0. Then, the
log-likelihood L(Y; γ) satisfies the large deviation property
P L(Y; γ) − L(Y; γ ∗ ) ≥ −LD1/2 (pγ , pγ ∗ )/2
≤ exp −LD1/2 (pγ , pγ ∗ )/4 . (24)
Proof. The
large deviation result is obtained by replacing
ψ ∗ − Lǫ in Theorem 2 by its lower bound − tǫ
L − ψ (t),
followed by setting t = 1/2 and ǫ = LD1/2 (pγ , pγ ∗ )/2.
Note that, in the above, we have used the sub-optimal
choice t = 1/2 for the Chernoff parameter t, since its
optimal value is not available in closed form. However, this
suboptimal selection of t is inconsequential as it figures only
as a multiplicative constant in the final sample complexity. By
using Corollary 1 in (22), we can upper bound P (ES ∗ ) as
X
∗
κcov exp −LD1/2
(S)/4 + P (G c ) , (25)
P(ES ∗ ) ≤
S∈SK \S ∗
∗
D1/2
(S)
where
is the minimum value of D1/2 (pγ , pγ ∗ ) across
all ǫ-net points γ in Θǫ (S)|G . A new constant κcov is introduced in (25) which denotes the maximum size of the ǫ-nets
Θǫ (S)|G , S ∈ SK , i.e.,
κcov , max Θǫ (S)|G .
S∈SK
(26)
∗
Proposition 5 presents a lower bound for D1/2
(S) for S ∈ SK .
Proposition 5. Let pγ denote the parameterized multivariate
Gaussian density with zero mean and covariance matrix Σγ =
σ 2 I + AΓAT , Γ = diag(γ). For any pair γ 1 , γ 2 ∈ ΘK such
that supp(γ 1 ) differs from supp(γ 2 ) in exactly kd locations,
the α-Rényi divergence between pγ 1 and pγ 2 satisfies
2
⊙
α
kd (1 − δ2K
)
γ min
,
Dα pγ 1 , pγ 2 ≥
T
2
2
4 σ + γ max Ka∗ sup
AS AS 2
S:|S|=2K
3 For any convex function f : X → R on a convex set X ⊆ Rn , its
Legendre transform is the function f ∗ defined by
f ∗ (z) = sup (hz, xi − f (x)) .
x∈X
7
⊙
where δ2K
is the 2K th order restricted isometry constant of
the self Khatri-Rao product A ⊙ A, and a∗ = ||vec(A)||∞ .
Proof. See Appendix E.
Using Proposition 5 with α = 1/2 in (25), we obtain
X
∗
η
log κcov
exp −LkdS,S
P(ES ∗ ) ≤
+ P (G c ) ,
−
4
L
∗
S∈SK \S
with η =
1
8
(27)
γ min
σ 2 + γ max
2
Ka2∗
⊙
1 − δ2K
ATS AS
sup
S:|S|=2K
.
(28)
2
Theorem 3. Suppose S ∗ is the true row support of the
unknown X satisfying assumption A1 and with |S ∗ | ≤ K.
Then, for any δ ∈ (0, 1), P(ES ∗ ) ≤ 2δ, if
8
2
1+δ
8
, log κcov , C log
,
log 3enK
L ≥ max
η
δ
η
δ
(29)
where κcov and η are as defined in (26) and (28), respectively, and C > 0 is a universal numerical constant.
c
Proof. Since L ≥ C log (2/δ), by Proposition 3, P(G ) ≤ δ.
Combined with L ≥ 8 log κcov /η, (27) can be rewritten as
X
∗
exp −ηLkdS,S /8 + δ.
P(ES ∗ ) ≤
(30)
S∈SK \S ∗
The total number of support sets belonging to SK \S ∗
∗
which differ
the true support
S in exactly kd locaPkd from
j+|S ∗ |
n−|S ∗ |
tions is j=0
, which is further upper
j
min(kd ,j+|S ∗ |)
k
bounded by (3enK) d . Thus, we can rewrite (30) as
P(ES ∗ )
≤ δ +
K+|S ∗ |
X
kd =1
∗
≤ δ+
K+|S |
X
X
ηLkd
exp −
8
S∈SK \S ∗ ,
|(S\S ∗ )∪(S ∗ \S)|=kd
kd
(3enK)
kd =1
ηL kd
e− 8
.
P(ES ∗ )
≤ δ+
X
kd =1
δ
1+δ
kd
by noting that for all S ⊆ [n], |S| = 2K,
q
ATS AS 2 ≤ tr AS ATS AS ATS = vec(AS ATS ) 2
q
√
⊙
= (AS ⊙ AS )1|S| 2 ≤ 1 + δ2K
2K. (33)
B. Upper Bound for κcov
∗
By setting ǫ = LD1/2
/2 in Proposition 4, we have
K
√
K
κcov ≤ max 1, 6 K(γ max − γ min )CL /ηL
,
(31)
≤ δ + δ = 2δ.
In Theorem 3, we finally have an abstract bound on the
sufficient number of MMVs, L, which guarantee vanishing
support error probability in cM-SBL(K), for a fixed true
support S ∗ . The MMV bound is meaningful only when η is
strictly positive. We now proceed to show that by imposing
certain isometry conditions on the self Khatri-Rao product
A ⊙ A, the following can be ensured.
i. η > 0 (a positive support error exponent),
ii. η and κcov scale favorably with the system dimensions.
(34)
where CLK = maxS∈SK CL,S denotes the Lipschitz constant
of L(Y; γ) with respect to γ over the bounded domain ΘK .
Proposition 6 characterizes the Lipschitz property of L(Y, γ).
Proposition 6. For S ∈ SK , the log-likelihood L(Y; γ) :
Θ(S) → R is Lipschitz continuous in γ as shown below.
|||RY |||2
LK
|L(Y,γ 2 )−L(Y,γ 1 )| ≤
1+
||γ 2 − γ 1 ||2 ,
γ min
σ2
for any γ 1 , γ 2 ∈ Θ(S). Here, RY ,
1
T
L YY .
Proof. See Appendix F.
Under 2K-RIP compliance of A ⊙ A√and for Y ∈ G
(as
⊙ 1/2
defined in (20)), |||RY |||2 ≤ 2 σ 2 +γ max K(1+δK
)
. By
using Proposition 6 in (34), combined with the lower bound
for η in (32) yields the following upper bound for κcov .
K
7/2 2
2
, (35)
κcov ≤ max 1, K a∗ ζ(γ max − γ min )/σ
√ 1+δ⊙ σ2 +γ 2 3σ2 +2γ
max
max
.
where ζ = 48 2 1−δ2K
⊙
γ
γ
2K
, P(ES ∗ ) can be upper
Since L ≥ η8 log 3enK 1+δ
δ
bounded by a geometric series as
K+|S ∗ |
A. Lower Bound for η
From (28), it is evident that if the self Khatri-Rao product
⊙
A ⊙ A satisfies the RIP condition: δ2K
< 1, then η is strictly
positive. Furthermore, η can be lower bounded as
2
⊙
1 − δ2K
γ min
1
√
q
,
(32)
η≥
2
⊙
8 2K 3/2 a2∗ σ + γ max
1 + δ2K
min
min
V. S UFFICIENT C ONDITIONS FOR S UPPORT R ECOVERY
Equipped with explicit bounds for η and κcov , we now
state the sufficient conditions for vanishing support error
probability in cM-SBL(K) as the following theorem.
Theorem 4. Under assumption (A1), suppose X has row
support S ∗ , |S ∗ | ≤ K. Let γ̂ be the solution of cMSBL(K). Then, supp (γ̂) = S ∗ with probability exceeding
1 − 2δ, for any δ ∈ (0, 1), provided
C1. A is non-degenerate, and A ⊙ A satisfies the RIP
⊙
of order 2K, i.e., δ2K
< 1.
C2. The number of MMVs, L, satisfies
L ≥ K 3/2 a2∗ ξ max {log (3enK(1 + δ)/δ) , log κcov }
2
2
⊙ 1/2
√
σ + γ max
(1 + δ2K
)
where ξ = 64 2
,
⊙
γ min
1 − δ2K
8
a∗ = ||vec(A)||∞ and κcov is bounded as in (35).
Proof. C1 ensures that η in (28) is strictly positive. Further,
C2 ensures that the abstract MMV bound in (29) is satisfied.
Therefore, by Theorem 3, P(ES ∗ ) ≤ 2δ.
In the following corollary, we state an extra condition
besides C1 and C2 which extends the above cM-SBL result
to the M-SBL cost function.
B. Continuous versus Binary Hyperparameters
In [31], Tang and Nehorai formulated support recovery
using MMVs as a multiple hypothesis testing problem by
assuming that each column of X is i.i.d. N (0, diag(1S ∗ )),
where S ∗ is the k-sparse support of X. For this choice of
signal prior, finding the true support via type-2 likelihood
maximization as in (5) is no longer a continuous variable
optimization but rather a combinatorial search over all k-sparse
n
vertices of the hypercube {0, 1} , as described below.
γ̂ =
Corollary 2. For the same setting as Theorem 4, let γ̂ be
the output of M-SBL optimization in (5). If γ̂ ∈ ΘK , and
the conditions C1 and C2 hold for a given δ ∈ (0, 1),
then supp(γ̂) = S ∗ with probability exceeding 1 − 2δ.
Proof. Since γ̂ belongs to ΘK and also maximizes the MSBL objective log p(Y; γ), it follows that γ̂ is a solution to
the constrained cM-SBL(K) optimization in (16). Hence, the
statement of Corollary 2 follows directly from Theorem 4.
According to Corollary 2, M-SBL has vanishing support error
probability under conditions C1 and C2, however only in a
retrospective sense, i.e., when the M-SBL output γ̂ belongs to
the bounded hyperparameter set ΘK .
So far, we have analyzed the support error probability for the
∗
case where the true row support of X
is nfixed
to S n. In reality,
n
R(X) can assume any one of the 1 + 2 +· · ·+ K possible
supports in the collection SK . The support error probability
averaged over all possible supports can be evaluated as
X
err
=
Pavg
P(R(X) = S ∗ )P (ES ∗ ) ,
(36)
S ∗ ∈SK
where ES ∗ is as defined in (17). For example, if all supports
err
in SK are equiprobable, then Pavg
≤ 2δ under C1 and C2.
VI. D ISCUSSION
In this section, we interpret the sufficiency conditions C1
and C2 in Theorem 4 in the context of various interesting
cases.
A. Column Normalized Measurement Matrices
According to C2 in Theorem 4, the number of sufficient
MMVs grows quadratically with a∗ , the largest absolute value
of entries in A. If A has unit norm columns, then
a∗ scales
inversely with the number of rows in A as O √1m . For
example, if A is a random matrix with i.i.d. N (0, √1m ) entries,
q
log n
with high probability. Likewise, if
then |a∗ | ≤ O
m
A is a partial DFT sensing matrix, then |a∗ | = √1m .
Often, column normalization is essential for accurate modelling of the norm-preserving nature of the application specific
measurement modality. In light of this, if A has normalized
columns, then one must account for O( √1m ) scaling of a∗
while interpreting the sufficient number of MMVs in C2.
arg max
log p (Y; γ).
(37)
γ∈{0,1}n ,||γ||0 =k
Binary valued hyperparameters can be accommodated as a
special case by setting γ min = γ max = 1 in our source signal
model. For γ min = γ max , according to Proposition 4, the ǫnet Θǫ (S) collapses to a single point for all S ∈ SK , which
ultimately amounts to κcov = 1. Thus, for binary γ and column
normalized A, by setting κcov = 1 and a∗ = √1m in C2, we
can conclude that the support error probability decays
3/2 exponentially fast with increasing L provided L ≥ O K mlog n .
For continuous valued γ, i.e. γ(i) ∈ [γ min , γ max ] ∪ {0},
the number
MMVs
must scale as L ≥
3/2of sufficient
5/2
O max K mlog n , K mlog K .
C. On the Restricted Isometry of A⊙A and Support Recovery
Guarantees for m ≤ K
After L surpasses the MMV threshold in C2, the support
error probability is guaranteed to decay exponentially with L,
provided η is strictly positive. The positivity of η (as defined
in (28)) is contingent upon the self Khatri-Rao product A ⊙ A
⊙
satisfying the 2K-RIP condition: δ2K
< 1.
Suppose A is an m × n sized random matrix comprising
i.i.d. subgaussian entries with zero mean and variance equal to
√1 . Then, according to [39, Theorem 3], the self Khatri-Rao
m
⊙
product A ⊙ A satisfies the RIP condition: δ2K
high
≤√ t with
K log3/2 n
.
probability for any t ∈ (0, 1) provided m ≥ O
t
Thus,
for
a
subgaussian
A,
it
is
sufficient
for
m
to
scale
as
√
O( K) in order to guarantee vanishing error probability for
M-SBL based reconstruction of all K or less sparse supports.
Furthermore, in [39, Theorem 3], it is also shown that the
self Khatri-Rao
A ⊙ A satisfies K-RIP for K as
2 product
m
high as O log
.
This
implies that for a fixed n, O(m2 )
3n
sized supports can be perfectly recovered by M-SBL with high
probability using only finitely many MMVs. The exact number
of sufficient MMVs is given by C2. An order-wise sufficient
MMV condition is reported in Table I.
D. Support Recovery from Noiseless Measurements
For K < spark(A) − 1, it can be shown that as the noise
variance σ 2 → 0, the support error exponent D1/2 (pγ , pγ ∗ ) in
(24) grows in an unbounded fashion for all γ ∈ ΘK \Θ(S ∗ ).
This is formally proved below.
The 1/2-Rényi divergence between two multivariate Gaussian densities pγ i (y) = N (0, Σγ i ), i = 1, 2 is given by
D1/2 pγ 1 , pγ 2
=
log
Σγ 1 + Σγ 2
1
− log Σγ 1 Σγ 2
2
2
9
=
1
2
1
1
log 0.5 H 2 + H− 2 .
(38)
1
2
where H , Σγ 1 Σ−1
γ 2 Σγ 1 is referred to as the discrimination
matrix. Since H is a normal matrix, it is unitarily diagonalizable. Let H = UΛUT , where Λ = diag(λ1 , . . . , λm ) with
λi ’s being the strictly positive eigenvalues of H, and U being
a unitary matrix with the eigenvectors of H as its columns.
The 1/2-Rényi divergence can be expressed in terms of λi as
m
X
−1/2
1/2
/2
log λi + λi
=
D1/2 pγ 1 , pγ 2
i=1
1
(λmax (H))1/2 + (λmax (H))−1/2 . (39)
≥ log
2
The above inequality is obtained by dropping all positive terms
in the summation except the one term which corresponds to
λmax (H), the maximum eigenvalue of H. Proposition 7 below
relates λmax (H) to the noise variance σ 2 .
Proposition 7. If K < spark(A) − 1, then, for any pair
γ 1 , γ 2 ∈ ΘK such that supp(γ 1 )\supp(γ 2 ) 6= φ, the maxi1
1
c1
2
mum eigenvalue of H , Σγ2 1 Σ−1
γ 2 Σγ 1 satisfies λmax (H) ≥ σ2
2
for some constant c1 > 0 independent of σ .
Proof. See Appendix G.
According to Proposition 7, in the limit σ 2 → 0,
λmax (H) → ∞, and consequently, D1/2 pγ 1 , pγ 2 grows
unbounded (due to (39)) whenever supp(γ 1 ) 6= supp(γ 2 ) and
K < spark(A) − 1. Based on this observation, we now state
Theorem 5 which lays forward the sufficient conditions for
exact support recovery in the noiseless case.
Theorem 5. Consider the noiseless MMV problem, with
observations Y = AX corresponding to an unknown
X satisfying assumption (A1). Suppose S ∗ is the true
nonzero row support of X with |S ∗ | ≤ K. Then, if a
solution γ̂ to the M-SBL optimization problem in (5)
satisfies γ̂ ∈ ΘK , then supp(γ̂) = S ∗ almost surely,
provided that K < spark(A) − 1. This result holds even
in the SMV case, i.e., when L = 1.
Proof. Under assumption A1, there exists a γ ∗ ∈ ΘK
such that every column in X is i.i.d. N (0, diag(γ ∗ )), and
supp(γ ∗ ) = S ∗ . Since γ̂ globally maximizes the M-SBL
objective L(Y; γ), it follows that L(Y; γ̂) ≥ L(Y; γ ∗ ) if
γ̂ 6= γ ∗ . Moreover, the following chain of implications holds.
{supp(γ̂) 6= S ∗ } =
⊆
{supp(γ̂) 6= supp(γ ∗ )} ⊆ {γ̂ 6= γ ∗ }
{L (Y; γ̂) ≥ L (Y; γ ∗ )} .
By applying Corollary 1, this further implies that
P (supp(γ̂) 6= S ∗ ) ≤ P (L (Y; γ̂) ≥ L (Y; γ ∗ ))
≤ exp −LD1/2 (pγ̂ , pγ ∗ ) /4 .
By using the lower bound (39) for D1/2 (pγ̂ , pγ ∗ ), we have
"
!#− L4
1 p
1
∗
P (supp(γ̂) 6= S ) ≤
λmax (H) + p
,
2
λmax (H)
(40)
1/2
1/2
∗
where H = Σγ̂ Σ−1
γ ∗ Σγ̂ . Since both γ̂ and γ belong to
ΘK , as long as K < spark(A) − 1, by Proposition 7, σ 2 → 0
results in λmax (H) → ∞ which in turn drives the RHS in (40)
to zero for L ≥ 1.
From Theorem 5, we conclude that, in the noiseless scenario
and for X satisfying assumption A1, M-SBL requires only a
single measurement vector (L = 1) to perfectly recover any
K < spark(A)−1 sized support. This result is in line with the
sufficient conditions identified for successful support recovery
by M-SBL in [25, Theorem 1]. However, unlike in [25], the
nonzero rows of X need not be orthogonal. Also, our result
improves over the k ≤ m/2 condition shown in [31].
E. Impact of Measurement Noise on Sufficient MMVs
∗
For σ 2 > 0, the error exponent term D1/2
in (25) is always
bounded. This implies that unlike in the noiseless case, a single
MMV is no longer sufficient, and multiple MMVs are needed
to drive the error probability close to zero.
A close inspection of the abstract MMV bound in (29)
reveals that the noise variance influences the error probability
in a twofold manner: (i) through η, and (ii) through the ǫ-net
cardinality bound κcov . As σ 2 increases, η decreases polynomially (see (28)) while log κcov increases at most logarithmically
(see (35)). The overall effect is captured by condition C2
in Theorem 4, which suggests that if the noise variance is
very high relative to γmax , it is sufficient to have a roughly
quadratically larger number of MMVs to guarantee the desired
probability of error. As the noise variance approaches zero, the
MMV bound in C2 loosens and is not informative.
F. Support Recovery using RIP Compliant A
If the measurement matrix A is itself 2K-RIP compliant,
then the support error probability vanishes for fewer MMVs
than specified by C2. In [31], the authors showed that for 2KRIP
compliant
A, and binary hyperparameters γ in M-SBL,
O logloglogn n MMVs suffice for error probability to vanish.
Here, we consider the general case of continuous valued γ.
⊙
A 2
If A has unit norm columns, then δ2K
≤ (δ2K
) by [39,
Theorem 1], which results in tighter bounds for η and κcov :
2
A 2
1
γ min
1 − (δ2K
)
, (41)
η ≥
A
8Ka2∗ σ 2 + γ max
1 + δ2K
(
K )
K 5/2 a2∗ (γ max − γ min )ζ ′
κcov ≤ max 1,
, (42)
A )
γ min (1 − δ2K
2
A
1+δK
(σ +γ max )3
where ζ ′ = γ144
2
A
σ2 +γ min . By using (41) and (42) in
1−δ
min
K
Theorem 3, we obtain the following condition on the number
of MMVs to guarantee P(ES ∗ ) ≤ 2δ:
2
3enK(1+δ)
64Ka2∗ σ 2 +γ max
, log κcov ,
max log
L≥
A )
γ min
δ
(1−δ2K
(43)
where κcov is bounded as in (42).
From (43), we conclude that for A with unit norm columns,
P(ES ∗ ) can be driven arbitrarily close to zero if L exceeds
10
TABLE I
S UFFICIENT CONDITIONS FOR EXACT SUPPORT RECOVERY OF K OR LESS
SPARSE SUPPORT IN M-SBL WHEN ITS OUTPUT γ̂ ∈ ΘK .
General A with
unit norm columns
√
|a∗ | = O(1/ m)
A < 1
δ2K
Sufficient number of MMVs
Binary γ
σ2
=0 L≥1
σ2 > 0 L ≥ O
⊙
δ2K
<1
σ2 ≥ 0 L ≥ O
A with i.i.d.
N (0, √1m ) entries
m = Ω(K log n)
Noise
Noise
Continuous γ
K log n
m
3
K 2 log n
m
L≥1
L≥O
L≥O
K 2 log K
m
5
K 2 log K
m
Binary γ
Continuous γ
L≥1
σ2 > 0 L ≥ O (log n)
L ≥ O (K log K)
2
√
√
3
log K
m = O( K log 2 n) σ2 ≥ 0 L ≥ O K log n L ≥ O K 3/2
log
n
n K 2 log K
,
. For A with i.i.d. N (0, √1m )
O max K log
m
m
entries, the sufficient MMV condition is reported in Table I.
VII. C ONCLUSIONS
In this work, we presented new sufficient conditions for
exact support recovery by M-SBL in the joint sparse support
recovery problem. The sufficient conditions are expressed in
terms of the number of MMVs and the restricted isometry
constants of the measurement matrix and its self Khatri-Rao
product. The new conditions cater to a wider, more useful
class of Gaussian signals, and dispenses with the restrictive
row-orthogonality condition on the signal matrix required in
the previous results [22], [25].
We have shown that M-SBL is capable of recovering ksparse support from fewer than k measurements per MMV
for both noiseless and noisy measurements. In the noiseless
case, a single MMV suffices for perfect recovery of any ksparse support, provided that spark(A) > k + 1. In case of
noisy measurements, M-SBL can perfectly recover any k-sized
support with vanishing error probability using finitely many
⊙
MMVs, provided that A⊙A satisfies the RIP condition: δ2k
<
⊙
1. For random subgaussian A, the RIP condition
δ
<
1
is
2k
m2
[39],
satisfied with high probability for k as high as O log
3n
which suggests that M-SBL can successfully recover O(m2 )
sized supports for fixed n.
There still remain the following open questions about MSBL’s support recovery performance:
(i) What are necessary conditions for exact support recovery
in terms of the number of required MMVs?
(ii) Under what conditions do all local maxima of the M-SBL
objective yield the correct support estimate?
(iii) How is the support recovery performance impacted by
inter and intra vector correlations in the signals?
A PPENDIX
A. Proof of Proposition 2
Let Σγ 1 , Σγ 2 ∈ MK . From Taylor’s expansion of
ψ(Σγ 2 ) = − log det Σγ 2 around Σγ 1 , we have
ψ(Σγ 2 )
= ψ(Σγ 1 ) + h∇ψ(Σγ 1 ), Σγ 2 − Σγ 1 i
∇2 ψ(Σγ t )[Σγ 2 − Σγ 1 ]
(44)
with Σγ t , tΣγ 1 + (1 − t)Σγ 2 for some t ∈ [0, 1].
2
The
Hessian operator ∇ ψ(Σγ t )[Σγ 2− Σγ 1 ] is evaluated as
−1
tr Σ−1
γ t (Σγ 2−Σγ 1 )Σγ t (Σγ 2−Σγ 1 ) . To prove the strong
convexity of ψ over Mk , one needs to show the existence
of a positive constant mψ such that
Sufficient number of MMVs
σ2 = 0 L ≥ 1
+
∇2 ψ(Σγ t )[Σγ 2 − Σγ 1 ] ≥ mψ Σγ 2 − Σγ 1
2
F
(45)
for all Σγ 1 , Σγ 2 ∈ MK . Note that Σγ t = σ 2 Im + AΓt AT ,
where Γt = diag(γ t ) and γ t = tγ 1 + (1 − t)γ 2 is at most
2K-sparse. Further, let S denote the nonzero support of ∆γ =
γ 2 − γ 1 , and let ∆Γ = diag(∆γ). Then, we have Σγ 2 −
Σγ 1 = AS ∆ΓS ATS . Using these new definitions, the Hessian
term in (44) can be lower bounded as
−1
)
−
Σ
)Σ
(Σ
−
Σ
tr Σ−1
(Σ
γ
γ
γ
γ
γt
γt
1
2
1
2
−1
T −1
= tr Σγ t AS ∆ΓS AS Σγ t AS ∆ΓS ATS
(a)
T
T −1
= vec (∆ΓS ) ATS Σ−1
γ t AS ⊗ AS Σγ t AS vec (∆ΓS )
T
(b)
−1
−1
−1
−1
= ∆γ TS Σγ t2 AS ⊙ Σγ t2 AS
Σγ t2 AS ⊙ Σγ t2 AS ∆γ S
(c)
T −1
= ∆γ TS ATS Σ−1
A
◦
A
Σ
A
S
S ∆γ S
γt
S γt
−1
(d)
T
(AS ◦ AS ) ∆γ S
= ∆γ TS (AS ◦ AS ) Σγ t ◦ Σγ t
T
T
≥ ∆γ S (AS ◦AS ) (AS ◦AS ) ∆γ S / Σγ t◦Σγ t 2
2
2
AS ∆ΓS ATS F
||(AS ⊙ AS ) ∆γ S ||2
=
Σγ t ◦ Σγ t 2
Σγ t ◦ Σγ t 2
−1
≥
sup
|||Σγ ◦Σγ |||2
Σγ 2−Σγ 1
(e)
≥
γ∈Rn
+ ,||γ||0 ≤2K
2
F
. (46)
In above, steps (a) and (d) follow from facts 7.4.9 and 8.21.49
in [40], respectively. Step (b) is due to Columnwise KhatroRao product (⊙) being a submatrix of the Kronecker product
(⊗). Step (c) is a consequence of the property: (A⊙A)T (A⊙
A) = AT A ◦ AT A satisfied by any matrix A. Step (e) is due
to Hadamard product (◦) being a submatrix of the columnwise
Khatri-Rao product matrix (⊙).
−1
From (46), mψ = supγ∈Rn+ ,||γ||0 ≤2K |||Σγ ◦Σγ |||2
>0
is a valid choice for the strong convexity constant of ψ over
domain MK , thereby establishing the Proposition.
B. Proof of Theorem 1
From Proposition 2, − log det(·) is a strongly convex function in Θ(Σ1 , Σ2 ) with strong convexity constant m∗ . Any
strongly convex function f : X → R with strong convexity
constant mf satisfies tf (x)+(1−t)f (y) ≥ f (x + (1 − t)y)+
0.5mf t(1 − t)kx − yk22 for t ∈ [0, 1] and x, y ∈ X . Thus, for
Σ1 , Σ2 ∈ Θ(Σ1 , Σ2 ),
− log |(1 − α)Σ1 + αΣ1 | ≤ −(1 − α) log |Σ1 | − α log |Σ2 |
1
2
− m∗ α(1 − α) ||Σ2 − Σ1 ||F .
2
11
By rearranging terms, we obtain
1
1
|(1 − α)Σ1 + αΣ1 |
≥ m∗ α ||Σ2 − Σ1 ||2F .
log
1−α
α
2(1 − α)
4
|Σ1 |
|Σ2 |
The left hand side in the above equation is precisely
Dα (p1 , p2 ) Thus, we have the desired result.
C. Proof of Proposition 4
The following stepwise procedure shows how to construct a
δ-net of Θ(S) (with respect to the Euclidean distance metric)
which is entirely contained in Θ(S).
1) Consider an δ-blow up of Θ(S), denoted by Θ↑δ (S) ,
{x : ∃x′ ∈ Θ(S) such that kx − x′ k2 ≤ δ}.
2) Let Θδ↑δ (S) be a δ-net of Θ↑δ (S). Some points in Θδ↑δ (S)
may lie outside Θ(S).
3) Let P denote the set containing the projections of all
points in Θδ↑δ (S) ∩ Θ(S)c onto the set Θ(S). By conδ
c
struction, P ⊂ Θ(S),
and |P| ≤ |Θ
↑δ (S) ∩ Θ(S) |.
4) Then, Θδ (S) , Θδ↑δ (S) ∩ Θ(S) ∪ P is a valid δ-net
of Θ(S) which is entirely contained in Θ(S).
To prove the validity of the above δ-net construction, we
need to show that for any γ ∈ Θ(S), there exists an element a
in Θδ (S) such that kγ−ak2 ≤ δ. Let γ be an arbitrary element
in Θ(S). Then, γ also belongs to the larger set Θ↑δ (S), and
consequently, there exists γ ′ ∈ Θδ↑δ (S) such that kγ − γ ′ k2 ≤
′
δ. Now, there are two
and (ii) γ ′ ∈
/ Θ(S).
cases. (i) γ ∈ Θ(S),
′
δ
In case (i), γ ∈ Θ↑δ (S) ∩ Θ(S) , and hence also belongs
to Θδ (S). Further, kγ − γ ′ k2 ≤ δ. Hence a = γ ′ will work.
In case (ii), γ ′ ∈ Θδ↑δ (S) ∩ Θ(S)c . Let γ ′′ be the projection
of γ ′ onto Θ(S), then γ ′′ must belong to P, and hence must
also belong to Θδ (S). Note that since γ ′′ is the projection
of γ ′ onto the convex set Θ(S), for any γ ∈ Θ(S), we have
hγ − γ ′′ , γ ′ − γ ′′ i ≤ 0. Further, we have
δ
≥
=
≥
2
2
||γ − γ ′ ||2 = ||(γ − γ ′′ ) + (γ ′′ − γ ′ )||2
2
2
||γ − γ ′′ ||2 + ||γ ′′ − γ ′ ||2 + 2hγ − γ ′′ , γ ′′ − γ ′ i
2
||γ − γ ′′ ||2 .
(47)
The last inequality is obtained by dropping the last two
nonnegative terms in the RHS. From (47), a = γ ′′ will work.
Since case (i) and (ii) together are exhaustive, Θδ (S) in
step-4 is a valid δ-net of Θ(S) which is entirely inside Θ(S).
δ
pCardinality of Θ (S): The diameter of Θ(S) is
|S|(γ max − γ min ). Based on the construction in step4, the cardinality of Θδ (S) can be upper bounded as:
|Θδ (S)|
= |Θδ↑δ (S) ∩ Θ(S)| + |Θδ↑δ (S) ∩ Θ(S)c |
= |Θδ↑δ (S)| ≤ |δ-net of ℓ2 ball of radius
p
|S|(γ max − γ min ) in R|S|
|S|
p
. (48)
≤ max 1, 3 |S|(γ max − γ min )/δ
The last step is an extension of the volumetric arguments in
[36] to show that the δ-covering number of a unit ball B1 (0) in
Rk with respect to the standard Euclidean norm ||·||2 satisfies
δ
Ncov
(B1 (0), ||·||2 ) ≤ (3/δ)k . The max operation with unity
covers the case when δ is larger than the diameter of Θ(S).
Now consider the modified net Θǫ/CL,S (S) obtained by
ǫ
setting δ = CL,S
in steps 1-4, where CL,S is the Lipschitz
constant of L(Y, γ) with respect to γ ∈ Θ(S). We claim
ǫ/C
that ΘS L,S is the desired set which simultaneously satisfies
conditions (i) and (ii) stated in Proposition 4.
ǫ/C
To show condition (i), we observe that since ΘS L,S is
an (ǫ/CL,S )-net of Θ(S) with respect to ||·||2 , for any γ ∈
Θ(S), there exists a γ ′ ∈ Θǫ/CL,S (S) such that ||γ − γ ′ ||2 ≤
ǫ/CL,S . Since L(Y, γ) is CL,S -Lipschitz in Θ(S), it follows
that |L(Y, γ) − L(Y, γ ′ )| ≤ CL,S ||γ − γ ′ ||2 ≤ ǫ.
Condition (ii) follows from (48) by setting δ = ǫ/CL,S .
D. Proof of Theorem 2
For continuous probability densities pγ and pγ ∗ defined on
the observation space Rm , the tail probability of the random
variable log (pγ (Y)/pγ ∗ (Y)) has a Chernoff upper bound
with parameter t > 0 as shown below.
L
X
pγ (Y)
p
(y
)
γ
j
P log
≥ −ǫ
= P
log
≥ −ǫ
∗ (yj )
pγ ∗ (Y)
p
γ
j=1
L
X
pγ (yj )
exp (tǫ)
log
≤ Epγ ∗ exp t
p
γ ∗ (yj )
j=1
L
pγ (y)
exp (tǫ)
= Epγ ∗ exp t log
pγ ∗ (y)
"
t #!L
pγ (y)
exp (tǫ)
= Epγ ∗
pγ ∗ (y)
Z
L
=
pγ (y)t pγ ∗ (y)1−t dy
exp (tǫ)
y∈Y
i
h ǫ
− (t − 1)Dt (pγ , pγ ∗ ) .(49)
= exp −L t −
L
In the above, the first and third steps follow from the independence of yj . The second step is the application of Chernoff
bound. The last step is obtained by using the definition of the
Rényi divergence and rearranging the terms in the exponent.
By introducing the function ψ(t) = (t − 1)Dt (pγ , pγ ∗ ), the
Chernoff bound (49) can be restated as
i
h ǫ
pγ (Y)
− ψ(t) .
≥ −ǫ ≤ exp −L t −
P log
pγ ∗ (Y)
L
(50)
For t = arg supt>0 t − Lǫ − ψ(t) , the upper bound in (50)
attains its tightest value exp (−Lψ ∗ (−ǫ/L)), where ψ ∗ is the
Legendre transform of ψ.
E. Proof of Proposition 5
∗
2
Σγ 2 − Σγ 1 F .
From Theorem 1, Dα pγ 1 , pγ 2 ≥ αm
4
In the following arguments, we derive separate lower bounds
2
for m∗ and Σγ 2 − Σγ 1 F .
2
The term Σγ 2 − Σγ 1 F can be lower bounded as:
Σγ 2 − Σγ 1
2
F
= A (Γ2 − Γ1 ) AT
2
F
12
⊙
= ||(A ⊙ A) (γ 2 − γ 1 )||22 ≥ (1 − δ2K
)kd γ 2min .(51)
−1
≤ L |||RY |||2 tr ATS Σ−1
γ Σγ AS
−1/2
−1/2
−1
= L |||RY |||2 tr ΓS ÃTS Σ−1
γ Σγ ÃS ΓS
−1
≤ (L |||RY |||2 /γ min ) tr ÃTS Σ−1
γ Σγ ÃS
T −1
tr
Ã
Σ
Ã
≤ L |||RY |||2 Σ−1
/γ
S
S
γ
γ
min
2
(a)
≤ L |||RY |||2 Σ−1
min
(m,
|S|)
/γ min
γ
2
The last inequality is a consequence of A⊙A satisfying RIP of
2
⊙
order 2K with RIC δ2K
and the fact that ||γ 2 −γ 1 ||2 ≥ kd γ 2min .
Next, we derive a lower bound for m∗ . We note that
(σ 2 Im + AΓAT ) ◦ (σ 2 Im + AΓAT )
|||Σγ ◦ Σγ |||2 =
≤ σ 4 + 2σ 2 γ max
+ γ 2max
AS ATS ◦ Im
AS ATS ◦ AS ATS
2
2
2
,
(52)
where S = supp(γ) and |S| ≤ 2K. The last step follows from
the triangle inequality. Let a∗ denote the maximum absolute
value of elements in A. Then, it can be shown that
AS ATS ◦ Im
AS ATS
◦
AS ATS
2
2
≤
≤
a2∗ K,
a2∗ K
(53)
ATS AS
2
. (54)
Using (53) and (54) in (52), and by Proposition 2, we have
−1
2
ATS AS 2
.
(55)
m∗ ≥ Ka2∗ σ 2 + γ max
Finally, we get the desired result by using Theorem 1, and
2
combining the lower bounds for m∗ and Σγ 2 − Σγ 1 F in
(55) and (51), respectively.
F. Proof of Proposition 6
The log-likelihood L(Y; γ) can be expressed as the sum
f (γ) + g(γ) with f (γ) = −L log |Σγ | and g(γ) =
2
T
−Ltr Σ−1
γ RY . Here, Σγ = σ Im + AΓA .
First, we derive an upper bound for the Lipschitz constant of
f (γ) = −L log |Σγ | for γ ∈ Θ(S). By mean value theorem,
any upper bound for k∇γ f (γ)k2 also serves as an upper bound
for the Lipschitz constant of f . So, we derive an upper bound
(γ)
for k∇γ f (γ)k2 . Note that ∂f
= L(ai Σ−1
γ ai ) for i ∈ S,
γ(i)
th
and 0 otherwise. Here, ai denotes the i column of A. Then,
k∇γ f (γ)k2 can be upper bounded as shown below.
X
k∇γ f (γ)k2 ≤ k∇γ f (γ)k1 = L
aTi Σ−1
γ ai
u∈S
= L tr ATS (σ 2 Im + AS ΓS ATS )−1 AS
−1/2
−1/2
= L tr ΓS ÃTS (σ 2 Im + ÃS ÃTS )−1 ÃS ΓS
(a)
tr ÃTS (σ 2 Im + ÃS ÃTS )−1 ÃS
= L Γ−1
S
2
(b)
≤ (L min (m, |S|)) /γ min .
1/2
AS ΓS .
(56)
In the above, step (a) follows from
where ÃS =
the trace inequality tr(AB) ≤ k|Ak|2 tr(B) for any positive
definite matrices A and B. Step (b) follows from the observation that input argument of trace operator has min (m, |S|)
nonzero eigenvalues, all of them less than unity.
We now shift focus to the second term g(γ) of the loglikeT −1
−1
lihood. Note that ∂g(γ)
∂γ(i) = L(ai Σγ RY Σγ ai ) for i ∈ S,
and 0 otherwise. Then, k∇γ g(γ)k2 can be upper bounded as
X
−1
k∇g(γ)k2 ≤ k∇γ g(γ)k1 = L
aTi Σ−1
γ RY Σγ ai
= L tr
i∈S
−1
ATS Σ−1
R
Σ
Y
γ
γ AS
(b)
≤ (L |||RY |||2 min (m, |S|)) /γ min σ 2 .
(57)
1/2
where ÃS , AS ΓS . The inequality in (57a) follows from
ÃTS Σγ−1 ÃS having min (m, |S|) nonzero eigenvalues, all
of them less than unity. The last inequality in (57b) is due to
Σ−1
≤ 1/σ 2 . Finally, the Lipschitz constant CL,S can
γ
2
be bounded as CL,S ≤ k∇γ f (γ)k2 + k∇γ g(γ)k2 . Thus, by
combining (56) and (57), and noting that min(m, |S|) ≤ K,
we obtain the desired result.
G. Proof of Proposition 7
1
1
2
Let µ∗ be the largest eigenvalue of Σγ2 1 Σ−1
γ 2 Σγ 1 . Then,
i
tr Σγ−1
Σ
γ
1
1 h 2 −1
2
T
µ∗ ≥
σ Tr Σγ 2 + Tr Σ−1
=
γ 2 AΓ1 A
m
m
1 −1
(58)
tr Σγ 2 AΓ1 AT .
≥
m
Here, the second step is setting Σγ 1 = σ 2 Im + AΓ1 AT . The
last inequality
is obtained by dropping the strictly positive
σ2
−1
tr
Σ
term.
γ2
m
Let S1 and S2 be the nonzero supports of γ 1 and γ 2 , respectively. Further, let the eigendecomposition of Σγ 2 be UΛUT ,
where Λ = diag(λ1 , . . . , λm ), λi ’s are the eigenvalues of Σγ 2
and U is a unitary matrix with columns as the eigenvectors
of Σγ 2 . Then, U can be partitioned as [U2 U2⊥ ], where the
columns of U2 and U2⊥ span the orthogonal complementary
subspaces Col(AS2 ) and Col(AS2 )⊥ , respectively. Further, let
Λ2 and Λ2⊥ be |S2 | × |S2 | and ((m − |S2 |) × (m − |S2 |))
sized diagonal matrices containing the eigenvalues in Λ corresponding to the eigenvectors in U2 and U2⊥ , respectively.
We observe that Λ2⊥ = σ 2 Im−|S2 | .
−1 T
−1 T
By setting Σ−1
γ 2 = U2 Λ2 U2 + U2⊥ Λ2⊥ U2⊥ in (58), we
get
1
T
T
tr U2 Λ−1
+ tr U2⊥ Λ−1
UT
µ∗ ≥
2 U2 AΓ1 A
2⊥ 2⊥
m
1
T
T
AΓ1 AT ≥ tr Λ−1
⊥ U2⊥ AΓ1 A U2⊥ ,
2
m
where the last inequality is due to nonnegativity of the dropped
first term. Since UT2⊥ AS2 = 0 by construction of U2⊥ ,
1
T
T
c
c
µ∗ ≥ tr Λ−1
⊥ U2⊥ AS2 Γ1,S2 AS c U2⊥
2
2
m
m−|S2 |
1 X
(u ⊥ )T AS2c Γ1,S2c ATS2c u2⊥ ,i
=
mσ 2 i=1 2 ,i
=
m−|S2 |
1 X
(u ⊥ )T AS1 \S2 Γ1,S1 \S2 ATS1 \S2 u2⊥ ,i . (59)
mσ 2 i=1 2 ,i
13
In the above, u2⊥ ,i denotes the ith column of U2⊥ . The last
equality is obtained by observing that the nonzero elements of
γ 1,S2c are located in the index set S1 \S2 .
We now prove that if K < spark(A) − 1, then there exists
at least one strictly positive term in the above summation. Let
us assume the contrary, i.e., let each term in the summation
in (59) be equal to zero. This implies that the columns of
1/2
U2⊥ belong to Null(Γ1,S1 \S2 ATS1 \S2 ), which means that they
also belong to Null(AS1 \S2 Γ1,S1 \S2 ATS1 \S2 ). Since, for a
symmetric matrix, the row and column spaces are equal and
orthogonal to the null space of the matrix, it follows that
1/2
Col(AS1 \S2 Γ1,S1 \S2 ATS1 \S2 ) (same as Col(AS1 \S2 Γ1,S1 \S2 ))
is spanned by the columns of U2 , or equivalently by columns
of AS2 . Thus, every column in AS1 \S2 can be expressed
as a linear combination of columns in AS2 . This contradicts
our initial assumption that K + 1 < spark(A), implying that
any K + 1 or fewer columns of A are linearly independent.
Therefore, we conclude that there is at least one strictly
positive term in the summation in (59), and consequently there
exists a constant c1 > 0 such that µ∗ ≥ c1 /σ 2 .
ACKNOWLEDGEMENT
The authors gratefully acknowledge Heng Qiao from UC
San Diego for helpful comments and discussions.
R EFERENCES
[1] M. Masood, L. Afify, and T. Al-Naffouri, “Efficient coordinated recovery
of sparse channels in massive MIMO,” IEEE Trans. Signal Process.,
vol. 63, no. 1, pp. 104–118, 2015.
[2] Y. Barbotin, A. Hormati, S. Rangan, and M. Vetterli, “Estimation of
sparse MIMO channels with common support,” IEEE Trans. Commun.,
vol. 60, no. 12, pp. 3705–3716, 2012.
[3] R. Prasad, C. Murthy, and B. Rao, “Joint channel estimation and
data detection in MIMO-OFDM systems: A sparse Bayesian learning
approach,” IEEE Trans. Signal Process., vol. 63, no. 20, pp. 5369–5382,
Oct. 2015.
[4] D. Baron, M. F. Duarte, M. B. Wakin, S. Sarvotham, and R. G.
Baraniuk, “Distributed compressive sensing,” CoRR, vol. abs/0901.3403,
2009. [Online]. Available: http://arxiv.org/abs/0901.3403
[5] H. Zhu, G. Giannakis, and A. Cano, “Distributed in-network channel
decoding,” IEEE Trans. Signal Process., vol. 57, no. 10, pp. 3970–3983,
Oct. 2009.
[6] S. Ji, D. Dunson, and L. Carin, “Multitask compressive sensing,” IEEE
Trans. Signal Process., vol. 57, no. 1, pp. 92–106, Jan. 2009.
[7] T. Erseghe, “A distributed and maximum-likelihood sensor network
localization algorithm based upon a nonconvex problem formulation,”
IEEE Trans. Signal Inf. Process. Netw., vol. 1, no. 4, pp. 247–258, Dec.
2015.
[8] Z. Tan, P. Yang, and A. Nehorai, “Joint sparse recovery method for
compressed sensing with structured dictionary mismatches,” IEEE Trans.
Signal Process., vol. 62, no. 19, pp. 4997–5008, Oct. 2014.
[9] Z. Fanzi, C. Li, and Z. Tian, “Distributed compressive spectrum sensing
in cooperative multihop cognitive networks,” IEEE J. Sel. Topics Signal
Process., vol. 5, no. 1, pp. 37–48, Feb. 2011.
[10] Y. Wang, A. Pandharipande, Y. L. Polo, and G. Leus, “Distributed
compressive wide-band spectrum sensing,” in Information Theory and
Applications Workshop, Feb. 2009, pp. 178–183.
[11] J. Bazerque and G. Giannakis, “Distributed spectrum sensing for cognitive radio networks by exploiting sparsity,” IEEE Trans. Signal Process.,
vol. 58, no. 3, pp. 1847–1862, Mar. 2010.
[12] S. Cotter, B. Rao, K. Engan, and K. Kreutz-Delgado, “Sparse solutions
to linear inverse problems with multiple measurement vectors,” IEEE
Trans. Signal Process., vol. 53, no. 7, pp. 2477–2488, Jul. 2005.
[13] J. Chen and X. Huo, “Theoretical results on sparse representations of
multiple-measurement vectors,” IEEE Trans. Signal Process., vol. 54,
no. 12, pp. 4634–4643, Dec. 2006.
[14] J. Tropp, A. Gilbert, and M. Strauss, “Simultaneous sparse approximation via greedy pursuit,” in Proc. ICASSP, vol. 5, 2005, pp. 721–724.
[15] S. Foucart and H. Rauhut, A Mathematical Introduction to Compressive
Sensing. Birkhäuser Basel, 2013.
[16] M. Duarte, S. Sarvotham, D. Baron, M. Wakin, and R. Baraniuk,
“Distributed compressed sensing of jointly sparse signals,” in Proc.
Asilomar Conf. on Signals, Syst., and Comput., Oct. 2005, pp. 1537–
1541.
[17] B. D. Rao, K. Engan, and S. Cotter, “Diversity measure minimization
based method for computing sparse solutions to linear inverse problems
with multiple measurement vectors,” in Proc. ICASSP, vol. 2, May 2004,
pp. 369–372.
[18] J. D. Blanchard, M. Cermak, D. Hanle, and Y. Jing, “Greedy algorithms
for joint sparse recovery,” IEEE Trans. Signal Process., vol. 62, no. 7,
pp. 1694–1704, Apr. 2014.
[19] K. Lee, Y. Bresler, and M. Junge, “Subspace methods for joint sparse
recovery,” IEEE Trans. Inf. Theory, vol. 58, no. 6, pp. 3613–3641, Jun.
2012.
[20] J. M. Kim, O. K. Lee, and J. C. Ye, “Compressive music: Revisiting the
link between compressive sensing and array signal processing,” IEEE
Trans. Inf. Theory, vol. 58, no. 1, pp. 278–301, Jan. 2012.
[21] M. Mishali and Y. C. Eldar, “Reduce and boost: Recovering arbitrary
sets of jointly sparse vectors,” IEEE Trans. Signal Process., vol. 56,
no. 10, pp. 4692–4702, Oct. 2008.
[22] O. Balkan, K. Kreutz-Delgado, and S. Makeig, “Localization of more
sources than sensors via jointly-sparse Bayesian learning,” IEEE Signal
Process. Lett., vol. 21, no. 2, pp. 131–134, Feb. 2014.
[23] P. Pal and P. P. Vaidyanathan, “Pushing the limits of sparse support
recovery using correlation information,” IEEE Trans. Signal Process.,
vol. 63, no. 3, pp. 711–726, Feb. 2015.
[24] S. Khanna and C. R. Murthy, “Rényi divergence based covariance
matching pursuit of joint sparse support,” in Sig. Process. Advances
in Wireless Commun. (SPAWC), Jul. 2017, pp. 1 – 5.
[25] D. Wipf and B. Rao, “An empirical Bayesian strategy for solving
the simultaneous sparse approximation problem,” IEEE Trans. Signal
Process., vol. 55, no. 7, pp. 3704–3716, Jul. 2007.
[26] S. E. Chiu and B. Rao, “Correlation learning on joint support recovery
for more sources than measurements,” in IEEE Sensor Array and
Multichannel Signal Processing Workshop (SAM), Jul. 2016, pp. 1–5.
[27] P. Pal and P. P. Vaidyanathan, “Parameter identifiability in sparse
Bayesian learning,” in Proc. ICASSP, May 2014, pp. 1851–1855.
[28] E. van den Berg and M. P. Friedlander, “Theoretical and empirical results
for recovery from multiple measurements,” IEEE Trans. Inf. Theory,
vol. 56, no. 5, pp. 2516–2527, May 2010.
[29] Y. C. Eldar and M. Mishali, “Robust recovery of signals from a
structured union of subspaces,” IEEE Trans. Inf. Theory, vol. 55, no. 11,
pp. 5302–5316, Nov. 2009.
[30] M. E. Davies and Y. C. Eldar, “Rank awareness in joint sparse recovery,”
IEEE Trans. Inf. Theory, vol. 58, no. 2, pp. 1135–1146, Feb. 2012.
[31] G. Tang and A. Nehorai, “Performance analysis for sparse support
recovery,” IEEE Trans. Inf. Theory, vol. 56, no. 3, pp. 1383–1399, Mar.
2010.
[32] Y. Jin and B. D. Rao, “Support recovery of sparse signals in the presence
of multiple measurement vectors,” IEEE Trans. Inf. Theory, vol. 59,
no. 5, pp. 3139–3157, May 2013.
[33] S. Park, N. Y. Yu, and H. N. Lee, “An information-theoretic study for
joint sparsity pattern recovery with different sensing matrices,” IEEE
Trans. Inf. Theory, vol. PP, no. 99, pp. 1–1, 2017.
[34] A. Banerjee, S. Merugu, I. S. Dhillon, and J. Ghosh, “Clustering with
Bregman divergences,” J. Mach. Learn. Res., vol. 6, pp. 1705–1749,
Dec. 2005.
[35] I. S. Dhillon and J. A. Tropp, “Matrix nearness problems with Bregman divergences,” SIAM Journal on Matrix Analysis and Applications,
vol. 29, no. 4, pp. 1120–1146, 2008.
[36] R. Vershynin, “On the role of sparsity in compressed sensing and random
matrix theory,” in Proc. CAMSAP, Dec. 2009, pp. 189–192.
[37] M. Gil, F. Alajaji, and T. Linder, “Rényi divergence measures for commonly used univariate continuous distributions,” Information Sciences,
vol. 249, pp. 124 – 131, Nov. 2013.
[38] R. Vershynin, “Introduction to the non-asymptotic analysis of random
matrices,” CoRR, vol. abs/1011.3027, 2011. [Online]. Available:
http://arxiv.org/abs/1011.3027
[39] S. Khanna and C. R. Murthy, “On the restricted isometry of the
columnwise Khatri-Rao product,” CoRR, vol. abs/1709.05789, 2017.
[Online]. Available: http://arxiv.org/abs/1709.05789
[40] D. S. Bernstein, Matrix Mathematics: Theory, Facts, and Formulas
(Second Edition). Princeton University Press, 2009.
| 7cs.IT
|
JOURNAL OF IEEE TRANSACTIONS ON XX, VOL. X, NO. X, XX 2018
1
Dual Asymmetric Deep Hashing Learning
arXiv:1801.08360v1 [cs.CV] 25 Jan 2018
Jinxing Li, Bob Zhang, Member, IEEE, Guangming Lu, David Zhang*, Fellow, IEEE
Abstract—Due to the impressive learning power, deep learning
has achieved a remarkable performance in supervised hash
function learning. In this paper, we propose a novel asymmetric
supervised deep hashing method to preserve the semantic structure among different categories and generate the binary codes
simultaneously. Specifically, two asymmetric deep networks are
constructed to reveal the similarity between each pair of images
according to their semantic labels. The deep hash functions
are then learned through two networks by minimizing the gap
between the learned features and discrete codes. Furthermore,
since the binary codes in the Hamming space also should keep
the semantic affinity existing in the original space, another
asymmetric pairwise loss is introduced to capture the similarity between the binary codes and real-value features. This
asymmetric loss not only improves the retrieval performance,
but also contributes to a quick convergence at the training
phase. By taking advantage of the two-stream deep structures
and two types of asymmetric pairwise functions, an alternating
algorithm is designed to optimize the deep features and highquality binary codes efficiently. Experimental results on three
real-world datasets substantiate the effectiveness and superiority
of our approach as compared with state-of-the-art.
Index Terms—deep learning, hashing learning, image retrieval,
similarity
I. I NTRODUCTION
W
ITH the rapid growth of multimedia data in search
engines and social networks, how to store these data
and make a fast search when an novel one such as the image
is given, plays a fundamental role in machine learning. Due
to the low storage cost and fast retrieval speed, hashing techniques have attracted much attention and are widely applied
in nearest neighbor search [1] for information retrieval on
large scale datasets. Hashing learning aims to project the data
from the original space into a Hamming space by generating
compact codes. These codes can not only dramatically reduce
the storage overhead and achieve a constant or sub-linear
time complexity in information search, but also preserve the
semantic affinity existing in the original space.
Many hashing methods have been studied [2] [3] [4] [5] [6]
[7] [8] [9] [10]. Generally, these approaches can be roughly
classified into two categories: data-independent and datadependent hashing methods. Locality Sensitive Hashing (LSH)
[2] and its extension Kernelized LSH (KLSH) [3], as the most
typical data-independent hashing methods, were proposed to
J. Li is with the Department of Computing, Hong Kong Polytechnic
University, Hung Hom, Kowloon (e-mail: [email protected]).
B. Zhang is with the Department of Computer and Information Science, University of Macau, Avenida da Universidade, Taipa, Macau (e-mail:
[email protected]).
G. Lu is with Department of Computer Science, Harbin Institute
of Technology Shenzhen Graduate School, Shenzhen, China (e-mail:
[email protected]).
D. Zhang is with the Department of Computing, Hong Kong Polytechnic
University, Hung Hom, Kowloon (e-mail: [email protected]).
obtain the hashing function by using random projections.
Although the designation of these data-independent methods is
quite simple, they often meet a performance degradation when
the length of the binary codes is relatively low. By contrary,
instead of randomly generating the hashing function like LSH
does, data-dependent methods aims to learn a data-specific
hashing function by using the training data, being capable of
generating shorter binary codes but achieving more remarkable performance. Therefore, various data-dependent hashing
approaches containing both unsupervised and supervised have
been proposed. Unsupervised hashing, e.g. Spectral Hashing
[6], Anchor Graph Hashing (AGH) [7], and Discrete Graph
Hashing (DGH) [8] etc., only try to utilize the data structure
to learn compact binary codes to improve the performance. By
taking the label information into account, supervised hashing
methods attempt to map the original data into a compact
Hamming space to preserve the similarity between each pair
samples. Many representative works including Fast Supervised
Hashing (FastH) [5], Kernel Supervised Hashing (KSH) [9],
and Supervised Discrete Hashing (SDH) [10] etc., demonstrate
that supervised hashing methods often obtain an outstanding
performance compared with unsupervised hashing methods.
Thus, we focus on studying the supervised hashing method in
this paper.
Although some traditional supervised hashing methods
achieve a good performance in some applications, most of
them only linearly map the original data into a Hamming space
by using the hand-crafted features, limited their application
for large-scale datasets which have complex distributions.
Fortunately, due to the powerful capability of data representation, deep learning [11] [12] provides a promising way
to jointly represent the data and learn hash codes. Some
existing deep learning based hashing methods have been
studied, such as Deep Supervised Hashing (DSH) [13] and
Deep Pairwise Supervised Hashing (DPSH) [14], etc. These
approaches demonstrate the effectiveness of the end-to-end
deep learning architecture for hashing learning.
Despite the wide applications of deep neural network on
hashing learning, most of them are symmetric structures in
which the similarity between each pair points are estimated
by the Hamming distance between the outputs of the same
hash function [15]. As described in [15], a crucial problem is
that this symmetric scheme would result in the difficulty of optimizing the discrete constraint. Thus in this paper, we propose
a novel asymmetric hashing method to address aforementioned
problem. Note that a similar work was described by Shen et
al. [15], named deep asymmetric pairwise hashing (DAPH).
However, our study is quite distinctive from DAPH. Shen et
al. tried to approximate the similarity affinity by exploiting
two different hashing functions, which can preserve more
similarity information among the real-value features. However,
JOURNAL OF IEEE TRANSACTIONS ON XX, VOL. X, NO. X, XX 2018
DAPH only exploits a simple Euclidean distance, but ignores
the semantic structure between the learned real-value features
and binary codes [16] [17]. One major deficiency is that it
is difficult to efficiently preserve the similarity in the learned
hash functions and discrete codes. Furthermore, in DAPH, two
different types of discrete hash codes corresponding to two
hash functions are estimated in the training time. However
this strategy would enlarge the gap between two schemes,
resulting in a performance degradation. By contrast, we not
only propose a novel asymmetric structure to learn two different hash functions and one consistent binary code for each
sample at the training phase, but also asymmetrically exploit
real-value and multiple integer values, which permits the better
preservation of similarity between the learned features and
hash codes. Experiments show that this novel asymmetric
structure can get a better performance in image retrieval and
quicker convergence at the training stage.
The main contributions of the proposed method are shown
as follows:
(1) A novel asymmetric deep structure are proposed. Two
streams of deep neural networks are trained to asymmetrically
learn two different hash functions. The similarity between each
pair images are utilized through a pairwise loss according to
their semantic/label information.
(2) The similarity between the learned features and binary
codes are also revealed through an additional asymmetric loss.
Real-value features and binary codes are bridged through an
inner product, which alleviates the binary limitation, better
preserves the similarity, and speeds up convergence at the
training phase.
(3) By taking advantage of these two asymmetric properties,
an alternative algorithm is designed to efficiently optimize the
real values and discrete values.
(4) Experimental results on three large-scale datasets substantiate the effectiveness and superiority of our approach as
compared with some existing state-of-the-art hashing methods
in image retrieval.
The rest of this paper is organized as follows. In Section
2, the related works including data-independent and datadependent hashing methods are briefly reviewed. In Section
3, the proposed Dual Asymmetric Deep Hashing Learning
(DADH) is then analyzed, followed by its optimization. In
Section 4, experiments are conducted on three real-world
datasets, and some comparisons, parameter sensitivity analysis
and convergence analysis are discussed. This paper is finally
concluded in Section 5.
II. R ELATED W ORKS
As mentioned before, the hashing method can be roughly
separated into data-independent and data-dependent hashing.
Locality Sensitive Hashing (LSH) [2] aims to use several
hash functions to randomly project the data into a Hamming
space, so as to ensure the probability of collision is much
higher for data points which are close to each other than for
those which are far apart. Consider the non-linearity existing in
many real-world datasets, LSH was generated to accommodate
arbitrary kernel functions (KLSH) in [3]. Some other priors,
2
such as p-stable distributions [18] and shift-invariant kernels
[19], are also embedded to extend LSH for performance
improvement.
Different from data-independent methods, the datadependent methods try to learn more compact codes from
a given dataset to achieve a satisfactory search accuracy.
According to whether the label information is available, datadependent hashing can also be classified into unsupervised and
supervised. Typical learning criteria for unsupervised hashing
methods contains graph learning [6] [7] [8] [20] and error
minimization [4] [21] [22]. Graph Learning: Yair et al. [6]
proved that finding a best code is associated with the problem
of graph partitioning. Thus, a spectral hashing (SH) was proposed to learn the hash function. Another graph hashing named
Anchor Graph Hashing (AGH) was presented by Liu et al.
[7], which is capable of capturing the neighborhood structure
inherent automatically. In order to avoid the high complexity
of existing graph hashing methods, Jiang et al. [20] proposed
a scalable graph hashing (SGH) which can be effectively
applied to the large-scale dataset search. Although SH, AGH
and SGH achieve a satisfactory performance in some datasets,
both of them relax the optimization by discarding the discrete
constraints, which results in an accumulated quantization error.
To address this problem, a discrete graph hashing (DGH)
[8] was proposed, which can find the neighborhood structure
inherent in a discrete code space. Error Minimization: A
typical method is the iterative quantization (ITQ) [4] which
aims to project the data to the vertices of a binary hypercube
and minimize the quantization error. Additionally, the method
in [21] makes a quantization by decomposing the input space
into a Cartesian product of low-dimensional subspaces, dramatically reducing the quantization noise. Different from most
single-bit quantization, a double-bit quantization hashing [23]
was also studied by quantizing each dimension into double
bits.
In contrast to unsupervised hashing methods, supervised
hashing learning utilizes the label information to encourage the
binary codes in the Hamming space to preserve the semantic
relationship existing in the raw data. For instance, Mohammad
et al. [24] introduced a hinge-like loss function to exploit the
semantic information. Besides, Li et al. [25] projected the
raw data into a latent subspace, and the label information is
embedded on this subspace to preserve the semantic structure.
The Jensen Shannon Divergence is also utilized in [26] to
learn the binary codes within a probabilistic framework, in
which an upper bound is derived for various hash functions.
Being similar to KLSH, Liu et al. [9] proposed a supervised
hashing with kernels, in which the similar pairs are minimized
while the dissimilar pairs are maximized. Consider the discrete
constraint, the supervised discrete hashing (SDH) [10] was
proposed to not only preserve the semantic structure, but
also discretely learn the hash codes without any relaxation.
However, this discrete optimization is time-consuming and
unscalable. To tackle this problem, a novel method named
column sample based discrete supervised hashing (COSDISH)
was presented to directly obtain the binary codes from semantic information.
Although various works mentioned above have been stud-
JOURNAL OF IEEE TRANSACTIONS ON XX, VOL. X, NO. X, XX 2018
ied, they only project the data into the Hamming space by
using the hand-crafted features. The main limitation is that
they would meet a performance degradation if the distribution
of a real-world dataset is complex. Fortunately, deep learning
provides a reasonable and promising solution. Liong et al.
[27] used the deep structure to hierarchically and non-linearly
learn the hash codes. Convolutional neural network (CNN)
was first applied by Xia and Yan et al. to the hashing learning
(CNNH) [28], which simultaneously represents the image and
learns a hash function. A novel deep structure [29] was then
proposed by modifying the fully-connected layer in CNNH
to a divide-and-encode module, in which the hash codes can
be obtained bit by bit. Also, Can et al. [30] combined the
quantization model with deep structure to gain a satisfactory
performance in image retrieval. Different from the triple loss
used in some deep hashing methods, Li et al. [14] studied
a pairwise loss (DPSH) which can effectively preserve the
semantic information between each pair outputs. Due to the
power of asymmetric structure, the asymmetric deep hashing
was also studied in recent years. For instance, Shen et al.
[15] (DAPH) tried to learn hash functions in an asymmetric network. However, DAPH only exploits two streams to
preserve the pairwise label information between the deep
neural network outputs, but ignores the similarity between
the real-value features and binary codes. Thus, in this paper,
we propose a novel deep hashing method to not only exploit
the label information between each two outputs through an
asymmetric deep structure, but also semantically associated
the learned real-value features with the binary codes.
III. T HE P ROPOSED M ETHOD
In this section, we first give some notations used in this
paper, as well as the problem definition. The proposed Dual
Asymmetric Deep Hashing Learning (DADH) is then described, followed by its optimization.
A. Notation and Problem Definition
In this paper, since there are two streams in
the proposed method, we use the uppercase letters
X
=
{x1 , · · · , xi , . . . , xN }
∈
RN ×d1 ×d2 ×3 and
N ×d1 ×d2 ×3
Y = {y1 , · · · , yi , . . . , yN } ∈ R
to denote
the input images in the first and second deep neural networks,
respectively, where N is the number of training samples,
d1 and d2 are the length and width for each image. Note
that, although X and Y are represented with different
symbols, both of them denote the same training data. In our
experiments, we only alternatively use training samples X
and Y in the first and second networks. Since our method
is supervised learning, the label information can be used.
Let the uppercase letter S ∈ {−1, +1} denote the similarity
between X and Y and Si,j is the element in the i-th row and
j-th column in S. Let Si,j = 1 if xi and yj share the same
semantic information or label, otherwise Si,j = −1.
Denote the binary codes as B = [b1 , · · · , bi , . . . , bN ]T ∈
N ×k
R
and the k-bit binary code of the i-th sample as
bi ∈ {−1, +1}k×1 . The purpose of our model is to learn
two mapping functions F and G to project X and Y into the
3
TABLE I
T HE NETWORK STRUCTURE OF CNN-F.
Layer
conv1
conv2
conv3
conv4
conv5
full6
full7
full8
Structure
f. 64 × 11 × 11; st. 4×4; pad. 0; LRN.; ×2 pool
f. 265 × 5 × 5; st. 1×1; pad. 2; LRN.; ×2 pool
f. 265 × 3 × 3; st. 1×1; pad. 1
f. 265 × 3 × 3; st. 1×1; pad. 1
f. 265 × 3 × 3; st. 1×1; pad. 1; ×2 pool
4096
4096
→ k-bit hash code
Hamming space B. bi = sign(F(xi )) and bj = sign(G(yj )),
where sign(·) is an element-wise sign function, and sign(x) =
1 if x ≥ 0, otherwise sign(x) = -1. The Hamming distance
distH (bi , bj ) between bi and bj should be as small as
possible if sij = 1 and vice versa. Due to the power of
deep neural network in data representation, we apply the
convolution neural work to learn the hash functions. Specifically, the CNN-F structure [31] is adopted to perform feature
learning. In CNN-F model, there are eight layers including five
convolutional layers as well as three fully-connected layers.
The network structure is listed in Table I, where ”f.” means
the filter, ”st.” means the convolution stride, ”LRN” means the
Local Response Normalization [11]. In order to get the final
binary code, we replace the last layer in CNN-F with a k-D
vector and the k-bit binary codes are obtained through a sign
operation on the output of the last layer. In this paper, CNN-F
model is applied to both streams in our proposed asymmetric
structure.
B. Dual Asymmetric Deep Hashing Learning
The main framework of the proposed method is shown
in Fig.1. As we can see, there are two end-to-end neural
networks to discriminatively represent the inputs. For a pair
of outputs F and G in these two streams, their semantic
information is exploit through a pairwise loss according to
their predefined similarity matrix. Since the purpose is to
obtain hash functions through the deep networks, the binary
code B is also generated by minimizing its distance between
F and G. Furthermore, in order to preserve the similarity
between the learned binary codes and real-value features, and
alleviate the binary limitation, another asymmetric pairwise
loss is introduced by using the inner product of the hash codes
B and learned features F (G).
Denote f (xi , Wf ) ∈ Rk×1 as the output of the i-th sample
in the last layer of the first stream, where Wf is the parameter
of the network. To simplify the notation, we use fi to replace
f (xi , Wf ). Similarly, we can obtain the output gj corresponding to the j-th sample under the parameter Wg in the second
network. Thus, the features F = [f1 , · · · , fi , fn ]T ∈ Rn×k and
G = [g1 , · · · , gi , gn ]T ∈ Rn×k corresponding to the first and
second networks are then gained.
To learn an accurate binary code, we set sign(fi ) and
sign(gi ) to be close to their corresponding hash code bi . A
general way is to minimize the L2 loss between them.
min ksign(fi ) − bi k22 + ksign(gi ) − bi k22 ,
s.t. bi ∈ {−1, +1}
(1)
JOURNAL OF IEEE TRANSACTIONS ON XX, VOL. X, NO. X, XX 2018
4
. ..
. ..
. ..
. ..
U
1, 0,
0,1,
1, 0,
Real-valued
Data Similarity
Preservation
, 0,1
, 0, 0
S
,1,1
V
Hash Code
Feature Extraction with
different Weights
B
1
1
1
1
Real-valued
Data and
Binary Codes
Similarity
Preservation
. ..
. ..
. ..
. ..
Fig. 1. The framework of the proposed method. Two streams with five convolution layers and two full-connected layers are used for feature extraction. For
the real-valued outputs from these two neural networks, their similarity is preserved by using a pairwise loss. Based on the outputs, a consistent hash code is
generated. Furthermore, an asymmetric loss is introduced to exploit the semantic information between the binary code and real-valued data.
However, it is difficult to make a back-propagation for the
gradient with respect to fi or gi in Eq.(1) since their gradients
are zero anywhere. In this paper, we apply tanh(·) to softly
approximate the sign(·) function. Thus, Eq.(1) is transformed
into
2
2
min ktanh(fi ) − bi k2 + ktanh(gi ) − bi k2 ,
(2)
s.t. bi ∈ {−1, +1}
prove that this kind of asymmetric inner product can quickly
make the network converge to a stable value for the real-valued
features and hash codes.
Jointly taking Eq.(2), Eq.(4) and Eq.(5) into account, the
objective function can be obtained as follows:
min L = tanh(F)BT − kS
F,G,B
Furthermore, to exploit the label information and keep
a consistent similarity between two outputs F and G, the
negative log likelihood of the dual-stream similarities with the
likelihood function is exploited.
σ(Θij )
Sij = 1
p(Sij |tanh(fi ), tanh(gj )) =
(3)
1 − σ(Θij ) Sij = 0
where Θij = 12 tanh(fiT )tanh(gj ), and σ(Θij ) = 1+e1−Θij .
Therefore, the pairwise loss for these two different outputs
is shown as follows.
n
X
min −
(Sij Θij − log(1 + eΘij )
(4)
i,j=1
Although Eq.(2) achieves to approximate discrete codes
and Eq.(4) exploits the intra- and inter-class information, the
similarity between the binary codes and real-value features is
ignored. To tackle this problem, another asymmetric pairwise
loss is introduced.
min tanh(fiT )bj − kSij
s.t. bi ∈ {−1, +1}
2
2
+ tanh(giT )bj − kSij
2
2
,
(5)
In Eq.(5), the similarity between the real-valued data and
binary codes is measured by their inner product. It is easy to
observe that Eq.(5) not only encourages the tanh(fi ) (tanh(gi ))
and bi to be consistent, but also preserve the similarity
between them. Additionally, our experiments in Section 4 also
−τ
n
X
2
F
+ tanh(G)BT − kS
2
F
(Sij Θij − log(1 + eΘij )+
i,j=1
2
2
+ γ(ktanh(F) − BkF + ktanh(G) − BkF )
+ η( tanh(F)T 1
2
F
+ tanh(G)T 1
2
)
F
s.t. bi ∈ {−1, +1}
(6)
where τ , γ and η are the non-negative parameters to make a
trade-off among various terms. Note that the purpose of the
2
2
forth term tanh(F)T 1 F + tanh(G)T 1 F in the objective
function Eq.(6) is to maximize the information provided by
each bit [32]. In detail, this term makes a balance for each bit,
which encourages the number of -1 and +1 to be approximately
similar among all training samples.
C. Optimization
From the objective function Eq.(6), we can see that the
real-valued features as well as the weights in two neural
networks (F, Wf ) / (G, Wg ), and discrete codes B need
to be optimized. Note that this NP-hard problem is highly
non-convex, and it is very difficult to directly get the optimal
solutions. In this paper, we design an efficient algorithm to
optimize them alternatively. Specially, we update one variable
by fixing other variables.
JOURNAL OF IEEE TRANSACTIONS ON XX, VOL. X, NO. X, XX 2018
5
1) Update (F, Wf ) with (G, Wg ) and B fixed: By fixing
(G, Wg ) and B, the objective function Eq.(6) can be transformed to
n
X
2
min tanh(F)BT − kS F − τ
(Sij Θij − log(1 + eΘij )+
F
i,j=1
+ γ(ktanh(F) −
2
BkF
+ η tanh(F)T 1
2
F
(7)
Then the back-propagation is exploited to update (F, Wf ).
Here denote U = tanh(F) and V = tanh(G). The gradient of
the objective function with respect to fi is
n
X
∂L
τ
={
[2bj (bTj ui − kSij ) + (σ(Θij )vj − Sij vj )]
∂fi
2
j=1
+ 2γ(ui − bi ) + 2ηUT 1}
(1 − u2i )
(8)
where
denotes the dot product. After getting the gradient
∂L
∂L
∂fi , the chain rule is used to obtain ∂Wf , and Wf is updated
by using back-propagation.
2) Update (G, Wg ) with (F, Wf ) and B fixed: Similarly,
by fixing (F, Wf ) and B, the back-propagation is exploited to
update (G, Wg ). The gradient of the objective function with
respect to gi is
n
X
∂L
τ
[2bj (bTj vi − kSij ) + (σ(Θij )uj − Sij uj )]
={
∂gi
2
j=1
+ 2γ(vi − bi ) + 2ηVT 1}
(1 − vi2 )
(9)
∂L
,
the
chain
rule
is
used
to
obtain
After getting the gradient ∂g
i
∂L
∂Wg , and Wg is updated by using back-propagation.
3) Update B with (F, Wf ) and (G, Wg ) fixed: By fixing
(F, Wf ) and (G, Wg ), we can get the following formulation.
min L(B) = UBT − kS
B
+ γ(kU −
2
BkF
2
F
+ VBT − kS
+ kV −
2
BkF )
s.t. bi ∈ {−1, +1}
(10)
minL(B) = −2Tr[B(k(UT S + VT S) + γ(UT + VT ))]
B
2
F
+ BVT
2
F
+ const s.t. bi ∈ {−1, +1}
(11)
where ’const’ means a constant value without any association
with B. For the sake of simplicity, let Q = −2k(ST U +
ST V) − 2γ(U + V). Eq.(11) can be simplified to
min L(B) = BUT
B
2
F
+ BVT
2
F
+ Tr[BQT ] + const
s.t. bi ∈ {−1, +1}
(12)
According to Eq.(12) and [17], B can be updated bit by bit.
In other words, we update one column in B with remaining
columns fixed. Let B∗c be the c-th column and B̂c be the
remaining columns in B. So do U∗c , Ûc , V∗c , V̂c , Q∗c , and
Q̂c . Eq.(12) can then be rewrote as:
T
min Tr(B∗c [2(UT∗c Ûc + V∗c
V̂c )B̂Tc + QT∗c ] + const
B∗c
s.t. B ∈ {−1, +1}n×k
Obviously, the optimal solution for B∗c is
B∗c = −sign(2B̂c (ÛTc U∗c + V̂cT V∗c ) + Q∗c )
(14)
After computing B∗c , we update B by replace the c-th column
with B∗c . Then we repeat Eq.(14) until all columns are
updated.
Overall, the optimization of the proposed method is listed
in Algorithm 1.
D. Query
2
F
Then Eq.(10) can be rewrote as:
+ BUT
Algorithm 1 Dual Asymmetric Deep Hashing Learning
(DADH)
Input: Training data X/Y; similarity matrix S; hash code
length k; predefined parameters τ , γ and η.
Output: Hashing functions F and G.
Initialization: Initialize weights of the first seven layers by
using the pretrained ImageNet model; the last layer is
initialized randomly; B is set to be a matrix whose
elements are zero.
1: while not converged or not reach the maximum iteration
do
2:
Update (F, Wf ):
Fix (G, Wg ) and B and update (F, Wf ) using backpropagation according to Eq.(8).
3:
Update (G, Wg ):
Fix (F, Wf ) and B and update (G, Wg ) using backpropagation according to Eq.(9).
4:
Update B:
Fix (F, Wf ) and (G, Wg ) and update B according to
Eq.(14).
5: end while
(13)
When Wf and Wg are learned, the hash functions corresponding to the two neural networks are subsequently obtained. For a given testing image x∗ , two kinds of binary
codes can be computed, which are b∗f = sign(f (x∗ , Wf )) and
b∗g = sign(f (x∗ , Wg )), respectively. Note that since tanh will
not influence the sign of each element at the testing phase,
we do not apply tanh for the output. From the experiments
we find that the performances computed through the first and
second networks are quite similar. To obtain a more robust
result, we use the average of two outputs as the final result in
our experiment.
b∗ = sign(0.5[f (x∗ , Wf ) + g(x∗ , Wg )])
(15)
IV. E XPERIMENTS
In this section, experiments are conducted on three largescale datasets to demonstrate the effectiveness of the proposed
method compared with some state-of-the-art approaches. We
first describe the datasets used in our experiments, followed
by the description of baselines, evaluation protocol, and implementation. We then make a comparison with other methods.
The parameter sensitivity as well as the convergence are
subsequently discussed.
JOURNAL OF IEEE TRANSACTIONS ON XX, VOL. X, NO. X, XX 2018
(a)
(d)
6
(b)
(c)
(e)
(f)
Fig. 2. The Precision-Recall curves computed by LSH, ITQ, DPLM, SDH, SGH, DPSH, ADSH, DAPH, and DADH on the IAPR TC-12 dataset. Figures
from (a) to (f) are associated with the code length 8-bit, 12-bit, 16-bit, 24-bit, 36-bit and 48-bit.
TABLE II
T HE MAP SCORES OBTAINED BY DIFFERENT METHODS ON THE IAPR
TC-12 DATASET.
Method
LSH
ITQ
DPLM
SDH
SGH
DPSH
ADSH
DAPH
DADH
8-bit
32.89
35.83
36.38
36.92
34.48
45.19
44.69
44.33
46.54
12-bit
33.39
36.00
36.86
37.81
35.01
46.03
46.98
44.48
49.27
16-bit
33.25
36.21
37.35
37.45
35.32
46.82
48.25
44.73
50.83
24-bit
34.33
36.55
37.71
38.12
35.49
47.37
49.06
45.12
52.71
36-bit
34.99
36.66
38.64
38.53
35.84
47.97
50.24
45.24
54.47
48-bit
35.15
36.75
38.83
38.27
36.21
48.60
50.59
45.52
55.39
labels, we also define two images be a ground-truth neighbor if
they share at least one same label. Additionally, 2000 images
are randomly selected as the testing data, and the rest is defined
as the retrieval data. Meanwhile, 5000 samples selected from
the retrieval data are used for training.
CIFAR-10 dataset [35] contains 60000 32×32 color images
with ten categories. Each image belongs to one of these ten
classes. Two images will be regarded as semantic neighbor
if they have the same label. Being similar to the setting in
[17], we randomly select 1000 samples to be the testing data.
In order to reduce the training time, we also randomly select
5000 images from the remaining 59000 images as the training
data. The rest is then regarded as the retrieval set.
A. Datasets
B. Baseline and Evaluation Protocol
Three datasets including IAPR TC-12 [33], MIRFLICKR25K [34] and CIFAR-10 [35] are used in this paper.
IAPR TC-12 [33] dataset consists of 20000 images associated with 255 categories. Since some samples have multiple
labels, we set Sij = 1 only if there is at least one same label
for the i-th and j-th sample. In this dataset, 2000 images are
used for testing and 5000 samples selected from the remaining
18000 (retrieval set) points are used for training to greatly
reduce the training time.
MIRFLICKR-25K dataset [34] is composed of 25000 images collected from the Flickr website. According to [32],
20015 images associated with 24 categories are selected. Being
similar to IAPR TC-12 dataset, some images has multiple
To demonstrate the superiority of DADH, some existing
hashing methods are used for comparison, including one
data-independent methods (LSH [2]), four traditional datadependent methods (ITQ [4], DPLM [36], SDH [10], SGH
[20]) and three deep learning based hashing methods (DPSH
[14], ADSH [17], DAPH [15]). Since LSH, ITQ, DPLM,
SDH, and SGH are not deep learning methods, features
should be extracted previously. For these three datasets, we
have extracted the 4096-D CNN feature 512-D GIST feature,
respectively. We have found that these five approaches often
achieve a better performance on the GIST feature. Thus we use
the GIST feature as the input for LSH, ITQ, DPLM, SDH, and
SGH. For DPSH, ADSH, and DAPH, the raw image is used
JOURNAL OF IEEE TRANSACTIONS ON XX, VOL. X, NO. X, XX 2018
7
TABLE III
T HE T OP -500 MAP AND T OP -500 P RECISION SCORES OBTAINED BY DIFFERENT METHODS ON THE IAPR TC-12 DATASET.
Evaluation
Method
LSH
ITQ
DPLM
SDH
SGH
DPSH
ADSH
DAPH
DADH
8-bit
36.19
41.16
42.21
44.08
40.04
57.07
53.70
56.26
57.12
12-bit
37.47
41.91
43.12
45.31
41.54
58.13
58.12
57.98
62.67
MAP@Top500
16-bit 24-bit
37.53
39.83
42.76
43.67
43.89
44.50
46.00
47.30
42.09
42.48
59.94
61.61
61.35
63.25
59.48
61.27
65.15
67.80
36-bit
40.50
44.07
45.86
48.17
43.07
63.05
64.90
62.57
70.11
as the input and all images are resized into 224×224×3. For
all deep learning methods, the CNN-F is used as the network
for feature extraction and the parameters in DPSH and ADSH
are set according to their descriptions in their publications.
Note that, since the code for DAPH is not released, we
implement it with the deep learning toolbox MatConvNet [37]
very carefully. Additionally, the original network structure in
DAPH is not CNN-F which means the parameters in [15] may
be not optimal. Thus, we try our best to tune the parameters
in DAPH.
To quantatively measure the proposed method and other
comparison methods, two widely used metrics containing
mean average precision (MAP) and precision-recall (PR) are
adopted. The definitions of MAP criteria is demonstrated as
follows: Given a query, the average precision (AP) is first
computed by searching a set of R retrieved results.
AP =
1
T
R
X
P (r)δ(r)
(16)
r=1
where T is the total number of document set in retrieved set,
P (r) is the precision of top r retrieved cases, and δ(r) denotes
whether the retrieved sample is relevant (if the instance is a
true neighbor of the query, δ(r) = 1, otherwise δ(r) = 0).
Additionally, being similar to some existing methods [15] [32],
Top-500 MAP and Top-500 Precision are also exploited to
evaluate the superiority of the proposed method.
C. Implementation
We implement DADH with the deep learning toolbox MatConvNet [37] on Titan X GPU. The pre-trained ImageNet
model is used to initialize the first seven layers in each stream
and the weights in the last layer are initialized randomly.
During the training time, we set the mini-batch size to be
64 and divide the learning rate among [10−6 , 10−4 ] into 150
iterations. In other words, the learning rate gradually reduces
from 10−4 to 10−6 and the stochastic gradient descent is used
to update the weights. Based on the cross-validation (a small
set for validation is randomly selected from the training data),
we set γ = 100, η = 10, and τ = 10 in the three datasets. We
will further demonstrate the insensitivity of these parameters
in the following subsection.
48-bit
41.52
44.30
46.42
48.06
43.97
64.49
65.59
63.94
70.93
8-bit
35.04
39.73
40.84
42.45
38.67
55.03
52.31
54.32
55.31
12-bit
36.10
40.30
41.65
43.54
39.70
55.90
56.47
55.44
60.20
Precision@Top500
16-bit 24-bit
36.12
38.10
40.88
41.60
42.33
42.82
43.94
45.00
40.08
40.33
57.73
58.96
58.97
60.29
56.52
57.92
62.52
64.93
36-bit
38.82
41.86
44.02
45.79
40.79
60.15
61.94
58.95
67.13
48-bit
39.55
42.04
44.44
45.49
41.45
61.40
62.50
60.03
67.81
TABLE IV
T HE MAP SCORES OBTAINED BY DIFFERENT METHODS ON THE
MIRFLICKR-25K DATASET.
Method
LSH
ITQ
DPLM
SDH
SGH
DPSH
ADSH
DAPH
DADHL
8-bit
56.06
57.59
60.42
60.17
57.35
73.48
75.39
72.79
77.15
12-bit
56.10
57.57
60.51
60.27
57.54
74.68
76.41
74.70
78.16
16-bit
56.72
57.70
60.53
60.46
57.57
75.58
76.98
74.30
78.64
24-bit
56.82
57.79
60.70
60.67
57.67
76.01
76.59
74.14
79.44
36-bit
57.56
57.84
60.91
60.96
57.80
76.09
76.20
73.81
79.72
48-bit
57.35
57.87
60.79
61.59
57.86
76.05
74.53
73.41
79.26
D. Comparison with Other Methods
1) IAPR TC-12: The MAP scores obtained by different
methods on the IAPR TC-12 dataset are shown in Tab.II.
It is easy to observe that DADH achieves a remarkable
improvement in MAP scores compared with other approaches.
In contrast to the data-independent method LSH, DADH
achieves more than 10%-20% percents higher in MAP scores.
Compared with ITQ, DPLM, SDH and SGH, there is also
an obvious enhancement. Specifically, our proposed method
obtains at least 46.54% MAP score and reaches as high as
55.39% when the bit length is 48, while the best result obtained
by ITQ, DPLM, SDH and SGH is only 38.83%, being far
below than our’s. Referring to DPSH, ADSH and DAPH,
DADH also has more or less improvement in MAP scores.
Particularly, compared with DAPH, the presented approach
gains about or more than 5% enhancement when the bit length
ranges from 12 to 48. In comparison to DPSH and ADSH,
the performance obtained by DADH also has about 3%-5%
improvement when the length of hashing bit is 24, 36, and
48, respectively.
The Top-500 MAP and Top-500 Precision scores on the
IAPR TC-12 dataset are listed in Tab.III. From this table we
can see that the experimental results obtained by the deep
learning based methods including DPSH, ADSH, DAPH and
DADH are remarkably better than that computed by other
traditional approaches. Specifically, there is about 10%-20%
improvement in MAP@Top500 and Precision@Top500 scores
in most cases. In contrast to DPSH, ADSH and DAPH,
the performance achieved by the proposed method DADH
JOURNAL OF IEEE TRANSACTIONS ON XX, VOL. X, NO. X, XX 2018
8
TABLE V
T HE T OP -500 MAP AND T OP -500 P RECISION SCORES OBTAINED BY DIFFERENT METHODS ON THE MIRFLICKR-25K DATASET.
Evaluation
Method
LSH
ITQ
DPLM
SDH
SGH
DPSH
ADSH
DAPH
DADHL
8-bit
57.57
61.00
63.08
65.16
60.74
82.88
82.14
81.08
85.80
12-bit
58.09
61.17
63.84
64.75
61.23
83.84
83.80
84.20
86.83
MAP@Top500
16-bit 24-bit
59.27
59.55
61.32
61.61
64.05
64.36
65.40
65.34
61.48
61.42
84.34
84.84
84.94
84.90
83.71
84.45
86.90
87.42
36-bit
60.84
61.64
64.73
65.57
61.91
85.77
84.20
84.02
87.98
48-bit
60.67
61.85
65.03
66.46
62.04
85.64
82.06
84.07
87.58
8-bit
56.94
59.98
62.36
64.05
59.81
81.85
81.50
80.37
84.73
12-bit
57.25
60.11
63.14
63.93
60.21
83.01
83.18
83.24
85.78
Precision@Top500
16-bit 24-bit
58.50
58.69
60.29
60.55
63.31
63.64
64.58
64.44
60.46
60.42
83.58
84.11
84.15
84.03
82.73
83.40
85.76
86.68
36-bit
59.93
60.56
64.05
64.80
60.85
84.97
83.52
82.93
87.08
(a)
(b)
(c)
(d)
(e)
(f)
48-bit
59.84
60.74
64.28
65.60
60.93
84.80
81.31
82.91
86.80
Fig. 3. The Precision-Recall curves computed by LSH, ITQ, DPLM, SDH, SGH, DPSH, ADSH, DAPH, and DADH on the MIRFLICKR-25K dataset.
Figures from (a) to (f) are associated with the code length 8-bit, 12-bit, 16-bit, 24-bit, 36-bit and 48-bit.
still reaches the best point. Except the case when the bit
length is 8, DADH always gain 4% or more enhancement in
MAP@Top500 and Precision@Top500 scores, indicating the
effectiveness of our method.
The Precision-Recall curves computed by different methods
on the IAPR TC-12 dataset are displayed in Fig.2, when the
bit length changes from 8 to 48. We can easily observe that
covered areas gained by DADH are much larger than that
obtained by other comparison methods. We can find that the
proposed method can dramatically outperform the traditional
data-independent and data-dependent strategies. Referring to
DPSH, ADSH and DAPH, there is also a better achievement
in all cases with different values of the code length.
2) MIRFLICKR-25K: The MAP results of the experiment
conducted on the MIRFLICKR-25K dataset are tabulated in
Tab.IV. We can see that DADH achieves the best performance
in all cases with different values of the code length. Being
similar to the results on IAPR TC-12 dataset, DPLM, SDH,
DPSH, ADSH, DAPH and DADH can obtain higher values
in MAP compared with LSH, ITQ and SGH. Referring to the
comparison between the traditional methods and deep learning
methods, DPSH, ADSH, DAPH and DADH dramatically
outperform LSH, ITQ, DPLM, SDH and SGH. Make a comparison between the proposed method with other deep hashing
approaches, DADH still has a more or less improvement. For
DADH, there is about 1.5%-3% enhancement on MAP scores
compared with these three deep hashing approaches.
The Top-500 MAP and Top-500 Precision scores on the
MIRFLICKR-25K dataset are displayed in Tab.V. It is easy
to observe that DADH obtains the best performance in both
JOURNAL OF IEEE TRANSACTIONS ON XX, VOL. X, NO. X, XX 2018
9
(a)
(b)
(c)
(d)
(e)
(f)
Fig. 4. The Precision-Recall curves computed by LSH, ITQ, DPLM, SDH, SGH, DPSH, ADSH, DAPH, and DADH on the CIFAR-10 dataset. Figures from
(a) to (f) are associated with the code length 8-bit, 12-bit, 16-bit, 24-bit, 36-bit and 48-bit.
TABLE VI
T HE MAP SCORES OBTAINED BY DIFFERENT METHODS ON THE
CIFAR-10 DATASET.
Method
LSH
ITQ
DPLM
SDH
SGH
DPSH
ADSH
DAPH
DADHL
8-bit
14.19
16.57
21.97
30.93
15.06
63.48
56.67
59.09
71.86
12-bit
13.26
17.12
22.76
32.50
15.43
66.97
71.41
61.17
75.12
16-bit
13.13
16.94
23.91
33.59
15.64
68.83
76.50
68.15
80.33
24-bit
13.84
17.08
25.89
35.36
16.07
73.45
80.40
69.22
81.70
36-bit
14.90
17.38
27.53
35.59
16.78
74.66
82.73
70.74
83.16
48-bit
15.13
17.58
28.85
36.53
16.88
75.02
82.73
70.28
83.90
MAP@Top500 and Precision@Top500, demonstrating the superiority compared with other existing strategies. With the
change of the code length, the MAP@Top500 and Precision@Top500 increase from (85.80%, 84.73%) to (87.58%,
86.80%), while the highest values obtained by LSH, ITQ and
SGH are only (62.04%, 60.93%), being much lower than ours’.
In contrast to DPLM and SDH, our strategy still gains more
than 20% enhancement in most cases. Furthermore, results
computed by DADH are much higher than that calculated by
DPSH, ADSH and DAPH. Concretely, scores of Top-500 MAP
and Top-500 Precision gained by DADH are almost always
higher than 85%, while these scores calculated by other deep
hashing methods are below than 85% in most cases.
Fig.3 show the Precision-Recall curves computed by differ-
ent methods on the MIRFLICKR-25K dataset, when the bit
length changes from 8 to 48. Note that we do not depict the
Precision-Recall curve obtained by LSH in Fig.3(f), since its
precision scores is far below than that of others. From Fig.3
we can observe that DADH remarkably outperforms LSH,
ITQ, DPLM, SDH, SGH, ADSH, and DAPH. Referring the
comparison between DPSH and DADH, the proposed method
is obviously superior to DPSH when the code length is 8 and
12, respectively. Although DPSH covers more areas when the
recall value is smaller than 0.4 in Fig.3(c)-(f), it is inferior
to DADH with the increase of the recall value. Overall, our
method still outperforms DPSH when the code length is 16,
24, 36 and 48.
3) CIFAR-10: Tab.VI lists the MAP scores obtained by the
proposed method and various comparison approaches on the
CIFAR-10 dataset. With the change of the code length from 8
to 48, the MAP scores computed by DADH rise from 71.86%
to 83.90%, being much higher than that obtained by traditional
hashing approaches, including LSH, ITQ, DPLM, SDH and
SGH. In contrast to DPSH and DAPH, it is easy to observe
that the presented method can achieve a better performance
under the different code length. Except the cases when code
length is 8 and 12, the MAP scores gained by DADH are
always higher than 80%, while the best performance computed
by DPSH and DAPH is only 75.02%. Also, ADSH is inferior
to the proposed method, especially when the code length is
small. This relatively indicates the effectiveness of our method
no matter the code length is small or large.
JOURNAL OF IEEE TRANSACTIONS ON XX, VOL. X, NO. X, XX 2018
10
TABLE VII
T HE T OP -500 MAP AND T OP -500 P RECISION SCORES OBTAINED BY DIFFERENT METHODS ON THE CIFAR-10 DATASET.
Evaluation
Method
LSH
ITQ
DPLM
SDH
SGH
DPSH
ADSH
DAPH
DADHL
8-bit
20.29
24.56
29.02
28.34
25.92
58.58
58.92
50.88
67.11
12-bit
19.92
27.72
34.07
34.78
25.11
66.85
70.07
66.24
73.08
MAP@Top500
16-bit 24-bit
19.81
21.17
27.55
28.65
35.66
38.86
37.92
41.69
25.91
27.13
71.48
75.74
74.95
78.09
72.43
77.31
78.42
82.02
36-bit
23.07
29.54
40.28
42.77
28.76
79.69
78.85
79.21
83.51
(a)
48-bit
24.42
30.24
41.67
44.57
29.46
80.62
77.61
80.40
84.17
8-bit
14.54
18.63
24.61
27.91
19.06
64.13
61.09
54.22
72.95
12-bit
15.26
21.88
29.72
33.83
20.10
70.80
73.71
68.28
77.53
Precision@Top500
16-bit 24-bit
16.19
18.13
22.53
24.59
31.93
35.62
37.31
42.39
21.53
23.46
74.71
78.34
78.85
81.84
74.42
77.36
82.68
84.18
(b)
36-bit
20.45
25.65
37.98
43.78
24.85
80.59
83.45
78.80
85.23
48-bit
21.44
26.40
39.84
45.08
25.42
81.55
82.86
79.55
85.59
(c)
Fig. 5. The MAP scores with the change of parameter τ , η and γ on three datasets, when the code length is 48.
(a)
(b)
(c)
Fig. 6. The change of objective function values and MAP scores with the increase of iterations.
The Top-500 MAP and Top-500 Precision scores computed
by different methods on the CIFAR-10 dataset are shown
in Tab.VII under the various code length. Obviously, four
deep hashing methods always achieve dramatic experimental results compared with rest traditional approaches. The
comparison between DADH and other deep hashing methods
also substantiates the superiority of the proposed strategy. In
contrast to DPSH, ADSH and DAPH, DADH has about 3%4% enhancement in both Top-500 MAP and Top-500 Precision
scores when the code length ranges from 12 to 48.
The Precision-Recall curves computed by different methods
on the CIFAR-10 dataset are depicted in Fig.4 under the bit
length ranging from 8 to 48. When the code length is 8, 12,
16, and 24, the Precision-Recall curves obtained by DADH
covers the most areas. Although the performance computed
by ADSH is competitive to ours’ when the code length is 36
and 48, DADH is still much superior to LSH, ITQ, DPLM,
SDH, SGH, DPSH, and DAPH.
E. Parameter Sensitivity Analysis
The MAP scores under the changes of different values of
τ , η and γ are shown in Fig.5. Note, we tune a parameter
with others fixed. For instance, we tune τ in the range of
[0.001, 1, 5, 10, 50, 100] by fixing η = 10 and γ = 100,
respectively. Similarly, we set τ = 10, γ = 100 in η tuning
and τ = 10, η = 10 in γ tuning. As we can see, our model is
JOURNAL OF IEEE TRANSACTIONS ON XX, VOL. X, NO. X, XX 2018
(a)
11
(b)
(c)
Fig. 7. The change of objective function values and MAP scores with the increase of iterations. Note that, the asymmetric terms tanh(F)BT − kS
2
tanh(G)BT − kS F are removed from the objective function.
insensitive to parameters. Specifically, τ , η and γ have a wide
range [1,50], [1,50] and [1,300], respectively. Our method
always achieves a satisfactory performance when τ , η and γ
are in these ranges. This relatively demonstrates the robustness
and effectiveness of the proposed method.
F. Convergence Analysis
To be honest, our proposed model can get a convergence
with a few of iterations. The change of the objective function
values and MAP scores on three datasets are displayed in
Fig.6 when the code length is 48-bit. It is easy to observe
that DADH converges to a stable value after less than 30
iterations. In fact, we further find that the asymmetric terms
2
2
tanh(F)BT − kS F and tanh(G)BT − kS F greatly contribute to the quick convergence. We try to remove these
two terms from our objective function to study the influence.
Note that, if the asymmetric terms are removed, the binary
code B are updated through sign(γ[tanh(G) + tanh(F)]).
2
As shown in Fig.7, if we remove tanh(F)BT − kS F and
2
tanh(G)BT − kS F from the objective function, not only
the MAP scores meet a degradation, but also our model
converges much slower compared with the original DADH,
indicating the necessity and significance of the asymmetric
terms.
V. C ONCLUSION
In this paper, we propose a novel deep hashing method
named dual asymmetric deep hashing learning (DADH) for
image retrieval. Specifically, two asymmetric networks are
designed to integrate the feature representation and hash function learning into the end-to-end framework. A pairwise loss
is introduced to exploit the semantic structure between each
pair outputs. Furthermore, another pairwise loss is proposed
to not only capture the similarity between the discrete binary
codes and learned real-value features, but also contribute to
a quick convergence at the training phase. Experiments are
conducted on three large-scale datasets and the outstanding
results substantiate the superiority of the proposed method.
2
F
and
ACKNOWLEDGMENT
The work is partially supported by the GRF fund from
the HKSAR Government, the central fund from Hong Kong
Polytechnic University, the NSFC fund (61332011, 61272292,
61271344, 61602540), Shenzhen Fundamental Research fund
(JCYJ20150403161923528, JCYJ20140508160910917), and
the Science and Technology Development Fund (FDCT) of
Macau 124/2014/A3.
R EFERENCES
[1] A. Andoni and P. Indyk, “Near-optimal hashing algorithms for approximate nearest neighbor in high dimensions,” in Foundations of Computer
Science, 2006. FOCS’06. 47th Annual IEEE Symposium on. IEEE, 2006,
pp. 459–468.
[2] A. Gionis, P. Indyk, R. Motwani et al., “Similarity search in high
dimensions via hashing,” in VLDB, vol. 99, no. 6, 1999, pp. 518–529.
[3] B. Kulis and K. Grauman, “Kernelized locality-sensitive hashing for
scalable image search,” in Computer Vision, 2009 IEEE 12th International
Conference on. IEEE, 2009, pp. 2130–2137.
[4] Y. Gong, S. Lazebnik, A. Gordo, and F. Perronnin, “Iterative quantization:
A procrustean approach to learning binary codes for large-scale image retrieval,” IEEE Transactions on Pattern Analysis and Machine Intelligence,
vol. 35, no. 12, pp. 2916–2929, 2013.
[5] G. Lin, C. Shen, Q. Shi, A. Van den Hengel, and D. Suter, “Fast
supervised hashing with decision trees for high-dimensional data,” in
Proceedings of the IEEE Conference on Computer Vision and Pattern
Recognition, 2014, pp. 1963–1970.
[6] Y. Weiss, A. Torralba, and R. Fergus, “Spectral hashing,” in Advances in
neural information processing systems, 2009, pp. 1753–1760.
[7] W. Liu, J. Wang, S. Kumar, and S.-F. Chang, “Hashing with graphs,”
in Proceedings of the 28th international conference on machine learning
(ICML-11). Citeseer, 2011, pp. 1–8.
[8] W. Liu, C. Mu, S. Kumar, and S.-F. Chang, “Discrete graph hashing,”
in Advances in Neural Information Processing Systems, 2014, pp. 3419–
3427.
[9] W. Liu, J. Wang, R. Ji, Y.-G. Jiang, and S.-F. Chang, “Supervised hashing
with kernels,” in Computer Vision and Pattern Recognition (CVPR), 2012
IEEE Conference on. IEEE, 2012, pp. 2074–2081.
[10] F. Shen, C. Shen, W. Liu, and H. Tao Shen, “Supervised discrete
hashing,” in Proceedings of the IEEE Conference on Computer Vision
and Pattern Recognition, 2015, pp. 37–45.
[11] A. Krizhevsky, I. Sutskever, and G. E. Hinton, “Imagenet classification
with deep convolutional neural networks,” in Advances in neural information processing systems, 2012, pp. 1097–1105.
[12] K. Simonyan and A. Zisserman, “Very deep convolutional networks for
large-scale image recognition,” arXiv preprint arXiv:1409.1556, 2014.
[13] H. Liu, R. Wang, S. Shan, and X. Chen, “Deep supervised hashing for
fast image retrieval,” in Proceedings of the IEEE Conference on Computer
Vision and Pattern Recognition, 2016, pp. 2064–2072.
JOURNAL OF IEEE TRANSACTIONS ON XX, VOL. X, NO. X, XX 2018
[14] W.-J. Li, S. Wang, and W.-C. Kang, “Feature learning based deep supervised hashing with pairwise labels,” arXiv preprint arXiv:1511.03855,
2015.
[15] F. Shen, X. Gao, L. Liu, Y. Yang, and H. T. Shen, “Deep asymmetric
pairwise hashing,” 2017.
[16] C. Da, S. Xu, K. Ding, G. Meng, S. Xiang, and C. Pan, “Amvh: Asymmetric multi-valued hashing,” in Proceedings of the IEEE Conference on
Computer Vision and Pattern Recognition, 2017, pp. 736–744.
[17] Q.-Y. Jiang and W.-J. Li, “Asymmetric deep supervised hashing,” arXiv
preprint arXiv:1707.08325, 2017.
[18] M. Datar, N. Immorlica, P. Indyk, and V. S. Mirrokni, “Locality-sensitive
hashing scheme based on p-stable distributions,” in Proceedings of the
twentieth annual symposium on Computational geometry. ACM, 2004,
pp. 253–262.
[19] M. Raginsky and S. Lazebnik, “Locality-sensitive binary codes from
shift-invariant kernels,” in Advances in neural information processing
systems, 2009, pp. 1509–1517.
[20] Q.-Y. Jiang and W.-J. Li, “Scalable graph hashing with feature transformation.” in IJCAI, 2015, pp. 2248–2254.
[21] H. Jegou, M. Douze, and C. Schmid, “Product quantization for nearest
neighbor search,” IEEE transactions on pattern analysis and machine
intelligence, vol. 33, no. 1, pp. 117–128, 2011.
[22] F. Shen, C. Shen, Q. Shi, A. Van Den Hengel, and Z. Tang, “Inductive
hashing on manifolds,” in Proceedings of the IEEE Conference on
Computer Vision and Pattern Recognition, 2013, pp. 1562–1569.
[23] W. Kong and W.-J. Li, “Double-bit quantization for hashing.” in AAAI,
vol. 1, no. 2, 2012, p. 5.
[24] M. Norouzi and D. M. Blei, “Minimal loss hashing for compact binary
codes,” in Proceedings of the 28th international conference on machine
learning (ICML-11). Citeseer, 2011, pp. 353–360.
[25] P. Zhang, W. Zhang, W.-J. Li, and M. Guo, “Supervised hashing with
latent factor models,” in Proceedings of the 37th international ACM SIGIR
conference on Research & development in information retrieval. ACM,
2014, pp. 173–182.
[26] L. Fan, “Supervised binary hash code learning with jensen shannon
divergence,” in Proceedings of the IEEE International Conference on
Computer Vision, 2013, pp. 2616–2623.
[27] V. Erin Liong, J. Lu, G. Wang, P. Moulin, and J. Zhou, “Deep hashing for
compact binary codes learning,” in Proceedings of the IEEE Conference
on Computer Vision and Pattern Recognition, 2015, pp. 2475–2483.
[28] R. Xia, Y. Pan, H. Lai, C. Liu, and S. Yan, “Supervised hashing for
image retrieval via image representation learning.” in AAAI, vol. 1, 2014,
pp. 2156–2162.
[29] H. Lai, Y. Pan, Y. Liu, and S. Yan, “Simultaneous feature learning and
hash coding with deep neural networks,” in Proceedings of the IEEE
Conference on Computer Vision and Pattern Recognition, 2015, pp. 3270–
3278.
[30] Y. Cao, M. Long, J. Wang, and S. Liu, “Deep visual-semantic quantization for efficient image retrieval,” in CVPR, vol. 2, 2017, p. 6.
[31] K. Chatfield, K. Simonyan, A. Vedaldi, and A. Zisserman, “Return of the
devil in the details: Delving deep into convolutional nets,” arXiv preprint
arXiv:1405.3531, 2014.
[32] Q.-Y. Jiang and W.-J. Li, “Deep cross-modal hashing,” arXiv preprint
arXiv:1602.02255, 2016.
[33] H. J. Escalante, C. A. Hernández, J. A. Gonzalez, A. López-López,
M. Montes, E. F. Morales, L. E. Sucar, L. Villaseñor, and M. Grubinger,
“The segmented and annotated iapr tc-12 benchmark,” Computer Vision
and Image Understanding, vol. 114, no. 4, pp. 419–428, 2010.
[34] M. J. Huiskes and M. S. Lew, “The mir flickr retrieval evaluation,”
in Proceedings of the 1st ACM international conference on Multimedia
information retrieval. ACM, 2008, pp. 39–43.
[35] A. Krizhevsky and G. Hinton, “Learning multiple layers of features from
tiny images,” 2009.
[36] F. Shen, X. Zhou, Y. Yang, J. Song, H. T. Shen, and D. Tao, “A fast
optimization method for general binary code learning,” IEEE Transactions
on Image Processing, vol. 25, no. 12, pp. 5610–5621, 2016.
[37] A. Vedaldi and K. Lenc, “Matconvnet: Convolutional neural networks
for matlab,” in Proceedings of the 23rd ACM international conference on
Multimedia. ACM, 2015, pp. 689–692.
12
| 1cs.CV
|
Aligned Drawings of Planar Graphs?
Tamara Mchedlidze1 , Marcel Radermacher1 , and Ignaz Rutter2
arXiv:1708.08778v3 [cs.DS] 21 Nov 2017
1
Department of Computer Science, Karlsruhe Institute of Technology, Germany
2
Department of Mathematics and Computer Science, TU Eindhoven, The
Netherlands,
[email protected], [email protected], [email protected]
Abstract. Let G be a graph topological embedded in the plane and
let A be an arrangement of pseudolines intersecting the drawing of G.
An aligned drawing of G and A is a planar polyline drawing Γ of G
with an arrangement A of lines so that Γ and A are homeomorphic
to G and A. We show that if A is stretchable and every edge e either
entirely lies on a pseudoline or intersects at most one pseudoline, then
G and A have a straight-line aligned drawing. In order to prove these
results, we strengthen the result of Da Lozzo et al. [5], and prove that a
planar graph G and a single pseudoline L have an aligned drawing with
a prescribed convex drawing of the outer face. We also study the more
general version of the problem where only a set of vertices is given and
we need to determine whether they can be collinear. We show that the
problem is N P-complete but fixed-parameter tractable.
1
Introduction
Two fundamental primitives for highlighting structural properties of a graph in
a drawing are alignment of vertices such that they are collinear, and geometric
separation of unrelated graph parts, e.g., by a straight line. Both these techniques
have been previously considered from a theoretical point of view in the case of
planar straight-line drawings.
Da Lozzo et al. [5] study the problem of producing a planar straight-line
drawing of a given embedded graph G = (V, E), i.e., G has a fixed combinatorial
embedding and a fixed outer face, such that a given set S ⊆ V of vertices is
collinear. It is clear that if such a drawing exists, then the line containing the
vertices in S is a simple curve starting and ending at infinity that for each edge
e of G either fully contains e or intersects e in at most one point, which may
be an endpoint. We call such a curve a pseudoline with respect to G. Da Lozzo
et al. [5] show that this is a full characterization of the alignment problem, i.e.,
a straight-line drawing where the vertices in S are collinear exists if and only
if there exists a pseudoline L with respect to G that contains the vertices in S.
?
Work was partially supported by grant WA 654/21-1 of the German Research Foundation (DFG). Preliminary work appeared in F. Frati and K. Ma, editors, Proceedings
of the 25th International Symposium on Graph Drawing (GD ’17).
(a)
(b)
(c)
Fig. 1: (Pseudo-) Lines are denoted by blue curves, edges are black. Aligned
Drawing (b) of a 2-aligned planar graph (a). (c) A non-stretchable arrangement
of 9 pseudolines (blue and black), which can be seen as a stretchable arrangement
of 8 pseudolines (blue) and an edge (black solid).
However testing whether such a pseudoline exists is an open problem, which we
consider in this paper.
Likewise, for the problem of separation, Biedl et al. [1] considered so-called
HH-drawings where, given an embedded graph G = (V, E) and a partition
˙
V = A∪B,
one seeks a planar drawing of G in which A and B can be separated
by a line. Again, it turns out that such a drawing exists if and only if there
exists a pseudoline L with respect to G such that the vertices in A and B are
separated by L. As a side-result Cano et al. [2] extend the result of Biedl et al.
to planar straight-line drawings with a given star-shaped outer face.
The aforementioned results of Da Lozzo et al. [5] show that given a pseudoline
L with respect to G one can always find a planar straight-line drawing of G such
that the vertices on L are collinear and the vertices contained in the halfplanes
defined by L are separated by a line L. In other words, a topological configuration
consisting of a planar graph G and a pseudoline with respect to G can always
be stretched. In this paper we initiate the study of this stretchability problem
with more than one given pseudoline.
More formally, a pair (G, A) is a k-aligned graph if G = (V, E) is a planar
embedded graph and A = {L1 , . . . , Lk } is an arrangement of (pairwise intersecting) pseudolines with respect to G. If the number k of curves is clear from the
context, we drop it from the notation and simply speak of aligned graphs. For
1-aligned graphs we write (G, L) instead of (G, {L}). Let A = {L1 , . . . , Lk } be
a line arrangement and Γ be planar drawing of G. A tuple (Γ, A) is an aligned
drawing of (G, A) if and only if the following properties hold; refer to Fig. 1(a-b).
(i) The arrangement of A is homeomorphic to the arrangement of A (i.e., A is
stretchable), (ii) Γ is homeomorphic to the planar embedding of G, (iii) each line
Li intersects in Γ the same vertices and edges as Li in G, and it does so in the
same order. We focus on straight-line aligned drawings. For brevity, unless stated
otherwise, the term aligned drawing refers to a straight-line drawing throughout
this paper.
Note that the stretchability of A is a necessary condition for the existence
of an aligned drawing. Since testing stretchability is N P-complete [14,15], we
2
assume that a geometric realization A of A is provided. Line arrangements of size
up to 8 are always stretchable [11], and only starting from 9 lines non-stretchable
arrangements exist; see the Pappus configuration [12] in Fig. 1c. This figure also
illustrates an example of an 8-aligned graph with a single edge that does not
have an aligned drawing. It is conceivable that in practical applications, e.g.,
stemming from user interactions, the number of lines to stretch is small, justifing
the stretchability assumption.
The aligned drawing convention generalizes the problems studied by Da Lozzo
et al. and Biedl et al. who concentrated on the case of a single line. We study a
natural extension of their setting and ask for alignment on general line arrangements.
In addition to the strongly related work mentioned above, there are several other works that are related to the alignment of vertices in drawings. Dujmović [6] shows that every n-vertex
√ planar graph G = (V, E) has a planar
straight-line drawing such that Ω( n) vertices are aligned, and Da Lozzo et
al. [5] show that in planar treewidth-3 and planar treewidth-k graphs, one can
align Θ(n) and Ω(k 2 ) vertices, respectively. Chaplik et al. [3] study the problem
of drawing planar graphs such that all edges can be covered by k lines. They
show that it is N P-hard to decide whether such a drawing exists. The complexity of deciding whether there exists a drawing where all vertices lie on k lines is
an open problem [4]. Drawings of graphs on n lines where a mapping between
the vertices and the lines is provided have been studied by Dujmović et al. [7,8].
Contribution & Outline. After introducing notation in Section 2, we first study
the topological setting where we are given a planar graph G and a set S of
vertices to align in Section 3. We show that the problem is N P-complete but
fixed-parameter tractable (FPT) with respect to |S|. Afterwards, in Section 4,
we consider the geometric setting where we seek an aligned drawing of an aligned
graph. Based on our proof strategy in Section 4.1, we strengthen the result of
Da Lozzo et al. and Biedl et al. in Section 4.2, and show that there exists a 1aligned drawing of G with a given convex drawing of the outer face. In Section 4.3
we consider k-aligned graphs with a stretchable pseudoline arrangement, where
every edge e either entirely lies on a pseudoline or intersects with at most one
pseudoline, which can either be in the interior or an endpoint of e. We utilize
the result of Section 4.2 to prove that every such k-aligned graph has an aligned
drawing, for any value of k.
2
Preliminaries
Let A be a pseudoline arrangement of a set of k pseudolines L1 , . . . , Lk and
(G, A) be an aligned graph with n vertices. The set of cells in A is denoted by
cells(A). A cell is empty if it does not contain a vertex of G. Removing from a
pseudoline its intersections with other pseudolines gives its pseudosegments.
Let G = (V, E) be a planar embedded graph with vertex set V and edge
set E. We call v ∈ V interior if v does not lie on the boundary of the outer
face of G. An edge e ∈ E is interior if e does not lie entirely on the boundary
3
(a) (1, 0, ⊥)
(b) (1, 0, 0)
(c) (2, 1, 0)
Fig. 2: Examples for the alignment complexity of an aligned graph.
of the outer face of G. An interior edge is a chord if it connects two vertices on
the outer face. A point p of an edge e is an interior point of e if p is not an
endpoint of e. A triangulation is a biconnected planar embedded graph whose
inner faces are all triangles and whose outer face is bounded by a simple cycle.
A triangulation of a graph G is a triangulation that contains G as a subgraph.
A k-aligned triangulation of (G, A) is a k-aligned graph (GT , A) with GT being
a triangulation of G. A Graph G0 is a subdivision of G if G0 is obtained by
placing subdivision vertices on edges of G. For a graph G and an edge e of G
the graph G/e is obtained from G by contracting e and merging the resulting
multiple edges and removing self-loops. A k-wheel is a simple cycle with k vertices
on the outer face and one interior vertex. Let Γ be a drawing of G and let C be
a cycle in G. We denote with Γ [C] the drawing of C in Γ .
A vertex is Li -aligned (or simply aligned to Li ) if it lies on the pseudoline Li .
A vertex that is not aligned is free. An edge e is Li -aligned (or simply aligned ) if it
completely lies on Li . Let Ealigned be the set of all aligned edges. An intersection
vertex lies on the intersection of two pseudolines Li and Lj . A non-aligned edge
is i-anchored (i = 0, 1, 2) if i of its endpoints are aligned to distinct curves.
An L-aligned edge is i-anchored (i = 0, 1, 2) if i of its endpoints are aligned to
distinct curves which are different from L. For example, the single aligned edge
in Fig. 2a is 1-anchored. Let Ei be the set of i-anchored edges; note that, the set
of edges is the disjoint union E0 ∪· E1 ∪· E2 . A 0-anchored, 0-crossed, non-aligned
edge is also called free. An edge e is (at most) l-crossed if (at most) l distinct
pseudolines intersect e in its interior. A non-empty edge set A ⊂ E is l-crossed
if l is the smallest number such that every edge in A is at most l-crossed.
The alignment complexity of an aligned graph G in a way describes how “complex” the relationship between the graph G and the line arrangement L1 , . . . , Lk
is and formally is defined as a triple (l0 , l1 , l2 ) , where li , i = 0, 1, 2, describes the
“complexity of i-anchored edges”, i.e., it indicates that Ei is at most li -crossed
or has to be empty, if li = ⊥. For example, an aligned graph where every vertex
is aligned and every edge has at most l interior intersections has the alignment
complexity (⊥, ⊥, l). For further examples, see Fig. 2.
4
3
Complexity and Fixed-Parameter Tractability
In this section we deal with the topological setting where we are given a planar
embedded graph G = (V, E) and a subset S ⊆ V to be collinear. According to
Da Lozzo et al. [5], this is equivalent to the existence of a pseudoline L with
respect to G passing exactly through the vertices in S. We refer to this problem
as pseudoline existence problem and the corresponding search problem is referred
to as pseudoline construction problem. Using techniques similar to Fößmeier and
Kaufmann [9], we can show that the pseudoline existence problem is N P-hard.
Let G? + V be the graph obtained from the dual graph G? = (V ? , E ? ) of
G = (V, E) by placing every vertex v ∈ V in its dual face v ? and connecting it
to every vertex on the boundary of the face v ? .
Lemma 1. Let G = (V, E) be a 3-connected 3-regular planar graph. There exists
a pseudoline through V with respect to the graph G? + V if and only if G is
Hamiltonian.
Proof. Recall that the dual of a 3-connected 3-regular graph is a triangulation.
Assume that there exists a pseudoline L through V with respect to G? + V .
Then the order of appearance of the vertices of G? + V on L defines a sequence
of adjacent faces in G? , i.e., vertices of the primal graph G that are connected
via primal edges. This yields a Hamiltonian cycle in G.
Let C be a Hamiltonian cycle of G and consider a simultaneous embedding
of G and G? + V on the plane, where each pair of a primal and its dual edge
intersects exactly once. Thus, the cycle C crosses each dual edge e at most once
and passes through exactly the vertices V . There is a vertex v on the cycle C
such that v lies in the unbounded face of G? + V . Thus, the cycle C can be
interpreted as a pseudoline L(V ) in G? + V through all vertices in V by splitting
it in the unbounded face of G? + V .
Since computing a Hamiltonian cycle in 3-connected 3-regular planar graphs
is N P-complete [10], we get that the pseudoline construction problem is N Phard. On the other hand, we can guess a sequence of vertices, edges and faces of
G, and then test in polynomial time whether this corresponds to a pseudoline L
with respect to G that traverses exactly the vertices in S. Thus, the pseudoline
construction problem is in N P. This proves the following theorem.
Theorem 1. The pseudoline existence problem is N P-complete.
In the following, we show that the pseudoline construction problem is fixedparameter tractable with respect to |S|. To this end, we construct a graph Gtr =
(V tr , E tr ) and a set S tr ⊆ V tr with |S tr | ≤ |S|+1 such that Gtr contains a simple
cycle traversing all vertices in S tr if and only if there exists a curve L that passes
exactly through the vertices of S such that (G, L) is an aligned graph.
We observe that if the vertices S of a positive instance are not independent,
they can only induce a linear forest, i.e., a set of paths, as otherwise there is
no pseudoline through the vertices S with respect to G. We call the edges on
5
e
ve
vfin
vfout
v
Fig. 3: The black and red edges depict a single face of the input graph G. Red
and blue edges build the transformed graph Gtr . Red squared vertices illustrate
the set S tr . Blue dashed edges sketch the clique edges between clique vertices
(filled blue).
these paths aligned edges. An edge that is not incident to a vertex in S is called
crossable, in the sense that only crossable edges can be crossed by L, otherwise
L is not pseudoline with respect to G. Let Sep ⊆ S be the subset of vertices
that are endpoints of the paths induced by S (an isolated vertex is a path of
length 0). We construct Gtr in several steps; refer to Fig. 3.
Step 1: Let G0 be the graph obtained from G by subdividing each aligned edge
e with a new vertex ve and let S tr be the set consisting of all isolated vertices
in S and the new subdivision vertices. Additionally, we add to G0 one new
vertex o that we embed in the outer face of G and also add to S tr . Observe
that by construction |S tr | ≤ |S| + 1. Finally, subdivide each crossable edge e
by a new vertex ve . We call these vertices traversal nodes and denote their
set by T = Sep ∪{ve | e is crossable}∪{o}. Intuitively, a curve will correspond
to a path that uses the vertices in Sep to hop onto paths of aligned edges
and the subdivision vertices of crossable edges to traverse from one face to
another. Moreover, the vertex o ∈ S tr plays a similar role, forcing the curve
to visit the outer face.
Step 2 For each face f of G we perform the following construction. Let T (f )
denote the traversal nodes that are incident to f . For each vertex v ∈ T (f )
we create two new vertices vfin and vfout , add the edges vvfin and vvfout to G0 ,
and embed them in the interior of f . Finally, we create a clique C(f ) on the
vertex set {vfin , vfout | v ∈ T (f )}, and embed its edges in the interior of f .
Step 3 To obtain Gtr remove all edges of G0 that correspond to edges of G
except those that stem from subdividing an aligned edge of G.
Lemma 2. There exists a curve L traversing exactly the vertices in S such that
(G, L) is an aligned graph if and only if there exists a simple cycle in Gtr that
traverses all vertices in S tr .
6
vfin
vfout
vfout
00
vfin00
ve
v
vfout
0
vfin0
(a)
(b)
Fig. 4: (a) A pseudoline (thick green) traversing a path of aligned edges (thin
red). (b) A path (thick green) in Gtr visiting consecutive vertices in S tr (red
squared).
Proof. Suppose C is a cycle in Gtr that visits all vertices in S tr . Without loss
of generality, we assume that there is no face f such that C contains a subpath
from vfin via v to vfout (or its reverse) for some vertex v ∈ T (f ) \ Sep . If this is
the case, we simply shortcut this path by the edge vfin vfout ∈ C(f ).
Consider a path P of aligned edges in G that contains at least one edge; refer
to Fig. 4. By definition, C visits all the subdivision vertices ve ∈ S tr of the edges
of P , and thus it enters P on an endpoint of P , traverses P and leaves P at the
other endpoint. All isolated vertices of S are contained in S tr , and therefore C
indeed traverses all vertices in S (and thus also all aligned edges). As described
above, Gtr is indeed a topological graph, and thus C corresponds to a closed
curve ρ that traverses exactly the vertices in S and the aligned edges. We now
show that ρ can be transformed to a pseudoline with respect to G.
Let e be a non-aligned edge of G that has a common point with ρ in its
interior; see Fig. 5. Thus, C contains the subdivision vertex ve . In particular,
this implies that e is crossable. Moreover, from our assumption on C, it follows
0
that C enters ve via vfin or vfout and leaves it via vfin0 or vfout
0 , where f and f are the
0
faces incident to e, and it is f 6= f as we could shortcut C otherwise. Therefore,
f0
e
vfin0
f
vfout
0
(a)
vfout
ve
vfin
(b)
Fig. 5: (a) A pseudoline (thick green) passing through a non-aligned edge. (b)
A path (thick green) in Gtr traversing a subdivision vertex ve (blue non-filled
square). Black (dashed) segments are edges of G.
7
β
α
γ
δ
Fig. 6: Resolving an intersection by exchanged the intersecting segments (red)
with non-intersecting segments (green).
ρ indeed intersects e and uses it to traverse to a different face of G. Moreover,
since e has only a single subdivision vertex in Gtr and C is simple, it follows
that e is intersected only once. Thus ρ is curve that intersects all vertices in S,
traverses all aligned edges, and crosses each edge of G (including the endpoints)
at most once. Moreover, ρ traverses the outer face since C contains o.
The only reasons why ρ is not necessarily a pseudoline with respect to G are
that it is a closed curve and it may cross itself . However, we can break ρ in
the outer face and route both ends to infinity, and remove such self-intersections
locally as follows; see Fig. 6. Consider a circle D around an intersection I that
neither contains a second self-intersection nor a vertex, nor an edge of G. Let
α, β, γ, δ be the intersections of D with L. We replace the pseudosegment αγ
with a pseudosegment αβ, and βδ with a pseudosegment γδ. We route the pseudosegments αβ and γδ through the interior of D such that they do not intersect.
respect to G that contains exactly the vertices in S such that (G, L) is an aligned
graph.
For the converse assume that L is a pseudoline that traverses exactly the
vertices in S such that (G, L) is an aligned graph. The pseudoline L can be split
into three parts L1 , L2 and L3 such that L1 and L3 have infinite length and
do not intersect with G, and L2 has its endpoints in the outer face of G. We
transform L into a closed curve L0 by removing L1 , L3 and adding a new piece
connecting the endpoints of L2 without intersecting L2 or G. Additionally, we
choose an arbitrary direction for L0 in order to determine an order of the crossed
edges and vertices.
We show that Gtr contains a simple cycle traversing the vertices in S tr . By
definition L0 consists of two different types of pieces, see Fig. 4. The first type
traverses a path of aligned edges between two vertices in Sep . The other type
traverses a face of G by entering and exiting it either via an edge or from a
vertex in Sep ; see Fig. 7. We show how to map these pieces to paths in Gtr ; the
cycle C is obtained by concatenating all these paths.
Pieces of the first type indeed correspond directly to a path in Gtr ; see Fig. 4.
Consider now a piece π of the second type traversing a face f ; refer to Fig. 7.
The piece π enters f either from a vertex in Sep or by crossing a crossable edge e.
In either case, T (f ) contains a corresponding traversal node u. Likewise, T (f )
contains a traversal node v for the edge or vertex that L0 intersects next. We
8
out
tr
map π to the path uuin
f vf v in G . By construction, paths corresponding to
0
consecutive pieces of L share a traversal node, and therefore concatenating all
paths yields a cycle C in G? . Moreover, C is simple, since L0 intersects each
edge and each vertex at most once. Note that C contains at least one edge of
the outer face f0 (as L0 traverses the outer face), and we modify C so that it
also traverses the special vertex o.
It remains to show that C contains all vertices in S tr . There are three types of
vertices in S tr ; the subdivision vertices of aligned edges, the isolated vertices in
S, and the special vertex o. The latter is in C by the last step of the construction.
The isolated vertices in S are traversed by L0 and contained in Sep , and they
are therefore visited also by C. Finally, the subdivision vertices of aligned edges
are traversed by the paths corresponding to the first type of pieces, since L0
traverses all aligned edges.
Theorem 2 (Wahlström [16]). Given a graph G = (V, E) and a subset S ⊆
V , it can be tested in O(2|S| poly(|V |)) time whether a simple cycle through the
vertices in S exists. If affirmative the cycle can be reported within the same
asymptotic time.
Theorem 3. The pseudoline construction problem can be solved in O(2|S| poly(|V |))
time.
Proof. Let G = (V, E) with S ⊆ V be an instance of the pseudoline construction
problem with n = |V |. By Lemma 2 the pseudoline construction problem is
equivalent to determining whether Gtr contains a simple cycle visiting all vertices
in S tr . Since the size of Gtr is O(n2 ) and it can be constructed in O(n2 ) time,
and |S tr | ≤ |S| + 1, Theorem 2 can be used to solve the latter problem in the
desired running time.
We note that indeed the construction of Gtr only allows to leave a path of
aligned edges at an endpoint in Sep . Therefore, a single vertex in S tr for each
path of aligned edges would be sufficient to ensure that C traverses the whole
path. Thus, by removing all but one vertex for each such path from S tr we obtain
an algorithm that is FPT with respect to the number of paths induced by S.
u
uout
f
uin
f
vfin
vfout
(a)
v
(b)
Fig. 7: (a) A pseudoline piece π (thick green) passing through a face f . (b) Path
(thick green) in Gtr corresponding to π.
9
Corollary 1. The pseudoline construction problem can be solved in O(2P poly(|V |))
time, where P is the number of paths induced by the vertex set S to be aligned.
4
Drawing Aligned Graphs
We show that every aligned graph where each edge either entirely lies on a
pseudoline or is intersected by at most one pseudoline, i.e., alignment complexity
(1, 0, ⊥), has an aligned drawing. For 1-aligned graphs we show the stronger
statement that every 1-aligned graph has an aligned drawing with a given aligned
convex drawing of the outer face. We first present our proof strategy and then
deal with 1- and k-aligned graphs.
4.1
Proof Strategy
Our general strategy for proving the existence of aligned drawings of an aligned
graph (G, A) is as follows. First, we show that we can triangulate (G, A) by
adding vertices and edges without invalidating its properties. We can thus assume
that our aligned graph (G, A) is an aligned triangulation. Second we show that,
unless G is a specific graph (e.g., a k-wheel or a triangle), it contains an aligned
or a free edge. Third, we exploit the existence of such an edge to reduce the
instance. Depending on whether the edge is contained in a separating triangle
or not, we either decompose along that triangle or contract the edge. In both
cases the problem reduces to smaller instances that are almost independent. In
order to combine solutions, it is, however, crucial to use the same arrangement
of lines A for both of them.
In the following we introduce the necessary tools used for all three steps on
k-aligned graphs of alignment complexity (1, 0, ⊥). Recall, that for this class
(i) every non-aligned edge is 1-crossed, (ii) every 1-anchored edge has no intersections, and (iii) there is no edge with its endpoints on two pseudolines.
Lemmas 3 – 6 shows that every aligned graph has an aligned triangulation with the same alignment complexity. If G contains a separating triangle,
Lemma 7 shows that (G, A) admits an aligned drawing if both split components have an aligned drawing. Finally, with Lemma 8 we obtain a drawing of
(G, A) from a drawing of the aligned graph (G/e, A) where one special edge e is
contracted.
Lemma 3. Let (G, A) be a k-aligned graph with two distinct vertices u, v incident to a common face f . Then there is a k-aligned graph (G0 , A) obtained from
G by adding a path P from u to v in the interior of f . The edge set E(P ) has
alignment complexity (1, 0, ⊥) and does not contain aligned edges. The size of P
is in O(k 2 n).
Proof. We obtain a k-aligned graph (G0 , A) by routing a path P from u to v
through the interior of f ; compare Fig. 8. We construct P such that it has a
vertex in every cell of A it traverses. As each of the O(k 2 ) cells in A may be
subdivided into O(n) smaller cells by f , a shortest path from u to v contains
O(k 2 n) vertices.
10
u
v
Fig. 8: Connecting two vertices in black pseudoline arrangement A and the interior of a green face f .
qi
qi
u
y
v
w
(a)
w
v
y
x
v
x u
v
y
(b)
(c)
x
(d)
Fig. 9: Black lines indicate a face f of G. Light green edges or vertices are newly
added into f . Blue lines denote the pseudoline arrangement. (a) Isolation of
an intersection. (b-c) Isolation of an aligned edge or vertex. (d) Isolation of a
pseudosegment.
Lemma 4. Let (G, A) be a k-aligned graph. Then there exists a biconnected kaligned graph (G0 , A) that contains G as a subgraph. The set E(G0 ) \ E(G) has
alignment complexity (1, 0, ⊥) and does not contain aligned edges. The size of
E(G0 ) \ E(G) is in O(k 2 n2 ).
Proof. While G is not connected, we use Lemma 3 to connect two vertices of
distinct connected components that lie on the boundary of a common face. Afterwards, we use Lemma 3 to iteratively connect vertices from different biconnected
components. This results in the claimed biconnected k-aligned graph (G0 , A).
Since there are O(n) connected components and O(n) cut vertices, it follows
that Lemma 3 is applied O(n) times, and hence G0 has size O(k 2 n2 ).
Lemma 5. Let (G, A) be a biconnected k-aligned graph of alignment complexity
(1, 0, ⊥). There exists a k-aligned triangulation (GT = (VT , ET ), A) of (G, A)
whose size is O(k 4 n). The set E(GT ) \ E(G) has alignment complexity (1, 0, ⊥)
and does not contain aligned edges.
Proof. We call a face non-triangular, if its boundary contains more than three
vertices. An aligned vertex v or an aligned edge e is isolated if all faces with v
11
or e on its boundary are triangles. A pseudosegment s is isolated if s does not
intersect the interior of a non-triangular face. Our proof distinguishes four cases.
Each case is applied exhaustively in this order. Let f be a non-triangular face.
1. If the interior of f contains the intersection of two or more pseudolines, we
split the face, so that there is a vertex that lies on the intersection.
2. If the boundary of a face has an aligned vertex or an aligned edge, we isolate
the vertex or the edge from f .
3. If the interior of a face f is intersected by a pseudoline L, then it subdivides
L into a set of pseudosegments. We isolate each of the psedudosegments
independently.
4. Finally, if none of the previous cases apply, i.e., neither the boundary nor
the interior of f contain parts of a pseudoline, the face f can be triangulated
with a set of additional free edges.
Let Af be the arrangement of A and f restricted to the interior of f .
1. Let f be a non-triangular face whose interior contains an intersection of
a set of pseudolines; see Fig. 9a. We place a vertex x on the intersection.
We obtain a biconnected aligned graph with the application of Lemma 4.
There are at most O(k 2 ) intersections. Isolating each uses at most O(k 2 n)
additional vertices and edges. Thus, in total the isolation of all intersections
increases the graph size by O(k 4 n).
2. Let f be a non-triangular face with an aligned vertex or an aligned edge uv
on its boundary. Further, the interior of f does not contain the intersection
of a set of pseudolines; see Fig. 9b and 9c. In case of an aligned vertex we
simply assume u = v. Since G is biconnected, there exist two edges xu, vy on
the boundary of f . There are at most 2k cells C 1 , . . . , C l ∈ cells(Af ), l ≤ 2k
with u or v on their boundary, such that C i is adjacent to C i+1 , i < l. We
construct an aligned graph (G0 , A) from (G, A) as follows. We place a vertex
qi in the interior of each cell C i , i ≤ l. Let q0 = x and ql+1 = y. We insert
edges ei = qi qi+1 , i = 0, . . . , l in the interior of f and so that the interior of
ei crosses the common boundary of C i and C i+1 exactly once and it crosses
no other boundary. Thus, each edge ei is either at most 1-anchored and 0crossed, if it is incident to x or y, or otherwise 0-anchored and 1-crossed. The
added path splits f into two face f 0 , f 00 with a unique face f 0 containing u
and v on its boundary. If w ∈ {u, v} is on the boundary of cell C i , we insert
an edge wqi . Each edge wqi is 1-anchored and 0-crossed. Let C i and C i+1
be two cells incident to w. Then, the vertices w, qi , qi+1 form a triangle. If
u 6= v, there is a unique cell C i incident to u and v. Hence, the vertices u, v, qi
form a triangle. Moreover, for 1 ≤ i ≤ l, every edge uqi and vqi is incident to
two triangles. Therefore, f 0 is triangulated. The number of aligned edges and
aligned vertices of f 00 is one less compared to f . Hence, we can inductively
proceed on f 00 .
Assume the aligned vertex v is a vertex on an intersection of a set of pseudolines. Then, v is incident to O(k) cells and lies on the boundary of at most
deg(v) faces. Hence, we can isolate it with O(k + deg(v)) additional vertices
and edges. Since there are O(k 2 ) intersections in the pseudoline arrangement,
12
we can isolate all intersection vertices with O(k 2 (k + n)) additional edges
and vertices.
Now consider an aligned vertex v that is not an intersection vertex. In this
case v is incident to at most two cells and O(deg(v)) faces. We can isolate
all such vertices with O(n) additional vertices and edges. The same bound
holds for aligned edges.
3. Let f be a non-triangular face whose interior is intersected by a pseudoline L
and has no aligned edge and no aligned vertex on its boundary. Further, the
interior of f does not contain the intersection of two or more pseudolines.
Then the face f subdivides L into a set of pseudosegments S; see Fig. 9d.
We iteratively isolate such a pseudosegment S. Since f does not contain
the intersection of a set of pseudolines in its interior, there are two distinct
cells C 1 ∈ cells(Af ) and C 2 ∈ cells(Af ) with S on their boundary. Since f
neither contains an aligned vertex nor an aligned edge and G is biconnected,
there are exactly two edges e1 = vw and e2 = xy with the endpoints of S
in the interior of these edges and v, x and w, y on the boundaries of C 1 and
C 2 , respectively. Since f does not have an l-crossed edge, l ≥ 2, and every
1-crossed edge is 0-anchored, the vertices v, w, x, y are free vertices. We
construct a graph G0 by placing a vertex u on s and insert edges uv, uw, ux
uy, vx and wy. We route each edge so that it the interior of an edge does
not intersect the boundary of a cell C i . Thus, the edges vx and wy are free
and the others are 1-anchored and 0-crossed.
Every edge in f is at most 1-crossed, thus the number of pseudosegments is
linear in the size of G. Thus, we add only a number of vertices and edges
linear in the size of G.
4. If none of the cases above apply to a non-triangular face f , then neither the
interior nor the boundary of the face intersect a pseudoline Li . Thus, we can
triangulate f with a number of free edges linear in the size of f , i.e., in total
with O(k 4 n) edges.
Thus, in total there is an aligned triangulation (GT , A) of G whose size is in
O(k 4 n + k 2 (k + n) + n + k 4 n) = O(k 4 n).
Observe that the previous triangulation only uses the fact that every nontriangular face contains at most 1-crossed edges. While Lemma 5 is sufficient for
our purposes, for the sake of generality, we show how to isolate l-crossed edges.
This allows us to triangulate arbitrary aligned graphs without increasing the
alignment complexity.
Theorem 4. Every biconnected aligned graph (G, A) of alignment complexity
(l0 , l1 , l2 ) has an aligned triangulation (GT , A). The alignment complexity of
E(GT ) \ E(G) is (max{l0 , 1}, l1 , l2 ) and the size of this set is O(k 4 n).
Proof. For l ≥ 1, we iteratively isolate l-crossed edges uv from a non-triangular
face f as sketched in Fig. 10. Let C 0 , C 1 , . . . , C l ∈ cells(A) be the cells in f that
occur in this order along uv. If one of both vertices is free, say v, we place a new
vertex x in the interior of C l−1 . We insert the two edges ux, xv and route both
edges close to uv. This isolates the edge uv from f . Notice that the edge vx is
13
C1
Cl−1
x
Cl
v
u
Fig. 10: An l-crossed edge uv in a (grey) face f and a pseudoline arrangement
(blue). The green edges isolate the edge uv.
0-anchored and 1-crossed. In case that l0 ≥ 1, the alignment complexity of the
new aligned graph is (l0 , l1 , l2 ). Otherwise, the alignment complexity is (1, l1 , l2 ).
If u and v are aligned, we place x on the boundary of C l−1 and C l and route
the edges ux and vx as before. The alignment complexity is not affected by this
operation.
This procedure introduces a new (l − 1)-crossed edge. Repeating the process
l − 2 generates a new face f 0 from f where edge uv is substituted by a path of
at most 1-crossed edges. To isolate all l-crossed edge in (G, A), we add O(kn)
vertices and edges.
By isolating all l-crossed edges in this way, we obtain an aligned graph where
every non-triangular face is bounded by at most 1-crossed edges. Thus, we can
apply the methods used in the proof of Lemma 5 to triangulate (G, A) with
O(k 4 n) number of additional vertices and edges.
We now return to the treatment of aligned graphs with alignment complexity
(1, 0, ⊥). To simplify the proofs, we augment the input graph with an additional
cycle in the outer face that contains all intersections of A in its interior, and we
add subdivision vertices on the intersections of Li -aligned edges with pseudolines
Lj , i 6= j. A k-aligned graph is proper if (i) every aligned edge is 0-crossed, (ii) for
k ≥ 2, every edge on the outer face is 1-crossed, (iii) the boundary of the outer
face intersects every pseudoline exactly twice, and (iv) the outer face does not
contain the intersections of A.
An aligned graph (G0 , A) is a rigid subdivision of an aligned graph (G, A) if
and only if G0 is a subdivision of G and every subdivision vertex is an intersection
vertex with respect to A. We show that we can extend every k-aligned graph
(G, A) to a proper k-aligned triangulation (G0 , A).
Lemma 6. For k ≥ 2 and every k-aligned graph (G, A) of alignment complexity
(1, 0, ⊥), let (Grs , A) be a rigid subdivision of (G, A). Then there exists a proper
k-aligned triangulation (G0 , A) of alignment complexity (1, 0, ⊥) such that Grs
is a subgraph of G0 . The size of G0 is in O(k 4 n2 ). The set E(G0 ) \ E(G) has
alignment complexity (1, 0, ⊥) and does not contain aligned edges.
Proof. We construct a rigid subdivision (Grs , A) from (G, A) by placing subdivision vertices on the intersections of Li -aligned edges with pseudolines Lj , i 6= j.
14
We obtain a proper k-aligned triangulation (G0 , A) by embedding a simple cycle
C in the outer face of Grs and applying Lemma 4 and Lemma 5. In order to
construct C, we place a vertex vc in each unbounded cell c of A and connect two
vertices vc and vc0 if the cells c and c0 share a common boundary.
The following two lemmas show that we can reduce the size of the aligned
graph and obtain a drawing by merging two drawings or by unpacking an edge.
Lemma 7. Let (G, A) be a k-aligned triangulation. Let T be a separating triangle splitting G into subgraphs Gin , Gout so that Gin ∩Gout = T and Gout contains
the outer face of G. Then, (i) (Gout , A) and (Gin , A) are k-aligned triangulations,
and (ii) (G, A) has an aligned drawing if and only if there exists a common line
arrangement A such that (Gout , A) has an aligned drawing (Γout , A) and (Gin , A)
has an aligned drawing (Γin , A) with the outer face drawn as Γout [T ].
Proof. It is easy to verify that (Gout , A) and (Gin , A) are aligned triangulations.
An aligned drawing (Γ, A) of (G, A) immediately implies the existence of an
aligned drawing (Γout , A) of (Gout , A) and (Γin , A) of (Gin , A).
Let (Γout , A) be an aligned drawing of (Gout , A). Since (Γout , A) is an aligned
drawing, (Γout [T ], A) is an aligned drawing of (T, A). Let (Γin , A) be an aligned
drawing of (Gin , A) with the outer face drawing as Γout [T ]. Let Γ be the drawing
obtained by merging the drawing Γout and Γin . Since, (Γout , A) and (Γin , A) are
aligned drawings on the same line arrangement A, (Γ, A) is an aligned drawing
of (G, A).
Lemma 8. Let (G, A) be a proper k-aligned triangulation of alignment complexity (1, 0, ⊥) and let e be an interior 0-anchored aligned or an interior free
edge of G that does not belong to a separating triangle and is not a chord. Then
(G/e, A) is a proper k-aligned triangulation of alignment complexity (1, 0, ⊥).
Further, (G, A) has an aligned drawing if (G/e, A) has an aligned drawing.
Proof. We first prove that (G/e, A) is a proper k-aligned triangulation. Consider
a topological drawing of the aligned graph (G, A). Let c be the vertex in G/e
obtained from contracting the edge e = uv. We place c at the position of u.
Thus, all the edges incident to u keep their topological properties. We route the
edges incident to v close to the edge uv within the cell from which they arrive
to v in (G, A). Since e is not an edge of a separating triangle, G/e is simple and
triangulated.
Consider a free edge e. Observe that the triangular faces incident to e do not
contain an intersection of two pseudolines in their interior, since (G, A) does not
contain l-crossed edges, l ≥ 2. Therefore, (G/e, A) is an aligned triangulation.
Since e is not a chord, (G/e, A) is proper. Further, u and v lie in the interior of
the same cell, thus, the edges incident to c have the same alignment complexity.
If e is aligned, it is also 0-crossed, since (G, A) is proper. Since e is also
0-anchored, the triangles incident to e do not contain an intersection of two
pseudolines and therefore (G/e, A) is a proper aligned triangulation. The routing of the edges incident to c, as described above, ensures that the alignment
complexity is (1, 0, ⊥).
15
Nu
Nu
Nv
Nv
(a)
(b)
Fig. 11: Unpacking an edge in a drawing Γ 0 of G/e (a) to obtain a drawing Γ
of G (b).
Let (Γ 0 , A) be an aligned drawing of (G/e, A). We now prove that (G, A) has
an aligned drawing. Let Γ 00 denote the drawing obtained from Γ 0 by removing
c together with its incident edges and let f denote the face of Γ 00 where c used
to lie. Since G/e is triangulated and e is an interior edge and not a chord, f is
star-shaped and c lies inside the kernel of f ; see Fig. 11. We construct a drawing
Γ of G as follows. If one of vertices u and v lie on the outer face, we assume,
without loss of generality, that vertex to be u. First, we place u at the position
of c and insert all edges incident to u. This results in a drawing of the face f 0 in
which we have to place v. Since, u is placed in the kernel of f , f 0 is star-shaped.
If e is a free edge, the vertex v has to be placed in the same cell as u. We then
place v inside f 0 sufficiently close to c so that it lies inside the kernel of f 0 and
in the same cell as u. All edges incident to v are at most 1-crossed, thus, (Γ, A)
is an aligned drawing of (G, A).
Likewise, if e is an L-aligned edge, then v has to be placed on the corresponding to L line L ∈ A. In this case also c and therefore u lie on L. Since e is an
interior edge, there exist two triangles uv, vx, xu and uv, vy, yu sharing the edge
uv. Since, e is not part of a separating triangle, x and y are on different sides of
L. Therefore the face f 0 contains a segment of the line L of positive length that
is within the kernel of f 0 . Thus, we can place v close to u on the line L such that
the resulting drawing is an aligned drawing of (G, A).
4.2
One Pseudoline
We show that every 1-aligned graph (G, R) has an aligned drawing (Γ, R), where
R is a single pseudoline and R is the corresponding straight-line. Using the
techniques from the previous section, we can assume that (G, R) is a proper 1aligned triangulation. We show that unless G is very small, it contains a special
edge. This enables an inductive proof to construct an aligned drawing of (G, R).
16
x
f0
x
v
v
(a)
(b)
f
Fig. 12: Transformation from a red vertex (a) to a gray vertex (b).
Lemma 9. Let (G, R) be a proper 1-aligned triangulation without chords and
with k vertices on the outer face. If G is neither a triangle nor a k-wheel whose
center is aligned, then (G, R) contains an interior aligned or an interior free
edge.
Proof. We first prove two useful claims.
Claim 1. Consider the order in which R intersects the vertices and edges of G.
If vertices u and v are consecutive on R, then the edge uv is aligned.
Observe that the edge uv can be inserted into G without creating crossings.
Since G is a triangulation, it therefore contains uv, and further, since no 1-crossed
edge can have both its endpoints on R, it follows that indeed uv is aligned. This
proves the claim.
Claim 2. Let (G, R) be an aligned triangulation without aligned edges and let x
be an interior free vertex of G, then x is incident to a free edge.
Assume for a contradiction that all neighbors of x lie either on R or on the
other side of R. First, we slightly modify R to a curve R0 that does not contain
any vertices. Assume v is an aligned vertex; see Fig. 12. Since there are no aligned
edges, R enters v from a face f incident to v and leaves it to a different face f 0
incident to v. We then reroute R from f to f 0 locally around v. If v is incident
to x, we choose the rerouting such that it crosses the edge vx.
Notice that if an edge e is intersected by R in its endpoints, then R0 either
does not intersect it, or intersects it in an interior point. Moreover, e cannot be
intersected by R0 twice as in such case R would pass through both its endpoints.
Now, since G is a triangulation and the outer face of G is proper, R0 is a simple
cycle in the dual G? of G, and hence corresponds to a cut C of G. Let H denote
the connected component of G − C that contains x and note that all edges of
H are free. By the assumption and the construction of R0 , x is the only vertex
in H. Thus, R0 intersects only the faces incident to x which are interior. This
contradicts the assumption that R0 passes through the outer face of G. This
finishes the proof of the claim.
We now prove the lemma. Assume that G is neither a triangle nor a k-wheel
whose center is aligned. If G is k-wheel whose center is free, we find a free edge
17
by Claim 2. Otherwise, G contains at least two interior vertices. If one of both
vertices is free, we find a free edge by Claim 2. Otherwise, there is no free internal
vertex, therefore the only edge which can intersect R is a chord of G. Since G
does not contain any chord, there is a pair of aligned vertices consecutive along
R. Thus by Claim 1 the instance (G, R) has an aligned edge.
Theorem 5. Let (G, R) be a proper aligned graph and let (ΓO , R) be a convex
aligned drawing of the aligned outer face (O, R) of G. There exists an aligned
drawing (Γ, R) of (G, R) with the same line R and the outer face drawn as ΓO .
Proof. Given an arbitrary proper aligned graph (G, R), we first complete it to a
biconnected graph and then triangulate it by applying Lemma 4 and Lemma 5,
respectively.
We prove the claim by induction on the size of G. If G is just a triangle,
then clearly (ΓO , R) is the desired drawing. If G is the k-wheel whose center is
aligned, placing the vertex on the line in the interior of ΓO yields an aligned
drawing of G. This finishes the base case.
If G contains a chord e, then e splits (G, R) into two graphs G1 , G2 with
G1 ∩ G2 = e. It is easy to verify that (Gi , R) is an aligned graph. Let (ΓOi , R)
be a drawing of the face of ΓO ∪ e whose interior contains Gi . By the inductive
hypothesis, there exists an aligned drawing of (Γi , R) with the outer face drawn
as (ΓOi , R). We obtain a drawing Γ by merging the drawings Γ1 and Γ2 . The
fact that both (Γ1 , R) and (Γ2 , R) are aligned drawings with a common line R
implies that (Γ, R) is an aligned drawing of (G, R).
If G contains a separating triangle T , let Gin and Gout be the respective
split components with Gin ∩ Gout = T . By Lemma 7, the graphs (Gin , R) and
(Gout , R) are aligned graphs. By the induction hypothesis there exists an aligned
drawing (Γout , R) of the aligned graphs (Gout , R) with the outer face drawn as
(ΓO , R). Let Γ [T ] be the drawing of T in Γout . Further, (Gin , R) has by induction
hypothesis an aligned drawing with the outer face drawn as Γ [T ]. Thus, by
Lemma 7 we obtain an aligned drawing of (G, R) with the outer face drawn as
ΓO .
If G is neither a triangle nor a k-wheel, by Lemma 9, it contains an interior
aligned or an interior free edge e. Since e is not a chord and does not belong
to a separating triangle, by Lemma 8, (G/e, R) is an aligned graph and by the
induction hypothesis it has an aligned drawing (Γ 0 , R) with the outer face drawn
as ΓO . It thus follows by Lemma 8 again that (G, R) has an aligned drawing
with the outer face drawn as ΓO .
4.3
Alignment Complexity (1, 0, ⊥)
We now consider k-aligned graphs (G, A) of alignment complexity (1, 0, ⊥), i.e.,
every edge has at most one interior intersection and 2-anchored edges are forbidden. In this section, we prove that every such k-aligned graph has an aligned
drawing. As before we can assume that (G, A) is a proper aligned triangulation.
We show that if the structure of the graph is not sufficiently simple, it contains
18
Fig. 13: All possible variations of vertices and edges in Lemma 10.
a special edge. Further, we prove that every graph of this structure indeed has
an aligned drawing. Together this again enables an inductive proof that (G, A)
has an aligned drawing. Fig. 13 illustrates the statement of the following lemma.
Lemma 10. Let (G, A) be a proper k-aligned triangulation of alignment complexity (1, 0, ⊥) that neither contains a free edge, nor a 0-anchored aligned edge,
nor a separating triangle. Then (i) every intersection contains a vertex, (ii) every
cell of the pseudoline arrangement contains exactly one free vertex, (iii) every
pseudosegment is either covered by two aligned edges or intersected by a single
edge.
Proof. The statement follows trivially from the following sequence of claims. We
refer to an aligned vertex that is not an intersection vertex as a flexible aligned
vertex.
Claim 1. Every pseudosegment alternately intersects flexible aligned vertices and
edges.
Let S be a pseudosegment in the pseudoline arrangement A. As in the proof
of Lemma 9 one can argue that if two vertices occur consecutively along S, then
we find an aligned edge. Assume that S intersects two edges e1 , e2 consecutively
as depicted in Fig. 14a. Since G is a triangulation, it follows that e1 and e2 share
an endpoint v. Every 1-crossed edge in G is 0-anchored, thus all endpoints of e1
and e2 must be free. Further e1 and e2 are consecutive in the circular order of
edges around v as otherwise we would either find an intersection with S between
e1 and e2 or a free edge. Thus, e1 and e2 bound a face and are 1-crossed, hence
their endpoints distinct from v are in the same cell and connected by an edge
e, which is thus free. In a proper aligned graph, the edges on the outer face are
1-crossed, thus, e is an interior edge, contradicting our assumptions.
Claim 2. Every cell contains at least one free vertex.
Observe that every triangle T containing the intersection of two pseudolines
has at least one 2-anchored edge or one l-crossed edge, with l ≥ 2. Since by
definition (G, A) neither containst 2-anchored aligned edges, T cannot contain
an empty cell in its interior, i.e., a triangle of aligned edges. Further, since G is
proper, the outer face of G contains the intersection of every pair of pseudolines
19
u
e
e1
(a)
B
e2
v
y
x
v
w
(b)
(c)
(d)
Fig. 14: Illustrations for the proof Lemma 10.
in its interior. Thus, since G is triangulated every cell contains at least one
vertex.
Claim 3. Every cell contains at most one free vertex.
The following proof is similar to Claim 2 in the proof of Lemma 9. Let
C be a cell and assume for the sake of a contradiction that C contains more
than one vertex in its interior; see Fig. 14b. These vertices are connected by
edges to adjacent cells. If C contains a vertex v on its boundary, we reroute the
corresponding pseudolines close to v such that v is now outside of C; refer to
Fig. 14c. Let C 0 be the resulting cell, it represents a cut in the graph with two
components A and B, where C 0 contains B in its interior. It is not difficult to
see that the modified pseudolines are still pseudolines with respect to G. Since
(G, A) neither contains 2-anchored edges nor l-crossed edges, l ≥ 2, every edge
of (G, A0 ) intersects the boundary of C 0 at most once. Hence, B is connected
and since it contains at least two vertices it also contains at least one free edge,
contradicting our initial assumption.
Claim 4. Every flexible aligned vertex is incident to two 1-anchored aligned edges.
Let v be a vertex lying in the interior of a pseudosegment S. Let u and w
be the anchored vertices incident to S, see Fig. 14d. Further, let x and y be the
vertices in the interior of the two cells incident to S. Our instance (G, A) is triangulated and every edge is at most 1-crossed. Thus, the vertices u, x, w, y build
a quadrangle containing v in its interior. Since G does not contain a separating
triangles, it cannot contain the edge xy. Moreover, S contains exactly v in its
interior, otherwise we would find a free aligned edge. Finally, since (G, A) is an
aligned triangulation, the vertex v is connected to all four vertices and, thus
incident to two 1-anchored aligned edges.
Since (G, A) is an aligned triangulation, Property (iii) immediately follows
from Claims 3 and 4.
Lemma 11. Let (G, A) be a proper k-aligned triangulation of alignment complexity (1, 0, ⊥) that does neither contain a free edge, nor a 0-anchored aligned
edge, nor a separating triangle. Let A be a line arrangement homeomorphic to
the pseudoline arrangement A. Then (G, A) has an aligned drawing (Γ, A).
20
Proof. We obtain a drawing (Γ, A) by placing a free vertex in its cell, an aligned
vertex on its pseudosegment and an intersection vertex on its intersection. According to Lemma 10 every cell and every intersection contains exactly one
vertex and each pseudosegment is either crossed by an edge or it is covered by
two aligned edges. Observe that the union of two adjacent cells of the arrangement A is convex. Thus, this drawing of G has an homeomorphic embedding to
(G, A) and every edge intersects in (Γ, A) the line L ∈ A corresponding to the
pseudoline L ∈ A in (G, A)
The following theorem can be proven along the same lines as Theorem 5.
Theorem 6. Every k-aligned graph (G, A) of alignment complexity (1, 0, ⊥)
with a stretchable pseudoline arrangement A has an aligned drawing.
Proof. Let (G, A) be an arbitrary aligned graph, such that A is a stretchable
pseudoline arrangement, let denote by A the corresponding line arrangement.
By Lemma 6, we obtain a proper k-aligned triangulation (GT , A) that contains
a rigid subdivision of G as a subgraph. Assume that (GT , A) has an aligned
drawing (ΓT , A). Let (Γ 0 , A) be the drawing obtained from (ΓT , A) by removing
all subdivision vertices v and merging the two edges incident to v at the common
endpoint. Recall, that a subdivision vertex in a rigid subdivision of (G, A) lies on
an intersection in A. Hence the drawing (Γ 0 , A) is a straight-line aligned drawing
and contains an aligned drawing (Γ, A) of (G, A).
We now show that (GT , A) indeed has an aligned drawing. We prove this
by induction on the size of the instance (GT , A). If (GT , A) neither contains
a free edge, nor a 0-anchored aligned edge, nor a separating triangle, then, by
Lemma 11 there is an aligned drawing (ΓT , A).
If G contains a separating triangle T , let Gin and Gout be the respective split
components with Gin ∩ Gout = T . Since the alignement complexity of (G, A)
is (1, 0, ⊥), triangle T is intersected by most one pseudoline L. It follows that,
(Gout , A) is a k-aligned triangulation and that (Gin , L) is a 1-aligned triangulation. By the induction hypothesis there exists an aligned drawing (Γout , A) of
(Gout , A). Let Γout [T ] be the drawing of T in Γout . By Theorem 5, we obtain an
aligned drawing (Γin , L) with T drawn as Γout [T ]. Moreover, since the drawing
of T is fixed and is intersected only by line L, (Γin , A) is an aligned drawing.
Thus, according to Lemma 7, there exists an aligned drawing of (G, A).
If GT does not contain separating triangles but contains either a free edge
or a 0-anchored aligned edge e, let GT /e be the graph after the contraction
of e. Observe that, since (GT , A) is proper, every edge on the outer face is 1crossed, and therefore every chord is `-crossed, ` ≥ 1. Thus, e is an interior
edge of (GT , A) and is not a chord. Thus, by Lemma 8, (GT /e, A) is a proper
aligned triangulation. By induction hypothesis, there exists an aligned drawing
of (GT /e, A), and thus, by the same lemma, there exists an aligned drawing of
(GT , A).
21
Fig. 15: Sketch of a 2-aligned triangulation without aligned or free edges. The
triangulation can be completed as indicated by the black edges.
5
Conclusion
In this paper we showed that if A is stretchable, then every k-aligned graph
(G, A) of alignment complexity (1, 0, ⊥) has a straight-line aligned drawing. As
an intermediate result we showed that a 1-aligned graph (G, R) has an aligned
drawing with a fixed convex drawing of the outer face. We showed that the less
restricted version of this problem, where we are only given a set of vertices to
be aligned, is N P-hard but fixed-parameter tractable.
The case of more general alignment complexities is widely open. Our techniques imply the existence of one-bend aligned drawings of general 2-aligned
graphs [13]. However, the existence of straight-line aligned drawings is unknown
even if in addition to 1-crossed edges, we only allow 2-anchored edges, i.e., in the
case of alignment complexity (1, 0, 0). In particular, there exist 2-aligned graphs
that neither contain a free edge nor an aligned edge but their size is unbounded
in the size of the arrangement; see Fig. 15. It seems that further reductions are
necessary to arrive at a base case that can easily be drawn. This motivates the
following questions.
1) What are all the combinations of line numbers k and alignment complexities
C such that for every k-aligned graph (G, A) of alignment complexity C there
exists a straight-line aligned drawing provided A is stretchable.
2) Given a k-aligned graph (G, A) and a line arrangement A homeomorphic to
A, what is the complexity of deciding whether (G, A) admits a straight-line
aligned drawing (Γ, A)?
References
1. T. C. Biedl, M. Kaufmann, and P. Mutzel. Drawing planar partitions II: HHdrawings. In J. Hromkovič and O. Sýkora, editors, WG’98, pages 124–136. Springer,
22
1998.
2. J. Cano, C. D. Tóth, and J. Urrutia. Upper bound constructions for untangling
planar geometric graphs. SIAM Journal on Discrete Mathematics, 28(4):1935–
1943, 2014.
3. S. Chaplick, K. Fleszar, F. Lipp, A. Ravsky, O. Verbitsky, and A. Wolff. Drawing
graphs on few lines and few planes. In Y. Hu and M. Nöllenburg, editors, GD’16,
pages 166–180. Springer, 2016.
4. S. Chaplick, K. Fleszar, F. Lipp, A. Ravsky, O. Verbitsky, and A. Wolff. The complexity of drawing graphs on few lines and few planes. In F. Ellen, A. Kolokolova,
and J.-R. Sack, editors, Algorithms and Data Structures: 15th International Symposium, WADS 2017, pages 265–276. Springer, 2017.
5. G. Da Lozzo, V. Dujmović, F. Frati, T. Mchedlidze, and V. Roselli. Drawing planar
graphs with many collinear vertices. In Y. Hu and M. Nöllenburg, editors, GD’16,
pages 152–165. Springer, 2016.
6. V. Dujmović. The utility of untangling. In E. Di Giacomo and A. Lubiw, editors,
GD’15, pages 321–332. Springer, 2015.
7. V. Dujmović, W. Evans, S. Kobourov, G. Liotta, C. Weibel, and S. Wismath. On
graphs supported by line sets. In U. Brandes and S. Cornelsen, editors, GD’10,
pages 177–182. Springer, 2011.
8. V. Dujmović and S. Langerman. A center transversal theorem for hyperplanes and
applications to graph drawing. Discrete & Computational Geometry, 49(1), 2013.
9. U. Fößmeier and M. Kaufmann. Nice drawings for planar bipartite graphs. In
Algorithms and Complexity, Third Italian Conference, CIAC ’97, pages 122–134,
1997.
10. M. R. Garey, D. S. Johnson, and R. Tarjan. The planar Hamiltonian circuit
problem is NP-complete. SIAM Journal on Computing, 5(4):704–714, 1976.
11. J. E. Goodman and R. Pollack. Proof of Grünbaum’s conjecture on the stretchability of certain arrangements of pseudolines. Journal of Combinatorial Theory,
Series A, 29(3):385 – 390, 1980.
12. F. Levi. Die Teilung der projektiven Ebene durch Geraden oder Pseudogerade.
Ber. Math.-Phys. Kl. Sächs. Akad. Wiss, 78:256–267, 1926.
13. T. Mchedlidze, M. Radermacher, and I. Rutter. Aligned drawings of planar graphs.
In Proceedings of the 33rd European Workshop on Computational Geometry (EuroCG 2017), 2017.
14. N. E. Mnëv. The universality theorems on the classification problem of configuration varieties and convex polytopes varieties. In O. Y. Viro and A. M. Vershik,
editors, Topology and Geometry — Rohlin Seminar, volume 1346 of LNCS, pages
527–543. Springer, 1988.
15. P. Shor. Stretchability of pseudolines is NP-hard. In Applied Geometry and Discrete
Mathematics–The Victor Klee Festschrift, volume 4 of DIMACS Series in Discrete
Mathematics and Theoretical Computer Science, pages 531–554. Amer. Math. Soc.,
1991.
16. M. Wahlström. Abusing the Tutte matrix: An algebraic instance compression for
the K-set-cycle problem. In 30th International Symposium on Theoretical Aspects
of Computer Science, STACS 2013, pages 341–352, 2013.
23
| 8cs.DS
|
Dominating Sets in Intersection Graphs
of Finite Groups
Selçuk Kayacan∗
arXiv:1602.03537v2 [math.GR] 15 Jul 2017
Department of Mathematics, Istanbul Technical University,
34469 Maslak, Istanbul, Turkey.
e-mail: [email protected]
Abstract
Let G be a group. The intersection graph Γ(G) of G is an undirected
graph without loops and multiple edges defined as follows: the vertex set is
the set of all proper non-trivial subgroups of G, and there is an edge between
two distinct vertices H and K if and only if H ∩ K 6= 1 where 1 denotes
the trivial subgroup of G. In this paper we studied the dominating sets in
intersection graphs of finite groups. It turns out a subset of the vertex set
is a dominating set if and only if the union of the corresponding subgroups
contains the union of all minimal subgroups. We classified abelian groups
by their domination number and find upper bounds for some specific classes
of groups. Subgroup intersection is related with Burnside rings. We introduce the notion of intersection graph of a G-set (somewhat generalizing the
ordinary definition of intersection graph of a group) and establish a general upper bound for the domination number of Γ(G) in terms of subgroups
satisfying a certain property in Burnside ring. Intersection graph of G is
the 1-skeleton of the simplicial complex whose faces are the sets of proper
subgroups which intersect non-trivially. We call this simplicial complex intersection complex of G and show that it shares the same homotopy type
with the order complex of proper non-trivial subgroups of G. We also prove
that if domination number of Γ(G) is 1, then intersection complex of G is
contractible.
2010 Mathematics Subject Classification. Primary: 20D99; Secondary: 05C69,
05C25, 20C05, 55U10.
Keywords: Finite groups; subgroup; intersection graph; dominating sets;
domination number; Burnside ring; order complex
1
Introduction
Let F be the set of proper subobjects of an object with an algebraic structure. In
[Yar13] the intersection graph of F is defined in the following way: there is a vertex
for each subobject in F other than the zero object, where the zero object is the object
∗
This work is supported by the TÜBİTAK 2214/A Grant Program: 1059B141401085.
1
having a unique endomorphism, and there is an edge between two vertices whenever
the intersection of the subobjects representing the vertices is not the zero object. In
particular, if F is the set of proper subgroups of a group G, then the zero object is the
trivial subgroup. The intersection graph of (the proper subgroups of) G will be denoted
by Γ(G).
Intersection graphs first defined for semigroups by Bosák in [Bos64]. Let S be a
semigroup. The intersection graph of the semigroup S is defined in the following way:
the vertex set is the set of proper subsemigroups of S and there is an edge between two
distinct vertices A and B if and only if A ∩ B 6= ∅. Interestingly, this definition is not in
the scope of the abstract generalization given in the preceding paragraph. Afterwards,
in [CP69] Csákány and Pollák adapted this definition into groups in the usual way.
Still there are analogous definitions. For example, in [CGMS09] authors studied the
intersection graphs of ideals of a ring. In particular, they determine the values of n
for which the intersection graph of the ideals of Zn is connected, complete, bipartite,
planar or has a cycle. For the corresponding literature the reader may also refer to
[JJR10, JJR11, LQ10, She10, Zel75] and some of the references therein.
As is well-known subgroups of a group G form a lattice L(G) ordered by set inclusion. Some of the structural properties of a group may be inferred by studying its
subgroup structure and those parts of the group theory form a part of the lattice theory. Intersection graphs of groups are natural objects and are intimately related with
subgroup lattices. Actually, given the subgroup lattice one can recover the intersection
graph but not vice versa in general. Intuitively, by passing from L(G) to Γ(G) a certain
amount of knowledge should be lost. In [KY15a] authors show that finite abelian groups
can almost be distinguished by their intersection graphs. The same result was proven
previously for subgroup lattices in [Bae39] (see also [Sch94, 1.2.8 Corollary]). Therefore,
rather surprisingly subgroup lattices and intersection graphs holds the same amount of
information on the subgroup structure if the group is abelian.
By defining intersection graphs we attach a graph to a group, like in the case of
Cayley graphs. So, there are two natural directions we may follow. First, we may study
the graph theoretical properties of intersection graphs by means of group theoretical
arguments. This is straightforward. For example we may ask for which groups their
intersection graphs are planar [KY15b, AT16] or connected [Luc03, She10, Kay17]. And
second, we may study the algebraic properties of groups by means of combinatorial
arguments applied to the intersection graphs though this part seems to require more
ingenuity.
In this paper, we study the dominating sets in intersection graphs. A dominating set
D of a graph Γ is a subset of the vertex set V such that any vertex not in D is adjacent
to some vertex in D. The domination number γ(Γ) of Γ is the smallest cardinal of the
dominating sets for Γ. Vizing’s conjecture from 1968 asserts that for any two graphs
Γ and Γ0 the product γ(Γ)γ(Γ0 ) is at most the domination number of the Cartesian
product of Γ and Γ0 . Despite the efforts of many mathematicians this conjecture is still
open (see [BDG+ 12]). Given a graph Γ and an integer n the dominating set problem
asks whether is there a dominating vertex set of size at most n. It is a classical instance
of a N P -complete decision problem. There are many papers on domination theory
covering algorithmic aspects as well. More can be found on this subject for example in
[BKR07, FGK05, HSH97] and references therein.
It is easy to observe that a subset D of the vertex set V (G) of the intersection graph
of the group G is a dominating set if and only if for any minimal subgroup A of G there
exists a H ∈ D such that A ≤ H. This in turn implies that D is a dominating set if and
only if the union of the subgroups in D contains all minimal subgroups of G as a subset,
2
or equivalently, if and only if the union of the subgroups in D contains all elements of
G of prime order. In particular, the set of minimal subgroups and the set of maximal
subgroups are dominating sets. We denote the domination number of Γ(G) simply by
γ(G) and call this invariant of the group the domination number of G. Let G be a finite
group. We shall note that a dominating set D of minimal size might be assumed to
consist of maximal subgroups as any proper subgroup of a finite group is contained in a
maximal subgroup. In particular, there exist a dominating set D such that each element
of D is a maximal subgroup and the cardinality of D is γ(G).
In [Coh94], Cohn defined a group as an n-sum group if it can be written as the union
of n of its proper subgroups and of no smaller number. Let G be an n-sum group. In the
light of the previous paragraphs it is clear that γ(G) ≤ n. Notice that any non-trivial
finite group can be written as the union of its proper subgroups unless it is not cyclic,
hence it is reasonable to call the cyclic group Cm of order m an ℵ0 -sum group. On
the other hand a cyclic group of order ps with p a prime contains a unique minimal
subgroup, therefore γ(Cps ) = 1 provided s > 1. The intersection graph Γ(G) is the
empty graph whenever G is trivial or isomorphic to a cyclic group of prime order and
in such case we adopt the convention γ(G) = ℵ0 . The reason for this will be justified
when we prove Lemma 2.4. Let G be an n-sum group. It can be easily observed that
γ(G) = n if G is isomorphic to the one of the following groups
Cp ,
Cp × Cp ,
Cp × Cq ,
Cp o Cq ,
where p and q are some distinct prime numbers. The reader may refer to for example
[Coh94, DL08, GL15] for the literature on n-sum groups.
Classifying groups by their domination number is a difficult problem. Even the
determination of groups with domination number 1 seems to be intractable. In Sections 3, 4, and 5 we determine upper bounds for the domination number of particular
classes of groups. For example, abelian groups can be classified by their domination
number (see Theorem 3.1) and the domination number of a supersolvable group is at
most p + 1 for some prime divisor p of its order (see Proposition 4.4). It turns out
symmetric groups forms an interesting class in our context. In Section 5 we find some
upper bounds for the symmetric groups by their degree (see Theorem 5.3) and show
that those bounds are applicable also for the primitive subgroups containing an odd
permutation (see Corollary 5.4).
In Section 6 we introduce intersection graphs of G-sets. This notion in a sense
generalize the ordinary definition of intersection graphs of groups (see Proposition 6.1).
Subgroup intersection is related with the multiplication operator of Burnside rings and
the ultimate aim in this section is to incorporate the Burnside ring context into our
discussion. We show that the domination number γ(G) can be bounded by the sum of
the indices of the normalizers of some subgroups in G satisfying a certain property as a
collection in the Burnside ring (see Proposition 6.2).
There is an extensive literature on combinatorial objects associated with algebraic
structures. An alternative path in this direction is to introduce order complexes of
subgroups and thereby rendering the use of topological terms possible (see, for example [Bro75, Qui78, HIÖ89, Smi11]). Similar work using subgroup lattices, frames,
coset posets, and quandles has also appeared in literature (see, [SW12, SW16, Fum09,
HSW15]).
A natural construction in which a simplicial complex K(G) is associated with a
group G is the following: the underlying set of K(G) is the vertex set of Γ(G) and for
each vertex H in Γ(G) there is an associated simplex σH in K(G) which is defined as the
set of proper subgroups of G containing H. Clearly, the common face of σH and σK is
3
σhH,Ki . Alternatively, K(G) is the simplicial complex whose faces are the sets of proper
subgroups of G which intersect non-trivially. Observe that Γ(G) is the 1-skeleton of
K(G). By an argument due to Volkmar Welker, K(G) shares the same homotopy type
with the order complex of proper non-trivial subgroups of G (see Proposition 7.1). In
Section 7 we study the intersection complex K(G) and prove that if domination number
of Γ(G) is 1, then intersection complex of G is contractible (see Corollary 7.4).
2
Preliminaries
First we recall some of the basic facts from standard group theory.
Remarks.
1. (Product Formula, see [Rot95, Theorem 2.20]) |XY ||X ∩ Y | = |X||Y | for any two
subgroups X and Y of a finite group.
2. (Sylow’s Theorem, see [Rot95, Theorem 4.12])
(i) If P is a Sylow p-subgroup of a finite group G, then all Sylow p-subgroups
of G are conjugate to P .
(ii) If there are r Sylow p-subgroups, then r is a divisor of |G| and r ≡ 1 (mod p).
3. (Hall’s Theorem, see [Gor80, Theorem 4.1]) If G is a finite solvable group, then
any π-subgroup is contained in a Hall π-subgroup. Moreover, any two Hall πsubgroups are conjugate.
4. (Correspondence Theorem, see [Rot95, Theorem 2.28]) Let N E G and let ν : G →
G/N be the canonical morphism. Then S 7→ ν(S) = S/N is a bijection from the
family of all those subgroups S of G which contain N to the family of all the
subgroups of G/N .
5. (Dedekind’s Lemma, see [Rot95, Exercise 2.49]) Let H, K, and L be subgroups of
G with H ≤ L. Then HK ∩ L = H(K ∩ L).
Let G be a finite group. We denote by NG the subgroup of G generated by its
∼ Cp with p a
minimal subgroups. Obviously, NG is a characteristic subgroup. If G =
prime, then we take NG = G. Adapting the module theoretical parlance we might call
a subgroup of a group essential provided that it contains all minimal subgroups. Thus
NG is the smallest essential subgroup. Notice that if G is abelian and G Cp , then NG
is the socle of G.
Lemma 2.1. For a finite group G, the following statements are equivalent.
(i) The domination number of Γ(G) is 1.
(ii) NG is a proper (normal) subgroup of G.
(iii) G is a (non-split) extension of NG by a non-trivial group H.
Proof. (i) =⇒ (ii): There must be a proper subgroup H of G such that H ∩ K 6= 1 for
any non-trivial subgroup K of G. In particular, H intersects non-trivially, and hence
contains, any minimal subgroup in NG . That is, H ≥ NG . However, H is a proper
subgroup of G, so is NG .
4
(ii) =⇒ (iii): Since NG is a proper normal subgroup, G is an extension of NG by a nontrivial group U . Notice that this extension cannot split, as otherwise, there would be a
subgroup of G isomorphic to H which intersects NG trivially. However, this contradicts
with the definition of NG .
(iii) =⇒ (i): Clearly NG is a proper subgroup intersecting any subgroup non-trivially;
hence {NG } is a dominating set for Γ(G).
Corollary 2.2. If G is a finite simple group, then γ(G) > 1.
In general, there is no relation between the domination number of a group and its
subgroups. As a simple example consider the dihedral group D8 = ha, b a4 = b2 =
1, bab = a3 i of order 8. It has three maximal subgroups ha2 , bi, ha2 , abi, hai and the
first two of them dominates Γ(D8 ). Moreover, as D8 = hab, bi, we have γ(D8 ) = 2 by
Lemma 2.1. On the other hand, ha2 , bi ∼
= C2 ×C2 and since its intersection graph consists
of isolated vertices, we have γ(ha2 , bi) = 3; whereas γ(hai) = 1 as it is isomorphic to the
cyclic group of order four. However, by imposing some conditions on the subgroup H
of G, it is easy to prove that γ(H) ≤ γ(G) holds.
Lemma 2.3. Let G be a group and D be a dominating set of Γ(G). Then γ(H) ≤ |D|
provided that none of the elements of D contains H. In particular, γ(H) ≤ γ(G) if there
is such a dominating set with cardinality γ(G).
Proof. Observe that if none of the elements of D contains H, then the set DH :=
{D ∩ H : D ∈ D} is a dominating set for Γ(H).
The following result will be very useful in our later arguments.
Lemma 2.4. Let N be a normal subgroup of G. If G/N Cp , then γ(G) ≤ γ(G/N ).
Moreover, the condition G/N Cp can be removed if G is a finite group.
Proof. Let D be a dominating set for Γ(G/N ) and set D := {N < W < G : W /N ∈ D}.
By the Correspondence Theorem, |D| = |D|. We want to show that D is a dominating
set for Γ(G). Let H be a proper non-trivial subgroup of G. If H ∩ N 6= 1, then any
element of D intersects H non-trivially. Suppose that H ∩N = 1. Take W ∈ D such that
(N H/N ) ∩ W := Y 6= 1. If Y = N H/N , then clearly W contains H where W ∈ D such
that W /N = W . Suppose Y 6= N H/N and let Y /N = Y . Obviously Y < W . Moreover,
By the Dedekind’s Lemma Y = N K, where K := Y ∩ H. That is W ∩ H ≥ K 6= 1.
Since H is an arbitrary subgroup, we see that D dominates Γ(G). This proves the first
part. The second part follows from the convention γ(Cp ) = ℵ0 .
Corollary 2.5. Let G ∼
= H × K be a finite group. Then γ(G) ≤ min{γ(H), γ(K)}.
Let S be a subset of the vertex set V (G) of Γ(G). It is natural to define the
intersection graph Γ(S) of the vertex set S in the following way. There is an edge between
two vertices H, K ∈ S if and only if H ∩ K ∈ S. We denote the domination number
of Γ(S) simply by γ(S). Let V (G)>N be the set of proper subgroups of G containing
∼
the normal subgroup N of G strictly. By the Correspondence Theorem Γ(V (G)>N ) =
Γ(G/N ), hence γ(V (G)>N ) = γ(G/N ). Moreover, by the proof of Lemma 2.4, a (nonempty) subset of V (G)>N dominates Γ(G) provided that it dominates Γ(V (G)>N ).
Let ϕN : V (G) ∪ {1, G} → V (G)>N ∪ {N, G} be the map taking H to N H; and let
D be a dominating set of Γ(G) consisting of maximal subgroups. Notice that even if
ϕN (D) ⊆ V (G)>N , the image set ϕN (D) might not be a dominating set for Γ(V (G)>N ).
5
Let Sp (G) be the set of all proper non-trivial p-subgroups of G. Observe that if
G is a p-group, Γ(G) and Γ(Sp (G)) coincides. It is also true in general that there is
an edge between two vertices of Γ(Sp (G)) if and only if there is an edge between the
corresponding vertices in Γ(G).
Lemma 2.6. Let G be a finite group and let N be a normal p-subgroup of G such that
the set Sp (G)>N is non-empty. Then
γ(Sp (G)) ≤ γ(Sp (G)>N ).
Proof. Let D be a dominating set of Sp (G)>N . As in the proof of Lemma 2.4, we want
to show that D dominates Γ(Sp (G)). Let H ∈ Sp (G). If H ∩ N 6= 1, then clearly
H ∩ W 6= 1 for any W ∈ D. Suppose H ∩ N = 1. Then N H ∈ Sp (G)>N . Let Y
be a minimal subgroup in Sp (G)>N contained by N H; and let W ∈ D contains Y . If
Y = N H, then H < W . Suppose Y 6= N H. Since N < Y < N H and N ∩ H = 1,
Y ∩ H ∈ Sp (G) by the Product Formula; hence, W ∩ H ∈ Sp (G) as well.
3
Abelian groups
In this section we classify finite abelian groups by their domination number. Recall
that the exponent of a group G, denoted by exp(G), is the least common multiple of
the orders of elements of G. Let G be a finite group and consider the function f from
the set of non-empty subsets of G to the set of positive integers taking X ⊆ G to the
lowest common multiple of the orders of elements of X. Clearly, the image of the whole
group G is exp(G). By a celebrated theorem of Frobenius if X is a maximal subset of
G satisfying the condition xk = 1 for all x ∈ X with k is a fixed integer dividing |G|,
then k divides |X|. Let g be the function taking the integer k to the maximal subset
Xk := {x ∈ G : xk = 1}. Then, f and g define a Galois connection between the poset of
non-empty subsets of G and the poset of positive integers ordered by divisibility relation.
In general such a maximal subset may not be a subgroup. For example, if the Sylow
p-subgroup P of G is not a normal subgroup of G, then the union of conjugates of P
cannot be a subgroup. However, if G is an abelian group then for any integer k the
subset Xk is actually a subgroup.
For a finite group G we denote by sfp(G) the square-free part of |G|, i.e. sfp(G) is
the product of distinct primes dividing |G|. Notice that a collection of proper subgroups
dominates to the intersection graph if and only if their union contains Xt , where t =
sfp(G).
Theorem 3.1. Let G be a finite abelian group with proper non-trivial subgroups. Then
1. γ(G) = 1 if and only if sfp(G) < exp(G).
2. γ(G) = 2 if and only if sfp(G) = exp(G) and sfp(G) is not a prime number.
3. γ(G) = p + 1 if and only if p = sfp(G) = exp(G) is a prime number.
Proof. Let t be the square-free part of |G| and m be the exponent of G.
Assertion 1. Observe that NG = {x ∈ G : xt = 1}. By Lemma 2.1, γ(G) = 1 if and
only if NG is a proper subgroup which is the case if and only if t < m.
For Assertion 2 and Assertion 3 it is enough to prove the sufficiency conditions, as
2 6= t + 1 for any prime number t.
6
Assertion 2. Suppose that t = m and t is not a prime number. Then there exist two
distinct primes p and q dividing t. Clearly, the subgroups H = {h ∈ G : ht/p = 1} and
K = {k ∈ G : k t/p = 1} dominates Γ(G). As there is no dominating set of cardinality
one (by virtue of Assertion 1), γ(G) = 2.
Assertion 3. Suppose that t = m and t is a prime number. We consider G as a
vector space over the field Ft of t elements of dimension d ≥ 2 and fix a basis for this
vector space in canonical way. Let Hi = hhi i, 1 ≤ i ≤ t + 1 where hi = (1, 0, . . . , 0, i)
for i ≤ t and ht+1 = (0, . . . , 0, 1). Then {Ki := Hi⊥ : 1 ≤ i ≤ t + 1} is a dominating
set for Γ(G). To see this first observe that for any g ∈ G there exists an hi such that
g ·hi = 0, i.e. g ∈ Ki for some 1 ≤ i ≤ t+1. Next, suppose that there exist a dominating
set D = {M1 , . . . , Ms } with cardinality s < t + 1. We want to derive a contradiction.
Without loss of generality elements of D can be taken as maximal subgroups. Let
Aj = Mj⊥ , 1 ≤ j ≤ s. Suppose Aj are generated by linearly independent vectors (if not,
we may take a maximal subset of {A1 , . . . , As } with this property and apply the same
arguments). By a change of basis if necessary, we may take Aj = Hj . However, that
means g = (1, 0, . . . , 0) is not contained in any Mi ∈ D as g · hj 6= 0 for 1 ≤ j ≤ s. This
contradiction completes the proof.
Remark. We may prove the first and second assertions by regarding G as a Z-module
(compare with [Yar13, Theorem 4.4]). Also for the third assertion we may argue as
follows. By Lemma 2.4, γ(G) ≤ t + 1 as γ(Cp × Cp ) = t + 1 and the rank of G, say
r, is greater than or equal to two. On the other hand, any non-identity element of G
belongs to exactly one minimal subgroup and there are tr − 1 of them. Any maximal
subgroup of G contains tr−1 − 1 non-identity elements and t maximal subgroups may
cover at most tr − t elements; hence, there is no dominating set for G of size < t + 1.
By the fundamental theorem of finite abelian groups any finite abelian group can
be written as the direct product of cyclic groups of prime power orders, thus we may
restate Theorem 3.1 as
γ(Cpα1 × · · · × Cpαk ) =
1
1
1
2
p+1
if αi ≥ 2 for some 1 ≤ i ≤ k
if αi = 1 for all 1 ≤ i ≤ k
and pj1 6= pj2 for some j1 6= j2
if αi = 1 and pi = p for all 1 ≤ i ≤ k
where p1 , . . . , pk are prime numbers and α1 , . . . , αk are positive integers with k = 1
implies α1 > 1.
4
Solvable groups
Though finite abelian groups can be classified by their domination number, it seems this
is not possible in general. Nevertheless, we may use the structural results to find upper
bounds for the domination number of groups belonging to larger families.
Proposition 4.1. Let G be a finite nilpotent group and p be a prime number. Suppose
G Cp . Then
1. γ(G) ≤ p + 1 if G is a p-group.
2. γ(G) ≤ 2 if G is not a p-group.
7
Proof. Assertion 1. There is a normal subgroup N of G of index p2 which is isomorphic
to either Cp2 or Cp × Cp . The assertion follows from Lemma 2.4.
Assertion 2. Let G be the internal direct product of P , Q and N , where P and Q
are the non-trivial Sylow p- and Sylow q- subgroups of G. Clearly, N P and N Q form a
dominating set of Γ(G).
Let G be a finite group. We denote by RG the intersection of the subgroups in the
lower central series of G. This subgroup is the smallest subgroup of G in which the
quotient group G/RG is nilpotent. Obviously, the nilpotent residual RG is a proper
subgroup whenever G is a solvable group.
Corollary 4.2. Let G be a finite group such that G/RG has proper non-trivial subgroups.
Then
1. γ(G) ≤ p + 1 if G/RG is a p-group.
2. γ(G) ≤ 2 if G/RG is not a p-group.
Notice that the condition given in Corollary 4.2 is not restrictive for solvable groups.
One important family of groups that is not comprised is dihedral groups. Let D2n
0 ∼ C and so the quotient
denotes the dihedral group of order 2n. Then RD2n = D2n
= n
D2n /RD2n has no proper non-trivial subgroups. Nevertheless, the structure of those
groups are fairly specific allowing us to determine exact formulas for their domination
number depending on the order 2n.
Lemma 4.3. Let p be the smallest prime dividing n. Then
(
γ(D2n ) =
p,
p + 1,
if p2 n
otherwise
Proof. Let D2n := ha, b an = b2 = 1, bab = a−1 i. Elements of D2n can be listed as
{1, a, a2 , . . . , an−1 , b, ab, a2 b, . . . , an−1 b}.
It can be easily observed that the latter half of those elements are all of order two;
and hence, the minimal subgroups of D2n consists of subgroups of hai that are of prime
order and subgroups haj bi, where j ∈ {0, 1, . . . , n − 1}. And the maximal subgroups of
D2n consists of hai and subgroups of the form hat , ar bi, where t n is a prime number.
Fix a prime t n. Observe that any element of the form aj b is contained exactly
one of the maximal subgroups Tr := hat , ar bi with r ∈ {0, 1, . . . , t − 1}. Let p be the
smallest prime dividing n. Obviously, the union of the subgroups Pr := hap , ar bi with
r ∈ {0, 1, . . . , p−1} contains all minimal subgroups of the form haj bi, j ∈ {0, 1, . . . , n−1},
and there are no possible way to cover them with fewer than p subgroups. Finally, if
p2 n, those subgroups contains all minimal subgroups of hai; and if p2 - n, we must
take Pr , r ∈ {0, 1, . . . , p − 1}, together with a subgroup containing han/p i to form a
dominating set with least cardinality.
Remark. It is difficult to find an n-sum group G such that γ(G) = n and sfp(G) <
exp(G). One example satisfying those conditions is the dihedral group of order 36. By
Lemma 4.3, γ(D36 ) = 3. Moreover, any cyclic subgroup of D36 is contained by those
three subgroups that are of index two; and hence, D36 is a 3-sum group.
8
Consider the normal series of subgroups
G = G0 ≥ G1 ≥ · · · ≥ Gk = 1.
0 /G2
By the third isomorphism theorem and Lemma 2.4, γ( G
G1 /G2 ) = γ(G0 /G1 ) ≥ γ(G0 /G2 ).
And, by a repeated application, we have
γ(G) ≤ γ(G/Gk−1 ) ≤ · · · ≤ γ(G/G1 ).
Proposition 4.4. Let G be a finite supersolvable group with proper non-trivial subgroups. Then γ(G) ≤ p + 1 for some prime divisor p of |G|.
Proof. Since G is a finite supersolvable group, it has a normal subgroup N of index m
where m is either a prime square or a product of two distinct primes. In the first case,
√
γ(G/N ) is at most p + 1 where p = m, and in the latter case γ(G/N ) is at most 2.
The proof follows from Lemma 2.4.
At this stage it is tempting to conjecture that the assertion of Proposition 4.4 holds
more generally for solvable groups. However, we may construct a counterexample in the
following way. Let G = N H be a Frobenius group with complement H such that the
kernel N is a minimal normal subgroup of G. Further, suppose that H ∼
= Cq for some
prime q. Notice that since Frobenius kernels are nilpotent, G must be a solvable group.
On the other hand, since N is a minimal normal subgroup, N has no characteristic
subgroup and in particular N ∼
= Cp × · · · × Cp for some prime p. Suppose that the rank
r of N is greater than 1. Now, since NG (H) = H and N is a minimal normal subgroup,
each conjugate of H is a maximal subgroup. That means there are totally |G : H| = pr
isolated vertices in Γ(G) which, in turn, implies that γ(G) = pr + 1. As pr = 1 + kq for
some integer k by Sylow’s Theorem, the domination number γ(G) is greater than both
p + 1 and q + 1.
Proposition 4.5. Let G be a finite solvable group and H, K be a pair of maximal
subgroups such that (|G : H|, |G : K|) = 1. Then, γ(G) ≤ |G : NG (H)| + |G : NG (K)|.
Notice that Hall’s Theorem guarantees the existence of such a pair of maximal
subgroups provided G is not a p-group. Notice also that one of the summands in the
stated inequality can always be taken 1.
Proof. Take a minimal subgroup A. If |A| |G : H|, then by Hall’s Theorem a conjugate
of K contains A. And if |A| - |G : H|, then clearly a conjugate of H contains A.
5
Permutation groups
We denote by SX the group of the permutations of the elements of X. If X =
{1, 2, . . . , n}, we simply write Sn ; and An will be the group of even permutations on
∼ C2 .
n letters. Notice that Sn /RSn = Sn /An =
Lemma 5.1. γ(Sn ) 6= 1 and γ(An ) 6= 1.
Proof. As S2 ∼
= C2 , we see that γ(S2 ) 6= 1. For n ≥ 3, since any transposition generates
a minimal subgroup of the symmetric group and since adjacent transpositions generate
the whole group, γ(Sn ) 6= 1 by Lemma 2.1. Similarly, γ(A3 ) 6= 1; and for n ≥ 4, since
3-cycles generate An , γ(An ) 6= 1 by the same lemma.
9
Let G be a permutation group acting faithfully on X = {1, 2, . . . , n}. An element
g ∈ G is called homogeneous if the associated permutation has cycle type is (pk , 1n−pk )
with p a prime. Since any minimal subgroup of G is a cyclic group of prime order, in
order for D ⊆ V (G) to be a dominating set of Γ(G), the union of the elements of D
must contain every homogeneous element of G and vice versa. The following standard
result will be used in the proof of Theorem 5.3.
Lemma 5.2 (see [DM96, Theorem 3.3A]). If G is a primitive subgroup of Sn containing
a transposition, then G = Sn .
Theorem 5.3. Let ϑ : N → N be the function given by
ϑ(n) =
n,
n + 1,
if
if
n+
1
,
if
n
if
2 ,
n
+
1
,
if
2
n = 2k + 1 and n is not a prime
n = 2k + 1 and n is a prime
n = 2k and k is even
n = 2k, k is odd and n − 1 is not a prime
n = 2k, k is odd and n − 1 is a prime
(∗)
(∗∗)
Then γ(Sn ) ≤ ϑ(n).
Proof. Let D be a dominating set for Γ(Sn ) of minimal size. Without loss of generality
we may assume that the elements of D are maximal subgroups of Sn . We know that every
homogeneous element of Sn must belong to some subgroup in D, in particular, every
involution (elements of order 2) must be covered by D. By Lemma 5.2, proper primitive
subgroups do not contain transpositions. Therefore, to cover transpositions the only
candidates are imprimitive maximal subgroups. The imprimitive maximal subgroups of
the symmetric group fall into two classes: (i) the intransitive maximal subgroups which
are isomorphic to Sk × Sn−k with k > 1 and k 6= n − k; or (ii) transitive imprimitive
maximal groups which are isomorphic to Sk o Sm with k, m > 1 and n = km. For
the classification of maximal subgroups of the symmetric group the reader may refer to
[Wil09, Chapter 2.5] or [DM96, Chapter 8.5].
Observe that if some collection of subgroups covers all involutions of type (2l , 1n−2l ),
with l is odd, but not all homogeneous elements, then by adding An to this collection
we would get a dominating set for Γ(Sn ), for An contains all cycles of prime length > 2.
Moreover, if the given set has the minimum cardinality, say δ, then since An does not
contain involutions of type (2l , 1n−2l ), l is odd, δ + 1 would be the domination number
of Γ(Sn ). For the rest, we try to form a collection of (as few as possible) maximal
imprimitive subgroups of Sn containing all involutions of type (2l , 1n−2l ) with l is an
odd number.
For n = 2k + 1, any involution must fix a point, hence must belong to some point
stabilizer. Moreover, those point stabilizers contain all homogeneous elements, unless n
is a prime number. In that case, γ(Sn ) ≤ n. If n is a prime number, then together with
An we have a dominating set of cardinality n + 1.
For n = 2k, k is even, point stabilizers contain all involutions of type (2l , 1n−2l ),
with l is odd; hence, together with An we have a dominating set of cardinality n + 1.
That is, γ(Sn ) ≤ n + 1. If k is odd, involutions of type (2k ) are contained neither in any
point stabilizer nor in An . Consider the maximal imprimitive subgroup S{1,2} × S{3,...,n} .
This subgroup and conjugates
of it contains all homogeneous
element, unless n − 1 is a
prime. Since there are n2 conjugates of it, γ(Sn ) ≤ n2 in that case. If n −1 is a prime
number, then together with An we have a dominating set of cardinality n2 + 1.
Remark. For S6 , Theorem 5.3 gives the upper bound 16. Since the exceptional outer
automorphism of S6 maps transpositions to triple transpositions, we may construct a
10
dominating set consisting of 7 elements. Namely, An together with the three point
stabilizers (which cover all transpositions) and the (non-inner) automorphic images of
them form a dominating set for Γ(G). A final note, those 3 subgroups containing all
triple transpositions are maximal primitive 2-transitive subgroups of S6 and they are
isomorphic to S5 (see [Hup67, Satz 5.5]).
Corollary 5.4. Let G be a primitive subgroup of Sn containing an odd permutation,
and let ϑ be the function defined in Theorem 5.3. Then, the inequality γ(G) ≤ ϑ(n)
holds. Moreover, the inequality is valid in cases (∗) and (∗∗) for all primitive subgroups.
Proof. Since G is a primitive subgroup, it is not contained by any imprimitive subgroup.
The proof follows from the fact that the only primitive subgroup used to form dominating
sets in the proof of Theorem 5.3 is An , and from Lemma 2.3.
6
Intersection graphs of G-sets
As is well-known any transitive G-set Ω is equivalent to a (left) coset space G/Gx , where
Gx is the stabilizer of a point x ∈ Ω. So, for example if Ω is a regular G-set, then it can
be represented with G/1; and if Ω is the trivial G-set, it is represented with G/G. Since
any G-set is the disjoint union of transitive G-sets, given a G-set Ω it can be represented
as the sum of coset spaces. For a subgroup H of G we denote by (H) the conjugacy
class of H in G, and by [G/H] the isomorphism class of the transitive G-set G/H. It is
well-known that [G/H] = [G/K] if and only if H = gK for some g ∈ G.
The Burnside ring B(G) of G is the ring generated by the isomorphism classes of
G-sets, where addition is the disjoint union and product is the Cartesian product of
G-sets. Therefore a typical element of B(G) is of the form
X
aj [G/Hj ]
j
where aj are integers and Hj are the representatives of the conjugacy classes of subgroups
of G. Let A and B be normal subgroups of a group G. Then the canonical map
gA ∩ B 7→ (gA, gB) from G/A ∩ B to G/A × G/B is an injective group homomorphism.
Now, let us consider two arbitrary subgroups (not necessarily normal) H1 , H2 of G. In
this case, G/H1 and G/H2 are still G-sets and the diagonal action of G on the Cartesian
product G/H1 × G/H2 yields the map
φ : G/H1 ∩ H2 −→ G/H1 × G/H2
gH1 ∩ H2 7−→ (gH1 , gH2 )
which is an injective G-equivariant map. That is to say, subgroup intersection is related
with the multiplication operator of the Burnside ring.
Let Ω1 and Ω2 be two G-sets and let R(Ω1 ,Ω2 ) be a set of representative elements
for the orbits of Ω1 × Ω2 . The Cartesian product Ω1 × Ω2 decomposes into the disjoint
union of transitive G-sets in a non-trivial fashion. More precisely,
Ω1 × Ω2 ∼
=
G
G/Gx ∩ Gy
(x,y)∈R(Ω1 ,Ω2 )
where Gx and Gy are the stabilizers of the points x ∈ Ω1 and y ∈ Ω2 respectively. Let
R(H,K) be a set of representatives for the equivalence classes of (H, K)-double cosets.
11
Setting Ω1 = G/H, Ω2 = G/K and using sigma notation for disjoint union, we may
write
X
[G/H][G/K] =
[G/H ∩ gK]
g∈R(H,K)
as the orbits of G/H × G/K are parametrized by the (H, K)-double cosets. For more
information the reader may refer to [tD79, Chapter I].
Let G be a fixed finite group. The following definition is suggested to me by Ergün
Yaraneri. The intersection graph Γ[Ω] of a G-set Ω is the simple graph with vertex
set the proper non-trivial stabilizers of points in Ω and there is an edge between two
distinct stabilizers if and only if their intersection is non-trivial. We used in this notation
“brackets” instead of “parentheses” to emphasize that the argument is a G-set. Observe
that since G is a finite group, there are at most finitely many intersection graphs Γ[Ω]
that can be associated with G-sets. Notice that both Γ[G/1] and Γ[G/G] are empty
graphs by definition.
Example. Take Ω := {1, 2, . . . , n} as the vertex set of a regular n-gon on plane. Since the
automorphism group of this polygon is isomorphic to the dihedral group D2n , considered
as a D2n -set a stabilizer of a point of Ω is corresponding to a unique involution of D2n .
Those involutions form a single conjugacy class if n is odd and form two conjugacy
classes otherwise. In both cases Γ[Ω] consists of n isolated vertices. Let G be a finite
group and H be a subgroup of G. As a general fact Γ[G/H] consists of |G : H| isolated
vertices if and only if G is a Frobenius group with complement H.
We denote by C(G) the set of conjugacy classes of subgroups of G. By the following
Proposition, intersection graphs of groups can be seen as particular cases of intersection
graphs of G-sets.
Proposition 6.1. Let G be a finite group and Σ :=
F
(H)∈C(G) [G/H].
Then
Γ[Σ] = Γ(G).
Proof. Since there is an edge between two vertices of Γ[Σ] if and only if their intersection
is non-trivial, it is enough to show that the vertex set of Γ[Σ] is V (G). However, this
is obvious as the set of subgroups of G can be partitioned into the conjugacy classes of
subgroups:
V (G) ∪ {1, G} = { gH : (H) ∈ C(G), g ∈ G}.
Notice that no vertices associated with 1 and G in Γ[Σ] by its definition.
Let G be a finite group. We denote by A(G) the set of minimal subgroups of G and
by M(G) the set of maximal subgroups. The following characterizations are easy to
deduce:
• N / G ⇐⇒ [G/K][G/N ] = |G : N K|[G/N ∩ K] for any K ≤ G.
• A ∈ A(G) ⇐⇒ [G/K][G/A] = |G : K|[G/A] unless K = 1.
• H ∈ M(G) ⇐⇒ [G/K][G/H] does not contain [G/H] as a summand unless
K = G or K ∈ (H).
Let A be a minimal subgroup, and H be a maximal subgroup of G. Observe that
(
[G/H][G/A] =
|G : H|[G/A] if A ≤ H
[G/1]
if A H
in case G is abelian. We may restate Theorem 3.1 in the following way:
12
1. γ(G) = 1 if and only if (∗) there exists an H ∈ M(G) such that |G : H| |[G/A]|
for every A ∈ A(G).
2. γ(G) = 2 if and only if (∗∗) for any H ∈ M(G) there exists an A ∈ A(G) such
that |G : H| - |[G/A]|.
3. γ(G) = p + 1 if and only if neither (∗) nor (∗∗) holds and p = |G : H| = |A|, where
H ∈ M(G) and A ∈ A(G).
Proposition 6.2. Let G be a finite group and let H = {Hi ∈ V (G) : 1 ≤ i ≤ s} be
a set having the property that for any K ∈ V (G) there exists an Hi ∈ H such that
[G/K][G/Hi ] 6= k[G/1] for any positive integer k. Then
γ(G) ≤
s
X
|G : NG (Hi )|.
i=1
Moreover, γ(G) = 1 if and only if there exists H / G such that [G/K][G/H] 6= k[G/1]
for any K ∈ V (G).
Proof. Let H be a set as in the statement of the Proposition. We want to show that
D := { gHi : Hi ∈ H, g ∈ G} is a dominating set for Γ(G). However this is obvious,
since for any K ∈ V (G), [G/K][G/Hi ] 6= k[G/1] implies that K ∩ gHi 6= 1 for some
g ∈ G. This completes the first part of the proof.
Now, suppose that γ(G) = 1. Let L < G such that L ∩ K 6= 1 for any K ∈ V (G).
P
P
−1
Obviously, [G/K][G/L] = [G/K ∩ gL] = [G/ g K ∩ L] does not contain a regular
summand, as L intersects any conjugate of K non-trivially. Take a conjugate aL of L.
P
−1
Since aL also forms a dominating set, [G/K][G/L ∩ aL] = [G/( g K ∩ L) ∩ aL] does
Tg
not contain a regular summand either. Let H :=
L. By a repeated application of
the above argument we see that [G/K][G/H] has no regular summand. However, H is
a normal subgroup which means [G/K][G/H] 6= k[G/1].
Conversely, suppose that there exists an H / G such that [G/K][G/H] 6= k[G/1] for
any K ∈ V (G). Since H is normal, this means [G/K][G/H] has no regular summand
which, in turn, implies that {H} is a dominating set for Γ(G). This completes the
proof.
Remark. Clearly, the elements of the set H in the statement of Proposition 6.2 can be
taken as maximal subgroups. Then, Proposition 4.5 tells us that there exists such a
2-element set if G is a solvable group but not a p-group.
7
Intersection complexes
Recall that a (abstract) simplicial complex is a collection S of finite non-empty sets,
such that if σ is an element of S, so is every non-empty subset of σ. The element σ of
S is called a simplex of S and each non-empty subset of σ is called a face of σ. The
underlying set of S is the union of one-point elements (singletons) of S. The k-skeleton
of S is the subcollection of elements of S having cardinality at most k + 1. For a group
G, we define the intersection complex K(G) of G as the simplicial complex whose faces
are the sets of proper subgroups of G which intersect non-trivially. As a graph the 1skeleton of K(G) is isomorphic to the intersection graph Γ(G). This notion is somewhat
between the two other notions in literature, namely the order complex and the clique
complex. In the first case, we begin with a poset (poset of proper non-trivial subgroups
13
of a group in our case) and construct its order complex by declaring chains of the poset
as the simplices. And, in the latter case, we take a graph (i.e. the intersection graph of
a group) and the underlying set of the corresponding clique complex is the vertex set of
the graph with simplices being the cliques.
Examples.
1. The quaternion group Q8 has three maximal subgroups, say hii, hji, and hki, of
order four intersecting at the unique minimal subgroup {−1, 1}. Thus, Γ(Q8 ) is
a complete graph K4 depicted in Figure 1a. Moreover, K(Q8 ) is a tetrahedron
as those four vertices form a simplex. However, the order complex of the poset
of proper non-trivial subgroups of Q8 is isomorphic to the star graph K1,3 as a
graph. Hence, order complexes and intersection complexes are different in general.
Notice that the clique complex of Γ(Q8 ) is same with K(Q8 ).
2. The intersection graph of the elementary abelian group of order eight is represented in Figure 1b. Here the vertices on the outer circle represents the minimal
subgroups and the vertices on the inner circle are the maximal subgroups. By the
Product Formula any two maximal subgroups intersects at a subgroup of order 2.
Therefore, the vertices in the inner circle form a complete subgraph and those vertices form a simplex in the clique complex whereas they do not in K(C2 ×C2 ×C2 ).
Thus, intersection complexes and clique complexes are not the same in general.
Notice that Γ(C2 × C2 × C2 ) is symmetrical enough to reflect the vector space
structure of the group.
(a) Γ(Q8 )
(b) Γ(C2 × C2 × C2 )
Figure 1: Intersection graphs of some groups of order 8
In Example 1, we remarked that order complexes and intersection complexes are
different in general. However, they are equivalent up to homotopy. The proof of the
following Proposition is due to Volkmar Welker.
Proposition 7.1. For a group G the intersection complex K(G) and the order complex
of the poset of proper non-trivial subgroups of G are homotopy equivalent.
Proof. Consider the face poset of K(G), i.e. the poset of simplices ordered by inclusion.
By the identification H 7→ σH , the poset of proper non-trivial subgroups of G becomes
a subposet (after reversing the order relation) of the face poset of K(G). The order
complex of the face poset of a simplicial complex is the barycentric subdivision of the
simplicial complex and therefore they are homeomorphic. We want to show that the
14
poset of the proper non-trivial subgroups of G and the face poset of K(G) are of the
same homotopy type as order complexes. Let f be the map taking σH to σK , where K
is the intersection of all maximal subgroups containing H. Then f is a closure operator
on the face poset of K(G). Let g be the map taking H to K, where K is the intersection
of all maximal subgroups containing H. Then g is a closure operator on the poset
of proper non-trivial subgroups of G. Since closure operations on posets preserve the
homotopy type of the order complex and since the images of f and g are isomorphic by
the identification K 7→ σK , the proof is completed.
Remark. Let S be a subposet of the poset of proper non-trivial subgroups of G. Recall
that in Section 2 we defined the intersection graph Γ(S) of S in the obvious way. We may
also define the intersection complex K(S) and then the above proof applies verbatim.
In particular, K(Sp (G)) and the order complex of Sp (G) are same up to homotopy.
An important result on the subject for our purposes (see Lemma 7.2 below) uses the
definitions of algebraic topology adapted to ‘poset’ context. By a covering C of a finite
poset P we mean a finite collection {Ci }i∈I of subsets of P, such that P = ∪i∈I Ci . The
nerve N (C) of C is the simplicial complex whose underlying set is I and the simplices
are the J ⊆ I such that CJ := ∩i∈J Ci 6= ∅. The covering C is called contractible if each
CJ is contractible considered as an order complex, where J is a simplex in N (C). We
say C is a downward closed covering if each Ci , i ∈ I, is a closed subset of P, i.e. for
each Ci if the condition
whenever x ∈ Ci and x0 ≤P x, then x0 ∈ Ci
is satisfied. Of course, we may define upward closed coverings dually.
Lemma 7.2 (see [Smi11, Theorem 4.5.2]). If C is a (upward or downward) closed
contractible covering of a poset P, then P is homotopy equivalent to nerve N (C) of C.
Let G be a finite group. For a proper non-trivial subgroup H of G, we denote by
V (G)≥H the set of proper subgroups of G containing H. Similarly, V (G)≤H is the set
non-trivial subgroups of G contained by H. In poset terminology V (G)≥H and V (G)≤H
are called cones and it is a standard fact that they are contractible. Consider the
following collections:
A := {V (G)≥H : H is a minimal subgroup of G}
M := {V (G)≤H : H is a maximal subgroup of G}
Observe that A is an upward closed contractible covering of V (G) considered as a poset
under set theoretical inclusion and M is a downward closed contractible covering. Therefore, by Lemma 7.2 N (A), N (M ), and the order complex of proper non-trivial subgroups
of G share the same homotopy type. Recall that the Frattini subgroup Φ(G) of G is the
intersection of all maximal subgroups of G.
Theorem 7.3. Let G be a finite group. Then
(i) Φ(G) 6= 1 if and only if N (M ) is a simplex.
(ii) γ(G) = 1 if and only if N (A) is a simplex.
Proof. Assertion 1. Clearly, N (M ) is a simplex if and only if for any subset S of
N (M ) intersection of the subgroups in S is non-trivial which is the case if and only if
intersection of all maximal subgroups of G is non-trivial.
Assertion 2. Similar to the previous case N (A) is a simplex if and only if NG is a
proper subgroup of G. The assertion follows from Lemma 2.1.
15
Remark. Using Theorem 3.1, one may conclude that for abelian groups γ(G) = 1 if and
only if Φ(G) 6= 1. Let P be a p-group with p being a prime number. It is also true that
γ(G) = 1 implies Φ(G) 6= 1; for if P is a non-cyclic p-group and γ(P ) = 1, then P/Φ(P )
is elementary abelian of rank > 1 and γ(P/Φ(P )) = p + 1 in that case. However,
the converse statement is not true even for p-groups. For example, Φ(D8 ) ∼
= C2 but
γ(D8 ) = 2.
Since N (A) and K(G) are of the same homotopy type by Proposition 7.1 and
Lemma 7.2, as a consequence of Theorem 7.3 we have the following
Corollary 7.4. Let G be a finite group. If γ(G) = 1, then K(G) is contractible.
References
[AT16]
H. Ahmadi and B. Taeri, Planarity of the intersection graph of subgroups of
a finite group, Journal of Algebra and Its Applications 15 (2016), no. 03,
1650040.
[Bae39]
R. Baer, The Significance of the System of Subgroups for the Structure of the
Group, Amer. J. Math. 61 (1939), no. 1, 1–44.
[BDG+ 12] B. Brešar, P. Dorbec, W. Goddard, B. L. Hartnell, M. A. Henning,
S. Klavžar, and D. F. Rall, Vizing’s conjecture: a survey and recent results,
Journal of Graph Theory 69 (2012), no. 1, 46–76.
[BKR07]
B. Brešar, S. Klavžar, and D. F. Rall, Dominating direct products of graphs,
Discrete Mathematics 307 (2007), no. 13, 1636 – 1642.
[Bos64]
J. Bosák, The graphs of semigroups, Theory of Graphs and its Applications (Proc. Sympos. Smolenice, 1963), Publ. House Czechoslovak Acad.
Sci., Prague, 1964, pp. 119–125.
[Bro75]
K. S. Brown, Euler characteristics of groups: the p-fractional part, Invent.
Math. 29 (1975), no. 1, 1–5.
[CGMS09] I. Chakrabarty, S. Ghosh, T. K. Mukherjee, and M. K. Sen, Intersection
graphs of ideals of rings, Discrete Math. 309 (2009), no. 17, 5381–5392.
[Coh94]
J. H. E. Cohn, On n-sum groups, Math. Scand. 75 (1994), no. 1, 44–58.
[CP69]
B. Csákány and G. Pollák, The graph of subgroups of a finite group,
Czechoslovak Math. J. 19 (94) (1969), 241–247.
[DL08]
E. Detomi and A. Lucchini, On the structure of primitive n-sum groups,
Cubo 10 (2008), no. 3, 195–210.
[DM96]
J. D. Dixon and B. Mortimer, Permutation groups, Graduate Texts in Mathematics, vol. 163, Springer-Verlag, New York, 1996.
[FGK05]
F. V. Fomin, F. Grandoni, and D. Kratsch, Measure and conquer: Domination – a case study, Springer Berlin Heidelberg, 2005.
[Fum09]
F. Fumagalli, A characterization of solvability for finite groups in terms of
their frame, Journal of Algebra 322 (2009), no. 6, 1919 – 1945.
16
[GL15]
M. Garonzi and A. Lucchini, Covers and normal covers of finite groups,
Journal of Algebra 422 (2015), 148 – 165.
[Gor80]
D. Gorenstein, Finite groups, second ed., Chelsea Publishing Co., New York,
1980.
[HIÖ89]
T. Hawkes, I. M. Isaacs, and M. Özaydin, On the Möbius function of a finite
group, Rocky Mountain J. Math. 19 (1989), no. 4, 1003–1034.
[HSH97]
T. W. Haynes, P. J. Slater, and S. T. Hedetniemi, Domination in graphs :
advanced topics, New York : Marcel Dekker, 1997.
[HSW15]
I. Heckenberger, J. Shareshian, and V. Welker, On the lattice of subracks of
the rack of a finite group, ArXiv e-prints (2015).
[Hup67]
B. Huppert, Endliche Gruppen. I, Die Grundlehren der Mathematischen
Wissenschaften, Band 134, Springer-Verlag, Berlin-New York, 1967.
[JJR10]
S. H. Jafari and N. Jafari Rad, Planarity of intersection graphs of ideals of
rings, Int. Electron. J. Algebra 8 (2010), 161–166.
[JJR11]
, Domination in the intersection graphs of rings and modules, Ital. J.
Pure Appl. Math. (2011), no. 28, 19–22.
[Kay17]
S. Kayacan, Connectivity of intersection graphs of finite groups, Communications in Algebra, to appear (2017).
[KY15a]
S. Kayacan and E. Yaraneri, Abelian groups with isomorphic intersection
graphs, Acta Math. Hungar. 146 (2015), no. 1, 107–127.
[KY15b]
Selçuk Kayacan and Ergün Yaraneri, Finite groups whose intersection graphs
are planar, J. Korean Math. Soc. 52 (2015), no. 1, 81–96.
[LQ10]
J. D. Laison and Y. Qing, Subspace intersection graphs, Discrete Math. 310
(2010), no. 23, 3413–3416.
[Luc03]
Maria Silvia Lucido, On the poset of non-trivial proper subgroups of a finite
group, J. Algebra Appl. 2 (2003), no. 2, 165–168.
[Qui78]
D. Quillen, Homotopy properties of the poset of nontrivial p-subgroups of a
group, Adv. in Math. 28 (1978), no. 2, 101–128.
[Rot95]
Joseph J. Rotman, An introduction to the theory of groups, fourth ed., Graduate Texts in Mathematics, vol. 148, Springer-Verlag, New York, 1995.
[Sch94]
R. Schmidt, Subgroup lattices of groups, de Gruyter Expositions in Mathematics, vol. 14, Walter de Gruyter & Co., Berlin, 1994.
[She10]
R. Shen, Intersection graphs of subgroups of finite groups, Czechoslovak
Math. J. 60(135) (2010), no. 4, 945–950.
[Smi11]
S. D. Smith, Subgroup complexes, Mathematical Surveys and Monographs,
vol. 179, American Mathematical Society, Providence, RI, 2011.
[SW12]
J. Shareshian and R. Woodroofe, A new subgroup lattice characterization of
finite solvable groups, Journal of Algebra 351 (2012), no. 1, 448 – 458.
17
[SW16]
, Order complexes of coset posets of finite groups are not contractible,
Advances in Mathematics 291 (2016), 758 – 773.
[tD79]
T. tom Dieck, Transformation groups and representation theory, Lecture
Notes in Mathematics, vol. 766, Springer, Berlin, 1979.
[Wil09]
R. A. Wilson, The finite simple groups, Graduate Texts in Mathematics, vol.
251, Springer-Verlag London, Ltd., London, 2009.
[Yar13]
E. Yaraneri, Intersection graph of a module, J. Algebra Appl. 12 (2013),
no. 5, 1250218, 30.
[Zel75]
B. Zelinka, Intersection graphs of finite abelian groups, Czechoslovak Math.
J. 25(100) (1975), 171–174.
18
| 4math.GR
|
arXiv:1605.02956v1 [math.CO] 10 May 2016
PROJECTIVE DIMENSION OF (HYPER)GRAPHS AND THE
CASTELNUOVO-MUMFORD REGULARITY OF BIPARTITE GRAPHS
TÜRKER BIYIKOĞLU AND YUSUF CIVAN
Abstract. We prove that proj-dimpHq ď regpLpHqq holds for any (hyper)graph H,
where LpHq denotes the Levi graph (the incidence bipartite graph) of H. This in particular brings the use of regularity’s upper bounds on the calculation of proj-dimpHq.
When G is just a (simple) graph, we prove that there exists an induced subgraph H of
G such that proj-dimpGq “ regpSpHqq, where LpHq “ SpHq is the subdivision graph of
H. Moreover, we show that known upper bounds on proj-dimpGq involving domination
parameters are in fact upper bounds to regpSpGqq. By way of application, we prove that
proj-dimpDompGqq “ ΓpGq for any graph G, where DompGq is the dominance complex
of G and ΓpGq is the upper domination number of G. As a counterpart of prime graphs
introduced for the regularity calculations of graphs, we call a connected graph G as a
projectively prime graph over a field k if proj-dimk pG ´ xq ă proj-dimk pGq holds for any
vertex x P V pGq. Such a notion allows us to create examples of graphs showing that
most of the known lower and upper bounds on the projective dimension of graphs are
far from being tight.
1. Introduction
Let R “ krV s be a polynomial ring with a finite set V of indeterminates over a
field k. When I is a monomial ideal in R, there are two central invariants associated
to I, the regularity regpIq :“ maxtj ´ i : βi,j pIq ‰ 0u and the projective dimension
proj-dimpIq :“ maxti : βi,j pIq ‰ 0 for some ju, that in a sense, they measure the complexity of computing the graded Betti numbers βi,j pIq of I. In particular, if I is a squarefree
monomial ideal, these invariants are related by a well-known duality result of Terai [24]
stating that proj-dimpIq “ regpR{I _ q, where I _ is the Alexander dual of I.
The main purpose of the current work is to explore the nature of the Terai’s duality
from the combinatorial point of view. Recall that if I is minimally generated by squarefree monomials m1 , . . . , mr , then the family ∆ :“ tK Ď V : Fi Ę K for any i P rrsu, where
Fi :“ txj : xj |mi u is a simplicial complex on V “ tx1 , . . . , xn u such that whose minimal
non-faces exactly correspond to minimal generators of I. Under such an association, the
ideal I is said to be the Stanley-Reisner ideal of the simplicial complex ∆, denoted by
I “ I∆ , and the simplicial complex ∆I :“ ∆ is called the Stanley-Reisner complex of the
ideal I. Furthermore, the pair HpIq :“ pV, tF1 , . . . , Fr uq is a (simple) hypergraph (known
Date: January 12, 2018.
2010 Mathematics Subject Classification. 13F55, 05E40, 05E45, 05C70, 05C75, 05C76.
Key words and phrases. Edge ring, projective dimension, Castelnuovo-Mumford regularity, dominance complex, induced matching number, Helly number, Levi graph, subdivision graph, vertex-edge
and edge-vertex domination numbers.
Both authors are supported by TÜBİTAK (grant no: 111T704), and the first author is also supported
by ESF EUROCORES TÜBİTAK (grant no: 210T173).
1
2
TÜRKER BIYIKOĞLU AND YUSUF CIVAN
also as a clutter ) on the set V , and under such a correspondence, the ideal I is called the
edge ideal of the hypergraph HpIq.
These interrelations can be reversible in the following way. Let H “ pV, Eq be a hypergraph on V . Then the family of subsets of V containing no edges of H forms a
simplicial complex IndpHq on V , the independence complex of H, and the corresponding
Stanley-Reisner ideal IH :“ IIndpHq is exactly the edge ideal of H. Under such settings,
the regularity of a simplicial complex is defined by regp∆q :“ regpR{I∆ q from which we
set regpHq :“ regpIndpHqq for any (hyper)graph H.
In the topological combinatorial side, the classical combinatorial Alexander duality has
already some fruitful consequences. For instance, Csorba [6] proved that the independence
complex IndpSpGqq of the subdivision graph SpGq of any graph G is homotopy equivalent
to the suspension of the Alexander dual IndpGq_ . However, his result becomes a special
case when we consider the result of Nagel and Reiner [21] on the homotopy type of the
suspension of an arbitrary simplicial complex. In detail, they show that if ∆ is a simplical
complex on V with the set of facets F∆ , and Bp∆q is the bipartite graph that can be
obtained by taking the bipartite complement of the Levi graph of ∆ (or the incidence
graph of the set system pV, F∆ q), then the homotopy equivalence IndpBp∆qq » Σp∆q
holds (see Section 3 for details). Even if the regularity of a simplicial complex is not a
homotopy invariant, we show that it behaves well under such a homotopy equivalence,
that is, we prove that the inequality regp∆q ` 1 ď regpBp∆qq holds for any simplicial
complex ∆ (Theorem 3.5). When we combine such a result with the Terai’s duality, we
conclude that proj-dimpHq ď regpLpHqq for any hypergraph H, where LpHq is the Levi
graph of H. In the case where H “ H is just a (simple) graph, such an inequality becomes
proj-dimpHq ď regpSpHqq that corresponds to the Csorba’s special case in the algebraic
language. In particular, we prove that every graph H has an induced subgraph H 1 such
that proj-dimpHq “ regpSpH 1qq.
Such interrelations naturally bring the question of whether the known lower and upper
bounds on the projective dimension of graphs and the regularity of (bipartite) graphs are
comparable? Firstly, we show that the induced matching number of the bipartite graph
Bp∆q is closely related to the Helly number of the simplicial complex ∆ (Corollary 3.10).
On the other hand, we verify that upper bounds on proj-dimpHq involving domination
parameters of H invented by Dao and Schweig in a series of papers [8, 9, 10] are in fact
upper bounds to regpSpHqq. Interestingly, the induced matching and cochordal cover
numbers of SpHq can be expressed in terms of the domination parameters of H that may
be of independent interest. Indeed, we show that the equalities impSpHqq “ |H| ´ γpHq
and cochordpSpHqq “ |H| ´ τ pHq hold for any graph H without any isolated vertex
(Proposition 4.5 and Corollary 4.11), where γpHq is the domination number of H, whereas
τ pHq is the independence domination number of H.
As a counterpart of prime graphs introduced for the regularity calculations of graphs [4],
we call a connected graph G as a projectively prime graph over a field k if proj-dimk pG ´
xq ă proj-dimk pGq holds for any vertex x P V pGq. Moreover, we say that G is a perfect
projectively prime graph if it is projectively prime graph over any field. In particular,
we verify that the join of any two graphs is a perfect projectively prime graph. In one
hand, such a notion turns the projective dimension calculations of graphs into generalized
(weighted) induced matching problems, on the other hand, it allows us to create examples
PROJECTIVE DIMENSION OF GRAPHS AND THE REGULARITY OF BIPARTITE GRAPHS
3
of graphs showing that most of the known lower and upper bounds on the projective
dimension is far from being tight.
2. Preliminaries
2.1. (Hyper)graphs. By a (simple) graph G, we will mean a finite undirected graph
without loops or multiple edges. If G is a graph, V pGq and EpGq (or simply V and E)
denote its vertex and edge sets. If U Ă V , the graph induced on U is written GrUs, and
in particular, we abbreviate GrV zUs to G ´ U, and write G ´ x whenever U “ txu. For a
given subset U Ď V , the (open) neighbourhood of U is defined by NG pUq :“ YuPU NG puq,
where NG puq :“ tv P V : uv P Eu, and similarly, NG rUs :“ NG pUq Y U is the closed
neighbourhood of U. Furthermore, if F “ te1 , . . . , ek u is a subset of edges of G, we write
NG rF s for the set NG rV pF qs, where V pF q is the set of vertices incident to edges in F .
We write G˝ for the graph obtained from G by removing isolated vertices (if any).
Throughout Kn , Cn and Pn will denote the complete, cycle and path graphs on n
vertices respectively.
We say that G is H-free if no induced subgraph of G is isomorphic to H. A graph G is
called chordal if it is Cr -free for any r ą 3. Moreover, a graph is said to be cochordal if
its complement is a chordal graph.
A hypergraph is a (finite) set system H “ pV, Eq, where V is a set, the set of vertices of
H, and E Ď 2V is the set of edges of H. The hypergraphs that we consider do not have
loops, that is, E can not contain singletons as edges. A hypergraph H “ pV, Eq is said to
be simple if A Ę B and B Ę A for any two edges A, B P E. Through our work, we only
deal with simple hypergraphs.
When S Ď V , the induced subhypergraph of H by the set S is defined to be the hypergraph HrSs :“ tA P E : A Ď Su.
Recall that a subset M Ď E is called a matching of H if no two edges in M share a
common vertex. Moreover, a matching M “ tF1 , . . . , Fk u of H is an induced matching if
the edges in M are exactly the edges of the induced subhypergraph of H over the vertices
contained in Yki“1 Fi , and the cardinality of a maximum induced matching is called the
induced matching number of H and denoted by impHq.
Definition 2.1. For a given hypergraph H “ pV, Eq, the bipartite graph LpHq on V Y E
defined by px, F q P EpLpHqq if and only if x P F is called the Levi graph (or the incidence
graph 1) of H.
If we drop the condition of being simple for hypergraphs, any bipartite graph appears
as the Levi graph of some hypergraph, even if such a representation is not unique. Since
we only deal with simple hypergraphs, the resulting bipartite graphs are of special types.
Let B be a bipartite graph with a bipartition V pBq “ X0 YX1 . Then B is called a Sperner
bipartite graph (or shortly an Sp-bipartite graph) on Xi , if Xi contains no pair of vertices
u and v such that NB puq Ď NB pvq. Observe that a bipartite graph B satisfies B – LpHq
for some (simple) hypergraph H if and only if B is a Sp-bipartite graph.
We also note that when H “ H is just a graph, its Levi graph LpHq is known as the
subdivision graph of H, denoted by SpHq, which can be obtained from H by subdividing
1It
seems that this terminology is not standard, both stemming from incidence geometry.
4
TÜRKER BIYIKOĞLU AND YUSUF CIVAN
each edge of H exactly once. The characterization of subdivision graphs is rather simple,
namely that there exists a graph G such that B – SpGq if and only if B is C4 -free and
degB pvq “ 2 for all v P Xi for some i “ 0, 1.
2.2. Domination parameters of graphs. We next recall the definitions of some domination parameters of graphs that will be in use throughout the sequel. We note that the
relationship between projective dimension of graphs and various domination parameters
firstly explored by Dao and Schweig in a series of papers [8, 9, 10]. However, some of their
newly defined notions turn out to be already known in the literature [1, 17, 22].
A subset A Ď V is called a dominating set for G if NG rAs “ V pGq, and the minimum
size γpGq of a dominating set for G is called the domination number of G, while the
maximum size ΓpGq of a minimal dominating set of G is known as the upper domination
number of G. Moreover, the least cardinality ipGq of an independent dominating set for
G is called the independent domination number of G.
When S Ď V , a vertex x P V zS is said to be a private neighbour of a vertex s P S
with respect to S in G if NG pxq X S “ tsu, and the set of private neighbours of s P S is
denoted by PGS psq.
A subset F Ď E is called an edge dominating set of G if each edge of G either belongs
to F or is incident to some edge in F , and the edge domination number γ 1 pGq of G is the
minimum cardinality of an edge dominating set of G. We remark that γ 1 pGq equals to
the minimum size of a maximal matching for any graph G [25].
For given subsets X, Y Ď V , we say that X dominates Y in G if Y Ď NG pXq, and let
γpY, Gq denote the least cardinality of a subset X that dominates Y in G.
τ pGq :“ maxtγpA, G˝ q : A Ď V pG˝ q is an independent setu
is called the independence domination number of G.
Remark 2.2. We note that the independence domination number τ pGq seems to first
appear in the work of Aharoni, Berger and Ziv [1] (see also [2]) in which it is denoted by
γ i pGq.
2.2.1. Vertex-edge and edge-vertex dominations in graphs. Most of the research on graph
parameters involved domination mainly concentrates either sets of vertices dominating
all other vertices or sets of edges dominating all other edges in graphs. On the contrary,
Peters [22] has introduced vertex-edge and edge-vertex dominations in graphs. Since
then there has been little research on such a mixing theory of graph dominations [5, 17].
However, the recent results of Dao and Schweig [8, 9, 10] have validated the usefulness of
such notions in the calculation of projective dimension of (hyper)graphs that we describe
next.
A vertex u P V is said to vertex-wise dominate an edge e “ xy P E, if u P NG res. A
subset S Ď V is called a vertex-wise dominating set, if for any edge e P E, there exists a
vertex s P S that vertex-wise dominates e. Furthermore, a vertex-wise dominating set S
of G is called minimal, if no proper subset of S is vertex-wise dominating for G. When S
is a minimal vertex-wise dominating set for G, every vertex in S has a private neighbour
in E, that is, e is a vertex-wise private neighbour of s P S if s vertex-wise dominates e
while no vertex in Sztsu vertex-wise dominates the edge e in G.
PROJECTIVE DIMENSION OF GRAPHS AND THE REGULARITY OF BIPARTITE GRAPHS
5
p1q ΥpGq :“ maxt|S| : S is a minimal vertex-wise dominating set of Gu is called the
upper vertex-wise domination number of G.
p2q βpGq :“ maxt|S| : S is a minimal independent vertex-wise dominating set of Gu is
called the upper independent vertex-wise domination number of G.
Observe that the inequality βpGq ď ΥpGq holds for any graph G as a result of the
definitions.
An edge e “ xy of G is said to edge-wise dominate a vertex u P V if u P NG res. A
subset F Ď E is called an edge-wise dominating set, if for any vertex u P V pG˝ q, there
exists an edge f P F that edge-wise dominates u. Moreover, a vertex v is said to be an
edge-wise private neighbour of f P F if f edge-wise dominates v while no edge in F ztf u
edge-wise dominates v.
p3q ǫpGq :“ mint|F | : F Ď E is an edge-wise dominating set of Gu is called the edgewise domination number of G.
We remark that the parameter ǫpGq can also be considered as the minimum size of an
edge-wise dominating matching of G (compare to Theorem 73 in [17]).
Remark 2.3. We note that Peters [22] chooses to call vertex-wise and edge-wise dominating sets by ve-dominating and ev-dominating sets respectively, and denotes parameters
ΥpGq and ǫpGq by Γve pGq and γev respectively. However, we prefer to follow the most
recent notations of Dao and Schweig [8] in order to overcome the difficulty of readability
caused by the prefixes ve- and ev-.
2.3. Simplicial complexes. An (abstract) simplicial complex ∆ on a finite set V is a
family of subsets of V such that tvu P ∆ for all v P V , and if F P ∆ and H Ă F , then
H P ∆. The elements of ∆ are called faces of it; the dimension of a face F is dimpF q :“
|F | ´ 1, and the dimension of ∆ is defined to be dimp∆q :“ maxtdimpF q : F P ∆u. The 0
and 1-dimensional faces of ∆ are called vertices and edges while maximal faces are called
facets. In particular, we denote by F∆ , the set of facets of ∆. For a given face A P ∆, the
subcomplex lk∆ pAq :“ tF P ∆ : F X A “ H and F Y A P ∆u is called the link of A in ∆.
When ∆ “ IndpGq for some graph G, the existence of vertices satisfying some extra
properties is useful when dealing with the homotopy type:
Theorem 2.4. [12, 18] If NG puq Ď NG pvq, then there is a homotopy equivalence IndpGq »
IndpG´vq. On the other hand, if NG rus Ď NG rvs, then the homotopy equivalence IndpGq »
IndpG ´ vq _ Σ IndpG ´ NG rvsq holds.
In [4] we introduced prime simplicial complexes that constitute building blocks for the
regularity calculations. In detail, a simplicial complex ∆ on V is said to be a prime
simplicial complex over a field k provided that regk p∆ ´ xq ă regk p∆q for any vertex
x P V . In particular, a graph G is called a prime graph (over k) whenever IndpGq is a
prime simplicial complex.
The notion of the Alexander dual of a simplicial complex plays a prominent role on
the calculation of the projective dimension of ∆ due to the Terai’s duality. However, for
its definition, we need to relax the definition of a simplicial complex first. A generalized
simplicial complex ∆ on a vertex set V is simply a family of subsets of V , closed under
inclusion. In particular, vertp∆q :“ tx P V : txu P ∆u is called the set of actual vertices of
∆. Observe that when ∆ “ IndpGq for some graph G, then v R vertpIndpGq_ q if and only
6
TÜRKER BIYIKOĞLU AND YUSUF CIVAN
if v is incident to any edge e P E, i.e., G – K1,n for some n ě 1, where v is the vertex
with degG pvq “ n.
Definition 2.5. Let ∆ be a (generalized) simplicial complex on V . Then the Alexander
dual of ∆ is defined to the simplicial complex ∆_ :“ tF Ď V : V zF R ∆u.
Observe that the equality p∆_ q_ “ ∆ always holds. Furthermore, we recall that
when ∆ “ IndpGq for some (hyper)graph G, the minimal non-faces of IndpGq_ exactly
correspond to minimal vertex covers of G, where a subset U Ď V is a vertex cover of G if
V zU P IndpGq. For this reason, the edge ideal IG_ :“ IIndpGq_ of G _ :“ HpIndpGq_ q is also
known as the vertex cover ideal of G.
3. Bounding the regularity of simplicial complexes from their Levi
graphs
In this section, we prove that the regularity of any simplicial complex (or any hypergraph) can be bounded from above by the regularity of a bipartite graph. This, in
particular, implies together with the Terai’s duality that the regularity of the Levi graph
LpHq of any hypergraph H provides an upper bound to the projective dimension of H.
For that purpose, we first provide a local version of Nagel and Reiner’s result [21]. Along
the way, we show that the induced matching number of the associated bipartite graph
can be derived from the combinatorics of the underlying simplicial complex.
Definition 3.1. Let ∆ be a simplicial complex on V . We define a bipartite graph BS p∆q
for each S Ď V by V pBS p∆qq :“ S Y F∆ and ps, F q P EpBS p∆qq if and only if s R F for
any s P S and F P F∆ , where F∆ is the set of facets of ∆. In particular, we abbreviate
BV p∆q to Bp∆q.
We note that if we consider ∆ as a hypergraph Hp∆q “ pV, F∆ q, the graph Bp∆q is
just the bipartite complement of the Levi graph of Hp∆q.
The following theorem of Nagel and Reiner [21] (see also Jonsson [13]) is our main
motivation for the definition of Bp∆q:
Theorem 3.2. [21] The homotopy equivalence IndpBp∆qq » Σp∆q holds for any simplicial complex ∆, where Σp∆q is the (unreduced) suspension ∆.
We first provide a local version of Theorem 3.2.
Lemma 3.3. Let ∆ be a simplicial complex. Then IndpBS p∆qq » IndpBp∆rSsqq for any
S ĎV.
Proof. Assume that F∆rSs “ tH1 , . . . , Hk u, and define Fi :“ tF P F∆ : Hi Ď F u for each
i P rks, AS :“ tC P F∆ : H ‰ C X S R F∆rSs u and BS :“ tD P F∆ : D X S “ Hu. Note
that if ps, Hi q P EpBp∆rSsqq for some s P S and i P rks, then ps, F q P EpBS p∆qq for
any F P Fi . Suppose otherwise Hi Y tsu P ∆ that in turn implies that Hi Y tsu P ∆rSs.
However, the latter is impossible due to the fact that Hi is a facet of ∆rSs. It then follows
that
NBS p∆q pF q “ NBS p∆q pF 1 q “ NBS p∆rSsq pHi q
for any F, F 1 P Fi and i P rks. If we fix a facet, say Fi P Fi for each i P rks, the homotopy
equivalence IndpBS p∆qq » IndpBS p∆qq ´ pFi zFi qq holds by Theorem 2.4. In other words,
PROJECTIVE DIMENSION OF GRAPHS AND THE REGULARITY OF BIPARTITE GRAPHS
7
we can remove the facets in Fi one by one until there remains only a single facet without
altering the homotopy type.
On the other hand, if C P AS , then C X S is a non empty face of ∆rSs which is not
maximal. So, it is contained by a facet, say Hi . However, the containment C X S Ď Hi
forces that NBS p∆q pFi q Ď NBS p∆q pCq, where Fi P Fi is the facet that we fixed earlier.
Therefore, the removal of any such vertex from IndpBS p∆qq does not alter the homotopy
type. Similarly, if D P BS , then we have NBS p∆q pFi q Ď NBS p∆q pDq for any i P rks so that
IndpBS p∆qq » IndpBS p∆q ´ BS q.
Finally, if we define CS :“ Yki“1 pFi zFi q Y AS Y BS , we conclude that
IndpBS p∆qq » IndpBS p∆q ´ CS q – IndpBS p∆qrS Y tF1 , . . . , Fk usq.
However, the graph BS p∆qrS Y tF1 , . . . , Fk us is isomorphic to BS p∆rSsq that proves the
claim.
Corollary 3.4. IndpBS p∆qq » Σp∆rSsq for any S Ď V .
Proof. Since Σp∆rSsq » IndpBS p∆rSsqq by Theorem 3.2. Thus, the result follows from
Lemma 3.3.
Theorem 3.5. The inequality regpBp∆qq ě regp∆q ` 1 holds for any simplicial complex
∆.
r n´1 p∆rSsq ‰ 0, we
Proof. Suppose that regp∆q “ n. If S Ď V is a subset such that H
r n pIndpBS p∆qqq ‰ 0 by Corollary 3.4 so that regpBp∆qq ě n ` 1.
then have H
We remark that the gap between regp∆q and regpBp∆qq could be arbitrarily large. For
instance, if we define Fn :“ W pKn q for some n ě 3, where W pKn q is the graph obtained
by attaching a pendant vertex to each vertex of Kn (the whisker of the complete graph),
then regpBpIndpFn qqq ě n, while regpIndpFn qq “ 1 since Fn is cochordal.
The relation between a simplicial complex ∆ and its bipartite graph Bp∆q can be
reversible over the class of Sperner bipartite graphs.
Proposition 3.6. If B is an Sp-bipartite graph on Xi , then the family
∆i pBq :“ tF Ď Xi : F Ď Xi zNB pvq for some v P X1´i u
is a simplicial complex on Xi for i “ 0, 1. In particular, we have regp∆i pBqq`1 ď regpBq.
Proof. The first claim is obvious. For the second, observe that the facets of ∆i pBq are
of the form Xi zNB pvq for v P X1´i ; hence, we have Bp∆i pBqq – B so that the stated
inequality follows from Theorem 3.5.
Our next goal is to determine the induced matching number of Bp∆q from the combinatorics of the simplicial complex ∆ that peculiarly involves the Helly number of ∆.
Definition 3.7. Let ∆ be a simplicial complex on V , and let m ě 3 be an integer. An
m-flower of ∆ is a set tpx1 , F1 q, . . . , pxm , Fm qu of pairs pxi , Fi q P V ˆ F∆ such that the set
tx1 , . . . , xm u is a face of ∆, and xi R Fj if and only if i “ j. We say that ∆ is flower-free,
if it has no k-flower for any k ě 3. We then define
#
0,
if ∆ is flower-free,
(3.8)
νp∆q :“
maxtk P N : ∆ has a k ´ floweru, otherwise,
8
TÜRKER BIYIKOĞLU AND YUSUF CIVAN
as the flower number of ∆.
Corollary 3.9. impBp∆qq ě maxt2, νp∆qu for any simplicial complex ∆ other than a
simplex.
Proof. Observe first that since ∆ is not a simplex, it has at least two facets F1 , F2 together
with x P F1 zF2 and y P F2 zF1 . However, we then have Bp∆qrx, y, F1, F2 s – 2K2 , which
forces impBp∆qq ě 2.
On the other hand, it is clear that any m-flower tpx1 , F1 q, . . . , pxm , Fm qu of ∆ gives rise
to an induced matching of Bp∆q.
Recall that a subset K Ď V is said to be a minimal non-face of ∆, if K is itself not
a face of ∆ while any proper subset of it is. The maximum size of minimal non-faces of
∆ is known as the Helly number of ∆, denoted by hp∆q [23]. This is consistent with the
classical definition of the Helly number for set systems [11]. Indeed, if F is a finite set
system, then its Helly number can be defined by hpF q :“ hpN pF qq, where N pF q is the
nerve complex of F . Observe that the inequality hp∆q ď regp∆q ` 1 ď regpBp∆qq holds2,
since if K is a minimal non-face in ∆ of size hp∆q, then ∆rKs – S hp∆q´2 , where S m is the
m-dimensional sphere.
Corollary 3.10. impBp∆qq “ maxt2, νp∆q, hp∆qu for any simplicial complex ∆ other
than a simplex.
Proof. Suppose that impBp∆qq “ m, and let tpx1 , F1 q, . . . , pxm , Fm qu is a maximum induced matching of size m ě 3. If K :“ tx1 , . . . , xm u P ∆, then tpx1 , F1 q, . . . , pxm , Fm qu
is an m-flower so that νp∆q ě m. On the other hand, if K R ∆, then K is a minimal
non-face of ∆ that implies hp∆q ě m.
One of the major open problem concerning our results in this section is the decision of
whether any possible gap between regp∆q and regpBp∆qq is caused because of a combinatorial or an algebraic property of the underlying simplicial complex ∆. In other words,
we wonder whether there exists a combinatorially defined positive integer ηp∆q such that
the equality regpBp∆qq “ regp∆q ` ηp∆q holds for any simplicial complex ∆ (over any
field k). In the particular case of prime simplicial complexes, we predict that ηp∆q “ 1.
Problem 3.11. Does the equality regk p∆q`1 “ regk pBp∆qq hold for any prime simplicial
complex ∆ over k?
3.1. Projective dimension and regularity of hypergraphs. Once we have Theorem 3.5 for arbitrary simplicial complexes, we may naturally interpret it for the projective dimension and the regularity of hypergraphs. For that purpose, if we set BpHq :“
BpIndpHqq, then Theorem 3.5 directly implies the following.
Corollary 3.12. regpHq ` 1 ď regpBpHqq for any hypergraph H.
We note that if B is an Sp-bipartite graph on Xi , then the hypergraph Hi pBq on Xi
with edges NB pvq for v P X1´i satisfies that LpHi pBqq – B for i “ 0, 1. This in particular
implies that regpHi pBqq ` 1 ď regpBq as a result of Corollary 3.12.
2The
first part of this inequality also appears in [16] for finite set systems.
PROJECTIVE DIMENSION OF GRAPHS AND THE REGULARITY OF BIPARTITE GRAPHS
9
We next show that the induced matching numbers of a hypergraph H and that of BpHq
are in fact comparable. We recall first that if tF1 , . . . , Fk u is an induced matching of H,
ř
then the inequality ki“1 p|Fi | ´ 1q ď regpHq holds [19].
ř
Proposition 3.13. If tF1 , . . . , Fk u is an induced matching of H, then ki“1 p|Fi | ´ 1q ď
impBpHqq.
Proof. We let Fi “ txi1 , . . . , xiki u for each i P rks, and define
Lij :“
k
ď
pFt ztxt1 uq Y pFi ztxij uq
t“1
t‰i
for any i P rks and 2 ď j ď ki . Since tF1 , . . . , Fk u is an induced matching, each set Lij is
an independent set in H. So, for each such set, there exists a maximal independent set
Sji in H such that Lij Ď Sji and xij R Sji . Now, the set tpxij , Sji q : 1 ď i ď k and 2 ď j ď ki u
ř
of edges forms an induced matching in BpHq of size ki“1 p|Fi | ´ 1q.
Remark 3.14. We note that the equality is possible in Proposition 3.13. For example,
if we consider the graph 2K2 , we have Bp2K2 q – C8 so that impBp2K2 qq “ imp2K2 q “ 2.
When we employ the Terai’s duality, Theorem 3.5 takes the following form:
Corollary 3.15. proj-dimpHq ď regpLpHqq for any hypergraph H.
_
q “ regpIndpHq_ q ` 1, where the first equality is
Proof. Since proj-dimpHq “ regpIH
the Terai’s duality, the claim follows from Theorem 3.5 together with the fact that
BpIndpHq_ q – LpHq, since any facet of IndpHq_ is of the form V zF , where F P E.
We remark that the Helly number hpIndpHq_ q of the Alexander dual of IndpHq exactly
equals to |V | ´ ipHq, where ipHq is the least cardinality of a maximal independent set of
H, since any facet F of IndpHq corresponds to a minimal non-face V zF of IndpHq_ . This
in particular implies that hpIndpHq_ q ď proj-dimpHq.
Our next target is to provide an upper bound on the regularity of the Levi graph LpHq
in terms of the upper independent vertex-wise domination number. However, we first
recall the following result of Há and Woodroofe:
Theorem 3.16. [14] For any simplicial complex ∆, we have regp∆q to be at most the
maximum size of a minimal face A with the property that lk∆ pAq is a simplex.
Let ∆ be a simplicial complex on V . We call a face A P ∆, an S-face of ∆ if lk∆ pAq is
a simplex. Furthermore, an S-face A is said to be minimal if no proper subset of A is an
S-face of ∆.
Proposition 3.17. An independent set L Ď V is a minimal S-face of IndpGq if and only
if L is a minimal vertex-wise dominating set for G.
Proof. Suppose first that L is a minimal S-face of IndpGq, while it is not vertex-wise
dominating for G. This means that there exists an edge e “ xy of G such that e is
not vertex-wise dominated by L. However, this forces that NG res X L “ H, that is,
txu, tyu P lkIndpGq pLq while tx, yu R lkIndpGq pLq, a contradiction.
10
TÜRKER BIYIKOĞLU AND YUSUF CIVAN
Assume now that L is a minimal vertex-wise dominating set of G. If L is not an S-face
of IndpGq, that is, lkIndpGq pLq is not a simplex, then there exist two vertices tuu, tvu P
lkIndpGq pLq with tu, vu R lkIndpGq pLq. Note that this can only be possible when f “ uv P E.
However, since L is vertex-wise dominating for G, we must have NG rf sXL ‰ H so that one
of the vertices x and y can not be contained by the link of L in IndpGq, a contradiction.
Corollary 3.18. For any graph G with E ‰ H, we have
βpGq “ maxt|L| : L is a minimal S-face of IndpGqu.
We are now ready to provide an upper bound on the regularity of any graph in terms
of a domination parameter of the underlying graph that may be of independent interest.
Theorem 3.19. regpGq ď βpGq for any graph G with E ‰ H.
Proof. This immediately follows from Theorem 3.16 and Corollary 3.18.
We remark that, in general, the invariants βpGq and cochordpGq are incomparable,
where cochordpGq is the cochordal cover number of G, i.e., the least number of cochordal
subgraphs H1 , . . . , Hk of G satisfying EpGq “ YEpHi q. For example, cochordpP4 q “ 1
and βpP4 q “ 2, while βpC7 q “ 2 and cochordpC7 q “ 3. Furthermore, it is not difficult to
construct examples of graphs showing that the gap between regpGq and βpGq could be
arbitrarily large.
Corollary 3.20. proj-dimpHq ď βpLpHqq for any hypergraph H with E ‰ H.
Proof. Since proj-dimpHq ď regpLpHqq for any hypergraph H by Corollary 3.15, the claim
follows from Theorem 3.19.
3.2. An application: Projective dimension of the dominance complex of graphs.
In this subsection, we compute the projective dimension of the dominance complex of any
graph G, or equally that of the closed neighbourhood hypergraph of G as an application
of Corollary 3.20.
We first recall that for a given graph G “ pV, Eq, its dominance complex DompGq is
the simplicial complex on V whose faces are those subsets A Ď V such that the set V zA
is a dominating set for G [18]. Observe that there exists a one-to-one correspondence
between facets of DompGq and minimal dominating sets of G. In order to compute the
projective dimension of DompGq, we first view DompGq as an independence complex of a
hypergraph associated to G.
The closed neighbourhood hypergraph N rGs of a graph G is defined to be the hypergraph
on V whose edges are the minimal elements of the poset tNG rxs : x P V u ordered with
respect to the inclusion.
Lemma 3.21. DompGq – IndpN rGsq for any graph G.
Proof. Since any face of DompGq is clearly an independent set of N rGs, to prove the
claim, it suffices to verify that when X Ď V is an independent set of N rGs, the set V zX
is dominating for G. So, suppose that x P X. It then follows from the definition that
there exists x1 P NG rxs such that NG rx1 s P EpN rGsq. However, since X is an independent
set, we must have NG rx1 s Ę X so that there exists v P NG rx1 s such that v R X. Now,
since NG rx1 s Ď NG rxs, we conclude that vx P E; hence, the vertex x is dominated by
v P V zX.
PROJECTIVE DIMENSION OF GRAPHS AND THE REGULARITY OF BIPARTITE GRAPHS
11
Proposition 3.22. ΥpLpN rGsqq “ ΓpGq for any graph G.
Proof. Suppose that A is a vertex-wise dominating set of LpN rGsq for which we write
A “ Av Y Ae , where Av Ď V and Ae Ď EpN rGsq. Observe that if F P Ae , then there
exists x P V such that F “ NG rxs. Now, we define B :“ tx P V : NG rxs P Ae u, and claim
that the set C :“ Av Y B is a minimal dominating set for G. Let v P V zC be given. It
then follows from the definition that there exists v 1 P NG rvs such that NG rv 1 s P EpN rGsq.
If v 1 P C, then v is dominated by the vertex v 1 so that we may further assume that v 1 R C.
Now, consider the edge pv, NG rv 1 sq of LpN rGsq. Since A is vertex-wise dominating, the
edge pv, NG rv 1 sq must be vertex-wise dominated by a vertex in A. If u P Av is a vertex
that vertex-wise dominates the edge pv, NG rv 1 sq, then u P NG rv 1 s Ď NG rvs so that uv P E,
that is, u dominates v in G. On the other hand, if there exists a vertex w P V such
that NG rws P Ae vertex-wise dominates the edge pv, NG rv 1 sq, then we have v P NG rws.
However, it means that the vertex w P B dominates v in G. This proves that C is a
dominating set for G.
For the other direction, suppose that S Ď V is a dominating set for G. We claim
that it is also a vertex-wise dominating set for LpN rGsq. Consider an edge px, NG rysq
of LpN rGsq. Since S is a dominating set for G, we must have NG rys X S ‰ H. This in
particular forces that there exists s P S such that sy P E. However, it then follows that
the vertex s vertex-wise dominates the edge px, NG rysq.
Theorem 3.23. proj-dimpDompGqq “ ΓpGq for any graph G.
Proof. We clearly have ipN rGsq “ |G| ´ ΓpGq so that hpDompGqq “ |G| ´ ipN rGsq “
ΓpGq ď proj-dimpDompGqq. On the other hand, we have
proj-dimpDompGqq “ proj-dimpN rGsq ď regpLpN rGsqq
ď βpLpN rGsqq ď ΥpLpN rGsqq “ ΓpGq
by Proposition 3.22.
4. Projective dimension of graphs and regularity of subdivisions
In this section, we provide a detail analysis on the consequences of Corollary 3.15 when
H “ H is just a graph. This will include the comparison of upper and lower bounds both
on projective dimension of a graph G and those of the regularity of its subdivision graph
SpGq. However, we first restate Corollary 3.15 in the graph’s case for completeness:
Corollary 4.1. proj-dimpGq ď regpSpGqq for any graph G.
Proposition 4.2. For any graph G, there exists a subgraph H (not necessarily induced)
of G such that proj-dimpHq “ regpSpGqq.
Proof. Suppose that regpSpGqq “ m, and let R Ď V and F Ď E be subsets satisfying
r m´1 pSpGqrR Y F sq ‰ 0 for which we may assume that both sets R and F are minimal
H
with this property. Now, we define H :“ pV pF q, F q and claim that it satisfies the required
property. In order to prove that it suffices to verify the homotopy equivalence SpHq »
SpGqrR Y F s, since we already have ΣpIndpHq_ q » SpHq. If x P V pF qzR and x P
NSpGqrRYF s pf q for some f P F , then degSpGqrRYF s pf q “ 1 by the minimality of the sets
R and F ; hence, if y is the only neighbour of f in SpGqrR Y F s, then we must have
12
TÜRKER BIYIKOĞLU AND YUSUF CIVAN
NSpHq pyq Ď NSpHq pxq, which in turn implies that SpHq » SpHq ´ x. It then follows that
we can remove any vertex in V pF qzR without altering the homotopy type of SpHq, that
is, SpHq » SpHq ´ pV pF qzRq – SpGqrR Y F s.
Lemma 4.3. Let G “ pV, Eq be a graph with regpSpGqq “ m ą 0. Then for any
m ą k ě 0, there exists an induced subgraph Hk of G such that regpSpHk qq “ k.
Proof. By the prime factorization theorem of [4], the graph SpGq has a prime factorization,
say tR1 , . . . , Rs u. Now, if v P V pR1 q X V , we have regpR1 q “ regpR1 ´ NR1 rvsq ` 1, since
R1 is a prime graph. However, since the disjoint union pR1 ´ NR1 rvsq Y R2 Y . . . Y Rs is
an induced subgraph of SpGq ´ NSpGq rvs – SpG ´ vq; it then follows that
regpSpGqq ě regpSpG ´ vqq ě regpR1 ´ NR1 rvsq ` regpR2 q ` . . . ` regpRs q
“pregpR1 q ´ 1q ` regpR2 q ` . . . ` regpRs q
“ regpSpGqq ´ 1.
Therefore, we have either regpSpGqq “ regpSpG´vqq or else regpSpGqq “ regpSpG´vqq`1
so that the claim follows from the induction on the order of G.
One of the immediate consequence of Lemma 4.3 is the following:
Corollary 4.4. Any graph G has an induced subgraph H such that proj-dimpGq “
regpSpHqq.
We note that even if Corollary 4.4 does not exactly determines the corresponding induced subgraph H of G satisfying proj-dimpGq “ regpSpHqq, it naturally translates the
projective dimension of graphs to that of the regularity of bipartite graphs.
Our next aim is to compare various upper and lower bounds on the projective dimension
and the regularity of graphs. We begin with interpreting the relevant invariants of a graph
G in terms of those of the subdivision graph SpGq.
Proposition 4.5. impSpGqq “ |G| ´ γpGq for any graph G without any isolated vertices.
Proof. Suppose that γpGq “ |D| for some dominating set D Ď V . Then for any vertex
x P V zD, there exists a vertex d such that xd P E. Now, the set tpx, xdq : x P V zDu of
edges of SpGq forms an induced matching in SpGq; hence, we have impSpGqq ě |G|´γpGq.
On the other hand, let M “ tpx1 , e1 q, . . . , pxn , en qu be a maximum induced matching of
SpGq. We claim that the set D :“ V ztx1 , . . . , xn u is a dominating set for G so that
impSpGqq ď |G| ´ γpGq. Indeed, for any vertex xi P V zD, the other end vertex of ei , say
yi , must be contained in D, since otherwise the set M would not be an induced matching;
hence, we have xi P NG pDq.
We recall that the equality hpIndpGq_ q “ |G| ´ ipGq holds for any graph G by our
previous remark for hypergraphs. On the other hand, the flower number νpSpGqq :“
νpIndpGq_ q is rather complicated to describe.
We call a subset X Ď V as an unstable dominating set of G, if X is a dominating set
for G with EpGrXsq ‰ H. The least cardinality of an unstable dominating set is called
the unstable domination number of G and denoted by γus pGq. Observe that the inequality
γpGq ď γus pGq ď γpGq ` 1 always holds for any graph G.
Lemma 4.6. νpSpGqq “ |G| ´ γus pGq for any graph G without any isolated vertices.
PROJECTIVE DIMENSION OF GRAPHS AND THE REGULARITY OF BIPARTITE GRAPHS
13
Proof. Suppose that tpx1 , e1 q, . . . , pxn , en qu is a maximum flower in SpGq. This in particular implies that the set X :“ tx1 , . . . , xn u is face of IndpGq_ . However, this means that
X misses at least one edge of G. It then follows that V zX is an unstable dominating set
for G, since each vertex xi is dominated by the other end vertex of ei contained in V zX;
hence, νpSpGqq ď |G| ´ γus pGq.
Assume next that Y Ď V is an unstable dominating set for G with γus pGq “ |Y |.
Then, for each u P V zY , there exists a vertex y P Y such that uy P E. However, the set
tpu, uyq : u P V zY u forms a flower for SpGq, since EpGrY sq ‰ H; hence, we conclude that
|G| ´ γus pGq ď νpSpGqq.
We note that the equality maxt|G| ´ γus pGq, |G| ´ ipGqu “ |G| ´ γpGq clearly holds for
any graph G so that Proposition 4.5 also follows from Corollary 3.10.
Corollary 4.7. If γpGq “ ipGq, then impSpGqq ď proj-dimpGq ď regpSpGqq.
4.1. Domination bounds and the regularity of subdivisions. We prove in this
subsection that all known upper bounds to projective dimension of graphs are in fact
upper bounds to the regularity of subdivision graphs. Before we proceed further, we state
the following upper and lower bounds on the projective dimension of graphs due to Dao
and Schweig [8].
Theorem 4.8. [8] For any graph G,
|G| ´ ipGq ď proj-dimpGq ď |G| ´ maxtτ pGq, ǫpGqu.
We next verify that both upper bounds of Theorem 4.8 are in fact upper bounds to
regpSpGqq. For that purpose, we reformulate the related graph parameters of G in terms
of those of the graph SpGq.
We recall that for any graph G “ pV, Eq, the square G2 of G is defined to be the graph
on V pG2 q “ V such that xy P EpG2 q if and only if distG px, yq ď 2. Moreover, we denote
by αpGq, the independence number of the graph G.
Lemma 4.9. For any graph G, the equality τ pGq “ αpG2 q holds.
Proof. Assume that αpG2 q “ k, and let A “ ta1 , . . . , ak u be an independent set of G2 .
Note that distG pai , aj q ě 3 for any i ‰ j, and in particular, the set A is also an independent
set of G. If X Ď V is a subset satisfying A Ď NG pXq, we necessarily have |A| ď |X|,
since |A X NG pxq| ď 1 for any x P X. However, it then follows that k ď |X| ď τ pGq, that
is, αpG2 q ď τ pGq.
Suppose now that τ pGq “ n, and let A Ď V be an independent subset of G of minimal
order satisfying γpA; Gq “ n. Furthermore, we let X Ď V be a subset such that A Ď
NG pXq and |X| “ n. Since X is minimal, each vertex in X has at least one private
neighbour in A. In other words, we have PAX pxq ‰ H for any x P X.
On the other hand, since A is minimal, every vertex of A must be the private neighbour
of a vertex in X. In other words, for each a P A, there exists a unique xa P X such that
a P PAX pxa q. We then claim that A is an independent set of G2 . Assume otherwise
that there exist a, b P A such that distG pa, bq “ 2, and let u P NG paq X NG pbq be any
vertex. Observe that u R X by the minimality of A. However, it then follows that
A Ď NG pY q, where Y :“ pXztxa , xb uq Y tuu, a contradiction. Therefore, we conclude that
αpG2 q ě |A| ě |X| “ n. This completes the proof.
14
TÜRKER BIYIKOĞLU AND YUSUF CIVAN
The following result of Horton and Kilakos [15] relates the independence number of the
square of a graph to the edge domination number of the subdivision graph.
Theorem 4.10. [15] If G is a graph, then γ 1 pSpGqq ` αpG2 q “ |G|.
We now arrive the proof of the first part of our claim regarding the upper bounds of
Theorem 4.8.
Corollary 4.11. The equality cochordpSpGqq “ |G| ´ τ pGq holds for any graph G. In
particular, we have proj-dimpGq ď cochordpSpGqq.
Proof. We proved in [3] that if G is a graph with girthpGq ě 5, then cochordpGq “ γ 1 pGq.
Since girthpSpGqq ě 6, the result follows from Lemma 4.9 and Theorem 4.10. The last
claim is the consequence of Corollary 4.1 together with the fact that regpHq ď cochordpHq
for any graph H.
Corollary 4.12. If G is a chordal graph, then regpSpGqq “ impSpGqq “ |G| ´ γpGq.
Proof. The equality γpGq “ τ pGq holds for any chordal graph [2]; hence, the claim follows
from Proposition 4.5 and Corollary 4.11.
We next verify that the edge-wise domination number of G is closely related to the
upper vertex-wise domination number of the subdivision graph of G.
Theorem 4.13. ΥpSpGqqq “ |G| ´ ǫpGq for any graph G without any isolated vertex.
Proof. Let M be an edge-wise dominating matching of G with |M| “ ǫpGq, and let U :“
UpMq be the set of unmatched vertices of G with respect to M. We then claim that the
set U Y M is a minimal vertex-wise dominating set for SpGq. Observe that it only suffices
to prove that U Y M is minimal, since it is clearly a vertex-wise dominating set. If u P U,
then there exists a v P V pMq such that eu “ uv P E, since M is an edge-wise dominating
matching of G. In particular, the edge pu, eu q of SpGq is only dominated by the vertex
u, that is, the edge pu, eu q is a vertex-wise private neighbour of u. On the other hand, if
f “ xy P M, then the edge px, f q of SpGq is a vertex-wise private neighbour of f ; hence,
the claim follows. However, we then have |G| ´ ǫpGq “ |G| ´ |M| “ |U| ` |M| ď ΥpSpGqq.
We prove the reversed inequality in three steps. Note that when A is a vertex-wise
dominating set of SpGq, we write A “ Av Y Ae such that Av Ď V and Ae Ď E. We first
verify that there exists a vertex-wise dominating set A of SpGq of appropriate size such
that Ae ‰ H. We then show that among any such vertex-wise dominating sets, we can
find one A for which the set Ae is an edge-wise dominating matching of G.
Claim 1. For any connected graph G, there exists a minimal vertex-wise dominating
set A with ΥpSpGqq “ |A| such that Ae ‰ H.
Proof of Claim 1: Suppose that SpGq admits a minimal vertex-wise dominating set A
with ΥpSpGqq “ |A| such that Ae “ H, that is, A “ Av Ď V . We note that for each
x P A, there exists a vertex yx P NG pxq such that yx R A, since otherwise A could not
be minimal. We claim that such a vertex yx P NG pxq is unique for each x P A. Suppose
otherwise that there exist x P A together with y1 , . . . , yk P NG pxq such that y1 , . . . , yk R A.
However, it then follows that the set pAztxuq Y te1 , . . . , ek u, where ei “ yi x for i P rks,
would be a minimal vertex-wise dominating set for SpGq having larger size than A, which
is not possible. Therefore, the vertex yx must be unique for each x P A. Now, if we
PROJECTIVE DIMENSION OF GRAPHS AND THE REGULARITY OF BIPARTITE GRAPHS
15
let B :“ pAztxuq Y tex u, where ex :“ yx x P E, then the set B is a minimal vertex-wise
dominating set of SpGq with Be ‰ H and |B| “ |A| “ ΥpSpGqq.
Claim 2. For any connected graph G, there exists a minimal vertex-wise dominating
set A with ΥpSpGqq “ |A| such that Ae is an edge-wise dominating set for G.
Proof of Claim 2: We first verify that such a set Ae can be turned into an edge-wise
dominating set for G. Suppose that Ae is not edge-wise dominating for G. So, there exists
a vertex x P V such that x is neither an end vertex of an edge in Ae nor it is adjacent
to any end vertex of an edge in Ae . Assume first that x R Av . This in particular implies
that NG pxq Ď Av . Pick a neighbour, say y P NG pxq. Observe that if f is an edge of G
incident to y, then f R Ae . If NG pyqztxu Ď Av , then the set pAztyuq Y txyu is a minimal
vertex-wise dominating set of SpGq having size ΥpSpGqq such that x is dominated by the
edge xy in G. Thus, we may assume that the vertex y has at least one neighbour in G
other than x not contained by Av . We denote by Ly , the set of such vertices. We note
that the set Ly can not contain a vertex w such that any edge of G incident to w is not
contained by Ae . Indeed, if we add the edge yw to pAztyuq Y txyu for any such vertex
w P Ly , the resulting set would be a minimal vertex-wise dominating set for SpGq, which
is not possible. This forces that every vertex w in Ly has at least one incident edge in Ae .
However, it then follows that the set pAztyuq Y txyu is a minimal vertex-wise dominating
set of SpGq for which the vertex x is now dominated by the edge xy in G.
Assume next that x P Av . In such a case there must exist at least one vertex in NG pxq
not contained in Av by the minimality of A. We denote by Tx , the set of such vertices.
Observe that no incident edges to a vertex z in Tx can be contained in Ae by the choice
of x. However, this forces that any vertex adjacent to a vertex in Tx must belong to Av .
Now, if we define A1 :“ pAztxuq Y txz : z P Tx u, then the set A1 is clearly a minimal
vertex-wise dominating set for SpGq. However, since |A| “ ΥpSpGqq, this could be only
possible when |Tx | “ 1. If Tx “ tzu, then the set pAztxuq Y txzu is a minimal vertex-wise
dominating set for SpGq for which the vertex x is dominated by the edge xz in G.
As a result, since we only replace a vertex in V by an edge in E for A, the resulting set
Ae is always an edge-wise dominating set of G.
Claim 3. For any connected graph G, there exists a minimal vertex-wise dominating
set A with ΥpSpGqq “ |A| such that Ae is an edge-wise dominating matching for G.
Proof of Claim 3: Suppose that A is a vertex-wise dominating set with ΥpSpGqq “ |A|
satisfying the following properties:
piq Ae is an edge-wise dominating set for G of minimum possible size among any minimal vertex-wise dominating sets for SpGq of maximum size,
piiq the set Ae contains least number of pairs of incident edges in G among sets satisfying the condition piq.
Let e “ xy and f “ yz be two such edges in Ae . This in particular implies that neither x
nor z can belong to Av . Moreover, no edge incident to either x or z can be contained in Ae
by the minimality of A. Note also that we must have degG pxq, degG pzq ą 1. For instance,
if degG pxq “ 1, then the set B :“ pAzteuq Y txu is a minimal vertex-wise dominating set
for SpGq in which Be has fewer pairs of adjacent edges than Ae , a contradiction. So, let
w be a neighbour of x in G. If w R Av , we then define C :“ pAzteuq Y txwu so that the
16
TÜRKER BIYIKOĞLU AND YUSUF CIVAN
set C provides a minimal vertex-wise dominating set for SpGq in which Ce has fewer pairs
of incident edges than Ae . Therefore, we must have NG pxqztyu Ď Av . We claim that
not every vertex in NG pxqztyu can be edge-wise dominated by some edge in Ae . Since
otherwise, the set B :“ pAzteuq Y txu would be a minimal vertex-wise dominating set for
SpGq for which |Be | ă |Ae |, violating the condition piq. So, we define Kx to be the subset
of NG pxqztyu consisting of those vertices being the edge-wise private neighbours of the
edge e “ xy. If we pick a vertex, say p P Kx , and define D :“ pAzteuq Y txpu, then the
set D is a minimal vertex-wise dominating set for SpGq with |D| “ |A|. In particular,
De is an edge-wise dominating set for G having fewer pairs of incident edges than Ae , a
contradiction. This completes the proof of Claim 3.
Finally, we conclude that ΥpSpGqq “ |A| “ |Av |`|Ae | ď |G|´|Ae |, where the inequality
is due to the fact that Ae is a matching for G. On the other hand, since Ae is an edgewise dominating set for G, we must have |Ae | ě ǫpGq from which we conclude that
ΥpSpGqq ď |G| ´ ǫpGq.
The first part of the proof of Theorem 4.13 together with the fact that βpHq ď ΥpHq
for any graph H give rise to the following fact.
Corollary 4.14. βpSpGqq “ ΥpSpGqq “ |G| ´ ǫpGq for any graph G.
We remark that it is not known in general whether the equality βpHq “ ΥpHq holds
for any graph H [17].
The following result completes the proof of our claim regarding the upper bounds of
Theorem 4.8:
Corollary 4.15. regpSpGqq ď ΥpSpGqq “ |G| ´ ǫpGq for any graph G.
Proof. The claim follows from Theorems 3.19 and 4.13.
We close this section by showing that the regularity of any bipartite graph can be
bounded from above by its induced matching number from which we may readily deduced
an upper bound on the projective dimension of graphs.
Theorem 4.16. If B is a bipartite graph with a bipartition V pBq “ X YY , then regpBq ď
1
pimpBq ` mint|X|, |Y |uq.
2
Proof. We only prove the inequality regpBq ď 12 pimpBq ` |Y |q from which the general
claim follows.
We start with the graph B0 :“ B having the bipartition X0 :“ X and Y0 :“ Y . We
then introduce a reduction process that creates a bipartite graph Bi in each state i ě 0,
and associate a pair of integers pprci , imci q to Bi by letting prc0 :“ 0 and imc0 :“ 0 at
the bottom.
Now, pick a vertex xi P Xi for i ě 0, and assume first that xi is not a prime vertex
of Bi . If degBi pxi q ě 2, we define Bi`1 to be the bipartite graph having the bipartition
V pBi`1 q “ Xi`1 Y Yi`1 , where Xi`1 :“ Xi ´ xi , Yi`1 :“ Yi together with pprci`1 , imci`1 q “
pprci , imci q.
On the other hand, if degBi pxi q “ 1 and yi is the only neighbour of it in Bi , we let
Bi`1 to be the bipartite graph with Xi`1 :“ Xi ´ NBi pyi q, Yi`1 :“ Yi ´ yi , and set
imci`1 :“ imci `1, while prci`1 :“ prci (compare to Lemma 3.25 of [4]).
PROJECTIVE DIMENSION OF GRAPHS AND THE REGULARITY OF BIPARTITE GRAPHS
17
If xi is a prime vertex of Bi , we take Bi`1 to be the bipartite graph with Xi`1 :“ Xi ´xi ,
Yi`1 :“ Yi ´NBi pxi q, and set prci`1 :“ prci `1 and imci`1 :“ imci . We proceed this process
until there exists a j ě 0 such that Xj “ H or Yj “ H.
Observe that regpBq “ prcj ` imcj and that the inequalities imcj ď impBq and prcj ď
1
p|Y | ´ imcj q hold. However, we then have
2
1
regpBq “ prcj ` imcj ď p|Y | ´ imcj q ` imcj
2
1
“ p|Y | ` imcj q
2
1
ď p|Y | ` impBqq.
2
Corollary 4.17. For any graph G, we have proj-dimpGq ď |G| ´ 21 γpGq.
Proof. Since impSpGqq “ |G| ´ γpGq by Proposition 4.5, it then follows from Corollary 4.1
together with Theorem 4.16 that
1
1
1
proj-dimpGq ď regpSpGqq ď pimpSpGqq ` |G|q “ p|G| ´ γpGq ` |G|q “ |G| ´ γpGq.
2
2
2
Observe that the bound of Corollary 4.17 is sharp as the graph C4 satisfies that
proj-dimpC4 q “ 3 “ regpC8 q “ |C4 | ´ 12 γpC4 q “ 3.
We leave it open the discussion whether the hypergraph versions of the domination
parameters, the independence domination number τ pHq and the edge-wise domination
number ǫpHq of a hypergraph H as they are introduced in [9, 10] provide upper bounds on
the regularity of the Levi graph of H. However, we prefer to state an interesting connection
between the parameter ǫpHq and the induced matching number of the corresponding Levi
graph.
Definition 4.18. Let B be a bipartite graph with a bipartition X Y Y containing no
isolated vertex. We define hX pBq to be the least cardinality of a subset S Ď X such that
for every y P Y there exists an s P S satisfying distB ps, yq ď 3 (the parameter hY pBq can
be defined analogously).
Lemma 4.19. ǫpHq “ hE pLpHqq ď impLpHqq for any hypergraph H “ pV, Eq.
Proof. The equality is a direct consequence of the definitions, and for the inequality,
suppose that tE1 , . . . , Ek u Ď E is a minimal edge-wise dominating set of cardinality ǫpHq.
By the minimality, there exists a vertex vi P Ei which is privately edge-wise dominated
by Ei in LpHq. However, the set tpv1 , E1 q, . . . , pvk , Ek qu forms an induced matching in
LpHq.
Observe that Theorem 4.16 naturally implies that regpLpHqq ď |H| ´ hE pHq whenever
impLpHqq ď |H|{3 or regpLpHqq ď |H|{2. On the other hand, the inequality τ pHq `
hV pLpHqq ď |H| trivially holds for any hypergraph H. In particular, we wonder whether
such a special notion of a distance-three domination on bipartite graphs always provides
an upper bound to the regularity of such graphs.
18
TÜRKER BIYIKOĞLU AND YUSUF CIVAN
Question 4.20. Does the inequality regpBq ď maxt|X| ´ hY pBq, |Y | ´ hX pBqu hold for
any bipartite graph with a bipartition X Y Y containing no isolated vertex?
5. Projectively prime graphs
In this section, in analogy with the notion of prime graphs introduced for the regularity
calculations of graphs [4], we demonstrate that a similar notion could be useful for the
projective dimension of graphs as well.
Definition 5.1. A connected graph G is called a projectively prime graph over a field
k, if proj-dimk pG ´ xq ă proj-dimk pGq for any vertex x P V pGq. Furthermore, we call a
connected graph G as a perfect projectively prime graph if it is a projectively prime graph
over any field.
Obvious examples of perfect projectively prime graphs can be given by the paths P3k
and P3k`2 for any k ě 1, and cycles Cn for any n ě 3. The primeness of the corresponding paths directly follows from Corollary 4.7 together with the fact that regpSpPm qq “
impSpPm qq “ t 2m´1
u for any m ě 2. Apart from these, our next result shows that the
3
join of any two graphs is a perfect projectively prime graph.
When G and H are two graphs on the disjoint set of vertices, we write G ˚ H, the
join of G and H, for the graph obtained from the disjoint union G Y H by inserting an
edge between any two vertices x P V pGq and y P V pHq. The calculation of projective
dimension of the graph G ˚ H seems to first appear in [20], while we here provide an
independent proof.
Proposition 5.2. If G and H are two graphs on the disjoint sets of vertices, then the
graph G˚H is a perfect projectively prime graph. In particular, we have proj-dimpG˚Hq “
regpSpG ˚ Hqq “ |G| ` |H| ´ 1.
Proof. We prove the second claim from which the projectively primeness of G ˚ H directly
follows. Since ǫpG ˚ Hq “ 1, we have proj-dimpG ˚ Hq ď regpSpG ˚ Hqq ď |G| ` |H| ´ 1
by Corollary 4.15. On the other hand, the Alexander dual of IndpG ˚ Hq triangulates a
p|G| ` |H| ´ 3q-dimensional sphere. Indeed, the complex IndpG ˚ Hq_ is just the simplicial
join of the boundary complexes of corresponding simplexes on V pGq and V pHq. Therefore,
we have |G| ` |H| ´ 1 ď proj-dimpG ˚ Hq from which we conclude the claim.
We note that the complete multipartite graph Kn1 ,...,nm for any ni ě 1 and m ě 2 is a
perfect projectively prime graph as a result of Proposition 5.2.
As in the case of prime graphs, the notion of projectively prime graphs allows us
to reformulate the projective dimension as a generalized (weighted) induced matching
problem that we describe next.
Definition 5.3. Let G be a graph and let R “ tR1 , . . . , Rr u be a set of pairwise vertex
disjoint induced subgraphs of G such that |V pRi q| ě 2 for each 1 ď i ď r. Ť
Then R is
said to be an induced decomposition of G if the induced subgraph of G on ri“1 V pRi q
contains no edge of G that is not contained in any of EpRi q, and and R is maximal with
this property. The set of induced decompositions of a graph G is denoted by IDpGq.
PROJECTIVE DIMENSION OF GRAPHS AND THE REGULARITY OF BIPARTITE GRAPHS
19
Definition 5.4. Let R “ tR1 , . . . , Rr u be an induced decomposition of a graph G. If each
Ri is a projectively prime graph, then we call R as a projectively prime decomposition of
G, and the set of projectively prime decompositions of a graph G is denoted by PDpGq.
Obviously, the set PDpGq is non-empty for any graph G.
Theorem 5.5. For any graph G and any field k, we have
r
ÿ
proj-dimk pGq “ maxt proj-dimk pHi q : tH1 , . . . , Hr u P PDk pGqu.
i“1
Proof. If G is itself a projectively prime graph, there is nothing to prove. Otherwise there
exists a vertex x P V such that proj-dimpGq “ proj-dimpG ´ xq. If G ´ x is a projectively
prime graph, then tG ´řxu P PDpGq so that the result follows. Otherwise, we have
proj-dimpG ´ xq “ maxt ti“1 proj-dimpSi q : tS1 , . . . , St u P PDpG ´ xqu by the induction.
However, since PDpG ´ xq Ď PDpGq for such a vertex, the claim follows.
By taking the advantage of Theorem 5.5, we next present examples of graphs showing
that most of the bounds on the projective dimension of graphs involving domination
parameters are in fact far from being tight. Beforehand, we need the following well-known
fact first.
Lemma 5.6. [7] For any vertex x of a graph G,
(5.7)
proj-dimpGq ď maxtproj-dimpG ´ NG rxsq ` degG pxq, proj-dimpG ´ xq ` 1u.
Proposition 5.8. For each integer a ě 1, there exists a graph Ga satisfying
(5.9)
|Ga | ´ ipGa q ` a ă proj-dimpGa q ă |Ga | ´ γpGa q ´ a.
Proof. Assume that k and r are two positive integers with k ą r ` 1. Let Tk :“ Kk,k
be the complete bipartite graph and fix a vertex v P V pTk q. We take Sr :“ rP3s`1 ,
which is the disjoint r copies of the path on 3s ` 1 vertices (s ě 1) together with a
leave vertex xi from each copy. Now, we from the graph G2k,r on V pTk q Y V pSr q such
that EpG2k,r q “ EpTk q Y EpSr q Y txi v : i P rrsu. Observe that ipG2k,r q “ sr ` k and
γpGr q “ sr ` 2. Since tKk,k , rP3s u is a projectively prime decomposition of G2k,r , we
have 2k ´ 1 ` 2rs ď proj-dimpG2k,r q by Theorem 5.5. Moreover, proj-dimpG2k,r ´ vq “
p2k ´2q`2rs and proj-dimpG2k,r ´NG2k,r rvsq “ 2rs so that proj-dimpGr q “ 2k `2rs´1 by
the inequality (5.7). Consequently, if we choose k “ 2r, we then have proj-dimpG4r,r q ´
p|G4r,r | ´ ipG4r,r qq “ r ´ 1 and p|G4r,r | ´ γpG4r,r qq ´ proj-dimpG4r,r q “ r ´ 1. This in
particular implies that if we define Ga :“ G4pa`2q,a`2 , then
p|Ga | ´ ipGa qq ` a ă proj-dimpGa q ă p|Ga | ´ γpGa qq ´ a
as claimed.
Proposition 5.10. For each integer b ě 1, there exists a graph Hb satisfying
(5.11)
|Hb | ´ γpHb q ` b ă proj-dimpHb q ă |Hb | ´ ǫpHb q ´ b.
Proof. We assume that k and r are two positive integers with k ą r ` 2.
Lk,r :“ rKk,k and pick a vertex vi from each copy of Kk,k . Then, define Rk,r
graph on V pLk,r q Y txu, where x is a vertex disjointly chosen from V pLk,r q,
EpRk,r q :“ EpLk,r q Y txvi : i P rrsu. Observe that |Rk,r | “ 2rk ` 1, γpRk,r q
We take
to be the
such that
“ 2r and
20
TÜRKER BIYIKOĞLU AND YUSUF CIVAN
ǫpRk,r q “ r. Furthermore, we have proj-dimpRk,r q “ p2k ´ 1qr, which follows from the
inequality (5.7) together with the fact that trKk,k u is a projectively prime decomposition
of Rk,r .
Secondly, we define another graph Zk,r on V ppr ` 1qKk,k q Y ty2 , . . . , yr`1 u by EpZk,r q “
Eppr ` 1qKk,k q Y tv1 yj , vj yj : j P t2, 3, . . . , r ` 1uu, where the vertex vl is chosen from the
lth -copy of Kk,k . Observe that |Zk,r | “ 2kpr ` 1q ` r, γpZk,r q “ 2pr ` 1q and ǫpZk,r q “ r ` 1.
Again, by considering the vertex v1 in the inequality (5.7) together with the projectively
prime decomposition tpr ` 1qKk,k u of Zk,r , we note that proj-dimpZk,r q “ p2k ´ 1qpr ` 1q.
Now, assuming that we construct Rk,r and Zk,r on the disjoint sets of vertices, we form
a new graph Hk,r on V pRk,r q Y V pZk,r q by EpHk,r q :“ EpRk,r q Y EpZk,r q Y txv1 u. It
then follows that |Hk,r | “ 2kp2r ` 1q ` r ` 1, γpHk,r q “ 2p2r ` 1q and ǫpHk,r q “ 2r ` 1.
Once again, the inequality (5.7) with respect to the vertex v1 of Hk,r together with the
projectively prime decomposition trKk,k , pr ` 1qKk,k u, where the first prime arises from
Rk,r and the second comes up from Zk,l yields that proj-dimpHk,r q “ p2k ´ 1qp2r ` 1q.
Finally, if we define Hb :“ Hb`4,b`1 , we conclude that proj-dimpHb q ´ p|Hb | ´ γpHb qq “
b ` 1 and p|Hb | ´ ǫpHb qq ´ proj-dimpHb q “ b ` 2 so that the claimed inequality
|Hb | ´ γpHb q ` b ă proj-dimpHb q ă |Hb | ´ ǫpHb q ´ b
holds.
Regarding the results of Propositions 5.8 and 5.10 together with Theorem 5.5, the most
prominent question would be investigation of the structural properties of projectively
prime graphs as well as operations on graphs that preserve projectively primeness. It
would be interesting to determine under what conditions on a projectively prime graph
G, the inequality proj-dimpGq ď |G|´γpGq holds. We predict that C3k`1 -free projectively
prime graphs satisfy such an inequality.
References
[1] R. Aharoni, E. Berger and R. Ziv, A tree version of König’s theorem, Combinatorica,
22, (2002), 335-343.
[2] R. Aharoni and T. Szabó, Vizing’s conjecture for chordal graphs, Discrete Math.,
309, (2009), 1766-1768.
[3] T. Bıyıkoğlu and Y. Civan, Vertex-decomposable graphs, codismantlability, CohenMacaulayness, and Castelnuovo-Mumford regularity, Electronic J. Combin., 21:1,
(2014), #P1, 1-17.
[4] T. Bıyıkoğlu and Y. Civan, Castelnuovo-Mumford regularity of graphs, available at
arXiv:1503.06018(v1), (2015), 43pp.
[5] R. Boutrig, M. Chellali, T.W Haynes and S.T. Hedetniemi, Vertex-edge domination
in graphs, Aequat. Math., 90, (2016), 355-366.
[6] P. Csorba, Subdivision yields Alexander duality on independence complexes, Electronic J. Combin., 16:2, (2009), #P11, 1-7.
[7] H. Dao, C. Huneke and J. Schweig, Bounds on the regularity and projective dimension
of ideals associated to graphs, J. Algebraic Combin., 38:1, (2013), 37-55.
PROJECTIVE DIMENSION OF GRAPHS AND THE REGULARITY OF BIPARTITE GRAPHS
21
[8] H. Dao and J. Schweig, Projective dimension, graph domination parameters, and
independence complex homology, J. Combin. Theory, Series A, 432:2, (2013), 453469.
[9] H. Dao and J. Schweig, Bounding the projective dimension of a squarefree monomial
ideal via domination in clutters, Proc. Amer. Math. Soc., 143, (2015), 555-565.
[10] H. Dao and J. Schweig, Further applications of clutter domination parameters to
projective dimension, J. Algebra, 432, (2015), 1-11.
[11] E.C. de Verdière, G. Ginot and X. Goaoc, Helly numbers of acyclic families, Advances
Math., 253, (2014), 163-193.
[12] A. Engström, Complexes of directed trees and independence complexes, Discrete
Math., 309:10, (2009), 3299-3309.
[13] J. Jonsson, On the topology of independence complexes of triangle-free graphs, unpublished manuscript, (2011), available at math.kth.se/ jakobj.
[14] H.T. Há and R. Woodroofe, Results on the regularity of squarefree monomial ideals,
Adv. Appl. Math, 58, (2014), 21-36.
[15] J.D. Horton and K. Kilakos, Minimum edge dominating sets, Siam J. Disc. Math.,
6:5, (1993), 375-387.
[16] G. Kalai and R. Meshulam, Leray numbers of projections and a topological Helly type
theorem, J. Topology, 1, (2008), 551-556.
[17] J. Lewis, Vertex-edge and Edge-vertex Parameters in Graphs, Phd Thesis, Clemson
University, (2007).
[18] M. Marietti and D. Testa, A uniform approach to complexes arising from forests,
Elect. J. Comb., 15, (2008), #R101.
[19] S. Morey and R.H. Villarreal, Edge ideals: algebraic and combinatorial properties, In
Progress in Commutative Algebra, Combinatorics and Homology, Vol.1, De Gruyter,
Berlin, (2012), pp. 85-126.
[20] A. Mousivand, Algebraic properties of product of graphs, Commun. Algebra, 40,
(2012), 4177-4194.
[21] U. Nagel and V. Reiner, Betti numbers of monomial ideals and shifted skew shapes,
Electronic J. Combin., 16:2, (2009), #R3.
[22] J.W. Peters, Theoretical and Algorithmic Results on Domination and Connectivity,
Phd Thesis, Clemson University, (1986).
[23] M. Tancer, Intersection patterns of convex sets via simplicial complexes, a survey,
Thirty Essays on Geometric Graph Theory, Springer, New York (2013), 521-540.
[24] N. Terai, Alexander duality theorem and Stanley-Reisner rings, Free resolution of
coordinate rings of projective varieties and related topics (Japanese) (Kyoto, 1998).
Surikaisekikenkyusho Kokyuroku, No:1078, (1999), 174-184.
[25] M. Yannakakis and F. Gavril. Edge dominating sets in graphs, SIAM J. on Applied
Math., 38:3, (1980), 364-372.
Department of Mathematics, Suleyman Demirel University, Isparta, 32260, Turkey.
E-mail address: [email protected] and [email protected]
| 0math.AC
|
1
Diffusion leaky LMS algorithm: analysis and implementation
Lu Lu, Haiquan Zhao*
School of Electrical Engineering, Southwest Jiaotong University, Chengdu, China.
https://doi.org/10.1016/j.sigpro.2017.05.015
COMMENT: This paper has been published in Signal Processing
ABSTRACT—The diffusion least-mean square (dLMS) algorithms have attracted much attention owing to its robustness for
distributed estimation problems. However, the performance of such filters may change when they are implemented for suppressing
noises from speech signals. To overcome this problem, a diffusion leaky dLMS algorithm is proposed in this work, which is
characterized by its numerical stability and small misadjustment for noisy speech signals when the unknown system is a lowpass
filter. Finally, two implementations of the leaky dLMS are introduced. It is demonstrated that the leaky dLMS can be effectively
introduced into a noise reduction network for speech signals.
Keywords: Distributed estimation; Diffusion LMS; Leaky LMS; Noisy speech signal.
1. Introduction
Since the simple structure and low computational burden, the least-mean square (LMS) algorithm has become a widely used
adaptive filter. However, in practice, it is well known that direct implementation of the conventional LMS algorithm can be
problematic. Such problems can be outlined as follows. (i) Numerical problem, caused by inadequacy of excitation in the input
signal; (ii) stagnation behaviour, often occur in low input signal scenarios, making the performance degradation of the algorithm. To
solve these problems, the leaky LMS algorithm [1] as its variants were proposed in diverse fields, such as active noise control (ANC)
[2], channel estimation [3], and nonlinear acoustic echo cancellation (NLAEC) [4].
Recently, to estimate some parameters of interest from the data collected at nodes distributed over a geographic region, several
distributed estimation algorithms were developed, including incremental [5-6], and diffusion algorithms [7-10], etc. In the
incremental strategy, a cyclic path is required the definition over the nodes, and this technique is also sensitive to link failures [11].
On the other hand, the diffusion method, is widely used because of its ease of implementation. In this strategy, each node
E-mail addresses: [email protected] (L. Lu), [email protected] (H. Zhao)
2
communicates with a subset of its neighbours, which achieves more data from their neighbours with moderate amount of
communications. Particularly, in [12], two versions of the diffusion LMS (dLMS) algorithms, adapt-then-combine (ATC) and
combine-then-adapt (CTA) were proposed, based on the different orders of adaptation and combination steps. Note that the ATC
method outperforms the CTA method in all cases, and better performance can be achieved if the measurement are shared. Due to its
merit, the ATC version of the dLMS algorithm (ATC dLMS) has been introduced to subband adaptive filter [13], information
theoretic learning (ITL) [14] to obtain improved performance.
Motivated by these considerations, in this paper, we proposed a leaky dLMS algorithm with two combination strategies-ATC and
CTA, resulting in ATC leaky dLMS and CTA leaky dLMS algorithms. Very recently, a distributed incremental leaky LMS was
proposed to surmount the drift problem of dLMS algorithm [15]. Unfortunately, it is derived from the incremental approach, which
prohibits its practical applications. Compared with the existing algorithms, the leaky dLMS algorithm is derived by minimizing the
instantaneous leaky objective function rather than the mean square error cost function. Moreover, it has a superiority performance
via ATC strategy.
Notation: Throughout this paper, we use {}
⋅ to represent a set, | ⋅ | denotes absolute value of a scalar, (⋅)T denotes transposition,
E{}
⋅ denotes expectation, tr{}
⋅ denotes trace operator, || ⋅ ||22 denotes l2-norm, and {}
⋅ o is the real value of parameter. Besides, we
use boldface and normal letters to denote the random quantities and deterministic quantities, respectively.
2. Diffusion LMS strategies
Consider a network of N sensor nodes distributed over a geographic area. At each time instant i, each sensor node k ∈ {1, 2,..., N }
has access to the realization {d k (i ), uk .i } of some zero-mean random process {d k (i ), u k .i } , and d k (i ) is a scalar and u k .i is an
regression vector with length M. Suppose these measurements follow a standard model given by:
d k (i) = u k ,i wo + ν k (i)
(1)
where wo ∈ ℂ M ×1 is the unknown parameter vector, and ν k (i ) is the measurement noise with variance σ v2, k . Here, we assume that
u k ,i and ν k (i ) are spatially independent and independent identically distributed (i.i.d.), and ν k (i ) is independent of u k ,i .
The dLMS algorithm is obtain by minimizing a linear combination of the local mean square error (MSE):
J kloc ( w) =
∑a
l ,k
l∈ N k
E | el ,i |2 =
∑a
l ,k
E | d l (i) − ul ,i w |2
(2)
l ∈N k
where N k is the set of nodes with which node k shares information (including k itself). The weighting coefficient {al , k } are real,
3
N
non-negative, and satisfy: ∑ al , k = 1 .
l =1
The dLMS algorithm obtains the estimation via two steps, adaptation and combination. According to the order of these two steps,
the diffusion LMS algorithm is classified into the ATC dLMS and CTA dLMS algorithms. The updation equation of ATC dLMS
can be expressed as
ϕk ,i = wk ,i −1 + µ ∑ cl , k ulT,i (d l (i ) − ul ,i wk ,i −1 ) (adaptation)
l ∈N k
(combination)
wk ,i = ∑ al , k ϕl ,i
l∈ N k
(3)
where µ is the step size (learning rate), and ϕk is the local estimates at node k. The weighting coefficients {cl ,k } is the real,
non-negative, satisfying the condition cl , k = al , k = 0 if l ∉ N k . Similarly, the CTA dLMS algorithm can be given as
ϕk ,i −1 = ∑ al , k wl ,i −1
(combination)
l ∈N k
T
wk ,i = ϕk ,i −1 + µ ∑ cl , k ul ,i (d l (i) − ul ,iϕ k ,i −1 ) (adaptation).
l ∈N k
(4)
3. Proposed diffusion leaky LMS algorithm
In this section, the leaky dLMS is proposed to address the two problems mentioned above. For each node k, we seek an estimate
of wo by minimizing the following cost function:
∑a
J kloc ( w) =
l ,k
E | el ,i |2 + γ wT w
l∈ N k
=
(5)
∑a
2
l ,k
T
E | dl (i ) − u l ,i w | + γ w w
l∈ N k
where γ>0 is the leakage coefficient, and w is the estimate of wo . Using the steepest descent algorithm, yields
∇ w J kloc ( w) =
∂{E | el ,i |2 +γ wT w}
∑a
l ,k
∂w
l ∈N k
.
(6)
Therefore, the updating of proposed algorithm for estimating wo at node k can be derived from the steepest descent recursion
wk ,i = wk ,i −1 − µ ∑ al , k
∂{E | el ,i |2 +γ wT w}
∂w
l ∈N k
|wk .i−1 .
(7)
Under the linear combination assumption [12,14], let us define the linear combination wk ,i at node k as
wk ,i −1 =
∑a
l ,k
l ∈N k
ϕl ,i −1 .
(8)
4
Introducing (8) to (7), we can obtain an iterative formula for the intermediate estimate
ϕl ,i = ϕl ,i −1 − µ
≈ ϕl ,i −1 − µ
∂E | el ,i |2 +γ wT w
∂w
∂E | el ,i |2 +γ wT w
∂w
= (1 − γµ )ϕl ,i −1 + µ ulT.i el ,i .
|wk .i−1
|wl .i−1
(9)
Hence, we obtain the leaky dLMS algorithm by transforming the steepest-descent type iteration of (7) into a two-step iteration
∂E | ek ,i |2 +γ wT w
=
w
−
|wk .i −1
ϕ
µ
k ,i
k , i −1
∂w
wk ,i = ∑ al , k ϕl ,i .
l∈ N k
(10)
In adaptation step of (10), the local estimate ϕk ,i −1 is replaced by linear combination wk ,i −1 . Such substitution is reasonable,
because the linear combination contains more data information from neighbor nodes than ϕk ,i −1 [12]. Then, we extend the leaky
dLMS algorithm to its ATC and CTA forms.
ATC leaky dLMS algorithm:
●
ϕk ,i = (1 − µγ ) wk ,i −1 + µ ∑ cl , k ulT,i ( dl (i) − ul ,i wk ,i −1 ) (adaptation)
l ∈N k
(combination).
wk ,i = ∑ al , k ϕl ,i
l∈ N k
(11)
ϕk ,i −1 = ∑ al , k wl ,i −1
(combination)
l ∈N k
T
wk ,i = (1 − µγ )ϕ k ,i −1 + µ ∑ cl , k ul ,i (d l (i ) − ul ,iϕk ,i −1 ) (adaptation).
l ∈N k
(12)
CTA leaky dLMS algorithm:
●
4. Simulation Results
In the following, we evaluate the performance of the proposed algorithm for noise reduction network with 20 nodes. The topology
of the network is shown in Fig. 1. The unknown vector of interest is a lowpass filter of order M=5, whose coefficients and
normalized frequency response are shown in Fig. 2(a-b). The regressors are zero-mean white Gaussian distributed with covariance
matrices Ru , k = σ uk2 I , with σ uk2 shown in Fig. 2(c). The performance of the different algorithms is measured in terms of the network
N
mean square deviation (MSD) MSD = 10 log10 (∑ || wl − wo || / N ) . All network MSD curves are obtained by ensemble averaging
l =1
over 50 independent trials.
5
Fig. 1. Network topology.
5.1 Example 1: Gaussian input
In this example, the Gaussian signal with zero mean and unit variance is employed as the input signal, al , k = 1/ nk is used for
{al , k } , and cl , k = 1 / nk (uniform, [16]), where nk is the degree of node k. The signal-to-noise ratio (SNR) of each node is SNR=0dB,
Amplitude
(dB)
Amplitude
corresponding to the high background noise power.
0.5
-0.5
0.064
-0.008
0
1
1.5
0.064
0.443
2
2.5
3
3.5
(a)
-0.008
4
4.5
5
Tap
-5
(b)
-10
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
Regressor
variances
Normalized Frequency (×π rad/sample)
1
0.8
0.6
0.4
0.2
0
(c)
2
4
6
8
10
12
14
16
18
20
Node number
Fig. 2. (a) Coefficients of the filter. (b) Frequency responses of the unknown system. (c) Regressor variances.
0
baseline
MSD(dB)
-5
-10
-15
node 7
node 20
node 5
CTA leaky dLMS
ATC leaky dLMS
-20
0.1
0.2
0.3
0.4
0.5
0.6
0.7
Step size
Fig. 3. Network MSD of the proposed algorithms versus different step sizes
6
-6
CTA leaky
CTA leaky
CTA leaky
ATC leaky
ATC leaky
ATC leaky
-8
MSD(dB)
-10
dLMS(γ=0.001)
dLMS(γ=0.002)
dLMS(γ=0.1)
dLMS(γ=0.001)
dLMS(γ=0.002)
dLMS(γ=0.1)
-12
-14
-16
-18
-20
0
200
400
600
800
1000
Number of iterations
Fig. 4. Network MSD of the proposed algorithms versus different γ (µ=0.08).
Fig. 3 shows the steady-state network MSD of the leaky dLMS with different step-size. As one can see, the steady-state MSDs of
the proposed algorithms increase as the step sizes increase. For µ>0.668 (node 5), the CTA leaky dLMS fails to work. Besides, by
using (31), we can obtain the theory upper bounds of step sizes from different nodes. The calculating data indicate that, the upper
bound of step size from node 1 is 1.1 (maximum upper bound among 20 nodes), and upper bound of step size from node 7 is about
0.4 (minimum upper bound among 20 nodes). To guarantee all the nodes work well, µ must be µ<0.4 to ensure that the stability. As
a result, we select µ=0.08 in the following simulations. Then, we investigate the performance of the algorithm in different γ, as
shown in Fig. 4. It can be easily observed that the CTA/ATC leaky dLMS algorithm is not sensitive to this choice, but it turns out that
the best option is γ=0.002. Fig. 5 displays a comparison of MSE from the proposed algorithm and existing algorithms. One can see
that the proposed algorithm outperforms the conventional dLMS algorithms in terms of steady-state error. After the about iteration
50, all the algorithm reach steady state. In addition, it is obviously shown from these curves that the proposed algorithms reach low
misadjustment (−14dB and −18dB) in about iteration 60, whereas the CAT dLMS and ATC dLMS algorithms finally settle at a
network MSD value of −12dB and −16dB. All the ATC-based algorithms achieve improved performance as compared with
CTA-based algorithms, with the similar initial convergence rate.
7
-6
CTA dLMS(µ=0.1)
CTA leaky dLMS(µ=0.08,γ=0.002)
-8
ATC dLMS(µ=0.1)
ATC leaky dLMS(µ=0.08,γ=0.002)
MSD(dB)
-10
-12
-14
-16
-18
-20
0
100
200
300
400
500
600
700
800
900
1000
Number of iterations
Fig. 5. Network MSD learning curves for Gaussian input.
5.2 Example 2: Speech input signal
In the second example, the input signal is a speech input (See Fig. 6). In piratical application, different nodes have different
speech signals. Therefore, we consider a input signal generated as u u , k = σ uk2 i speech input . The environment and parameters of all
the algorithms are same as those of example 1.
In Fig. 7, it demonstrates a comparison of network MSD performance of different algorithms. In this cases, superior performance
of the proposed algorithms are obtained because that the leaky method can stabilize the system [1]. The proposed algorithm, which
adopts the ATC and CTA approaches, reduces the misadjustment when the speech is selected as the input signal. Meanwhile, the fast
convergence rate is obtained. Furthermore, it can be obviously observed from this figure that the misadjustment for the ATC version
is less than that of CTA version. To further demonstrate the performance of proposed algorithm, Fig. 8 depicts the waveforms of the
simulation result at node 14 ( σ uk2 = 0.35 ). The ATC leaky dLMS algorithm, inherits the advantages of the leaky algorithm, and
achieves a stability performance in the presence of strong disturbances.
5
Amplitude
3
1
-1
-3
-5
0
100
200
300
400
500
600
700
Number of iterations
Fig. 6. Input speech signals
800
900
1000
8
20
CTA dLMS(µ=0.1)
CTA leaky dLMS(µ=0.08,γ=0.002)
ATC dLMS(µ=0.1)
ATC leaky dLMS(µ=0.08,γ=0.002)
15
10
MSD(dB)
5
0
-5
-10
-15
-20
0
100
200
300
400
500
600
700
800
900
1000
Number of iterations
Fig. 7. Network MSD learning curves for speech input.
3
(a)
1
Amplitude
-1
-3
0
100
200
300
400
500
600
700
800
900
1000
3
(b)
1
-1
-3
0
100
200
300
400
500
600
700
800
900
1000
Number of iterations
Fig. 8. Waveforms of the simulation result at node 14. (a) Noisy speech (green line: noise signal with SNR=0dB; blue line: clean speech). (b) Filtered speech with
the ATC leaky dLMS algorithm.
Table 1 Computation time
Algorithms
Computation
Time(s)
ATC dLMS
0.6962
ATC leaky dLMS
CTA dLMS
1.0026
1.1945
CTA leaky dLMS
1.5701
To evaluate the computational burden, we have measured the average run execution time of the algorithm on a 2.1-GHz AMD
processor with 8GB of RAM, running Matlab R2013a on Windows 7. The computation time of the algorithms is outlined in Table.
1. One can see that the ATC-dLMS algorithm is the fastest one among these distributed algorithms. The proposed algorithm slightly
increases the execution time, but the final performance is lower than other algorithms. The CTA algorithms achieve the slower than
the ATC algorithms, still with an affordable computation time.
9
5. Conclusion
In this paper, a leaky dLMS algorithm with ATC and CTA strategies has been proposed for distributed estimation. The proposed
algorithm updates the local estimation based on leaky method, which can not only maintain the stability for distributed system, but
also enhance the performance for noise suppression. Simulation results in the context of noise-reduction system show that the
proposed algorithm achieves an improved performance as compared with existing algorithms.
Acknowledgement
This work was partially supported by National Science Foundation of P.R. China (Grant nos: 61271340, 61571374, 61433011).
The first author would also like to acknowledge the China Scholarship Council (CSC) for providing him with financial support to
study abroad (No. 201607000050).
References
[1] K. Mayyas, T. Aboulnasr, Leaky LMS algorithm: MSE analysis for Gaussian data, IEEE Trans. Signal Process. 45 (4) (1997) 927–934.
[2] J. Cheer, S.J. Elliott, Active noise control of a diesel generator in a luxury yacht, Appl. Acoust. 105 (2016) 209–214.
[3] D.B. Bhoyar, C.G. Dethe, M.M. Mushrif, A.P. Narkhede, Leaky least mean square (LLMS) algorithm for channel estimation in
BPSK-QPSK-PSK MIMO-OFDM system, International Multi-Conference on Automation, Computing, Communication, Control and
Compressed Sensing., Kottayam, 2013, pp. 623–629.
[4] J.M. Gil-Cacho, M. Signoretto, T.V. Waterschoot, M. Moonen, S.H. Jensen, Nonlinear acoustic echo cancellation based on a
sliding-window leaky kernel affine projection algorithm, IEEE Trans. Audio Speech Lang. Process. 21 (9) (2013) 1867–1878.
[5] L. Li, J.A. Chambers, C.G. Lopes, A.H. Sayed, Distributed estimation over an adaptive incremental network based on the affine projection
algorithm, IEEE Trans. Signal Process. 58 (1) (2010) 151–164.
[6] Y. Liu, W.K.S. Tang, Enhanced incremental LMS with norm constraints for distributed in-network estimation, Signal Process. 94 (2014)
373–385.
[7] N. Takahashi, I. Yamada, A.H. Sayed, Diffusion least-mean squares with adaptive combiners: Formulation and performance analysis, IEEE
Trans. Signal Process. 58 (9) (2010) 4795–4810.
[8] J. Chen, A.H. Sayed, Diffusion adaptation strategies for distributed optimization and learning over networks, IEEE Trans. Signal Process. 60
(8) (2012) 4289–4305.
[9] R. Abdolee, B. Champagne, A.H. Sayed, Estimation of space-time varying parameters using a diffusion LMS algorithm, IEEE Trans. Signal
Process. 62 (2) (2014) 403–418.
[10] R. Abdolee, B. Champagne, A.H. Sayed, Diffusion adaptation over multi-agent networks with wireless link impairments, IEEE Trans.
Mobile Computing 2015 DOI: 10.1109/TMC.2015.2460251.
10
[11] R. Abdolee, B. Champagne, Diffusion LMS strategies in sensor networks with noisy input data, IEEE/ACM Trans. Networking 24 (1) (2016)
3–14.
[12] F.S. Cattivelli, A.H. Sayed, Diffusion LMS strategies for distributed estimation, IEEE Trans. Signal Process. 58 (3) (2010) 1035–1048.
[13] J. Ni, Diffusion sign subband adaptive filtering algorithm for distributed estimation, IEEE Signal Process. Lett. 22 (11) (2015) 2029–2033.
[14] C. Li, P. Shen, Y. Liu, Z. Zhang, Diffusion information theoretic learning for distributed estimation over network, IEEE Trans. Signal
Process. 61 (16) (2013) 4011–4024.
[15] M. Sowjanya, A.K. Sahoo, S. Kumar, Distributed incremental Leaky LMS, International Conference on Communications and Signal
Processing (ICCSP), Melmaruvathur, 2015, pp. 1753–1757.
[16] V.D. Blondel, J.M. Hendrickx, A. Olshevsky, J.N. Tsitsiklis, Convergence in multiagent coordination, consensus, and flocking, in Proc.
Joint 44th IEEE Conf. Decision Control Eur. Control Conf.(CDC-ECC), Seville, Spain, 2005, pp. 2996–3000
| 3cs.SY
|
Learning Networks from Random Walk-Based Node Similarities
Jeremy G. Hoskins∗
Cameron Musco†
Christopher Musco‡
Charalampos E. Tsourakakis§
arXiv:1801.07386v1 [cs.SI] 23 Jan 2018
January 24, 2018
Abstract
Digital presence in the world of online social media entails significant privacy risks [6, 10,
33, 58, 34]. In this work we consider a privacy threat to a social network in which an attacker
has access to a subset of random walk-based node similarities, such as effective resistances (i.e.,
commute times) or personalized PageRank scores. Using these similarities, the attacker’s goal is
to infer as much information as possible about the underlying network, including any remaining
unknown pairwise node similarities and edges.
For the effective resistance metric, we show that with just a small subset of measurements,
the attacker can learn a large fraction of edges in a social network (and in some cases all edges),
even when the measurements are noisy. We also show that it is possible to learn a graph
which accurately matches the underlying network on all other effective resistances. This second
observation is interesting from a data mining perspective, since it can be expensive to accurately
compute all effective resistances or other random walk-based similarities. As an alternative, our
graphs learned from just a subset of approximate effective resistances can be used as surrogates
in a wide range of applications that use effective resistances to probe graph structure, including
for graph clustering, node centrality evaluation, and anomaly detection.
We obtain our results by formalizing the graph learning objective mathematically, using
two optimization problems. One formulation is convex and can be solved provably in polynomial time. The other is not, but we solve it efficiently with projected gradient and coordinate
descent. We demonstrate the effectiveness of these methods on a number of social networks
obtained from Facebook. We also discuss how our methods can be generalized to other random walk-based similarities, such as personalized PageRank scores. Our code is available at
https://github.com/cnmusco/graph-similarity-learning.
∗
Yale University. [email protected]
Massachusetts Institute of Technology. [email protected]
‡
Massachusetts Institute of Technology. [email protected]
§
Boston University & Harvard University. [email protected]
†
1
1
Introduction
In graph mining and social network science, a variety of measures are used to quantify the similarity
between nodes in a graph, including the shortest path distance, Jaccard’s coefficient between node
neighborhoods, the Adamic-Adar coefficient [2], and hub-authority-based metrics [32, 11]. An important family of similarity measures are based on random walks, including SimRank [25], random
walks with restarts [52], commute times [20], personalized PageRank [42, 26, 7], and DeepWalk embeddings [43]. These measures capture both local and global graph structure and hence are widely
used in graph clustering and community detection [4, 46], anomaly detection [44], collaborative
filtering [20, 47, 57], link prediction [38], and many other applications (including outside of network
science, such as in computer vision [22]).
In this work we focus on these random walk-based similarity metrics. We initiate the study of
a fundamental question:
How much information about a network can be learned given access to a subset of
potentially noisy estimates of pairwise node similarities?
This question is important from a privacy perspective. A common privacy breach is social link
disclosure [58], in which an attacker attempts to learn potentially sensitive links between nodes
in a network. Such attacks are very common; fake accounts with engineered profiles are used to
infiltrate and spy on social groups, potential employers may want to inspect the social network of
a job candidate, and advertisers may wish to probe the demographic and interest information of a
user to offer targeted ads. Thus, characterizing the ability of an attacker to reveal link information
using pairwise node similarities is important in understanding the privacy implications of releasing
such similarities, or information which can be used to compute them.
From a data mining perspective, computing all pairwise node similarities can be infeasible
for large networks since the number of similarities grows quadratically in the number of nodes.
Additionally, when the network cannot be accessed in full but can only be probed via crawling
with random walks (e.g., a by third party studying a social network [30]), we may only have
access to estimates of pairwise similarities rather than their exact values. Thus, understanding
what information can still be learned from a partial, potentially noisy, set of node similarities is
important when using these metrics in large scale graph mining.
1.1
Learning from Effective Resistances
In this paper, we focus on commute times, which are one of the most widely used random walkbased similarities. Commute times are a scaled version of effective resistances, they form a metric,
and have major algorithmic applications, such as spectral graph sparsification [49]. Our ideas can
be extended to related similarity measures, such as personalized PageRank, which we discuss in
Section 4.4. It was shown in the seminal work of Liben-Nowell and Kleinberg [38] that effective
resistances can be used to predict a significant fraction of future links appearing in networks from
existing links, typically ranging from 5% up to 33%.
A difficulty associated with this task is that, in contrast to local similarity measures such as the
number of common neighbors or the Adamic-Adar coefficient [2], node similarity under the effective
resistance metric does not necessarily imply local connectivity. For example, two nodes connected
by many long paths may be more similar than two nodes directly connected by a single edge.
Furthermore, in certain cases, the effective resistance between two nodes u, v tends to correlate
1
1
+ d(v)
) [39, 54, 55], and it
well with a simple function of the degree sequence (specifically, d(u)
2
Figure 1: Grid graph learned from small, randomly selected fractions of effective resistance pairs
using our proposed method. Edge thickness is proportional to edge weight.
is known that there are many graphs with the same degree sequence but very different global
structures. Nevertheless, considered in aggregate, effective resistances encode global structure in a
very strong way. For any graph, given all pairs effective resistances, it is possible to provably recover
the full graph in polynomial time [48, 56]! This contrasts with purely local similarity metrics, which
can be used heuristically for link prediction, but do not give network reconstruction in general. For
instance, all-pairwise counts of common neighbors in any triangle free graph equal 0. Thus, these
counts reveal no information about graph structure.
While the full information case is well understood, when all exact effective resistances are not
available, little is known about what graph information can be learned. Some work considers
reconstruction of trees based on a subset of effective resistances [17, 8, 50]. However outside of this
special case, essentially nothing is known.
1.2
Our Contributions
We study in depth what can be learned about a graph given a subset of potentially noisy effective
resistance estimates, from both a theoretical and empirical viewpoint. Our main contributions are:
Mathematical formulation. We provide an optimization-based formulation of the problem of
learning a graph from effective resistances. Specifically, given a set of effective resistance measurements, we consider the problem of finding a graph whose effective resistances match the given
resistances as closely as possible.
In general, there may be many different graphs which match any subset of all pairs effective
resistances, and hence many minimizers to our optimization problem. If the resistances additionally
have some noise, there may be no graph which matches them exactly but many which match them
approximately. Nevertheless, as we show empirically, the graph obtained via our optimization
approach typically recovers significant information about the underlying graph, including a large
fraction of its edges, its global structure, and good approximations to all of its effective resistances.
Algorithms. We prove that, in some cases, the optimization problem we present can be solved
exactly, in polynomial time. However, in general, the problem is non-convex and does not admit
an obvious polynomial time solution. We give expressions for its gradient and Hessian, and show
that it can be solved efficiently via iterative methods. In particular, we employ projected gradient
and coordinate descent, as well as a powerful initialization strategy which allows us to find high
quality solutions in most instances.
We also show that the problem can be relaxed to a convex formulation. Instead of searching for a
3
graph that matches all given effective resistance measurements, we just find a graph whose effective
resistances are upper bounded by those given and which has minimum total edge weight. This
modified problem is convex and can be solved in polynomial time via semidefinite programming.
Experimental Results. We evaluate our algorithms on several synthetic graphs and real Facebook
ego networks, which contain all nodes in the social circle of a user. We demonstrate that given
a small randomly selected fraction of all effective resistance pairs (10%-25%) we can learn a large
fraction of a network – typically between 20% and 60% of edges. We show that this recovery is
robust to adding Gaussian noise to the given effective resistances. A visual sample of our results
is included in Figure 1 for a simple grid graph. We observe that using a small fraction of noisy
pairwise similarities we can extract a lot of structural information about the graph. Here, noise
is distributed for each edge independently as a Gaussian random variable with zero mean, and
variance σ 2 = 0.1. We observe that as the size of the fraction of revealed similarities increases, an
attacker can recover most of the graph.
We also demonstrate that by finding a graph which closely matches the given set of effective
resistances (via our optimization approach), we in fact find a graph which closely matches the
underlying network on all effective resistance pairs. This indicates that a significant amount of the
information contained in all pairs effective resistances can be learned from just a small subset of
these pairs, even when they are corrupted by noise.
2
Related Work
Link prediction and privacy in social networks. The link prediction problem was popularized
by Liben-Nowell and Kleinberg [38]. The goal of link prediction is to infer which edges are likely to
appear in the near-future, given a snapshot of the network. In [38] various node similarity measures
are used to predict a non-trivial fraction of future interactions. Other works focus on predicting
positive or negative links [35, 53]. For an extensive survey, see [3].
While closely related, link prediction differs from the problem we consider since, typically, one
is given full access to a network snapshot (and in particular, could compute exact pairwise node
similarities for this network) and seeks to make predictions about future evolutions of the network.
In our setting, we are given partial information about a network (via a partial set of noisy similarity
measures) and seek to learn existing links,
While link prediction is useful in applications ranging from understanding network evolution,
to link recommendation, and predicting interactions between terrorists, it entails privacy risks.
For instance, a malicious attacker can use link prediction to disclose information about sensitive
connections in a sexual contact graph [33, 58]. Private link prediction is not adequately explored
in the literature. Abebe and Nakos suggest a possible formalization [1].
Learning graphs. Learning graphs from data is central to many disciplines including machine
learning [27], network tomography [13], bioinformatics [18], and phylogenetics [19]. The general
framework is that there exists a hidden graph that we wish to discover by exploiting some kind of
data, e.g., answers from a blackbox oracle returning certain (possibly noisy) graph measurements.
Theoretical work in this area has focused on worst case query complexity. Two representative
examples include Angluin’s et al. work on learning graphs using edge detecting queries [5], and the
recent work of Kannan, Mathieu, and Zhou using distance queries [28]. A number of works also
consider learning graph parameters such as node count and mixing time by examining random walks
4
traces [30, 16, 9]. Experimental work has focused on recovering graphs from noisy measurements
such as GPS traces [15], and distances between cell populations based on genetic differences [19, 18].
Learning trees. While a tree is special type of a graph, the case of learning trees deserves
special mention since significant work on learning graphs from data has focused on trees. Distancebased reconstruction of
trees aims to reconstruct a phylogenetic tree whose leaves correspond to n
species, given their n2 pairwise distances [19]. Note that on a tree, pairwise distances are identical
to pairwise
effective resistanes. Batagelj et al. study tree realizability assuming access to fewer
n
than 2 leaves’ pairwise distances [8]. A spectral method has been proposed by Stone and Griffing
[50]. Culberson and Rudnicki [17] consider the problem of reconstructing a degree restricted tree
given its distance matrix, see also [45].
3
3.1
Proposed Method
Notation and Preliminaries
For an undirected, weighted graph G = (V, E, w) with n nodes, we let A be the n × n adjacency
matrix. L denotes the (unnormalized) graph Laplacian: L = D − A, where D is a diagonal
matrix with Di,i equal to the weighted degree of node i. For an integer n > 0, [n] denotes the set
{1, 2, ..., n}. ei denotes the ith standard basis vector. For a matrix M , Mi,j denotes the entry in its
ith row and j th column.
Commute time and effective resistance. For two nodes u, v ∈ V , the hitting time hG (u, v)
is the expected time it takes a random walk to travel from u to v. The commute time is its
symmetrized version cG (u, v) = hG (u, v) + hG (v, u), i.e., the time to move from u to v and then
back to u. For connected graphs, the effective
P resistance between u, v is a scaling of the commute
cG (u,v)
time: rG (u, v) = vol(G) where vol(G) = 2 e∈E we . Effective resistance has a natural electrical
interpretation. When G is viewed as an electrical network on n nodes where each edge e corresponds
to a link of conductance we (equivalently to a resistor of resistance w1e ), the effective resistance is the
voltage difference that appears across u, v when a unit current source is applied to them. Effective
resistances (and hence commute times) always form a metric [31].
Letting χu,v = eu − ev , the effective resistance between nodes u and v in a graph G with
Laplacian matrix L can be computed as:
rG (u, v) = χTu,v L+ χu,v .
(1)
Here L+ denotes the Moore-Penrose pseudoinverse of L.
3.2
Problem Definition
We begin by providing a mathematic formulation of the problem introduced in Section 1 – that of
learning the structure of a graph from partial and possibly noisy measurements of pairwise effective
resistances. An analogous problem can be defined for other random walk-based similarities, such
as personalized PageRank. We discuss initial results in this direction in Section 4.4.
5
Problem 1 (Graph Reconstruction From Effective Resistances). Reconstruct an unknown
graph G given a set of noisy effective resistance measurements,
r̄(u, v) = rG (u, v) + nuv
for each (u, v) ∈ S, where S ⊆ [n] × [n] is a set of node pairs and nuv is a potentially random
noise term.
We focus on three interesting cases of Problem 1:
Problem 1.1 S = [n] × [n] and nuv = 0 for all (u, v) ∈ S. This is the full information recovery
problem.
Problem 1.2 S is a subset of [n] × [n] and nuv = 0 for all (u, v) ∈ S. In this setting we must learn
G from a limited number of exact effective resistances.
Problem 1.3 S is a subset of [n] × [n] and nuv is a random term, e.g. a mean 0 normal random
variable with variance σ 2 : nuv ∼ N (0, σ 2 ).
It is known that there in a unique graph consistent with any full set of effective resistance measurements (see e.g., [48] or the related problem in [56]). Additionally, this graph can be computed
by solving a fully determined linear system. So, we can solve Problem 1.1 exactly in polynomial
time. We illustrate this in Section 4.1.
From a privacy and data mining perspective, the limited information settings of Problems 1.2
and 1.3 are more interesting. In Section 4.1 we demonstrate that, when G is a tree, exact recovery
is possible for Problem 1.2 whenever S is a superset of G’s edges. However, in general, there is no
simple closed form solution to these problems, and exact recovery of G is typically impossible. In
particular, several graphs may be consistent with the measurements given. Thus, we address these
cases by reposing Problem 1 as an optimization problem, in which we attempt to recover a graph
matching the given effective resistances as best as possible.
3.3
Optimization Formulation
A natural formalization of Problem 1 is as a least squares problem.
Problem 2. Given a set of vertex pairs S ⊆ [n] × [n] and a target effective resistance r̄(u, v)
for each (u, v) ∈ S:
def
minimize F (H) =
graph H
X
[rH (u, v) − r̄(u, v)]2 .
(2)
(u,v)∈S
Using formula (1) for effective resistances, Problem 2 can equivalently be viewed as an opti2
P
mization problem over the set of graph Laplacians: we minimize (u,v)∈S χTu,v L+ χu,v − r̄(u, v) .
While this set is convex, the objective function is not and thus it is unclear if it can be minimized
provably in polynomial time. Nevertheless, we demonstrate that it is possible to solve the problem
approximately using iterative methods. In Section 4.2 we derive projected gradient and coordinate
6
descent algorithms for the problem. Combined with a powerful initialization heuristic, our experiments show that these methods quickly converge to near global minimums of the objective function
for many networks.
For Problem 1.2, where r̄(u, v) comprise a subset of the exact effective resistances for some
graph G, minH F (H) = 0. This minimum may be achieved by multiple graphs (including G itself)
if S does not contain all effective resistance pairs. Nevertheless, we demonstrate experimentally in
Section 5 that even when S contains a small fraction of these pairs, an approximate solution to
Problem 2 often recovers significant information about G, including a large fraction of its edges.
Interestingly, we find that, while Problem 2 only minimizes over the subset S, the recovered graph
typically matches G on all effective resistances, explaining why it contains so much structural
information about G. For Problem 1.3, if S = [n] × [n] and the noise terms nuv are distributed as
i.i.d. Gaussians, it is not hard to see that Problem 2 gives the maximum likelihood estimator for
G. We again show that an approximate solution can recover a large fraction of G’s edges.
We note that while we can solve Problem 2 quickly via iterative methods, we leave open provable
polynomial time algorithms for solving this problem in the settings of both Problems 1.2 and 1.3.
Convex relaxation. As an alternative to Problem 2, we introduce an optimization formulation of
Problem 1 that is convex. It is convenient here to optimize over the convex set of graph Laplacians.
Problem 3. Let L be the convex set of n × n graph Laplacians. Given a set of vertex pairs
S ⊆ [n] × [n] and a target effective resistance r̄(u,v) for every (u, v) ∈ S,
minimize
Tr(L)
L∈L
subject to
χTu,v L+ χu,v ≤ r̄(u, v)
∀(u, v) ∈ S
Observe that we can trivially find a feasible solution for Problem 3 by setting L to be a large
complete graph, in which case all effective resistances will be close to 0. By Rayleigh’s monotonicity
law, decreasing the weight on edges in L increases effective resistances. Tr(L) is equal to the total
degree of the graph corresponding to L, so the problem asks us to find a graph with as little total
edge weight as possible that still satisfies the effective resistance constraints.
The disadvantage of this formulation is that it only encodes the target resistances r̄(u, v) as
upper bounds on the resistances of L. The advantage is that we can solve Problem 3 provably in
polynomial time via semidefinite programming (see Section 4.3). In practice, we find that it can
sometimes effectively learn graph edges and structure from limited measurements.
Problem 3 is related to work on convex methods for minimizing total effective resistance or
relatedly, mixing time in graphs [12, 51, 21, 21]. However, prior work does not consider pairwise
resistance constraints and so is not suited to the graph learning problem.
4
4.1
Analytical Results and Algorithms
Full Graph Reconstruction – Problem 1
Problem 1 can be solved exactly in polynomial time when S contains all resistance pairs of some
graph G (i.e. Problem 1.1). In this case, there is a closed form solution for G’s Laplacian L and the
solution is unique. This was pointed out in [48], however we include our own proof for completeness.
7
Theorem 1. If there is a feasible solution to Problem 1.1 then it is unique and can be found in
O(n3 ) time. Specifically, the Laplacian L of the unique solution G is given by
J +
J
(3)
R I−
−2 · I −
n
n
where R is the matrix with Ru,v = rG (u, v) for all u, v ∈ [n], I is the n × n identity matrix , and J
is the n × n all ones matrix.
Proof. For an n node graph, the number of possible edges is m = n2 . Let B ∈ Rm×n be the vertex
edge incidence matrix of the complete graph with a row equal to χu,v = eu − ev for every (u, v).
The Laplacian of any n vertex graph can be written as L = B T W B, for some W ∈ Rm×m which
is a nonnegative diagonal matrix with entries corresponding to the edge weights.
We can rewrite the effective resistance formula in (1) as:
+
+
+
+
rL (u, v) = χ⊤
u,v L χu,v = (L )u,u + (L )v,v − 2(L )u,v .
(4)
Since L+ is symmetric we need only determine n(n+1)
unknown entries to determine the full
2
matrix. Moreover, since the all ones vector is in the null space of L and therefore L+ , we see that:
X
L+
L+
(5)
u,u = −
u,v ,
v6=u
and hence we can rewrite (4) as:
ru,v = −
X
v′ 6=v,u
X
(L+ )u,v′ −
(L+ )u′ ,v − 4(L+ )u,v .
(6)
u′ 6=u,v
Let M be the m × m matrix with rows and columns indexed by pairs u, v ∈ [n] with u 6= v and
(u1 , v1 ), (u2 , v2 ) entry given by:
−4 if u1 = u2 and v1 = v2
M(u1 ,v1 ),(u2 ,v2 ) = −1 if u1 = u2 or v1 = v2
0 otherwise.
Let r ∈ Rm contain each effective resistance rG (u, v). We can see from (6) that if we solve the
linear system M x = −r, as long as M is full rank and so the solution is unique, the entries of x will
give us each (L+ )u,v with u 6= v. We can then use these entries to recover the remaining diagonal
entries of L+ using (5).
We can verify that M is in fact always full rank by writing M = −|B||B|T − 2I, where |B|
denotes the matrix formed from B by taking the absolute value of each of its entries. We note
that the non-zero eigenvalues of |B||B|T are equal to the non-zero eigenvalues of |B T ||B| which
is the n × n matrix M = (n − 2)I + J. M has eigenvalues 2n − 2 with multiplicity 1 and n − 2
with multiplicity n − 1. The remaining m − n eigenvalues of |B||B|T are zero. Consequently, the
eigenvalues of M are −2n with multiplicity 1, −n with multiplicity n − 1 and −2 with multiplicity
m − n. Thus M is full rank, proving that the effective resistances fully determine L+ and thus L.
Solving for L via the linear system M x = −r would require O(n6 ) time, however, the closed
form solution (3) given in Lemma 9.4.1 of [48] allows us to solve this problem in O(n3 ) time.
8
Reconstruction from hitting times. The above immediately generalizes to graph reconstruction
from hitting times since, as discussed, for connected G, the effective resistance between u, v can
G (v,u)
G (u,v)
= hG (u,v)+h
. Thus, by Theorem 1, we can recover G up to a
be written as rG (u, v) = cvol(G)
vol(G)
scaling from all pairs hitting times. This recovers a result in [56]. Note that if we scale all edge
weights in G by a fixed factor, the hitting times do not change. Thus recovery up to a scaling is
the best we can hope for in this setting.
Reconstruction from other similarity measures. An analogous result to Theorem 1 holds for
graph recovery from all pairs personalized PageRank scores, and for related measures such as Katz
similarity scores [29]. We discuss this direction in Section 4.4.
Are all pairs always necessary for perfect reconstruction?
For general graphs, Problem
1 can only be solved exactly when S contains all n2 true effective resistances. However, given
additional constraints on G, recovery is possible with much less information. In particular, when
G is a tree, we can recover it (i.e., solve Problem 1.2) whenever S is a superset of its edge set.
Roughly, since G is a tree, the effective resistance rG (u, v) is equal to the length of the unique
path connecting u and v. As long as S includes all edges in G, it fully determines all path lengths
and hence the effective resistances for all pairs u, v. We can thus recover G via Theorem 1. Formally:
Theorem 2. If G is a tree and a feasible solution to Problem 1.2 with edge set E ⊆ S then G is
unique and can be found in O(n3 ) time.
Proof. Let Puv be the unique path between u, v in G. It is well known [14] that:
X
1/we .
rG (u, v) =
(7)
e∈Puv
For (u, v) ∈ S set r̄(u, v) = rG (u, v). Let G be an undirected graph with an edge for each (u, v) ∈ S
with length rG (u, v). For all (u, v) ∈
/ S, set r̄(u, v) to the shortest path distance between u and v
in G.
Claim. r̄(u, v) = rG (u, v) for all (u, v) ∈ [n] × [n].
For any pair (u, v), we have r̄(u, v) ≤ rG (u, v). The length of shortest path between u, v in G is
certainly
at mostPthe length of Puv , which is contained P
in G since E ⊆ S. Puv ’s length in G is:
P
e∈Puv 1/we = rG (u, v), again using (7).
e∈Puv 1/we using (7). Thus, r̄(u, v) ≤
e∈Puv rG (e) =
Further, Puv is in fact a shortest path between u, v in G, giving that r̄(u, v) = rG (u, v). This is
because the length of every edge (u, v) ∈ S that is not in E just equals the length of path Puv in G
(i.e., rG (u, v)) and so removing this edge from G does not change any shortest path distance. So
we can assume that G just contains the edges in E, and so Puv is the unique path between u, v.
Given the above claim, the theorem follows since we can compute each r̄(u, v) from the effective
resistances of the edges in S and can then compute G from these resistances by Theorem 1.
The problem of recovering trees from pairwise distance measurements is a central problem in
phylogenetics. There are other cases when just a subset of effective resistances is known to allow
full recovery, for example when the effective resistances between any pair of leaves is known [50, 48].
Also related to our analysis for trees lies is the work of Mathieu and Zhou [40], which reconstructs
graphs with bounded degree from pairwise distance measurements.
9
4.2
Graph Learning via Least Squares Minimization – Problem 2
When Problem 1 cannot be solved exactly, e.g. in the settings of Problems 1.2 and 1.3, an effective
surrogate is to solve Problem 2 to find a graph with effective resistances close to the given target
resistances. As we demonstrate experimentally in Section 5, this yields good solutions to Problems
1.2 and 1.3 in many cases. Problem 2 is non-convex, however we show that a good solution can
often be found efficiently via projected gradient descent.
Optimizing over edge weights. Let m = n2 . We write the Laplacian of the graph H as
def
L(w) = B T diag(w)B, where w ∈ Rm is a non-negative vector whose entries correspond to the edge
weights in H, diag(w) is the m × m matrix with w as its diagonal, and B ∈ Rm×n is the vertex
edge incidence matrix with a row equal to χu,v = eu − ev for every possible edge (u, v) ∈ [n] × [n].
Optimizing the objective function F (H) in Problem 2 is equivalent to optimizing F (w) over
def
the edge weight vector w, where we define F (w) = F (H) for the unique H with Laplacian equal
to L(w).
We restrict wi ≥ 0 for all i and project to this constraint after each gradient step simply by
setting wi := max(wi , 0). The gradient of F (w) can be computed in closed form. We first define
an auxiliary variable, R(w) ∈ Rm×m , whose diagonal contains all pairwise effective resistances of
H with weight vector w:
Definition 1. For w ∈ Rm with wi ≥ 0 for all i, define
R(w) = BL(w)+ B T .
Using R(w) we can compute the gradient of F (w) by:
Proposition 1. Let ◦ denote the Hadamard (entrywise) product for matrices. Define the error
vector ∆(w) ∈ Rm as having ∆(w)i = r̄(i) − [R(w)]i,i for all i ∈ S and 0s elsewhere. We have:
∇F (w) = 2 (R ◦ R) ∆(w)
Proof. We begin by observing that, letting ei be the ith standard basis vector in Rm , for any weight
vector w, the graph Laplacian corresponding to the weight vector w + ǫei , is L(w) + ǫbi bTi . The
Sherman-Morrison formula for the matrix pseudoinverse yields:
(L(w) + ǫbi bTi )+ = L(w)+ − ǫ
L(w)+ bi bTi L(w)+
,
1 + ǫbTi L(w)+ bi
and, hence thinking of L(w)+ as a matrix-valued function of w,
∂L(w)+
1
L(w)+ bi bTi L(w)+
+
+
− L(w)
= lim
L(w) − ǫ
ǫ→0 ǫ
∂wi
1 + ǫbTi L(w)+ bi
= −L(w)+ bi bTi L(w)+ .
Let Ri denote the ith column of R(w) = BL(w)+ B T . By linearity:
∂R
= −B L(w)+ bi bTi L(w)+ B T = −Ri RiT .
∂wi
10
Thus,
X
∂F
=2
(r̄(j) − [R(w)]j,j ) · [R(w)]2i,j ,
∂wi
j∈S
and so we obtain that the gradient equals ∇F (w) = 2 (R ◦ R) ∆(w).
While gradient descent works well in our experiments, one may also apply second order methods,
which require F (w)’s Hessian. Using similar computations to those in Proposition 1 we obtain:
Proposition 2. Let IS ∈ Rm×m be the diagonal matrix with a 1 at each entry corresponding to
i ∈ S and 0s elsewhere and ∆(w) be as defined in Proposition 1. The Hessian matrix of F (w) is:
HF (w) = −4 [R diag(∆(w))R] ◦ R + 2(R ◦ R)IS (R ◦ R).
Acceleration via coordinate descent. Naively computing the gradient ∇F (w) via Proposition 1
requires computing the full m × m matrix R(w), which can be prohibitively expensive for large
graphs – recall that m = n2 = O(n2 ). Note however, that the error vector ∆(w) only has nonzero
entries at positions corresponding to the node pairs in S. Thus, it suffices to compute just |S|
columns of R corresponding to these pairs, which can give a significant savings.
We obtain further savings using block coordinate descent. At each step we restrict our updates
to a subset of edges B ⊆ [n] × [n]. Let IB be the matrix with a 1 at diagonal entries corresponding
to elements of B and 0’s elsewhere. We step in the direction of IB ∇F (w). Computing this step
only requires forming the rows of R corresponding to edges in B. A typical way to choose B is at
random. See Section 5.1 for the actual implementation details.
Initialization. A good initialization for gradient descent can significantly accelerate the solution
of Problem 2. We use a strategy based on the exact solution to Problem 1.1 in Theorem 1.
Since effective resistances form a metric, by triangle inequality, for any u, v, w ∈ [n], rH (u, v) ≤
rH (u, w) + rH (w, v). Guided by this fact, given target resistances r̄(u, v) for (u, v) ∈ S, we first
“fill in” the constraint set. For (w, z) ∈
/ S, we set r̄(w, z) equal to the shortest path distance in the
graph Ḡ which has an edge for each pair in S with length r̄(u, v).
We thus obtain a full set of target effective resistances. We can form R with Ru,v = r̄(u, v) and
initialize the Laplacian of H using the formula given in (3) in Theorem 1. However, this formula is
quite unstable and generally yields an output which is far from a graph Laplacian even when R is
corrupted by a small amount of noise. So we instead compute a regularized estimate,
+
J
J
R I−
+ λI ,
L̃ = −2 · I −
n
n
where λ > 0 can be chosen e.g. by line search. Generally, L̃ will not be a valid graph Laplacian,
but by removing negative edge weights, we typically obtain a good initialization for Problem 2.
4.3
Graph Learning via Convex Optimization – Problem 3
We finally discuss how to efficiently solve our convex formulation, Problem 3. We express this
problem as a semidefinite program (SDP) which can be solved via a number of available packages.
We can re-express our effective resistance constraint as a positive semidefinite constraint using
the Schur complement condition:
L
χu,v
T
+
χu,v L χu,v ≤ r̄(u, v) iff
0 and L 0.
χTu,v r̄(u, v)
11
Doing so yields the following program:
Problem 4 (SDP Form of Problem 3). Given vertex pairs S ⊆ [n] × [n], and target effective
resistance r̄(u, v) for every (u, v) ∈ S,
minimize Tr(L)
L∈L
subject to
L 0 and ∀ (u, v) ∈ S,
L
χu,v
0
χTu,v r̄(u, v)
We require L to be a valid graph Laplacian (i.e., constrain L ∈ L) by adding linear constraints of
the form:
X
∀ i, Li,i = −
Li,j and ∀ i 6= j, Li,j ≤ 0.
j6=i
4.4
Extensions to Other Similarity Measures
As discussed, our results generalize to random walk-based node similarities beyond effective resistances, such as personalized PageRank scores. Given localization parameter α ≥ 0, the personalized
PageRank score pαG (u, v) is the probability that a lazy random walk on graph G which jumps back
to u with probability α in each step is at node v in its stationary distribution [42, 24, 4].
Letting W = 12 (I + AD −1 ) be the lazy random walk matrix, pαG (u, v) is the v th entry of the
personalized PageRank vector:
pαG (u) = α(I − (1 − α)W )−1 eu .
(8)
This vector gives the stationary distribution for the random walk on G and thus the personalized
PageRank pαG (u, v) is its v th entry.
It is not hard to show an analogous result to Theorem 1, that given a full set of exact personalized
PageRank scores, full recovery of G is possible. Roughly, if we let P be the matrix with pαG (u) as
its uth column, we have P = α(I − (1 − α)W )−1 and can thus solve for the random walk matrix
W , and the graph G. This gives:
Theorem 3. For any connected graph G, given personalized PageRank score pαG (u, v) for each
(u, v) ∈ [n] × [n], there is algorithm returning G (up to a scaling of its edge weights) in O(n3 ) time.
Further, it is possible to formulate a problem analogous to Problem 2 and solve for a graph
matching a subset of personalized PageRank measurements as closely as possible. As shown in
Figure 2, personalized PageRank often gives a stronger signal of global graph structure than effective
resistance. To create the plot, nodes are sorted by their value in the Laplacian Fielder vector, which
corresponds roughly to residence in different clusters. In an extended version of our work, we will
provide detailed empirical results with personalized PageRank, and other random walk measures.
5
Empirical results
In this section we present an in depth experimental study of how well our methods can learn a graph
given a set of (noisy) effective resistance measurements. We seek to answer two key questions:
12
Figure 2: Personalized PageRank correlates better than commute times with the cluster structure
in the FB Small C network (see Table 1). Heatmaps are shown in log scale.
1. Given a set of effective resistance measurements, can we find a graph matching these measurements via the optimization formulations posed in Problems 2 and 3 and the corresponding
algorithms discussed in Sections 4.2 and 4.3?
2. What structure does the graph we learn via our optimization approach share with the underlying network that produced the resistance measurements? Can it be used to predict links in
the network? Does it approximately match the network on effective resistances outside the
given set of measurements, or share other global structure?
We address these questions by examining a variety of synthetic, and social network graphs.
5.1
Experimental Setup
Table 1 lists the networks analyzed in our experiments. These include two synthetic examples: an
8 × 8 two dimensional grid graph and a k-nearest neighbor graph constructed for vectors drawn
from a Gaussian mixture model with two clusters. The other networks are Facebook ‘ego networks’
obtained from the Stanford Network Analysis Project (SNAP) collection [36, 37]. Each of these
networks is formed by taking the largest connected component in the social circle of a specific user
(whose nodeId is shown in Table 1).
For all experiments, we provide
our algorithms with effective resistances that are uniformly
def
× 100% of
sampled from the set of all n2 effective resistances. We sample a fixed fraction f = |S|
(n2 )
all possible measurements. We typically use f ∈ {10, 25, 50, 100}%. In some cases, these resistances
are corrupted with i.i.d. Gaussian noise η ∼ N (0, σ 2 ). We experiment with different values of
variance σ 2 .
For Problem 2 we implemented gradient decent based on the closed-from gradient calculation
in Section 4.2. Line search was used to optimize step size at each iteration since it significantly
outperformed implementations with fixed step sizes. For larger problems, block coordinate descent
was used as described in Section 4.2, with the coordinate set chosen uniformly at random in each
iteration. We set the block size |B| = 5000. For Problem 3 we used MOSEK convex optimization
software, accessed through the CVX interface [41, 23]. All experiments were run on a computer
with a 2.6 GHz Intel Core i7 processor and 16 GB of main memory.
13
Name
Grid (synthetic)
k-nn (synthetic)
FB Small A (nodeId 698)
FB Small B (nodeId 3980)
FB Small C (nodeId 414)
FB Small D (nodeId 686)
FB Medium A (nodeId 348)
FB Medium B (nodeId 0)
FB Large A (nodeId 3437)
FB Large B (nodeId 1912)
# of nodes, n
64
80
40
44
148
168
224
324
532
795
# of edges, m
224
560
220
138
1692
1656
3192
2514
4812
30023
Table 1: Datasets for experiments. FB denotes “Facebook”.
Figure 3: Graphs learned by solving Problem 2 with gradient descent run to convergence for
uniformly sampled effective resistances with varying levels of Gaussian noise. Edge width is proportional to edge weight in the plots.
5.2
Learning Synthetic Graphs
We first evaluate our graph learning algorithms on Grid and k-nn, which are simple synthetic
graphs with clear structure.
Least squares formulation. We first observe that gradient descent effectively minimizes the
objective function of Problem 2 on the Grid and k-nn graphs. We consider the normalized
objective for a constraint set S and output graph H:
P
2
(u,v)∈S [rH (u, v) − r̄(u, v)]
P
Fb (H) =
.
(9)
2
(u,v)∈S r̄(u, v)
For noise variance 0, minH Fb(H) = 0 and in Figure 4 we see that for Grid we in fact find H with
Fb(H) ≈ 0 for varying sizes of S. Convergence is notably faster when 100% of effective resistances
are included in S, but otherwise does not correlate strongly with the number of constraints.
14
10
0
10
0
10
-2
10
-2
10
-4
10
-4
10 -6
10 -6
10 -8
10 -8
10 -10
10 -10
10
-12
10
-14
10
0
2000
4000
6000
8000
10000
0
10
-12
10
-14
10
10 -1
10 -2
10 -2
10 -3
10 -3
0
2000
4000
6000
8000
10 -4
10000
2000
4000
6000
8000
10000
0
2000
4000
6000
8000
10000
0
10 -1
10 -4
0
Figure 4: Objective error and generalization error for Problem 2, as defined in (10) for Grid. For
details, see Section 5.2.
In Figure 4 we also plot the generalization error :
P
2
(u,v)∈[n]×[n] [rH (u, v) − rG (u, v)]
P
,
Fgen (H) =
2
(u,v)∈[n]×[n] rG (u, v)
(10)
where rG (u, v) is the true effective resistance, uncorrupted by noise. Fgen (H) measures how well the
graph obtained by solving Problem 2 matches all effective resistances of the original network. We
confirm that generalization decreases with improved objective function performance, indicating that
optimizing Problem 2 effectively extracts network structure from a small set of effective resistances.
We observe that the generalization error is small even when f = 10%, and becomes negligible as
we increase the fraction f of measurements, even in the presence of noise.
We repeat the same experiments with Gaussian noise added to each resistance measurement.
The variance of the noise, σ 2 , is scaled relatively to the mean effective resistance in the graph, i.e.,
we set r̄(u, v) = rG (u, v) + N (0, σ̄ 2 ) where:
X
σ2
σ̄ 2 = n ·
rG (u, v).
(11)
2
(u,v)∈[n]×[n]
While generally minH Fb(H) > 0 when r̄(u, v) is noisy (it is likely that there is no graph consistent
with these noisy measurements), we see that the objective value still decreases steadily with a larger
number of iterations. Generalization error also decreases as desired.
15
10
0
10
10 -1
10
0
10 -1
-2
10
-2
10 -3
10 -3
10 -4
0
10
2000
4000
6000
8000
10
10000
0
-4
10
10 -1
10 -2
10 -2
10 -3
10 -3
0
2000
4000
6000
8000
10 -4
10000
2000
4000
6000
8000
10000
0
2000
4000
6000
8000
10000
0
10 -1
10 -4
0
Figure 5: Objective error and generalization error for Problem 2, as defined in (10) for k-nn.
Objective function error decreases steadily, leading to improved generalization error.
We obtain similar results by applying Problem 2 to a k-nearest neighbor graph with two clear
clusters of data points. Again, it is apparent in Figure 5 that gradient descent converges easily for
a variety of noise levels and constraint sets. Convergence on the least squares objective leads to
improved generalization error.
Figure 3 shows the graphs obtained from solving the problem for varying σ 2 and f . For both
graphs, when σ 2 = 0 and f = 100%, the original network is recovered exactly. Reconstruction
accuracy decreases with increasing noise and a decreasing number of constraints. For Grid, even
with 25% of constraints, nearly full recovery is possible for σ 2 = 0 and recovery of approximately
half of true edges is possible for σ 2 = 0.1. For k-nn, for σ 2 = 0 and σ 2 = 0.1 we observe that
cluster structure is recovered. Detailed quantitative results for both networks are given in Table 3.
Convex formulation. We next evaluate the performance graph learning via the convex relaxation
in Problem 3. In this case, we do not focus on convergence as we solve the problem directly using a
semidefinite programming (SDP) routine. Unlike for Problem 2, solving Problem 3 does not recover
the exact input graph, even in the noiseless all pairs effective resistance case. This is because the
input graph does not necessarily minimize the objective of Problem 3 since there can be other
graphs with smaller total edge weight and lower effective resistances.
However, the learned graphs do capture information about edges in the original: their heaviest
edges typically align with true edges in the target graph. This property is captured in the quantitative results of Table 3. Qualitatively, it is very apparent for Grid: in Figure 6 we mark the 224
16
Grid, 64 nodes
# of constraints Time (min.)
202
238.8
504
761.6
1008
1309.5
2016
2183.7
k-nn, 80 nodes
# of constraints Time (min.)
316
1216.0
790
3673.2
1580
8008.5
3160
16192.1
Table 2: Semidefinite program (SDP) optimization for Problem 3. Runtime is averaged over noise
levels σ 2 = 0, 0.1, 1.
heaviest edges in the learned graph in red and note that this set converges exactly on the grid.
The convex formulation never significantly outperforms the least squares formulation, and significantly underperforms for small constraint sets. Additionally, the semidefinite program scales
poorly with the number of nodes and effective resistance constraints. Sample runtimes are included
in Table 2. Due to these considerations, we use the least squares formulation of Problem 2 instead
of the convex formulation in our experiments on real social networks. However, we believe there is
further opportunity for exploring Problem 3, especially given its provable runtime guarantees.
Figure 6: Graphs learned from solving the convex program in Problem 3 for uniformly sampled
effective resistances from Grid with varying f, σ 2 . Heaviest edges marked in red.
5.3
Learning Social Network Graphs
We conclude by demonstrating the effectiveness of the least squares formulation of Problem 2 in
learning Facebook ego networks from limited effective resistance measurements. We consider three
metrics of performance, shown in Table 3 for a number of networks learned from randomly sampled
subsets of effective resistances, corrupting with varying levels of noise.
17
2
Network
Algorithm
σ
Grid
GD
GD
SDP
SDP
GD
GD
SDP
SDP
GD
GD
GD
GD
GD
GD
GD
GD
GD
CD
CD
CD
CD
CD
0
.1
0
.1
0
.1
0
.1
0
.1
0
.1
0
.1
0
.1
0
0
.1
0
0
.1
k-nn
FB Small A
FB Small B
FB Small C
FB Small D
FB Medium A
FB Medium B
FB Large A
FB Large B
Objective
function error
f = 10% f = 25%
.00001
.00001
.00090
.00514
na
na
na
na
.00001
.00002
.00197
.00447
na
na
na
na
.01345
.00001
.00017
.00204
.00002
.00003
.00032
.00206
.00162
.00166
.00532
.00644
.00335
.00434
.00610
.18384
.00447
.00665
.00224
.01255
.01174
.03182
.00516
.00796
.00524
.00440
.12745
.34646
Effect. resistance
generalization error
f = 10% f = 25%
.06559
.00099
.08129
.01336
.08758
.07422
.09549
.09343
.01122
.00117
.05536
.00709
.09314
.10399
.11899
.14097
.21097
.00984
.07964
.01687
.01515
.00623
.02229
.01291
.00217
.00203
.01542
.00218
.00821
.00830
.00923
.21426
.02910
.01713
.00862
.01471
.01687
.03295
.00682
.00862
.00635
.00580
.14532
.36095
% Edges
learned
baseline
5.56
11.58
28.20
14.59
15.55
11.80
12.78
4.80
3.41
9.51
% Edges
learned
f = 10% f = 25%
20.54
88.39
25.89
50.00
16.07
25.00
12.50
26.79
44.54
72.68
25.96
41.53
27.05
48.36
24.32
39.89
44.54
75.00
53.64
60.00
42.75
48.55
36.23
43.48
57.03
59.51
52.66
57.51
21.92
24.52
21.38
21.20
23.50
25.59
18.97
22.15
17.50
16.03
10.52
12.45
20.26
24.95
19.43
16.97
Table 3: Graph recovery results. All results use a randomly sampled subset of f = 10% or f = 25%
of all effective resistances. For “Algorithm”, GD denotes gradient descent and CD denotes block
coordinate descent with random batches of size 5000. “Noise level, σ 2 ” indicates that the target
resistances were set to r̄(u, v) = rG (u, v) + N (0, σ 2 · meanu,v (rG (u, v))). “% Edges baseline”, is
the edge density of the underlying network, equivalent to the expected edge prediction accuracy
achieved with random guessing.
1. Objective Function Value: the value of the objective function of Problem 2, normalized as
described in (9).
2. Generalization Error: the error in reconstructing the full set of effective resistances of the
true graph, as defined in (10).
3. Edges Learned: the rate of recovery for edges in the true graph. We utilize a standard metric
from the link prediction literature [32]: given underlying graph G with m edges and learned
graph H, we consider the m heaviest edges of H and compute the percentage of G’s edges
contained in this set.
Results. We find that as for the synthetic Grid and k-nn graphs, we can effectively minimize the
objective function of Problem 2 for the Facebook ego networks. Moreover, this minimization leads
to very good generalization error in nearly all cases. i.e., we can effectively learn a graph matching
our input on all effective resistances, even when we consider just a small subset.
For all graphs, we are able to recover a significant fraction of edges, even when just considering
10% or 25% of effective resistance pairs. We obtain the best recovery for small graphs, learning
over half of the true edges in FB Small A and FB Small C. Even for larger graphs, we can
typically recover over 20% of the true edges.
18
Typically, the number of edges learned increases as we increase the number of constraints and
decrease the noise variance. However, occasionally, considering fewer effective resistances in fact
improves learning, possibly because we more effectively solve the underlying optimization problem.
6
Conclusion
In this work, we show that a small subset of noisy effective resistances can be used to learn significant
information about a network, including predicting a large percentage of edges and recovering global
structure, such as accurate approximations to all pairwise effective resistances. From a privacy
standpoint, our results raise major concerns about releasing random walk-based pairwise node
similarity information as it entails significant privacy risk. From a data mining perspective, our
methods can be used for graph mining, even when computing all effective resistances exactly is
infeasible.
Our work leaves a number of future research directions open, including giving a provable polynomial time algorithm for the least squares formulation (Problem 2), extending our work to other
similarity metrics, and scaling our methods to larger social networks.
References
[1] R. Abebe and V. Nakos. Private link prediction in social networks. 2014.
[2] L. A. Adamic and E. Adar. Friends and neighbors on the web. Social Networks, 25(3):211–230,
2003.
[3] M. Al Hasan and M. J. Zaki. A survey of link prediction in social networks. In Social Network
Data Analytics, pages 243–275. Springer, 2011.
[4] R. Andersen, F. Chung, and K. Lang. Local graph partitioning using PageRank vectors. In
47th Annual IEEE Symposium on Foundations of Computer Science, pages 475–486, 2006.
[5] D. Angluin and J. Chen. Learning a hidden graph using o(logn) queries per edge. Journal of
Computer and System Sciences, 74(4):546–556, 2008.
[6] L. Backstrom, C. Dwork, and J. Kleinberg. Wherefore art thou?: Anonymized social networks,
hidden patterns, and structural steganography. In Proceedings of the 16th International Conference on World Wide Web, pages 181–190. ACM, 2007.
[7] B. Bahmani, A. Chowdhury, and A. Goel. Fast incremental and personalized PageRank.
Proceedings of the VLDB Endowment, 4(3):173–184, 2010.
[8] V. Batagelj, T. Pisanski, and J. M. S. S. os Pereira. An algorithm for tree-realizability of
distance matrices. International Journal of Computer Mathematics, 34(3-4):171–176, 1990.
[9] A. Ben-Hamou, R. I. Oliveira, and Y. Peres. Estimating graph parameters via random walks
with restarts. In Proceedings of the ACM-SIAM Symposium on Discrete Algorithms, 2018.
[10] S. Bhagat, G. Cormode, B. Krishnamurthy, and D. Srivastava. Class-based graph anonymization for social network data. Proceedings of the VLDB Endowment, 2(1):766–777, 2009.
19
[11] V. D. Blondel, A. Gajardo, M. Heymans, P. Senellart, and P. Van Dooren. A measure of
similarity between graph vertices: Applications to synonym extraction and web searching.
SIAM Review, 46(4):647–666, 2004.
[12] S. Boyd, P. Diaconis, and L. Xiao. Fastest mixing Markov chain on a graph. SIAM Review,
46(4):667–689, 2004.
[13] R. Castro, M. Coates, G. Liang, R. Nowak, and B. Yu. Network tomography: Recent developments. Statistical Science, pages 499–517, 2004.
[14] A. K. Chandra, P. Raghavan, W. L. Ruzzo, R. Smolensky, and P. Tiwari. The electrical resistance of a graph captures its commute and cover times. Computational Complexity, 6(4):312–
340, 1996.
[15] D. Chen, L. J. Guibas, J. Hershberger, and J. Sun. Road network reconstruction for organizing paths. In Proceedings of the Twenty-First Annual ACM-SIAM Symposium on Discrete
Algorithms, pages 1309–1320. Society for Industrial and Applied Mathematics, 2010.
[16] C. Cooper, T. Radzik, and Y. Siantos. Estimating network parameters using random walks.
Social Network Analysis and Mining, 4(1):168, 2014.
[17] J. C. Culberson and P. Rudnicki. A fast algorithm for constructing trees from distance matrices.
Information Processing Letters, 30(4):215 – 220, 1989.
[18] R. Desper, F. Jiang, O.-P. Kallioniemi, H. Moch, C. H. Papadimitriou, and A. A. Schäffer.
Inferring tree models for oncogenesis from comparative genome hybridization data. Journal of
Computational Biology, 6(1):37–51, 1999.
[19] J. Felsenstein. Confidence limits on phylogenies: an approach using the bootstrap. Evolution,
39(4):783–791, 1985.
[20] F. Fouss, A. Pirotte, J.-M. Renders, and M. Saerens. Random-walk computation of similarities between nodes of a graph with application to collaborative recommendation. IEEE
Transactions on Knowledge and Data Engineering, 19(3):355–369, 2007.
[21] A. Ghosh, S. Boyd, and A. Saberi. Minimizing effective resistance of a graph. SIAM Review,
50(1):37–66, 2008.
[22] L. Grady. Random walks for image segmentation. IEEE Transactions on Pattern Analysis
and Machine Intelligence, 28(11):1768–1783, 2006.
[23] M. Grant and S. Boyd. CVX: Matlab software for disciplined convex programming, version
2.1. http://cvxr.com/cvx, Mar. 2014.
[24] T. H. Haveliwala. Topic-sensitive PageRank: A context-sensitive ranking algorithm for web
search. IEEE Transactions on Knowledge and Data Engineering, 15(4):784–796, 2003.
[25] G. Jeh and J. Widom. Simrank: a measure of structural-context similarity. In Proceedings
of the Eighth ACM SIGKDD International Conference on Knowledge Discovery and Data
Mining, pages 538–543. ACM, 2002.
20
[26] G. Jeh and J. Widom. Scaling personalized web search. In Proceedings of the 12th International
Conference on World Wide Web, pages 271–279. ACM, 2003.
[27] V. Kalofolias. How to learn a graph from smooth signals. In The 19th International Conference on Artificial Intelligence and Statistics (AISTATS 2016). Journal of Machine Learning
Research (JMLR), 2016.
[28] S. Kannan, C. Mathieu, and H. Zhou. Near-linear query complexity for graph inference. In
International Colloquium on Automata, Languages, and Programming. Springer, 2015.
[29] L. Katz. A new status index derived from sociometric analysis. Psychometrika, 18(1):39–43,
1953.
[30] L. Katzir, E. Liberty, and O. Somekh. Estimating sizes of social networks via biased sampling.
In Proceedings of the 20th International Conference on World Wide Web. ACM, 2011.
[31] D. J. Klein and M. Randić. Resistance distance. Journal of Mathematical Chemistry, 12(1):81–
95, 1993.
[32] J. Kleinberg, R. Kumar, P. Raghavan, S. Rajagopalan, and A. Tomkins. The web as a graph:
measurements, models, and methods. Computing and Combinatorics, pages 1–17, 1999.
[33] A. Korolova, R. Motwani, S. U. Nabar, and Y. Xu. Link privacy in social networks. In
Proceedings of the 17th ACM Conference on Information and Knowledge Management, pages
289–298. ACM, 2008.
[34] N. Korula and S. Lattanzi. An efficient reconciliation algorithm for social networks. Proceedings
of the VLDB Endowment, 7(5):377–388, 2014.
[35] J. Leskovec, D. Huttenlocher, and J. Kleinberg. Predicting positive and negative links in
online social networks. In Proceedings of the 19th International Conference on World Wide
Web (WWW), pages 641–650. ACM, 2010.
[36] J. Leskovec and A. Krevl. SNAP Datasets: Stanford large network dataset collection.
http://snap.stanford.edu/data, June 2014.
[37] J. Leskovec and J. J. Mcauley. Learning to discover social circles in ego networks. In Advances
in Neural Information Processing Systems, pages 539–547, 2012.
[38] D. Liben-Nowell and J. Kleinberg. The link-prediction problem for social networks. Journal
of the Association for Information Science and Technology, 58(7):1019–1031, 2007.
[39] U. V. Luxburg, A. Radl, and M. Hein. Getting lost in space: Large sample analysis of the
resistance distance. In Advances in Neural Information Processing Systems, 2010.
[40] C. Mathieu and H. Zhou. Graph reconstruction via distance oracles. In International Colloquium on Automata, Languages, and Programming, pages 733–744. Springer, 2013.
[41] MOSEK ApS. The MOSEK Optimization Suite, 2017.
[42] L. Page, S. Brin, R. Motwani, and T. Winograd. The PageRank citation ranking: Bringing
order to the web. Technical report, Stanford InfoLab, 1999.
21
[43] B. Perozzi, R. Al-Rfou, and S. Skiena. DeepWalk: Online learning of social representations.
In Proceedings of the 20th ACM SIGKDD International Conference on Knowledge Discovery
and Data Mining, pages 701–710. ACM, 2014.
[44] M. J. Rattigan and D. Jensen. The case for anomalous link discovery. ACM SIGKDD Explorations Newsletter, 7(2):41–47, 2005.
[45] L. Reyzin and N. Srivastava. On the longest path algorithm for reconstructing trees from
distance matrices. Information Processing Letters, 101(3):98–100, 2007.
[46] M. Saerens, F. Fouss, L. Yen, and P. Dupont. The principal components analysis of a graph,
and its relationships to spectral clustering. In ECML, volume 3201, pages 371–383, 2004.
[47] P. Sarkar and A. Moore. A tractable approach to finding closest truncated-commute-time
neighbors in large graphs. In Proceedings of the Twenty-Third Conference Annual Conference
on Uncertainty in Artificial Intelligence (UAI-07), pages 335–343, 2007.
[48] D. A. Spielman. Trees and distances. University Lecture, 2012.
[49] D. A. Spielman and N. Srivastava. Graph sparsification by effective resistances. SIAM Journal
on Computing, 40(6):1913–1926, 2011.
[50] E. A. Stone and A. R. Griffing. On the Fiedler vectors of graphs that arise from trees by Schur
complementation of the Laplacian. Linear Algebra and its Applications, 431(10):1869 – 1880,
2009.
[51] J. Sun, S. Boyd, L. Xiao, and P. Diaconis. The fastest mixing Markov process on a graph and
a connection to a maximum variance unfolding problem. SIAM Review, 48(4):681–699, 2006.
[52] H. Tong, C. Faloutsos, and J.-Y. Pan. Fast random walk with restart and its applications.
2006.
[53] C. E. Tsourakakis, M. Mitzenmacher, J. Blasiok, B. Lawson, P. Nakkiran, and V. Nakos.
Predicting positive and negative links with noisy queries: Theory & practice. arXiv preprint
arXiv:1709.07308, 2017.
[54] U. Von Luxburg, A. Radl, and M. Hein. Hitting and commute times in large graphs are often
misleading. arXiv preprint arXiv:1003.1266, 2010.
[55] U. Von Luxburg, A. Radl, and M. Hein. Hitting and commute times in large random neighborhood graphs. Journal of Machine Learning Research, 15(1):1751–1798, 2014.
[56] D. M. Wittmann, D. Schmidl, F. Blöchl, and F. J. Theis. Reconstruction of graphs based on
random walks. Theoretical Computer Science, 410(38):3826 – 3838, 2009.
[57] L. Yen, F. Fouss, C. Decaestecker, P. Francq, and M. Saerens. Graph nodes clustering based
on the commute-time kernel. Advances in Knowledge Discovery and Data Mining, 2007.
[58] E. Zheleva, E. Terzi, and L. Getoor. Privacy in social networks. Synthesis Lectures on Data
Mining and Knowledge Discovery, 3(1):1–85, 2012.
22
| 8cs.DS
|
Interactive Timetabling
Tomáš Müller, Roman Barták*
Charles University
Department of Theoretical Computer Science
Malostranské náměstí 2/25, Praha 1, Czech Republic
[email protected]
[email protected]
Abstract. Timetabling is a typical application of constraint programming
whose task is to allocate activities to slots in available resources respecting
various constraints like precedence and capacity. In this paper we present a
basic concept, a constraint model, and the solving algorithms for interactive
timetabling. Interactive timetabling combines automated timetabling (the
machine allocates the activities) with user interaction (the user can interfere
with the process of timetabling). Because the user can see how the timetabling
proceeds and can intervene this process, we believe that such approach is more
convenient than full automated timetabling which behaves like a black-box.
The contribution of this paper is twofold: we present a generic model to
describe timetabling (and scheduling in general) problems and we propose an
interactive algorithm for solving such problems.
Introduction
Timetabling is a form of scheduling where the task is to allocate activities to available
slots in resources respecting some constraints. Typically, the activity is described by
its duration and by resources required to process the activity (conjunction and
disjunction of resources can be used). There are also precedence constraints between
the activities stating which activity must proceed before another activity and resource
constraints stating when and how many activities can be processed by a given
resource. In addition to the above hard constraints, that cannot be violated, there exist
many preferences - soft constraints describing users' wishes about the timetable.
Constraint programming (CP) is a natural tool for describing and solving such
problems and there exist many timetabling systems based on CP [1, 2, 5, 9, 12].
In the current personal computing environment, an interactive behaviour of
software is a feature requested by many users. Interactivity manifests itself in two
directions: the user can observe what is doing inside the system (this is important for
debugging but the final users like the animation of problem solving process as well
[personal communication to Helmut Simonis]) and he or she can immediately
influence the system. Interactive timetabling is a concept where the system presents to
the user how the timetable is being built and the user can interfere with this process
*
Supported by the Grant Agency of the Czech Republic under the contract no. 201/01/0942.
by changing the task on flow. Naturally, this requires a special interactive solving
algorithm that is building a timetable step by step and that can present a sound (sub-)
result anytime. Moreover, such algorithm can be exposed to task change, so it should
not start re-scheduling from scratch after such a change but the new solution is built
using a former solution. By the task change we mean that the user can schedule the
activity manually (and fix it in the slot) and vice versa (remove the activity from the
slot to which it is currently allocated). Moreover, he/she can alter some other
parameters like activity duration, precedence relation etc.
In this paper we define a generic timetabling problem motivated by school
timetabling and we propose an interactive algorithm to solve this problem. It may
seem that local search algorithms are best suited for this type of problems but note
that typically local search is exploring the space of complete schedules which are not
sound and it reduces the number of violations. However, we require a sound (perhaps
incomplete) schedule to be presented to the user every time. On the other side
backtracking based search, which is a method of extending sound but incomplete
solutions to a complete solution, does not support interactivity in the sense of
changing the task during search. Therefore we propose a mixture of both approaches
that borrows techniques both from backtracking based search (extending a partial
solution to a complete solution and value and variable selection criteria) as well as
from local search (tabu list to prevent cycling). In some sense, the presented algorithm
mimics the behaviour of a human scheduler and therefore the animation of the
scheduling/timetabling process looks naturally. To show capabilities of the proposed
algorithm we implemented a timetabling system in Java that consists of independent
modules with a generic scheduling engine and with a graphical user interface.
Motivation and Problem Description
A traditional scheduling problem is defined by a set of activities and by a set of
resources to which the activities are allocated. Moreover, the activities must be
positioned in time (otherwise, the problem is called resource allocation). Timetabling
can be seen as a special form of the scheduling problem with slightly simplified view
of time. In the timetabling problem, the time slots are defined uniformly for all the
resources so the activities are being allocated to these slots (rather than to particular
time).
As the basic motivation of our research we took a school timetabling problem and
we extracted the basic features to model timetabling problems in general. The basic
object to be scheduled in a typical school timetabling problem is a lecture. The
scheduled time interval, typically a week, is divided into the time slots with the same
duration. For example, we have ten slots per day where every slot takes 45 minutes.
Every lecture has assigned its duration, expressed as a number of the time slots; this
number is usually rather small - from one to three. Next, we have a set of resources,
e.g. classrooms where the lectures take place, teachers who teach the lectures,
classes for which the lectures are given, and other resources like overhead projector
etc. Capacity of every resource is limited to one lecture. It may seem strange to
consider classes as a resource but doing this allows us to define naturally constraints
like no overlap of the lectures for a single class or a common lecture for several
classes. Note that the set of slots is assigned to every resource so when we are
allocating the lectures to the slots, it means allocation to a particular slot of a
particular resource. It is possible to disable some slots in the resource to express
forbidden time when no activity can be scheduled, e.g. the teacher is not available, the
classroom is occupied by some external activity etc. Naturally, it is also possible to
put preferences to the slots expressing to which slots the activities should be
preferentially allocated.
Every lecture has assigned a teacher, a class, a set of possible classrooms (one of
them is selected), and a set of special resources (overhead projector etc.). So when we
are allocating a lecture, we are putting the lecture to equivalent time slots in all
requested resources. Thus, the task is to choose a start time (a start lot) of every
lecture as well as to select one resource among alternative resources (e.g. classrooms).
Until now the relationship between the lectures is indirect only, via resources (no two
different lectures can be allocated to a single slot of the same resource). Sometimes,
the relation between the lectures should be explicit, for example to express that two
alternative lectures are taught in parallel or that the lesson is before the exercises. We
call this relation a dependency.
Fig. 1. In interactive timetabling, the user see continuous improvements of the schedule (left)
and he or she can interfere with this process, e.g. by manual allocation of the activity (right).
In the following section, we propose a general model motivated by the above school
timetabling problem. However, we believe that this model can be applied to other
timetabling problems different from this particular school timetabling problem. Still,
there is a space for other extensions of this model, like introduction of new
preferences or dependencies.
The Model
We propose a generic model for timetabling problems consisting of a set of resources,
a set of activities, and a set of dependencies between the activities. Time is divided
into the time slots with the same duration. Every slot may have assigned a constraint,
either hard or soft one: a hard constraint indicates that the slot is forbidden for any
activity, a soft constraint indicates that the slot is not preferred. We call these
constraints “time preferences”. Every activity and every resource may have assigned a
set of time preferences, which indicate forbidden and not preferred time slots.
Activity (which can be directly mapped to a lecture) is identified by its name.
Every activity is described by its duration (expressed as a number of time slots), by
time preferences, and by a set of resources. This set of resources determines which
resources are required by the activity. To model alternative as well as required
resources, we divide the set of resources into several subsets – resource groups. Each
group is either conjunctive or disjunctive: the conjunctive group of resources means
that the activity needs all the resources from the group, the disjunctive group means
that the activity needs just one of the resources (we can select among several
alternatives). An example can be a lecture, which will take place in one of the
possible classrooms and it will be taught for all of the selected classes. Note that we
need not model conjunctive groups explicitly because we can use a set of disjunctive
groups containing exactly one resource instead (the set of required resources can be
described in a conjunctive normal form). However, usage of both conjunctive and
disjunctive groups simplifies modelling for the users.
Resource is also identified by its name and it is fully described by time
preferences. There is a hard restriction, that only one activity can use the resource in
the same time. As we mentioned in the previous section, the resource represents a
teacher, a class, a classroom or another special resource at the school timetabling
problem.
Finally, we need a mechanism for defining and handling direct dependencies
between the activities. It seems sufficient to use binary dependencies only that define
relationship between two activities. Currently, we use a selection of three temporal
constraints: the activity finishes before another activity, the activity finishes exactly at
the time when the second activity starts, and two activities run concurrently (they
have the same start time). The scheduling engine provides an interface for
introduction of other binary dependencies.
The solution of the problem defined by the above model is a timetable, where
every scheduled activity has assigned its start time and a set of reserved resources,
which are needed for its execution (the activity is allocated to respective slots). This
timetable must satisfy all the hard constraints, namely:
• every scheduled activity has all required resources reserved, i.e., all
resources from the conjunctive groups and one resource from each
disjunctive group of resources,
• two scheduled activities cannot use the same resource at the same time,
• no activity is scheduled into a time slot where the activity or some of its
reserved resources has a hard constraint in the time preferences,
• all dependencies between the scheduled activities must be satisfied.
Furthermore, we want to minimise the number of violated soft constraints in the time
preferences of resources and activities. We do not formally express this objective
function; these preferences will be used as a guide during search for the solution.
Last but not least, in interactive timetabling we want to present some solution to
the user at each time. Therefore, we will work with partial solutions where some of
the activities are not scheduled yet. Still, these partial solutions must satisfy the above
constraints. Note that using a partial solution allows us to provide some reasonable
result even in case of over-constrained problems.
An Interactive Solver
To satisfy needs of interactive timetabling we designed an ad-hoc algorithm for
solving the above described timetabling problem. Nevertheless, we believe that this
algorithm can be generalised to solve other interactive problems as well.
Let us first summarise the requirements to such interactive solving algorithm. The
algorithm should provide some (partial) solution at each step. This (partial) solution
must be sound (all required constraints are satisfied) and it should be good enough in
respect to satisfaction of soft constraints. It is not necessary to find the optimal
solution (minimal number of violated soft constraints), a good enough solution is OK.
We insist more on interactivity of the algorithm that means two things: the difference
between the solutions provided by the algorithm in two consecutive steps should be
minimal and the user can interfere with the solution process. The user may change
parameters of activities and resources, add or remove dependencies, allocate or detach
any activity from the time slot. The algorithm must be able to start from such
modified timetable, make it sound and continue in extending the partial solution to a
complete solution.
There exist two basic approaches how to solve problems defined by means of
constraints: backtracking based search that extends a partial sound solution to a
complete solution and local search that decreases the number of violations in a
complete solution. The idea of local search follows our requirement that there is some
solution provided at each step and the difference between the solutions in consecutive
steps is rather small. However, the provided solutions are not sound. On the other side
backtracking based search can provide a partial sound solution at each step but it is
not easy to add interactive behaviour to it. In particular, it does not support external
changes of the partial solution and the difference between two consecutive solutions
can be large after long backtrack.
To satisfy needs of an interactive solving algorithm we propose to mix features of
above two approaches. Our algorithm uses forward based search that extends a partial
sound solution to a complete solution. The next solution is derived from the previous
solution by allocating some activity and removing conflicting activities from the
partial schedule. Every step is guided by a heuristic that combines minimal
perturbation and minimal violation of soft constraints requirements. If the user
changes somehow the partial solution, the algorithm first removes all the violated
activities and then it continues from a given sound solution.
A Basic Concept of the Interactive Solver
We propose an interactive scheduling algorithm that works in iterations. The
algorithm uses two basic data structures: a set of activities that are not scheduled and
a partial sound schedule. At each iteration step, the algorithm tries to improve the
current partial schedule towards a complete schedule. The scheduling starts with an
empty partial schedule (which is a sound schedule), i.e. all the activities are in the list
of non-scheduled activities. Then it goes repeatedly from one partial solution to
another sound solution until all the activities are scheduled or the maximum number
of iterations is reached.
The user may interrupt the algorithm after arbitrary iteration and he or she can
acquire a solution (latest or the best ever found) where perhaps not all the activities
are scheduled but all the constraints on the scheduled activities are satisfied. During
this interruption, the user may allocate manually some of the non-scheduled activities,
weaken some constraints or make other changes (add new activities etc.) and then let
the automated scheduling continue.
Look now what is going on in the iteration step. First, the algorithm selects one
non-scheduled activity and computes the locations where the activity can be allocated
(locations with the hard constraint are not assumed). Every location is described by a
start time (the number of the first slot for the activity) and by a set of required
resources. Then every location is evaluated using a heuristic function over the current
partial schedule. Finally, the activity is placed to the best location that may cause
some conflicts with already scheduled activities. Such conflicting activities are
removed from the schedule and they are put back to the list of non-scheduled
activities.
procedure solve(unscheduled, schedule, max_iter)
// unscheduled is a list of non-scheduled activities
// schedule is a partial schedule (empty at the beginning)
iterations=0;
while unscheduled non empty & iterations<max_iter
& non user interruption do
iterations ++;
activity = selectActivityToSchedule(unscheduled);
unscheduled -= activity;
location = selectPlaceToSchedule(schedule, activity);
unscheduled += place(schedule, activity, location)
// place the activity and return violated activities
end while;
return schedule;
end solve
Fig. 2. A kernel of the interactive scheduling algorithm
The above algorithm schema is parameterised by two functions: activity selection and
location selection. We can see these functions in a broader view of constraint
programming as variable selection and value selection functions. Note that
value/variable selection functions can be found both in backtracking based search as
well as in local search algorithms and there exist some general techniques used to
define these functions. Our usage of these functions is somewhere in between
backtracking and local search. We are selecting non-scheduled activity (i.e., non-
instantiated variables) but during this selection we use information about the previous
locations of the activity. Remind that the activity might be removed from the schedule
during some previous iteration step. Moreover, because of removing activities from
the partial schedule we need some mechanism to prevent cycles. Techniques used for
activity and location selection and for escaping from a cycle are described in next
sections.
Activity Selection (a variable selection criterion)
As mentioned above, the proposed algorithm requires a function that selects a nonscheduled activity to be placed in the partial schedule. This problem is equivalent to
what is known as a variable selection criterion in constraint programming. There exist
several general guidelines how to select a variable. In local search, usually a variable
participating in the largest number of violations is selected first. In backtracking based
algorithms, a first-fail principle is often used, i.e., a variable whose instantiation is the
most complicated is selected first. This could be a variable involved in the largest set
of constraints (static criterion) or a variable with the smallest domain (dynamic
criterion) etc. [3, 4, 7, 8, 11]
We follow the general variable selection rules and we recommend selecting the
hardest to schedule activity first. For simplicity reasons, we call it the worst activity.
The question is how to decide between two activities which one is the worse activity.
We can use both static and dynamic information when deciding about the worst
activity. An example of static information is a number of dependencies in which the
activity participates. More dependencies mean worse activity. Dynamic information is
information which arises from the partial schedule, for example a number of locations
in the timetable that don’t conflict with any already scheduled activity. It is more
expensive to acquire dynamic information but this information is usually more
accurate than static information. Moreover, if static information is used alone then the
solving algorithm can be easily trapped in a cycle.
In our algorithm we currently use the following criteria when selecting the activity:
o How many times was the activity removed from the timetable? (N#Rm)
o In how many dependencies does the activity participate? (N#Dep)
o In how many places can this activity be placed? (N#Plc)
Note: Places, where another activity lays and can’t be removed (the
activity was fixed by the user), are not counted.
o In how many places this activity can be placed with no conflict? (N#PlcNC)
A weighted-sum of above criteria is used as follows:
valactivity = − w1 N # Rm − w2 N # Dep + w3 N # Plc + w4 N # PlcNC
Activity with the minimal heuristic value is selected. Notice that the first two criteria
are used with a negative weight, this is because a larger value means worse activity.
Using such formula gives us more flexibility when tuning the system for a particular
problem. Moreover, it allows studying influence of a particular criterion to efficiency
of the scheduling.
It is possible to select the worst activity among the all non-scheduled activities but
due to complexity of computing the heuristic value, this could be rather expensive.
Therefore we propose to select a subset of non-scheduled activities randomly (we use
a probability of selection 0.2) and then choose the worst activity from this subset. The
results will not be much worse and we can select the activity approximately five times
faster (see section Results).
Location Selection (a value selection criterion)
After selecting the activity we need to find a location where to place it. This problem
is usually called a value selection in constraint programming. Typically, the most
useful advice is to select the best-fit place [3, 4, 7, 8, 11]. So, we are looking for a
place, which is the most preferred for the activity and also where the activity causes
less trouble. It means that we need to find a place with minimal potential future
conflicts with other activities. Note that we are not using constraint propagation
explicitly in our algorithm, the power of constraint propagation is hidden in the
location selection.
To find the best place for the activity we explore the space of all possible start
times and for each start time we explore the possible sets of resources that the activity
needs for its execution. Remind that for a single start time we can have several sets of
resources (there can be alternatives among the required resources). We evaluate each
such location using the following criteria:
o How many scheduled activities will be in the conflict with the current
activity if it is placed to the selected location? (N#CnfAct)
o Will the activity remove another activity repeatedly? (N#rep, the number of
such repeatedly removed activities)
For every activity we can register the activity which caused its removal
from the schedule in any previous iteration. By including these activities
in the next decision we can avoid some short cycles.
o How many scheduled activities, which are in conflict with the location,
can’t be rescheduled without another conflict elsewhere? (N#ConflNoRsh)
o How many soft constraints are violated (both in activity and selected
resources time preferences)? (N#soft)
o How far is the location from the location where the activity was placed
before? (NdiffPl)
o How good is the location from the user point of view? (Nuser)
This is a user-defined preference for location used to model real-life
problems (for example, prefer morning slots to evening slots).
Again, we use a weighted sum of above criteria to evaluate the location:
val place = w1′N #CnfAct + w′2 N # rep + w3′ N #ConfNoRsh + w′4 N # soft + w5′ N diffPl + w6′ N user
Location with the minimal heuristic value is selected. To remove cycling, it is
possible to randomise this location selection procedure. For example, it is possible to
select five best locations for the activity and then choose one randomly. Or, it is
possible to select a set of best locations such that the heuristic value for the worst
location in this group is maximally twice as large as the heuristic value of the best
location. Again, the location is selected randomly from this group. This second rule
inhibits randomness if there is a single very good location.
How to Escape a Cycle?
In the previous two sections about value and variable selection we proposed some
mechanisms how to avoid cycling, but we can do more. In the current implementation
of the scheduling engine we use a technique based on a tabu list [3, 6, 10].
Tabu list is a FIFO (first in first out) queue of pairs (variable, value) with a fixed
length. When a new value is assigned to a variable, a new pair is inserted into the tabu
list and the last pair is removed from the tabu list. Now, when we are selecting a new
value for the variable X we can avoid a repeated selection of the same value for the
variable (value V is not selected if the pair (X,V) is in the tabu list,). This tabu rule
prevents short cycles (a cycle length corresponds to the length of the tabu list). There
also exist other rules, called aspiration criteria, which can brake the above tabu rule of
assignment, i.e., we can assign a value satisfying the aspiration criterion even if it is in
the tabu list.
In our scheduling algorithm we adapted a slightly modified version of the tabu list.
We use pairs (activity, location) in the tabu list. When we choose an activity A and
before we compute a heuristic value for the location L we explore the tabu list. If the
pair (A,L) appears twice in the tabu list then we do not use the location L for the
activity A in this iteration. If the pair (A,L) appears exactly once in the tabu list then
the location L is assumed only if this location has the minimal heuristic value (it is the
best location). This is also the case when the pair (A,L) could be introduced to the
tabu list for the second time (if the location L is selected). Finally, if the pair (A,L) is
not in the tabu list then the location is processed as described in the previous section.
To summarise it, our aspiration criterion allows us to select some location for the
activity maximally two times in the period defined by the length of the tabu list. The
second assignment is allowed only if the location is the best for the activity.
Preliminary Results
We have implemented the above described scheduling algorithm in Java and added a
user interface for solving school timetabling problems (see Figure 1). In this section
we present some preliminary results achieved using randomly generated problems
with the size 20 classes, rooms, and teachers and 10 slots per day (5 days). In these
tests, we used fix weights for all the criteria described in the sections about activity
and location selection. We compared the three basic activity selection criteria, namely
random selection of the activity, selection of the worst activity among 20% randomly
selected non-scheduled activities, and selection of the worst activity among all nonscheduled activities. Feasible problems with different number of activities were used
(the number of activities is measured in % of the timetable space).
It is not surprising that the selection of the activity from all non-scheduled
activities leads to the smallest number of iterations and to schedules with the largest
number of scheduled activities. However, this method also brings some overhead so it
needs more time to find a schedule. Random selection of the activity is very fast but it
requires higher number of iterations to find a schedule and in cases with more
activities the number of scheduled activities decreases. Our experiments confirm that
the idea of combination of random selection with activity selection heuristic could
bring interesting results. The number of iterations and successfully scheduled
activities is comparable to the method where all non-scheduled activities are explored.
However, the overall time is lower because we are choosing from about 20% of
randomly selected activities only.
150.000
2000
140.000
1750
130.000
120.000
1500
110.000
100.000
90.000
Time [ms]
Iterations
1250
1000
80.000
70.000
60.000
750
50.000
40.000
500
30.000
250
20.000
10.000
0
0
50
55
60
65
70
75
80
85
90
95
100
50
55
60
Fullness of the timetable [%]
65
70
75
80
85
90
95
100
Fullness of the timetable [%]
Fig. 3. Comparison of the number of iterations and time for three basic variable selection
criteria (× randomly selected activity, ◊ the worst activity among 20% randomly selected
activities, • the worst activity among all activities).
100
90
Planned activities [%]
80
70
60
50
40
30
20
10
0
50
55
60
65
70
75
80
85
90
95
100
Fullness of the timetable [%]
Fig. 4. Comparison of the number of scheduled activities for three basic variable selection
criteria (× randomly selected activity, ◊ the worst activity among 20% randomly selected
activities, • the worst activity among all activities).
Conclusions and Future Work
We presented a promising algorithm for solving timetabling problems, which
combines principles of the local search and other techniques for solving constraint
satisfaction problems. The presented algorithm can be used in interactive environment
that is its greatest advantage. Moreover, the algorithm can be easily extended by
adding more heuristics describing other soft constraints. This algorithm was proposed
to solve timetabling problems with motivation in school timetabling. However, we
believe that its principles can be used in other constraint satisfaction problems
especially when interactive behaviour is required.
Currently, we are working on further empirical studies of this algorithm with a
particular emphasis on studies how the weights influence efficiency. Further research
is oriented both theoretically, to formalise the techniques and to put them in a wider
context of constraint programming, and practically, to implement the engine as a
system of co-operating parallel threads.
References
[1]
S. Abdennadher and M. Marte. University timetabling using constraint handling
rules. Journal of Applied Artificial Intelligence, Special Issue on Constraint
Handling Rules, 1999.
[2] R. Barták. Dynamic Constraint Models for Planning and Scheduling Problems.
In New Trends in Constraints, LNAI 1865, pp. 237-255, Springer, 2000.
[3] R. Barták. On-line Guide to Constraint Programming,
http://kti.mff.cuni.cz/~bartak/constraints, 1998.
[4] D. Clements, J. Crawford, D. Joslin, G. Nemhauser, M. Puttlitz, M.
Savelsbergh. Heuristic optimization: A hybrid AI/OR approach. In Workshop
on Industrial Constraint-Directed Scheduling, 1997.
[5] J. M. Crawford. An approach to resource constrained project scheduling. In
Artificial Intelligence and Manufacturing Research Workshop, 1996.
[6] P. Galinier and J.K. Hao. Tabu search for maximal constraint satisfaction
problems. In Proceedings of CP'97, G. Smolka ed., LNCS 1330, pp. 196-208,
Schloss Hagenberg, Austria, Springer, 1997.
[7] J.-M. Labat, L. Mynard, Oscillation, Heuristic Ordering and Pruning in
Neighborhood Search. In Proceedings of CP'97, G. Smolka ed., LNCS 1330,
pp. 506-518, Schloss Hagenberg, Austria, Springer, 1997.
[8] Z. Michalewicz, D. B. Fogel. How to Solve It: Modern Heuristics. Springer,
2000.
[9] A. Schaerf. A survey of automated timetabling. Technical Report CS-R9567,
Centrum voor Wiskunde en Informatica (CWI), Amsterdam, The Netherlands,
1996.
[10] A. Schaerf. Tabu search techniques for large high-school timetabling problems.
In Proceedings of the Fourteenth National Conference on Artificial Intelligence,
pp. 363-368, Portland, Oregon, USA,1996.
[11] E. Tsang and C. Voudouris. Fast Local Search and Guided Local Search and
Their Application to British Telecom's Workforce Scheduling Problem.
Technical Report CSM-246, Department of Computer Science, University of
Essex, 1995.
[12] M. Wallace. Applying constraints for scheduling. In Constraint Programming,
volume 131 of NATO ASI Series Advanced Science Institute Series, Springer,
1994.
| 6cs.PL
|
arXiv:1803.03938v1 [math.AC] 11 Mar 2018
Про моногеннi функцiї, визначенi в рiзних
комутативних алгебрах
Вiталiй С. Шпакiвський
Анотацiя. Встановлено вiдповiднiсть мiж моногенною функцiєю в довiльнiй скiнченновимiрнiй комутативнiй асоцiативнiй алгебрi i скiнченним набором моногенних функцiй в спецiальнiй комутативнiй асоцiативнiй алгебрi.
MSC 2010. 30G35, 57R35
Ключовi слова. Комутативна асоцiативна алгебра, моногенна
функцiя, характеристичне рiвняння, iнтегральне представлення
1
Вступ
Напевно першим хто використав аналiтичнi функцiї, що приймають
значення в комутативнiй алгебрi для побудови розв’язкiв тривимiрного рiвняння Лапласа був П. Кетчум [1]. Вiн показав, що кожна аналiтична функцiя Φ(ζ) змiнної ζ = xe1 + ye2 + ze3 задовольняє тривимiрне рiвняння Лапласа, якщо лiнiйно незалежнi елементи
e1 , e2 , e3 комутативної алгебри задовольняють умову
e21 + e22 + e23 = 0 ,
(1.1)
оскiльки
∆3 Φ :=
∂2Φ ∂2Φ ∂2Φ
+
+
≡ Φ′′ (ζ) (e21 + e22 + e23 ) = 0 ,
∂x2
∂y 2
∂z 2
(1.2)
де Φ′′ := (Φ′ )′ i Φ′ (ζ) визначається рiвнiстю dΦ = Φ′ (ζ)dζ.
Узагальнюючи П. Кетчума, М. Рошкулець [2, 3] використовував
аналiтичнi функцiї зi значеннями в комутативних алгебрах для дослiдження рiвнянь вигляду
LN U (x, y, z) :=
X
α+β+γ=N
Cα,β,γ
∂N U
= 0,
∂xα ∂y β ∂z γ
1
Cα,β,γ ∈ R. (1.3)
Розглядаючи змiнну ζ = xe1 + ye2 + ze3 i аналiтичну функцiю
Φ(ζ), отримуємо наступну рiвнiсть для мiшаної похiдної:
∂ α+β+γ Φ
= eα1 eβ2 eγ3 Φ(α+β+γ) (ζ) = eα1 eβ2 eγ3 Φ(N ) (ζ).
∂xα ∂y β ∂z γ
(1.4)
Пiдставляючи (1.4) в рiвняння (1.3), маємо рiвнiсть
X
LN Φ(ζ) = Φ(N ) (ζ)
Cα,β,γ eα1 eβ2 eγ3 .
α+β+γ=N
Приходимо до висновку, що для виконання рiвностi LN Φ(ζ) = 0
елементи алгебри e1 = 1, e2 , e3 мають задовольняти характеристичне рiвняння
X
X (1, e2 , e3 ) :=
Cα,β,γ eβ2 eγ3 = 0 .
(1.5)
α+β+γ=N
Якщо лiву частину рiвняння (1.5) розкласти за базисом алгебри,
то характеристичне рiвняння (1.5) рiвносильне характеристичнiй
системi рiвнянь, породженiй рiвнянням (1.5).
Таким чином, при виконаннi умови (1.5) кожна аналiтична функцiя Φ зi значеннями в довiльнiй комутативнiй асоцiативнiй алгебрi
задовольняє рiвняння (1.3), i, вiдповiдно, усi дiйснозначнi компоненти функцiї Φ є розв’язками рiвняння (1.3).
В роботi [4] розглядаються диференцiальнi рiвняння в частинних
похiдних вiд декiлькох змiнних i наведено ряд прикладiв на застосування описаного вище методу.
I. Мельниченко [5] запропонував розглядати в рiвностях (1.2) i
(1.4) функцiї Φ, двiчi диференцiйовнi за Гато, при цьому описав усi
базиси {e1 , e2 , e3 } тривимiрних комутативних алгебр з одиницею над
полем C, якi задовольняють рiвнiсть (1.1), див. [6].
Для цих тривимiрних комутативних алгебр, асоцiйованих з тривимiрним рiвнянням Лапласа, в роботах [7–9] отримано конструктивний опис усiх моногенних (тобто неперервних i диференцiйовних за
Гато) функцiй за допомогою трьох вiдповiдних голоморфних функцiй комплексної змiнної.
В роботах [10, 11] встановлено конструктивний опис моногенних
функцiй (зв’язаних з рiвнянням ∆3 Φ = 0) зi значеннями в деяких
n-вимiрних комутативних алгебрах за допомогою вiдповiдних n голоморфних функцiй комплексної змiнної i, спираючись на одержанi
представлення моногенних функцiй, доведено аналоги ряду класичних результатiв комплексного аналiзу.
2
Нарештi в роботi [14] отримано конструктивний опис моногенних функцiй (зв’язаних з рiвнянням (1.3)) зi значеннями в довiльнiй
комутативнiй асоцiативнiй алгебрi над полем C за допомогою голоморфних функцiй комплексної змiнної.
У цiй роботi буде показано, що для побудови розв’язкiв рiвняння (1.3) у виглядi компонент моногенних функцiй зi значеннями в
скiнченновимiрних комутативних асоцiативних алгебрах достатньо
обмежитись вивченням моногенних функцiй у алгебрах певного виду.
2
Алгебра Am
n
Нехай N — множина натуральних чисел i m, n ∈ N такi, що m ≤ n.
Нехай Am
n — довiльна комутативна асоцiативна алгебра з одиницею
над полем комплексних чисел C. Е. Картан [12, с. 33] довiв, що в
n
алгебрi Am
n iснує базис {Ik }k=1 , який задовольняє наступнi правила
множення:
(
0 при r 6= s,
1. ∀ r, s ∈ [1, m] ∩ N :
Ir Is =
Ir при r = s;
2.
∀ r, s ∈ [m + 1, n] ∩ N :
Ir Is =
n
P
k=max{r,s}+1
Υsr,k Ik ;
3. ∀ s ∈ [m + 1, n] ∩ N ∃! us ∈ [1, m] ∩ N ∀ r ∈ [1, m] ∩ N :
(
0 при r 6= us ,
Ir Is =
Is при r = us .
Крiм того, структурнi константи Υsr,k ∈ C задовольняють умови асоцiативностi:
(A 1). (Ir Is )Ip = Ir (Is Ip ) ∀ r, s, p ∈ [m + 1, n] ∩ N;
(A 2). (Iu Is )Ip = Iu (Is Ip ) ∀ u ∈ [1, m] ∩ N ∀ s, p ∈ [m + 1, n] ∩ N.
Очевидно, що першi m базисних векторiв {Iu }m
u=1 є iдемпотентами i породжують напiвпросту пiдалгебру S алгебри Am
n , а вектори
{Ir }nr=m+1 породжують нiльпотентну пiдалгебру N цiєї алгебри. З
m
правил множення алгебри Am
n випливає, що An є напiвпрямою сумою m-вимiрної напiвпростої пiдалгебри S i (n − m)-вимiрної нiльпотентної пiдалгебри N , тобто
Am
n = S ⊕s N.
3
Pm
Одиницею алгебри Am
n є елемент 1 =
u=1 Iu .
Алгебра Am
мiстить
m
максимальних
iдеалiв
n
(
)
n
X
Iu :=
λk Ik : λk ∈ C , u = 1, 2, . . . , m,
k=1, k6=u
перетином яких є радикал
R :=
n
n X
k=m+1
o
λk I k : λk ∈ C .
(2.1)
Визначимо m лiнiйних функцiоналiв fu : Am
n → C рiвностями
fu (Iu ) = 1,
fu (ω) = 0 ∀ ω ∈ Iu ,
u = 1, 2, . . . , m.
(2.2)
Оскiльки ядрами функцiоналiв fu є вiдповiдно максимальнi iдеали
Iu , то цi функцiонали є також неперервними i мультиплiкативними
(див. [13, с. 147]).
3
Моногенi функцiї
Нехай
e1 = 1,
e2 =
n
X
ar I r ,
e3 =
n
X
br I r
(3.1)
r=1
r=1
при ar , br ∈ C — трiйка векторiв в алгебрi Am
n , якi лiнiйно незалежнi
над полем R. Це означає, що рiвнiсть
α1 e1 + α2 e2 + α3 e3 = 0,
α1 , α2 , α3 ∈ R,
виконується тодi i тiльки тодi, коли α1 = α2 = α3 = 0.
Нехай ζ := xe1 + ye2 + ze3 , де x, y, z ∈ R. Очевидно, що ξu :=
fu (ζ) = x + yau + zbu , u = 1, 2, . . . , m. Видiлимо в алгебрi Am
n лiнiйну оболонку E3 := {ζ = xe1 + ye2 + ze3 : x, y, z ∈ R}, породжену
векторами e1 , e2 , e3 .
Далi iстотним є припущення: fu (E3 ) = C при всiх u = 1, 2, . . . , m,
де fu (E3 ) — образ множини E3 при вiдображеннi fu . Очевидно, що
це має мiсце тодi i тiльки тодi, коли при кожному фiксованому u =
1, 2, . . . , m хоча б одне з чисел au чи bu належить C \ R. В теоремi
7.1 роботи [14] встановлено пiдклас рiвнянь вигляду (1.3) для яких
умова fu (E3 ) = C виконується при всiх u = 1, 2, . . . , m.
Областi Ω тривимiрного простору R3 поставимо у вiдповiднiсть
область Ωζ := {ζ = xe1 + ye2 + ze3 : (x, y, z) ∈ Ω} в E3 .
4
Неперервну функцiю Φ : Ωζ → Am
n називатимемо моногенною в
областi Ωζ ⊂ E3 , якщо Φ диференцiйовна за Гато в кожнiй точцi цiєї
областi, тобто якщо для кожного ζ ∈ Ωζ iснує елемент Φ′ (ζ) алгебри
Am
n такий, що виконується рiвнiсть
lim (Φ(ζ + εh) − Φ(ζ)) ε−1 = hΦ′ (ζ) ∀ h ∈ E3 .
ε→0+0
Φ′ (ζ) називається похiдною Гато функцiї Φ в точцi ζ.
n
Розглянемо розклад функцiї Φ : Ωζ → Am
n за базисом {Ik }k=1 :
Φ(ζ) =
n
X
(3.2)
Uk (x, y, z) Ik .
k=1
У випадку, коли функцiї Uk : Ω → C є R-диференцiйовними в
областi Ω, тобто для довiльного (x, y, z) ∈ Ω
Uk (x+∆x, y +∆y, z +∆z)−Uk (x, y, z) =
+o
p
(∆x)2 + (∆y)2 + (∆z)2 ,
∂Uk
∂Uk
∂Uk
∆x+
∆y +
∆z+
∂x
∂y
∂z
(∆x)2 + (∆y)2 + (∆z)2 → 0 ,
функцiя Φ моногенна в областi Ωζ тодi i тiльки тодi, коли у кожнiй
точцi областi Ωζ виконуються умови:
∂Φ
∂Φ
=
e2 ,
∂y
∂x
∂Φ
∂Φ
=
e3 .
∂z
∂x
(3.3)
Вiдмiтимо, що розклад резольвенти має вигляд
(te1 − ζ)−1 =
m
X
u=1
n
X
1
Iu +
t − ξu
s=m+1
∀ t ∈ C : t 6= ξu ,
s−m+1
X
k=2
Qk,s
(t − ξus )k
Is
(3.4)
u = 1, 2, . . . , m,
де Qk,s визначенi наступними рекурентними спiввiдношеннями:
Q2,s := Ts ,
Qk,s =
s−1
X
Qk−1,r Br, s , k = 3, 4, . . . , s − m + 1.
r=k+m−2
при
Ts := yas + zbs , Br,s :=
s−1
X
k=m+1
5
Tk Υkr,s , s = m + 2, . . . , n,
а натуральнi числа us визначенi у правилi 3 таблицi множення алгебри Am
n.
Iз спiввiдношень (3.4) випливає, що точки (x, y, z) ∈ R3 , якi вiдповiдають необоротним елементам ζ ∈ Am
n , лежать на прямих
(
x + y Re au + z Re bu = 0,
Lu :
(3.5)
y Im au + z Im bu = 0
в тривимiрному просторi R3 .
Нехай область Ω ⊂ R3 є опуклою в напрямку прямих Lu , u =
1, 2, . . . , m. Позначимо через Du область комплексної площини C , на
яку область Ωζ вiдображається функцiоналом fu .
Теорема А [14]. Нехай область Ω ⊂ R3 є опуклою в напрямку
прямих Lu i fu (E3 ) = C при всiх u = 1, 2, . . . , m. Тодi кожна моногенна функцiя Φ : Ωζ → Am
n подається у виглядi
m
X
1
Iu
Φ(ζ) =
2πi
u=1
Z
Γu
−1
Fu (t)(t−ζ)
n
X
1
Is
dt+
2πi
s=m+1
Z
Gs (t)(t−ζ)−1 dt,
Γus
(3.6)
де Fu — деяка голоморфна функцiя в областi Du i Gs — деяка голоморфна функцiя в областi Dus , а Γq — замкнена жорданова спрямлювана крива, яка лежить в областi Dq , охоплює точку ξq i не
мiстить точок ξℓ , ℓ = 1, 2, . . . , m, ℓ 6= q.
Оскiльки за умов теореми А кожна моногенна функцiя Φ : Ωζ →
Am
n продовжується до функцiї, моногенної в областi
Πζ := {ζ ∈ E3 : fu (ζ) = Du , u = 1, 2, . . . , m},
(3.7)
то надалi будемо розглядати моногеннi функцiї Φ, визначенi в областях виду Πζ .
4
Характеристичне рiвняння в рiзних комутативних алгебрах
Скажемо, що система полiномiальних над полем C рiвнянь Q1 редукується до системи полiномiальних рiвнянь Q2 , якщо система Q2
отримується з системи Q1 шляхом вiдкидання деякої кiлькостi рiвнянь. В свою чергу, система Q2 є редукцiєю системи Q1 . Вiдмiтимо,
що для заданої системи полiномiальних рiвнянь Q1 редукована система Q2 не єдина. Очевидним є наступне твердження.
6
Твердження 4.1. Нехай система полiномiальних рiвнянь Q1 з
комплексними невiдомими t1 , t2 , . . . , tn має розв’язки i Q2 — будьяка її редукована система з невiдомими ti1 , ti2 , . . . , tik , де i1 , i2 , . . . , ik ,
k ≤ n, — попарно рiзнi елементи множини {1, 2, . . . , n}. Тодi усi
ti1 , ti2 , . . . , tik , якi задовольняють систему Q1 є розв’язками системи Q2 .
Наприклад, система рiвнянь
1 + a21 + b21 = 0,
1 + a22 + b22 = 0,
(4.1)
a2 a3 + b2 b3 = 0
редукується до системи рiвнянь
1 + a22 + b22 = 0,
(4.2)
a2 a3 + b2 b3 = 0.
Твердження 4.1 означає, що всi значення a2 , b2 , a3 , b3 , якi задовольняють систему (4.1) є розв’язками системи (4.2).
Встановимо допомiжнi твердження.
Лема 4.1. Нехай в алгебрi Am
n iснує трiйка лiнiйно незалежних над
R векторiв 1, e2 , e3 , якi задовольняють характеристичне рiвняння
(1.5). Тодi для кожного u ∈ {1, 2, . . . , m} характеристична система,
породжена рiвнянням X (Iu , e2 Iu , e3 Iu ) = 0 є редукцiєю характеристичної системи, породженої рiвнянням (1.5).
Доказательство. Нехай лiва частина рiвняння (1.5) в базисi алгебри
має вигляд
X (1, e2 , e3 ) =
X
Cα,β,γ eβ2 eγ3 =
α+β+γ=N
n
X
Vk Ik = 0.
k=1
Вiдповiдно, характеристична система, породжена рiвнянням (1.5),
має вигляд
V1 = 0,
......
(4.3)
Vn = 0.
Тепер розглянемо характеристичну систему, породжену рiвнянням X (Iu , e2 Iu , e3 Iu ) = 0. Маємо
X
X (Iu , e2 Iu , e3 Iu ) =
Cα,β,γ Iu (e2 Iu )β (e3 Iu )γ =
α+β+γ=N
7
= Iu
X
Cα,β,γ eβ2 eγ3 = Iu
α+β+γ=N
n
X
Vk I k = Vu + I u
k=1
n
X
Vk Ik = 0.
k=m+1
(4.4)
Вiдповiдно до правила 3 таблицi множення алгебри Am
добуток
n
n
P
Vk Ik належить радикалу R. Таким чином, рiвняння (4.4)
Iu
k=m+1
рiвносильне такiй характеристичнiй системi:
Vu = 0,
(4.5)
......
Vk = 0 ∀ k ∈ {m + 1, . . . , n} : Iu Ik = Ik .
Очевидно, що система (4.5) є редукцiєю системи (4.3).
Позначимо через Rad e2 частину вектора e2 з розкладу (3.1), яка
n
P
мiститься в його радикалi, тобто Rad e2 :=
ar Ir . Аналогiчно,
r=m+1
Rad e3 :=
n
P
br I r .
r=m+1
Лема 4.2. Нехай в алгебрi Am
n = S ⊕s N iснує трiйка лiнiйно
незалежних над R векторiв 1, e2 , e3 , якi задовольняють характеристичне рiвняння (1.5). Тодi в алгебрi A1n−m+1 = 1 ⊕s N (де нiльпотентна пiдалгебра N та ж сама що й в алгебрi Am
n ) для кожного
u ∈ {1, 2, . . . , m} iснує трiйка векторiв
ee1 (u) = 1,
(4.6)
ee2 (u) := au + Iu Rad e2 ,
ee3 (u) := bu + Iu Rad e3
така, що характеристична система, породжена рiвнянням
X (1, ee2 (u), ee3 (u)) = 0 є редукцiєю характеристичної системи,
породженої рiвнянням (1.5).
Доказательство. Наслiдком рiвностей (4.6) є рiвностi
eeβ2 (u) = aβu + Iu
eeγ3 (u)
=
bγu
+ Iu
β
P
k=1
γ
P
k=1
Cβk auβ−k (Rad e2 )k ,
Cγk bγ−k
u
8
(4.7)
k
(Rad e3 ) .
Враховуючи
формули
(4.7),
характеристичний
многочлен
X (1, ee2 (u), ee3 (u)) = 0 набуває вигляду
X
X
β
γ
Cα,β,γ ee2 (u) ee3 (u) =
Cα,β,γ aβu bγu +
α+β+γ=N
+Iu bγu
β
X
α+β+γ=N
Cβk
auβ−k
k
Iu aβu
(Rad e2 ) +
+ Iu
Cγk bγ−k
(Rad e3 )k +
u
k=1
k=1
β
X
γ
X
Cβk
auβ−k
γ
X
k
(Rad e2 )
Cγp bγ−p
(Rad e3 )p
u
p=1
k=1
= 0.
(4.8)
Далi покажемо, що характеристичнi системи, породженi рiвняннями X (1, ee2 (u), ee3 (u)) = 0 i X (Iu , e2 Iu , e3 Iu ) = 0 спiвпадають.
З цiєю метою зауважимо, що наслiдком розкладiв (3.1) є подання
e2 = a1 I1 + · · · + am Im + Rad e2 ,
e3 = b1 I1 + · · · + bm Im + Rad e3 ,
з яких випливають спiввiдношення
e2 Iu = au Iu + Iu Rad e2 ,
e3 Iu = bu Iu + Iu Rad e3 .
(4.9)
Тепер з (4.9) випливають рiвностi
eβ2 Iu = aβu Iu + Iu
eγ3 Iu
=
bγu Iu
+ Iu
β
P
k=1
γ
P
k=1
Cβk auβ−k (Rad e2 )k ,
(4.10)
Cγk bγ−k
u
k
(Rad e3 ) .
Беручи до уваги формули (4.10), характеристичне рiвняння
X (Iu , e2 Iu , e3 Iu ) = 0 набуває вигляду
X
X
β γ
Iu
Cα,β,γ e2 e3 =
Cα,β,γ aβu bγu Iu +
α+β+γ=N
+Iu bγu
β
X
Cβk
α+β+γ=N
auβ−k
k
(Rad e2 ) +
k=1
+ Iu
β
X
k=1
Cβk
Iu aβu
γ
X
Cγk bγ−k
(Rad e3 )k +
u
k=1
auβ−k
k
(Rad e2 )
γ
X
p=1
Cγp bγ−p
(Rad e3 )p
u
= 0.
(4.11)
З рiвностей (4.8), (4.11) очевидним чином випливає, що характеристичнi системи, породженi рiвняннями X (1, ee2 (u), ee3 (u)) = 0 i
X (Iu , e2 Iu , e3 Iu ) = 0 спiвпадають. Тепер доведення леми випливає з
леми 4.1.
9
Зауваження 4.1. Вiдмiтимо, що алгебра A1n−m+1 = 1⊕s N з базисом
{1, Im+1 , . . . , In } є пiдалгеброю алгебри Am
n = S ⊕s N . Дiйсно, будьякий елемент a алгебри Am
=
S
⊕
N
вигляду
s
n
a = a0 I1 + a0 I2 + · · · + a0 Im + am+1 Im+1 + · · · + an In =
= a0 (I1 +· · · +Im )+am+1 Im+1 +· · · +an In = a0 +am+1 Im+1 +· · · +an In
є представленням довiльного елементу алгебри A1n−m+1 = 1 ⊕s N .
З твердження 4.1 i леми 4.2 випливає така
Теорема 4.1. Нехай в алгебрi Am
n = S ⊕s N iснує трiйка лiнiйно
незалежних над R векторiв 1, e2 , e3 , якi задовольняють характеристичне рiвняння (1.5). Тодi в алгебрi A1n−m+1 = 1 ⊕s N (де нiльпотентна пiдалгебра N та ж сама що й в алгебрi Am
n ) для кожного u ∈ {1, 2, . . . , m} трiйка векторiв (4.6) задовольняє характеристичне рiвняння X (1, ee2 (u), ee3 (u)) = 0.
Приклад 4.1. Розглянемо над полем C алгебру A23 з таблицею множення (див., наприклад, [6, c. 32], [8])
·
I1
I2
I3
I1
I1
0
0
I2
0
I2
I3
I3
0
.
I3
0
(4.12)
Очевидно, що напiвпростою пiдалгеброю S є пiдалгебра, породжена
iдемпотентами I1 , I2 , а нiльпотентною пiдалгеброю N є пiдалгебра
{αI3 : α ∈ C}. Тодi алгебра A12 := 1 ⊕s N спiвпадає з вiдомою бiгармонiчною алгеброю B (див., наприклад, [15]) i має таку таблицю
множення:
·
1 I3
(4.13)
1
1 I3 .
I3 I3 0
Нехай в алгебрi A23 задане характеристичне рiвняння (1.1). Як
вiдомо (див. теорему 1.8 в [6]), умова гармонiчностi (1.1) векторiв
e1 = 1, e2 = a1 I1 + a2 I2 + a3 I3 , e3 = b1 I1 + b2 I2 + b3 I3 алгебри A23
рiвносильна системi рiвнянь (4.1).
Оскiльки для алгебри A23 m = 2, то в алгебрi B ми будуємо двi
трiйки векторiв виду (4.6):
ee1 (1) = 1, ee2 (1) = a1 + I1 (a3 I3 ) = a1 , ee3 (1) = b1 + I1 (b3 I3 ) = b1 (4.14)
10
та
ee1 (2) = 1,
ee2 (2) = a2 + I2 (a3 I3 ) = a2 + a3 I3 ,
(4.15)
ee3 (2) = b2 + I2 (b3 I3 ) = b2 + b3 I3 .
За теоремою 4.1 трiйки (4.14) та (4.15) гармонiчнi в алгебрi B (тобто задовольняють умову (1.1)). Справдi, гармонiчнiсть трiйки (4.14)
рiвносильна першому рiвнянню системи (4.1), а гармонiчнiсть трiйки
(4.15) рiвносильна системi (4.2).
Приклад 4.2. Розглянемо над полем
цею множення
·
I1 I2 I3
I1 I1 0 0
I2
0 I2 0
I3
0 0 I3
I4
0 0 I4
I5 I5 0 0
C алгебру A35 з такою таблиI4
0
0
I4
0
0
I5
I5
0
.
0
0
0
(4.16)
Вiдмiтимо, що напiвпростою пiдалгеброю S є пiдалгебра, породжена
iдемпотентами I1 , I2 , I3 , а нiльпотентною пiдалгеброю N є пiдалгебра з базисом {I4 , I5 }. Тодi алгебра A13 := 1 ⊕s N спiвпадає з вiдомою
алгеброю A4 (див., наприклад, [6, c. 26]) i має таку таблицю множення:
1 I4 I5
·
1
1 I4 I5
.
(4.17)
I4 I4 0 0
I5 I5 0 0
Нехай в алгебрi A35 задане характеристичне рiвняння (1.1). Умова
гармонiчностi (1.1) векторiв вигляду (3.1) алгебри A35 рiвносильна
наступнiй системi рiвнянь
1 + a2u + b2u = 0,
u = 1, 2, 3,
a3 a4 + b3 b4 = 0,
(4.18)
a1 a5 + b1 b5 = 0.
Оскiльки для алгебри A35 m = 3, то в алгебрi A4 ми будуємо три
трiйки векторiв виду (4.6):
ee1 (1) = 1,
ee2 (1) = a1 + I1 (a4 I4 + a5 I5 ) = a1 + a5 I5 ,
ee3 (1) = b1 + I1 (b4 I4 + b5 I5 ) = b1 + b5 I5 ,
11
(4.19)
ee1 (2) = 1,
ee2 (2) = a2 + I2 (a4 I4 + a5 I5 ) = a2 ,
(4.20)
ee3 (2) = b2 + I2 (b4 I4 + b5 I5 ) = b2 ,
та
ee1 (3) = 1,
ee2 (3) = a3 + I3 (a4 I4 + a5 I5 ) = a3 + a4 I4 ,
(4.21)
ee3 (3) = b3 + I3 (b4 I4 + b5 I5 ) = b3 + b4 I4 .
За теоремою 4.1 трiйки (4.19), (4.20) та (4.21) гармонiчнi в алгебрi A4 (тобто задовольняють умову (1.1)). Справдi, гармонiчнiсть
трiйки (4.19) рiвносильна системi з першого i п’ятого рiвняння системи (4.15); гармонiчнiсть трiйки (4.20) рiвносильна другому рiвнянню
системи (4.15), а гармонiчнiсть трiйки (4.21) рiвносильна системi з
третього i четвертого рiвняння системи (4.15).
4.1
Лiнiйна незалежнiсть векторiв 1, e
e2 (u), e
e3 (u)
З наведених прикладiв видно, що вектори 1, ee2 (u), ee3 (u) при деяких
u ∈ {1, 2, . . . , m} можуть бути лiнiйно залежними над полем R. Так,
трiйки (4.14), (4.20) завжди лiнiйно залежнi над полем R.
Встановимо необхiднi i достатнi умови лiнiйної незалежностi над
полем R векторiв 1, ee2 (u), ee3 (u) алгебри A1n−m+1 = 1 ⊕s N .
Лема 4.3. Нехай вектори (3.1) алгебри Am
n = S ⊕s N лiнiйно незалежнi над полем R i нехай u ∈ {1, 2, . . . , m} фiксоване. Тодi
1. якщо вектори Iu Rad e2 , Iu Rad e3 ∈ Am
n лiнiйно незалежнi над
полем R, то вектори 1, ee2 (u), ee3 (u) алгебри A1n−m+1 = 1 ⊕s N також
лiнiйно незалежнi над полем R;
2. якщо ж вектори Iu Rad e2 , Iu Rad e3 ∈ Am
n лiнiйно залежнi
над полем R, то вектори 1, ee2 (u), ee3 (u) алгебри A1n−m+1 = 1 ⊕s N
лiнiйно незалежнi над полем R тодi i тiльки тодi, коли iснує
r ∈ {m + 1, . . . , n} таке, що Iu Ir = Ir i виконується хоча б одне
спiввiдношення
Im au Re br 6= Im bu Re ar
або
Im au Im br 6= Im bu Im ar .
(4.22)
Доказательство. Доведемо перше твердження леми. За умовою рiвнiсть
β2 Iu Rad e2 + β3 Iu Rad e3 = 0, β2 , β3 ∈ R
(4.23)
виконується тодi i тiльки тодi, коли β2 = β3 = 0.
Розглянемо лiнiйну комбiнацiю
α1 + α2 ee2 (u) + α3 ee3 (u) = (α1 + α2 au + α3 bu )+
12
+ (α2 Iu Rad e2 + α3 Iu Rad e3 ) = 0,
α1 , α2 , α3 ∈ R.
(4.24)
Оскiльки вираз у другiй дужцi в рiвностi (4.24) приймає значення
в радикалi R алгебри, а перша дужка комплекснозначна, то умова
(4.24) рiвносильна системi рiвнянь
α1 + α2 au + α3 bu = 0,
α2 Iu Rad e2 + α3 Iu Rad e3 = 0.
(4.25)
З другого рiвняння системи (4.25) i умови (4.23) випливає, що α2 =
α3 = 0. А тодi з першого рiвняння системи (4.25) отримуємо α1 = 0.
Отже, вектори 1, ee2 (u), ee3 (u) лiнiйно незалежнi над R.
Доведемо друге твердження леми. Розглянемо рiвнiсть
n
m
X
X
Is (β1 +β2 as +β3 bs )+
Ik (β2 ak +β3 bk ) = 0,
β1 +β2 e2 +β3 e3 =
s=1
k=m+1
яка рiвносильна системi рiвнянь
β1 + β2 Re as + β3 Re bs = 0,
β2 Im as + β3 Im bs = 0,
s = 1, 2, . . . , m,
β2 Re ak + β3 Re bk = 0,
β2 Im ak + β3 Im bk = 0,
(4.26)
k = m + 1, . . . , n.
Лiнiйна незалежнiсть над R векторiв 1, e2 , e3 означає, що серед усiх
рiвнянь системи (4.26), окрiм першого, iснує хоча б два рiвняння, якi
мiж собою не пропорцiйнi.
Тепер запишемо умову лiнiйної незалежностi над R векторiв
1, ee2 (u), ee3 (u). Для цього систему (4.25) запишемо в розгорнутому
виглядi
α1 + α2 Re au + α3 Re bu = 0,
α2 Im au + α3 Im bu = 0,
α2 Re ar + α3 Re br = 0,
(4.27)
α2 Im ar + α3 Im br = 0
∀ r ∈ {m + 1, . . . , n} : Iu Ir = Ir .
За умовою пункту 2 леми вектори Iu Rad e2 , Iu Rad e3 лiнiйно залежнi над R. Це означає, що в системi (4.27) всi рiвностi, окрiм перших двох, пропорцiйнi мiж собою. Очевидно, що для лiнiйної незалежностi над R векторiв 1, ee2 (u), ee3 (u) необхiдно i достатньо, щоб
друге рiвняння системи (4.27) було не пропорцiйне хоча б з одним
iншим рiвнянням (крiм першого) системи (4.27). А це рiвносильно
умовам (4.22).
13
5
Моногеннi функцiї, визначенi в рiзних комутативних алгебрах
В алгебрi Am
n = S ⊕s N будемо розглядати моногеннi функцiї Φ,
визначенi в деякiй областi Πζ ⊂ E3 виду (3.7). Геометрично область
Π ⊂ R3 , яка конгуентна областi Πζ ⊂ E3 , є перетином m нескiнченних цилiндрiв, кожен з яких паралельний деякiй з m прямих Lu ,
3
u = 1, 2, . . . , m вигляду (3.5). Тобто, Π = ∩m
u=1 Π(u), де R ⊃ Π(u) —
нескiнченний цилiндр, паралельний прямiй Lu . I те ж саме маємо
для конгруентних областей в E3 :
Πζ =
m
\
(5.1)
Πζ (u).
u=1
Аналiтично цилiндр Πζ (u) визначається рiвнiстю
Πζ (u) = {ζu := Iu ζ : ζ ∈ Πζ }.
Тепер розглянемо моногенну в областi Πζ функцiю Φ : Πζ → Am
n.
Введемо позначення
Φu (ζ) := Iu Φ(ζ),
u = 1, 2, . . . , m.
(5.2)
Тодi очевидною є рiвнiсть
Φ = (I1 + · · · + Im )Φ =
m
X
Φu .
(5.3)
u=1
Крiм того, з рiвностi (3.6) i таблицi множення алгебри Am
n випливає,
що при кожному u ∈ {1, 2, . . . , m} функцiя Φu моногенна у всьому
нескiнченному цилiндрi Πζ (u).
Таким чином, кожна моногенна в областi (5.1) функцiя Φ : Πζ →
Am
n подається у виглядi суми (5.3), де функцiя Φu моногенна у всьому
цилiндрi Πζ (u).
e в алгебрi
Тепер перейдемо до розгляду моногенних функцiй Φ
1
An−m+1 = 1 ⊕s N . Оскiльки, вiдповiдно до зауваження 4.1, алгебра
1
A1n−m+1 є пiдалгеброю алгебри Am
n , то в алгебрi An−m+1 усi цилiндри Πζ (u) з рiвностi (5.1) спiвпадають мiж собою. Тобто, в алгебрах
вигляду A1n−m+1 кожна моногенна функцiя буде моногенною в деякому одному нескiнченному цилiндрi.
В наступнiй теоремi встановлюється зв’язок мiж моногенними
1
функцiями в алгебрах Am
n = S ⊕s N та An−m+1 = 1 ⊕s N . Для формулювання результату введемо деякi позначення.
14
На вектори вигляду (4.6) алгебри A1n−m+1 натягнемо лiнiйний
e
e3 (u) := {ζ(u)
простiр E
= x + ye
e2 (u) + ze
e3 (u) : x, y, z ∈ R}. Трiйка
e
векторiв (4.6) визначає одну пряму L(u)
виду (3.5), яка вiдповiдає
e
e3 (u). Нехай Π
ee —
множинi необоротних елементiв ζ(u)
простору E
ζ(u)
e
e
деякий нескiнченний цилiндр в E3 (u), паралельний прямiй L(u).
Теорема 5.1. Нехай в алгебрi Am
n = S ⊕s N iснує трiйка лiнiйно
незалежних над R векторiв 1, e2 , e3 , якi задовольняють характеристичне рiвняння (1.5) i нехай fu (E3 ) = C при всiх u = 1, 2, . . . , m.
Крiм того, нехай функцiя Φ : Πζ → Am
n змiнної ζ = x + ye2 + ze3 моногенна в областi Πζ ⊂ E3 виду (5.1). Тодi в алгебрi A1n−m+1 = 1⊕s N
(де нiльпотентна пiдалгебра N та ж сама що й в алгебрi Am
n ) для
кожного u ∈ {1, 2, . . . , m} iснує трiйка векторiв (4.6), яка задовольняє характеристичне рiвняння X (1, ee2 (u), ee3 (u)) = 0 i iснує функцiя
e
eu : Π
e e → A1
Φ
n−m+1 змiнної ζ(u), яка моногенна в цилiндрi
ζ(u)
n
o
e
e
e e = ζ(u)
e3 (u) : fu ζ(u)
Π
∈
E
=
f
(ζ)
,
ζ
∈
Π
(u)
u
ζ
ζ(u)
така, що
e
e u ζ(u)
.
Φu (ζ) = Iu Φ
(5.4)
Доказательство. Iснування
трiйки
(4.6)
з
властивiстю
X (1, ee2 (u), ee3 (u)) = 0 доведено в теоремi 4.1. Нехай надалi
u ∈ {1, 2, . . . , m} фiксоване. Доведемо iснування i моногеннiсть
e u , яка задовольняє рiвнiсть (5.4). З цiєю
e e функцiї Φ
в областi Π
ζ(u)
метою спочатку доведемо рiвнiсть
Iu ζ −1 = Iu ζe−1 (u)
(5.5)
e
∀ ζ = x + ye2 + ze3 ∀ ζ(u)
= x + ye
e2 (u) + ze
e3 (u), x ∈ C, y, z ∈ R.
З рiвностей (4.10), (4.6) випливають спiввiдношення
Iu e2 = Iu ee2 (u) ,
Iu e3 = Iu ee3 (u) ,
з яких, в свою чергу, випливає рiвнiсть
e
Iu ζ = Iu ζ(u).
(5.6)
Розглянемо рiзницю Iu ζ −1 − Iu ζe−1 (u). За формулою Гiльберта
(див., наприклад, теорему 4.8.2 в [13]), маємо
−1
e
e
ζ ζ(u)
Iu ζ −1 − Iu ζe−1 (u) = Iu ζ − Iu ζ(u)
= 0,
15
внаслiдок рiвностi (5.6). Отже, рiвнiсть (5.5) доведено. Тепер з (5.5)
маємо спiввiдношення
−1
e
(5.7)
Iu (t − ζ)−1 = Iu t − ζ(u)
e
ee .
∀ t ∈ C : t 6= ξu = fu (ζ) ∀ ζ ∈ Πζ (u) ∀ ζ(u)
∈Π
ζ(u)
З таблицi множення алгебри Am
n i формули (3.6) для моногенної
в областi Πζ (u) функцiї Φu (ζ) маємо представлення
1
Φu (ζ) = Iu
2πi
Z
Γu
Fu (t) +
n
X
s=m+1
Is Gs (t) (t − ζ)−1 dt,
(5.8)
де функцiї Fu , Gs визначенi в теоремi A.
Враховуючи спiввiдношення (5.7), представлення (5.8) перепишемо у виглядi
Φu (ζ) = Iu
1
2πi
Z
n
X
−1
e
Is Gs (t) t − ζ(u)
dt.
Fu (t) +
Γu
(5.9)
s=m+1
Оскiльки в алгебрi A1n−m+1 мiститься єдиний максимальний iдеал
I, який спiвпадає з радикалом (2.1) цiєї алгебри R, то на цiй алгебрi
визначений єдиний лiнiйний неперервний мультиплiкативний функцiонал f : A1n−m+1 → C ядром якого є радикал R. А це означає, що
e
e
e3 (u). Беручи до уваги
f ζ(u)
= x + au y + bu z для кожного ζ(u)
∈E
рiвнiсть fu (ζ) = x + au y + bu z для довiльного ζ ∈ E3 , маємо рiвнiсть
e
f ζ(u)
= fu (ζ).
(5.10)
З рiвностi (5.10)
i умови теореми fu (E3 ) = C отримуємо спiввiдноe
e
e3 (u).
шення f ζ(u)
= C для довiльного ζ(u)
∈E
Щойно ми показали, що виконуються умови теореми A для моногенних функцiй в алгебрi A1n−m+1 . Тодi в алгебрi A1n−m+1 формула
e
e e функцiї Φ
e u ζ(u)
має вигляд
(3.6) для моногенної в областi Π
ζ(u)
1
e
e u ζ(u)
Φ
=
2πi
Z
n
X
−1
e
es (t) t − ζ(u)
Is G
dt.
Fe(t) +
γ
(5.11)
s=m+1
Потрiбна нам формула (5.4) буде прямим наслiдком рiвностей
(5.9) та (5.11), якщо ми покажемо, що можна покласти γ ≡ Γu , Fu ≡
es для таких s, що Iu Is = Is . Покажемо це.
Fe i Gs ≡ G
16
ee ⊂
З рiвностi (5.10) випливає, що цилiндри Πζ (u) ⊂ E3 та Π
ζ(u)
e3 (u) вiдповiдними функцiоналами fu та f вiдображаються в одну i
E
ту ж область D комплексної площини C. А це означає, що функцiї
es голоморфнi в однiй i тiй самiй областi
Fu i Fe , а також функцiї Gs i G
es в D.
D. Отже, ми можемо покласти Fu ≡ Fe i Gs ≡ G
Оскiльки кривi iнтегрування γ i Γu лежать в областi D, то ми
можемо взяти γ ≡ Γu . Бiльше того, оскiльки за теоремою А крива
Γu в рiвностi (5.8) охоплює точку fu (ζ) = x + au y + bu z, то внаслiдок
e
рiвностi
(5.10) крива γ ≡ Γu охоплює спектр точки ζ(u) — точку
e
f ζ(u)
= x + au y + bu z. А це нам i потрiбно. Теорему доведено.
Зауваження 5.1. З рiвностей (5.3), (5.4) випливає представлення
e
e
e m ζ(m)
e 1 ζ(1)
.
(5.12)
+ · · · + Im Φ
Φ(ζ) = I1 Φ
e u при
Зауваження 5.2. Теорема 4.1 означає, що функцiї Φ i Iu Φ
всiх u = 1, 2, . . . , m задовольняють одне й те ж саме диференцiальне
рiвняння виду (1.3).
Зауваження 5.3. Теорема 5.1 стверджує, що для побудови
розв’язкiв диференцiального рiвняння (1.3) у виглядi компонент моногенних функцiй зi значеннями в комутативних алгебрах, достатньо
обмежитись вивченням моногенних функцiй в алгебрах з базисом
{1, η1 , η2 , . . . , ηn }, де η1 , η2 , . . . , ηn — нiльпотенти. Тобто кiлькiсть таких n-вимiрних комутативних асоцiативних алгебр з одиницею над
полем C в яких потрiбно вивчати моногеннi функцiї рiвна кiлькостi
(n − 1)-вимiрних комутативних асоцiативних комплексних нiльпотентних алгебр.
Зокрема, серед двовимiрних комутативних асоцiативних алгебр
з одиницею над полем C (яких iснує всього двi) достатньо обмежитись вивченням моногенних функцiй в бiгармонiчнiй алгебрi B. Серед тривимiрних комутативних асоцiативних алгебр з одиницею над
полем C (яких iснує всього чотири) достатньо обмежитись вивченням моногенних функцiй в двох iз них (це алгебри A3 i A4 в термiнах
роботи [6]). А серед чотиривимiрних комутативних асоцiативних алгебр з одиницею над полем C (яких iснує всього 9, див. [16]) достатньо обмежитись вивченням моногенних функцiй в чотирьох iз них
e3,1 , A
e3,2 , A
e3,3 , A
e3,4 з таблицi 9 роботи [17], див. також
(це алгебри A
теорему 5.1 в роботi [18]). Серед усiх п’ятивимiрних комутативних
асоцiативних алгебр з одиницею над полем C (яких iснує всього 25,
див. [16]) достатньо обмежитись вивченням моногенних функцiй в
дев’яти iз них (таблицi множення усiх цих 9 нiльпотентних чотиривимiрних алгебр наведено в теоремi 6.1 з роботи [18]). I нарештi серед
17
усiх шестивимiрних комутативних асоцiативних алгебр з одиницею
над полем C достатньо обмежитись вивченням моногенних функцiй
в 25-ти iз них (усi цi 25 нiльпотентних п’ятивимiрних алгебр наведено в таблицi 1 з роботи [19]). Вiдомо також (див. [20]), що починаючи
з розмiрностi 6 множина усiх попарно неiзоморфних нiльпотентних
комутативних алгебр над C є нескiнченною.
Зауваження 5.4. Теорема 5.1 залишається справедливою для випадку, коли ми будемо розглядати функцiї Φ : Πζ → Am
n змiнної
k
P
ζ :=
xr er , 2 ≤ k ≤ 2n, яка моногенна в областi Πζ ⊂ Ek . При
r=1
цьому замiсть теореми A необхiдно використовувати теорему 1 з роботи [21].
Продемоструємо теорему 5.1 на алгебрах, якi розглядалися в прикладах 4.1 та 4.2.
Приклад 5.1. Отже, розглядаємо алгебру A23 з таблицею множення
(4.12). Для алгебри A23 алгеброю виду 1 ⊕s N є бiгармонiчнi алгебра
B з таблицею множення (4.13).
Вiдповiдно до представлення (3.6), кожна моногенна функцiя Φ
зi значеннями в алгебрi A23 подається у виглядi
Φ(ζ) = F1 (ξ1 )I1 + F2 (ξ2 )I2 + (a3 y + b3 z)F2′ (ξ2 ) + G3 (ξ2 ) I3 (5.13)
∀ ζ ∈ Πζ ,
ξu = x + au y + bu z, u = 1, 2,
де F1 — деяка голоморфна функцiя в областi D1 , а F2 , G3 — деякi голоморфнi функцiї в областi D2 . Оскiльки в A23 m = 2, то
геометрично область Πζ є перетином двох нескiнченних цилiндрiв:
Πζ = Πζ (1) ∩ Πζ (2).
Зауважимо, що представлення (5.13) ранiше було отримано в роботi [8]. Крiм того, функцiя (5.13) задовольняє деяке диференцiальне
рiвняння вигляду (1.3).
Подамо функцiю (5.13) у виглядi (5.3):
Φ(ζ) = Φ(ζ)I1 + Φ(ζ)I2 =: Φ1 (ζ) + Φ2 (ζ),
(5.14)
де Φ1 (ζ) = F1 (ξ1 )I1 — моногенна функцiя в цилiндрi Πζ (1), а функцiя
Φ2 (ζ) = F2 (ξ2 )I2 + (a3 y + b3 z)F2′ (ξ2 ) + G3 (ξ2 ) I3
моногенна в цилiндрi Πζ (2).
18
Перейдемо до розгляду моногенних функцiй в алгебрi B. З предe зi значенставлення (3.6) випливає, що кожна моногенна функцiя Φ
нями в алгебрi B подається у виглядi
e = Fe(ξ)+
e
e G(
e I3 ∀ ζe ∈ Π
e (5.15)
e ζ)
e ξ)
e e , ξe = f (ζ),
Φ(
(a3 y+b3 z)Fe′ (ξ)+
ζ
e — деякi голоморфнi функцiї в областi D. Область Π
e e є нескiнде Fe, G
ζ
ченним цилiндром. Рiвнiсть (5.15) для спецiального випадку встановлена в роботi [15].
Теорема 5.1 стверджує наступне:
1) в алгебрi B iснує трiйка векторiв 1, ee2 (1) , ee3 (1), яка задовольняє
те ж саме характеристичне рiвняння що й трiйка 1, e2 , e3 ∈ A23 . При
e D1 ≡ D i, крiм
цьому, будуть виконуватись спiввiдношення ξ1 ≡ ξ,
e
того, iснує моногенна в B функцiя Φ така, що
e
e 1 ζ(1)
= Φ1 (ζ) .
(5.16)
I1 Φ
2) в алгебрi B iснує трiйка векторiв 1, ee2 (2) , ee3 (2), яка задовольняє
те ж саме характеристичне рiвняння що й трiйка 1, e2 , e3 ∈ A23 . При
e D2 ≡ D i, крiм
цьому, будуть виконуватись спiввiдношення ξ2 ≡ ξ,
e
того, iснує моногенна функцiя Φ така, що
e
e 2 ζ(2)
I2 Φ
= Φ2 (ζ) .
(5.17)
Потрiбнi трiйки векторiв 1, ee2 (1) , ee3 (1) та 1, ee2 (2) , ee3 (2) були знайденi у прикладi 4.1. Розглянемо випадок 1). Дiйсно, для трiйки (4.14)
e
e D1 ≡ D. Покладемо Fe ≡ F1 ,
маємо ζ(1)
= x + a1 y + b1 z ≡ ξ1 ≡ ξ,
e
G ≡ G3 в D. Тодi рiвнiсть (5.15) перепишеться у виглядi
e
e 1 (ζ(1))
Φ
= F1 (ξ1 ) + (a3 y + b3 z)F1′ (ξ1 ) + G3 (ξ1 ) I3 .
(5.18)
Помноживши рiвнiсть (5.18) на I1 , переконуємось у справедливостi
рiвностi (5.16).
Розглянемо випадок 2). Дiйсно, для трiйки (4.15) маємо
e = x + ye
ζ(2)
e2 (2) + ze
e3 (2) = x + a2 y + b2 z + a3 xI3 + b3 yI3 .
e
e D2 ≡ D. Покладемо
Очевидно, що f (ζ(2))
= x + a2 y + b2 z = ξ2 ≡ ξ,
e ≡ G3 в D. Тодi рiвнiсть (5.15) перепишеться у виглядi
Fe ≡ F2 , G
e
e 2 (ζ(2))
Φ
= F2 (ξ2 ) + (a3 y + b3 z)F1′ (ξ2 ) + G3 (ξ2 ) I3 .
(5.19)
Помноживши рiвнiсть (5.19) на I2 , переконуємось у справедливостi
рiвностi (5.17).
19
Таким чином, справедлива рiвнiсть (5.12):
e
e
e 1 ζ(1)
e 2 ζ(2)
Φ(ζ) = I1 Φ
+ I2 Φ
,
e
e
e 2 ζ(2)
e 1 ζ(1)
приймають
,Φ
де Φ приймає значення в алгебрi A23 , а Φ
значення в B.
Приклад 5.2. Отже, розглядаємо алгебру A35 з таблицею множення
(4.16). Для алгебри A35 алгеброю виду 1⊕s N є алгебра A4 з таблицею
множення (4.17).
Вiдповiдно до представлення (3.6), кожна моногенна функцiя Φ
зi значеннями в алгебрi A35 подається у виглядi
Φ(ζ) = F1 (ξ1 )I1 + F2 (ξ2 )I2 + F3 (ξ3 )I3 + (a4 y + b4 z)F3′ (ξ3 ) + G3 (ξ3 ) I4 +
+ (a5 y
∀ ζ ∈ Πζ ,
+ b5 z)F1′ (ξ1 )
+ G5 (ξ1 ) I5
(5.20)
ξu = x + au y + bu z, u = 1, 2, 3,
де F1 , G5 — деякi голоморфнi функцiї в областi D1 , F2 — деяка голоморфна функцiя в областi D2 , а F3 , G3 — деякi голоморфнi функцiї
в областi D3 . Оскiльки для A35 m = 3, то геометрично область Πζ є
перетином трьох нескiнченних цилiндрiв: Πζ = Πζ (1) ∩ Πζ (2) ∩ Πζ (3).
Подамо функцiю (5.20) у виглядi (5.3):
Φ(ζ) = Φ(ζ)I1 + Φ(ζ)I2 + Φ(ζ)I3 =: Φ1 (ζ) + Φ2 (ζ) + Φ3 (ζ),
де
(5.21)
Φ1 (ζ) = F1 (ξ1 )I1 + (a5 y + b5 z)F1′ (ξ1 ) + G5 (ξ1 ) I5
— моногенна функцiя в цилiндрi Πζ (1), функцiя Φ2 (ζ) = F2 (ξ2 )I2
моногенна в цилiндрi Πζ (2) , а функцiя
′
Φ3 (ζ) = F3 (ξ3 )I3 + (a4 y + b4 z)F3 (ξ3 ) + G3 (ξ3 ) I4
моногенна в цилiндрi Πζ (3).
Перейдемо до розгляду моногенних функцiй в алгебрi A4 . З предe зi значенставлення (3.6) випливає, що кожна моногенна функцiя Φ
нями в алгебрi A4 подається у виглядi
e = Fe(ξ)
e + (a4 y + b4 z)Fe′ (ξ)
e +G
e I4 +
e ζ)
e3 (ξ)
Φ(
e +G
e I5
e5 (ξ)
+ (a5 y + b5 z)Fe′ (ξ)
20
e
e e , ξe = f (ζ),
∀ ζe ∈ Π
ζ
(5.22)
e3 , G
e5 — деякi голоморфнi функцiї в областi D ⊂ C. Область
де Fe, G
e
Πζe є нескiнченним цилiндром.
Для трiйки (4.19) алгебри A4 маємо
e = x + ye
ζ(1)
e2 (1) + ze
e3 (1) = x + a1 y + b1 z + a5 xI5 + b5 yI5 .
e
e D1 ≡ D. Покладемо
Очевидно, що f (ζ(1))
= x + a1 y + b1 z = ξ1 ≡ ξ,
e
e
e
F ≡ F1 , G3 ≡ G3 , G5 ≡ G5 в D. Тодi рiвнiсть (5.22) перепишеться у
виглядi
′
e
e
Φ1 (ζ(1)) = F1 (ξ1 ) + (a4 y + b4 z)F1 (ξ1 ) + G3 (ξ1 ) I4 +
+ (a5 y + b5 z)F1′ (ξ1 ) + G5 (ξ1 ) I5 .
(5.23)
Помноживши рiвнiсть (5.23) на I1 , переконуємось у справедливостi
рiвностi
e
e 1 ζ(1)
I1 Φ
= Φ1 (ζ) .
Для трiйки (4.20) алгебри A4 маємо
e = x + ye
ζ(2)
e2 (2) + ze
e3 (2) = x + a2 y + b2 z.
e
e
e D2 ≡ D.
Очевидно, що f (ζ(2))
= ζ(2)
= x + a2 y + b2 z = ξ2 ≡ ξ,
e3 ≡ G3 , G
e5 ≡ G5 в D. Тодi рiвнiсть (5.22)
Покладемо Fe ≡ F2 , G
набуде вигляду
e
e 2 (ζ(2))
Φ
= F2 (ξ2 ) + (a4 y + b4 z)F2′ (ξ2 ) + G3 (ξ2 ) I4 +
+ (a5 y + b5 z)F2′ (ξ2 ) + G5 (ξ2 ) I5 .
(5.24)
Помноживши рiвнiсть (5.24) на I2 , переконуємось у справедливостi
рiвностi
e
e 2 ζ(2)
= Φ2 (ζ) .
I2 Φ
Нарештi, для трiйки (4.21) отримуємо
e = x + ye
ζ(3)
e2 (3) + ze
e3 (3) = x + a3 y + b3 z + a4 xI4 + b4 yI4 .
e
e D3 ≡ D. Покладемо
Очевидно, що f (ζ(3))
= x + a3 y + b3 z = ξ3 ≡ ξ,
e
e
e
F ≡ F3 , G3 ≡ G3 , G5 ≡ G5 в D. Тодi рiвнiсть (5.22) перепишеться у
виглядi
e
e 3 (ζ(3))
Φ
= F3 (ξ3 ) + (a4 y + b4 z)F3′ (ξ3 ) + G3 (ξ3 ) I4 +
+ (a5 y + b5 z)F3′ (ξ3 ) + G5 (ξ3 ) I5 .
21
(5.25)
Помноживши рiвнiсть (5.25) на I3 , переконуємось у справедливостi
рiвностi
e
e 3 ζ(3)
= Φ3 (ζ) .
I3 Φ
Таким чином, справедлива рiвнiсть (5.12):
e
e
e
e 3 ζ(3)
e 2 ζ(2)
e 1 ζ(1)
,
+ I3 Φ
+ I2 Φ
Φ(ζ) = I1 Φ
e
e
e
e 3 ζ(3)
e 2 ζ(2)
e 1 ζ(1)
,Φ
,Φ
де Φ приймає значення в алгебрi A35 , а Φ
приймають значення в A4 .
Список литературы
[1] P. W. Ketchum, Analytic functions of hypercomplex variables //
Trans. Amer. Math. Soc., 30(4) (1928), 641–667.
[2] M. N. Roşculeţ, Algebre infinite asociate la ecuaţii cu derivate
parţiale, omogene, cu coeficienţi constanţi de ordin oarecare //
Studii şi Cercetǎri Matematice, 6(3–4) (1955), 567–643.
[3] M. N. Roşculeţ, Algebre infinite, comutative, asociate la sisteme de
ecuaţii cu derivate parţiale // Studii şi Cercetǎri Matematice, 7(3–
4) (1956), 321–371.
[4] A. Pogorui, R. M. Rodriguez-Dagnino and M. Shapiro, Solutions for
PDEs with constant coefficients and derivability of functions ranged
in commutative algebras // Math. Meth. Appl. Sci., 37(17) (2014),
2799–2810.
[5] I. P. Mel’nichenko, The representation of harmonic mappings by
monogenic functions // Ukr. Math. J., 27(5) (1975), 499–505.
[6] И. П. Мельниченко, С. А. Плакса, Коммутативные алгебры и
пространственные потенциальные поля. – К.: Ин-т математики
НАН Украины, 2008. – 230 с.
[7] S. A. Plaksa, V. S. Shpakovskii, Constructive description of
monogenic functions in a harmonic algebra of the third rank // Ukr.
Math. J., 62(8) (2011), 1251–1266.
[8] S. A. Plaksa, R. P. Pukhtaevich, Constructive description of
monogenic functions in a three-dimensional harmonic algebra with
one-dimensional radical // Ukr. Math. J., 65(5) (2013), 740–751.
22
[9] R. P. Pukhtaievych, Monogenic functions in a three-dimensional
harmonic semi-simple algebra // Zb. Pr. Inst. Mat. NAN Ukr.,
10(4–5) (2013), 352–361.
[10] S. A. Plaksa, V. S. Shpakivskyi, Monogenic functions in a finitedimensional algebra with unit and radical of maximal dimensionality
// J. Algerian Math. Soc., 1 (2014), 1–13.
[11] S. A. Plaksa, R. P. Pukhtaievych, Constructive description of
monogenic functions in n-dimensional semi-simple algebra // An.
Şt. Univ. Ovidius Constanţa, 22(1) (2014), 221–235.
[12] E. Cartan, Les groupes bilinéares et les systèmes de nombres
complexes // Annales de la faculté des sciences de Toulouse, 12(1)
(1898), 1–64.
[13] Э. Хилле, Р. Филлипс, Функциональный анализ и полугруппы.
— М.: Из-во иностр. лит., 1962. — 829 с.
[14] V. S. Shpakivskyi, Constructive description of monogenic functions
in a finite-dimensional commutative associative algebra // Adv.
Pure Appl. Math., 7(1)(2016), 63–75.
[15] S. V. Grishchuk, S. A. Plaksa, Monogenic functions in a biharmonic
algebra // Ukr. Math. J., 61(12) (2009), 1865–1876.
[16] G. Mazzola, Generic finite schemes and Hochschiid cocycles //
Comment. Math. Helvetici, 55(1980), 267–293.
[17] D. Burde, W. de Graaf, Classification of Novicov algebras
// Applicable Algebra in Engineering, Communication and
Computing, 24(1) (2013), 1–15.
[18] A. S. Hegazim, H. Abdelwahab, Classification of five-dimensional
nilpotent Jordan algebras // Linear Algebra and its Applications,
494(2016), 165–218.
[19] B. Poonen, Isomorphism types of commutative algebras of nite rank
over an algebraically closed field // Contemp. Math., 463(2008),
111–120.
[20] Д. А. Супруненко, О максимальных коммутативных подалгебрах полной линейной алгебры // Успехи мат. наук, 11(3)
(1956), 181–184.
23
[21] V. S. Shpakivskyi, Monogenic functions in finite-dimensional
commutative associative algebras // Zb. Pr. Inst. Mat. NAN Ukr.,
12(3) (2015), 251–268.
IНФОРМАЦIЯ ПРО АВТОРА
Вiталiй Станiславович Шпакiвський
Iнститут математики НАН України,
вул. Терещенкiвська, 3, Київ, Україна
[email protected]
24
| 0math.AC
|
Data Rate for Distributed Consensus of Multi-agent
Systems with High Order Oscillator Dynamics
Zhirong Qiu1 , Lihua Xie†,1 , and Yiguang Hong2
arXiv:1609.09206v1 [cs.SY] 29 Sep 2016
Abstract
Distributed consensus with data rate constraint is an important research topic of multi-agent systems. Some results have been
obtained for consensus of multi-agent systems with integrator dynamics, but it remains challenging for general high-order systems,
especially in the presence of unmeasurable states. In this paper, we study the quantized consensus problem for a special kind of
high-order systems and investigate the corresponding data rate required for achieving consensus. The state matrix of each agent is
a 2m-th order real Jordan block admitting m identical pairs of conjugate poles on the unit circle; each agent has a single input,
and only the first state variable can be measured. The case of harmonic oscillators corresponding to m = 1 is first investigated
under a directed communication topology which contains a spanning tree, while the general case of m ≥ 2 is considered for a
connected and undirected network. In both cases it is concluded that the sufficient number of communication bits to guarantee
the consensus at an exponential convergence rate is an integer between m and 2m, depending on the location of the poles.
I. I NTRODUCTION
Distributed consensus is a basic problem in distributed control of multi-agent systems, which aims to reach an interested
common value of the states for a team of agents or subsystems by exchanging information with their neighbors. A variety
of consensus protocols have been proposed for different kinds of applications; see the survey papers [2], [3], [4] and the
reference therein. Nonetheless, to apply the consensus protocol in a digital network with limited bandwidth, it is necessary to
introduce quantization and devise the corresponding encoding-decoding scheme. With static uniform quantization, quantized
consensus was first studied in [5] to achieve the approximate average consensus for integer-valued agents by applying gossip
algorithms. For a large class of averaging algorithms of real-valued agents, [6] established the bounds of the steady-state error
and the convergence times, as well as their dependence on the number of quantization levels. Logarithmic quantizers with
infinite quantization levels were adopted in [7] to guarantee the asymptotic average consensus. To achieve the asymptotic
average consensus with finite quantization levels, a static finite-level uniform quantizer with a dynamic encoding scheme was
proposed in [8], and used to shown that an exponentially fast consensus can be ensured by finite-level quantizers for multi-agent
systems with general linear dynamics, whether the state is fully measurable [9], or the state is only partially measurable and
yet detectable [10]. However, the lower bound of sufficient data rate for the consensus obtained in these works are overly
conservative, and it is more appealing to achieve the consensus with fewer bits of information exchange from the perspective
of reducing communication load.
Some works have been devoted to exploring the sufficient data rate to guarantee the consensus of multi-agent systems with
integrator dynamics, and single-integrator systems receive the most attention. With a presumed bound of the initial state of
each agent, Li et. al. [8] showed that the average consensus can be achieved by 1 bit of information exchange for a fixed and
undirected network, which was further extended to the case when the network is balanced and contains a spanning tree [11]. In
an undirected network where the duration of link failure is bounded, 5-level quantizers suffices for the consensus [8], which also
holds when the network is periodically strongly connected [12]. With a novel update protocol carefully screening the quantized
message, the presumed bound of initial values was shown to be unnecessary in [13] and it was concluded that ternary messages
are sufficient for the average consensus under a periodically connected network. Then, for double-integrator systems with only
position being measurable, [14] concluded that 2 bits of communications suffice for the consensus. By employing a totally
different technique based on matrix perturbation, n bits were found to be sufficient to achieve the consensus of multi-agent
systems with n-th integrator agent dynamics in [15]. Still, it is unclear about the sufficient data rate to guarantee the consensus
for general high-order systems, especially when the state variables are only partially measured.
In this paper, we explore the data rate problem in achieving quantized consensus of another kind of discrete-time high-order
critical systems as a complement of integrator systems. The dynamics of each agent is described by a 2m-th order real Jordan
block admitting m identical pairs of conjugate poles on the unit circle with single input, and only the first state variable
can be measured. We design the encoding-decoding scheme on the basis of the constructability of the state variables of each
*This work is partially supported by National Research Foundation of Singapore under grant NRF-CRP8-2011-03 and NNSF of China under Grant 61174071.
The work in this paper has been partially presented in [1].
† Author for correspondence.
1 Z. Qiu and L. Xie are with School of Electrical and Electronic Engineering, Nanyang Technological University, Singapore 639798. Email:
[email protected]; [email protected]
2 Y.
Hong is with the Key Laboratory of Systems and Control, Institute of Systems Science, Chinese Academy of Sciences, Beijing 100190, China. Email:
[email protected]
individual system: at each time instant the quantizer will produce a signal to make an estimate for the current measurable
state, which is combined with the previous 2m − 1 estimates of the measurable state to obtain the estimate of the current full
state. The same quantized signal will also be sent to neighbor agents to generate the identical estimate of state. The control
input is constructed in terms of the estimate of its own state, as well as those of its neighbor agents. For harmonic oscillators
(m = 1), it is shown that 2 bits of communications suffice to guarantee the exponentially fast consensus for a directed network
containing a spanning tree. For higher-order case of m ≥ 2, the exponentially fast consensus can be achieved with at most
2m bits under an undirected network, provided that the undirected communication topology is connected. The exact number
of bits for achieving consensus in both the cases is an integer between m and 2m, depending on the frequency of oscillators
or the location of poles on the unit circle.
Although the analysis of consensus and data rate in this paper employs similar perturbation techniques as in [15], the problem
posed here is much different, and it is much more challenging to obtain an explicit data rate required for consensus in the
oscillator case (corresponding to complex eigenvalues). In contrast to [15] where the special structure of integrator dynamics
enables a direct connection between the encoder’s past outputs and those at the present moment which leads to a convenient
iteration in the encoding scheme, a similar iteration is no longer available for the estimation of state variables in the case of
oscillator dynamics. As such, a new observer-based encoding scheme is devised. However, such an encoding scheme leads
to the involvement of control inputs into the estimation error, which makes the consensus analysis challenging. Furthermore,
the expression of data rate for the oscillator case requires calculating a linear combination of some rows of a matrix which
is a multiplication of the (2m − 1)-th power of the system matrix and the inverse of the observability matrix, and is hard to
obtain by a direct computation. To overcome this difficulty, we transform the linear combination into a set of linear equations
and employ techniques of combinatorics. It is shown that a data rate between m and 2m, depending on the frequencies of the
oscillations, suffices to achieve the consensus. It is worthy noting that the result not only provides a sufficient data rate for
consensus of the systems under consideration but also reveals an interesting connection between the data rate and the system
dynamics. We believe it will shed some further light on the data rate problem for multi-agent systems of general dynamics.
The rest of the paper is organized as follows. Some preliminaries about graph theory and the problem formulation are
presented in Section II. Then the data rate problem for distributed consensus of the coupled harmonic oscillators is conducted
in Section III, which is followed by the general case of m ≥ 2 in Section IV. For illustration, a numeric example is given in
Section V. Some concluding remarks are drawn in Section VI. The proofs of the main lemmas can be found in the Appendix.
Some notations listed below will be used throughout this paper. For a matrix U , U (i, j) and U (i, ·) respectively denote its
(i, j)-th entry and i-th row; U T is its transpose, and ||U || is its infinity-norm. N+ is the set of positive integers, and ⌈a⌉, ⌊a⌋
respectively denote the smallest integer not less than a, and the largest integer not greater than a. C(n, k) is the number of
k-combinations from a given set of n√ elements. 1N is the N dimensional vector with every component being 1, and Im is
the identity matrix of order m. = −1 is the unit imaginary number. Jλ,n denotes the n dimensional Jordan block with
eigenvalue λ. A ⊗ B denotes the Kronecker product between matrices A and B. h·, ·i denotes the standard inner product in
Euclidean spaces.
II. P ROBLEM F ORMULATION
Consider a multi-agent system in the following form:
i
x (t + 1) = Axi (t) + bui (t),
yi (t) = xi1 (t),
T
where xi (t) = [ xi1 (t), xi2 (t), . . . , xi,2m
(t)] ∈
Q I2
Q
agent i, respectively. Moreover, A =
(1)
R2m , yi (t),
ui (t) ∈ R, i = 1, . . . , N represent the state, output and input of
..
∈ R2m×2m is a real Jordan form consisting of m pairs of conjugate
..
. I2
Q
cos θ
sin θ
T
eigenvalues cos θ + sin θ with sin θ 6= 0 and Q =
; b = [ 0, . . . , 0, 1] ∈ R2m .
− sin θ cos θ
Suppose that the total number of agents is N . Assumed to be error-free, the digital communication channels between agents
are modeled as edges of a directed or undirected graph. A graph G consists of a node set V = {1, . . . , N } and an edge
set E = {(i, v) : i, v ∈ V} where self-loop (i, i) is excluded. An edge (i, v) of a directed graph implies that node v can
receive information from node i, but not necessarily vice versa. In contrast, for an undirected graph, (i, v) ∈ E means mutual
communications between i and v. For node i, Ni+ = {v : (v, i) ∈ E} and Ni− = {v : (i, v) ∈ E} respectively denote its inneighbors and out-neighbors, which coincide if G is undirected, and will be denoted as Ni . A directed path (i1 , i2 ), (i2 , i3 ), . . .
is formed by a sequence of edges. For a directed graph G, if there exists a directed path connecting all the nodes, then G is
said to contain a spanning tree, which is equivalent to the case of being connected when G is undirected.
Usually, a nonnegetive matrix G = [giv ] ∈ RN ×N is assigned to the weighted graph G, where giv > 0 if and only if
(v, i) ∈ E, and giv = gvi is further required for an undirected graph. The connectivity of G can be examined from an algebraic
.
PN
G
G
point of view, by introducing the Laplacian matrix L = DG − G, where DG = diag(dG
1 , . . . , dN ) and di =
v=1 giv . By
L1N = 0, L has at least one zero eigenvalue, with the other non-zero eigenvalues on the right half plane. L has only one zero
eigenvalue √
if and only if G contains a spanning tree [16]. We can always find a nonsingular matrix UL = [φ1 φ2 . . . φN ] with
φ1 = 1N / N and ||φi || = 1, such that UL−1 LUL = diag{J0,N1 , Jλ2 ,N2 , . . . , Jλl ,Nl } , LJ , where 0 ≤ Reλ2 ≤ · · · ≤ Reλl
with λi being an eigenvalue of L. In particular, we denote ψi = UL−1 (i, ·)T . Moreover, LJ = diag{0, λ2 , . . . , λN } with
0 ≤ λ2 ≤ · · · ≤ λN and ψi = φTi if G is undirected.
We adopt the following finite-level uniform quantizer qt (·) in the encoding scheme, where M (t) ∈ N+ :
0,
− 21 < y < 21 ;
j,
2j−1
≤ y < 2j+1
2
2 , j = 1, . . . , M (t) − 1;
(2)
qt (y) =
2M(t)−1
;
M (t),
y≥
2
−qt (−y), y ≤ − 12 .
Remark 2.1: Clearly, the total number of quantization levels of qt (·) is 2M (t) + 1. Demanding that agent i does not send
out any signal when the output is zero, it is enough to use ⌈log2 (2M (t))⌉ bits to represent all the signals.
The problem of distributed quantized consensus is solved if we can design a distributed control protocol based on the outputs
of the encoding-decoding scheme, making the states of different agents reach the agreement asymptotically:
lim [xi (t) − xj (t)] = 0,
t→∞
i, j = 1, 2, . . . , N.
(3)
III. H ARMONIC O SCILLATOR C ASE
In this section, we will start with the harmonic oscillator case as an example to investigate how many bits of information
exchange are enough to achieve consensus exponentially fast with quantized neighbor-based control. We separate it from
higher-order cases due to its speciality and simplicity: the solution of this basic case not only provides a result under a directed
communication topology, but also serves to facilitate the understanding of higher-order cases. Some relevant remarks will be
included in the next section, as a comparison
between second-order
and higher-order cases, or a summary of general cases.
cos θ sin θ
Note that now the system matrix A =
.
− sin θ cos θ
A. Encoding-decoding scheme and distributed control law
An encoding-decoding scheme has a paramount importance in the quantized consensus, which should not only provide
estimates for all the states from the partially measurable states, but also help reduce the data rate. Accordingly, the encoder
should serve as an observer based on iterations. To be specific, inspired by the constructability in the sense that the present
state of the system can be recovered from the present and past outputs and inputs, namely
xi1 (t)
0
1
xi1 (t − 1)
0
=
+
,
(4)
xi2 (t)
− csc θ cot θ
xi1 (t)
ui (t − 1)
we propose the following encoder ϕi for agent i:
i (1)
), x̂i1 (1) = p(0)si (1);
si (1) = qt ( yp(0)
y
(2)
i
si (2) = qt ( p(1)
), x̂i1 (2) = p(1)si (2),
x̂i2 (2) = cot θx̂i1 (2) − csc θx̂i1 (1);
yi (t)−[cos θ x̂i1 (t−1)+sin θ x̂i2 (t−1)]
),
si (t) = qt (
p(t−1)
x̂
(t)
=
cos
θx̂
(t
−
1)
+
sin
θx̂
(t
−
1)
+
p(t − 1)si (t),
i1
i2
i1
x̂i2 (t) = cot θx̂i1 (t) − csc θx̂i1 (t − 1), t ≥ 2,
where p(t) = p0 γ t , 0 < γ < 1 is a decaying scaling function.
After si (t) is received by one of the i-th agent’s out-neighbors, say v ∈ Ni− , a decoder ϕiv will be activated:
x̂iv1 (1) = p(0)si (1);
x̂ (2) = p(1)s (2), x̂ (2) = cot θx̂ (2) − csc θx̂ (1);
iv1
i
iv2
iv1
iv1
x̂
(t)
=
cos
θx̂
(t
−
1)
+
sin
θx̂
(t
−
1)
+
p(t
−
1)si (t),
iv1
iv1
iv2
x̂iv2 (t) = cot θx̂iv1 (t) − csc θx̂iv1 (t − 1), t ≥ 2.
(5)
(6)
Remark 3.1: As in [15], a scaled “prediction error” is quantized to generate the signal si (t), in an effort to reduce the number
of quantization levels. si (t) is then used to construct the estimate x̂i1 (t) of the first component xi1 (t), which is combined with
x̂i1 (t − 1) to obtain the estimate x̂i2 (t) for xi2 (t). Denote ∆i (t) = si (t) − di (t) as the quantization error, where
( y (t)
i
p(t−1) , t = 1, 2;
(7)
di (t) =
yi (t)−[cos θ x̂i1 (t−1)+sin θ x̂i2 (t−1)]
, t > 2,
p(t−1)
and eij (t) = x̂ij (t) − xij (t) as the estimation for xij (t), j = 1, 2. Then comparing (4), (5) and (6) we have
ei1 (t) = x̂i1 (t) − xi1 (t) = p(t − 1)∆i (t), t ≥ 1;
ei2 (t) = x̂i2 (t) − xi2 (t) = cot θei1 (t) − csc θei1 (t − 1) − ui (t − 1), t ≥ 2.
(8)
Evidently the estimation error is related with control inputs in addition to quantization errors, which may impair the consensus.
But as shown in the consensus analysis below, the influence of the control inputs can be ignored by making the control gains
arbitrarily small.
Based on the outputs of the encoding-decoding scheme, the distributed control law of agent i is given by
(
0, t = 0, 1;
P
P2
ui (t) =
(9)
j=1 kj
v∈N + giv [x̂vij (t) − x̂ij (t)], t ≥ 2.
i
B. Consensus Analysis and Data Rate
Some notations are defined as follows:
u(t) = [u1 (t), . . . , uN (t)]T ,
∆(t) = [∆1 (t), . . . , ∆N (t)]T ,
d(t) = [d1 (t), . . . , dN (t)]T ,
xj (t) = [x1j (t), . . . , xN j (t)]T ,
δj (t) = (IN − φ1 ψ1T )xj (t) = [δ1j (t), . . . , δN j (t)],
ej (t) = [e1j (t), . . . , eN j (t)]T .
(10)
We adopt the following two assumptions in the subsequent analysis.
Assumption 3.1: The communication graph G contains a spanning tree.
Assumption 3.2: There exist known positive constants C ∗ and Cδ∗ such that maxj=1,2 ||xj (0)|| ≤ C ∗ and maxj=1,2 ||δj (0)|| ≤
Cδ∗ .
Remark 3.2: Assumption 3.1 is a standard assumption, under which we have 0 < Reλ2 ≤ · · · ≤ Reλl , with 0 as the simple
eigenvalue. Assumption 3.2 enables us to make the quantizer qt (·) unsaturated at initial steps.
The following lemma is critical
in theconsensus analysis.
0 0
and Ai = A − λi K with Reλi > 0. Let kj = cj ε, j = 1, 2 and ε > 0. Then the
Lemma 3.1: Denote K =
k1 k2
following results hold with sufficiently small ε:
1). The spectral radius ρi of Ai is less than 1 if c2 cos θ − c1 sin θ > 0 and c1 cos θ + c2 sin θ = 0. Moreover, ρi =
1 − 21 (Reλi )(c2 cos θ − c1 sin θ)ε + o(ε).
2). Take c1 , c2 as in 1). For any vector ξ ∈ R2 , the entries of Asi ξ, which are denoted as ξs1 and ξs2 , satisfy that |ξsj | ≤ 5/2ρsi
for j = 1, 2.
θ
1 −
e
1 1
1
−1
−1
P
with
P
=
, we have
Proof: 1). Noticing that A = P
and
P
=
2
1
−
e−θ
µ − eθ
µI − A = P
P −1
µ − e−θ
(11)
(µ − eθ ) + (µ − e−θ ) −(µ − eθ ) + (µ − e−θ )
= 12
.
(µ − eθ ) − (µ − e−θ )
(µ − eθ ) + (µ − e−θ )
Consequently the characteristic polynomial of Ai can be obtained as
λi
ε[(c2 + c1 )(µ − eθ ) + (c2 − c1 )(µ − e−θ )].
2
By perturbation theory [17] it is readily seen that the two perturbed roots of (12) are given by
χi (µ) = (µ − eθ )(µ − e−θ ) +
µi1 = eθ + µi11 ε + o(ε), µi2 = e−θ + µi21 ε + o(ε).
(12)
(13)
Substituting µ = µi1 into χi (µ) = 0 and comparing the coefficient of ε yield
1
µi11 (2 sin θ) + λi (c2 − c1 )(2 sin θ) = 0
2
and µi11 = − 12 λi (c2 − c1 ) follows immediately. Direct computation shows that |µi1 |2 = 1 + 2Re(µi11 e−θ )ε + o(ε), where
1
Re(µi11 e−θ ) = − [ai (c2 cos θ − c1 sin θ) + bi (c1 cos θ + c2 sin θ)]
2
if we let λi = ai + bi . Clearly |µi1 | = 1 − 12 (Reλi )(c2 cos θ − c1 sin θ)ε + o(ε) when c1 cos θ + c2 sin θ = 0. Similarly we
can show µi21 = − 21 λi (c2 + c1 ) and |µi2 | = 1 − 12 Reλi (c2 cos θ − c1 sin θ)ε + o(ε), which implies the conclusion.
2). Here we need to compute the Jordan decomposition of Ai . The eigenvector corresponding to the eigenvalue µi1 is
given by wi1 = wi10 + wi11 ε + o(ε). Substituting it into the equation Ai wi1 = µi1 wi1 and comparing the coefficients of
constant term, we have Awi10 = eθ wi10 . With the normalization condition v T wi1 = 1 where v T = 21 (1 − ), wi10 = (1 )T .
Similarly, the eigenvector corresponding
to the eigenvalue µi2 is given by wi2 = wi20 + wi21 ε + o(ε) with wi20 = (1 −
1 + O(ε) 1 + O(ε)
− + O(ε) −1 + O(ε)
T
) . Letting Ri = (wi1 wi2 ) =
, it is clear that Ri−1 = det1Ri
=
+
O(ε)
−
+
O(ε)
−
+ O(ε) 1 + O(ε)
µi1
1 + O(ε) − + O(ε)
1
Ri−1 .
. The result follows directly by noticing that Ai = Ri
2
µi2
1 + O(ε) + O(ε)
Remark 3.3: Denote ρ = max ρi and let h be a constant in (0, Reλ2 ]. Taking c1 = − sin θ/h and c2 = cos θ/h, we have
i=2,...,l
ρ ≤ 1 − ε + o(ε) < 1 − ε/2 with sufficiently small ε.
We also need to define some constants as follows:
1
C0 = 12 |c1 | + 23 |c2 csc θ| = 2h
(| sin θ| + 3| cot θ|),
Λ = maxi=2,...,l |λi |,
C(1) = ||UL−1 || + 2C0 Λ||UL ||,
C(k) = ||UL−1 || + 2C0 (Λ + 1)||UL || + 10(|c1 | + |c2 |)C(k − 1), k ≥ 2,
C̄ = 5(|c1 | + |c2 |)C(Nmax ) + C0 ||UL ||,
(14)
where Nmax = maxi=2,...,l Nl .
Lemma 3.2: Let γ = 1 − ε/4. Then we can choose sufficiently small ε to satisfy the following inequalities:
(Λ + 1)C̄ε ≤ 12 γ| csc θ|||UL ||;
+ γ1 ) ≤ 2| cos θ| + 1 + 21 ;
(N − 1)C̄(Λ + 1)ε ≤ 14 | csc θ|γ 3 .
1
γ (2| cos θ|
(15a)
(15b)
(15c)
Theorem 3.1: Take cj ’s as in Remark 3.3 and let γ = 1 − ε/4. Select sufficiently small ε to satisfy Lemmas 3.1 and 3.2
with ρ < 1 − ε/2. Then under Assumptions 3.1 and 3.2, consensus can be achieved at a convergence rate of O(γ t ) provided
4
C ∗ , Cδ∗ } and M (t) satisfies
that g0 ≥ max{ 3γ
M (t) ≥ 1, t = 1, 2;
(16)
M (t) ≥ | cos θ| + 1/2, t = 2m + 1, . . . .
Therefore, the number of bits used to achieve the consensus is ⌈log2 2⌈| cos θ| + 1/2⌉⌉.
Proof: 1) Preparation. The closed-loop system of disagreement vectors can be established as
δ1 (t)
0
δ1 (t + 1)
cos θIN
sin θIN
+
=
δ2 (t)
− sin θIN cos θIN
u(t)
δ2 (t + 1)
with
u(t) =
0, t = 0, 1;
P2
− j=1 kj L(δj (t) + ej (t)), t ≥ 2,
(17)
by noticing (9) and L = L(IN − φ1 ψ1T ). Letting δ̃j (t) = UL−1 δj (t) = [δ̃1j (t), . . . , δ̃N,j (t)]T , we obtain
0
cos θIN
sin θIN
δ̃1 (t)
δ̃1 (t + 1)
+
,
=
η(t)
− sin θIN − k1 LJ cos θIN − k2 LJ
δ̃2 (t)
δ̃2 (t + 1)
where η(t) = −LJ UL−1 (k1 e1 (t) + k2 e2 (t)). Denote δ̃ i (t) = [δ̃i1 (t), δ̃i2 (t)]T for i = 1, . . . , N . Clearly δ̃ 1 (t) ≡ 0 due to that
δ̃1j (t) = ψ1T (IN − φ1 ψ1T )xj (t) = 0 for j = 1, 2. Without loss of generality we assume that N2 = 2 (the Jordan block with
respect to λ2 is two-dimensional) and consequently δ̃ 2 (t) and δ̃ 3 (t) are coupled in the following way:
δ̃ i (t + 1) = Aδ̃ i (t), t = 0, 1, i = 2, 3;
δ̃ 2 (t + 1) = A2 δ̃ 2 (t) − K δ̃ 3 (t) − η2 (t),
δ̃ 3 (t + 1) = A2 δ̃ 3 (t) − η3 (t), t ≥ 2,
(18)
where A2 and K have been defined in Lemma 3.1 and η2 (t) = [0, (λ2 ψ2T +ψ3T )(k1 e1 (t)+k2 e2 (t))]T , η3 (t) = [0, λ2 ψ3T (k1 e1 (t)+
k2 e2 (t))]T .
2) Estimation error and exponential convergence. Remember that e2 (t) = cot θe1 (t) − csc θe1 (t − 1) − u(t − 1) is dependent
on the control input by (8), we have to first make an estimate for u(t) before establishing the consensus result. Below we shall
show |ψiT u(t)| ≤ ε(Λ + 1)C̄p0 γ t−2 , t ≥ 2 for i ≥ 2 by induction.
With the choice of p0 and γ it is easy to see |si (t)| ≤ 3/2 when t ≤ 2 by noticing |yi (t)| ≤ 2C ∗ , hence we obtain
max ||∆(t)|| ≤ 1/2 if M (1), M (2) ≥ 1. For i = 2, we have ψ2T L = λ2 ψ2T + ψ3T and as a result
t=1,2
P2
|ψ2T u(2)| = |ψ2T L j=1 kj (δj (2) + ej (2))|
≤ ε(Λ + 1)||UL || |c1 |(2||δ(0)|| + p(1)||∆(2)||)
+ |c2 |(2||δ(0)|| + | cot θ|p(1)||∆(2)|| + | csc θ|p(0)||∆(1)||)
≤ ε(Λ + 1)||UL ||(|c1 | + |c2 |)(2||δ(0)|| + | csc θ|p0 )
≤ ε(Λ + 1)||UL ||(|c1 | + |c2 |)(2 + | csc θ|)p0
≤ ε(Λ + 1)C̄p0 ,
which also holds for |ψiT u(2)| for i > 2.
Now assume that
|ψiT u(t)| ≤ε(Λ + 1)C̄p0 γ t−2 , t ≥ 2;
||∆(τ )|| ≤ 1/2(⇒ |ei1 (τ )| ≤ 21 p0 γ τ −1 ), 1 ≤ τ ≤ t.
(19)
Then by combining (19) and (15a) it follows that
||η3 (τ )|| = |λ2 ||ψ3T (k1 e1 (τ ) + k2 e2 (τ ))|
≤ ε|λ2 |p0 γ τ −2 ||UL ||( 12 |c1 | + |c2 || csc θ|) + ε|λ2 |c2 ||ψ3T u(τ − 1)|
≤ ε|λ2 |||UL ||C0 p0 γ τ −2 .
Recalling (18) we get that for t ≥ 2
3
δ̃ 3 (t + 1) = At−1
2 δ̃ (2) −
which produces the following estimate by Lemma 3.1 and (20)
Pt−1
τ =1
(20)
At−1−τ
η3 (τ + 1),
2
3
t−1
||δ̃ 3 (t + 1)|| ≤ 25 (ρt−1
)
2 ||δ̃ (2)|| + 4|λ2 |C0 ||UL ||p0 γ
−1
t−1
≤ 5γ (||UL ||||δ(0)|| + 2|λ2 |C0 ||UL ||p0 )
≤ 5C(1)p0 γ t−1 .
(21)
Similarly, an estimate for ||δ̃2 (t + 1)|| can be found as ||δ̃ 2 (t + 1)|| ≤ 5C(2)p0 γ t−1 , if we notice that ||K δ̃ 3 (τ )|| ≤ 5ε(|c1 | +
|c2 |)C(1)p0 γ τ −2 and ||η2 (τ )|| ≤ ε(|λ2 | + 1)||UL ||C0 p0 γ τ −2 for 2 ≤ τ ≤ t. For any i ≥ 2, by proceeding along the same line
as in the above it is concluded that
||δ̃ i (t + 1)|| ≤ 5C(Nmax )p0 γ t−1 .
(22)
P
2
3) Data rate. Now we are able to discuss the estimation for |ψiT u(t+1)|, which is bounded by the sum of | j=1 kj ψiT Lδj (t+
P2
1)| and | j=1 kj ψiT Lej (t + 1)|. For the first term, by (22) it is readily seen that
P2
P2
| j=1 kj ψiT Lδj (t + 1)| ≤ | j=1 kj ψiT UL LJ δ̃j (t + 1)|
(23)
≤ 5ε(|c1 | + |c2 |)C(Nmax )p0 γ t−1 ;
while the second term is essentially related with ej (t + 1), or more exactly ∆(t + 1). By (7) and (8) we have
d(t + 1) =
=
1
p(t) (− cos θe1 (t) − sin θe2 (t))
1
−2 cos θ
∆(t) + γ12 ∆(t − 1) + p(t)
γ
sin θu(t − 1),
(24)
which is obviously dependent on the previous quantization errors ∆(t) and ∆(t − 1), as well as the previous control input
u(t − 1). Hence with the induction assumption (19) the quantizer can be made unsaturated with sufficiently many bits at time
t + 1, and ||∆(t + 1)|| ≤ 1/2 follows directly. Consequently
P2
| j=1 kj ψiT Lej (t + 1)| ≤ ε|λ2 |||UL ||C0 p0 γ t−1
(25)
as in (20). The induction is then established by combining (23). Moreover, by (22) the consensus can be achieved at a
convergence rate of O(γ t ).
Below we are to calculate the number of required quantization levels at each time step. The situation when t ≤ 2 has been
discussed. When t > 2, from (24) we can see that
||d(t)|| ≤
≤
≤
by noticing (15b), (15c) and u(t) =
PN
i=2
| sin θ|
1
1
2γ (2| cos θ| + γ ) + p(t−1) ||u(t − 2)||
| sin θ|
1
1
2 (2| cos θ| + 1) + 4 + γ 3 (N − 1)(Λ
1
1
2 (2| cos θ| + 1) + 2
+ 1)C̄
φi ψiT u(t) (ψ1T u(t) = 0). In summary, the proof is completed.
Remark 3.4: For the coupling system shown in (18), we divide it into two subsystems with disturbance. Each subsystem
can be stabilized as long as the disturbance decays exponentially at a speed slower than ρ2 , i.e. ||η3 (t)|| ∼ O(γ t ) and
||K δ̃ 3 (t) + η2 (t)|| ∼ O(γ t ), with ρ2 < γ < 1. The interference of u(t) in the estimation error e(t) can be ignored, as long
as ||u(t)|| ∼ O(εα )p0 γ t with α > 0, yielding that ||η3 (t)|| ∼ O(ε)p0 γ t , and then ||δ̃ 3 (t)|| ∼ O(γ t ) follows. As a result,
||K δ̃ 3 (t)|| ∼ O(ε)γ t and ||δ̃ 3 (t)|| ∼ O(γ t ) follows by combining ||η3 (t)|| ∼ O(ε)p0 γ t . Such a reasoning still applies when
(18) involves more than two subsystems. Finally we show that ||u(t)|| ∼ O(ε)p0 γ t , and by (24) we conclude that the control
input does not consume extra bits in exchanging the information when the control gains are sufficiently small.
IV. H IGHER - ORDER
CASES
In this section, we will conduct the same task as in the last section for general higher-order cases. The analysis actually
proceeds along a similar line, but the assignment of control gains to achieve consensus is much more challenging, and we
have to resort to combinatorial identities for an explicit data rate. As before, we first provide an encoding-decoding scheme
for all the agents and devise a control protocol in terms of the outputs of the scheme. Then we present some lemmas, which
will play a crucial role in the convergence analysis and the derivation of the data rate in the final part.
A. Encoding-decoding scheme and distributed control law
As pointed out in the last section, the construction of the encoding scheme should follow two principles: firstly, the encoder
is able to estimate other state variables given that only the first component is measurable; secondly, the estimation should be
based on iterations in an effort to reduce quantization levels. Such an idea can be stated more clearly as follows. At each time
step, the scaled difference between the output yi (t) and its estimate is quantized to obtain a signal si (t). Based on si (t) we
construct an estimate x̂i1 (t) of the first component xi1 (t), and combine previous estimates x̂i1 (t − 1) through x̂i1 (t − 2m + 1)
to obtain estimates of the other components xi2 (t) through
xi,2m (t).
I2m (1, ·)
A(1, ·)
To be detailed, denote the observability matrix O =
,
..
.
2m−1
A
(1, ·)
x̄i (t) = [xi1 (t − 2m + 1), xi1 (t − 2m + 2), . . . , xi1 (t)]T ,
bn (θ) = [0, . . . , 0, A(1, 2m), . . . , An−1 (1, 2m)]T ∈ R2m .
We have
x̄i (t) = Oxi (t − 2m + 1) +
if we notice by (1) that
xi (t − 2m + k) = Ak−1 xi (t − 2m + 1) +
k = 1, . . . , 2m.
2m−1
X
n=1
k−2
P
n=0
As a result,
xi (t − 2m + 1) = O−1 [x̄i (t) −
and
bn (θ)ui (t − n)
Ak−2−n bui (t − 2m + 1 + n),
2m−1
X
n=1
bn (θ)ui (t − n)]
P2m−2
xi (t) = A2m−1 xi (t − 2m + 1) + j=0 A2m−2−j bui (t − 2m + 1 + j)
P2m−1
= S x̄i (t) + j=1
b̃j (θ)ui (t − j),
(26)
(27)
(28)
(29)
where S = A2m−1 O−1 (the existence of O−1 can be easily verified by PBH test [18] if sin θ 6= 0) and b̃n (θ) = −Sbn (θ) +
An−1 (·, 2m). Inspired by (29), the encoding scheme for agent i is implemented below:
for t ≤ 2m,
yi (t)
), x̂i1 (t) = p(t − 1)si (t);
si (t) = qt ( p(t−1)
x̂ (2m)
x̂i1 (1)
i2
(30)
..
..
= Sm
;
.
.
x̂i,2m (2m)
x̂i1 (2m)
for t > 2m,
θ x̂i2 (t−1)+x̂i3 (t−1)]
),
si (t) = qt ( yi (t)−[cos θx̂i1 (t−1)+sin
p(t−1)
x̂
(t)
=
cos
θx̂
(t
−
1)
+
sin
θx̂
(t
−
1)
+
x̂
(t
−
1) + p(t − 1)si (t),
i1
i1
i2
i3
x̂i1 (t − 2m + 1)
x̂i2 (t)
..
..
,
= Sm
.
.
x̂i1 (t)
x̂i,2m (t)
(31)
where Sm = S(2 : 2m, ·) is a submatrix of S obtained by deleting the first row, and p(t) = p0 γ t , 0 < γ < 1 is a decaying
scaling function.
After si (t) is generated, transmitted and received by one of agent i’s out-neighbors, say v ∈ Ni− , a decoder will be activated:
for t ≤ 2m,
x̂
iv1 (t) = p(t − 1)si (t);
x̂iv1 (1)
x̂iv2 (2m)
(32)
..
..
;
= Sm
.
.
x̂iv1 (2m)
x̂iv,2m (2m)
for t > 2m,
x̂iv1 (t) = cosθx̂iv1 (t −
+ x̂iv3 (t − 1) + p(t − 1)si (t);
1) + sin θx̂iv2 (t − 1)
x̂iv1 (t − 2m + 1)
x̂iv2 (t)
..
..
.
= Sm
.
.
x̂iv1 (t)
x̂iv,2m (t)
(33)
Remark 4.1: Comparing (30) with (32), (31) with (33), it is clear that x̂ivj (t) ≡ x̂ij (t), j = 1, . . . , 2m, for v ∈ Ni− , i =
1, . . . , N . Denote eij (t) = x̂ij (t) − xij (t) as the estimation error, ∆i (t) = si (t) − di (t) as the quantization error, where
( y (t)
i
p(t−1) , t = 1, 2, . . . , 2m;
di (t) =
(34)
yi (t)−[cos θ x̂i1 (t−1)+sin θ x̂i2 (t−1)+x̂i3 (t−1)]
, t > 2m.
p(t−1)
Comparing (29) with (31), the estimation errors are given by the following:
− 1)∆i (t), t ≥ 1;
ei1 (t) = p(t
P2m−1
P2m
eij (t) = n=1 S(j, n)ei1 (t − 2m + n) − n=1 b̃nj (θ)ui (t − n),
t ≥ 2m, j = 2, . . . , 2m,
(35)
where b̃nj is the j-th entry of b̃n .
Remark 4.2: The encoding schemes (5) and (31) proposed in our work is different from those in [10] or [15]. Actually, to
address the general dynamics with unmeasurable states, [10] designed the encoding scheme respectively for the output and
control input, and used Luenberger observer to estimate the unmeasurable states. If we compare with [15], we can also see a
big difference: the special structure of n-th order integrator dynamics enables it to easily “recover” the control input at n steps
earlier, based on which an estimate of the unmeasurable components can be made with time delay, and the encoding scheme
can be designed accordingly. However, in our case it is unlikely to achieve the same task and we resort to the constructability
of the system, namely we estimate the unmeasurable states directly from x̂i1 (t) through x̂i1 (t − 2m + 1). Although such a
method introduces the control input into the estimation errors, it is able to make an estimation without time delay, and hence
avoids the stabilization of a time-delayed closed-loop system in the consensus analysis.
For agent i, the outputs of encoder are x̂i1 (t), . . . , x̂i,2m (t), while the outputs of decoders are x̂vi1 (t), . . . , x̂vi,2m (t) for
v ∈ Ni+ . Based on these outputs, the distributed control law of agent i is proposed as
(
0, t = 0, 1, . . . , 2m − 1;
P
P2m
ui (t) =
(36)
j=1 kj
v∈N + giv [x̂vij (t) − x̂ij (t)], t ≥ 2m.
i
B. Lemmas
The following two lemmas are respectively needed in analyzing consensus and data rate. The first one is to stabilize the
closed-loop system of disagreements, and the second one is used for estimating the magnitude of ui (t) and di (t).
Lemma 4.1: Denote Ai = A − λi K with λi > 0, where K ∈ R2m×2m and its nonzero entries are only at the last row
[k1 , k2 , . . . , k2m−1 , k2m ]. Take
c2j−1 εm−j , j = 1, . . . , m − 1;
k2j−1 =
ε, j = m,
c2j−1m−j
(37)
c2j ε
, j = 1, . . . , m − 1;
k2j =
c2j ε, j = m.
Then we can find constants c2j−1 and c2j (j = 1, . . . , m) such that, when ε is sufficiently small, the spectral radius ρi of Ai
is less than 1 with distinct eigenvalues. Moreover, denote
Rm =
c2m−5 − c2m−4 −θ
1 1
e ].
+ (c2m−1 sin θ − c2m cos θ), H = Re[
2 2
c2m−3 − c2m−2
(38)
The requirements about c2j−1 ’s and c2j ’s corresponding to different m’s are listed below.
1). m = 2: let c1 = − sin 2θ and c2 = cos 2θ. If R2 < 0, then ρi = 1 + 12 λi R2 ε + o(ε);
2). m ≥ 3: let c2m−3 = − sin 2θ and c2m−2 = cos 2θ. If λi Rm + H < 0 and Re(ϑn1 e−θ ) < 0 with ϑn1 , n = 3, . . . , m
denoting the m − 2 distinct roots of the equation
ϑm−2
(c2m−2 − c2m−3 ) + · · · + ϑ1 (c4 − c3 ) + (c2 − c1 ) = 0,
1
1
2
(39)
−θ
then ρi = 1 + maxn=3,...,n {λi Rm + H, 2Re(ϑn1 e )}ε + o(ε).
Lemma 4.2: Assume that Lemma 4.1 holds. When ε is sufficiently small, for any vector ξ ∈ R2m , the entries of Asi ξ, which
are denoted as ξs,2j−1 and ξs,2j , j = 1, . . . , m, satisfy that
||ξ||Mij ρsi εj−(m−1) , j = 1, 2, . . . , m − 2;
|ξs,2j−1 |, |ξs,2j | ≤
(40)
||ξ||Mij ρsi ε(j−m)/2 , j = m − 1, m,
where
Mij =
Pm
5
2λi (
Pm
5
2λi (
j−1
Q|ϑn1 |
|ϑk1 −ϑn1 | ),
n=3
3≤k≤m,k6=n
n=3
m−3
Q|ϑn1 |
|ϑk1 −ϑn1 |
3≤k≤m,k6=n
j = 1, . . . , m − 3;
+ 1), j = m − 2;
√3 ,
2λi
j = m − 1;
5/2, j = m.
Remark 4.3: The proofs of the above lemmas can be found in the Appendix. As in [15], the basic idea is to combine
the bifurcation analysis of the roots of characteristic polynomials and the Jordan basis of a perturbed matrix [19]. However,
the situation here is much different. On one hand, the complex conjugate eigenvalues of the original matrix A complicates
the analysis of the perturbed eigenvalues, as seen from the proof of Lemma 4.1. On the other hand, unlike [15] where the
unperturbed matrix admits multiple eigenvalues of 0 and 1, the unperturbed matrix here admits eigenvalues of m identical pairs
of complex conjugate numbers, which allows a less cumbersome calculation of the perturbed Jordan basis, as in the proof of
Lemma 4.2.
Remark 4.4: Assume 0 < λ2 ≤ · · · ≤ λN and let ρ = max ρi , h ∈ (0, λ2 ]. Given c2m−3 = − sin 2θ and c2m−2 = cos 2θ,
i=2,...,N
the other constants c2j−1 and c2j can be selected as follows such that ρ ≤ 1 − ε + o(ε) < 1 − ε/2 holds with sufficiently small
ε:
1). m = 2: select c3 = −(4/h + 1) sin θ, c4 = (4/h + 1) cos θ such that R2 = −2/h < 0;
2). m ≥ 3: first select c2m−4 , c2m−5 , · · · , c1 such that the solutions of (39) are given by ϑn1 = −(n − 2)eθ , n = 3, . . . , m
and H is determined by (38). In fact, direct computation shows that c2m−4 − c2m−5 = 21 (m − 1)(m − 2)e3θ and consequently
H = (m − 1)(m − 2)/2 > 0. Now let c2m−1 = −[(2H + 4)/h + 1] sin θ, c2m = [(2H + 4)/h + 1] cos θ such that
λi Rm + H = H(1 − λi /h) − 2λi /h < −2λi /h ≤ −2.
P
j−1
Qn
With such a selection, Mij = 2λ5 i ( m−2
n=1
|k−n| ), j = 1, . . . , m−3 and Mi,m−2 =
1≤k≤m−2,k6=n
1).
To explicitly express the data rate, another lemma is required.
Lemma 4.3: Denote
Pm−2
5
2λi (
m−3
Qn
n=1
1≤k≤m−2,k6=n
|k−n| +
l(θ) = [l0 (θ), l1 (θ), . . . , l2m−2 (θ), l2m−1 (θ)] = cos θS(1, ·) + sin θS(2, ·) + S(3, ·).
Then
lk (θ) = (−1)k−1
⌊k/2⌋
P
h=0
C(m, k − 2h)C(m − (k − 2h), h)(2 cos θ)k−2h , k = 0, 1, . . . , 2m − 1.
P2m−1
Moreover, k=0 |lk (θ)| = [2(1 + | cos θ|)]m − 1.
The proof of Lemma 4.3 can be found in the Appendix. The idea is simple and direct: by viewing l(θ) as the solution to
an equation, we are left to the verification when l(θ) takes the form in the lemma. Still, the computation is quite complicated
and requires special techniques from combinatorics [20], [21].
C. Convergence analysis and data rate
The notations in (10) will still be used, except that ψ1 is replaced by φ1 . The following assumptions are adopted in the
subsequent analysis.
Assumption 4.1: The communication graph G is undirected and connected.
Assumption 4.2: There exist known positive constants C ∗ and Cδ∗ such that max ||xj (0)|| ≤ C ∗ and max ||δj (0)|| ≤
j=1,...,2m
j=1,...,2m
Cδ∗ .
Remark 4.5: Assumption 4.1 is a standard assumption, under which the eigenvalues of L can be rearranged as 0 = λ1 <
λ2 ≤ · · · ≤ λN . The reason that we only consider the undirected graph will be clarified in Remark 4.7. Assumption 4.2 serves
the same purpose as Assumption 3.2.
We also need the following constants:
b∗ = max{|b̃lj |}, c∗ = max {|cj |},
l,j
1/2
3/2
Λi = max{λi , λi }, C̄ =
1≤j≤2m
√9 [||U −1 ||
L
2
+ 5c∗ mN (||S|| + 2)].
Lemma 4.4: Let γ = 1 − ε/4. Then we can choose sufficiently small ε to satisfy the following inequalities:
P
2c∗ j6=m−1 Mij ε1/2 ≤ Mi,m−1 , i = 2, . . . , N ;
P2m−1
2m
≤ [2(1 + | cos θ|)]m − 1/2;
k=0 |lk (θ)|/γ
∗
(2m − 1)b (N − 1)ΛC̄ε1/2 ≤ 81 γ 4m−1 ,
(41)
(42a)
(42b)
(42c)
where Λ = maxi Λi .
Theorem 4.1: Take kj ’s as in (37), cj ’s as in Remark 4.4 and γ = 1 − ε/4. Select sufficiently small ε to satisfy Lemma 4.4
and ρi < 1 − ε/2 for i = 2, . . . , N . Then under Assumptions 4.1 and 4.2, consensus can be achieved at a convergence rate of
O(γ t ) provided that M (t) satisfies
M (t) ≥ 1, t = 1, . . . , 2m;
(43)
M (t) ≥ 2m−1 (1 + | cos θ|)m − 12 , t = 2m + 1, . . . ,
√
and p0 ≥ ( 2 + 1)2m max{C ∗ , Cδ∗ }.
Therefore, we can use ⌈log2 2⌈2m−1 (1 + | cos θ|)m − 12 ⌉⌉ bits of information exchange to achieve the consensus.
Proof: 1) Preparation. By (36) we have
0, t = 0, 1, . . . , 2m − 1;
2m
P
(44)
u(t) =
kj L(δj (t) + ej (t)), t ≥ 2m.
−
j=1
Direct computation shows
δ1 (t + 1)
δ1 (t)
0
..
..
..
= (A ⊗ IN )
+ .
.
.
δ2m (t + 1)
δ2m (t)
u(t)
Let δ̃j (t) = UL−1 δj (t) = [δ̃1j (t), . . . , δ̃N,j (t)]T and δ̃ i (t) = [δ̃i1 (t), . . . , δ̃i,2m (t)]T . Then we obtain δ̃ 1 (t) ≡ 0, and for
i = 2, . . . , N
Aδ̃ i (t),
t = 0, 1 . . . , 2m − 1;
i
(45)
δ̃ (t + 1) =
Ai δ̃ i (t) − ǫi (t), t ≥ 2m,
P2m
where ǫi (t) = [0, . . . , 0, j=1 kj λi φTi ej (t)]T ∈ R2m .
2) Estimation error and exponential convergence. To analyze the influence of u(t) on the error term ej (t), we will show
|φTi u(t)| ≤ Λi C̄p0 γ t−2m ε1/2 , t ≥ 2m by induction.
√
2m ∗
C , hence we
With the choice of p0 and γ it’s easy to see |si (t)| ≤ 3/2 when t ≤ 2m by noticing |y√
i (t)| ≤ ( 2 + 1)
2m ∗
obtain max ||∆(t)|| ≤ 1/2 provided M (t) ≥ 1, t = 1, . . . , 2m. Moreover, ||δj (2m)|| ≤ ( 2 + 1) Cδ . Recalling (36) and
1≤t≤2m
ej (2m) ≤ p0 ||S|| max ||∆(t)|| we have
1≤t≤2m
P2m
(2m) + ej (2m))|
|φTi u(2m)| = |φTi
j=1 kj L(δ
√ j
≤ 2λi mN c∗ ε(( 2 + 1)2m Cδ∗ + p0 ||S||/2)
≤ 2λi mN c∗ ε(1 + ||S||/2)p0
≤ Λi C̄p0 ε1/2
√
by noticing p0 ≥ ( 2 + 1)2m Cδ∗ and ε < 1.
Assume that
we have
and
|φTi u(τ )| ≤ Λi C̄p0 γ τ −2m ε1/2 , 2m ≤ τ ≤ t;
||∆(τ )|| ≤ 12 (⇒ |ei1 (τ )| ≤ 21 p0 γ τ −1 ), 1 ≤ τ ≤ t.
S(j, 1)
e11 (τ − 2m + 1) . . . e11 (τ )
P2m−1
..
..
.
..
ej (τ ) =
− n=1 b̃nj (θ)u(τ − n),
. ..
.
.
S(j, 2m)
eN 1 (τ − 2m + 1) . . . eN 1 (τ )
|φTi ej (τ )| ≤ N p0
||ǫi (τ )||
max
τ −2m+1≤s≤τ
≤ λi c∗ ε2mN [p0
≤
(46)
||∆(s)||γ τ −2m ||S|| + N (2m − 1)b∗ λi ε1/2 C̄p0 γ τ −4m+1
max
||∆(s)||γ τ −2m ||S|| + (2m − 1)b∗ λi ε1/2 C̄p0 γ τ −4m+1 ]
τ −2m+1≤s≤τ
2λi c∗ mN (||S||/2 + 1)p0 γ τ −2mε
(47)
by (42c), if ||∆(s)|| ≤ 1/2 for τ − 2m + 1 ≤ s ≤ τ . Recalling (45) we obtain
i
δ̃ (t + 1) =
At+1−2m
δ̃ i (2m)
i
−
t−2m
X
At−2m−τ
ǫi (2m + τ ).
i
τ =0
By applying Lemma 4.2 and taking into account (47), it yields that
|
δ̃i,2j−1 (t + 1)|, |δ̃i,2j (t + 1)|
Mij εj−(m−1) γ t+1−2m ||δ̃ i (2m)|| + 4λi c∗ mN (||S|| + 2)p0 ,
j = 1, . . . , m − 2;
≤
M ε(j−m)/2 γ t+1−2m ||δ̃ i (2m)|| + 4λi c∗ mN (||S|| + 2)p0 ,
ij
j = m − 1, m,
(48)
due to ε/(γ − ρi ) < 4.
P
P2m
T
T
With (48) it is ready to estimate φTi u(t + 1), which is a sum of 2m
j=1 kj φi Lδj (t + 1) and
j=1 kj φi ej (t + 1). For the
−1
i
first part, by (48) and ||δ̃ (2m)|| ≤ ||UL ||||δi (2m)|| we have
P2m
| j=1 kj φTi Lδj (t + 1)|
P
= λi | 2m
j=1 kj δ̃ij (t + 1)|
P
(49)
≤ λi p0 γ t−2m+1 (2Mi,m−1 ε1/2 + 2c∗ ε j6=m−1 Mij )·
i
∗
[||δ̃ (2m)|| + 4λi c mN (||S|| + 2)p0 ]
≤ 3λi Mi,m−1 [||UL−1 || + 4λi c∗ mN (||S|| + 2)]p0 γ t−2m+1 ε1/2
if we note that |c2m−2 |, |c2m−3 | ≤ 1. For the second part, as in the second order case, it is closely related with ||∆(t + 1)|| and
similarly it can be inferred from (34) that d(t + 1) is only dependent on the past quantization errors ∆(τ ), t − 2m + 1 ≤ τ ≤ t
and the past control inputs u(τ ), t − 2m + 1 ≤ τ ≤ t − 1. Hence with the induction assumption (46) the quantizer can be made
unsaturated at time t + 1 with finite bits, namely ||∆(t + 1)|| ≤ 1/2. In consequence we get an estimation similar to (47) that
P2m
(50)
| j=1 kj φTi Lej (t + 1)| ≤ 2λi c∗ mN (||S||/2 + 1)p0 γ t−2m+1 ε.
Combining (49) and (50), it is clear that |φTi u(t + 1)| ≤ Λi C̄p0 γ t−2m+1 ε1/2 , which establishes the induction. Furthermore,
by (48) clearly the consensus can be achieved at a convergence rate of O(γ t ).
3) Data rate. Below we are to discuss the number of quantization levels at each time step. The situation when t ≤ 2m has
been discussed. When t > 2m, we have
||d(t)||
θ
≤ || cos
γ ∆(t − 1) +
j=1
sin θS(2,j)+S(3,j)
∆(t
γ 2m−j+1
− 1 − 2m + j)|| + 2b∗
1
γ 2m−j+1 [cos θS(1, j) + sin θS(2, j) + S(3, j)]∆(t − 1
j=1
P2m−1 ||u(t−1−j)||
+2b∗ j=1
p(t−1)
P
2m−1
1
|l
(θ)|
+ 2b∗ (2m − 1)(N − 1)ΛC̄γ 1−4m ε1/2
k
k=0
2γ 2m
= ||
≤
2m
P
2m
P
P2m−1
j=1
||u(t−1−j)||
p(t−1)
− 2m + j)||
PN
by noticing S(1, ·) = [0, . . . , 0, 1] and u(t) = i=1 φi φTi u(t), Λ1 = 0. By taking into account (42b) and (42c) it can be seen
that ||d(t)|| is bounded by 2m−1 (1 + | cos θ|)m and the proof is completed by remembering (43).
Fig. 1.
Communication topology
Q
Remark 4.6: Noticing that 1≤k≤m−2,k6=n |k−n| = (n−1)!(m−2−n)! attains the minimum at n = ⌊ m−2
2 ⌋ and multiplying
by a positive λi on both sides does not change the direction of an inequality, (42a) can be substituted by the following stronger
one, which is easier to check:
q
Pm−2 Pm−2
λ2
nj−1
1/2
(51)
+
1
+
λ
)ε
≤
3
5c∗ ( j=1
m
m
N
n=1 (⌊ ⌋−2)!(m−1−⌊ ⌋)!
2 .
2
2
Remark 4.7: From the proof it is readily seen that we can still use the same number of bits to achieve the quantized consensus
once the Laplacian of the directed topology satisfies that 0 < λ2 ≤ · · · ≤ λN . However, unlike the case of the 2nd-order
oscillator, it does not hold for the general topology, when the Laplacian contains complex eigenvalues, or real Jordan blocks of
multiple dimensions. For one reason, note that Lemma 4.1 does not hold for a complex λi . For another one, note the disparity
in the order of ε between the disturbance term and the weighted sum of disagreement entries, i.e. ||ǫi (t)|| ∼ O(ε)p0 γ t and
||K δ̃ i (t)|| ∼ O(ε1/2 )p0 γ t . Therefore, if we assume m = 2 and the Jordan block corresponding to λ2 > 0 is two-dimensional
as in (18), then it follows from ||K δ̃ 3 (t)|| ∼ O(ε1/2 )p0 γ t that ||δ̃ 2 (t)|| ∼ O(ε−1 )p0 γ t and ||u(t)|| ∼ O(1)p0 γ t , suggesting
that the input term can no longer be neglected in the estimation errors, nor in the quantization input d(t). Such a situation is
also encountered in [15].
Remark 4.8: At the first glance it may seem doubtful that the data rate is dependent on | cos θ|; but a little further inspection
is enough to clarify. Similar to the situation of the n-th order integrator system investigated in [15], the control input does
not consume any bit in exchanging the estimates of the states when ε is sufficiently small. In other words, we only need to
focus on how many bits it needs to estimate the output of an individual open loop system. Take the second-order case as an
example. Noticing that yi (t) = cos θxi1 (t − 1) + sin θxi2 (t − 1) = 2 cos θxi1 (t − 1) − xi1 (t − 2), we can estimate yi (t) based
on x̂i1 (t − 1) and x̂i1 (t − 2) with an error bound no larger than 12 (2| cos θ| + 1) + 21 . Generally speaking, when | cos θ| ≈ 0
or equivalently | sin θ| ≈ 1, xi,2j−1 (t) and xi,2j (t) are tightly coupled, and it needs only m bits of information exchange
to achieve the consensus; in the case of | cos θ| ≈ 1, after rearranging of states A can be approximated by I2 ⊗ J1,m , and
2m bits are sufficient. Anyway, for a 2m-th order system studied in this paper, 2m bits are enough to realize the consensus
asymptotically, which is consistent with the conclusion for n-th order integrator systems [15].
V. N UMERICAL E XAMPLE
For simplicity we only show an example of m = 2. Consider a 5-node network with 4-th order dynamics, where the edges
are generated randomly according to probability P (i, j) ∈ E = 0.5 with 0-1 weights. The initial states are randomly chosen
as xij (0) ∈ (0, j), i = 1, . . . , 5, j = 1, .. . , 4. Given
3 bitsof information exchange to realize the
√ it is enough
√ to use √
√ θ = π/3,
−4/(3 3) 2/ 3 −4/(3 3) 5/(3 3)
to construct the encoder and decoder
−1/3
1√
−1√
2/3
consensus, and we can compute S2 =
√
−1/ 3
1/ 3
−1/ 3
0
respectively as (30)-(33). The communication topology is generated as in Figure 1 with λ2 = 0.8299, and cj ’s are determined
as in Remark 4.4 by choosing h = λ2 . Moreover, let ε = 0.01, p0 = 10, γ = 0.9975 to satisfy the conditions in Theorem 4.1.
From Figure 2 which depicts the trajectory of δjmax (t) = {δnj (t) : n = arg maxi |δij (t)|}, we can see that the consensus is
achieved asymptotically.
VI. C ONCLUDING R EMARKS
In this paper, we explored the data rate problem for quantized consensus of a special kind of multi-agent systems. The
dynamics of each agent is described by a 2m-th order real Jordan form consisting of m pairs of conjugate poles on the unit
circle with single input, and only the first state can be measured. The encoding-decoding scheme was based on the observability
matrix. Perturbation techniques were employed in the consensus analysis and the data rate analysis, and combinatorial techniques
15
15
10
10
5
5
δ
1max
δ
2max
0
−5
−5
−10
−10
−15
0
200
400
600
800
1000
−15
1200
2
2
1.5
1.5
1
1
0.5
δ3max
0
200
400
600
800
1000
1200
0
200
400
600
800
1000
1200
0.5
δ4max
0
0
−0.5
−0.5
−1
−1
−1.5
−1.5
−2
Fig. 2.
0
0
200
400
600
800
1000
1200
−2
Trajectories of disagreements δjmax (t)
were used to explicitly obtain the data rate. The second-order case of m = 1 and higher-order cases of m ≥ 2 were investigated
separately. For the second-order case, we showed that at most 2 bits of information exchange suffice to achieve the consensus
at an exponential rate, if the communication topology has a spanning tree. For the higher-order cases, consensus was achieved
with at most 2m bits, provided that the undirected communication topology is connected. The exact number of bits for achieving
consensus in both cases is an integer which increases from m to 2m when | cos θ| increases from 0 to 1. The case of switching
directed topology is still under investigation, and noisy communication channels will be considered in the future work. As for
general unstable systems with poles outside the unit circle, perturbation techniques no longer apply and new methods need to
be developed to serve the same purpose of stabilizing the dynamics of disagreements.
A PPENDIX
Proof of Lemma 4.1 Here we mainly deal with the case of m ≥ 3, since the proof can be slightly adapted if m = 2 and the
modification will be pointed out accordingly. The characteristic equation of Ai can be computed as
χi (µ) = det[(µI − Q)m + λi Km (µI − Q)m−1 + · · · + λi K2 (µI − Q) + λi K1 ],
P
0
m−1
2
Pm−1
1
P2m
0
Fig. 3.
P
2m−1
1
P
m−2
P
P
m
2m−2
2
m−1
m
m+1
m+2
2m
Newton diagram
cos θ
where Q =
− sin θ
we rewrite χi (µ) as
sin θ
cos θ
and Kj =
0
k2j−1
0
k2j
for j = 1, . . . , m. By employing (11) in the proof of Lemma 3.1,
χi (µ) = (µ − eθ )m (µ − e−θ )m + 21 [(µ − eθ )m + (µ − e−θ )m ] λ2i ·
m
m
P
P
(k2j−1 + k2j )(µ − e−θ )j−1 ]
[ (−k2j−1 + k2j )(µ − eθ )j−1 + 2k2 +
j=2
j=2
+ 2 [(µ − eθ )m − (µ − e−θ )m ] λ2i ·
m
m
P
P
(k2j−1 − k2j )(µ − e−θ )j−1 ]
[ (k2j−1 + k2j )(µ − eθ )j−1 + 2k1 +
(52)
j=2
j=2
θ m
= (µ − e ) [(µ − e
−θ m
) +
λi
2 (k2
+ k1 )] +
λi
2 (µ
− e−θ )m
m
P
j=1
(µ − eθ )j−1 (−k2j−1 + k2j ).
With Ai being real, we only need to focus on the perturbed roots around eθ , which are denoted by µ = eθ + ∆µ. Noticing
that µ − e−θ = µ − eθ + 2 sin θ, we substitute µ = eθ + ∆µ into (52) and obtain
χi (eθ + ∆µ) =
m
X
ain (ε)(∆µ)m−n +
n=1
m
X
(C(m, m − n)(2 sin θ)m−n + O(ε))(∆µ)m+n
(53)
n=0
with the selection of k2j−1 and kj in (37), where
(2 sin θ)m−1 [2 sin θ(c2m − c2m−1 ) + m(c2m−2 − c2m−3 )]ε + o(ε), n = 1;
ain (ε) =
(2 sin θ)m (c2(m−n+1) − c2(m−n)+1 )εn−1 + o(εn−1 ), n = 2, . . . , m.
Now the Newton diagram [17] can be depicted as in Fig. 3, by first plotting points P2m−j (j, α2m−j ), j = 0, . . . , 2m and then
connecting the segments on the lower boundary of the convex hull of the above points, where α2m−j is the leading exponent
of ε in the coefficient of (∆µ)2m−j . The slopes of the two non-horizontal segments are 1/2, 1 respectively, implying that ∆µ
has the following two forms of expansions:
1
∆µ = µ1 ε 2 + µ2 εβ + o(εβ ), µ1 6= 0;
∆µ = υ1 ε + o(ε), ϑ1 =
6 0.
(54a)
(54b)
m−2 λi
m
m
Substituting (54a) into (53) and finding the coefficients of the term εm/2 , it yields that µm
1 (2 sin θ) +µ1
2 (2 sin θ) (c2m−2 −
c2m−3 ) = 0, and thus
r q
α
λi
µ1 = ±
c22m−3 + c22m−2 e 2 ,
(55)
2
where α = arg(c2m−3 − c2m−2 ). Moreover, to determine µ2 and β, we substitute (54a) into (53) again and find the lowest
order term as
mµm−1
µ2 (2 sin θ)m ε(m−1)/2 εβ
1
λi m−1
+ 2 µ1 (2 sin θ)m (c2m − c2m−1 )ε(m+1)/2
(56)
+ λ2i (m − 2)µm−3
µ2 (2 sin θ)m (c2m−2 − c2m−3 )ε(m−1)/2 εβ
1
(2 sin θ)m (c2m−4 − c2m−5 )ε(m+1)/2 = 0,
+ λ2i µm−3
1
which implies β = 1 and µ2 =
(c2m−5 −c2m−4 )
λi
4 (c2m−1 − c2m ) + 2(c2m−3 −c2m−2 ) .
1
In the form of (54a), the module of µ is determined as
|µ|2 = µµ̄ = 1 + 2Re(µ1 e−θ )ε 2 + (|µ1 |2 + 2Re(µ2 e−θ ))ε + o(ε),
(57)
with Re(µ1 e
−θ
) = ±
r
λi
2
q
c22m−3 + c22m−2 cos( α2 − θ). In order that |µ| < 1 with sufficiently small ε, we must have
− θ = π2 , and hence it suffices to let |µ1 |2 + 2Re(µ2 e−θ ) < 0. Combining these arguments gives rise to a sufficient
condition as
α
2
−c2m−3 /c2m−2 = tan(2θ + π), c22m−3 + c22m−2 6= 0;
q
−c2m−4 −θ
λi
e ] < 0.
c22m−3 + c22m−2 + λ2i (c2m−1 sin θ − c2m cos θ) + Re[ cc2m−5
2
2m−3 −c2m−2
(58b)
|µ|2 = 1 + 2Re(ϑ1 e−θ )ε + o(ε)
(59)
(58a)
With c2m−3 = − sin 2θ and c2m−2 = cos 2θ satisfying (58a), (58b) is equivalent to λi Rm + H < 0. When m = 2, µ only
takes the form of (54a) and µ2 = λ4i (c2m−1 − c2m ), leading to the sufficient condition R2 < 0 for |µ| < 1.
On the other hand, substituting (54b) into (52) and finding the coefficients of the term εm−1 , we obtain the equation (39).
Similarly, the module of µ with the form (54b) is determined by
and it suffices to let Re(ϑ1 e−θ ) to be negative such that |µ| < 1 with sufficiently small ε. For prescribed c2m−3 and c2m−2 ,
the roots of (39) can be assigned arbitrarily such that Re(ϑ1 e−θ ) < 0 with m − 2 distinct ϑ1 ; after determining c2m−4 and
c2m−5 , (58b) can always be satisfied by properly chosen c2m and c2m−1 since λ2i (c2m−1 sin θ − c2m cos θ) can be assigned
to any number. In summary, the proof is completed.
Proof of Lemma 4.2 As in the last proof, we only focus on the case of m ≥ 3 which essentially includes the case of m = 2.
For Ai , we are to find the following Jordan decomposition:
Ai = A +
m−1
X
Aij εj = Ri Ãi Ri−1 ,
(60)
j=1
where Ãi is a diagonal matrix consisting of 2m different eigenvalues determined in Lemma 4.1. To find an appropriate Ri
and the corresponding Ri−1 , we first determine the Jordan basis of the unperturbed matrix A. The Jordan chain corresponding
to the eigenvalue µ0 = eθ is given by
A−µ0 I
A−µ0 I
A−µ0 I
A−µ0 I
um−1 −−−−→ um−2 −−−−→ . . . −−−−→ u1 −−−−→ u0 ,
where uj = e2j+1 + e2j+2 , j = 0, . . . , m − 1 and en ∈ R2m denotes the vector with a 1 in the n-th coordinate and 0’s
elsewhere. Similarly, the Jordan chain corresponding to the eigenvalue µ̄0 = e−θ is given by
A−µ̄0 I
A−µ̄0 I
A−µ̄0 I
A−µ̄0 I
ūm−1 −−−−→ ūm−2 −−−−→ . . . −−−−→ ū1 −−−−→ ū0 .
1 1
Hence the two Jordan chains of A can be rearranged as R0 = (u0 ū0 . . . um−1 ūm−1 ) = Im ⊗ P with P =
.
−
With Ai being real, once we obtain the eigenvectors corresponding to the m different perturbed eigenvalues around µ0 , the
other eigenvectors can be obtained by taking conjugates. Hence we only need to find the eigenvectors corresponding to the m
different perturbed eigenvalues around µ0 .
The eigenvectors corresponding to the m perturbed eigenvalues around µ0 have the following form of Puiseux series [19]:
P∞
P∞
k/2
µin = µ0 + Pk=1 µink εk/2 , uin = uin0 +
P∞k=1 uinkkε , n = 1, 2;
∞
k
µin = µ0 + k=1 ϑink ε , uin = uin0 + k=1 uink ε , n = 3, . . . , m,
r q
α
λi
c22m−3 + c22m−2 e 2 , µi21 = −µi11 and ϑin1 = ϑn1 , n = 3, . . . , m have been defined in Lemma 4.1.
where µi11 =
2
Substituting µin , uin into the equation Ai uin = µin uin respectively, and collecting coefficients of equal powers of ε; moreover,
noticing the fact that Aij uk = 0, j = 1, . . . , m − 2; k = 0, . . . , m − 2 − j, where Aij has
been defined in (60) and imposing
T
T
the normalization condition as vm−1
uin = 1, where vm−1
= 21 1 − 0 . . . 0 is the left associated eigenvector of
T
T
T
A with respect to the eigenvalue µ0 such that vm−1
u0 = 1, vm−1
u1 = · · · = vm−1
um−1 = 0, the m eigenvectors can be
obtained as:
m−1
P k/2 k
′
ε (µi1n uk + uikn ) + o(εm/2 ), n = 1, 2;
uin = u0 +
uin = u0 +
k=1
m−1
P
k=1
′
′
′
εk (ϑk1n uk + uikn ) + o(εm ), n = 3, . . . , m,
where ui1n = 0 and uikn ∈ span{u1 , . . . , uk−1}, k = 2, . . . , m − 1 for n = 1, . . . , m.
Letting Ri = ui1 ūi1 . . . uim ūim , we are to investigate the magnitude of each entry in Ri−1 by adjoint method.
Therefore we need to find the order of det Ri and the corresponding cofactor, both of which can be expressed as Puiseux
series. The following facts should be mentioned before the calculation:
1). Determinant is a multi-linear function of column vectors, and it vanishes when two or more columns coincide.
2). There exist two types of series in the columns of Ri , and we categorize ui1 , ui2 and their conjugates for type I, the others
for type II.
With these facts, we can see that the lowest degree can be obtained by taking out terms with ε(m−2)/2 um−2 and ε(m−1)/2 um−1
respectively from ui1 and ui1 , terms with u0 , εu1 , . . . , εm−3 um−3 respectively from ui3 , . . . , uim , as well as the corresponding
m−1
2
conjugates from ūi1 , . . . , ūim , and calculated by 2(0 + 1 + · · · + m − 3 + m−2
2 + 2 ) = m − 3m + 3. Moreover,
2
m−2
m−2 m−1 2
2
m −3m+3
| det Ri | = |µm−1
(1 + o(1))
i11 µi21 − µi11 µi21 | | det V0 | | det R0 |ε
(61)
m−2 m−2
2
2 m m2 −3m+3
= |µi11 µi21 (µi11 − µi21 )| | det V0 | 2 ε
(1 + o(1)),
1 ϑ31 . . . ϑm−3
31
1 ϑ41 . . . ϑm−3
41
where V0 = V (ϑ31 , . . . , ϑm1 ) = .
is a Vandermonde matrix of order m − 2.
..
..
..
..
.
.
.
m−3
1 ϑm1 . . . ϑm1
(i)
On the other hand, we need to determinate the order of the cofactor Cs,t of the (s, t) entry, and we illustrate it by calculating
(i)
C1,1 with m = 3. After deleting the first column ui1 , we delete the first row and use the same notations u1 , u2 and e2 , . . . , e6 .
(i)
Now Ri has been reduced to a square matrix R1,1 consisting of the following 5 columns:
a1 = −je2 + ε1/2 µ̄i11 ū1 + εµ̄2i11 ū2 + O(ε)ū1 + O(ε2 ),
a2 = je2 + ε1/2 µi21 u1 + εµ2i21 u2 + O(ε)u1 + O(ε2 ), a3 = ā2 ,
a4 = je2 + εϑ31 u1 + ε2 ϑ231 u2 + O(ε2 )u1 + O(ε2 ), a5 = ā4 .
(i)
Consequently the order of C1,1 is found in such a way: take out terms with ε1/2 ū1 , εū2 respectively from a1 , a3 , terms with
e2 from a5 , terms with εu2 from a2 , terms with εu1 from a4 . Now that a1 , a3 , a5 jointly contribute the same degree of
1
2
2 (m − 3m + 3) as ūi1 , ūi2 , ūi3 , we are left to choose terms with u1 and u2 respectively from a2 and a4 . The above can
(i)
be conducted similarly for calculating the order of C1,1 when m > 3, and actually for every cofactor. Moreover, by the
(i)
(i)
(i)
(i)
(i)
symmetry of conjugates, C2k−1,2n−1 , C2k−1,2n , C2k,2n−1 , C2k,2n have an identical order. So we only focus on C2k−1,2n−1
(i)
below. Reminded by the case of C1,1 when m = 3, we suffice to choose linearly independent terms with a lowest sum
of degrees from the modified columns uij for j 6= n, where uk−1 has been subtracted from each column. Recall that in
finding the order of | det Ri |, terms with εu0 , . . . , εm−3 um−3 from type II columns are first selected, and then terms with
ε(m−2)/2 um−2 , ε(m−1)/2 um−1 from type I columns. Such a method still applies in finding the order of cofactors, and we
(i)
conclude that C2k−1,2n−1 has the lowest order for fixed n if and only if k = m. In other words, for any row in adjRi , the
entries at the 2m − 1-th and 2m-th column exclusively have the lowest order when compared with other entries at the same
row. To be detailed,
(i)
(i)
(i)
(i)
|C2m−1,2n−1 |, |C2m−1,2n |, |C2m,2n−1 |, |C2m,2n |
m−1
m−2
m−2
2
| det V0 |2 |µm−2
i11 µi21 (µi11 − µi21 )||µin1 |
m2 −3m+3−(m−1)/2
·ε
(1 + o(1)), n = 1, 2;
=
m−2
m−1
2
|
det
V
||
det
V
||µm−2
0
n
i11 µi21 (µi11 − µi21 )|
m−3 m−3
m2 −3m+3−(m−2)
·|µi11 µi21 (µi11 − µi21 )|ε
(1 + o(1)), n = 3, . . . , m,
(62)
where Vn = V (ϑ31 , . . . , ϑn−1,1 , ϑn1 , . . . , ϑm1 ) is a Vandermonde matrix of order m − 3. Together with (61) it yields that by
µi11 = −µi21
−1
−1
−1
−1
|R
( i (2n1 − 1, 2m − 1)|, |Ri (2n, 2m − 1)|, |Ri (2n − 1, 2m)|, |Ri (2n, 2m)|
−(m−1)/2
(1 + o(1)), n = 1, 2;
4|µi11 |m−1 ε
=
| det Vn | −(m−1)
1
ε
(1 + o(1)), n = 3, . . . , m;
2|µi11 |2 | det V0 |
(63)
−1
−1
|R
− 1, k)|, |Ri (2n, k)|
i (2n
o(ε−(m−1)/2 ), n = 1, 2;
=
for 1 ≤ k ≤ 2m.
o(ε−(m−1) ), n = 3, . . . , m;
In the meanwhile, the following holds for 1 ≤ j ≤ m:
|Ri (2j − 1, 2n − 1)|, |Ri (2j − 1, 2n)|, |Ri (2j, 2n − 1)|, |Ri (2j, 2n)|
j−1 (j−1)/2
|µin1 |ε
+ O(εj/2 ), n = 1, 2;
=
j−1 j−1
|ϑn1 |ε
+ O(εj ), n = 3, . . . , m;
(64)
Combining (60), (63) and (64) we can obtain
|ξs,2j−1 |
Pm
−1
−1
≤ ρsi ||ξ||
n=1 |Ri (2j − 1, 2n − 1)|(|Ri (2n − 1, 2m − 1)| + |Ri (2n − 1, 2m)|)
Pm
−1
−1
+ n=1 |Ri (2j − 1, 2n)|(|Ri (2n, 2m − 1)| + |Ri (2n, 2m)|) (1 + o(1))
P2
Pm |ϑj−1 |
|µj−1
Vn | j−(m−1)
(j−m)/2
in1 |
+ n=3 2|µn12 | || det
(1 + o(1)).
≤ ρsi ||ξ||2
n=1 4|µm−1 | ε
det V0 | ε
i11
i11
< j−m
p 2
and the conclusion follows by noticing that j − (m − 1)
for j = m − 1, m, as well as |µi11 | =
j − (m − 1) > j−m
2
for j < m − 2, j − (m − 1) = j−m
for j = m − 2 and
2
Q
Vn |
|ϑ
−
ϑ
|.
=
λi /2, || det
k1
n1
det V0 |
3≤k≤m,k6=n
Proof of Lemma 4.3 The proofPof Lemma 4.3 relies on the following combinatorial identity.
∞
Lemma A.1: [21] Let f (t) = k=0 fk tk be a formal power series [20]. Then the following rule holds if b = 0 and f (t) is
a polynomial:
X
C(n + ak, m + bk)z m+bk fk = [tm ](1 + zt)n f (t−b (1 + zt)a ),
k
m
where [t ]g(t) denotes the extraction of the coefficient of tm from the formal power series g(t).
Now let we return to the proof. Denoting
v(θ) = cos θA2m−1 (1, ·) + sin θA2m−1 (2, ·) + A2m−1 (3, ·)
= [v1 (θ) v2 (θ) . . . v2m−1 (θ) v2m (θ)]
and recalling S = A2m−1 O−1 , the original equation is equivalent to l(θ)O = v(θ). Direct computation shows that the entries
of v(θ) are given by
v2j−1 (θ) = C(2m, j − 1) cos(2m − j + 1)θ,
v2j (θ) = C(2m, j − 1) sin(2m − j + 1)θ, j = 1, . . . , m;
and the entries of O are given by
O(k, 2j − 1) = C(k − 1, j − 1) cos(k − j)θ,
O(k, 2j) = C(k − 1, j − 1) sin(k − j)θ, k = 1, . . . , 2m, j = 1, . . . , m.
As a result, the equation l(θ)O = v(θ) is equivalent to the following m equations:
2m−1
X
k=0
or equally
lk (θ)C(k, h)e(k−h)θ = C(2m, h)e(2m−h)θ , h = 0, 1, . . . , m − 1,
2m
X
k=0
lk (θ)C(k, h)e(k−h)θ = 0, h = 0, 1, . . . , m − 1,
(65)
(65′ )
if we let l2m (θ) = −1. Noticing that 2 cos θ = eθ + e−θ , we substitute
the expression of lk (θ)P
into the left-hand side of the
P
m
(k−h)θ
(2w−h)θ
above h-th equation, and expand it into a power series of eθ as 2m
l
(θ)C(k,
h)e
=
, with
k
k=0
w=0 αw,h e
αw,h =
w m−k
X
X
k=0 j=w−k
C(m, j)C(m − j, k)C(j, w − k)C(j + 2k, h)(−1)j+2k−1 .
Therefore, if we can show that αw,h for w = 0, 1, . . . , m and h = 0, 1, . . . , m − 1, then the prescribed l(θ) is a solution of
(65), and by the nonsigularity of O it is also unique.
We first transform αw,h as follows. By remembering that
C(m, j)C(m − j, k)C(j, w − k) = C(m, k)C(m − k, j)C(j, w − k)
= C(m, k)C(m − k, w − k)C(m − w, j − (w − k))
and letting s = j − (w − k), it is clear that
αw,h =
w
X
k=0
C(m, k)C(m − k, w − k)(−1)k+w−1
m−w
X
s=0
C(m − w, s)C(s + w + k, h)(−1)s .
NowPwe claim that
s
m−w
, w = 0, 1, . . . , m,
1). Pm−w
s=0 C(m − w, s)C(s + w + k, h)(−1) = C(w + k, h − (m − w))(−1)
w
m+k−1
= 0, h = 0, 1, . . . , m − 1,
2). k=0 C(m, k)C(m − k, w − k)C(w + k, h − (m − w))(−1)
and the proof of the first part is completed by combining
these two claims.
P
1). Let fs = (−1)s C(m − w, s) and f (t) = fs ts = (1 − t)m−w . Applying Lemma A.1, we have
s
Pm−w
C(m − w, s)C(s + w + k, h)(−1)s
Ps=0
m−w
= s=0 C(w + k + 1 · s, h + 0 · s) · 1 · fs
= [th ](1 + t)w+k f (1 + t)
= [th ](1 + t)w+k (−t)m−w
= (−1)m−w [th−(m−w) ](1 + t)w+k
= (−1)m−w C(w + k, h − (m − w)),
which establishes the first claim.
2). For the second claim,
Pw
=
=
=
=
=
k)C(m − k, w − k)C(w + k, h − (m − w))(−1)m+k−1
k=0 C(m,
Pm
m−1
k
(−1)
k=0 C(m, k)(−1) C(m − k, w − k)C(w + k, h − (m − w))
P
m
(−1)m−1 k=0 [tk ](1 − t)m [v m−w ](1 + v)m−k [uh−(m−w) ](1 + u)w+k
Pm
k
m
(−1)m−1 [v m−w ](1 + v)m [uh−(m−w) ](1 + u)w k=0 [tk ]( 1+u
1+v ) (1 − t)
m
(−1)m−1 [v m−w ](1 + v)m [uh−(m−w) ](1 + u)w (1 − 1+u
1+v )
m
m−1 m−w
m h−(m−w)
w (v−u)
(−1)
[v
](1 + v) [u
](1 + u) (1+v)m
m−1 m−w
h−(m−w)
w
m
= (−1)
[v
r=h−(m−w)
][u
](1 + u) (v − u)
========== (−1)m−1 [v m−w ][ur ](1 + u)w (v − u)m
Pr
= (−1)m−1 [v m−w ] k=0 C(w, r − k)C(m, k)v m−k (−1)k
P
= (−1)m−1 [v k−w ] rk=0 C(w, r − k)C(m, k)(−1)k ,
where t, v, u are indeterminates. Noticing that r ≥ w ⇔ h − (m − w) ≥ w ⇔ h ≥ m is contradictory to h = 0, 1, . . . , m − 1,
we have r < w and k ≤ r < w, which suggests the vanishing of the last equation in the above, and the proof for the first part
is complete.
As for the second part, by noting that the exponents of cos θ in lk (θ) are even when k is an even number, while the exponents
are odd when k is an odd number, it can be noted that the sign of each term in lk (θ) is the same. Therefore we obtain
P2m−1
P2m−1 P⌊k/2⌋
k−2h
k=0 |lk (θ)| = Pk=0
h=0 C(m, k − 2h)C(m − (k − 2h), h)|2 cos θ|
m
j
= j=0 βj | cos θ| ,
P
m
with β0 = 2m − 1 and βj = 2j C(m, j) m−j
h=0 C(m − j, h) = 2 C(m, j) for j = 1, . . . , m, and the conclusion follows
directly.
Acknowledgement The authors would like to thank Dr. Shuai Liu for his valuable suggestions.
R EFERENCES
[1] Z. Qiu, L. Xie, and Y. Hong, “Data rate for quantized consensus of high-order multi-agent systems with poles on the unit circle,” in 53rd IEEE Conference
on Decision and Control, Dec. 2014, pp. 3771–3776.
[2] R. Olfati-Saber, J. A. Fax, and R. M. Murray, “Consensus and Cooperation in Networked Multi-Agent Systems,” Proceedings of the IEEE, vol. 95,
no. 1, pp. 215–233, Jan. 2007.
[3] Y. Cao, W. Yu, W. Ren, and G. Chen, “An Overview of Recent Progress in the Study of Distributed Multi-Agent Coordination,” IEEE Transactions on
Industrial Informatics, vol. 9, no. 1, pp. 427–438, Feb. 2013.
[4] S. Knorn, Z. Chen, and R. Middleton, “Overview: Collective Control of Multi-agent Systems,” IEEE Transactions on Control of Network Systems,
vol. PP, no. 99, pp. 1–1, 2015.
[5] A. Kashyap, T. Başar, and R. Srikant, “Quantized consensus,” Automatica, vol. 43, no. 7, pp. 1192–1203, July 2007.
[6] A. Nedic, A. Olshevsky, A. Ozdaglar, and J. N. Tsitsiklis, “On distributed averaging algorithms and quantization effects,” IEEE Transactions on
Automatic Control, vol. 54, no. 11, pp. 2506–2517, 2009.
[7] R. Carli, F. Bullo, and S. Zampieri, “Quantized average consensus via dynamic coding/decoding schemes,” International Journal of Robust and
Nonlinear Control, vol. 20, no. 2, pp. 156–175, Jan. 2010.
[8] T. Li, M. Fu, L. Xie, and J.-F. Zhang, “Distributed Consensus With Limited Communication Data Rate,” IEEE Transactions on Automatic Control,
vol. 56, no. 2, pp. 279–292, Feb. 2011.
[9] Keyou You and Lihua Xie, “Network Topology and Communication Data Rate for Consensusability of Discrete-Time Multi-Agent Systems,” IEEE
Transactions on Automatic Control, vol. 56, no. 10, pp. 2262–2275, Oct. 2011.
[10] Y. Meng, T. Li, and J.-F. Zhang, “Coordination Over Multi-Agent Networks With Unmeasurable States and Finite-Level Quantization,” arXiv:1505.03259
[cs, math], May 2015.
[11] Q. Zhang and J.-F. Zhang, “Quantized Data–Based Distributed Consensus under Directed Time-Varying Communication Topology,” SIAM Journal on
Control and Optimization, vol. 51, no. 1, pp. 332–352, Jan. 2013.
[12] D. Li, Q. Liu, X. Wang, and Z. Yin, “Quantized consensus over directed networks with switching topologies,” Systems & Control Letters, vol. 65, pp.
13–22, Mar. 2014.
[13] A. Olshevsky, “Consensus with Ternary Messages,” SIAM Journal on Control and Optimization, vol. 52, no. 2, pp. 987–1009, Jan. 2014.
[14] Tao Li and Lihua Xie, “Distributed Coordination of Multi-Agent Systems With Quantized-Observer Based Encoding-Decoding,” IEEE Transactions on
Automatic Control, vol. 57, no. 12, pp. 3023–3037, Dec. 2012.
[15] Z. Qiu, L. Xie, and Y. Hong, “Quantized Leaderless and Leader-Following Consensus of High-Order Multi-Agent Systems With Limited Data Rate,”
IEEE Transactions on Automatic Control, vol. 61, no. 9, pp. 2432–2447, Sept. 2016.
[16] Wei Ren and R. Beard, “Consensus seeking in multiagent systems under dynamically changing interaction topologies,” IEEE Transactions on Automatic
Control, vol. 50, no. 5, pp. 655–661, May 2005.
[17] A. P. Seyranian and A. A. Mailybaev, Multiparameter stability theory with mechanical applications. World Scientific, 2003, vol. 13.
[18] C.-T. Chen, Linear system theory and design. Oxford University Press, Inc., 1995.
[19] H. Baumgärtel, Analytic perturbation theory for matrices and operators. Springer, 1985, vol. 15.
[20] R. Sprugnoli, “An introduction to mathematical methods in combinatorics,” Dipartimento di Sistemi e Informatica Viale Morgagni, 2006.
[21] ——, “Riordan array proofs of identities in Gould’s book,” Published electronically at http://www. dsi. unifi. it/resp/GouldBK. pdf, 2007.
| 3cs.SY
|
Optimal Warping Paths are unique for almost
every Pair of Time Series
arXiv:1705.05681v2 [cs.LG] 2 Mar 2018
Brijnesh J. Jain and David Schultz
Technische Universität Berlin, Germany
e-mail: [email protected]
Update rules for learning in dynamic time warping spaces are based on optimal warping
paths between parameter and input time series. In general, optimal warping paths are not
unique resulting in adverse effects in theory and practice. Under the assumption of squared
error local costs, we show that no two warping paths have identical costs almost everywhere
in a measure-theoretic sense. Two direct consequences of this result are: (i) optimal warping
paths are unique almost everywhere, and (ii) the set of all pairs of time series with multiple
equal-cost warping paths coincides with the union of exponentially many zero sets of quadratic
forms. One implication of the proposed results is that typical distance-based cost functions
such as the k-means objective are differentiable almost everywhere and can be minimized by
subgradient methods.
1. Introduction
1.1. Dynamic time warping
Time series such as audio, video, and other sensory signals represent a collection of time-dependent values
that may vary in speed (see Fig. 1). Since the Euclidean distance is sensitive to such variations, its
application to time series related data mining tasks may give unsatisfactory results [6, 8, 24]. Consequently,
the preferred approaches to compare time series apply elastic transformations that filter out the variations
in speed. Among various techniques, one of the most common elastic transformation is dynamic time
warping (DTW) [20].
Dynamic time warping is based on the concept of warping path. A warping path p determines how to
stretch two given time series x and y to warped time series x0 and y 0 under certain constraints. The cost
of warping x and y along warping path p measures how dissimilar the warped time series x0 and y 0 are.
There are exponential many different warping paths [1] each of which determines the cost of warping
time series x and y. An optimal warping path of x and y is a warping path with minimum cost. Optimal
warping paths exist but are not unique in general (see Fig. 2).
1.2. The problem of non-uniqueness
Recent research is directed towards extending standard statistical concepts and machine learning methods
to time series spaces endowed with the DTW distance. Examples include time series averaging [3, 5, 16,
17, 21], k-means clustering [10, 18, 22], self-organizing maps [14], learning vector quantization [23, 12],
and warped-linear classifiers [11, 13].
The lowest common denominator of these approaches is that they repeatedly update one or more
parameter time series. In addition, update directions are based on optimal warping paths such that the
following properties hold:
• If an optimal warping path is unique, then the update direction is well-defined.
1
Figure 1: Variations in speed and their implications for comparing two time series shown in blue and
yellow. Both time series are similar in shape but differ in speed along their flat regions. Plot
(a): The Euclidean distance warps the i-th points of both time series onto one another as shown
by the gray lines. Since the Euclidean warping is sensitive to variations in speed the shapes are
not preserved and both peaks are aligned to points in flat regions. Plot (c): Euclidean warping
leaves the time series unchanged and results in a large dissimilarity score as indicated by the
gray shaded area. Plot (b): An optimal warping that preserves the shape of both time series.
Double arrows indicate segments stretched by dynamic time warping. Plot (d): Warped time
series obtained by the optimal warping in (b). The left flat segment of the yellow and the right
flat segment of the blue time series have been stretched to align both time series. The resulting
dissimilarity is zero and better reflects the similarity in shape than the Euclidean distance.
• If an optimal warping path is non-unique, then there are several update directions.
Non-uniqueness of optimal warping paths complicates the algorithmic design of learning methods in DTW
spaces and their theoretical analysis. In some situations, non-uniqueness may result in adverse effects.
For example, repulsive updating in learning vector quantization finds a theoretical justification only in
cases where the corresponding optimal warping path is unique [12].
Given the problems caused by non-uniqueness, it is desirable that optimal warping paths are unique
almost everywhere. In this case, non-unique optimal warping paths occur exceptionally and are easier to
handle as we will shortly. Therefore, we are interested in how prevalent unique optimal warping paths are.
1.3. Almost everywhere
The colloquial term “almost everywhere” has a precise measure-theoretic meaning. A measure quantifies
the size of a set. It generalizes the concepts of length, area, and volume of a solid body defined in one,
two, and three dimensions, respectively. The term “almost everywhere” finds its roots in the notion of a
“negligible set”. Negligible sets are sets contained in a set of measure zero. For example, the function
1 : x 6= 0
f (x) =
0 : x=0
is discontinuous on the negligible set {0} with measure zero. We say, function f is continuous almost
everywhere, because the set where f is not continuous is negligible. More generally, a property P is
said to be true almost everywhere if the set where P is false is negligible. The property that an optimal
warping is unique almost everywhere means that the set of all pairs of time series with non-unique optimal
warping path is negligible.
When working in a measure space, a negligible set contains the exceptional cases we can handle or
even do not care about and often ignore. For example, we do not care about the behavior of the above
function f on its negligible set {0} when computing its Lebesgue integral over [−1, 1]. Another example
is that cost functions of some machine learning methods in Euclidean spaces such as k-means or learning
vector quantization are non-differentiable on a negligible set. In such cases, it is common practice to
ignore such points or to resort to subgradient methods.
2
Figure 2: Example of non-unique warping paths (see Section 2.1 for definitions). Section (a) Time series
x and y. (b) Local cost matrix with elements (xi − yj )2 . (c)√Accumulated cost obtained by
dynamic programming [20] gives the DTW distance δ(x, y) = 15. The gray shaded cells show
two different optimal warping paths. (d) Alignments of x and y by the two different optimal
warping paths.
1.4. Contributions
Consider the following property P on the set F m × F n of pairs of time series of length m and n: The pair
of time series x ∈ F m and y ∈ F n satisfies P if there are two different (not necessarily optimal) warping
paths between x and y with identical costs. Under the assumption of a squared error local cost function,
the main result of this article is Theorem 1:
Property P is negligible on F m × F n .
Direct consequences of Theorem 1 are (i) optimal warping paths are unique almost everywhere, and (ii)
property P holds on the union of exponentially many zero sets of quadratic forms. The results hold for
uni- as well as multivariate time series.
An implication of non-unique optimal warping paths is that adverse effects in learning are exceptional
cases that can be safely handled. For example, learning amounts in (stochastic) gradient descent update
rules almost everywhere.
2. Background
This section first introduces warping paths and then defines the notions of negligible and almost everywhere
from measure theory.
2.1. Time Series and Warping Paths
We first define time series. Let F = Rd denote the d-dimensional Euclidean space. A d-variate time series
of length m is a sequence x = (x1 , . . . , xm ) consisting of elements xi ∈ F. By F m we denote the set of all
time series of length m with elements from F.
Next, we describe warping paths. Let [n] = {1, . . . , n}, where n ∈ N. An (m × n)-lattice is a set of
the form Lm,n = [m] × [n]. A warping path in lattice Lm,n is a sequence p = (p1 , . . . , pL ) of L points
pl = (il , jl ) ∈ Lm,n such that
1. p1 = (1, 1) and pL = (m, n)
2. pl+1 − pl ∈ {(1, 0), (0, 1), (1, 1)} for all l ∈ [L − 1].
The first condition is called boundary condition and the second one is the step condition.
3
By Pm,n we denote the set of all warping paths in Lm,n . A warping path departs at the upper left
corner (1, 1) and ends at the lower right corner (m, n) of the lattice. Only east (0, 1), south (1, 0), and
southeast (1, 1) steps are allowed to move from a given point pl to the next point pl+1 for all 1 ≤ l < L.
Finally, we introduce optimal warping paths. A warping path p ∈ Pm,n defines an alignment (warping)
between time series x ∈ F m and y ∈ F n by relating elements xi and yj if (i, j) ∈ p. The cost of aligning
time series x and y along warping path p is defined by
X
2
Cp (x, y) =
kxi − yj k ,
(i,j)∈p
where k·k denotes the Euclidean norm on F. A warping path p∗ ∈ Pm,n between x and y is optimal if
p∗ ∈ argmin Cp (x, y).
p∈Pm,n
By P∗ (x, y) we denote the set of all optimal warping paths between time series x and y. The DTW
distance is defined by
q
δ(x, y) = min
Cp (x, y).
p ∈ Pm,n
p
is the DTW distance. Observe that δ(x, y) = Cp (x, y) for all p ∈ P∗ (x, y).
2.2. Measure-Theoretic Concepts
We introduce the necessary measure-theoretic concepts to define the notions of negligible and almost
everywhere. For details, we refer to [9].
One issue in measure theory is that not every subset of a given set X is measurable. A family A
of measurable subsets of a set X is called σ-algebra in X . A measure is a function µ : A → R+ that
assigns a non-negative value to every measurable subset of X such that certain conditions are satisfied.
To introduce these concepts formally, we assume that P(X ) denotes the power set of a set X , that is the
set of all subsets of X . A system A ⊂ P(X ) is called a σ-algebra in X if it has the following properties:
1. X ∈ A
2. U ∈ A implies X \ U ∈ A
S
3. (Ui )i∈N ∈ A implies i∈N Ui ∈ A.
A measure on A is a function µ : A → [0, +∞] that satisfies the following properties:
1. µ(U) ≥ 0 for all U ∈ A
2. µ(∅) = 0
3. For a countable collection of disjoint sets (Ui )i∈N ∈ A, we have
!
[
X
µ
Ui =
µ (Ui ) .
i∈N
i∈N
A triple (X , A, µ) consisting of a set X , a σ-algebra A in X and a measure µ on A is called a measure space.
The Borel-algebra B in Rd is the σ-algebra generated by the open sets of Rd . The Lebesgue-measure µ
on B generalizes the concept of d-volume of a box in Rd . The triple Rd , B, µ is called Borel-Lebesgue
measure space.
Let (X , A, µ) be a measure space, where X is a set, A is a σ-algebra in X , and µ is a measure defined
on A. A set N ⊂ X is µ-negligible if there is a set N 0 ∈ A such that µ(N 0 ) = 0 and N ⊆ N 0 . A property
of X is said to hold µ-almost everywhere if the set of points in X where this property fails is µ-negligible.
4
3. Results
We first show that optimal warping paths are unique almost everywhere. Then we geometrically describe
the location of the non-unique set. Finally, we discuss the implications of the proposed results on learning
in DTW spaces.
Let X = F m × F n bet the set of all pairs (x, y) of time series, where x has length m ∈ N and y has
length n ∈ N. We regard the set X as a Euclidean space and assume the Lebesgue-Borel measure space
(X , B, µ).1 The multi optimal-path set of X is defined by
NX∗ = {(x, y) ∈ X : |P∗ (x, y)| > 1} .
This set consists of all pairs (x, y) ∈ X with non-unique optimal warping path. To assert that NX∗ is
µ-negligible, we show that NX∗ is a subset of a set of measure zero. For this, consider the multi-path set
[
NX =
{(x, y) ∈ X : Cp (x, y) = Cq (x, y)},
p,q∈Pm,n
p6=q
The set NX consists of all pairs (x, y) that can be aligned along different warping paths with identical
cost. Obviously, the set NX∗ is a subset of NX . The next theorem states that NX is a set of measure zero.
Theorem 1. Let (X , B, µ) be the Lebesgue-Borel measure space. Then µ(NX ) = 0.
From Theorem 1 and NX∗ ⊆ NX immediately follows that NX∗ is µ-negligible.
Corollary 2. Under the assumptions of Theorem 1 the set NX∗ is µ-negligible.
Thus, optimal warping paths are unique µ-almost everywhere in X . Even more generally: The property
that all warping paths have different cost holds µ-almost everywhere in X .
We describe the geometric form of the multi-path set NX . For this we identify F m × F n with F k ,
where k = m + n. Thus, pairs (x, y) ∈ X of time series are summarized to z ∈ F k , henceforth denoted as
z ∈ X . By X 2 = F k×k we denote the set of all (k × k)-matrices with elements from F. Finally, the zero
set of a function f : X → R is of the form
Z(f ) = {z ∈ X : f (z) = 0} .
From the proof of Theorem 1 directly follows that the set NX is the union of zero sets of quadratic forms.
Corollary 3. Under the assumptions of Theorem 1, there is an integer D ∈ N and symmetric matrices
A1 , . . . , AD ∈ X 2 such that
D
[
NX =
Z zT Ai z .
i=1
The number D of zero sets in Corollary 3 grows exponentially in m and n. From the proof of Theorem
1 follows that D = Dm,n (Dm,n − 1)/2, where
min {m,n}
Dm,n =
X
m n
2
.
i
i
i
i=0
is the Delannoy number [1]. The Delannoy number Dm,n = |Pm,n | counts the number of all warping
paths in lattice Lm,n . Table 1 presents the first Delannoy numbers up to m = 10 and n = 9. We see
that there are more than half a million warping paths in a (10 × 9)-lattice showing that NX is the union
of more than 178 billion zero sets. For two time series of length 20, the number of warping paths is
D20,20 = 260, 543, 813, 797, 441, which is more than 260 trillions. Thus, the multi-path set NX of two
time series of length 20 is the union of more than 33 octillion zero sets, that is d > 33 · 1027 . An open
question is the number D∗ ≤ D of zero sets that form the multi optimal-path set NX∗ . The example in
Figure 3 indicates that the multi optimal-path set NX∗ can be much smaller than the multi-path set NX .
1 See
Remark 7 for an explanation of why we regard X as a Euclidean space.
5
m\n
1
2
3
4
5
6
7
8
9
1
2
3
4
5
6
7
8
9
10
1
1
1
1
1
1
1
1
1
1
1
3
5
7
9
11
13
15
17
19
1
5
13
25
41
61
85
113
145
181
1
7
25
63
129
231
377
575
833
1,159
1
9
41
129
321
681
1,289
2,241
3,649
5,641
1
11
61
231
681
1,683
3,653
7,183
13,073
22,363
1
13
85
377
1,289
3,653
8,989
19,825
40,081
75,517
1
15
113
575
2,241
7,183
19,825
48,639
108,545
224,143
1
17
145
833
3,649
13,073
40,081
108,545
265,729
598,417
Table 1: Delannoy numbers Dm,n for m ∈ [10] (rows) and n ∈ [9] (columns).
Figure 3: Illustration of the multi-path set NX and multi optimal-path sets NX∗ for the pair of time series
x = (x1 , x2 , 0) and y = (4, 8, 7) with varying values for x1 and x2 . The Delannoy number D3,3 is
13. Hence, the multi-path set NX consists of D = 78 zero sets of quadratic forms as indicated by
colored curves. The subset of multi optimal-path set NX∗ is highlighted by fat curve segments.
3.1. Discussion
We discuss the implications of Theorem 1 for learning in DTW spaces.
3.1.1. Learning
This section shows that almost-everywhere uniqueness implies almost-everywhere differentiability of the
underlying cost function.
To convey the line of argument, it is sufficient to restrict to the problem of averaging time series
as representative for other, more complex learning problems. In contrast to computing the average in
Euclidean spaces, time series averaging is a non-trivial task for which the complexity class is currently
unknown [3].
Let x(1) , . . . , x(N ) be a sample of N time series, possibly of varying length. Consider the cost function
J : F n → R of the form
N
1 X (k)
J(z) =
` δ x ,z ,
N
k=1
where ` : R → R is a loss function. Common loss functions for averaging time series are the identity
`(a) = a and the squared loss `(a) = a2 . The goal of time series averaging is to find a time series z∗ ∈ F n
of length n that minimizes the cost J(z).
6
The challenge of time series averaging is to minimize the non-differentiable cost function J(z). We show
that almost-everywhere uniqueness of optimal warping paths implies almost-everywhere differentiability
of the cost function J(z) and provides a stochastic (incremental) update rule.
We express the DTW distance δ(x, z) as a parametrized function. Suppose that x ∈ F m is a time series.
Then the parametrized DTW function δx : F n → R restricted to the set F n is of the form δx (z) = δ(x, z).
We have the following result:
Proposition 4. Suppose that x ∈ F m and z ∈ F n are two time series with unique optimal warping path.
Then the function δx (z) = δ(x, z) is differentiable at z and its gradient is a time series of the form
∇z δx (z) = ∇z Cp∗ (x, z) ∈ F n ,
where p∗ ∈ P∗ (x, z) is an optimal warping path.
The proof follows from [19] after reducing δx (z) to a piecewise smooth function. By construction, the
cost Cp (x, z) of warping x and z along warping path p ∈ Pm,n is differentiable as a function of the second
argument z ∈ F n . Non-differentiability of δx (z) is caused by non-uniqueness of an optimal warping path
between x and z. In this case we have
δx (z) = Cp∗ (x, z) = Cq∗ (x, z),
where p∗ , q∗ ∈ P∗ (x, z) are two distinct optimal warping paths. Then it can happen that
∇z Cp∗ (x, z) 6= ∇y Cq∗ (x, z)
showing that δx is non-differentiable at z.
Next, suppose that the loss function ` : R → R is differentiable and an optimal warping path p∗ between
time series x and z is unique. Then the individual cost Jx (z) = ` (δ(x, z)) is also differentiable at z with
gradient
∇z Jx (z) = `0 (Cp∗ (x, z))∇z Cp∗ (x, z).
Differentiability of Jx (z) gives rise to a stochastic (incremental) update rule of the form
z ← z − η `0 (Cp∗ (x, z))∇z Cp∗ (x, z),
(1)
where η is the step size and p∗ is an optimal warping path between x and z. We can also apply the
stochastic update rule (1) in cases where an optimal warping path between x and z is not unique. In this
case, we first (randomly) select an optimal warping path p∗ from the set P∗ (x, z). Then we update z by
applying update rule (1).
Updating at non-differentiable points according to the rule (1) is not well-defined. In addition, it is
unclear whether the update directions are always directions of descent, for learning problems in general.
The next result confines both issues to a negligible set.
Corollary 5. Suppose that X = F m × F n and ` : R → R is a differentiable loss function. Then the
functions
• δ(x, z) = δx (z)
• Jx (z) = ` (δ(x, z))
are differentiable µ-almost everywhere on X .
Corollary 5 directly follows from Prop. 4 together with Theorem 1. In summary, almost-everywhere
uniqueness of optimal warping paths implies almost-everywhere differentiability of the individual cost
Jx (z). The latter in turn implies that update rule (1) is a well-defined stochastic gradient step almost
everywhere.
The arguments in this section essentially carry over to other learning problems in DTW spaces such
as k-means, self-organizing maps, and learning vector quantization. We assume that it should not be
a problem to transfer the proposed results to learning based on DTW similarity scores as applied in
warped-linar classifiers [11, 13].
7
3.1.2. Learning Vector Quantization in DTW Spaces
Learning vector quantization (LVQ) is a supervised classification scheme introduced by Kohonen [15]. A
basic principle shared by most LVQ variants is the margin-growth principle [12]. This principle justifies
the different learning rules and corresponds to stochastic gradient update rules if a differentiable cost
function exists. As the k-means algorithm, the LVQ scheme has been generalized to DTW spaces [12, 23].
In this section, we illustrate that a unique optimal-warping path is a necessary condition to satisfy the
margin-growth principle in DTW spaces as proved in [12], Theorem 12.
As a representative example, we describe LVQ1, the simplest of all LVQ algorithms [15]. Let X = Rd be
the d-dimensional Euclidean space and let Y = {1, . . . , C} be a set consisting of C class labels. The LVQ1
algorithm assumes a codebook C = {(p1 , z1 ), . . . , (pK , zk )} of K prototypes pk ∈ X with corresponding
class labels zk ∈ Y. As a classifier, LVQ1 assigns an input point x ∈ X to the class zc of its closest
prototype pc ∈ C, where
2
c ∈ argmin kpk − xk .
k
LVQ1 learns a codebook C on the basis of a training set D = {(x1 , y1 ), . . . , (xN , yN )} ⊆ X × Y. After
initialization of C, the algorithm repeats the following steps until termination: (i) Randomly select a
training example (xi , yi ) ∈ D; (ii) determine the prototype pc ∈ C closest to xi ; and (iii) attract pc to xi
if their class labels agree and repel pc from xi otherwise. Step (iii) adjusts pc according to the rule
pc ← pc ± η (xi − pc ),
where η is the learning rate, the sign ± is positive if the labels of pc and xi agree (zc = yi ), and negative
otherwise (zc 6= yi ). The update rule guarantees that adjusting pc makes an incorrect classification of xi
more insecure. Formally, if the learning rate η < θ is bounded by some threshold θ, the LVQ1 update
rule guarantees to increase the hypothesis margin
(
2
2
kxi − p− k − kxi − pc k : yi = zc
,
µc (xi ) =
2
2
kxi − pc k − kxi − p+ k : yi 6= zc
where p+ (p− ) is the closest prototypes of x with the same (different) class label.
The different variants of LVQ have been extended to DTW spaces by replacing the squared Euclidean
distance with the squared DTW distance [12, 23]. The update rule is based on an optimal warping
path between the current input time series and its closest prototpye. In asymmetric learning [12], the
margin-growth principle always holds for the attractive force and for the repulsive force only when the
optimal warping path is unique (as a necessary condition).
3.1.3. Comments
We conclude this sections with two remarks.
Remark 6. Proposition 4 states that uniqueness of an optimal warping path between x and y implies
differentiability of δx at y. The converse statement does not hold, in general. A more general approach to
arrive at Prop. 4 and Corollary 5 is as follows: First, show that a function f is locally Lipschitz continuous
(llc). Then invoke Rademacher’s Theorem [7] to assert almost-everywhere differentiability of f . By the
rule of calculus of llc functions, we have:
• δx (y) is llc on F n , because the minimum of continuously differentiable functions is llc.
• If the loss ` is llc, then Jx (z) is llc, because the composition of llc functions is llc.
Remark 7. Measure-theoretic, geometric, and analytical concepts are all based on Euclidean spaces
rather than DTW spaces. The reason is that contemporary learning algorithms are formulated in such
a way that the current solution and input time series are first projected into the Euclidean space via
optimally warping to the same length. Then an update step is performed and finally the updated solution
is projected back to the DTW space. Therefore, to understand this form of learning under warping, we
study the DTW distance δ(x, y) as a function restricted to the Euclidean space F m × F n , where m is the
length of x and n is the length of y.
8
4. Conclusion
The multi-path set is negligible and corresponds to the union of zero sets of exponentially many quadratic
forms. As a subset of the multi-path set, the multi optimal-path set is also negligible. Therefore optimal
warping paths are unique almost everywhere. The implications of the proposed results are that adverse
effects on learning in DTW spaces caused by non-unique optimal warping paths can be controlled and
learning in DTW spaces amounts in minimizing the respective cost function by (stochastic) gradient
descent almost everywhere.
Acknowledgements.
B. Jain was funded by the DFG Sachbeihilfe JA 2109/4-1.
A. Proofs
The appendix presents the proof of Theorem 1 and Proposition 4. We first consider the univariate case (d = 1) in
Sections A.1 and A.2. Section A.1 introduces a more useful representation for proving the main results of this
contribution and derives some auxiliary results. Section A.2 proves the proposed results for the univariate case.
Finally, Section A.3 generalizes the proofs to the multivariate case.
A.1. Preliminaries
We assume that elements are from F = R, that is d = 1. We write Rm instead of F m to denote a time series of
length m. By ek ∈ Rm we denote the k-th standard basis vector of Rm with elements
1 : i=k
k
ei =
.
0 : i 6= k
Definition 8. Let p = (p1 , . . . , pL ) ∈ Pm,n be a warping path with points pl = (il , jl ). Then
T
Φ = ei1 , . . . , eiL ∈ RL×m
T
Ψ = ej1 , . . . , ejL ∈ RL×n
is the pair of embedding matrices induced by warping path p.
The embedding matrices have full column rank n due to the boundary and step condition of the warping
path. Thus, we can regard the embedding matrices of warping path p as injective linear maps Φ : Rm → RL and
Ψ : Rn → RL that embed time series x ∈ Rm and y ∈ Rn into RL by matrix multiplication Φx and Ψy. We can
express the cost Cp (x, y) of aligning time series x and y along warping path p by the squared Euclidean distance
between their induced embeddings.
Proposition 9. Let Φ and Ψ be the embeddings induced by warping path p ∈ Pm,n . Then
Cp (x, y) = kΦx − Ψyk2
for all x ∈ Rm and all y ∈ Rn .
Proof. [21], Proposition A.2.
Next, we define the warping and valence matrix of a warping path.
Definition 10. Let Φ and Ψ be the pair of embedding matrices induced by warping path p ∈ Pm,n . Then the
valence matrix V ∈ Rm×m and warping matrix W ∈ Rm×n of warping path p are defined by
V = ΦT Φ
W = ΦT Ψ.
The definition of valence and warping matrix are oriented in the following sense: The warping matrix W ∈ Rm×n
aligns a time series y ∈ Rn to the time axis of time series x ∈ Rm . The diagonal elements vii of the valence matrix
9
V ∈ Rm×m count the number of elements of y warped onto the same element xi of x. Alternatively, we can define
the complementary valence and warping matrix of w by
V = ΨT Ψ
W = ΨT Φ = W T .
The complementary warping matrix W ∈ Rn×m warps time series x ∈ Rm to the time axis of time series y ∈ Rn .
The diagonal elements v ii of the complementary valence matrix V ∈ Rn×n counts the number of elements of x
warped onto the same element yi of y.
Let p ∈ Pm,n be a warping path of length L with induced embedding matrices Φ ∈ RL×m and Ψ ∈ RL×n . The
aggregated embedding matrix Θ induced by warping path p is defined by
Θ = (Φ, −Ψ) ∈ RL×k ,
where k = m + n. Then the symmetric matrix ΘT Θ is of the form
V
−W
.
ΘT Θ =
V
−W
We use the following notations:
X = Rm × Rn = Rk
X 2 = Rk×k .
The next result expresses the cost Cp (x, y) by the matrix ΘT Θ.
Lemma 11. Let Θ be the aggregated embedding matrix induced by warping path p ∈ Pm,n . Then we have
Cp (z) = zT ΘT Θz
for all z ∈ X .
Proof. Suppose that Θ = (Φ, −Ψ), where Φ and Ψ are the embedding matrices induced by p. Let z = (x, y) ∈ X .
Then we have
Cp (z) = kΦx − Ψyk2
= xT ΦT Φx − xT ΦT Ψy − yT ΨT Φx + yT ΨT Ψy
= xT V x − xT W y − yT W x + yT V y
V
−W
x
= xT , yT
y
−W
V
= zT ΘT Θz.
The last auxiliary result shows that the zero set of a non-zero quadratic form has measure zero.
Lemma 12. Let matrix A ∈ Rn×n be non-zero and symmetric. Then
µ x ∈ Rn : xT Ax = 0 = 0,
where µ is the Lebesgue measure on Rn .
Proof. Since A is symmetric, there is an orthogonal matrix Q ∈ Rn×n such that Λ = QT AQ is a diagonal matrix.
Consider the function
n
X
f (x) = xT Λx =
λii x2i ,
i=1
where the λii are the diagonal elements of Λ. Since A is non-zero, there is at least one λii 6= 0. Hence, f (x) is a
non-zero polynomial on Rn . Then
the set U = {x ∈ Rn : f (x) = 0} is measurable and has measure zero [4].
We show that the set Ũ = x̃ ∈ Rn : x̃T Ax̃ = 0 is also a set of measure zero. Consider the linear map
φ : U → Rn ,
First, we show that φ(U) = Ũ.
10
x 7→ QT x.
• Ũ ⊆ φ(U): Suppose that x̃ ∈ Ũ. With Qx̃ = x we have
0 = x̃T Ax̃ = x̃T QT ΛQx̃ = xT Λx.
This shows that x ∈ U . From φ(x) = QT x = x̃ follows that x̃ ∈ φ(U).
• φ(U) ⊆ Ũ: Let x̃ ∈ φ(U). Then x̃ = QT x = φ(x) for some x ∈ U. Hence, x = Qx̃ and we have
0 = xT Λx = x̃T QT ΛQx̃ = x̃T Ax̃.
This shows that x̃ ∈ Ũ.
Next, we show that µ(Ũ) = 0. Observe that the linear map φ is continuously differentiable on a measurable set
U with Jacobian Jφ (x) = QT . Applying [2], Prop. 3.7.3 gives
Z
µ (φ(U)) ≤
det QT dx.
U
Since QT is orthogonal, we have det QT = 1. Thus, we find that
Z
µ (φ(U)) ≤
dx = µ(U) = 0.
U
Finally, the assertion µ(Ũ) = 0 follows from Ũ = φ(U).
A.2. Proof of Theorem 1 and Proposition 4
This section assumes the univariate case (d = 1).
Proof of Theorem 1:
Suppose that Pm,n = {p1 , . . . , pD }. We use the following notations for all i ∈ [D]:
1.
2.
3.
4.
Θi denotes the aggregated embedding matrix induced by warping path pi .
Vi and Wi are the valence and warping matrices of pi .
V i and W i are the complementary valence and warping matrices of pi .
Ci (z) with z = (x, y) denotes the cost Cpi (x, y) of aligning x and y along warping path pi .
For every i, j ∈ [D] with i 6= j and for every z = (x, y) ∈ X , we have
Ci (z) − Cj (z) = zT ΘTi Θi z − zT ΘTj Θj z = zT A(ij) z,
where A(ij) ∈ X 2 is a symmetric matrix of the form
(ij)
A
=
ΘTi Θi
−
ΘTj Θj
=
Vi − Vj
−W i + W j
−Wi + Wj
Vi−Vj
!
.
For i 6= j the warping paths pi and pj are different implying that the warping
n matrices Wi and Wj ,oresp., are also
different. Hence, A(ij) is non-zero and from Lemma 12 follows that Uij = z ∈ X : zT A(ij) z = 0 has measure
zero. Then the union
[
U=
Uij
i<j
of finitely many measure zero sets also has measure zero. It remains to show that NX = U.
• NX ⊆ U: Suppose that z = (x, y) ∈ NX . Then there are indices i, j ∈ [D] with i < j such that the costs of
aligning x and y along warping paths pi and pj are identical, that is Ci (x, y) = Cj (x, y). Setting z = (x, y)
gives
0 = Ci (z) − Cj (z) = zT A(ij) z.
Hence, z ∈ Uij ⊆ U and therefore NX ⊆ U.
• U ⊆ NX : Let z = (x, y) ∈ U. Then there is a set Uij containing z. From Ci (z) − Cj (z) = 0 follows that pi
and pj are two warping paths between x and y with identical costs. Hence, (x, y) ∈ NX . This proves the
assertion.
11
Proof of Proposition 4:
To show the proposition, we first define the notion of piecewise smooth function. A function f : Rn → R is
piecewise smooth if it is continuous on Rn and for each x0 ∈ Rn there is a neighborhood N (x0 ) of x0 and a finite
collection (fi )i∈I of continuously differentiable functions fi : N (x0 ) → R such that
f (x) ∈ {fi (x) : i ∈ I}
for all x ∈ N (x0 ).
Proof. We show that the function δx (y) = minp Cp (x, y) is piecewise smooth. The function δx (y) is continuous,
because all Cp are continuous and continuity is closed under the min-operation. In addition, the functions Cp
are continuously differentiable as functions in the second argument. Let y0 ∈ F n and let N (y0 ) ⊆ F n be a
neighborhood of y0 . Consider the index set
I = {p ∈ Pm,n : ∃y ∈ N (y0 ) s.t. δx (y) = Cp (x, y)} .
By construction, we have δx (y) ∈ {Cp (x, y) : p ∈ I} for all y ∈ N (y0 ). This shows that δx (y) is piecewise smooth.
Then the assertion follows from [19], Lemma 2.
A.3. Generalization to the Multivariate Time Series
We briefly sketch how to generalize the results from the univariate to the multivariate case. The basic idea is to
reduce the multivariate case to the univariate case. In the following, we assume that x ∈ F m and y ∈ F n are two
d-variate time series and p = (p1 , . . . , pL ) ∈ Pm,n is a warping path between x and y with elements pl = (il , jl ).
First observe that a d-variate time series x ∈ F m consists of d individual component time series x(1) , . . . , x(d) ∈
Rm . Next, we construct the embeddings of a warping path. The d-variate time warping embeddings Φd : F n → F L
and Ψd : F m → F L induced by p are maps of the form
yj1
xi1
Φd (x) = ... ,
Ψd (y) = ... .
xiL
yjL
The maps Φd and Ψd can be written as
Φd (x) = Φx(1) , . . . , Φx(d)
Ψd (x) = Ψy (1) , . . . , Ψy (d) ,
where Φ and Ψ are the embedding matrices induced by p. Since Φ and Ψ are linear, the maps Φd and Ψd are also
linear maps. We show the multivariate formulation of Prop. 9.
Proposition 13. Let Φd and Ψd be the d-variate embeddings induced by warping path p ∈ Pm,n . Then
Cp (x, y) = kΦd (x) − Ψd (y)k2 .
for all x ∈ F m and all y ∈ F n .
Proof. The assertion follows from
kΦd (x) − Ψd (y)k2 =
d
X
2
Φx(k) − Ψy (k)
k=1
=
d
X
X
(k)
2
(k)
2
(k)
− yj
(k)
− yj
xi
k=1 (i,j)∈p
=
d
X X
xi
(i,j)∈p k=1
=
X
kxi − yj k 2 = Cp (x, y).
(i,j)∈p
Due to the properties of product spaces and product measures, the proofs of all other results can be carried out
componentwise.
12
References
[1] C. Banderier and S. Schwer. Why Delannoy numbers?
135(1):40–54, 2005.
Journal of Statistical Planning and Inference,
[2] V. Bogachev. Measure Theory. Springer-Verlag Berlin Heidelberg, 2007.
[3] M. Brill, T. Fluschnik, V. Froese, B. Jain, R. Niedermeier, D. Schultz, Exact Mean Computation in Dynamic
Time Warping Spaces. SIAM International Conference on Data Mining, 2018 (accepted).
[4] R. Caron and T. Traynor. The zero set of a polynomial. WSMR Report, University of Windsor, 2005.
[5] M. Cuturi and M. Blondel. Soft-DTW: a differentiable loss function for time-series. International Conference
on Machine Learning, 2017.
[6] P. Esling and C. Agon. Time-series data mining. ACM Computing Surveys, 45:1–34, 2012.
[7] L.C. Evans and R.F. Gariepy. Measure theory and fine properties of functions. CRC Press, 1992.
[8] T. Fu. A review on time series data mining. Engineering Applications of Artificial Intelligence, 24(1):164–181,
2011.
[9] P.R. Halmos. Measure Theory. Springer-Verlag, 2013.
[10] V. Hautamaki, P. Nykanen, P. Franti. Time-series clustering by approximate prototypes. International
Conference on Pattern Recognition, 2008.
[11] B. Jain. Generalized gradient learning on time series. Machine Learning 100(2-3):587–608, 2015.
[12] B. Jain and D. Schultz. Asymmetric learning vector quantization for efficient nearest neighbor classification
in dynamic time warping spaces. Pattern Recognition, 76:349-366, 2018.
[13] B. Jain. Warped-Linear Models for Time Series Classification arXiv preprint, arXiv:1711.09156, 2017.
[14] T. Kohonen and P. Somervuo. Self-organizing maps of symbol strings. Neurocomputing, 21(1-3):19–30, 1998.
[15] T. Kohonen. Self-Organizing Maps Springer-Verlag Berlin Heidelberg, 2001.
[16] J.B. Kruskal and M. Liberman. The symmetric time-warping problem: From continuous to discrete. Time
warps, string edits and macromolecules: The theory and practice of sequence comparison, 1983.
[17] F. Petitjean, A. Ketterlin, and P. Gancarski. A global averaging method for dynamic time warping, with
applications to clustering. Pattern Recognition 44(3):678–693, 2011.
[18] F. Petitjean, G. Forestier, G.I. Webb, A.E. Nicholson, Y. Chen, and E. Keogh. Faster and more accurate
classification of time series by exploiting a novel dynamic time warping averaging algorithm. Knowledge and
Information Systems, 47(1):1–26, 2016.
[19] R.T. Rockafellar. A Property of Piecewise Smooth Functions. Computational Optimization and Applications,
25:247–250, 2003.
[20] H. Sakoe and S. Chiba. Dynamic programming algorithm optimization for spoken word recognition. IEEE
Transactions on Acoustics, Speech, and Signal Processing, 26(1):43–49, 1978.
[21] D. Schultz and B. Jain. Nonsmooth analysis and subgradient methods for averaging in dynamic time warping
spaces. Pattern Recognition, 74:340–358, 2018.
[22] S. Soheily-Khah, A. Douzal-Chouakria, and E. Gaussier. Progressive and Iterative Approaches for Time
Series Averaging. Workshop on Advanced Analytics and Learning on Temporal Data, 2015.
[23] P. Somervuo and T. Kohonen, Self-organizing maps and learning vector quantization for feature sequences.
Neural Processing Letters, 10(2):151–159, 1999.
[24] Z. Xing, J. Pei, and E. Keogh. A brief survey on sequence classification. ACM SIGKDD Explorations
Newsletter, 12(1):40–48, 2010.
13
| 2cs.AI
|
arXiv:0906.3920v1 [cs.PL] 22 Jun 2009
Reasoning About a Service-oriented Programming Paradigm
Claudio Guidi1
Fabrizio Montesi1
Department of Computer Science, University of Bologna, Italy
italianaSoftware s.r.l., Italy
[email protected]
[email protected]
This paper is about a new way for programming distributed applications: the service-oriented one. It
is a concept paper based upon our experience in developing a theory and a language for programming
services. Both the theoretical formalization and the language interpreter showed us the evidence that
a new programming paradigm exists. In this paper we illustrate the basic features it is characterized
by.
1
Introduction
This paper is about a new way for programming distributed applications: the service-oriented one. It
is a concept paper based upon our experience in developing a theory and a language for programming
services. Our work started some years ago when we began to formalize the basic mechanisms of the
Web Services technology in a process calculus. We chose such an approach because process calculi
were naturally born for describing concurrent processes, as Web Services are. The need to address Web
Services with a formal approach was motivated by the high level of complexity they are characterized
by: we wanted a simple and precise means for catching their essentials and, at the same time, strong
foundations for developing concrete tools for designing and implementing service systems.
When we started to build our formal model we took inspiration from foundational calculi such as CCS
[22] and π -calculus [23], by enriching those approaches with specific mechanisms which came from the
Web Services technology. We developed SOCK [16, 13], that is a process calculus where those aspects of
service-oriented computing which deal with communication primitives, work-flow composition, service
session management and service networks are considered. SOCK is structured on three layers, each
one representing a specific feature of the service-oriented approach: the behaviour of a service, the
execution of service sessions into a service engine and the connection of services within a network. Such
a categorization allowed us to handle the complexity of service-oriented computing without loosing the
important details they are characterize by. Differently from our approach, other authors proposed SCC [6]
and COWS [20]. The main difference between SCC and SOCK can be found in the session identification
mechanism. In SOCK we identify sessions by means of correlation sets whereas in SCC sessions are
identified by freshly generated names. In our opinion, correlation set represents a key mechanism for
the service-oriented programming paradigm which is also modelled in COWS and it is provided by the
most credited orchestration language for Web Services: WS-BPEL [26]. The main difference between
SOCK and COWS is the fact that SOCK explicitly supports a state whereas COWS does not. Moroever,
SOCK encoded the RequestResponse communication primitive which is not supported by COWS. Both
state and the RequestResponse primitive made SOCK close to the technologies. This fact allowed us
to reason about fault handling issues which led us to propose a new way, the dynamic handling, for
managing faults [14, 24].
1 These
authors contributed equally to this work.
M.H. ter Beek (Ed.): Young Researchers Workshop
on Service-Oriented Computing 2009 (YR-SOC’09).
EPTCS 2, 2009, pp. 67–81, doi:10.4204/EPTCS.2.6
c C. Guidi & F. Montesi
This work is licensed under the
Creative Commons Attribution License.
Reasoning About a Service-oriented Programming Paradigm
68
At the time we created SOCK, we did not imagine that a new way for approaching distributed system
design could be developed starting from it. The evidence of this came some years later, when we started
to develop the JOLIE programming language [17, 25]. JOLIE was born as a strict implementation of
the semantics contained in SOCK. The syntax was inspired by that of the process calculus, blended with
some common constructs which are familiar to those accustomed to languages such as C and Java. The
JOLIE language allowed us to apply the concepts studied in SOCK in real world application development, and this raised new issues and in turn made new software design patterns to emerge. Addressing
these issues brought to the definition of new mechanisms for service system composition such as embedding and redirecting. Embedding increases the granularity level of services into a system, whereas
redirecting allows for grouping services under a unique endpoint. Both the approaches can be freely
mixed together in order to obtain new systems of services. Today, at the best of our knowledge, JOLIE
is the first language which allows for the designing of a distributed system completely composed by
services.
At the present, we can state that SOCK and JOLIE form a framework that offers the possibility
to study service-oriented computing issues from the theoretical, architectural and practical points of
view. The sum of these experiences made us aware of the fact that we were facing a new way for
designing, developing and studying distributed applications: the service-oriented paradigm. It is difficult
and ambitious to highlight the distinctive features of a new programming paradigm, but with this paper
we would like to share our strong impressions that the service-oriented paradigm exists.
In the following we describe the basic concepts that we have extracted and we try to show how they
can be considered the foundations of the service-oriented programming paradigm. Our work is strongly
influenced by that made by the industrial and scientific communities on service-oriented computing,
and strives to consider the most relevant results in our definition of the service-oriented approach. In
section 2 we present our definition of service, after introducing the main concepts that are behind it.
In section 3 we expose how one can compose services in order to obtain a distributed system in which
they communicate with each other. Section 4 contains some design pattern examples and remarks that
emerged from our experience in using the programming paradigm we propose. Finally, in section 5 we
report our conclusions and future works.
2
Definition of Service
The first thing to address is the definition of the term service. In order to provide evidence that a serviceoriented paradigm exists we need to define what a service is, as services are the most important component of the paradigm. The definition of service for the W3C Working Group [35] follows:
“A service is an abstract resource that represents a capability of performing tasks that form
a coherent functionality from the point of view of providers entities and requesters entities.
To be used, a service must be realized by a concrete provider agent.”
We agree with this definition but we argue that it is too abstract because too many things could be a
service. At the end of this section we present our definition of service, which is based upon the concepts
of behaviour, engine and service description.
2.1 Behaviour
Here we present the definition of behaviour of a service, which introduces two basic concepts: service
activities and their composition in a work-flow. Activities represent the basic functional elements of a
C. Guidi & F. Montesi
69
behaviour, whereas their composition represents the logical order in which they can be executed. Workflow composition is a key aspect of the service-oriented programming paradigm, which comes from the
most credited business process language for Web Services, WS-BPEL. In the following we present the
definition of behaviour:
The behaviour of a service is the description of the service activities composed in a workflow.
We distinguish three basic kinds of service activities:
• communication activities: they deal with message exchange between services;
• functional activities: they deal with data manipulation;
• fault activities: they deal with faults and error recovery.
2.1.1
Communication activities.
Communication activities are called operations. We inherit them from the Web Services Description
Language specifications (WSDL) [32]. Operations can be divided into input operations and output operations. The former operations provide a means for receiving messages from an external service whereas
the latter ones are used for message sending.
• Input operations:
– One-Way: it is devoted to receive a request message.
– Request-Response : it is devoted to receive a request message and to send a response message back to the invoker.
• Output operations
– Notification: it is devoted to send a request message.
– Solicit-Response: it is devoted to send a request message and to receive a response message
from the invoked service.
Such a categorization is also presented in [5, 4], even if other authors consider only the single message
exchange pattern (represented by One-Way and Notification) as in [20]. Here we consider the RequestResponse and Solicit-Response key interaction patterns for the service-oriented programming paradigm
because they introduce specific issues from an architectural point of view. We will clarify such an aspect
in Section 4.
Output operations require the specification of a target endpoint to which the message has to be sent.
At the level of behaviour such an endpoint abstractly refers to a service. Here, we call it receiving service.
2.1.2
Functional activities.
They allow for the manipulation of internal data by providing all the basic operators for expressing
computable functions.
2.1.3
Fault activities.
They allow for the management of faults. This is a fundamental aspect of service-oriented programming.
The following list of basic activities for managing faults has been extracted from the experience we have
developed on dynamic handling in SOCK and JOLIE [14, 24] and other models and languages such as
StAC [11], SAGAS [8] and WS-BPEL.
Reasoning About a Service-oriented Programming Paradigm
70
• Fault raising: it deals with the signaling of a fault
• Fault handler: it defines the activities to be performed when a fault must be handled
• Termination handler: it defines the activities to be performed when an executing activity must be
terminated before its ending.
• Compensation handler: it defines the activities to be performed for recovering a successfully
finished activity.
2.2 Engine
Here we present the concept of engine that we introduced in SOCK for dealing with those aspects of the
service-oriented programming paradigm related to the actual execution of a service into a network. To
the best of our knowledge there is no clear and precise definition of engine as that which we presented
in SOCK. The motivations of such a lack probably reside in the fact that it is usually considered an
implementation detail that does not add anything relevant to service-oriented models. As far as serviceoriented computing is concerned, engines are generally associated to those which execute WS-BPEL
such as, for example, ActiveVOS [1] or the Oracle one [27], but what about simple Web Services developed in Java, Python or .NET? Can we consider usual web servers such as IIS [21], Apache [2] or
Zope [36] as service engines? Can we identify some characteristics on these applications and consider
them basic features of the service-oriented programming paradigm? In our perception the answer is yes
and this section is devoted to highlight the main features a service engine is characterized by. In general we say that an engine is a machinery able to create, execute and manage service sessions. A more
detailed definition of engine will be provided at the end of this section, but we need to introduce some
concepts first. Let us see the concept of session.
2.2.1
Session.
The definition of session follows:
A service session is an executing instance of a service behaviour equipped with its own local
state.
A key element of the service-oriented programming paradigm is session identification. In general a
session is identified by its own local state or a part of it. The part of the local state which identifies a
session can be programmed and it is called correlation set. Correlation sets is a mechanism provided
by WS-BPEL and it has been formalized in SOCK, COWS and in [29]. We chose to allow for the
definition of correlation sets even in JOLIE. In order to explain such a mechanism, let us introduce a
simple notation, where a session is represented by a couple of terms (P, S) where P represents a behaviour
in a given formalism and S represents the local state here modelled as a function from variables to names.
S : Var → Values where Var is the set of variables and Values the set of values1 . Now, let us consider
two sessions with same behaviour but different local states S1 and S2 :
s1 := (P, S1 ) s2 := (P, S2 )
1 For the sake of brevity, here we consider both states and messages as a flat mappings from variables to values. The
introduction of structured and typed values does not alter the correlation set insights presented in this section.
C. Guidi & F. Montesi
71
We say that s1 is not distinguishable from s2 if S1 = S2 2 . Now, let us consider both S1 and S2 as a
composition of states defined on disjoint domains:
S1 = S11 ⊕ S12
S2 = S21 ⊕ S22
where the operator ⊕ represents a composition operator over states3 . Let us consider S11 and S21 as the
correlation sets for s1 and s2 respectively. We say that s1 is not distinguishable by correlation from s2 if
S11 = S21 .
Such a session identification mechanism differs from that of the object-oriented paradigm for objects.
Indeed, in the object-oriented approach an object is always identified by the reference issued at the
moment of the object creation. Such a reference cannot be used in service-oriented computing, because
of its loosely coupled nature. This fact represents a major difference between the two approaches.
2.2.2
Session management.
Session management involves all the actions performed by a service engine in order to create and handle
sessions. In order to achieve this task, a service engine provides the following functionalities:
• Session creation. Sessions can be created in two different ways:
– when an external message is received on a particular operation of the behaviour. Some operations can be marked as session initiators. When a message is received on a session initiator
operation, a session can be started.
– when a user manually starts it. A user can launch a service engine which immediately executes a session without waiting for an external message. We denote such a session as firing
session.
• State support. The service engine also provides the support for accessing data which are not
resident into session local states: the global state and the storage state. Summarizing, it is possible
to distinguish three different kind of data resources, which we call states, that can be accessed and
modified by a session:
– a local state, which is private and not visible to other sessions. This state is deleted when the
session finishes.
– a global state which is shared among all the running sessions. This state is deleted when the
engine stops.
– a storage state which is shared among all the running sessions and whose persistence is
independent from the execution of a service engine (e.g. a database or a file).
• Message routing. Since a session is identified by its correlation set, the engine must provide
the mechanisms for routing the incoming messages to the right session. The session identification issue is raised every time a message is received. For the sake of generality, in the serviceoriented paradigm we cannot assume that some underlying application protocol such as WSAddressing [34] or other transport protocol identification mechanisms such as HTTP cookies [18]
are always used for identifying sessions. We consider correlation sets as the representative mechanism for routing incoming messages. Its functioning can be summarized as it follows. A message
can be seen as a function from variables to values: M ∈ Σ. As we have done for the states we
2 Let
Σ be the set of states, we define =: (Σ × Σ) where S1 = S2 if S1 (x) = S2 (x) and Dom(S1 ) = Dom(S2 )
→ Σ where S1 ⊕ S2 (x) = S1 (x) i f x ∈ Dom(S1 ), S2 (x) i f x ∈ Dom(S2 ) ∧ x ∈
/ Dom(S1 ), unde f ined otherwise
3⊕ : Σ × Σ
Reasoning About a Service-oriented Programming Paradigm
72
can define a correlation set also for a message. Let us consider M = M1 ⊕ M2 where M1 is the
correlation set for the message M. We define the correlation function c : Var → Var which allows
us to map message variables to state variables. We say that a message M must be routed to the
session s whose state is S = S1 ⊕ S2 where S1 is the correlation set, if:
∀x ∈ Dom(M1 ), c(x) ∈ Dom(S1 ) ⇒ S(c(x)) = M(x) ∨ S(c(x)) is unde f ined
Informally, a message can be routed to a session only if its correlated data correspond to those of
the session. The correlation function c is the concrete means used by the programmers for defining
correlation. For each incoming message it will be possible to define a specific function c and the
correlation set, which identifies the session, is indirectly defined by the union of the codomains of
all the defined c functions. It is worth noting that, if the correlation set is not correctly programmed,
more than one running session could be correlated to an incoming message. In this case the session
which has to receive the message is non-deterministically selected.
• Session execution. Session execution deals with the actual running of a created session behaviour
equipped with all the required state supports. Sessions can be executed sequentially or concurrently. The majority of existing technologies share the idea that sessions are to be executed concurrently, but the sequential case allows for the controlling of some specific hardware resource
which needs to be accessed sequentially. As an example, consider a cash withdrawal machine
which starts sessions sequentially due to its hardware nature. We consider such an aspect fundamental from an architectural point of view because it can raise system deadlock issues if not
considered properly, as we have shown in [15] by means of SOCK.
2.2.3
Engine definition.
Now we can provide our definition for service engine:
An engine is a machinery able to manage service sessions by providing session creation,
state support, message routing and session execution capabilities.
2.3 Service Description
A service description provides all the necessary information for interacting with a service. Service descriptions are composed by two parts: interface and deployment.
2.3.1
Interface.
Service interfaces contain abstract information for performing compatibility checks between services,
abstracting from low-level details such as communication data protocols and transports. We identify
three different levels of service interface:
• Functional. It reports all the input operations used by the behaviour for receiving messages from
other services or applications. An operation description is characterized by a name, and its request
and response message types. If we trace a comparison with the Web Services technology, such
an interface level is well represented by the WSDL specifications v1.1 [32]. At this level, only
message type checks on the interface are required in order to interact with the service.
C. Guidi & F. Montesi
73
• Work-flow. It describes the work-flow of the behaviour. In a work-flow, input operations could
not be always available to be invoked but they could be enabled by other message exchanges by
implementing a sort of high level application protocol. Thus, it is fundamental to know how a
service work-flow behaves in order to interact with it correctly. In the Web Services technology
such an interface could be provided by means of an Abstract-BPEL [26] description; WSDL 2.0
specifications [33] provide Message Exchange Patterns (MEP) which allows for the description
of custom service interaction patterns, too. At this level, other approaches such as choreography
must be considered. Choreography languages, such as WS-CDL [31], allow for the designing of
a service system from a global point of view. In this case, a choreography can be exploited for
describing the work-flow behaviour of a service by highlighting its role into the choreography.
This is a wide area of investigation that involves works on contracts [7] and conformance [10], but
it is out of the focus of this paper. The reader interested in this topic may consult [12, 9, 3, 19].
• Semantics. It offers semantic information about the service and the specific functionalities provided by it. It is usually provided by using some kind of ontology such as OWL-S [30].
Service interfaces are strictly related to service discovery, which is a key element of the serviceoriented programming paradigm. Discovery issues are strictly related to search and compatibility check
algorithms over interface repositories, also called registries. These algorithms differ depending on which
interface type is considered. It is out of the scope of this paper to discuss the different methodologies
used for implementing service discovery. However, here we want to highlight the fact that, at the present,
some information related to service engine do not appear on the current interface standard proposals, such
as the sequential execution modality or the correlation set of a service. These information are relevant
and should be specified into the interface because they could influence other services in the system, as
we have hinted in the previous section.
2.4 Deployment
The deployment phase is in charge of binding the service interface with network locations and protocols.
A service, for example, could receive messages by exploiting the HTTP protocol or the SOAP over
HTTP one, but the choice is potentially unlimited because new protocols may be created. Such a task
is achieved by means of port declarations. There are two kind of ports: input ports and output ports.
The former allows for the declaration of all the input endpoints able to receive messages exhibited by
the service engine, whereas the latter bind target location and transport protocol to the receiving services
of the behaviour. In other words the output port allows for the concrete connection with the services to
invoke. In general, we define a port as it follows:
A port is an endpoint equipped with a network address and a communication protocol joined
to an interface whose operations will become able to receive or send requests. Ports which
enables operations to receive requests are called input ports, output ports otherwise.
A service engine needs to be joined with deployment information in order to receive and send messages.
2.5 Service
Now, we are able to provide our definition of service which follows:
A service is a deployed service engine whose sessions animate a given service behaviour.
Reasoning About a Service-oriented Programming Paradigm
74
Figure 1: A firing session starting a system of services.
Such a definition is not enough if we consider the concrete execution of a service in a complex system.
In order to complete our model we introduce the concept of container. The definition of container is
fundamental in the analysis of the system composition mechanisms described in the next section.
A service container is an application able to execute one or more services.
3
Service Composition
Now that we have a definition of service and service container we can discuss the meaning of service
composition. The term composition discussed in Section 2.1 concerns only the inner activities of a behaviour, which can be composed in a work-flow. Here we extend the usage of the term composition at
the level of service system, by showing the different mechanisms we experimented for putting services
together into a system. To the best of our knowledge, this is the first attempt to classify different kinds of
composition at the level of service system. These mechanisms became evident when we tried to implement the concepts of behaviour and engine, discussed in the previous section, in JOLIE, where we faced
the challenge to offer to the programmer an easy way for building modular applications based solely on
the concept of service. The following kinds of service composition emerged from our experience:
• simple composition
• embedding
• redirecting
• aggregation
These composition techniques can be freely mixed together in order to obtain different architectures. The
different composition techniques do not alter the functionalities of a service system but they allow for
the engineering of the system architecture depending on the design needs. It is important to notice that,
regardless of the kinds of composition that are used, one must ensure that at least one service present in
the composition starts with a firing session. This is necessary because the services not containing a firing
session always have to wait for an input from an external process in order to be started. In order to clarify
the concept of firing session let us consider Fig. 1 where we represent the firing session as a ball which
starts a set of domino cards. The spatial placement of the cards represents the dependencies among the
services of the system started by the firing session. In the following we use the term client for referring
to a service that calls another service.
C. Guidi & F. Montesi
75
Container
Service B
Service A
Service C
Figure 2: Services A and B are embedded and their interactions do not exploit the network but they are
performed within the container. Service C is not embedded and its interaction with service A exploits
the network.
3.1 Simple Composition
Simple composition is the parallel execution of service containers into a network where they
are able to communicate with each other.
In a simple composition the resulting system will behave accordingly to the behaviour of each service
involved in. This is the most obvious way for composing services.
3.2 Embedding
Embedding is the composition of more than one service into the same container.
The main advantage of such a composition is the fact that embedded services could communicate with
each other without using the network but exploiting inner communication mechanisms depending on the
implementing technology (e.g. RAM communications). Embedding is particularly suitable for increasing the level of granularity of a service-oriented system. Sending and receiving messages over a network
indeed can be a strong limitation when we consider auxiliary services which provide basic functionalities
(e.g. services which provide mathematical functions, services that manage time and dates, etc.). By composing with embedding it is possible to use auxiliary services without deploying them into the network.
In Fig. 2 is represented the case where a service A requires the functionalities of one other service B but
it does not make sense to deploy B on the network and consequentially expose it to external processes. A
and B are embedded and their interactions are performed within the container. On the other hand, A can
also interact via network with another service C which is not embedded.
3.3 Redirecting
Redirecting allows for the creation of a master service acting as a single communication
endpoint to multiple services which are called resources.
In redirecting a master service receives all the messages of a system and then forward them to system
services. It is obtained by binding an input port of the master service to multiple output ports, each one
identifying a service by means of a resource name. The client will send messages to the master service
specifying the final resource service to invoke. The main advantages of such an approach are:
• the possibility to provide a unique access point to the system clients. In this way the services of
the system could be relocated and/or replaced transparently w.r.t. the clients;
• the possibility to provide transparent communication protocol transformations between the client
and the master and the master and the rest of the system.
Reasoning About a Service-oriented Programming Paradigm
76
/A
A
/B
B
M/?
M
/C
C
Figure 3: Service M redirects messages to services A, B and C depending on the target destination of the
message (M/A, M/B or M/C).
op1
op1@A
A
op1
op2
op3
op2
B
op1@M
M
op3
C
Figure 4: In aggregation the master publishes the union of all the service interfaces it aggregates. Interfaces are here represented with dotted rectangles. The message on operation op1 to service M is actually
redirected to service A.
In order to understand the second advantage better, consider Fig. 3 and suppose that A speaks a certain
protocol pa . Now suppose that a client needs to interact with A, but it does know only a different protocol:
pm . The client could then call M with destination M/A using protocol pm (known by M), and leave to M
the task of transforming the call message into an instance of pa before sending it to A.
3.4 Aggregation
Aggregation is a redirecting composition of services whose interfaces are joined together
and published as unique.
Aggregation deals with the grouping of more services under the same interface. It is similar to redirecting, but the resource services are not visible from the point of view of the client. The client sees a unique
service, the master one, which exhibits an interface by providing the functionalities of the resource services. Differently from redirecting, which maintains the different interfaces of each composed service
separated, in this case the client looses the details of each single service used behind aggregation. The
main advantage of such a composition approach deals with the possibility to completely hide the system
components to the client.
C. Guidi & F. Montesi
77
3.5 Dynamic System Composition
The aforementioned service composition techniques (simple, embedding, redirection and aggregation)
can be used both statically and at runtime. In the static case all the services are composed before their
execution and the composition never changes during the execution of all the system. On the contrary,
if the composition of the system changes at runtime, we say that the system is composed dynamically.
Dynamic composition is strictly related to the concept of service mobility. Service mobility deals with
the representation of a service in some data format, its transmission from one service to another and then
its execution in the service container of the receiver. It is worth noting that here we do not consider the
case of running service migration but only the case of static service behaviour mobility. In the following
we briefly describe some case of dynamic composition:
3.5.1
Dynamic embedding.
Let us consider a service which needs to receive software updates for a certain functionality. One may
encapsulate that functionality in an embedded service. Then, when a software update is issued, the
embedder service may unload the embedded one, receive the updated service to embed and dynamically
embed the received service.
3.5.2
Dynamic redirecting and aggregation.
Let us consider the case that a resource service faults or needs maintenance without affecting the service
availability from the client point of view. It is sufficient to install a spare part resource service and
registering it to the master service in place of the faulty one.
4
Discussion
The definitions of service and service composition shown so far are the key elements of the serviceoriented programming paradigm. Such a paradigm becomes evident when building a system of services
where it is possible to construct distributed system architectures by following new design patterns. In
this section we present and discuss some of them.
Request-Response.
The Request-Response interaction pattern plays a fundamental role when designing a service system
architecture. Such a message exchange between two peers can be performed in two different ways: by
means of a callback or a Request-Response. In the former case both services exhibit a port for receiving
the message, whereas in the latter case only the port of the Request-Response receiver is exhibited and
the response message is sent on the same channel4 on which the request message is received. In Fig. 5
we graphically show the differences between a callback configuration and a Request-Response one. The
main difference is that in the callback configuration the service A must be aware of the fact that it has to
receive the response message in a specific port. In other words, service A must be aware of the workflow behaviour of B, so it needs to know the work-flow interface of B in order to interact with it. In the
4 It is out of the scope of this paper to provide a precise definition of channel. For the sake of our discussion, it is sufficient
to informally consider a channel as a socket or an abstract channel like the one described in the WS-Addressing specification.
Reasoning About a Service-oriented Programming Paradigm
78
Callback
A
Request-Response
B
A
B
Figure 5: Callback and Request-Response configurations. Black rectangles represent ports.
Master service mobility
Slave service mobility
M
op1
op2
...
S
S
op1
op2
...
M
Figure 6: Slave service mobility and Master service mobility.
Request-Response case, it is sufficient for A to know the functional interface of B where the RequestResponse operation is declared in order to interact with it. Such architectural difference motivated us to
consider both the approaches as fundamental for the service oriented programming paradigm. Moreover,
the Request-Response pattern raises interesting issues from the fault handling point of view but, for the
sake of brevity, we do not discuss them here and we invite the interested reader to consult [14, 24].
Web client-server pattern.
It is easy to model the web client-server pattern in the service-oriented programming paradigm. A web
server is a service deployed on an HTTP port, where it exhibits some general purpose Request-Response
operations, such as GET and POST. Web servers usually manage session identification by means of
cookies or query strings, which can be easily modelled with the correlation set mechanism.
Slave service mobility and Master service mobility.
Here, we show two design patterns we experimented: slave service mobility and the master service
mobility. We consider slave the service which provides simple functionalities by means of RequestResponse operations and master the service which makes use of multiple slave services in its work-flow
behaviour. In the slave service mobility pattern the slave services are moved and embedded to the master
service container, whereas in the master service mobility the master service is moved and embedded into
the container of the slave. In the following we explain both the patterns by referring to Fig. 6
• Slave service mobility. Consider the case in which a service M (the master service) defines a workflow that is dependent on some functionalities that cannot be provided statically before execution
time. M, instead, needs to obtain these functionalities at runtime and to use them. In order for this
to work, M must define an appropriate output port for the functionalities it is looking for. Then,
M asks a service repository for downloading the functionalities it needs. The repository sends a
service S to M, and M dynamically embeds S. M has now access to the functionalities offered by
S, the slave service, and exploit them to complete its work-flow.
C. Guidi & F. Montesi
79
• Master service mobility. Consider the case in which a service S (the slave service) possesses
the functionalities that are needed for the actualization of a work-flow, but the work-flow cannot
be provided statically before execution time. S needs to obtain the work-flow at runtime and
to execute it, ensuring that the work-flow makes use of the functionalities provided by S. We
exploit such a pattern for implementing the SENSORIA automotive scenario [28] where a car
experiments a failure and starts a recovery work-flow for booking some services such as the garage,
the car rental and the truck one. We implemented it with a slave service on the car and the master
work-flow which is downloaded from the car factory assistance service. In this way we obtain
that the recovery work-flow can be changed and maintained by the assistance car service without
updating all the car softwares periodically and, at the same time, we guarantee transaction security
by isolating some functionalities such as the bank payment, into the slave service of the car. In this
case the downloaded work-flow is able to search for all the services it needs but it relies upon the
slave service car functionalities for the payment.
SoS: Service of services pattern.
The SoS pattern exploits both dynamic embedding and dynamic redirecting. A service is embedded
at run-time if a client performs a resource request to the master service. In this case the master service
embeds the requested service by downloading it from a repository and make it available to the client with
a private resource name. From now to the end, the client will be able to access to its own resource by
simply addressing its requests to the resource name it has received. The main advantage of this approach
is that we can provide an entire service as a resource to a specific client instead of a single session of a
service. We exploit this pattern for implementing the SENSORIA finance case study [28] which models
a finance institute where several employees works on the same data. We use the SoS pattern for loading
a service for each employee which maintain its private data and, at the same time, can offer a set of
functionalities. The main advantage in this approach is that each functionality offered to the employee
is able to open a session on its own service thus obtaining a private complex resource made of a set of
service sessions.
5
Conclusions
We have presented the results of our experience in investigating service-oriented applications both from a
theoretical and a practical point of view. It is our opinion that a service-oriented programming paradigm
exists and that it is characterized by the concepts of behaviour, session, session execution, correlation set,
engine, interface, deployment, service, container, embedding, redirecting and aggregation. Upon these
key concepts we experimented some interesting design patterns from an architectural viewpoint, such
as the slave service mobility, the master service mobility and the SoS. They probably are only some of
the possible design patterns which can be built with the service-oriented programming paradigm. In the
future we will continue to investigate the service-oriented programming paradigm by also considering
choreography languages as complementary means for designing service-oriented systems. As far as
JOLIE is concerned, we will continue in its development by studying and implementing all the necessary
mechanisms for obtaining the aggregation service composition pattern and tools for improving usability
and property checking.
80
Reasoning About a Service-oriented Programming Paradigm
Acknowledgments
Research partially funded by EU Integrated Project Sensoria, contract n. 016004.
References
[1] Active Endpoint: ActiveVOS. http://www.activevos.com/.
[2] Apache software foundation: Apache. http://www.apache.org/.
[3] M. Baldoni, C. Baroglio, A. Martelli, V. Patti & C. Schifanella (2005): Verifying the Conformance of Web
Services to Global Interaction Protocols: A First Step. In: EPEW/WS-FM, Lecture Notes in Computer
Science 3670. Springer-Verlag, pp. 257–271.
[4] A. Barros & E. Borger (2005): A Compositional Framework for Service Interaction Patterns and Interaction
Flows. In: Proc. of International conference on formal engineering methods (ICFM’05), LNCS. Springer
Verlag, pp. 5–35.
[5] A. Barros, M. Dumas & A.H.M. ter Hofstede (2005): Service Interaction Patterns: Towards a Reference
Framework for Service-Based Business Process Interconnection. Technical Report FIT-TR-2005-02, Faculty
of information Technology, Queensland University of Technology, Brisbane, Australia.
[6] M. Boreale, R. Bruni, L. Caires, R. De Nicola, I. Lanese, M. Loreti, F. Martins, U. Montanari, A. Ravara,
D. Sangiorgi, V. Vasconcelos & G. Zavattaro (2006): SCC: A Service Centered Calculus. In: Proc. of Web
Services and Formal Methods Workshop (WS-FM’06), LNCS 4184. Springer-Verlag, pp. 38–56.
[7] M. Bravetti & G. Zavattaro (2007): Towards a Unifying Theory for Choreography Conformance and Contract
Compliance. In: 6th International Symposium on Software Composition (SC’07), LNCS. pp. 34–50.
[8] R. Bruni, H. Melgratti & U. Montanari (2005): Theoretical foundations for compensations in flow composition languages. In: Proc. of POPL’05. ACM Press, pp. 209–220.
[9] N. Busi, R. Gorrieri, C. Guidi, R. Lucchi & G. Zavattaro (2005): Towards a formal framework for Choreography. In: Proc. of 3rd International Workshop on Distributed and Mobile Collaboration (DMC’05). IEEE
Computer Society Press, pp. 107–112.
[10] N. Busi, R. Gorrieri, C. Guidi, R. Lucchi & G. Zavattaro (2006): Choreography and Orchestration conformance for system design. In: Proc. of 8th International Conference on Coordination Models and Languages
(COORDINATION’06), LNCS 4038. pp. 63–81.
[11] M.J. Butler & C. Ferreira (2004): An Operational Semantics for StAC, a Language for Modelling LongRunning Business Transactions. In: Proc. of COORDINATION’04, Lecture Notes in Computer Science
2949. springer, pp. 87–104.
[12] M. Carbone, K. Honda & N. Yoshida (2007): Structured Communication-Centred Programming for Web
Services. In: Proc. of ESOP’07, Lecture Notes in Computer Science 4421. Springer-Verlag, pp. 2–17.
[13] C. Guidi (2007): Formalizing languages for Service Oriented Computing. PhD. thesis, Department of Computer Science, University of Bologna. http://www.cs.unibo.it/pub/TR/UBLCS/2007/2007-07.pdf.
[14] C. Guidi, I. Lanese, F. Montesi & G. Zavattaro (2008): On the interplay between fault handling and requestresponse service invocations. In: Proc. of ACSD’08. IEEE Press, pp. 190–199.
[15] C. Guidi & R. Lucchi (2008): Programming service oriented applications. Technical Report UBLCS-200811, Department of Computer Science, University of Bologna. http://www.cs.unibo.it/pub/TR/UBLCS/
2008/2008-11.pdf.
[16] C. Guidi, R. Lucchi, R. Gorrieri, N. Busi & G. Zavattaro (2006): SOCK: A Calculus for Service Oriented Computing. In: Proceedings of the 4th International Conference on Service-Oriented Computing
(ICSOC’06), Chicago, IL, USA, LNCS 4294. pp. 327–338.
[17] Jolie—Java Orchestration Language Interpreter Engine: http://www.jolie-lang.org.
[18] D. Kristol & L. Montulli: HTTP State Management Mechanism. http://www.w3.org/Protocols/
rfc2109/rfc2109.
C. Guidi & F. Montesi
81
[19] I. Lanese, C. Guidi, F. Montesi & G. Zavattaro (2008): Bridging the Gap between Interaction- and ProcessOriented Choreographies. In: 6th IEEE International Conferences on Software Engineering and Formal
Methods (SEFM’08). IEEE Computer Society. To appear.
[20] A. Lapadula, R. Pugliese & F. Tiezzi (2006): A WSDL-Based Type System for WS-BPEL. In: SpringerVerlag, editor: Proc. of 8th International Conference on Coordination Models and Languages (COORDINATION’06), LNCS 4038. pp. 145–163.
[21] Microsoft: Internet Information Services. http://www.microsoft.com/windowsserver2008/en/us/
internet-information-services.aspx.
[22] R. Milner (1989): Communication and Concurrency. Prentice Hall.
[23] R. Milner, J. Parrow & J. Walker (1992): A Calculus of Mobile Processes, I and II. Information and Computation 100(1), pp. 1–40, 41–77.
[24] F. Montesi, C. Guidi, I. Lanese & G. Zavattaro: Dynamic fault handling mechanisms for service oriented
applications. In: Proc. of 6th IEEE European Conference on Web Services (ECOWS’08). IEEE Computer
Society.
[25] F. Montesi, C. Guidi, R. Lucchi & G. Zavattaro (2007): JOLIE: a Java Orchestration Language Interpreter
Engine. Electr. Notes Theor. Comput. Sci. 181, pp. 19–33.
[26] OASIS: Web Services Business Process Execution Language Version 2.0. http://docs.oasis-open.org/
wsbpel/.
[27] Oracle: Oracle BPEL Process Manager. http://www.oracle.com/technology/products/ias/bpel/
index.html.
[28] Software Engineering for Service-Oriented Overlay Computers. http://www.sensoria-ist.eu/.
[29] M. Viroli (2004): Towards a Formal Foundation to Orchestration Languages. In: M. Bravetti & G. Zavattaro,
editors: Proc. of 1st International Workshop on Web Services and Formal Methods (WS-FM’04), Electr.
Notes Theor. Comput. Sci. 105.
[30] W3C: OWL-S: Semantic Markup for Web Services. http://www.w3.org/Submission/OWL-S/.
[31] W3C: Web Services Choreography Description Language Version 1.0. http://www.w3.org/TR/2004/
WD-ws-cdl-10-20040427/.
[32] W3C: Web Services Description Language (WSDL) 1.1. http://www.w3.org/TR/wsdl.
[33] W3C: Web Services Description Language (WSDL) Version 2.0 Part 1: Core Language. http://www.w3.
org/TR/wsdl20/.
[34] W3C Recommendation 9 May 2006: Web Services Addressing, Core 1.0. http://www.w3.org/TR/
ws-addr-core/.
[35] Web Services Glossary: http://www.w3.org/TR/ws-gloss/.
[36] Zope: http://www.zope.org/.
| 6cs.PL
|
Autonomous Driving in Reality with Reinforcement Learning and Image
Translation
Nayun Xu
Bowen Tan
Bingyu Kong
Shanghai Jiao Tong University
arXiv:1801.05299v1 [cs.CV] 13 Jan 2018
{xunayun, tanbowen, crystal mis}@sjtu.edu.cn
Abstract
Supervised learning is widely used in training autonomous driving vehicle. However, it is trained with large
amount of supervised labeled data. Reinforcement learning can be trained without abundant labeled data, but we
cannot train it in reality because it would involve many unpredictable accidents. Nevertheless, training an agent with
good performance in virtual environment is relatively much
easier. Because of the huge difference between virtual and
real, how to fill the gap between virtual and real is challenging. In this paper, we proposed a novel framework of
reinforcement learning with image semantic segmentation
network to make the whole model adaptable to reality. The
agent is trained in TORCS, a car racing simulator.
1. Introduction
In the artificial intelligence field, autonomous driving is a
significant task and closely relevant to computer vision. The
intention of this task is to design a system to autonomously
control vehicles to do actions, such as steering, accelerating,
and braking. Essentially, autonomous driving is an interactive model with the environment. In this task, on one hand,
computer vision techniques can help the system extract and
analyze information from driving scene, On the other hand,
this task also requires many other techniques to do the decision making.
There are mainly two categories of methodologies to
deal with this task, supervised learning and reinforcement
learning. Both methods face obstacles. When using supervised learning to do the autonomous driving, the most crucial problem is training relies on a large amount of labeled
data which requires much human effort. Also, it is hard
to develop an end-to-end model because the ground truth
for action is not objectively determined and the label on action would include personal bias. And state-of-the-art methods not using an end-to-end style are mostly very complex
involving many empirical rules. Therefore, reinforcement
learning seems better fits this task because it can be trained
without human labeled training data. Reinforcement learning avoids the problem of a large amount of labeled data
and the potential shortcoming of human bias.
Nevertheless, reinforcement learning also meets many
problems. The most fundamental problem is that reinforcement learning model cannot be trained in reality because the
training process would involve many collisions and other
unpredictable situations. So most reinforcement learning
models on this task are trained in virtual simulators. The
approach of training also brings about some problems. The
models performance when it is applied in reality will largely
depend on how real the simulator is. In other words, the real
driving scene is more complicated than the virtual simulator, which also challenges the generalization ability of the
model.
In this paper, our approach is focused on how to fill the
gap between virtual and real. We want to solve mainly two
problems. One is the difference between the virtual simulator and real in terms of driving scenes. The other is the
complexity and noise of reality scenes. Therefore, we use
a translation network to transfer the virtual driving scene to
semantic segmentation image and use these semantic images as state input to the agent. When applying the model
to reality, we do semantic segmentation on the real driving
scene and use the segmentation result as input to our model.
We consider this translation would fill the gap between virtual and real. Also, we consider the semantic segmentation
would be an appropriate level of abstraction of the real driving scene which reduces the complexity and still holds most
useful information such as lanes and barriers.
Our framework has several advantages as below:
• Compared with the huge demand for labeled data in
state-of-art supervised learning, our framework does
not relies on any labeled data.
• Training in a virtual environment and transfer to the
real world, we do not need to confront the danger and
enormous loss of failure.
Figure 1. The reinforcement learning agent is trained as following: Initially, a virtual image is produced by simulator TORCS.Then this
image go through an image semantic segmentation network. Finally the semantic representation of current driving scene is fed into the
reinforcement learning agent. In reinforcement learning framework, agent observes the semantic image and chooses an action. After agent
taking action, the environment will give a reward to agent to help the agent adjust its parameters and give an image of next state to the
semantic segmentation network
• The input of the reinforcement learning agent is semantic segmentation image. Semantic image contains
less information compared to original image, but includes most information needed by agent to take actions. In other words, semantic image neglects useless
information in original image.
2. Related Work
2.1. Supervised Learning for Autonomous Driving
Supervised learning has been used in autonomous driving for decades. And these work can be categorized into
two major styles, perception-based approaches and end-toend approaches. The perception-based approaches detect
some mediate information to help the agent make decisions.
In early years, they detect driving-relevant objects such as
lanes, cars, pedestrians, etc. Recently, approaches like deep
driving involve CNNs [2] to detect more direct information
like distances between cars and lanes.
The end-to-end approaches seem more direct. They want
to directly map images input to driving action predictions.
ALVINN [16] shows an early attempt of end-to-end approaches. It learns the direct mapping with a shallow neural network. And in recent years, the shallow network has
been replaced by more powerful deep neural networks like
CNNs. NVIDIA [5]recently presented an end-to-end system with deep learning methods [3].
However, whichever styles of supervised learning approaches are employed, the training process requires large
quantities of labeled data. And the model performance is
highly relevant to the quality and quantity of data. Besides,
supervised models have limited generalization abilities because the real driving environments are numerous which are
far beyond the training data.
2.2. Reinforcement learning for Autonomous Driving
With a lot of variations, reinforcement learning has been
a common technique for many scenarios such as computer
games [15] and robot control[10, 8]. Recently, plenty
of work [1, 19] contributed to building an autonomous
driving system with reliable security. However, highdimensionality of state space and non-trivial large action
range in the real world’s practical driving environments are
challenging the training of reinforcement learning. It is
time-consuming to get an optimal policy over such high
complexity. With the power of deep neural networks and
deep reinforcement learning[11, 15, 18, 12, 14], a great step
forward in such complexity is made recently. Nonetheless,
not only deep Q-learning[15] method but policy gradient
method[12] , they both require the interaction between the
agent and environment to get feedback and reward. Obviously, training agents of an autonomous vehicle in realworld scenes is unrealistic because of the huge cost for every wrong action.
In order to avoid damage to the real world, reinforcement
learning with driving simulators and transfer learning models appear. The training process of reinforcement learning
with a driving simulator is safe and fast. However, in order
to drive autonomously in the real world, the driving agent
Figure 2. Examples of translation from the original output got from simulator to our intended input for the reinforcement learning agent.
The first column is the original scene displayed by TORCS, the second column is the corresponding first-person perspective scene got by
hacking from the source code, the third column is the semantic segmentation of the first-person perspective scene, the fourth column is the
gray scale semantic perspective.
must be able to take actions according to an intricate visual
field, which is much different from the virtual images we
get from a driving simulator. Models trained on virtual data
and simulator cannot perform well in real-world data.
For the past decade, there are many models [17, 9, 20].
These models either first train a model in virtual environment and then fine-tune in the real environment [17], or
learn an alignment between virtual images and real images
by finding representations that are shared between the two
domains [21] contribute to the transferring reinforcement
learning. Some of these models first trained on virtual data
to reduce time training in real-world significantly. But these
models still have to train in real-world, they cannot avoid
the risk of damage to real-world radically. Some other models try to learn an alignment between the virtual image and
real image. In reality, the real visual field is much more
complicated and more noisy than virtual image. Challenge
of these models are, under a certain condition, an alignment
which is good enough cannot be found between virtual images and real images.
There is a recent work [23] managing to train the rein-
forcement learning only in environment created by simulator TORCS[22]. This novel framework demonstrated that
after training on a simulator, the autonomous agent is able to
realize a collision-free flight. Nonetheless, the work needs
a nontrivial training environment to achieve its goal.
2.3. Scene Parsing
Semantic image segmentation is one part of our model. It
can be seen as a pixel-level prediction task. Based on a deep
convolutional neural network and fully convolutional neural network[13], many works achieved good performance
in the field of image segmentation[4]. What we used in
our framework to do image segmentation is PSPNet[24].
It extends the pixel-pixel level feature to specially designed
global pooling one. And it also proposes an optimization
strategy with a deeply supervised loss. This work achieves
state-of-the-art performance on various datasets.
3. Proposed Framework
Our goal is to develop an autonomous driving model
trained entirely in a virtual environment which can be ap-
plied in real-world driving scenes with good performance.
One of the major challenges is that the training environment
is generated by a simulator, which means the training environment would be quite different from real-world scenes in
terms of their appearance. To tackle this problem, we proposed to use an image translation process to convert virtual
images to semantic layouts which resembles the semantic
segmentation of its supposed corresponding real world image. This idea is inspired by the work of [23] which tries
to fill the gap between virtual and real on synthesized images. Our framework contains two parts, the image translation process and the reinforcement learning. The image
translation process intends to translate the virtual driving
scene to a semantic representation. We use PSPNet as the
semantic segmentation network in our model. In this part,
in order to get required information such as the first-person
perspective from the TORCS simulator, we use some hacking techniques. The sample process of this translation part
is presented in Figure 1. Finally, we train an autonomous
driving car using reinforcement learning on the semantic
layouts obtained by the translation network. In the reinforcement learning part, we use the asynchronous advantage actor-critic reinforcement learning algorithm. In this
section, we will present the image translation process and
how to apply reinforcement learning to train an autonomous
driving agent.
3.1. Image Translation Process
In order to ensure our autonomous model entirely trained
in a virtual environment has good performance in the real
world, we have to fill the gap between the training environment generated by simulators and the real-world scenes.
From our point of view, the semantic segmentation of realworld visual field contains enough information the agent
needs to take driving actions. Inspired by the work of [23],
we adopt a translation process to translate the virtual image into a semantic segmentation image. This translation
is based on the hypothesis that the semantic segmentation
of real-world visual field is similar to the segmentation of a
virtual image.
The image translation process mainly includes an image
segmentation network.
First, we get the first-person perspective driving scene
from the TORCS simulator. Then we use PSPNet to translate the virtual driving image into semantic segmentation
image. The output of this part is the semantic layout, and
the semantic layout will be fed into the reinforcement learning agent.
There is an obvious obstacle in this part. The appearance of virtual driving images are different from the realworld images, we cannot apply the segmentation tool pretrained on real-world datasets like Cityscapes[7] directly to
the virtual images. And there are no semantic annotations
Figure 3. The convolutional neural network in PSPNet is used to
get the feature map from the last convolutional layer. After that,
in order to harvest different sub-region representation, the PSPNet
applies a pyramid parsing module. Then, it uses upsampling and
concatenation layers to get the final feature representation. Finally,
the convolution layer, which gets the final feature representation as
input, outputs per-pixel prediction.
for TORCS virtual images. We tackle this problem by some
hacking techniques which will be explained in the experiment part.
3.2. Reinforcement Learning for Training Autonomous Driving
39
58
159
480
119
8
29
f(a|s)
3
4
3
4
8
Action Probability
78
Stride of 1
Stride of 2
32
32
Fully Connected Layer
32
640
Stride of 4
3
Figure 4. Reinforcement learning network architecture: an end-toend network mapping state representations to action probability
outputs.
We use Asynchronous Advantage Actor-Critic(A3c) to
train the autonomous-driving vehicle and get the best performance comparing with other reinforcement learning
structures, which is a conceptually simple and lightweight
framework for deep reinforcement learning that uses asynchronous gradient descent for optimization of deep neural
network controllers. A3C utilizes multiple incarnations of
the above in order to learn more efficiently.
In A3C there is a global network and multiple worker
agents which each have their own set of network parameters. Each of these agents interacts with its own copy of the
environment at the same time as the other agents are interacting with their environments.
The experience of each agent is independent of the experience of the others, making speed up and better performance. In this way, the overall experience available for
training becomes more diverse. Critically, the agent uses
the value estimate to update the policy more intelligently
than traditional policy gradient methods.
In [14] there are more details of implement A3C algorithm.
In order to encourage the agent to drive faster and avoid
trained agent on a real-world driving data to evaluate its performance. When we apply the agent to the real-world driving data, we also use the semantic segmentation network
to get the semantic input to the agent. To have a comparison, we also trained another reinforcement learning method
without image translation network in the TORCS simulator. This model is same as our proposed model except the
translation network. We call it Basic RL.
4.1.1
Figure 5. Diagram of A3C high-level architecture.
collisions, we define the reward function as
(t)
(vt · cos α − distcenter ) · β no collision,
rt =
γ
collision,
(1)
where vt is the speed (in m/s) of the agent at time step t, α
is the angle (in rad) between the agent’s speed and the tan(t)
gent line of the track, and distcenter is the distance between
the center of the agent and the middle of the track. β, γ are
constants and are determined at the beginning of training.
We take β = 0.006, γ = −0.025 in our training.
4. Experiments
We perform experiments to compare the performance of
our method and other existing methods of autonomous driving on real-world driving data. This set of experiments aim
to evaluate our model’s performance in the real-world driving scene. And also we perform experiments to compare
our method and basic reinforcement learning without image translation network on TORCS simulator. This set of
experiments aim to show advantages our model bring to reinforcement learning process.
4.1. Autonomous Driving with Image Translation
Network and RL on Real-world Driving Data
In this experiment, we trained our proposed reinforcement learning model with image translation process. We
first trained the semantic segmentation network(PSPNet)
and then apply the trained network to generate semantic
parsing images to feed into our A3C reinforcement learning agent to train a driving policy. And finally, we apply the
Dataset
The real-world driving data is from [6], which is collected
with detailed steering angle autonomous per frame. There
are in total around 45k images in the dataset. To train the
image translation network, we use two datasets separately.
We collect 2k images from the TORCS simulator and use
hacking techniques to get semantic labels for these collected
virtual images. We modified the source code of TORCS
guided by the work to control each category of objects to
show or hide. In detail, we compare the original image with
the image after we hide all the trees, and get the exact pixels covered by trees. And we do the same for other objects.
We process the collected images in this way to get their semantic labels. The other dataset we use is Cityscape[7]. It
contains around 25k real-world images with semantic segmentation annotations.
4.1.2
Image Translation Process
Both scene segmentation parts to translating virtual images
and real-world images in our framework are PSPNet.
We train the network translating virtual images to semantic ones which would be applied in training part with the
dataset we collected from TORCS. We don’t simply use the
ground truth label we hacked from the simulator as the output of this part because we want to avoid the bias between
PSPNet result and hacked ground truth. More specifically,
we want the training part in TORCS and testing part on
real-world data share more similarity through using same
segmentation tool for both parts.
And we train the network translating real-world images
to semantic ones which would be applied in real-world testing part with the Cityscape dataset.
We do experiments to determine whether use the original
RGB semantic result as the output of this part and feed it
into the agent or translate the semantic result into a gray
scale image before fed into the agent. We finally choose to
use the gray scale images as the output of this part. This
will be presented in the result section.
4.1.3
Reinforcement Training
We use such structure in our A3C algorithm: the actor network is a 4-layer convolutional network, with ReLU as ac-
Figure 6. The training process statistics of RGB observation.
Figure 7. The training process statistics of gray scale observation.
tivation function. It’s input is 4 consecutive frames and
there are 9 discrete actions can be its output(“go straight
with acceleration”, “go left with acceleration”, “go right
with acceleration”, “go straight and brake”, “go left and
brake”, “go right and brake”, “go straight”, “go left”, and
“go right”). 12 asynchronous threads with RMSProp optimizer is what we used to train our reinforcement agent,
whose initial learning rate is 0.01, γ = 0.9 and = 0.1.
4.1.4
5. Result
Evaluation on Real-World Dataset
The real world driving dataset [6] provides the steering angle annotations per frame. However, the actions performed
in the TORCS virtual environment only contain ”going
left”, ”going right”, and ”going straight” and their combination with ”brake” and ”acceleration”. Therefore we come
up with a mapping in Table 1 from the steering angle to the
action space of our RL agent. With this mapping, we evaluate the accuracy of action prediction of our model.
Table 1. the steering angles and corresponding actions
angle(degree)
action
[−15, 15]
less than −15
more than 15
going straight
going left
going right
Figure 8. Real images from the driving data and their semantic
segmentations
5.1. Result of Image Translation Process
We use PSPNet to translate virtual images and real-world
images into semantic images. Figure 2 is examples of the
result of translation process on virtual images. We tried two
alternative output of this part - the RGB image or gray scale
image. We do experiments for both, compare the performance and finally choose the gray scale image.
Figure 8 is examples of the translation result of realworld driving images.
Table 2. Accuracy of different models
Accuracy
References
Our Model
Supervised Model
Basic RL
36.6%
52.6%
28.1%
5.2. Result of Reinforcement Learning
Figure 6. shows the training process using RGB semantic image as agent observation. Figure 7. shows the training
process using gray scale semantic image as agent observation.
5.3. Testing on Real-World Driving Data
We extract 4 consecutive frames from the real-world
driving data and parse them into semantic images with PSPNet. And feed these images into our agent to get predictions. Figure 5 shows the prediction our agent made to corresponding input.
We test all the frames in the driving data set and finally
got an accuracy of 36.6%. The comparison with the basic
reinforcement learning and the supervised model trained on
the dataset is shown in Table 2.
And examples of predictions and ground truth are shown
in Table 3.
Table 3. The predictions on real-world driving data and their
ground truth
origin
semantic
prediction
label
straight
45.180000
left
−12.400000
straight
−8.970000
left
−95.700000
6. Conclusion and Future Work
Our proposed autonomous driving model tries to transfer the reinforcement learning agent developed in a virtual
environment to real-world tasks. We use semantic segmentation as the tool to fill the gap between virtual and real. But
result reveals that its performance is limited by the result of
segmentation.If the segmentation techniques develop, our
model would have better capacity. Future work can combine our proposed model with supervised models on realworld data to achieve better results.
[1] P. Abbeel, A. Coates, M. Quigley, , and A. Y. Ng. An application of reinforcement learning to aerobatic helicopter.
flight. Advances in neural information processing systems,
19:1, 2007.
[2] A. Alpher. Frobnication. Journal of Foo, 12(1):234–778,
2002.
[3] A. Alpher and J. P. N. Fotheringham-Smythe. Frobnication
revisited. Journal of Foo, 13(1):234–778, 2003.
[4] V. Badrinarayanan, A. Kendall, , and R. Cipolla. Segnet: A
deep convolutional encoder-decoder architecture for image.
segmentation. arXiv preprint arXiv:1511.00561, 2015.
[5] M. Bojarski, D. D. Testa, D. Dworakowski, B. Firner,
B. Flepp, P. Goyal, L. D. Jackel, M. Monfort, U. Muller,
J. Zhang, X. Zhang, J. Zhao, , and K. Zieba. End to end
learning for self-driving cars. CoRR, abs/1604.07316, 2016.
URL http://arxiv.org/abs/1604.07316.
[6] S. Chen.
Autopilot-tensorflow, 2016.
URL
https://github.com/SullyChen/Autopilot-TensorFlow.
[7] M. Cordts, M. Omran, S. Ramos, T. Rehfeld, M. Enzweiler, R. Benenson, U. Franke, S. Roth, , and
B. Schiele. The cityscapes dataset for semantic urban
scene understanding. CoRR, abs/1604.01685, 2016. URL
http://arxiv.org/abs/1604.01685.
[8] G. Endo, J. Morimoto, T. Matsubara, J. Nakanishi, , and
G. Cheng. Learning cpg-based biped locomotion with a policy gradient method:. Application to a humanoid robot. The
International Journal of Robotics Research, 27(2):213–228,
2008.
[9] A. Gupta, C. Devin, Y. Liu, P. Abbeel, , and S. Levine. Learning invariant feature spaces to transfer skills with. reinforcement learning. arXiv preprint arXiv:1703.02949, 2017.
[10] N. Kohl and P. Stone. Policy gradient reinforcement learning
for fast quadrupedal. locomotion. In Robotics and Automation, 2004. Proceedings. ICRA’04. 2004 IEEE International
Conference on, volume 3, pages 2619–2624. IEEE, 2004.
[11] J. Koutnı́k, G. Cuccu, J. Schmidhuber, , and F. Gomez.
Evolving large-scale neural networks for vision-based reinforcement. learning. In Proceedings of the 15th annual
conference on Genetic and evolutionary computation, pages
1061–1068. ACM, 2013.
[12] T. P. Lillicrap, J. J. Hunt, A. Pritzel, N. Heess, T. Erez,
Y. Tassa, D. Silver, , and D. Wierstra. Continuous control with deep reinforcement learning. arXiv preprint
arXiv:1509.02971, 2015.
[13] J. Long, E. Shelhamer, , and T. Darrell. Fully convolutional
networks for semantic segmentation. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR),
June 2015.
[14] V. Mnih, A. P. Badia, M. Mirza, A. Graves, T. P. Lillicrap, T. Harley, D. Silver, , and K. Kavukcuoglu.
Asynchronous
methods
for
deep
reinforcement
learning.
CoRR, abs/1602.01783, 2016. URL
http://arxiv.org/abs/1602.01783.
[15] V. Mnih, K. Kavukcuoglu, D. Silver, A. A. Rusu, J. Veness,
M. G. Bellemare, A. Graves, M. Riedmiller, A. K. Fidjeland,
[16]
[17]
[18]
[19]
[20]
[21]
[22]
[23]
[24]
G. Ostrovski, and et al. Human-level control through deep
reinforcement learning. Nature, 518(7540):529–533, 2015.
D. A. Pomerleau. Alvinn, an autonomous land vehicle in a
neural network. Technical report, Carnegie Mellon University, Computer Science Department, 1989.
A. A. Rusu, M. Vecerik, T. Rothörl, N. Heess, R. Pascanu,
, and R. Hadsell. Sim-to-real robot learning from pixels
with progressive nets. CoRR, abs/1610.04286, 2016. URL
http://arxiv.org/abs/1610.04286.
J. Schulman, S. Levine, P. Abbeel, M. I. Jordan, , and
P. Moritz. Trust region policy optimization. In ICML, pages
1889–1897, 2015.
S. Shalev-Shwartz, S. Shammah, , and A. Shashua.
Safe, multi-agent, reinforcement learning for autonomous driving. CoRR, abs/1610.03295, 2016. URL
http://arxiv.org/abs/1610.03295.
J. Tobin, R. Fong, A. Ray, J. Schneider, W. Zaremba, ,
and P. Abbeel. Domain randomization for transferring deep
neural networks from. simulation to the real world. arXiv
preprint arXiv:1703.06907, 2017.
E. Tzeng, C. Devin, J. Hoffman, C. Finn, P. Abbeel,
S. Levine, K. Saenko, , and T. Darrell. Adapting deep visuomotor representations with weak pairwise. constraints.
In Workshop on the Algorithmic Foundations of Robotics
(WAFR), 2016.
B. Wymann, E. Espié, C. Guionneau, C. Dimitrakakis,
R. Coulom, , and A. Sumner. Torcs, the open racing car simulator. Software available at http://torcs. sourceforge. net,
2000.
Y. You, X. Pan, Z. Wang, and C. Lu. Virtual to real
reinforcement learning for autonomous driving. CoRR,
abs/1704.03952, 2017.
H. Zhao, J. Shi, X. Qi, X. Wang, and J. Jia. Pyramid scene
parsing network. CoRR, abs/1612.01105, 2016.
| 1cs.CV
|
Accepted as a workshop contribution at ICLR 2015
T RAINING DEEP NEURAL NETWORKS
ON NOISY LABELS WITH BOOTSTRAPPING
arXiv:1412.6596v3 [cs.CV] 15 Apr 2015
Scott E. Reed & Honglak Lee
Dept. of Electrical Engineering and Computer Science, University of Michigan
Ann Arbor, MI, USA
{reedscot,honglak}@umich.edu
Dragomir Anguelov, Christian Szegedy, Dumitru Erhan & Andrew Rabinovich
Google, Inc.
Mountain View, CA, USA
{dragomir,szegedy,dumitru,amrabino}@google.com
A BSTRACT
Current state-of-the-art deep learning systems for visual object recognition and
detection use purely supervised training with regularization such as dropout to
avoid overfitting. The performance depends critically on the amount of labeled
examples, and in current practice the labels are assumed to be unambiguous and
accurate. However, this assumption often does not hold; e.g. in recognition, class
labels may be missing; in detection, objects in the image may not be localized;
and in general, the labeling may be subjective. In this work we propose a generic
way to handle noisy and incomplete labeling by augmenting the prediction objective with a notion of consistency. We consider a prediction consistent if the
same prediction is made given similar percepts, where the notion of similarity is
between deep network features computed from the input data. In experiments we
demonstrate that our approach yields substantial robustness to label noise on several datasets. On MNIST handwritten digits, we show that our model is robust to
label corruption. On the Toronto Face Database, we show that our model handles
well the case of subjective labels in emotion recognition, achieving state-of-theart results, and can also benefit from unlabeled face images with no modification
to our method. On the ILSVRC2014 detection challenge data, we show that our
approach extends to very deep networks, high resolution images and structured
outputs, and results in improved scalable detection.
1
I NTRODUCTION
Currently the predominant systems for visual object recognition and detection (Krizhevsky et al.,
2012; Zeiler & Fergus, 2013; Girshick et al., 2013; Sermanet et al., 2013; Szegedy et al., 2014)
use purely supervised training with regularization such as dropout (Hinton et al., 2012) to avoid
overfitting. These systems do not account for missing labels, subjective labeling or inexhaustivelyannotated images. However, this assumption often does not hold, especially for very large datasets
and in high-resolution images with complex scenes. For example, in recognition, the class labels
may be missing; in detection, the objects in the image may not all be localized; in subjective tasks
such as facial emotion recognition, humans may not even agree on the class label. As training sets
for deep networks become larger (as they should), the problem of missing and noisy labels becomes
more acute, and so we argue it is a fundamental problem for scaling up vision.
In this work we propose a simple approach to hande noisy and incomplete labeling in weaklysupervised deep learning, by augmenting the usual prediction objective with a notion of perceptual
consistency. We consider a prediction consistent if the same prediction is made given similar percepts, where the notion of similarity incorporates features learned by the deep network.
One interpretation of the perceptual consistency objective is that the learner makes use of its representation of the world (implicit in the network parameters) to match incoming percepts to known
1
Accepted as a workshop contribution at ICLR 2015
categories, or in general structured outputs. This provides the learner justification to “disagree”
with a perceptually-inconsistent training label, and effectively re-label the data while training. More
accurate labels may lead to a better model, which allows further label clean-up, and the learner
bootstraps itself in this way. Of course, too much skepticism of the labels carries the risk of ending
up with a delusional agent, so it is important to balance the trade-off between prediction and the
learner’s perceptual consistency.
In our experiments we demonstrate that our approach yields substantial robustness to several types
of label noise on several datasets. On MNIST handwritten digits (LeCun & Cortes, 1998) we show
that our model is robust to label corruption. On the Toronto Face Database (Susskind et al., 2010)
we show that our model handles well the case of subjective labels in emotion recognition, achieving
state-of-the-art results, and can also benefit from unlabeled face images with no modification to our
method. On the ILSVRC2014 detection challenge data (Russakovsky et al., 2014), we show that
our approach improves single-shot person detection using a MultiBox network (Erhan et al., 2014),
and also improves performance in full 200-way detection using MultiBox for region proposal and a
deep CNN for post-classification.
In section 2 we discuss related work, in section 3 we describe our method along with a probabilistic
interpretation and in section 4 we present our results.
2
R ELATED W ORK
The literature on semi-supervised and weakly-supervised learning is vast (see Zhu (2005) for a
survey), and so in this section we focus on the key previous papers that inspired this work and on
other papers on weakly- and semi-supervised deep learning.
The notion of bootstrapping, or “self-training” a learning agent was proposed in (Yarowsky, 1995)
as a way to do word-sense disambiguation with only unlabeled examples and a small list of seed
example sentences with labels. The algorithm proceeds by building an initial classifier using the
seed examples, and then iteratively classifying unlabeled examples, extracting new seed rules for
the classifier using the now expanded training data, and repeating these steps until convergence. The
algorithm was analyzed by Abney (Abney, 2004) and more recently by (Haffari & Sarkar, 2012).
Co-training (Blum & Mitchell, 1998; Nigam et al., 2000; Nigam & Ghani, 2000) was similarlymotivated but used a pair of classifiers with separate views of the data to iteratively learn and generate additional training labels. Whitney & Sarkar (2012) proposed bootstrapping labeled training
examples with graph-based label propagation. Brodley & Friedl (1999) developed statistical methods for identifying mislabeled training data.
Rosenberg et al. (2005) also trained an object detection system in a weakly-supervised manner
using self-training, and demonstrated that their proposed model achieved comparable performance
to models trained with a much larger set of labels. However, that approach works as a wrapper
around an existing detection system, whereas in this work we integrate a consistency objective for
bootstrapping into the training of the deep network itself.
Our work shares a similar motivation to these earlier works, but instead of explicitly generating new
training labels and adding new examples to the training set in an outer loop, we incorporate our
consistency objective directly into the model. In addition, we consider not only the case of learning
from unlabeled examples, but also from noisy labels and inexhaustively-annotated examples.
Mnih & Hinton (2012) developed deep neural networks for improved labeling of aerial images, with
robust loss functions to handle label omission and registration errors. This work shares a similar
motivation of robustness to noisy labels, but rather than formulating loss functions for specific types
of noise, we add a generic consistency objective to the loss to achieve robustness.
Minimum entropy regularization, proposed in (Grandvalet & Bengio, 2005; 2006), performs semisupervised learning by augmenting cross-entropy loss with a term encouraging the classifier to make
predictions with high confidence on the unlabeled examples1 . This is notable because in their approach training on unlabeled examples does not require a generative model, which is beneficial for
training on high-resolution images and other sensory data. We take a similar approach by side1
see eq. 9.7 in (Grandvalet & Bengio, 2006)
2
Accepted as a workshop contribution at ICLR 2015
stepping the difficulty of fully-generative models of high-dimensional sensory data. However, we
extend beyond shallow models to deep networks, and to structured output prediction.
Never ending language learning (NELL) (Carlson et al., 2010) and never ending image learning
(NEIL) (Chen et al., 2013; 2014) are lifelong-learning systems for language and image understanding, respectively. They continuously bootstrap themselves using a cycle of data collection, propagation of labels to the newly collected data, and self-improvement by training on the new data. Our
work is complementary to these efforts, and focuses on building robustness to noisy and missing
labels into the model for weakly-supervised deep learning.
Larochelle & Bengio (2008) developed an RBM for classification that uses a hybrid generative and
discriminative training objective. Deep Boltmann Machines (Salakhutdinov & Hinton, 2009) can
also be trained in a semi-supervised manner with labels connected to the top layer. More recently,
multi-prediction DBM training (Goodfellow et al., 2013) and Generative Stochastic Networks (Bengio & Thibodeau-Laufer, 2013) improved the performance and simplified the training of deep generative models, enabling training via backpropagation much like in standard deep supervised networks.
However, fully-generative unsupervised training on high-dimensional sensory data, e.g. ImageNet
images, is still far behind supervised methods in terms of performance, and so in this work we do
not follow the generative approach directly. Instead, this work focuses on a way to benefit from
unlabeled and weakly-labeled examples with minimal modification to existing deep supervised networks. We demonstrate increased robustness to label noise and performance improvements from
unlabeled data for a minimal engineering effort.
More recently, the problem of deep learning from noisy labels has begun to receive attention. Lee
(2013) also followed the idea of minimum entropy regularization, and proposed generating “pseudolabels” as training targets for unlabeled data, and showed improved performance on MNIST with
few labeled examples. Sukhbaatar & Fergus (2014) developed two deep learning techniques for
handling noisy labels, learning to model the noise distribution in a top-down and bottom-up fashion.
In this work, we push further by extending beyond class labels to structured outputs, and we achieve
state-of-the-art scalable detection performance on ILSVRC2014, despite the fact that our method
does not require explicitly modeling the noise distribution.
3
M ETHOD
In this section we describe two approaches: section 3.1 uses reconstruction error as a consistency
objective and explicitly models the noise distribution as a matrix mapping model predictions to
training labels. A reconstruction loss is added to promote top-down consistency of model predictions
with the observations, which allows the model to discover the pattern of noise in the data.
The method presented in section 3.2 (bootstrapping) uses a convex combination of training labels
and the current model’s predictions to generate the training targets, and thereby avoids directly
modeling the noise distribution. This property is well-suited to the case of structured outputs, for
which modeling dense interactions among all pairs of output units may be neither practical nor
useful. These two approaches are compared empirically in section 4.
In section 3.3 we show how to apply our bootstrapping approach to structured outputs by using the
MultiBox (Erhan et al., 2014) region proposal network to handle the case of inexhaustive structured
output labeling for single-shot person detection and for class-agnostic region proposal.
3.1
C ONSISTENCY IN MULTI - CLASS PREDICTION VIA RECONSTRUCTION
P
Let x ∈ {0, 1}D be the data (or deep features computed from the data) and t ∈ {0, 1}L , k tk = 1
the observed noisy multinomial labels. The standard softmax regresses x onto t without taking into
account noisy or missing labels. In addition to optimizing the conditional log-likelihood log P (t|x),
we add a regularization term encouraging the class prediction to be perceptually consistent.
We first introduce into our model the “true” class
P label (as opposed to the noisy label observations)
as a latent multinomial variable q ∈ {0, 1}L , j qj = 1. Our deep feed-forward network models
3
Accepted as a workshop contribution at ICLR 2015
the posterior over q using the usual softmax regression:
PD
(1)
(1)
exp( i=1 Wij xi + bi )
P (qj = 1|x) = PL
= PL
PD
(1)
(1)
0
j 0 =1 P̃ (qj = 1|x)
j 0 =1 exp(
i=1 Wij 0 xi + bi )
P̃ (qj = 1|x)
(1)
where P̃ denotes the unnormalized probability distribution. Given the true label q, the label noise
can be modeled using another softmax with logits as follows:
log P̃ (tk = 1|q) =
L
X
(2)
(2)
Wkj qj + bk
(2)
j=1
(2)
Roughly, Wkj learns the log-probability of observing true label j as noisy label k. Given only an
observation x, we can marginalize over q to compute the posterior of target t given x.
P (tk = 1|x) =
L
X
P (tk = 1, qj = 1|x) =
j=1
L
X
P (tk = 1|qj = 1)P (qj = 1|x)
(3)
j=1
where the label noise distribution and posterior over true labels are defined above. We can perform discriminative training by gradient ascent on log P (t|x). However, this purely discriminative
training does not yet incorporate perceptual consistency, and there is no explicit incentive for the
model to treat q as the “true” label; it can be viewed as another hidden layer, with the multinomial
constraint resulting in an information bottleneck.
In unpublished work2 , Hinton & Mnih (2009) developed a Restricted Boltzmann Machine
(RBM) (Smolensky, 1986) variant with hidden multinomial output unit q and observed noisy label unit t as described above. The associated energy function can be written as
E(x, t, q) = −
L
D X
X
(1)
Wij xi qj −
i=1 j=1
L
L X
X
(2)
Wkj tk qj −
L
X
(1)
bj q j −
j=1
k=1 j=1
L
X
(2)
bk tk −
k=1
D
X
(3)
bi xi (4)
i=1
Due to the bipartite structure of the RBM, t and x are conditionally independent given q, and so the
energy function in eq. (4) leads to a similar form of the posterior as in eq. (3), marginalizing out the
hidden multinomial unit. The probability distribution arising from (4) is given by
P
q exp(−E(x, t, q))
P (x, t) =
Z
P
where Z =
x,t,q exp(−E(x, t, q)) is the partition function. The model can be trained with
a generative objective, e.g. by approximate gradient ascent on log P (x, t) via contrastive divergence (Hinton, 2002). Generative training naturally provides a notion of consistency between observations x and predictions q because the model learns to draw sample observations via the conditional
PL
(1)
(3)
likelihood P (xi = 1|q) = σ( j=1 Wij qj + bi ), assuming binary observations.
Observed noisy label t
Data reconstruction
True output labels q
True output
labels q
Observed data x
Observed noisy label t
Observed data x
Figure 1: Left: Restricted Boltzmann Machine with hidden multinomial output unit. Right: Analogous feed-forward autoencoder version.
2
Known from personal correspondence.
4
Accepted as a workshop contribution at ICLR 2015
However, fully-generative training is complicated by the fact that the exact likelihood gradient is
intractable due to computing the partition function Z, and in practice MCMC is used. Generative
training is further complicated (though certainly still possible) in cases where the features x are nonbinary. To avoid these complications, and to make our approach rapidly applicable to existing deep
networks using rectified linear activations, and trainable via exact gradient descent, we propose an
analogous autoencoder version.
Figure 1 compares the RBM and autoencoder approaches to the multiclass prediction problem with
perceptual consistency. The overall objective in the feed-forward version is as follows:
Lrecon (x, t) = −
L
X
tk log P (tk = 1|x) + β||x − W (2) q(x)||22
(5)
k=1
where q(x)j = P (qj = 1|x) as in equation 1. The parameter β can be found via cross-validation.
Experimental results using this method are presented in sections 4.1 and 4.2.
3.2
C ONSISTENCY IN MULTI - CLASS PREDICTION VIA BOOTSTRAPPING
In this section we develop a simple consistency objective that does not require an explicit noise
distribution or a reconstruction term. The idea is to dynamically update the targets of the prediction
objective based on the current state of the model. The resulting targets are a convex combination of
(1) the noisy training label, and (2) the current prediction of the model. Intuitively, as the learner
improves over time, its predictions can be trusted more. This mitigates the damage of incorrect
labeling, because incorrect labels are likely to be eventually highly inconsistent with other stimuli
predicted to have the same label by the model.
By paying less heed to inconsistent labels, the learner can develop a more coherent model, which
further improves its ability to evaluate the consistency of noisy labels. We refer to this approach
as “bootstrapping”, in the sense of pulling oneself up by one’s own bootstraps, and also due to
inspiration from the work of Yarowsky (1995) which is also referred to as bootstrapping.
Concretely, we use a cross-entropy objective as before, but generate new regression targets for each
SGD mini-batch based on the current state of the model. We empirically evaluated two types of bootstrapping. “Soft” bootstrapping uses predicted class probabilities q directly to generate regression
targets for each batch as follows:
Lsof t (q, t) =
L
X
[βtk + (1 − β)qk ] log(qk )
(6)
k=1
In fact, it can be shown that the resulting objective is equivalent to softmax regression with minimum
entropy regularization, which was previously studied in (Grandvalet & Bengio, 2006). Intuitively,
minimum entropy regularization encourages the model to have a high confidence in predicting labels
(even for the unlabeled examples, which enables semi-supervised learning).
“Hard” bootstrapping modifies regression targets using the MAP estimate of q given x, which we
denote as zk := 1[k = argmax qi , i = 1...L]:
Lhard (q, t) =
L
X
[βtk + (1 − β)zk ] log(qk )
(7)
k=1
When used with mini-batch stochastic gradient descent, this leads to an EM-like algorithm: In the
E-step, estimate the “true” confidence targets as a convex combination of training labels and model
predictions; in the M-step, update the model parameters to better predict those generated targets.
Both hard and soft bootstrapping can be viewed as instances of a more general approach in which
model-generated regression targets are modulated by a softmax temperature parameter T ; i.e.
PD
(1)
(1)
exp(T · ( i=1 Wij xi + bj ))
P (qj = 1|x) = PL
(8)
PD
(1)
(1)
j 0 =1 exp(T · (
i=1 Wij 0 xi + bj 0 ))
Setting T = 1 recovers soft boostrapping, and T = ∞ recovers hard bootstrapping. We only use
these two operating points in our experiments, but it may be worthwhile to explore other values for
T , and learning T for each dataset.
5
Accepted as a workshop contribution at ICLR 2015
3.3
C ONSISTENCY WITH STRUCTURED OUTPUT PREDICTION
Noisy labels also occur in structured output prediction problems such as object detection. Current
state-of-the-art object detection systems train on images annotated with bounding box labels of the
relevant objects in each image, and the class label for each box. However, it is expensive to exhaustively annotate each image, and for some commonly-appearing categories the data may be prone to
missing annotations. In this section, we modify the training objective of the MultiBox (Erhan et al.,
2014) network for object detection to incorporate a notion of perceptual consistency into the loss.
In the MultiBox approach, ground-truth bounding boxes are clustered and the resulting centroids are
used as “priors” for predicting object location. A deep neural network is trained to predict, for each
groundtruth object in an image, a residual of that groundtruth bounding box to the best-matching
bounding box prior. The network also outputs a logistic confidence score for each prior, indicating
the model’s belief of whether or not an object appears in the corresponding location. Because
MultiBox gives proposals with confidence scores, it enables very efficient runtime-quality tradeoffs
for detection via thresholding the top-scoring proposals within budget. Thus it is an attractive target
for further quality improvements, as we pursue in this section.
Denote the confidence score training targets as t ∈ {0, 1}L and the predicted confidence scores as
c ∈ [0, 1]L . The objective for MultiBox 3 can be written as the following cross-entropy loss:
Lmultibox (c, t) = −
L
X
(tk log(ck ) + (1 − tk ) log(1 − ck ))
(9)
k=1
Note that the sum here is over object locations, not class labels as in the case of sections 3.1 and 3.2.
If there is an object at location k, but tk = 0 due to inexhaustive annotation, the model pays a
large cost for correctly predicting ck = 1. Training naively on the noisy labels leads to perverse
learning situations such as the following: two objects of the same category (potentially within the
same image) appear in the training data, but only one of them is labeled. To reduce the loss, the
confidence prediction layer must learn to distinguish the two objects, which is exactly contrary to
the objective of visual invariance to category-preserving differences.
To incorporate a notion of perceptual consistency into the loss, we follow the same approach as in
the case of multi-class classification: augment the regression targets using the model’s current state.
In the “hard” case, MAP estimates can be obtained by thresholding ck > 1/2.
Lmultibox−hard (c, t) = −
L
X
[βtk + (1 − β)1ck >0.5 ] log(ck )
(10)
k=1
−
L
X
[β(1 − tk ) + (1 − β)(1 − 1ck >0.5 )] log(1 − ck )
k=1
Lmultibox−sof t (c, t) = −
L
X
[βtk + (1 − β)ck ] log(ck )
(11)
k=1
−
L
X
[β(1 − tk ) + (1 − β)(1 − ck )] log(1 − ck )
k=1
With the bootstrap variants of the MultiBox objective, unlabeled positives pose less of a problem
because penalties for large ck are down-scaled by factor β in the first term and (1 − ck ) in the second
term. By mitigating penalties due to missing positives in the data, our approach allows the model to
learn to predict ck with high confidence even if the objects at location k are often unlabeled.
4
E XPERIMENTS
We perform experiments on three image understanding tasks: MNIST handwritten digits recognition, Toroto Faces Database facial emotion recognition, and ILSVRC2014 detection. In all tasks,
3
We omit the bounding box regression term for simplicity, see (Erhan et al., 2014) for full details.
6
Accepted as a workshop contribution at ICLR 2015
we train a deep neural network with our proposed consistency objective. In our figures, “bootstraprecon” refers to training as described in section 3.1, using reconstruction as a consistency objective.
“bootstrap-soft” and “bootstrap-hard” refer to our method described in sections 3.2 and 3.3.
4.1
MNIST WITH NOISY LABELS
In this section we train using our reconstruction-based objective (detailed in section 3.1) on MNIST
handwritten digits with varying degrees of noise in the labels. Specifically, we used a fixed random
permutation of the labels as visualized in figure 2, and we perform control experiments while varying
the probability of applying the label permutation to each training example.
All models were trained with mini-batch SGD, with the same architecture: 784-500-300-10 neural
network with rectified linear units. We used L2 weight decay of 0.0001. We found that β = 0.8
worked best for bootstrap-hard, 0.95 for bootstrap-soft, and 0.005 for bootstrap-recon. We initialize
W (2) to the identity matrix.
For the network trained with our proposed consistency objective, we initialized the network layers
from the baseline prediction-only model. It is also possible to initialize from scratch using our
approach, but we found that with pre-training we could use a larger β and more quickly converge
to a good result. Intuitively, this is similar to the initial collection of “seed” rules in the original
bootstrapping algorithm of (Yarowsky, 1995). During the fine-tuning training phase, all network
weights are updated by backpropagating gradients through the layers.
Figure 2 shows that our bootstrapping method provides a very significant benefit in the case of
permuted labels. The bootstrap-recon method performs the best, and bootstrap-hard nearly as well.
The bootstrap-soft method provides some benefit in the high-noise regime, but only slightly better
than the baseline overall.
True Label
1
2 3 4
5
True Label
6 7 8 9
0
1
2
3
4
5
6
7
8
9
0
Noisy Label
Noisy Label
0
Label noise pattern
1
2 3 4
5
6 7 8 9
0
1
2
3
4
5
6
7
8
9
Learned
Figure 2: Left: Digit recognition accuracy versus percent corrupted labels. Middle: a visualization
of the noise pattern. If there is a white entry in row r, column c, then label c is mapped to to label
r with some probability during training. Right: a visualization of P (tr = 1|qc = 1) learned by our
(2)
bootstrap-recon model (as parameters Wrc ) trained with 40% label noise.
Figure 2 also shows that bootstrap-recon effectively learns the noise distribution P (t|q) via the
(2)
parameters Wkj . Intuitively, the loss from the reconstruction term provides the learner a basis on
which predictions q may disagree with training labels t. Since x must be able to be reconstructed
from q in bootstrap-recon, learning a non-identity W (2) allows q to flexibly vary from t to better
reconstruct x without incurring a penalty from prediction error.
However, it is interesting to note that bootstrap-hard achieves nearly equivalent performance without
explicitly parameterizing the noise distribution. This is useful because reconstruction may be challenging in many cases, such as when x is drawn from a complicated, high-dimensional distribution,
and bootstrap-hard is trivial to implement on top of existing deep supervised networks.
7
Accepted as a workshop contribution at ICLR 2015
4.2
T ORONTO FACES DATABASE EMOTION RECOGNITION
In this section we present results on emotion recognition. The Toronto Faces Database has 112,234
images, 4,178 of which have emotion labels. In all experiments we first extracted spatial-pyramidpooled OMP-1 features as described in (Coates & Ng, 2011) to get 3200-dimensional features. We
then trained a 3200-1000-500-7 network to predict the 1-of-7 emotion labels for each image.
As in the case for our MNIST experiments, we initialize our model from the network pre-trained
with prediction only, and the fine-tuned all layers with our hybrid objective.
Figure 3 summarizes our TFD results. As in the case of MNIST, bootstrap-recon and bootstrap-hard
perform the best, significantly outperforming the softmax baseline, and bootstrap-soft provides a
more modest improvement.
Predicted True Label
r
ge
An
t
us
sg
Di
id
ra
Af
y
pp
Ha
d
Sa
r
Su
Training
baseline
bootstrap-recon
bootstrap-hard
bootstrap-soft
disBM a
CDA+CCA b
d
ise tral
u
Ne
pr
Noisy (Weak) Label
Anger
Disgust
Afraid
Happy
Accuracy (%)
85.3
86.8
86.8
85.6
85.4
85.0
Sad
Table 1: Emotion recognition results
on Toronto Faces Database compared
to state-of-the-art methods.
Surprised
Neutral
a
Figure 3: Predicted-to-noisy emotion label connection W (2) learned by our model.
b
(Reed et al., 2014)
(Rifai et al., 2012)
There is significant off-diagonal weight in W (2) learned by bootstrap-recon on TFD, suggesting
that the model learns to “hedge” its emotion prediction during training by spreading probability
mass from the predicted class to commonly-confused classes, such as “afraid” and “surprised”. The
strongest off diagonals are in the “happy” column, which may be due to the fact that “happy” is
the most common expression, or perhaps that happy expressions have a large visual diversity. Our
method improves the emotion recognition performance, which to our knowledge is state-of-the-art.
The performance improvement from all three bootstrap methods on TFD suggests that our approach
can be useful not just for mistaken labels, but also for semi-supervised learning (missing labels) and
learning from weak labels such as emotion categories.
4.3
ILSVRC 2014 FAST SINGLE - SHOT PERSON DETECTION
In this section we apply our method to detecting persons using a MultiBox network built on top of
the Inception architecture proposed in (Szegedy et al., 2014). We first pre-trained MultiBox on classagnostic localization using the full ILSVRC2014 training set, since there are only several thousand
images labeled with persons in ILSVRC, and then fine-tuned on person images only.
An important point for comparison is the top-K bootstrapping heuristic introduced for MultiBox
training in (Szegedy et al., 2014) for person detection in the presence of missing annotations. In
that approach, the top-K largest confidence predictions are dropped from the loss (which we show
here in eq. (9)), and the setting used was K = 4. In other words, there is no gradient coming from
the top-K most confident location predictions. In fact, it can be viewed as a form of bootstrapping
where only the top-K most confident locations modify their targets, which become the predictions
themselves. In this work, we aim to achieve similar or better performance in a more general way
that can be applied to MultiBox and other discriminative models.
The precision-recall curves in figure 4 show that our proposed bootstrapping improves substantially
over the prediction-only baseline. At the high-precision end of the PR curve, the approaches introduced in this paper perform better, while the top-K heuristic is slightly better at high-recall.
8
Accepted as a workshop contribution at ICLR 2015
ILSVRC Person Detection
1
baseline
top−K heuristic
bootstrap−hard
bootstrap−soft
0.9
0.8
Training
baseline
top-K heuristic
bootstrap-hard
bootstrap-soft
Precision
0.7
0.6
0.5
0.4
AP (%)
30.9
44.1
44.6
43.6
Recall @ 60p
14.3
43.4
43.9
42.8
0.3
Table 2: Our proposed bootstrap-hard and
bootstrap-soft, and the top-K heuristic significantly improve average precision compared to
the prediction-only baseline. Recall@60p is the
recall achievable at 60% precision.
0.2
0.1
0
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
Recall
Figure 4: Precision-recall curves for our methods compared to the baseline prediction-only
network, and the top-K heuristic.
4.4
ILSVRC 2014 200- CATEGORY DETECTION
In this section we apply our method to the case of object detection on the large-scale ImageNet data.
Our proposed method is applied in two ways: first to the MultiBox network for region proposal, and
second to the classifier network that predicts labels for each cropped image region. We follow the
approach in (Szegedy et al., 2014) and combine image crops from MultiBox region proposals with
deep network context features as the input to the classifier for each proposed region.
We trained the MultiBox network as described in 3.3, and the post-classifier network as described
in section 3.2. We found that “hard” performed better than the “soft” form of bootstrapping.
MultiBox
baseline
baseline
bootstrap-hard
bootstrap-hard
-
Postclassifier
baseline
bootstrap-hard
baseline
bootstrap-hard
GoogLeNet single modela
DeepID-Net single modelb
mAP (%)
39.8
40.0
40.3
40.3
38.8
40.1
Recall @ 60p
38.4
38.6
39.3
39.1
-
Using our proposed bootstrapping method, we observe modest improvement on the ILSVRC2014
detection “val2” data4 , mainly attributable to bootstrapping in MultiBox training.
5
C ONCLUSIONS
In this paper we developed novel training methods for weakly-supervised deep learning, and demonstrated the effectiveness of our approach on multi-class prediction and structured output prediction
for several datasets. Our method is exceedingly simple and can be applied with very little engineering effort to existing networks trained using a purely-supervised objective. The improvements
that we show even with very simple methods, suggest that moving beyond purely-supervised deep
learning is worthy of further research attention. In addition to achieving better performance with the
data we already have, our results suggest that performance gains may be achieved from collecting
more data at a cheaper price, since image annotation need not be as exhasutive and mistaken labels
are not as harmful to the performance.
In future work, it may be promising to consider learning a time-dependent policy for tuning β,
the scaling factor between prediction and perceptual consistency objectives, and also to extend our
approach to the case of a situated agent. Another promising direction is to augment large-scale
training for detection (e.g. ILSVRC) with unlabeled and more weakly-labeled images, to further
benefit from our proposed perceptual consistency objective.
4
In a previous version of this draft we included numbers on the full validation set. However, we discovered
that context and post-classifier models used “val1” data, so we re-ran experiments only on the “val2” subset
9
Accepted as a workshop contribution at ICLR 2015
R EFERENCES
Abney, Steven. Understanding the yarowsky algorithm. Computational Linguistics, 30(3):365–395, 2004. 2
Bengio, Yoshua and Thibodeau-Laufer, Eric. Deep generative stochastic networks trainable by backprop. arXiv
preprint arXiv:1306.1091, 2013. 3
Blum, Avrim and Mitchell, Tom. Combining labeled and unlabeled data with co-training. In Proceedings of
the eleventh annual conference on Computational learning theory, pp. 92–100. ACM, 1998. 2
Brodley, Carla E and Friedl, Mark A. Identifying mislabeled training data. Journal of Artificial Intelligence
Research, 11:131–167, 1999. 2
Carlson, Andrew, Betteridge, Justin, Kisiel, Bryan, Settles, Burr, Hruschka Jr, Estevam R, and Mitchell, Tom M.
Toward an architecture for never-ending language learning. In AAAI, volume 5, pp. 3, 2010. 3
Chen, Xinlei, Shrivastava, Abhinav, and Gupta, Abhinav. Neil: Extracting visual knowledge from web data. In
Computer Vision (ICCV), 2013 IEEE International Conference on, pp. 1409–1416. IEEE, 2013. 3
Chen, Xinlei, Shrivastava, Abhinav, and Gupta, Abhinav. Enriching visual knowledge bases via object discovery and segmentation. CVPR, 2014. 3
Coates, Adam and Ng, Andrew Y. The importance of encoding versus training with sparse coding and vector
quantization. In Proceedings of the 28th International Conference on Machine Learning (ICML-11), pp.
921–928, 2011. 8
Erhan, Dumitru, Szegedy, Christian, Toshev, Alexander, and Anguelov, Dragomir. Scalable Object Detection
Using Deep Neural Networks. In CVPR, pp. 2155–2162, 2014. 2, 3, 6
Girshick, Ross, Donahue, Jeff, Darrell, Trevor, and Malik, Jitendra. Rich feature hierarchies for accurate object
detection and semantic segmentation. arXiv preprint arXiv:1311.2524, 2013. 1
Goodfellow, Ian, Mirza, Mehdi, Courville, Aaron, and Bengio, Yoshua. Multi-prediction deep boltzmann
machines. In Advances in Neural Information Processing Systems, pp. 548–556, 2013. 3
Grandvalet, Yves and Bengio, Yoshua. Semi-supervised learning by entropy minimization. In Advances in
Neural Information Processing Systems, pp. 529–536, 2005. 2
Grandvalet, Yves and Bengio, Yoshua. 9 entropy regularization. 2006. 2, 5
Haffari, Gholam Reza and Sarkar, Anoop. Analysis of semi-supervised learning with the yarowsky algorithm.
arXiv preprint arXiv:1206.5240, 2012. 2
Hinton, Geoffrey E. Training products of experts by minimizing contrastive divergence. Neural computation,
14(8):1771–1800, 2002. 4
Hinton, Geoffrey E and Mnih, Volodymyr. Restricted boltzmann machine with hidden multinomial output unit.
Unpublished work, 2009. 4
Hinton, Geoffrey E, Srivastava, Nitish, Krizhevsky, Alex, Sutskever, Ilya, and Salakhutdinov, Ruslan R. Improving neural networks by preventing co-adaptation of feature detectors. arXiv preprint arXiv:1207.0580,
2012. 1
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. 1
Larochelle, Hugo and Bengio, Yoshua. Classification using discriminative restricted boltzmann machines. In
Proceedings of the 25th international conference on Machine learning, pp. 536–543. ACM, 2008. 3
LeCun, Yann and Cortes, Corinna. The mnist database of handwritten digits, 1998. 2
Lee, Dong-Hyun. Pseudo-label: The simple and efficient semi-supervised learning method for deep neural
networks. In Workshop on Challenges in Representation Learning, ICML, 2013. 3
Mnih, Volodymyr and Hinton, Geoffrey E. Learning to label aerial images from noisy data. In Proceedings of
the 29th International Conference on Machine Learning (ICML-12), pp. 567–574, 2012. 2
Nigam, Kamal and Ghani, Rayid. Analyzing the effectiveness and applicability of co-training. In Proceedings
of the ninth international conference on Information and knowledge management, pp. 86–93. ACM, 2000. 2
10
Accepted as a workshop contribution at ICLR 2015
Nigam, Kamal, McCallum, Andrew Kachites, Thrun, Sebastian, and Mitchell, Tom. Text classification from
labeled and unlabeled documents using em. Machine learning, 39(2-3):103–134, 2000. 2
Ouyang, Wanli, Luo, Ping, Zeng, Xingyu, Qiu, Shi, Tian, Yonglong, Li, Hongsheng, Yang, Shuo, Wang, Zhe,
Xiong, Yuanjun, Qian, Chen, et al. Deepid-net: multi-stage and deformable deep convolutional neural
networks for object detection. arXiv preprint arXiv:1409.3505, 2014.
Reed, Scott, Sohn, Kihyuk, Zhang, Yuting, and Lee, Honglak. Learning to disentangle factors of variation with
manifold interaction. In Proceedings of The 31st International Conference on Machine Learning, 2014. 8
Rifai, Salah, Bengio, Yoshua, Courville, Aaron, Vincent, Pascal, and Mirza, Mehdi. Disentangling factors of
variation for facial expression recognition. In Computer Vision–ECCV 2012, pp. 808–822. Springer, 2012.
8
Rosenberg, Chuck, Hebert, Martial, and Schneiderman, Henry. Semi-supervised self-training of object detection models. 2005. 2
Russakovsky, Olga, Deng, Jia, Su, Hao, Krause, Jonathan, Satheesh, Sanjeev, Ma, Sean, Huang, Zhiheng,
Karpathy, Andrej, Khosla, Aditya, Bernstein, Michael, Berg, Alexander C., and Fei-Fei, Li. ImageNet Large
Scale Visual Recognition Challenge, 2014. 2
Salakhutdinov, Ruslan and Hinton, Geoffrey E. Deep boltzmann machines. In International Conference on
Artificial Intelligence and Statistics, pp. 448–455, 2009. 3
Sermanet, Pierre, Eigen, David, Zhang, Xiang, Mathieu, Michaël, Fergus, Rob, and LeCun, Yann. Overfeat: Integrated recognition, localization and detection using convolutional networks. arXiv preprint
arXiv:1312.6229, 2013. 1
Smolensky, Paul. Information processing in dynamical systems: Foundations of harmony theory. 1986. 4
Sukhbaatar, Sainbayar and Fergus, Rob. Learning from noisy labels with deep neural networks. arXiv preprint
arXiv:1406.2080, 2014. 3
Susskind, Josh M, Anderson, Adam K, and Hinton, Geoffrey E. The toronto face database. Department of
Computer Science, University of Toronto, Toronto, ON, Canada, Tech. Rep, 2010. 2
Szegedy, C., Reed, S., Erhan, D., and Anguelov, D. Scalable, High-Quality Object Detection. ArXiv e-prints,
December 2014. 1, 8, 9
Szegedy, Christian, Liu, Wei, Jia, Yangqing, Sermanet, Pierre, Reed, Scott, Anguelov, Dragomir, Erhan, Dumitru, Vanhoucke, Vincent, and Rabinovich, Andrew. Going deeper with convolutions. arXiv preprint
arXiv:1409.4842, 2014. 8
Whitney, Max and Sarkar, Anoop. Bootstrapping via graph propagation. In Proceedings of the 50th Annual
Meeting of the Association for Computational Linguistics: Long Papers-Volume 1, pp. 620–628. Association
for Computational Linguistics, 2012. 2
Yarowsky, David. Unsupervised word sense disambiguation rivaling supervised methods. In Proceedings
of the 33rd annual meeting on Association for Computational Linguistics, pp. 189–196. Association for
Computational Linguistics, 1995. 2, 5, 7
Zeiler, Matthew D and Fergus, Rob. Visualizing and understanding convolutional neural networks. arXiv
preprint arXiv:1311.2901, 2013. 1
Zhu, Xiaojin. Semi-supervised learning literature survey. 2005. 2
11
| 9cs.NE
|
arXiv:1604.08267v1 [math.GT] 27 Apr 2016
RANK GRADIENTS OF INFINITE CYCLIC COVERS OF KÄHLER
MANIFOLDS
STEFAN FRIEDL AND STEFANO VIDUSSI
Abstract. Given a Kähler group G and a primitive class φ ∈ H 1 (G; Z), we show
that the rank gradient of (G, φ) is zero if and only if Ker φ ≤ G finitely generated.
Using this approach, we give a quick proof of the fact (originally due to Napier
and Ramachandran) that Kähler groups are not properly ascending or descending HNN extensions. Further investigation of the properties of Bieri–Neumann–
Strebel invariants of Kähler groups allows us to show that a large class of groups
of orientation–preserving PL homeomorphisms of an interval (customarily denoted
1
F (ℓ, Z[ n1 ···n
], hn1 , . . . , nk i)), which generalize Thompson’s group F , are not Kähler.
k
1. Introduction
In [Lac05 ] M. Lackenby introduced the notion of rank gradient rg (G, {Gi }) associated to a lattice {Gi } of finite index subgroups Gi ≤f G (i.e. a collection of subgroups
which is closed with respect to intersection). This rank gradient is defined as
(1)
rg (G, {Gi }) = lim inf
i→∞
rk(Gi )
,
[G : Gi ]
where given a group H we denote by rk(H) the rank of H, i.e. the minimal number
of generators. Lackenby computed the rank gradient for various classes of groups and
he illustrated its rôle in understanding properties of the lattice and, in turn, of the
group G itself. Successive work of many authors has related this notion with various
invariants of the group G.
The explicit calculation of the rank gradient appears rarely straightforward, and
usually requires some relatively sophisticated information on the group.
Here we will focus on the rank gradient defined by lattices associated to the infinite
cyclic quotients of G, namely collections of the type
φ
{Gi = Ker(G −
→ Z → Zi )}
determined by a primitive class φ ∈ H 1 (G; Z) = Hom(G, Z); this rank gradient will
be denoted, for simplicity, rg (G, φ).
Date: November 13, 2017.
2010 Mathematics Subject Classification. 53C55, 20F65.
Key words and phrases. Kähler groups, rank gradient, Thompson–type groups.
1
2
STEFAN FRIEDL AND STEFANO VIDUSSI
One example where this calculation can be carried out in some detail is that of the
fundamental group π = π1 (N) of a 3–manifold N with b1 (N) ≥ 1: in [DFV14] the
authors proved that rg (π, φ) vanishes if and only if φ is represented by a locally–trivial
fibration of N over S 1 with surface fibers. More precisely rg (π, φ) = 0 if and only if
Ker φ ≤ π is finitely generated, which is equivalent to what is stated above by virtue
of Stallings’ Fibration Theorem (see [Sta62]). While one direction of the proof is
straightforward, the other originates from some rather deep properties of 3–manifold
groups, which can be interpreted as a consequence of the separability results of Agol
[Ag13] and Wise [Wi12].
In this paper, we will show that a similar result holds for fundamental groups of
Kähler manifolds, or Kähler groups. (The standard reference for the study of this
class of groups is [ABCKT96]; more recent results are surveyed in [Bu11].) Much like
in the case of 3–manifold groups, the non–straightforward direction of the proof of
our main theorem will come from suitably exploiting nontrivial properties of Kähler
groups. Interestingly enough, the key tool here will also be separability, this time for
the fundamental groups of surfaces.
For sake of notation, we define a Kähler pair to be the pair (G, φ) where G is
a Kähler group and φ ∈ H 1 (G; Z) = Hom(G, Z) is a primitive (i.e. surjective)
cohomology class. Our main theorem is the following:
Theorem A. Let (G, φ) be a Kähler pair. Then rg (G, φ) = 0 if and only if Ker φ ≤ G
is finitely generated.
Note that for Kähler groups there is no analog (to the best of our understanding)
to Stallings’ Fibration Theorem, for a class with rg (G, φ) = 0; to start, there is no
reason to expect that the kernel Ker φ is finitely presented, as Kähler groups, in
general, are not coherent. Instead, in this case it is the classes φ ∈ Hom(G, Z) with
rg (G, φ) > 0 that admit a geometric interpretation; in fact, as we discuss in Section
2 (equation 2 and following remarks) these are the classes that (virtually) factorize
through the epimorphism induced, in homotopy, by a genus g ≥ 2 pencil on a Kähler
manifold X such that G = π1 (X). In fact, as we will see in the proof of Theorem A,
the existence of that pencil (at homotopy level) can be thought of as the very reason
why for these classes we have rg (G, φ) > 0.
As a corollary of this theorem, we will give a quick proof of a result of [NR08],
Theorem 0.3(ii):
Corollary B. Let G be a Kähler group; then G cannot be a properly ascending or
descending HNN extension of a finitely generated subgroup.
The latter result was originally proven by Napier and Ramachandran in order to
show that Thompson’s group F (and some of its generalizations) are not Kähler,
answering a question originally posed by Ross Geoghegan.
Several classes of generalizations of the group F have been studied. We will be interested here in a rather broad collection of finitely presented subgroups, usually denoted
RANK GRADIENTS OF INFINITE CYCLIC COVERS OF KÄHLER MANIFOLDS
3
1
as F (ℓ, Z[ n1 ···n
], hn1 , . . . , nk i), of the group of orientation–preserving PL homeomork
phisms of the interval [0, ℓ] that was first investigated by Brown in [Bro87a] (and in
unpublished work), and further discussed in [Ste92]; see Section 3 for the definition.
This collection of groups (which for notational simplicity will be referred to here as
generalized Thompson groups) include Thompson’s group F = F (1, Z[ 12 ], h2i). However, of relevance to us, for “most” of them Corollary B yields no information to
decide if they are Kähler. Using Delzant’s work [De10] on Bieri–Neumann–Strebel
invariants ([BNS87]) of Kähler groups, we prove the following:
1
], hn1 , . . . , nk i) be a generalized Thompson group;
Theorem C. Let G = F (ℓ, Z[ n1 ···n
k
then G is not Kähler.
The structure of the paper is the following. Section 2 is devoted to the proof of
Theorem A and Corollary B. In Section 3 we prove that the generalized Thompson
groups are not Kähler (Theorem C) and we also discuss the fact that for “most”
of these groups Corollary B (even applied to finite index subgroups) is inconclusive.
Finally in Section 4 we discuss some examples to illustrate the intricacy of the relation
between the rank gradient and the BNS invariant of groups in general. The Appendix
is devoted to the proof of some basic results on generalized Thompson groups used
in Section 3.
Acknowledgments. The authors would like to thank the referee for carefully reading the manuscript and providing us with useful comments. The first author was
supported by the SFB 1085 ‘higher invariants’ funded by the Deutsche Forschungsgemeinschaft DFG.
2. Rank Gradient of Kähler Groups
We start with two lemmata. The first is straightforward.
Lemma 2.1. Consider the short exact sequence of finitely generated groups
p
1→K→G−
→ Γ → 1.
Let ρ ∈ H 1 (Γ; Z) be a primitive class; then p∗ ρ ∈ H 1 (G; Z) is primitive and rg (G, p∗ ρ) =
rg (Γ, ρ).
Proof. The fact that p∗ ρ = ρ◦p is primitive is obvious from the fact that p is surjective,
hence so is p∗ ρ thought of as an element of Hom(G, Z). The rank gradients in question
are associated respectively to the lattices
p∗ ρ
ρ
{Gi = Ker(G −−→ Z → Zi )} and {Γi = Ker(Γ −
→ Z → Zi )}.
These groups are related, for each i ≥ 1, by a short exact sequence
p|G
i
1 → K → Gi −−→
Γi → 1
4
STEFAN FRIEDL AND STEFANO VIDUSSI
hence the inequalities
rk(Γi ) ≤ rk(Gi ) ≤ rk(Γi ) + rk(K).
As the second summand in the rightmost term is independent of i, a check of the
definition in (1) yields the desired equality.
The proof of the second lemma is more elaborate. Remember that a finitely presented group Γ is LERF if, given any finitely generated subgroup A ≤ Γ, and any
element γ ∈ Γ \ A, there exists an epimorphism α : Γ → Q to a finite group such that
α(γ) ∈
/ α(A).
Lemma 2.2. Let Γ be a finitely presented LERF group, and let ρ ∈ Hom(Γ, Z) =
H 1 (Γ; Z) be a primitive class. Then the following are equivalent:
(1) Ker ρ ≤ Γ is infinitely generated;
e ≤f Γ and an epimorphism qe: Γ
e→
(2) there exists a normal finite index subgroup Γ
e → Z factorizes through qe,
Fn to a free nonabelian group such that ρ|Γe : Γ
namely we have
1
/
e
Γ
❃
/
❃❃
❃❃
❃
qe ❃❃
/Q
❄❄
❄❄ρ
❄❄
❄
/ Z.
Γ❄
Fn
/
1
ψ
with Q a finite group.
Proof. Let Γ be a finitely presented group and ρ ∈ H 1 (Γ; Z) a primitive class. We
can write Γ as an HNN extension of a finitely generated base group B ≤ Ker ρ with
finitely generated associated subgroups A± ≤ B, namely we have
Γ = hB, t|t−1 A+ t = A− i,
with ρ(t) = 1 ∈ Z. Moreover, Ker ρ is finitely generated if and only if A± = B = Ker ρ
(see [BS78]).
Assume that Ker ρ is infinitely generated; then one (or both) of A± is a proper
subgroup of B, say A+ B. Since Γ is LERF, for any element b ∈ B \ A+ there exists
an epimorphism to a finite group α : Γ → S such that α(b) ∈
/ α(A+ ). In particular,
α(A+ ) is strictly contained in α(B). We can now define an epimorphism from Γ = B∗A
to the HNN extension α(B)∗α(A) of the finite base group α(B) with finite associated
(isomorphic) subgroups α(A± ) α(B). (We use here standard shorthand notation
for HNN extensions in terms of base and abstract associated subgroups.) Precisely,
we have a map
q : hB, t|t−1 A+ t = A− i −→ hα(B), t|t−1α(A+ )t = α(A− )i.
Clearly the map ρ : Γ → Z descends to α(B)∗α(A) . By [Se80, Prop 11 p. 120 and
Exercise 3 p. 123] this group admits a finite index free nonabelian normal subgroup,
RANK GRADIENTS OF INFINITE CYCLIC COVERS OF KÄHLER MANIFOLDS
5
and we get the commutative diagram
1
/
q̃
e
Γ
1
/
Fn
/Q
✺✺
✺
∼
q ✺✺
=
✺✺
✺✺ ρ
❚❚❚❚ / α(B)∗α(A) ✺✺ / Q
❚❚❚❚
■■ ✺✺
■
❚❚❚❚
✺
❚❚❚❚ ■■■■ ✺✺
❚
■
❚❚❚❚ ■■✺
ψ
❚❚$*
/
Γ✺
1
/
/
1
Z
which yields the virtual factorization of ρ ∈ Hom(Γ, Z).
To prove the other direction, assume that ρ admits a virtual factorization as in (2).
To show that (1) holds, it is enough to prove that Ker ρ|Γe is an infinitely generated
e as the latter is finite index in Γ. But the group Ker ρ|e surjects
subgroup of Γ,
Γ
to the kernel of a nontrivial homomorphism Fn → Z. This kernel must be infinitely
generated, as finitely generated normal subgroups of a free nonabelian group are finite
index (see e.g. [Ca03, Lemma 3.3]).
We are now in position to prove our main theorem, that we restate for sake of
convenience.
Theorem 2.3. Let (G, φ) be a Kähler pair. Then rg (G, φ) = 0 if and only if Ker φ ≤
G is finitely generated.
Proof. The proof of the “if” direction is straightforward: if φ : G → Z is an epimorphism with finitely generated kernel Ker φ, and we denote by {g1 , . . . , gk } a generating
set for Ker φ, then the group G is the semidirect product Ker φ ⋊ hti and admits the
generating set {g1 , . . . , gk , t}, so that the rank of G is bounded above by k + 1. It is
well–known that
φ
Gi := Ker(G −
→ Z → Z/i)
equals Ker φ ⋊ hti i, hence rk(Gi ) is uniformly bounded above by k + 1. This immediately entails the vanishing of rg (G, φ).
For the “only if” direction, we proceed as follows. Let φ : G → Z be an epimorphism
with infinitely generated kernel. By [NR01, Theorem 4.3], such an epimorphism
factorizes through an epimorphism p : G → Γ with finitely generated kernel K where
Γ is a cocompact Fuchsian group of positive first Betti number, which arises as the
fundamental group of a hyperbolic Riemann orbisurface Σorb of positive genus:
(2)
1
/
K
/
p
G ❖❖❖ / Γ ❄
❄
❖
/
❖❖❖ ❄❄ ρ
❖❖❖ ❄❄
❖❖❖❄❄
φ
❖'
1
Z
Although not strictly necessary for the proof, we can integrate this picture and interpret (following [Ca03, NR06]) the maps of the diagram in terms of induced maps
6
STEFAN FRIEDL AND STEFANO VIDUSSI
of an irrational pencil f : X → Σ, where X is a Kähler manifold with G = π1 (X)
and Σ a Riemann surface of positive genus. In the presence of multiple fibers, to
obtain a surjection of fundamental groups with finitely generated kernel, we promote
the fundamental group of Σ to the fundamental group Γ of the orbisurface Σorb with
orbifold points determined by the multiple fibers.
The proof now boils down to the use of the lemmata above and some basic results
from [Lac05 ]. By Lemma 2.1, rg (G, φ) = rg (G, p∗ρ) = rg (Γ, ρ). By definition, the
ρ
latter is the rank gradient rg (Γ, {Γi }) of the lattice {Γi = Ker(Γ −
→ Z → Zi )}. Rank
e ≤f Γ be a finite
gradients behave well under passing to finite index subgroup: let Γ
index subgroup; [Lac05 , Lemma 3.1] asserts that rg (Γ, {Γi }) is nonzero if and only if
e {Γ
e ∩ Γi }) is nonzero. We can now apply Lemma 2.2 to Γ, as the latter is LERF
rg (Γ,
by [Sc78] and for any ρ ∈ H 1 (Γ; Z), Ker ρ is infinitely generated (see e.g. [Ca03,
Lemma 3.4]). Keeping the notation from Lemma 2.2 we have
e
Γ
❄
/
❄❄
❄❄
❄
qe ❄❄
Γ❄
Fn
❄❄
❄❄ρ
❄❄
❄
/Z
ψ
ψ
e → Fn restricts to an epimorphism from Γ
e ∩ Γi to Ker(Fn −
so qe: Γ
→ Z → Zi )}. It
follows that we have the inequality
(3)
ψ
e ∩ Γi ) ≥ rk(Ker(Fn −
rk(Γ
→ Z → Zi )).
We endeavor now to compute the term on the right hand side. In general, the map
ψ : Fn → Z identified in Lemma 2.2 may fail to be surjective, when m is not primitive
and has some divisibility m ∈ Z. In that case the image of ψ is a subgroup mZ ⊂ Z.
ψ
.
A simple exercise shows that the index of Ker(Fn −
→ Z → Zi ) in Fn equals lcm(m,i)
m
The Reidemeister–Schreier formula gives then the rank of this free group:
(4)
ψ
rk(Ker(Fn −
→ Z → Zi )) =
lcm(m, i)
(n − 1) + 1.
m
Combining equations (3) and (4) in the calculation of the rank gradient we get
e ∩ Γi )
e ∩ Γi )
rk(Γ
rk(Γ
1 lcm(m, i)
1
1
≥
≥
(n − 1) + 1 ≥ (n − 1) +
e:Γ
e ∩ Γi ]
[Γ : Γi ]
i
m
m
i
[Γ
from which it follows that
e {Γ
e ∩ Γi }) ≥ 1 (n − 1) > 0.
rg (Γ,
m
As observed above, this entails rg (G, φ) > 0.
RANK GRADIENTS OF INFINITE CYCLIC COVERS OF KÄHLER MANIFOLDS
7
Remark. While there is no reason to expect that Kähler groups are LERF (or even
satisfy weaker forms of separability, that would be still sufficient for Lemma 2.2
to hold) we see that the presence of an irrational pencil, in the form of (2) yields
essentially the same consequences, using the fact that cocompact Fuchsian groups are
LERF. In that sense, the proof above and the proof for 3–manifold groups in [DFV14]
originate from similar ingredients.
We can now give a proof of Corollary B. Recall that a finitely presented G is an
ascending (respectively descending) HNN extension if it can be written as
G = hB, t|t−1 A+ t = A− i
where B is finitely generated and where A+ = B (respectively A− = B).
Proof. Let G be a Kähler group that can be written as an ascending or descending
HNN extension with stable letter t, and denote by φ : G → Z the homomorphism
sending t to the generator of Z and the base of the extension (a subgroup B ≤ G
with generating set {g1, . . . , gk }) to 0. We claim that rg (G, φ) vanishes. The proof is
almost verbatim the proof of the “if” direction of Theorem A, once we observe (using
φ
Schreier’s rewriting process) the well–known fact that Gi := Ker(G −
→ Z → Z/i) is
i
generated by {g1 , . . . , gk , t }, i.e. the rank of Gi is uniformly bounded. By Theorem
A it follows that Ker φ is finitely generated. But as already mentioned in the proof of
Lemma 2.2 this can occur if and only if the HNN extension is simultaneously ascending
and descending, i.e. both A± = B.
1
3. The Groups F (ℓ, Z[ n1 ···n
], hn1 , . . . , nk i) are not Kähler
k
As observed in [NR06, NR08] Corollary B implies that Thompson’s group F (and
some of its generalizations) is not Kähler, as that group is a properly ascending HNN
extension. There are several families of generalizations of Thompson’s group, that
arise as finitely presented subgroups of the group of orientation–preserving PL homeomorphisms of an interval, for which we can expect as well that they fail to be Kähler.
For “most” of them, however, Corollary B is not conclusive: there are examples for
which every class φ in the first integral cohomology group has the property that Ker φ
is finitely generated (see below).
We will be able to decide the problem above by determining, in more generality,
some constraints that groups with vanishing rank gradient function rg (G, φ) need to
satisfy in order to be Kähler.
In order to determine these constraints we will make use of the Bieri–Neumann–
Strebel invariant of G (for which we refer the reader to [BNS87]), together with the
generalization of [NR01, Theorem 4.3] due to Delzant.
Proposition 3.1. Let G be a Kähler group whose rank gradient function rg (G, φ)
vanishes identically for all φ ∈ H 1 (G; Z). Then the commutator subgroup [G, G] ≤ G
of G is finitely generated.
8
STEFAN FRIEDL AND STEFANO VIDUSSI
Proof. Let G be as in the statement and X a Kähler manifold with G = π1 (X).
Assume, by contradiction, that the commutator subgroup is not finitely generated.
By [BNS87, Theorem B1] there exists an exceptional character χ : G → R. But then
there exists an epimorphism p : G → Γ with finitely generated kernel (see [Ca03]) to
a cocompact Fuchsian group, induced by the irrational pencil f : X → Σ of [De10,
Théorème 1.1]. The induced map p∗ : H 1 (Γ; Z) → H 1 (G Z) is then injective, and
for all φ ∈ im(p∗ ) ⊂ H 1 (G; Z) we have rg (G, φ) > 0 by Lemma 2.1, which yields a
contradiction to our assumption on G.
There is a case where we can use a simple criterion to certify that a group has rank
gradient that vanishes identically for all φ ∈ H 1 (G; Z), and this will be sufficient to
apply Proposition 3.1 to Thompson’s group and its generalization.
Lemma 3.2. Let G be a finitely presented group which does not contain any F2
subgroup. Then the rank gradient function rg (G, φ) vanishes identically for all φ ∈
H 1 (G; Z).
Proof. Given any primitive class φ ∈ H 1 (G; Z), by [BS78] there exists a finitely
generated subgroup B ≤ Ker φ ≤ G and finitely generated associated subgroups
A± ≤ B so that G = hB, t|t−1 A+ t = A− i, so that φ(t) = 1. It is well–known (see e.g.
[BM07] for a detailed proof) that unless this extension is ascending or descending,
then G contains an F2 subgroup. If G does not contain any F2 subgroup, repeating
the argument of the proof of Corollary B for all primitive classes φ ∈ H 1(G; Z), we
deduce the vanishing of rg(G, φ).
Remark. In light of the results above, we can give a group-theoretic relic of the result
of Delzant ([De10, Théorème 1.1]): either a Kähler group G has finitely generated
commutator subgroup, or it can be written as a non–ascending and non–descending
HNN extension. This alternative is decided by the vanishing of the rank gradient
function. A straightforward (but not completely trivial) exercise in 3–manifold group
theory shows that this alternative is the same that occurs for a 3–manifold group π =
π1 (N). (The nontriviality comes from the need to use results on the Thurston norm
of N, that has no complete analog in the case of other groups.) For 3–manifolds with
b1 (π) ≥ 2, furthermore, the first case corresponds to N a nilmanifold or Euclidean.
For a general group, it is quite possible that the assumption that the rank gradient
function rg (G, φ) vanishes identically does not yield particularly strong consequences
(see e.g. the examples discussed in Section 4).
We are now in a position to show that a large class of generalizations of Thompson’s group are not Kähler. Omitting further generality, we will limit ourselves to
1
the groups of type F (ℓ, Z[ n1 ···n
], hn1 , . . . , nk i) discussed by Brown in [Bro87a] and
k
further elucidated by Stein in [Ste92]. These groups are finitely presented, and can
be described as follows: Given a nontrivial multiplicative subgroup
hn1 , . . . , nk i = {nj11 · · · njkk |ji ∈ Z} ≤ R+
RANK GRADIENTS OF INFINITE CYCLIC COVERS OF KÄHLER MANIFOLDS
9
(where we assume that {n1 , . . . , nk } is a positive integer basis of the multiplicative
group), this groups is defined as the group of orientation–preserving PL homeomor1
phism of the interval [0, ℓ], where 0 < ℓ ∈ Z[ n1 ···n
], with finitely many singularities
k
1
all in Z[ n1 ···nk ] and slopes in hn1 , . . . , nk i. In this notation, Thompson’s group F is
F (1, Z[ 21 ], h2i). The proof that these groups are not Kähler is a consequence of the
knowledge of the BNS invariants for the group in question, which follows from the
results of [BNS87, Section 8]:
1
Theorem 3.3. Let G = F (ℓ, Z[ n1 ···n
], hn1 , . . . , nk i); then G is not Kähler.
k
1
Proof. By [BrSq85, Theorem 3.1] the group G = F (ℓ, Z[ n1 ···n
], hn1 , . . . , nk i) does not
k
contain F2 subgroups, hence its rank gradient function rg (G, φ) identically vanishes.
On the other hand, as we show in detail in the Appendix, [BNS87, Theorem 8.1]
implies that it has exceptional characters, hence its commutator subgroup cannot be
finitely generated. By Proposition 3.1 G is not Kähler.
As observed in [BNS87, Section 8] “most” of the groups above have the property
that for all primitive classes φ ∈ H 1 (G; Z), Ker φ is finitely generated: an explicit
example is the group G = F (1, Z[ 16 ], h2, 3i) of homeomorphisms of the unit interval
with slopes in {2j 3k |j, k ∈ Z} and singularities in Z[1/6]; here H 1 (G; Z) = Z4 , as
computed in [Ste92]. It follows from this observation that for these groups Corollary
B is inconclusive. A less obvious fact is that we cannot even apply the “covering trick”
to use Corollary B, namely fish for a finite index subgroup H ≤f G that is a properly
ascending or descending HNN extension, hence fails to be Kähler by Corollary B:
1
Lemma 3.4. Let G = F (ℓ, Z[ n1 ···n
], hn1 , . . . , nk i), and assume that G cannot be
k
written as a properly ascending or descending HNN extension; then neither can any
of the finite index subgroups of G.
Proof. Let H ≤f G be a finite index subgroup, and denote by f : H → G the inclusion
map. We start by showing that f ∗ : H 1 (G; R) → H 1 (H; R) is an isomorphism. By
the existence of the transfer map, injectivity is clear, so we just need to show that
b1 (G) = b1 (H). As it suffices to prove this result for the normal core of H, we will
assume without loss of generality that H Ef G. Then [H, H] is a normal subgroup of
[G, G]. And as the latter is simple (see e.g. [Ste92, Theorem 5.1], based also on work
of Bieri and Strebel) we must have [H, H] = [G, G]. The map
H1 (H; Z) = H/[H, H] = H/[G, G] → G/[G, G] = H1 (G; Z)
is therefore injective, and the result follows. At this point an endless amount of
patience would allow a proof of the statement directly in terms of generators and
relations, but we can proceed in a simpler way using BNS invariants. In fact, if we
assume by contradiction that there exists a primitive class χ : H → Z that corresponds
to a properly ascending extension, then the (projective class) of χ must satisfy [χ] ∈
10
STEFAN FRIEDL AND STEFANO VIDUSSI
Σ(H) and [χ] ∈
/ −Σ(H), where
Σ(H) ⊂ S(H) := (H 1 (H; R) \ 0)/R∗
is the BNS invariant of H. Surjectivity of f ∗ : H 1 (G; R) → H 1 (H; R) entails that
[χ] ∈ S(H) is the image of some class [φ] ∈ S(G), where we can assume that φ
is a primitive class in H 1 (G; Z). But [BNS87, Theorem H] guarantees now that
f ∗ (S(G) \ Σ(G)) = S(H) \ Σ(H) (we are using here the fact that [H, H] = [G, G]). It
follows that [φ] ∈ Σ(G) and [φ] ∈
/ −Σ(G). These two conditions characterize properly
ascending HNN extensions (see e.g. [BGK10, Theorem 2.1]), hence G can be written
as an ascending HNN extension. (The proof in the descending case is verbatim.) This
contradicts the assumption.
4. Rank Gradient and BNS Invariant
While the rank gradient of infinite cyclic covers of 3–manifold groups and, now,
Kähler manifolds follow the dichotomy of Theorem A, this property should be counted
more of as a property of these very peculiar classes of groups than a common phenomenon. Stated otherwise, the vanishing of the rank gradient rg (G, φ) is a necessary
but, as we discuss below, not a sufficient condition to decide that φ belongs to the
BNS invariant of G.
The first example of this fact comes from Thompson’s group F . As discussed above
(cf. Lemma 3.2) this group has vanishing rank gradient for all classes φ ∈ H 1 (F ; Z),
but two pairs of classes {±φi , i = 1, 2} (corresponding to properly ascending HNN
extensions) fail to satisfy the statement of Theorem A. In this case, {φi , i = 1, 2}
belong to the BNS invariant of F , but {−φi , i = 1, 2} do not.
A second, more compelling example is the following, that shows that the discrepancy between the information contained in the rank gradient and in the BNS invariants is not limited to properly ascending or descending extensions. Consider the
group
(5)
π = ht, a, b | t−1at = a2 , t−1 b2 t = bi.
Here H1 (π; Z) = Z. We claim that π cannot be written as a properly ascending or
descending HNN extension. To prove this we need the following lemma. Here we
denote by ∆π,φ ∈ Z[t±1 ] the Alexander polynomial corresponding to a pair (π, φ), i.e.
the order of the Alexander Z[t±1 ]–module
H1 (π; Z[t±1 ]) = H1 (Ker(φ)).
Lemma 4.1. Let π be a group together with an epimorphism φ : π → Z. If the pair
(π, φ) corresponds to an ascending (resp. descending) HNN extension of a finitely
generated group, then the top (resp. bottom) coefficient of ∆π,φ equals ±1.
Proof. Let π be a group together with an epimorphism φ : π → Z = hti. We denote
by φ also the corresponding ring homomorphism Z[π] → Z[Z] = Z[t±1 ]. Assume π
can be written as an ascending HNN extension of a finitely generated group B. In
RANK GRADIENTS OF INFINITE CYCLIC COVERS OF KÄHLER MANIFOLDS
11
this case it is convenient to write π = ht, B | t−1 Bt = ϕ(B)i where ϕ : B → B is a
monomorphism, B = C and D = ϕ(B) ≤ B. Now let
B = hg1 , . . . , gk | r1 , r2 , . . .i
be a presentation for B with finitely many generators. We denote by n ∈ N ∪ {∞}
the number of relations. We can write
π = ht, g1 , . . . , gk | t−1 g1 t = ϕ(g1 ), . . . , t−1 gk t = ϕ(gk ), r1 , r2 , . . .i.
We continue with the following claim.
Claim. The matrix
φ
∂t−1 gi t
∂gj
i,j=1,...,k
∂ri
φ
∂gj
i=1,...,j=1,...,k
!
is a presentation matrix for H1 (π; Z[t±1 ]).
We denote by X the usual CW-complex corresponding to the presentation for π
and we denote by Y the CW-complex corresponding to the presentation hti. By Fox
calculus the chain complex C∗ (X; Z[t±1 ]) is given as follows.
∂t−1 gi t
∂ri
φ ∂g
φ
j
∂gj
∂t−1 gi t
∂ri
φ ∂t
φ
1−x1 ... 1−xk 1−t
∂t
Z[t±1 ]k ⊕Z[t±1 ]n −−−−−−−−−−−−−−−−−→ Z[t±1 ]k ⊕Z[t±1 ] −−−−−−−−−−−−→ Z[t±1 ] → 0.
Here, for space reasons we left out the range of the indices i and j. The chain complex
1−t
C∗ (Y ; Z[t±1 ]) = 0 → Z[t±1 ] −−→ Z[t±1 ] → 0
is an obvious subcomplex of C∗ (X; Z[t±1 ]). It is now straightforward to see that the
matrix in the claim is a presentation matrix for the relative Alexander module
±1
±1
±1
H1 (π, hti; Z[t ]) = H1 C∗ (X; Z[t ])/C∗ (Y ; Z[t ]) .
But considering the long exact sequence in homology with Z[t±1 ]-coefficients of the
pair (π, hti) shows that H1 (π, hti; Z[t±1 ]) ∼
= H1 (π; Z[t±1 ]). This concludes the proof
of the claim.
We now recall that Z[t±1 ] is a Noetherian ring. This implies that every ascending
chain of submodules of a given finitely generated module stabilizes. This in turn
implies that every finitely generated module over Z[t±1 ] which is given by infinitely
many relations is already described by a finite subset of those relations. (We refer to
[Lan02, Chapter X.§1] for details.)
It thus follows that there exists an l ∈ N such that the following matrix is already
a presentation matrix for H1 (π; Z[t±1 ]).
!
−1
∂t gi t
∂ri
φ
φ
.
∂gj
∂gj i=1,...,l,j=1,...,k
i,j=1,...,k
12
STEFAN FRIEDL AND STEFANO VIDUSSI
The block matrix on the right is a matrix over Z, whereas the block matrix on the
left is a matrix of the form tidk − B where B is a matrix over Z. If we now take the
gcd of the k × k-minors of the matrix we see that the top coefficient is 1. The proof
for the descending case is identical.
We have now the following:
Proposition 4.2. Let π be the finitely presented group of Eq. (5) and let φ ∈
H 1 (π; Z) ∼
= Z be the generator given by the epimorphism of Hom(π, Z) determined
by φ(t) = 1 and φ(a) = φ(b) = 0. The pair (π, φ) has zero rank gradient but it cannot
be written as an ascending or descending HNN extension.
φ
Proof. We denote by πn the kernel of the epimorphism π −
→ Z → Z/n. It follows
from the Reidemeister-Schreier process, that πn has a presentation of the form
hs, a0 , . . . , an−1 , b0 , . . . , bn−1 | a1 = a20 , . . . , an−1 = a2n−2 , s−1 a0 s = a2n−1 ,
b21 = b0 , . . . , b2n−1 = bn−2 , s−1 b20 s = b2n−1 i
Here, the map πn → π is given by sending s to tn and sending ai 7→ t−i ati and
bi 7→ t−i bti . The fact that this is a correct presentation for πn can also be seen easily
by thinking of π as the fundamental group of a 2-complex with one 0-cell, three 1-cells
and two 2-cells.
Note that in the presentation of πn we can remove the generators a1 , . . . , an−1 and
b1 , . . . , bn−1 . We can thus find a presentation for πn with three generators. It is now
immediate that rg (π, φ) is zero.
A straightforward calculation using Fox calculus shows that the Alexander polynomial of (π, φ) is (t − 2)(2t − 1) = 2 − 5t + 2t2 . It now follows from Lemma 4.1 that the
pair (π, φ) corresponds neither to an ascending nor descending HNN extension.
Appendix
We include in this appendix definitions and results needed in the proof of Theorem
3.3. The purpose is to show that the class of groups discussed in Section 3 satisfies the
assumptions of Theorem 8.1 of [BNS87], which thus determines their BNS invariant.
This is arguably standard fare for the practitioners of Thompson–type groups, but
may be less obvious for others, so we provide here the details.
Let G be a subgroup of the group of orientation–preserving PL homeomorphisms
of an interval [0, ℓ]. We say that such a group is irreducible if it has no fixed point
in (0, ℓ). Given such a group, we can define two homomorphisms λ, ρ : G → R+ with
values in the multiplicative group R+ as follows:
λ(g) :=
dg
dg
(0), ρ(g) := (ℓ),
dt
dt
RANK GRADIENTS OF INFINITE CYCLIC COVERS OF KÄHLER MANIFOLDS
13
which return the slope at the endpoints. We say that λ and ρ are independent if
λ(G) = λ(Kerρ) and ρ(G) = ρ(Kerλ). Equivalently,
(ρ,λ)
G −−→ λ(G) × ρ(G)
is an epimorphism. Then we have the following theorem (see [BNS87, Theorem 8.1],
as well as [Bro87b, Proposition 8.2]).
Theorem. (Bieri–Neumann–Strebel) Let G be a finitely generated, irreducible subgroup of the group of orientation–preserving PL homeomorphisms of an interval [0, ℓ].
Then if λ and ρ are independent, the set of exceptional characters of G is given by
[log(λ)], [log(ρ)] ∈ Hom(G, R) = H 1 (G; R).
(The statement in [BNS87] covers subgroups of the group of orientation–preserving
PL homeomorphisms of the unit interval [0, 1], but the proof applies with obvious
modifications to the case stated here.)
We will apply this result to the study of the BNS invariants for the case of the
1
generalized Thompson groups F (ℓ, Z[ n1 ...n
], hn1 , . . . , nk i) which, we recall, are defined
k
as follows. Given a nontrivial multiplicative subgroup
hn1 , . . . , nk i = {nj11 . . . njkk |ji ∈ Z} ≤ R+
(where {n1 , . . . , nk } is a positive integer basis of the multiplicative group), we de1
], hn1 , . . . , nk i) the group of orientation–preserving PL homeonote by F (ℓ, Z[ n1 ...n
k
1
morphism of the interval [0, ℓ], where 0 < ℓ ∈ Z[ n1 ...n
], with finitely many singularik
1
ties all in Z[ n1 ···nk ] and slopes in hn1 , . . . , nk i.
1
Corollary. Let G = F (ℓ, Z[ n1 ···n
], hn1 , . . . , nk i) be a generalized Thompson group;
k
then G admits exceptional characters.
Proof. The proof of the corollary consists in verifying that the assumptions of the
Bieri–Neumann-Strebel Theorem are satisfied. First, we show that the group G
is irreducible: we can do so by constructing an explicit element g ∈ G such that
g : [0, ℓ] → [0, ℓ] has no other fixed points but {0, ℓ}. To do so, choose an element
Q
ν ∈ hn1 , . . . , nk i such that ν > 1, and define n := ki=1 ni ≥ 2. Consider the pairs
(x1 , y1 ), (x2 , y2 ) ⊂ [0, ℓ] × [0, ℓ] with coordinates
1 ℓ ℓ
ℓ
1ℓ
(x1 , y1 ) =
, (x2 , y2 ) = ℓ − , ℓ −
.
,
νn n
n
νn
then x1 < x2 ∈ (0, ℓ) ∩ Z[ n1 ]. The pieceIt is immediate to verify that as ℓ ∈
wise linear graph joining the four points (0, 0), (x1 , y1 ), (x2 , y2 ), (ℓ, ℓ) ∈ [0, ℓ] × [0, ℓ]
is composed of three segments of slope respectively ν, 1, ν1 ∈ hn1 , · · · , nk i. Therefore,
this represents the graph of an element g ∈ G that has no fixed points in the interval
(0, ℓ), hence the proof that G is irreducible. On the left hand side of Figure 1 we show
the graph of such an element for the generalized Thompson group F (1, Z[ 61 ], h2, 3i),
with ν = 2.
Z[ n1 ],
14
STEFAN FRIEDL AND STEFANO VIDUSSI
Next, let’s show that λ and ρ are independent. We will limit ourselves to show that
λ(G) = λ(Kerρ), the proof of the symmetric relation being identical. Namely, we will
show that for any element ν ∈ hn1 , . . . , nk i there exists an element h ∈ Kerρ (i.e. for
(l) = 1) such that dh
(0) = ν. The proof is based on a suitable scaled
which ρ(h) = dh
dt
dt
version of the construction above. Define ℓ̃ = ℓ − nℓ , and assume first ν > 1. We
consider now the pairs
(x1 , y1), (x2 , y2 ) ⊂ [0, ℓ̃] × [0, ℓ̃] ⊂ [0, ℓ] × [0, ℓ]
with coordinates
(x1 , y1 ) =
1 ℓ̃ ℓ̃
1 ℓ̃
ℓ̃
,
, (x2 , y2 ) = ℓ̃ − , ℓ̃ −
.
νn n
n
νn
Similar to above, it is immediate to verify that x1 < x2 < ℓ̃ ∈ (0, ℓ) ∩ Z[ n1 ]. The
piecewise linear graph joining the five points (0, 0), (x1 , y1), (x2 , y2), (ℓ̃, ℓ̃), (ℓ, ℓ) ∈
[0, ℓ] ×[0, ℓ] is composed of four segments of slope respectively ν, 1, ν1 , 1 ∈ hn1 , . . . , nk i.
It follows that this represents the graph of an element h ∈ G such that λ(h) = ν and
ρ(h) = 1. The right hand side of Figure 1 show the graph of h for the generalized
Thompson group F (1, Z[ 61 ], h2, 3i), with ν = 2.
y
y
x
x
g ∈ F (1, Z[ 61 ], h2, 3i); λ(g) = 2, ρ(g) =
1
2
h ∈ F (1, Z[ 16 ], h2, 3i); λ(h) = 2, ρ(h) = 1
Figure 1
The proof for the case of ν < 1 is similar, defining this time
ℓ ℓ
ℓ
ℓ
, ν , (x2 , y2) = ℓ − ν , ℓ −
(x1 , y1 ) =
n
n
n
n
and performing similar elementary calculations. It now follows from the theorem of
Bieri–Neumann–Strebel that the set of exceptional characters is nonempty as claimed
in the statement.
RANK GRADIENTS OF INFINITE CYCLIC COVERS OF KÄHLER MANIFOLDS
15
References
[ABCKT96] J. Amorós, M. Burger, K. Corlette, D. Kotschick and D. Toledo, Fundamental groups of
compact Kähler manifolds, Mathematical Surveys and Monographs, 44. American Mathematical
Society, Providence, RI, 1996.
[Ag13] I. Agol, The virtual Haken conjecture, with an appendix by I. Agol, D. Groves and J. Manning, Documenta Math. 18 (2013) 1045–1087.
[BM07] G. Baumslag and C. F. Miller III, Finitely presented extensions by free groups, J. Group
Theory 10 (2007), 723–729.
[BNS87] R. Bieri, W. Neumann and R. Strebel, A geometric invariant of discrete groups, Invent.
Math. 90 (1987), no. 3, 451–477.
[BGK10] R. Bieri, R. Geoghegan and D. Kochloukova, The sigma invariant of Thompson’s group
F , Groups Geom. Dyn. 4 no 2 (2010), 263–273.
[BS78] R. Bieri and R. Strebel, Almost finitely presented soluble groups, Comment. Math. Helv. 53
(1978), 258–278.
[BrSq85] M. Brin and C. Squier, Groups of piecewise linear homeomorphisms of the real line, Invent.
Math. 79 (1985), no. 3, 485–498.
[Bro87a] K. S. Brown, Finiteness properties of groups, Proceedings of the Northwestern conference
on cohomology of groups (Evanston, Ill., 1985), J. Pure Appl. Algebra 44 (1987), no. 1-3, 45–75.
[Bro87b] K. S. Brown, Trees, valuations, and the Bieri–Neumann–Strebel invariant, Invent. Math.
90 (1987), no. 3, 479–504.
[Bu11] M. Burger, Fundamental groups of Kähler manifolds and geometric group theory, Séminaire
Bourbaki 62 (2009/2010). Exposé 1022. Astérisque No. 339 (2011)
[Ca03] F. Catanese, Fibred Kähler and quasi–projective groups, Special issue dedicated to Adriano
Barlotti, Adv. Geom. suppl. (2003), 13–27.
[De10] T. Delzant, L’invariant de Bieri Neumann Strebel des groupes fondamentaux des variétés
kähleriennes, Math. Annalen 348 (2010), 119–125.
[DFV14] J. DeBlois, S. Friedl, and S. Vidussi, Rank gradients of infinite cyclic covers of 3-manifolds,
Michigan Math. J. 63 (2014), no. 1, 65–81.
[Lan02] S. Lang, Algebra, third revised edition, Graduate Texts in Mathematics. 211. (2002)
[Lac05 ] M. Lackenby, Expanders, rank and graphs of groups, Israel J. Math., 146 (2005), 357–370.
[Lu96] A. Lubotzky, Free quotients and the first Betti number of some hyperbolic manifolds, Transform. Groups 1, no. 1-2 (1996), 71–82.
[NR01] T. Napier and M. Ramachandran, Hyperbolic Kähler manifolds and proper holomorphic
mappings to Riemann surfaces, Geom. Funct. Anal. 11, no 2 (2001), 382–406.
[NR06] T. Napier and M. Ramachandran, Thompson’s group F is not Kähler, Topological and
asymptotic aspects of group theory, 197–201, Contemp. Math., 394, Amer. Math. Soc., Providence, RI, 2006.
[NR08] T. Napier, and M. Ramachandran, Filtered ends, proper holomorphic mappings of Kähler
manifolds to Riemann surfaces, and Kähler groups, Geom. Funct. Anal. 17, no 5 (2008), 1621–
1654.
[Sc78] P. Scott, Subgroups of surface groups are almost geometric, J. London Math. Soc. (2) 17 no.
3 (1978), 555–565.
[Se80] J.-P. Serre, Trees, Springer-Verlag, Berlin-New York, 1980.
[Sta62] J. Stallings, On fibering certain 3–manifolds, 1962 Topology of 3–manifolds and related
topics (Proc. The Univ. of Georgia Institute, 1961) pp. 95–100 Prentice-Hall, Englewood Cliffs,
N.J. (1962)
[Ste92] M. Stein, Groups of piecewise linear homeomorphisms, Trans. Amer. Math. Soc. 332 (1992),
no. 2, 477–514.
16
STEFAN FRIEDL AND STEFANO VIDUSSI
[Wi12] D. Wise, The structure of groups with a quasi-convex hierarchy, 189 pages, preprint (2012),
downloaded on October 29, 2012 from http://www.math.mcgill.ca/wise/papers.html
Fakultät für Mathematik, Universität Regensburg, Germany
E-mail address: [email protected]
Department of Mathematics, University of California, Riverside, CA 92521, USA
E-mail address: [email protected]
| 4math.GR
|
SpCoSLAM 2.0: An Improved and Scalable Online Learning of
Spatial Concepts and Language Models with Mapping*
arXiv:1803.03481v1 [cs.RO] 9 Mar 2018
Akira Taniguchi1 , Yoshinobu Hagiwara1 , Tadahiro Taniguchi1 , and Tetsunari Inamura2
Abstract— In this paper, we propose a novel online learning
algorithm, SpCoSLAM 2.0 for spatial concepts and lexical
acquisition with higher accuracy and scalability. In previous
work, we proposed SpCoSLAM as an online learning algorithm
based on the Rao–Blackwellized particle filter. However, this
conventional algorithm had problems such as the decrease
of the estimation accuracy due to the influence of the early
stages of learning as well as the increase of the computational
complexity with the increase of the training data. Therefore,
we first develop an improved algorithm by introducing new
techniques such as rejuvenation. Next, we develop a scalable
algorithm to reduce the calculation time while maintaining
a higher accuracy than the conventional algorithm. In the
experiment, we evaluate and compare the estimation accuracy
and calculation time of the proposed algorithm, conventional
online algorithm, and batch learning. The experimental results
demonstrate that the proposed algorithm not only exceeds the
accuracy of the conventional algorithm but also capable of
achieving an accuracy comparable to that of batch learning. In
addition, the proposed algorithm showed that the calculation
time does not depend on the amount of training data and
becomes constant for each step with the scalable algorithm.
I. INTRODUCTION
Robots operating in various human-living environments
are required to adaptively and sequentially acquire the spatial
concepts and vocabulary related to various places as well
as the environmental map. It is difficult to manually design
spatial concepts in advance, thus it is desirable for robots
to autonomously acquire spatial concepts and vocabulary
based on their own experiences. Related research in the fields
of semantic mapping and place categorization [1]–[4] have
attracted considerable interests in recent years. However,
the conventional approaches of most of these studies could
not learn unknown words without the pre-set vocabulary.
In addition, the processes for simultaneous localization and
mapping (SLAM) [5] and the estimation of semantics related
to places have been addressed separately. However, in our
proposed approach, the robot can automatically and sequentially perform place categorization and learn unknown words
without any prior knowledge. In addition, the proposed un*This work was partially supported by JST CREST grant number
JPMJCR15E3, JSPS KAKENHI grant number JP17J07842, and a Grantin-Aid for Scientific Research on Innovative Areas (16H06561) funded by
the Ministry of Education, Culture, Sports, Science, and Technology, Japan.
1 Akira Taniguchi, Yoshinobu Hagiwara, and Tadahiro Taniguchi
are with Ritsumeikan University, 1-1-1 Noji-Higashi, Kusatsu, Shiga
525-8577, Japan. {a.taniguchi, yhagiwara, taniguchi}
@em.ci.ritsumei.ac.jp
2 Tetsunari Inamura is with the National Institute of Informatics / SOKENDAI (The Graduate University for Advanced Studies), 2-1-2 Hitotsubashi, Chiyoda-ku, Tokyo 101-8430, Japan. [email protected]
“kyouiNkenkyuushitsunikimashita”
(We come to the faculty lab.)
“konobashogaroboqtookibadayo”
(This place is the robot storage space.)
OK! I learn
quickly!
OK! I learn
correctly!!
Online learning
/robotookiba/
/kimashita/
/idi/
/idomari/
/pyoigegyushitsu/
/kyoiNkeNkiyushitsu/
Fig. 1. Overview of an improved and scalable online learning; The robot
can additionally learn spatial concepts while sequentially correcting mistakes
in previous learning based on interaction with an user and an environment,
as shown from the bottom left to the bottom right.
supervised Bayesian generative model integrates multimodal
place categorization, lexical acquisition, and SLAM.
Taniguchi et al. [7], [8] proposed the nonparametric
Bayesian spatial concept acquisition method (SpCoA) using
an unsupervised word-segmentation method latticelm [9] and
SpCoA++, which enables highly accurate lexical acquisition by updating the language model. Isobe et al. [10]
proposed a learning method for the relationship between
objects and places using image features obtained by the
convolutional neural network (CNN) [11]. However, these
methods [7], [8], [10] cannot learn spatial concepts from
unknown environments without a map because they rely
on batch-learning algorithms. Therefore, we developed in
previous work an online algorithm, SpCoSLAM [6], that
can sequentially learn a map, spatial concepts that integrate
positions, speech signals, and scene images.
One of the advantages of online learning is that the spatial
concepts mistakenly learned by the robot can be corrected
sequentially, something which could not be achieved thus
far. Moreover, we consider that it can flexibly deal with
changes in the environment and the names of places. However, to enable online learning based, long-term human–
robot interactions with limited computational resources, the
following problems need to be solved: (i) the increase in
calculation cost owing to an increase in the amount of
data, and (ii) the decrease in estimation accuracy by online
learning when compared with batch learning. More specifically, the conventional algorithm [6] caused the accuracy of
m
ut −1
zt −1
ut
μk
V0 ,ν 0
Σk
α
λ
π
Ct
AM
Fig. 2.
zt +1
ut +1
xt
xt −1
m0 ,κ0
zt
x t +1
it
φl
γ
ft
θl
χ
St
Wl
β
yt
LM
λ
λ
m
xt
zt
ut
ft
yt
it
Ct
St
µk , Σk
π
φl
θl
Wl
LM
AM
α, β, γ, χ, λ
m0 , κ0 , V0 , ν0
Environmental map
Self-position of a robot
Depth data
Control data
Image feature
Speech signal
Index of position distributions
Index of spatial concepts
Word sequence (word segmentation result)
Parameters of Gaussian distribution (position distribution)
Parameter of multinomial distribution for index Ct of spatial concepts
Parameter of multinomial distribution for index it of position distribution
Parameter of multinomial distribution for image feature
Parameter of multinomial distribution for the names of places
Language model (word dictionary)
Acoustic model for speech recognition
Hyperparameters of prior distributions
Left: Graphical model representation of SpCoSLAM [6]; Gray nodes indicate observation variables. Right: Description of random variables
categorization and word segmentation to decrease because
sufficient statistical information could not be used at the early
stages of learning. In addition, the computational complexity
of the processes of speech recognition and unsupervised
word segmentation increased with an increase in the training
data. In the online learning algorithms for mobile robots, the
increase in scalability without reducing accuracy is extremely
important but difficult to achieve. Therefore, we aim to
develop improved and scalable algorithms to solve the abovementioned problems. An overview is shown in Fig. 1.
In the one hand, we assume that the robot has not acquired
any vocabulary in advance and can recognize only phonemes
or syllables. In the other hand, we represent the spatial
area of the environment in terms of a position distribution.
Furthermore, we define a spatial concept as a place category
that includes the place names, the scene image features, and
the position distributions.
The main contributions of this paper are as follows:
• We propose an improved and scalable online learning
algorithm with several novel extensions from the conventional algorithm.
• The improved algorithm achieves accuracy comparable
to batch learning.
• The scalable algorithm reduces the order of the computational complexity.
II. S P C O SLAM: ONLINE SPATIAL CONCEPT
ACQUISITION
A. Overview
SpCoSLAM [6] can perform online spatial concept acquisition, lexical acquisition, and simultaneous localization and
mapping in an integrative manner. Fig. 2 shows the graphical
model of SpCoSLAM and lists each variable of the graphical
model. The details of the formulation of the generation
process represented by the graphical model are described
in [6]. This method can learn sequential spatial concepts for
unknown environments without maps. It can also learn the
many-to-many correspondences between names and places
via spatial concepts and can mutually complement the
uncertainty of information using multimodal information.
The procedure of SpCoSLAM for each step is described
as follows. (a) The robot obtains the weighted finite-state
transducer (WFST) speech recognition results from the user’s
speech signals using a language model. (b) The robot obtains
the likelihood of self-localization by performing FastSLAM.
(c) The robot segments the WFST speech recognition results
using the unsupervised word segmentation, latticelm [9]. (d)
The robot obtains the latent variables of spatial concepts by
sampling. (e) The robot obtains the marginal likelihood of
the observed data as the importance weight. (f) The robot
updates the environmental map. (g) The robot estimates
the set of model parameters of the spatial concepts from
the observed data and the sampled variables. (h) The robot
updates the language model of the maximum weight for the
next step. (i) The particles are resampled according to their
weights. Steps (b)–(g) are performed for each particle.
B. Online Spatial Concept Acquisition and Mapping
The online learning algorithm of SpCoSLAM introduced
sequential equation updates to estimate the parameters of the
spatial concepts into the formulation of Rao-Blackwellized
particle filter (RBPF) in the FastSLAM 2.0 algorithm [12].
The particle filter is advantageous in that parallel processing
can be easily applied because each particle can be calculated
independently.
In the formulation of SpCoSLAM, the joint posterior
distribution can be factorized to the probability distributions
of a language model LM , a map m, the set of model
parameters of spatial concepts Θ = {W, µ, Σ, θ, φ, π}, the
joint distribution of the self-position trajectory x0:t , and the
set of latent variables C1:t = {i1:t , C1:t , S1:t }. We describe
the joint posterior distribution of SpCoSLAM as follows:
p(x0:t , C1:t , LM, Θ, m | u1:t , z1:t , y1:t , f1:t , AM, h)
= p(LM | S1:t , λ)p(Θ | x0:t , C1:t , f1:t , h)p(m | x0:t , z1:t )
(1)
· p(x0:t , C1:t | u1:t , z1:t , y1:t , f1:t , AM, h)
{z
}
|
Particle filter
where the set of hyperparameters is denoted as h =
{α, β, γ, χ, λ, m0 , κ0 , V0 , ν0 }. It is noteworthy that the
speech signal yt is not observed at all time steps. Herein,
the proposed method is equivalent to FastSLAM when yt is
not observed, i.e., when the speech signal is a trigger for the
place categorization.
The particle filter algorithm uses sampling importance
resampling (SIR). The importance weight is denoted by
[r]
[r]
[r]
ωt = Pt /Qt for each particle, where r is the particle
index. The number of particles is R. Henceforth, the following equations are also calculated for each particle r; however,
the subscripts representing the particle index are omitted.
We describe the target distribution Pt as follows:
Pt
=
p(x0:t , C1:t | u1:t , z1:t , y1:t , f1:t , AM, h)
≈
p(it , Ct | x0:t , i1:t−1 , C1:t−1 , S1:t , f1:t , h)
q(x0:t , C1:t | u1:t , z1:t , y1:t , f1:t , AM, h)
=
p(xt | xt−1 , zt , mt−1 , ut )
· p(it , Ct | x0:t , i1:t−1 , C1:t−1 , S1:t , f1:t , h)
· p(St | S1:t−1 , y1:t , AM, λ) · Qt−1 .
(3)
Then, p(xt | xt−1 , zt , mt−1 , ut ) is the equivalent to the
proposal distribution of FastSLAM 2.0. The probability distribution of it and Ct is the marginal distribution pertaining
to the set of model parameters Θ. This distribution can be
calculated by a formula equivalent to the collapsed Gibbs
sampling. The details are described in [6]. We approximate
the probability distribution of St as the speech recognition
with the language model LMt−1 and the unsupervised word
segmentation using the WFST speech recognition results by
latticelm [9] as follows:
p(St | S1:t−1 , y1:t , AM, λ)
≈ latticelm(S1:t | L1:t , λ)SR(L1:t | y1:t , AM, LMt−1 )
(4)
where SR() denotes the function of speech recognition, L1:t
denotes the speech recognition results in the WFST format,
which is a word graph representing the speech recognition
results.
Finally, ωt is represented as follows:
≈ p(zt | mt−1 , xt−1 , ut )p(ft | C1:t−1 , f1:t−1 , h)
p(St | S1:t−1 , C1:t−1 , α, β)
·
· ωt−1 .
(5)
p(St | S1:t−1 , β)
In the formulation of (1), it is desirable to estimate the
language model LMt for each particle. In other words,
speech recognition of the amount of data multiplied by
the number of particles for each teaching utterance must
be performed. In this paper, to reduce the computational
cost, we use a language model LMt of a particle with the
maximum weight for speech recognition.
ωt
Pt
≈
p(it , Ct | x0:t , i1:t−1 , C1:t−1 , S1:t , f1:t , h)
· p(zt | xt , mt−1 )p(ft | C1:t−1 , f1:t−1 , h)
· p(xt | xt−1 , ut )p(St | S1:t−1 , y1:t , AM, λ)
· p(xt | x0:t−1 , i1:t−1 , C1:t−1 , h)
p(St | S1:t−1 , C1:t−1 , α, β)
·
· Pt−1 .
p(St | S1:t−1 , β)
ωt
· p(xt | xt−1 , ut )p(St | S1:t−1 , y1:t , AM, λ)
p(St | S1:t−1 , C1:t−1 , α, β)
· Pt−1 .
(2)
·
p(St | S1:t−1 , β)
We describe the proposal distribution Qt as follows:
=
Herein, we applied two modifications related to the
weighting of the original SpCoSLAM algorithm [6].
1) Additional Weight for it , Ct and xt (AW): We modify
the derivation of (2) in the target distribution Pt as follows:
(6)
Finally, ωt is modified as follows:
· p(zt | xt , mt−1 )p(ft | C1:t−1 , f1:t−1 , h)
Qt
C. Modification
≈ p(zt | mt−1 , xt−1 , ut )p(ft | C1:t−1 , f1:t−1 , h)
Xh
p(xt | x0:t−1 , i1:t−1 , it = k, h)
·
it =k
·
X
i
p(it = k, Ct = l | C1:t−1 , i1:t−1 , h)
Ct =l
p(St | S1:t−1 , C1:t−1 , α, β)
· ωt−1 .
·
p(St | S1:t−1 , β)
(7)
Unlike the original algorithm of SpCoSLAM, the marginal
likelihood for it and Ct weighted by the marginal likelihood
for the position distribution was added to the third term
on the right side. The calculation amount does not increase
because the most of the formulas for weight ωt were already
calculated when it and Ct were sampled. This is described
in Algorithm 1 (line 16) and Algorithm 2 (line 17).
2) Weight for the Selection of a Language Model LM
(WS): We modify the weight for the selection of a language
model from the whole weight ωt of the model to the weight
ωS related to the word information as follows:
ωS = p(S1:t | C1:t−1 , α, β)/p(S1:t | β).
(8)
The segmentation result of all of the uttered sentences for
each particle changes at every step because of the word segmentation processes that use all previous data. Therefore, we
consider that better word segmentation results can be selected
by the weight that considers not only the current data but
also all previous data. In addition, this weight corresponds
to the selection of the word segmentation results from the
mutual information in SpCoA++ [8]. This is described in
Algorithm 1 (line 23) and Algorithm 2 (line 24).
III. S P C O SLAM 2.0: IMPROVED AND SCALABLE
ONLINE LEARNING ALGORITHM
In this section, we describe an improved and scalable
online learning algorithm to overcome the problems in
SpCoSLAM. First, we develop an improved algorithm by
introducing new techniques such as rejuvenation as shown
in Section III-A. Next, we develop a scalable algorithm to
reduce the calculation time while maintaining a higher accuracy than the conventional algorithm as shown in Section IIIB. The proposed algorithms are theoretically reasonable
compared to heuristic.
Algorithm 1 SpCoSLAM 2.0 algorithm (Improved)
A. Improvement
We describe in this section the details of the improved
algorithm. A pseudo-code for the improved algorithm is
given by Algorithm 1.
1) Fixed-lag Rejuvenation of it and Ct (FLR–it , Ct ):
Canini et al. [13] improved the accuracy with the rejuvenation that resamples some previous samples randomly.
However, in a completely random choice, all of the previous
samples need to be held. In the proposed algorithm, we introduce the fixed-lag rejuvenation (FLR). The FLR, inspired
by the fixed-lag smoothing [14], performs the resampling of
latent variables iτ and Cτ by iterating τ times from the step
t − TL + 1 to the current step t as follows:
iτ , Cτ ∼ p(iτ , Cτ | x0:t , S1:t , f1:t , i{1:t|¬τ } , C{1:t|¬τ } , h)
(9)
where i{1:t|¬τ } and C{1:t|¬τ } denote the sets of elements
from 1 to t without the elements of step τ . In this case, the
latent variables of step t − TL can be sampled using the data
up to step t, as described in Algorithm 1 (lines 12–14).
2) Re-segmentation (RS): We introduce the resegmentation of word sequences to improve the accuracy
of word segmentation. In the conventional algorithm, we
approximated the left side of (10) by registering the word
sequences segmented by latticelm to the word dictionary.
However, this can be considered as a process of sampling
∗
a language model LM from word sequences S1:t
and a
hyperparameter λ of a language model. Therefore, we adopt
the nested Pitman–Yor language model (NPYLM) [15] to
estimate a language model from the word sequences as
follows:
∗
∗
p(LM | S1:t
, λ) = NPYLM(LM | S1:t
, λ).
B. Scalability
We describe in this section the details of the scalable
algorithm. A pseudo-code for the scalable algorithm is given
in Algorithm 2.
1) WFST Speech Recognition and Unsupervised Word
Segmentation using FLR (FLR–St ): FLR can also be extended to the sampling of St . The data used for speech
recognition and word segmentation are changed from all data
such as in (4) to the data of the fixed-lag interval. In this case,
word segmentation can be performed based on the statistical
information collected from the WFSTs of the number of lag
values TL . FLR performs the sampling of word sequences
St−TL +1:t of time steps from t − TL + 1 to the current step
t as follows:
11:
12:
13:
14:
15:
16:
17:
18:
19:
20:
21:
22:
23:
24:
25:
26:
27:
28:
29:
30:
[r]
[r]
[r]
xt = scan matching(zt , x́t , mt−1 )
for j = 1 to J do
[r]
xj = sample motion model(ut , xt−1 )
end forP
[r]
[r]
ωz = Jj=1 measurement model(zt , xj , mt−1 )
[r]
S1:t ∼ latticelm(S1:t | L1:t , λ)
for τ = t − TL + 1 to t do
[r]
[r]
[r]
[r]
iτ , Cτ ∼ p(iτ , Cτ | x0:t , S1:t , f1:t ,
[r]
[r]
i{1:t|¬τ } , C{1:t|¬τ } , h)
end for
[r]
[r]
ωf = p(ft | C1:t−1
, f1:t−1 , α, χ)
h
P
[r]
[r]
[r]
[r]
ωic =
| x0:t−1 , i1:t−1 , it = k, h)
it =k p(xt
i
P
[r]
[r]
· Ct =l p(it = k, Ct = l | C1:t−1 , i1:t−1 , h)
[r]
[r]
[r]
[r]
[r]
[r]
ωs = p(St | S1:t−1 , C1:t−1 , α, β)/p(St | S1:t−1 , β)
[r]
[r]
[r]
[r]
[r]
ωt = ωz · ωf · ωs · ωic
[r]
[r]
[r]
mt = updated occupancy grid(zt , xt , mt−1 )
[r]
[r]
[r]
Θt = E[p(Θ | x0:t , C1:t , f1:t , h)]
[r]
[r]
[r]
[r]
[r]
X̄t = X̄t ∪ hx0:t , C1:t , mt , Θt , ωt i
end for
P
[r]
[r]
∗
S1:t
= argmaxS [r] R
r=1 ωS δ(S1:t − S1:t )
1:t
∗
LMt ∼ NPYLM(LM | S1:t
, λ)
for r = 1 to R do
[i]
draw i with probability ∝ ωt
[i]
[i]
[i]
[i]
add hx0:t , C1:t , mt , Θt , LMt i to Xt
end for
return Xt
end procedure
2) Sequential Bayesian Update of Parameters in Posterior
Distribution (SBU): We change the algorithm to the sequential Bayesian update (SBU) of the set of hyperparameters H
in the posterior distribution. In the conventional algorithm,
the model parameters Θ are estimated from all the data D1:t
and the set of latent variables C1:t for each step. However,
in FastSLAM, a map mt is updated from xt , zt , and mt−1
sequentially. Similarly, the posterior hyperparameters Ht can
be calculated from the new data Dt , latent variables Ct , and
posterior hyperparameters Ht−1 at the previous time. Doing
so, both the computational and memory efficiency, important
in long-term learning using real robots, can be significantly
improved. The SBU for the posterior hyperparameters is
calculated as follows:
p(Θ | Ht )
= p(Θ | D1:t , C1:t , h)
= p(Θ | Dt , Ct , {D1:t−1 , C1:t−1 , h})
= p(Θ | Dt , Ct , Ht−1 )
∝ p(Dt | Ct , Θ)p(Θ | Ht−1 ).
∼ latticelm(St−TL +1:t | Lt−TL +1:t , λ)
· SR(Lt−TL +1:t | yt−TL +1:t , AM, LMt−1 )
(11)
This is described in Algorithm 2 (lines 4 and 12).
6:
7:
8:
9:
10:
(10)
Then, we convert the sentences into syllable sequences, and
segment them into word sequences by NPYLM, as described
in Algorithm 1 (line 24).
St−TL +1:t
1: procedure SpCoSLAM(Xt−1 , ut , zt , f1:t , y1:t )
2:
X̄t = Xt = ∅
3:
L1:t = SR(L1:t | y1:t , AM, LMt−1 )
4:
for r = 1 to R do
[r]
[r]
5:
x́t = sample motion model(ut , xt−1 )
(12)
These posterior hyperparameters Ht can also be used in the
sampling of Ct . In the implementation, it is sufficient to hold
the statistics obtained during the calculation of the posterior
distribution.
TABLE I
C OMPUTATIONAL COMPLEXITY OF LEARNING ALGORITHMS
Algorithm
SpCpSLAM [6]
SpCoSLAM 2.0 (Improved)
SpCoSLAM 2.0 (Scalable)
SpCoA [7] (Batch learning)
SpCoA++ [8] (Batch learning)
Algorithm 2 SpCoSLAM 2.0 algorithm (Scalable)
1: procedure SpCoSLAM(Xt−1 , ut , zt , ft−TL +1:t , yt−TL +1:t )
2:
X̄t = Xt = ∅
3:
t0 = t − TL
4:
Lt0 +1:t = SR(Lt0 +1:t | yt0 +1:t , AM, LMt0 )
5:
for r = 1 to R do
[r]
[r]
6:
x́t = sample motion model(ut , xt−1 )
Order
O(N R)
O(N R)
O(TL R)
O(N G)
O(N GM I)
In addition, we consider that the word segmentation problem can be improved by a combined method of FLR and the
SBU update. In this case, speech recognition is performed
using the initial syllable dictionary in the steps earlier than
step TL and using a word dictionary of the step t − TL in
the steps later than step TL + 1. When used together with
FLR, the SBU equation is shown as follows:
p(Θ | Ht )
7:
8:
9:
10:
11:
12:
13:
14:
15:
16:
17:
=
p(Θ | D1:t , C1:t , h)
=
p(Θ | Dt0 +1:t , Ct0 +1:t , {D1:t0 , C1:t0 , h})
=
p(Θ | Dt0 +1:t , Ct0 +1:t , Ht0 )
[r]
[r]
Ct =l
[r]
(13)
18:
where a time step before the lag value is t0 = t − TL . In
this case, it is only necessary to hold the observed data and
posterior hyperparameters of the number corresponding to
the lag value TL . Equation (13) is applied to Algorithm 2.
Table I shows the order of computational complexity for
each learning algorithm. The data number is denoted N ,
the number of particles R, the value of fixed-lag TL , the
number of iterations for Gibbs sampling in batch learning
G, the number of candidates of word segmentation results
for updating the language model in SpCoA++ M , and
the number of iterations for the parameter estimation in
SpCoA++ I. In this case, the computational efficiency of the
scalable algorithm is higher than the previous SpCoSLAM
algorithm when TL < N .
19:
∝ p(Dt0 +1:t | Ct0 +1:t , Θ)p(Θ | Ht0 ),
[r]
xt = scan matching(zt , x́t , mt−1 )
for j = 1 to J do
[r]
xj = sample motion model(ut , xt−1 )
end forP
[r]
[r]
ωz = Jj=1 measurement model(zt , xj , mt−1 )
[r]
St0 +1:t ∼ latticelm(St0 +1:t | Lt0 +1:t , λ)
for τ = t0 + 1 to t do
[r]
[r]
[r]
[r]
iτ , Cτ ∼ p(iτ , Cτ | xt0 +1:t , St+1:t , ft0 +1:t ,
[r]
[r]
[r]
i{t0 +1:t|¬τ } , C{t0 +1:t|¬τ } , Ht0 )
end for
[r]
[r]
[r]
ωf = p(ft h| Ct0 +1:t−1 , ft0 +1:t−1 , Ht0 )
X
[r]
[r]
[r]
[r]
[r]
ωic =
p(xt | xt0 +1:t−1 , it0 +1:t−1 , it = k, Ht0 )
i
Xit =k
[r]
[r]
[r]
·
p(it = k, Ct = l | Ct0 +1:t−1 , it0 +1:t−1 , Ht0 )
IV. EXPERIMENTS
In this section, we performed the experiments to demonstrate the online learning of spatial concepts from a novel
environment. In addition, we performed the evaluations of
place categorization and lexical acquisition related to places.
We compared the performances of the following methods:
(A) SpCoSLAM [6]
(B) SpCoSLAM with AW (Section II-C.1)
(C) SpCoSLAM with AW + WS (Section II-C.2)
(D) SpCoSLAM 2.0 (FLR–it , Ct )
(E) SpCoSLAM 2.0 (RS)
(F) SpCoSLAM 2.0 (FLR–it , Ct + RS)
(G) SpCoSLAM 2.0 (FLR–it , Ct , St + SBU)
(H) SpCoA (Batch learning) [7]
(I) SpCoA++ (Batch learning) [8]
Methods (A)–(C) performed the conventional and modified SpCoSLAM algorithms. Methods (D)–(F) performed
the proposed improved algorithms in different conditions.
In methods (D) and (F), the lag value for FLR was set
20:
21:
22:
23:
24:
[r]
ωs
[r]
=
[r]
[r]
[r]
p(St |S 0
,C 0
,H 0 )
t +1:t−1
t +1:t−1
t
[r]
[r]
[r]
p(St |S 0
,H 0 )
t +1:t−1
t
[r]
[r]
[r]
[r]
ωt = ωz · ωf · ωs · ωic
[r]
[r]
[r]
mt = updated occupancy grid(zt , xt , mt−1 )
[r]
[r]
[r]
[r]
Ht = F [p(Θ | xt0 +1:t , Ct0 +1:t , ft0 +1:t , Ht0 )]
[r]
[r]
[r]
[r]
[r]
X̄t = X̄t ∪ hxt0 +1:t , Ct0 +1:t , mt , Ht0 +1:t , ωt i
end for
PR
[r]
[r]
0
St∗0 +1:t = argmaxS [r]
r=1 ωS δ(St +1:t − St0 +1:t )
t0 +1:t
25:
LMt = argmaxLM p(LM | St∗0 +1:t , LMt0 , λ)
26:
for r = 1 to R do
[i]
27:
draw i with probability ∝ ωt
[i]
[i]
[i]
[i]
28:
add hxt0 +1:t , Ct0 +1:t , mt , Ht0 +1:t , LMt0 +1:t i to Xt
29:
end for
30:
return Xt
31: end procedure
as TL = 10. Method (G) performed the proposed scalable
algorithm in three different conditions, i.e., the lag values for
the FLR was set as 1, 10, and 20. Batch-learning methods
(H) and (I) were estimated by Gibbs sampling based on a
weak-limit approximation [16] of the stick-breaking process
(SBP) [17]. The upper limits of the spatial concepts and
position distributions were set to L = 50 and K = 50,
respectively. We set the number of iterations for Gibbs
sampling to G = 100. In method (I), we set the number
of candidates of the word segmentation results for updating
the language model to M = 6, and the number of iterative
estimation procedures to I = 10.
A. Online Learning
We conducted experiments for the online spatial concept acquisition in a real environment. We implemented
SpCoSLAM 2.0 based on the open-source SpCoSLAM1 that
extended the gmapping package implementing the grid-based
1 https://github.com/EmergentSystemLabStudent/
SpCoSLAM
Step
15
30
50
it :
1:
1:
1:
/kyouyuuseki/ (Sharing desk)
/nobasyanyanamae/, /bawakyoo/, /yuseki/
/yusekinikibashita/, /kyoyuseki/, /kyoyusekidayo/
/kibashita/, /ni/, /kyouyuseki/
it :
3:
3:
3:
/ikidomari/ (The end of corridor)
/namaewa/, /enikimashita/, /gonobashiha/
/dayo/, /ikidomari/, /fokoga/
/idomari/, /miriiNgusupesu/, /koko/
it :
4:
6:
6:
/roboqtookiba/ (Robot storage space)
/robotokiba/, /robotokibanya/, /rimasu/
/kochirawaagaro/, /botoki/, /baninarimasu/
/wabotookiba/, /robotookiba/, /kochiraga/
Fig. 3. Top: Learning results of position distributions in a generated map; ellipses denoting the position distributions are drawn on the map at steps 15,
30, and 50. The colors of the ellipses were randomly determined for each index number it = k. Bottom: The examples of the correct word (English) and
estimated words for each position distribution at steps 15, 30, and 50.
FastSLAM 2.0 [12] in the robot operating system (ROS). We
used an open dataset, albert-b-laser-vision, i.e., a rosbag file
containing the odometry, laser range data, and image data.
This dataset was obtained from the Robotics Data Set Repository (Radish) [18]. We prepared Japanese speech data corresponding to the movement of the robot of above-mentioned
dataset because speech data were not initially included. The
total number of taught utterances was N = 50, including
10 types of phrases. The numbers of taught places and
place names were 10 and nine, respectively. The employed
microphone was a SHURE PG27-USB. Julius dictation-kitv4.4 (DNN-HMM decoding) [19] was used as a speech
recognizer. The initial word dictionary contains 115 Japanese
syllables. The unsupervised word segmentation system used
the latticelm [9]. The image feature extractor is implemented
with the deep-learning framework Caffe [20]. We used a
pre-trained CNN model, Places365-ResNet, trained with 365
scene categories of the Places2 Database with 1.8 million
images [21]. The number of particles was R = 30. The
hyperparameters for the online learning were set as follows:
α = 20, γ = 0.1, β = 0.1, χ = 0.1, m0 = [0, 0]T ,
κ0 = 0.001, V0 = diag(2, 2), and ν0 = 3. The abovementioned parameters were set such that all online methods
in comparison were tested under the same conditions. The
hyperparameters for the batch learning were set as follows:
α = 10, γ = 10, β = 0.1, m0 = [0, 0]T , κ0 = 0.001,
V0 = diag(2, 2), and ν0 = 3.
Fig. 3 shows the position distributions in the environmental
maps at steps 15, 30, and 50 in (F). The bottom part of this
figure shows an example of the correct phoneme sequence
of the place name, and the 3-best words estimated by the
probability distribution p(St | it , Θt , LMt ) at step t. Fig. 3
shows how the spatial concepts are acquired during sequential mapping the environment. The details of the online
learning experiment can be viewed in the video attached.
B. Evaluation Metrics
We evaluated the methods of different algorithms according to the following metrics: normalized mutual information
(NMI), estimation accuracy rate (EAR), phoneme accuracy
rate of uttered sentences (PARs), and PAR of acquired words
(PARw). We conducted six learning trials for each algorithm
condition. The details of the evaluation metrics are described
in the following sections.
1) Estimation Accuracy of Spatial Concepts: We compared the matching rate for the estimated indices C1:N of
the spatial concept and the classification results of the correct
answers given by a person. In this experiment, the evaluation
metric adopts the NMI, which is a measure of the similarity
degree between two clustering results. The matching rate for
the estimated indices i1:N of the position distributions is also
evaluated in the same manner.
In addition, we evaluated the estimated number of spatial
concepts L and position distributions K using the EAR. The
EAR was calculated as follows:
EAR = max(1 − | nT − nE |/nT , 0)
T
(14)
E
where n is the correct number and n is the estimated
number. We used the average EAR value of all steps as the
evaluation value.
2) PAR of Uttered Sentences: We compared in this section the accuracy rate for phoneme recognition and word
segmentation for all the recognized sentences. We calculated
the PAR of the uttered sentences (PARs) with the correct
phoneme sequence StP , and a phoneme sequence StR of the
recognition result of each uttered sentence. This experiment
considered the position of a delimiter as a single letter. PAR
was calculated as follows:
PAR = max(1 − LD(StP , StR )/np , 0)
(15)
where LD() is calculated using the Levenshtein distance
between StP and StR . Here, nP the number of phonemes
of the correct phoneme sequence.
3) PAR of Acquired Words: We evaluated in this section
whether a phoneme sequence has learned the properly segmented place names. This experiment assumes a request for
(best)
the best phoneme sequence, St,b , representing the selfposition xt of the robot. We compared the PARw with the
correct place name and a selected word for each teaching
place. PAR is calculated by (15).
0.6
(A) SpCoSLAM
(B) SpCoSLAM with AW
(C) SpCoSLAM with AW+WS
(D) 2.0 (10 FLR)
(E) 2.0 (RS)
(F) 2.0 (10 FLR + RS)
(G1) 2.0 (1 FLR + SBU)
(G2) 2.0 (10 FLR + SBU)
(G3) 2.0 (20 FLR + SBU)
PARw
0.4
0.3
Calculation time [sec.]
0.5
800
0.2
700
600
500
(A) SpCoSLAM
(B) SpCoSLAM with AW
(C) SpCoSLAM with AW+WS
(D) 2.0 (10 FLR)
(E) 2.0 (RS)
(F) 2.0 (10 FLR + RS)
(G1) 2.0 (1 FLR + SBU)
(G2) 2.0 (10 FLR + SBU)
(G3) 2.0 (20 FLR + SBU)
400
300
200
0.1
100
0.0
10
20
30
40
50
0
10
step
Fig. 4.
(best)
(best)
St,b
Fig. 5.
is calculated as follows:
= argmax p(St,b | xt , Θt , LMt ).
30
40
50
step
Changes in the PARw values for each step
The selection of St,b
20
(16)
St,b
In this experiment, we used the self-position xt that was not
included in the training data for the evaluation of PARw. The
more a method achieves accurate recognition of words and
acquisition of spatial concepts, the higher is the PARw value.
We consider this evaluation metric as an overall measure for
evaluating the proposed method.
C. Results and Discussion on Improvement and Scalability
In this section, we discuss the improvement and scalability
of the proposed learning algorithms. Table II lists the averages of the evaluation values calculated using the metrics
NMI, EAR, PARs, and PARw at step 50. Fig. 4 shows the
results of the PARw values for each step. Table III presents
the examples of the word segmentation results of the four
methods. “|” denotes a word segment position. Morpheme
was suitably segmented into Japanese morphemes using
MeCab, which is an off-the-shelf Japanese morphological
analyzer that is widely used for natural language processing.
Although the modified SpCoSLAM (B) and (C) are
theoretically appropriate than the original algorithm, little
differences were found. Time-driven self-localization and
event-driven spatial concept acquisition were solved by the
same particle filter. Therefore, strong influence of the weight
for self-localization is suggested rather than the weight of the
spatial concept. This will be investigated deeper in future
work.
From the results of (D) and (F), EAR, i.e., the accuracy of
the estimated number of clusters, was markedly improved by
introducing the FLR of Ct and it . From the results of (E) and
(F), the PARs and PARw were markedly improved by the resegmentation of the word sequences. These results showed
that the robot can accurately segment the names of the places
and learn the relationship between the places and the words
more precisely. In particular, method (F) that combines the
FLR and RS achieved an overall improvement compared to
the evaluation values of the other online algorithms.
Calculation times for evaluation of scalability
Fig. 5 shows the results of the calculation times between
the online learning algorithms. In batch learning, the calculation time of SpCoA was 1,063.736 s and SpCoA++ was
13,850.873 s. In the previous SpCoSLAM algorithm (A)–
(C) and the improved SpCoSLAM 2.0 algorithm (D)–(F),
the calculation time increased with the number of steps,
i.e., the amount of data increased. However, the scalable
SpCoSLAM 2.0 algorithm (G1)–(G3) could be kept at a
constant calculation time regardless of the increase in the
amount of data. As shown in Table II and Fig. 5, the
evaluation values improved overall when the lag value was
increased. In particular, when the lag value was 20, it showed
relatively high evaluation values approaching the improved
algorithm.
Because a trade-off exists between the fixed-lag size
and accuracy, the algorithm needs to be set appropriately
according to both the computational power embedded in the
robot, and the duration requirements on the actual operation.
In this experiment, we did not evaluate the scalability of
the algorithm on parallel processing. However, we consider
that the proposed algorithm could be executed even faster by
parallelizing the process of particles and using GPU.
V. CONCLUSION
This paper discussed an improved and scalable online
learning to address the problems encountered in the conventional SpCoSLAM algorithm. We proposed an online
learning algorithm, SpCoSLAM 2.0, for spatial concept and
lexical acquisition, to render higher accuracy and scalability. In the experiments, we conducted online learning with
a robot in a novel environment without any pre-existing
lexicon and map. In addition, we evaluated comparatively
the proposed algorithm, the conventional online algorithm,
and batch learning in terms of estimation accuracy and
calculation time. As the results shown, the proposed algorithm demonstrated a higher accuracy than the conventional
algorithm and achieved an accuracy comparable to batch
learning. Moreover, the calculation time of the proposed scalable algorithm becomes constant for each step independently
from the amount of training data.
TABLE II
E VALUATION VALUES FOR COMPARATIVE METHODS
Methods
(A) SpCoSLAM [6]
(B) SpCoSLAM with AW
(C) SpCoSLAM with AW + WS
(D) SpCoSLAM 2.0 (10 FLR–it , Ct )
(E) SpCoSLAM 2.0 (RS)
(F) SpCoSLAM 2.0 (10 FLR–it , Ct + RS)
(G1) SpCoSLAM 2.0 (1 FLR–it , Ct , St + SBU)
(G2) SpCoSLAM 2.0 (10 FLR–it , Ct , St + SBU)
(G3) SpCoSLAM 2.0 (20 FLR–it , Ct , St + SBU)
(H) SpCoA [7] (Batch learning)
(I) SpCoA++ [8] (Batch learning)
Improved
Scalable
X
X
X
X
X
X
X
X
X
NMI Ct
0.666
0.622
0.588
0.622
0.582
0.619
0.591
0.638
0.670
0.611
0.686
NMI it
0.811
0.764
0.730
0.821
0.733
0.801
0.743
0.805
0.852
0.842
0.848
EAR L
0.596
0.585
0.557
0.837
0.638
0.782
0.633
0.589
0.649
0.283
0.767
EAR K
0.770
0.758
0.725
0.915
0.796
0.871
0.792
0.779
0.833
0.519
0.648
PARs
0.524
0.516
0.496
0.533
0.789
0.801
0.648
0.690
0.704
0.708
0.787
PARw
0.154
0.132
0.086
0.157
0.358
0.419
0.158
0.262
0.292
0.140
0.524
TABLE III
E XAMPLES OF WORD SEGMENTATION RESULTS OF UTTERED SENTENCES
English
Morpheme
(A)
(F)
(G3)
(I)
“This place is the sharing desk.”
kochira | ga | kyouyuuseki | ni | nari | masu
a | kochiragagyoyusekiNni | narimasu
kochira | ga | kyouyuseki | ninarimasu
uo | kochi | ra | ga | kyoyuseki | nina | ri | ma | su
ochiraga | kyoyuseki | ninarimasu
“This is the meeting space.”
koko | wa | miitiNgusupeisu | desu
kokowaga | midigisupesudesujoouya
kokowa | miriiNgusupesu | desu
kowa | midigyusu | pesu | desu
gokoga | miidiNgusupesu | desu
We expect this work to contribute greatly to the realization
of long-term spatial language interactions between humans
and robots. As future work, we would like to perform a
long-term online learning experiment of spatial concepts,
in large-scale environments, based on the proposed scalable
algorithm of this paper. In addition, the generative model
of SpCoSLAM is connected with the SLAM and lexical
acquisition via the latent variables related to the spatial
concepts. Therefore, we will also conduct experiments on
loop-closure based on speech signals, and investigate how
spatial concepts can positively affect mapping.
ACKNOWLEDGMENT
The authors thank Cyrill Stachniss for providing the data.
R EFERENCES
[1] A. Pronobis and P. Jensfelt, “Large-scale semantic mapping and
reasoning with heterogeneous modalities,” in Proceedings of ICRA.
IEEE, 2012, pp. 3515–3522.
[2] I. Kostavelis and A. Gasteratos, “Semantic mapping for mobile
robotics tasks: A survey,” Robotics and Autonomous Systems, vol. 66,
pp. 86–103, 2015.
[3] N. Sünderhauf, F. Dayoub, S. McMahon, B. Talbot, R. Schulz,
P. Corke, G. Wyeth, B. Upcroft, and M. Milford, “Place categorization
and semantic mapping on a mobile robot,” in Proceedings of ICRA.
IEEE, 2016, pp. 5729–5736.
[4] C. Landsiedel, V. Rieser, M. Walter, and D. Wollherr, “A review of
spatial reasoning and interaction for real-world robotics,” Advanced
Robotics, vol. 31, no. 5, pp. 222–242, 2017.
[5] S. Thrun, W. Burgard, and D. Fox, Probabilistic Robotics. MIT Press,
2005.
[6] A. Taniguchi, Y. Hagiwara, T. Taniguchi, and T. Inamura, “Online
spatial concept and lexical acquisition with simultaneous localization
and mapping,” in Proceedings of IROS, 2017.
[7] A. Taniguchi, T. Taniguchi, and T. Inamura, “Spatial concept acquisition for a mobile robot that integrates self-localization and unsupervised word discovery from spoken sentences,” IEEE Transactions
on Cognitive and Developmental Systems, vol. 8, no. 4, pp. 285–297,
2016.
[8] ——, “Unsupervised spatial lexical acquisition by updating a language
model with place clues,” Robotics and Autonomous Systems, vol. 99,
pp. 166–180, 2018.
“The printer room is here.”
puriNtaabeya | wa | kochira | desu
ioporiNtabeaakochiragadesuduuryuzuqaqo
puriNtabeya | kochira | desu
puriNpabeya | ta | kochiradesu
poriNpabeya | wakochiradesu
[9] G. Neubig, M. Mimura, and T. Kawahara, “Bayesian learning of
a language model from continuous speech,” IEICE Transactions on
Information and Systems, vol. 95, no. 2, pp. 614–625, 2012.
[10] S. Isobe, A. Taniguchi, Y. Hagiwara, and T. Taniguchi, “Learning
relationships between objects and places by multimodal spatial concept
with bag of objects,” in Proceedings of ICSR, 2017, pp. 115–125.
[11] A. Krizhevsky, I. Sutskever, and G. Hinton, “Imagenet classification
with deep convolutional neural networks,” in Proceedings of NIPS,
2012, pp. 1097–1105.
[12] G. Grisetti, C. Stachniss, and W. Burgard, “Improved techniques for
grid mapping with Rao-Blackwellized particle filters,” IEEE Transactions on Robotics, vol. 23, pp. 34–46, 2007.
[13] K. R. Canini, L. Shi, and T. L. Griffiths, “Online inference of topics
with latent Dirichlet allocation.” in Proceedings of AISTATS, vol. 9,
2009, pp. 65–72.
[14] G. Kitagawa, “Computational aspects of sequential Monte Carlo filter
and smoother,” Annals of the Institute of Statistical Mathematics,
vol. 66, no. 3, pp. 443–471, 2014.
[15] D. Mochihashi, T. Yamada, and N. Ueda, “Bayesian unsupervised
word segmentation with nested Pitman-Yor language modeling,” in
Proceedings of ACL-IJCNLP, 2009, pp. 100–108.
[16] E. B. Fox, E. B. Sudderth, M. I. Jordan, and A. S. Willsky, “A sticky
HDP-HMM with application to speaker diarization,” The Annals of
Applied Statistics, pp. 1020–1056, 2011.
[17] J. Sethuraman, “A constructive definition of Dirichlet priors,” Statistica
Sinica, vol. 4, pp. 639–650, 1994.
[18] A. Howard and N. Roy, “The robotics data set repository (radish),”
2003. [Online]. Available: http://radish.sourceforge.net/
[19] A. Lee and T. Kawahara, “Recent development of open-source speech
recognition engine Julius,” in Proceedings of the APSIPA ASC, 2009,
pp. 131–137.
[20] 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.
[21] B. Zhou, A. Lapedriza, A. Khosla, A. Oliva, and A. Torralba, “Places:
A 10 million image database for scene recognition,” IEEE TPAMI,
2017.
| 2cs.AI
|
Linear Convergence of a Frank-Wolfe Type Algorithm over
Trace-Norm Balls
arXiv:1708.02105v3 [cs.LG] 9 Nov 2017
Zeyuan Allen-Zhu
Microsoft Research, Redmond
[email protected]
Wei Hu
Princeton University
[email protected]
Elad Hazan
Princeton University
[email protected]
Yuanzhi Li
Princeton University
[email protected]
November 8, 2017
Abstract
We propose a rank-k variant of the classical Frank-Wolfe algorithm to solve convex optimization
over a trace-norm ball. Our algorithm replaces the top singular-vector computation (1-SVD) in
Frank-Wolfe with a top-k singular-vector computation (k-SVD), which can be done by repeatedly
applying 1-SVD k times. Alternatively, our algorithm can be viewed as a rank-k restricted version
of projected gradient descent. We show that our algorithm has a linear convergence rate when the
objective function is smooth and strongly convex, and the optimal solution has rank at most k. This
improves the convergence rate and the total time complexity of the Frank-Wolfe method and its
variants.
1
Introduction
Minimizing a convex matrix function over a trace-norm ball, which is: (recall that the trace norm kXk∗
of a matrix X equals the sum of its singular values)
minX∈Rm×n f (X) : kXk∗ ≤ θ ,
(1.1)
is an important optimization problem that serves as a convex surrogate to many low-rank machine learning
tasks, including matrix completion [2, 10, 16], multiclass classification [4], phase retrieval [3], polynomial
neural nets [12], and more. In this paper we assume without loss of generality that θ = 1.
One natural algorithm for Problem (1.1) is projected gradient descent (PGD). In each iteration, PGD
first moves X in the direction of the gradient, and then projects it onto the trace-norm ball. Unfortunately,
computing this projection requires the full singular value decomposition (SVD) of the matrix, which takes
O(mn min{m, n}) time in general. This prevents PGD from being efficiently applied to problems with
large m and n.
Alternatively, one can use projection-free algorithms. As first proposed by Frank and Wolfe [5], one
can select a search direction (which is usually the gradient direction) and perform a linear optimization
over the constraint set in this direction. In the case of Problem (1.1), performing linear optimization over
a trace-norm ball amounts to computing the top (left and right) singular vectors of a matrix, which can
be done much faster than full SVD. Therefore, projection-free algorithms become attractive for convex
minimization over trace-norm balls.
Unfortunately, despite its low per-iteration complexity, the Frank-Wolfe (FW) algorithm suffers from
slower convergence rate compared with PGD. When the objective f (X) is smooth, FW requires O(1/ε)
iterations to convergence to an ε-approximate minimizer, and
√ this 1/ε rate is tight even if the objective
is also strongly convex [6]. In contrast, PGD achieves 1/ ε rate if f (X) is smooth (under Nesterov’s
acceleration [14]), and log(1/ε) rate if f (X) is both smooth and strongly convex.
1
Recently, there were several results to revise the FW method to improve its convergence rate for
strongly-convex
functions. The log(1/ε) rate was obtained when the constraint set is a polyhedron [7, 11],
√
and the 1/ ε rate was obtained when the constraint set is strongly convex [8] or is a spectrahedron [6].
Among these results, the spectrahedron constraint (i.e., for all positive semidefinite matrices X with
Tr(X) = 1) studied by Garber [6] is almost identical to Problem (1.1), but slightly weaker.1 When stating
the result of Garber [6], we assume for simplicity that it also applies to Problem (1.1).
Our Question. In this paper, we propose to study the following general question:
Can we design a “rank-k variant” of Frank-Wolfe to improve the convergence rate?
(That is, in each iteration it computes the top k singular vectors – i.e., k-SVD – of some matrix.)
Our motivation to study the above question can be summarized as follows:
• Since FW computes a 1-SVD and PGD computes a full SVD in each iteration, is there a value
k min{n, m} such that a rank-k variant of FW can achieve the convergence rate of PGD?
• Since computing k-SVD costs roughly the same (sequential) time as “computing 1-SVD for k times”
(see recent work [1, 13]),2 if using a rank-k variant of FW, can the number of iterations be reduced by
a factor more than k? If so, then we can improve the sequential running time of FW.
• k-SVD can be computed in a more distributed manner than 1-SVD. For instance, using block Krylov [13],
one can distribute the computation of k-SVD to k machines, each in charge of independent matrix-vector
multiplications. Therefore, it is beneficial to study a rank-k variant of FW in such settings.
1.1
Our Results
We propose blockFW, a rank-k variant of Frank-Wolfe. Given a convex function f (X) that is β-smooth, in
each iteration t, blockFW performs an update Xt+1 ← Xt + η(Vt − Xt ), where η > 0 is a constant step size
and Vt is a rank-k matrix computed from the k-SVD of (−∇f (Xt ) + βηXt ). If k = min{n, m}, blockFW
can be shown to coincide with PGD, so it can also be viewed as a rank-k restricted version of PGD.
Convergence. Suppose f (X) is also α-strongly convex and suppose the optimal solution X ∗ of
Problem (1.1) has rank k, then we show that blockFW achieves linear convergence: it finds an ε-approximate
β
minimizer within O( α
log 1ε ) iterations, or equivalently, in
kβ
1
log
computations of 1-SVD.
T =O
α
ε
We denote by T the number of 1-SVD computations throughout this paper. In contrast,
TFW = O βε
for Frank-Wolfe
n
o
√
β 1/2 β 1/2
β 1/4 β 3/4
1
TGar = O min βε ,
k ,
for Garber [6].
α
ε
α
ε
σmin (X ∗ )
Above, σmin (X ∗ ) is the minimum non-zero singular value of X ∗ . Note that σmin (X ∗ ) ≤
kX ∗ k∗
rank(X ∗ )
≤ k1 .
We note that TGar is always outperformed by min{T, TFW }: ignoring the log(1/ε) factor, we have
√
β 1/4 β 3/4 1/4
β 1/4 β 3/4
≤ α
k
< α
k, and
• min βε , kβ
α
ε
ε
β kβ
β 1/2 β 1/2 1/2
β 1/2 β 1/2
1
• min ε , α ≤ α
k
< α
ε
ε
σmin (X ∗ ) .
Remark. The low-rank assumption on X ∗ should be reasonable: as we mentioned, in most applications
of Problem (1.1), the ultimate reason for imposing a trace-norm constraint is to ensure that the optimal
1 The the best of our knowledge, given an algorithm that works for spectrahedron, to solve Problem (1.1), one has to
define a function g(Y ) over (n + m) × (n + m) matrices, by setting g(Y ) = f (2Y1:m,m+1:m+n ) [10]. After this transformation,
the function g(Y ) is no longer strongly convex, even if f (X) is strongly convex. In contrast, most algorithms for trace-norm
balls, including FW and our later proposed algorithm, work as well for spectrahedron after minor changes to the analysis.
2 Using block Krylov [13], Lanszos [1], or SVRG [1], at least when k is small, the time complexity of (approximately)
computing the top k singular vectors of a matrix is no more than k times the complexity of (approximately) computing the
top singular vector of the same matrix. We refer interested readers to [1] for details.
2
algorithm
# rank
# iterations
PGD [14]
min{m, n}
κ log(1/ε)
accelerated
PGD [14]
min{m, n}
FrankWolfe [9]
1
Garber [6]
1
blockFW
k
√
κ log(1/ε)
β
ε
κ
κ
1
4
1
2
√
β3
4
k , or
ε
β1
1
2
ε
σmin (X ∗ )
κ log(1/ε)
time complexity per iteration
O mn min{m, n}
O mn min{m, n}
Õ nnz(∇)
Õ nnz(∇) + (m + n)
× min
1/2
k∇k2
ε1/2
1/2
,
k∇k2
(σ1 (∇)−σ2
(∇))1/2
1/2
1/2
k∇k2
k∇k2
× min
,
1/2
1/2
ε
(σ1 (∇)−σ2 (∇))
k · Õ nnz(∇) + k(m + n)κ
(k∇k2 +α)1/2 κ(k∇k2 +α)1/2
× min
,
1/2
1/2
∗
ε
α
σmin (X )
Table 1: Comparison of first-order methods to minimize a β-smooth, α-strongly convex function over the unit-trace
β
norm ball in Rm×n . In the table, k is the rank of X ∗ , κ = α
is the condition number, ∇ = ∇f (Xt ) is the
gradient matrix, nnz(∇) is the complexity to multiply ∇ to a vector, σi (X) is the i-th largest singular
value of X, and σmin (X) is the minimum non-zero singular value of X.
solution is low-rank; otherwise the minimization problem may not be interesting to solve in the first place.
Also, the immediate prior work [6] also assumes X ∗ to have low rank.
k-SVD Complexity. For theoreticians who are concerned about the time complexity of k-SVD, we also
compare it with the 1-SVD complexity of FW and Garber. If one uses LazySVD [1]3 to compute k-SVD in
each iteration of blockFW, then the per-iteration k-SVD complexity can be bounded by
(k∇k2 + α)1/2 κ(k∇k2 + α)1/2
k · Õ nnz(∇) + k(m + n)κ × min
, 1/2
.
(1.2)
ε1/2
α σmin (X ∗ )
β
Above, κ = α
is the condition number of f , ∇ = ∇f (Xt ) is the gradient matrix of the current iteration t,
nnz(∇) is the complexity to multiply ∇ to a vector, σmin (X ∗ ) is the minimum non-zero singular value of
X ∗ , and Õ hides poly-logarithmic factors.
In contrast, if using Lanczos, the 1-SVD complexity for FW and Garber can be bounded as (see [6])
1/2
n k∇k1/2
o
k∇k2
2
Õ nnz(∇) × min
,
.
(1.3)
ε1/2 (σ1 (∇) − σ2 (∇))1/2
Above, σ1 (∇) and σ2 (∇) are the top two singular values of ∇, and the gap σ1 (∇) − σ2 (∇) can be as small
as zero.
We emphasize that our k-SVD complexity (1.2) can be upper bounded by a quantity that only depends
poly-logarithmically on 1/ε. In contrast, the worst-case 1-SVD complexity (1.3) of FW and Garber
depends on ε−1/2 because the gap σ1 − σ2 can be as small as zero. Therefore, if one takes this additional
ε dependency into consideration for the convergence rate, then blockFW has rate polylog(1/ε), but FW
and Garber have rates ε−3/2 and ε−1 respectively. The convergence rates and per-iteration running times
of different algorithms for solving Problem (1.1) are summarized in Table 1.
Practical Implementation. Besides our theoretical results above, we also provide practical suggestions
for implementing blockFW. Roughly speaking, one can automatically select a different “good” rank k for
each iteration. This can be done by iteratively finding the 1st, 2nd, 3rd, etc., top singular vectors of
the underlying matrix, and then stop this process whenever the objective decrease is not worth further
increasing the value k. We discuss the details in Section 6.
2
Preliminaries and Notation
For a positive integer n, we define [n] := {1, 2, . . . , n}. For a matrix A, we denote by kAkF , kAk2 and
kAk∗ respectively the Frobenius norm, the spectral norm, and the trace norm of A. We use h·, ·i to
3 In fact, LazySVD is a general framework that says, with a meaningful theoretical support, one can apply a reasonable
1-SVD algorithm k times in order to compute k-SVD. For simplicity, in this paper, whenever referring to LazySVD, we mean
to apply the Lanczos method k times.
3
denote the (Euclidean) inner products between vectors, or the (trace) inner products between matrices
(i.e., hA, Bi = Tr(AB > )). We denote by σi (A) the i-th largest singular value of a matrix A, and by
σmin (A) the minimum non-zero singular value of A. We use nnz(A) to denote the time complexity of
multiplying matrix A to a vector (which is at most the number of non-zero entries of A). We define the
(unit) trace-norm ball Bm,n in Rm×n as Bm,n := {X ∈ Rm×n : kXk∗ ≤ 1}.
Definition 2.1. For a differentiable convex function f : K → R over a convex set K ⊆ Rm×n , we say
• f is β-smooth if f (Y ) ≤ f (X) + h∇f (X), Y − Xi + β2 kX − Y k2F for all X, Y ∈ K;
• f is α-strongly convex if f (Y ) ≥ f (X) + h∇f (X), Y − Xi + α2 kX − Y k2F for all X, Y ∈ K.
For Problem (1.1), we assume f is differentiable, β-smooth, and α-strongly convex over Bm,n . We
β
denote by κ = α
the condition number of f , and by X ∗ the minimizer of f (X) over the trace-norm ball
Bm,n . The strong convexity of f (X) implies:
Fact 2.2. f (X) − f (X ∗ ) ≥
α
2 kX
− X ∗ k2F for all X ∈ K.
Proof. The minimality of X ∗ implies h∇f (X ∗ ), X − X ∗ i ≥ 0 for all X ∈ K. The fact follows then from
the α-strong convexity of f .
The Frank-Wolfe Algorithm. We now quickly review the Frank-Wolfe algorithm (see Algorithm 1)
and its relation to PGD.
Algorithm 1 Frank-Wolfe
Input: Step sizes {ηt }t≥1 (ηt ∈ [0, 1]), starting point X1 ∈ Bm,n
1: for t = 1, 2, . . . do
2:
Vt ← argminV ∈Bm,n h∇f (Xt ), V i
by finding the top left/right singular vectors ut , vt of −∇f (Xt ), and taking Vt = ut vt> .
3:
Xt+1 ← Xt + ηt (Vt − Xt )
4: end for
Let ht = f (Xt ) − f (X ∗ ) be the approximation error of Xt . The convergence analysis of Algorithm 1
is based on the following relation:
¬
β
ht+1 = f (Xt + ηt (Vt − Xt )) − f (X ∗ ) ≤ ht + ηt h∇f (Xt ), Vt − Xt i + ηt2 kVt − Xt k2F
2
(2.1)
®
β
β
≤ ht + ηt h∇f (Xt ), X ∗ − Xt i + ηt2 kVt − Xt k2F ≤ (1 − ηt )ht + ηt2 kVt − Xt k2F .
2
2
Above, inequality ¬ uses the β-smoothness of f , inequality is due to the choice of Vt in Line 2, and
inequality ® follows from the convexity of f . Based on (2.1), a suitable choice of the step size ηt = Θ(1/t)
gives the convergence rate O(β/ε) for the Frank-Wolfe algorithm.
If f is also α-strongly convex, a linear convergence rate can be achieved if we replace the linear
optimization step (Line 2) in Algorithm 1 with a constrained quadratic minimization:
β
Vt ← argminh∇f (Xt ), V − Xt i + ηt kV − Xt k2F .
(2.2)
2
V ∈Bm,n
In fact, if Vt is defined as above, we have the following relation similar to (2.1):
β
ht+1 ≤ ht + ηt h∇f (Xt ), Vt − Xt i + ηt2 kVt − Xt k2F
2
(2.3)
β
∗
≤ ht + ηt h∇f (Xt ), X − Xt i + ηt2 kX ∗ − Xt k2F ≤ (1 − ηt + κηt2 )ht ,
2
1
where the last inequality follows from Fact 2.2. Given (2.3), we can choose ηt = 2κ
to obtain a linear
convergence rate because ht+1 ≤ (1 − 1/4κ)ht . This is the main idea behind the projected gradient descent
(PGD) method. Unfortunately, optimizing Vt from (2.2) requires a projection operation onto Bm,n , and
this further requires a full singular value decomposition of the matrix ∇f (Xt ) − βηt Xt .
4
3
A Rank-k Variant of Frank-Wolfe
Our main idea comes from the following simple observation. Suppose we choose ηt = η =
iterations, and suppose rank(X ∗ ) ≤ k. Then we can add a low-rank constraint to Vt in (2.2):
Vt ←
argmin
V ∈Bm,n , rank(V )≤k
β
ηkV − Xt k2F .
2
h∇f (Xt ), V − Xt i +
1
2κ
for all
(3.1)
Under this new choice of Vt , it is obvious that the same inequalities in (2.3) remain to hold, and thus the
linear convergence rate of PGD can be preserved. Let us now discuss how to solve (3.1).
3.1
Solving the Low-Rank Quadratic Minimization (3.1)
Although (3.1) is non-convex, we prove that it can be solved efficiently. To achieve this, we first show that
Vt is in the span of the top k singular vectors of βηXt − ∇f (Xt ).
Pk
Lemma 3.1. The minimizer Vt of (3.1) can be written as Vt = i=1 ai ui vi> , where a1 , . . . , ak are
nonnegative scalars, and (ui , vi ) is the pair of the left and right singular vectors of At := βηXt − ∇f (Xt )
corresponding to its i-th largest singular value.
The proof of Lemma 3.1 is given in Appendix B. Now,
Pkowing to Lemma 3.1, we can perform k-SVD on
At to compute {(ui , vi )}i∈[k] , plug the expression Vt = i=1 ai ui vi> into the objective of (3.1), and then
Pk
Pk
search for the optimal values {ai }i∈[k] . The last step is equivalent to minimizing − i=1 σi ai + β2 η i=1 a2i
k
(where σi = u>
i At vi ) over the simplex ∆ := a ∈ R : a1 , . . . , ak ≥ 0, kak1 ≤ 1 , which is the same as
1
projecting the vector βη
(σ1 , . . . , σk ) onto the simplex ∆. It can be easily solved in O(k log k) time (see
for instance the applications in [15]).
3.2
Our Algorithm and Its Convergence
We summarize our algorithm in Algorithm 2 and call it blockFW.
Algorithm 2 blockFW
Input: Rank parameter k, starting point X1 = 0
1
1: η ← 2κ
.
2: for t = 1, 2, . . . do
3:
At ← βηXt − ∇f (Xt )
4:
(u1 , v1 , . . . , uk , vk ) ← k-SVD(At )
(ui , vi ) is the i-th largest pair of left/right singular vectors of At
1
k
5:
a ← argmina∈Rk ,a≥0,kak1 ≤1 ka − βη
σk2
where σ := (u>
i At vi )i=1
Pk
>
6:
Vt ← i=1 ai ui vi
7:
Xt+1 ← Xt + η(Vt − Xt )
8: end for
Since the state-of-the-art algorithms for k-SVD are iterative methods, which in theory can only give
approximate solutions, we now study the convergence of blockFW given approximate k-SVD solvers.
We introduce the following notion of an approximate solution to the low-rank quadratic minimization
problem (3.1).
Definition 3.2. Let gt (V ) = h∇f (Xt ), V − Xt i + β2 ηkV − Xt k2F be the objective function in (3.1), and let
gt∗ = gt (X ∗ ). Given parameters γ ≥ 0 and ε ≥ 0, a feasible solution V to (3.1) is called (γ, ε)-approximate
if it satisfies g(V ) ≤ (1 − γ)gt∗ + ε.
Note that the above multiplicative-additive definition makes sense because gt∗ ≤ 0:
Fact 3.3. If rank(X ∗ ) ≤ k, for our choice of step size η =
− h2t ≤ 0 according to (2.3).
1
2κ ,
we have gt∗ = gt (X ∗ ) ≤ −(1 − κη)ht =
The next theorem gives the linear convergence of blockFW under the above approximate solutions to
5
(3.1). Its proof is simple and uses a variant of (2.3) (see Appendix B).
Theorem 3.4. Suppose rank(X ∗ ) ≤ k and ε > 0. If each Vt computed in blockFW is a ( 12 , 8ε )-approximate
solution to (3.1), then for every t, the error ht = f (Xt ) − f (X ∗ ) satisfies
1 t−1
h1 + 2ε .
ht ≤ 1 − 8κ
As a consequence, it takes O(κ log
h1
ε )
iterations to achieve the target error ht ≤ ε.
Based on Theorem 3.4, the per-iteration running time of blockFW is dominated by the time necessary
to produce a ( 12 , 8ε )-approximate solution Vt to (3.1), which we study in Section 4.
4
Per-Iteration Running Time Analysis
In this section, we study the running time necessary to produce a ( 12 , ε)-approximate solution Vt to (3.1).
In particular, we wish to show a running time that depends only poly-logarithmically on 1/ε. The reason
is that, since we are concerning about the linear convergence rate (i.e., log(1/ε)) in this paper, it is not
meaningful to have a per-iteration complexity that scales polynomially with 1/ε.
Remark 4.1. To the best of our knowledge, the Frank-Wolfe method and Garber’s method [6] have their
worst-case per-iteration complexities scaling polynomially with 1/ε. In theory, this also slows down their
overall performance in terms of the dependency on 1/ε.
4.1
Step 1: The Necessary k-SVD Accuracy
We first show that if the k-SVD in Line 4 of blockFW is solved sufficiently accurate, then Vt obtained in
Line 6 will be a sufficiently good approximate solution to (3.1). For notational simplicity, in this section
we denote Gt := k∇f (Xt )k2 + α, and we let k ∗ = rank(X ∗ ) ≤ k.
Lemma 4.2. Suppose γ ∈ [0, 1] and ε ≥ 0. In each iteration t of blockFW, if the vectors u1 , v1 , . . . , uk , vk
Pk
∗
>
returned by k-SVD in Line 4 satisfy
u>
i At vi ≥ (1 − γ)σi (At ) − ε for all i ∈ [k ], then Vt =
i=1 ai ui vi
6Gt
obtained in Line 6 is
ht + 2 γ, ε -approximate to (3.1).
The proof of Lemma 4.2 is in Appendix C, and is based on our earlier characterization Lemma 3.1.
4.2
Step 2: The Time Complexity of k-SVD
We recall the following complexity statement for k-SVD:
Theorem 4.3 ([1]). The running time to compute the k-SVD of A ∈ Rm×n using LazySVD is4
2
(m+n)
k·nnz(A)+k2 (m+n)
√
√
Õ k·nnz(A)+k
or
Õ
.
γ
gap
>
In
the former case,i we can have ui Avi ≥ (1 − γ)σi (A) for all i ∈ [k]; in the latter case, if gap ∈
σk∗ (A)−σk∗ +1 (A)
∗
0,
for some k ∗ ∈ [k], then we can guarantee u>
i Avi ≥ σi (A) − ε for all i ∈ [k ].
σk∗ (A)
The First Attempt. Recall that we need a ( 12 , ε)-approximate solution to (3.1). Using Lemma 4.2, it
suffices to obtain a (1 − γ)-multiplicative approximation to the k-SVD of At (i.e., u>
i At vi ≥ (1 − γ)σi (At )
1
for all i ∈ [k]), as long as γ ≤ 12Gt /h
.
Therefore,
we
can
directly
apply
the
first running time in
t +4
k·nnz(At )+k2 (m+n)
√
Theorem 4.3: Õ
. However, when ht is very small, this running time can be unbounded.
γ
In that case, we observe that γ = Gεt (independent of ht ) also suffices: since kAt k2 = α2 Xt − ∇f (Xt ) 2 ≤
α
ε
>
>
2 + k∇f (Xt )k2 ≤ Gt , from ui At vi ≥ (1 − ε/Gt )σi (At ) we have ui At vi ≥ σi (At ) − Gt σi (At ) ≥ σi (At ) −
ε
Gt kAt k2 ≥ σi (At ) − ε; then according to Lemma 4.2 we can obtain (0, ε)-approximation to (3.1), which
is stronger than ( 12 , ε)-approximation. We summarize this running time (using γ = Gεt ) in Claim 4.5; the
running time depends polynomially on 1ε .
The Second Attempt. To make our linear convergence rate (i.e., the log(1/ε) rate) meaningful, we
4 The first is known as the gap-free result because it does not depend on the gap between any two singular values. The
second is known as the gap-dependent result, and it requires a k × k full SVD after the k approximate singular vectors are
computed one by one. The Õ notation hides poly-log factors in 1/ε, 1/γ, m, n, and 1/gap.
6
want the k-SVD running time to depend poly-logarithmically on 1/ε. Therefore, when ht is small, we wish
to instead apply the second running time in Theorem 4.3.
Recall that X ∗ has rank k ∗ so σk∗ (X ∗ ) − σk∗ +1 (X ∗ ) = σmin (X ∗ ). We can show that this implies
A := α2 X ∗ − ∇f (X ∗ ) also has a large gap σk∗ (A∗ ) − σk∗ +1 (A∗ ). (See Lemma C.1.) Now, according to
Fact 2.2, when ht is small, Xt and X ∗ are sufficiently close. This means At = α2 Xt − ∇f (Xt ) is also close
to A∗ , and thus has a large gap σk∗ (At ) − σk∗ +1 (At ). Then we can apply the second running time in
Theorem 4.3.
∗
4.2.1
Formal Running Time Statements
Fact 4.4. We can store Xt as a decomposition into at most rank(Xt ) ≤ kt rank-1 components.5 Therefore,
for At = α2 Xt − ∇f (Xt ), we have nnz(At ) ≤ nnz(∇f (Xt )) + (m + n)rank(Xt ) ≤ nnz(∇f (Xt )) + (m + n)kt.
If we always use the first running time in Theorem 4.3, then Fact 4.4 implies:
Claim 4.5. The k-SVD computation in the t-th iteration of blockFW can be implemented in Õ k ·
p
nnz(∇f (Xt )) + k 2 (m + n)t Gt /ε time.
Remark 4.6. As long
p as (m + n)kt ≤ nnz(∇f (Xt )), the k-SVD running time in Claim 4.5
p becomes
Õ k · nnz(∇f (Xt )) Gt /ε , which roughly equals k-times the 1-SVD running time Õ nnz(∇) k∇k2 /ε)
of FW and Garber [6]. Since in practice, it suffices to run blockFW and FW for a few hundred 1-SVD
computations, the relation (m + n)kt ≤ nnz(∇f (Xt )) is often satisfied.
If, as discussed above, we apply the first running time in Theorem 4.3 only for large ht , and apply the
second running time in Theorem 4.3 for small ht , then we obtain the following theorem whose proof is
given in Appendix C.
Theorem 4.7. The k-SVD comuputation
in the t-th iteration of blockFW can be implemented in
√
κ Gt /α
2
Õ k · nnz(∇f (Xt )) + k (m + n)t σmin (X ∗ ) time.
Remark 4.8. Since according to Theorem 3.4 we only need to run blockFW for O(κ log(1/ε)) iterations,
we can plug t = O(κ log(1/ε)) into Claim 4.5 and Theorem 4.7, and obtain the running time presented in
(1.2). The per-iteration running time of blockFW depends poly-logarithmically on 1/ε. In contrast, the
per-iteration running times of Garber [6] and FW depend polynomially on 1/ε, making their total running
times even worse in terms of dependency on 1/ε.
5
Maintaining Low-Rank Iterates
One of the main reasons to impose trace-norm constraints is to produce low-rank solutions. However, the
rank of iterate Xt in our algorithm blockFW can be as large as kt, which is much larger than k, the rank
of the optimal solution X ∗ . In this section, we show that by adding a simple modification to blockFW, we
can make sure the rank of Xt is O(kκ log κ) in all iterations t, without hurting the convergence rate much.
We modify blockFW as follows. Whenever t − 1 is a multiple of S = d8κ(log κ + 1)e, we compute (note
that this is the same as setting η = 1 in (3.1))
β
Wt ←
argmin
h∇f (Xt ), W − Xt i + kW − Xt k2F ,
2
W ∈Bm,n , rank(W )≤k
and let the next iterate Xt+1 be Wt . In all other iterations the algorithm is unchanged. After this change,
the function value f (Xt+1 ) may be greater than f (Xt ), but can be bounded as follows:
Lemma 5.1. Suppose rank(X ∗ ) ≤ k. Then we have f (Wt ) − f (X ∗ ) ≤ κht .
Proof. We have the following relation similar to (2.3):
β
kWt − Xt k2F
2
β
≤ ht + h∇f (Xt ), X ∗ − Xt i + kX ∗ − Xt k2F
2
f (Wt ) − f (X ∗ ) ≤ ht + h∇f (Xt ), Wt − Xt i +
5
In Section 5, we show how to ensure that rank(Xt ) is always O(kκ log κ), a quantity independent of t.
7
≤ ht − ht +
β 2
· ht
2 α
= κht .
1 S
1 8κ(log κ+1)
h1 + 2ε ≤ e−(log κ+1) h1 + 2ε =
From Theorem 3.4 we know that hS+1 ≤ (1− 8κ
) h1 + 2ε ≤ (1− 8κ
)
1
κε
Therefore, after setting XS+2 = WS+1 , we still have hS+2 ≤ e h1 + 2 (according to Lemma 5.1).
Continuing this analysis (letting the κε here be the “new ε”), we know
that this modified version of
blockFW converges to an ε-approximate minimizer in O κ log κ · log hε1 iterations.
Remark 5.2. Since in each iteration the rank of Xt is increased by at most k, if we do the modified step
every S = O(κ log κ) iterations, we have that throughout the algorithm, rank(Xt ) is never more than
O(kκ log κ). Furthermore we can always store Xt using O(kκ log κ) vectors, instead of storing all the
singular vectors obtained in previous iterations.
1
eκ h1 +ε/2.
6
Preliminary Empirical Evaluation
We conclude this paper with some preliminary experiments to test the performance of blockFW. We first
recall two machine learning tasks that fall into Problem (1.1).
Matrix Completion. Suppose there is an unknown matrix M ∈ Rm×n close to low-rank, and we
observe a subset Ω of its entries – that is, we observe Mi,j for every (i, j) ∈ Ω. (Think of Mi,j as user i’s
rating of movie j.) One can recover M by solving the following convex program:
P
minX∈Rm×n 21 (i,j)∈Ω (Xi,j − Mi,j )2 | kXk∗ ≤ θ .
(6.1)
Although Problem (6.1) is not strongly convex, our experiments show the effectiveness of blockFW on this
problem.
Polynomial Neural Networks. Polynomial networks are neural networks with quadratic activation
function σ(a) = a2 . Livni et al. [12] showed that such networks can express any function computed by a
Turing machine, similar to networks with ReLU or sigmoid activations. Following [12], we consider the
class of 2-layer polynomial networks with inputs from Rd and k hidden neurons:
n
o
Pk
V
Pk = x 7→ j=1 aj (wj> x)2 ∀j ∈ [k], wj ∈ Rd , kwj k2 = 1 a ∈ Rk .
Pk
If we write A = i=1 aj wj wj> , we have the following equivalent formulation:
Pk = x 7→ x> Ax A ∈ Rd×d , rank(A) ≤ k .
Therefore, if replace the hard rank constraint with trace norm kAk∗ ≤ θ, the task of empirical risk
minimization (ERM) given training data {(x1 , y1 ), . . . , (xN , yN )} ⊂ Rd × R can be formulated as6
PN
2
kAk∗ ≤ θ .
(6.2)
minA∈Rd×d 21 i=1 (x>
i Axi − yi )
P
N
2
Since f (A) = 12 i=1 (x>
i Axi − yi ) is convex in A, the above problem falls into Problem (1.1). Again,
this objective f (A) might not be strongly convex, but we still perform experiments on it.
6.1
Preliminary Evaluation 1: Matrix Completion on Synthetic Data
We consider the following synthetic experiment for matrix completion. We generate a random rank-10
matrix in dimension 1000 × 1000, plus some small noise. We include each entry into Ω with probability
1/2. We scale M to kM k∗ = 10000, so we set θ = 10000 in (6.1).
We compare blockFW with FW and Garber [6]. When implementing the three algorithms, we use
exact line search. For Garber’s algorithm, we tune its parameter ηt = ct with different constant values c,
and then exactly search for the optimum η̃t . When implementing blockFW, we use k = 10 and η = 0.2.
We use the Matlab built-in solver for 1-SVD and k-SVD.
In Figure 1(a), we compare the numbers of 1-SVD computations for the three algorithms. The plot
confirms that it suffices to apply a rank-k variant FW in order to achieve linear convergence.
6.2
Auto Selection of k
In practice, it is often unrealistic to know k in advance. Although one can simultaneously try k =
1, 2, 4, 8, . . . and output the best possible solution, this can be unpleasant to work with. We propose the
6 We
consider square loss for simplicity. It can be any loss function `(x>
i Axi , yi ) convex in its first argument.
8
8
FW
Garber
This paper
7
6
4
7
FW
Garber
This paper
6
2
1
0
5
Log(error)
3
4
3
2
1.5
1
0.5
0
2
-1
-2
3
2.5
4
Log(error)
Log(error)
5
FW
Garber
This paper
3.5
-0.5
1
0
20
40
60
# 1-SVD computations
80
100
0
50
100
150
# 1-SVD computations
(a) matrix completion on synthetic data (b) matrix completion
on MovieLens1M, θ = 10000
200
-1
0
100
200
300
400
500
# 1-SVD computations
(c) polynomial neural network
on Mnist, θ = 0.03
Figure 1: Partial experimental results. The full 6 plots for MovieLens and 3 plots for Mnist are included in
Figure 2 and Figure 3 in Appendix A.
following modification to blockFW which automatically chooses k.
In each iteration t, we first run 1-SVD and compute the objective decrease, denoted by d1 ≥ 0. Now,
given any approximate k-SVD decomposition of the matrix At = βηXt − ∇f (Xt ), we can compute its
(k + 1)-SVD using one additional 1-SVD computation according to the LazySVD framework [1]. We
compute the new objective decrease dk+1 . We stop this process and move to the next iteration t + 1
k+1
< dkk . In other words, we stop whenever it “appears” not worth further increasing k. We
whenever dk+1
count this iteration t as using k + 1 computations of 1-SVD.
All the experiments on real-life datasets are performed using this above auto-k process.
6.3
Preliminary Evaluation 2: Matrix Completion on MovieLens
We study the same experiment in Garber [6], the matrix completion Problem (6.1) on datasets MovieLens100K (m = 943, n = 1862 and |Ω| = 105 ) and MovieLens1M (m = 6040, n = 3952 and |Ω| ≈ 106 ).
In the second dataset, following [6], we further subsample Ω so it contains about half of the original entries.
For each dataset, we run FW, Garber, and blockFW with three different choices of θ.7 We present the six
plots side-by-side in Figure 2 in Appendix A.
We observe that when θ is large, there is no significant advantage for using blockFW. This is because
the rank of the optimal solution X ∗ is also high for large θ. In contrast, when θ is small (so X ∗ is of low
rank), as demonstrated for instance by Figure 1(b), it is indeed beneficial to apply blockFW.
6.4
Preliminary Evaluation 3: Polynomial Neural Network on Mnist
We use the 2-layer neural network Problem (6.2) to train a binary classifier on the Mnist dataset of
handwritten digits, where the goal is to distinguish images of digit “0” from images of other digits. The
training set contains N = 60000 examples each of dimension d = 28 × 28 = 784. We set yi = 1 if that
example belongs to digit “0” and yi = 0 otherwise. We divide the original grey levels by 256 so xi ∈ [0, 1]d .
We again try three different values of θ, and compare FW, Garber, and blockFW.8 We present the three
plots side-by-side in Figure 3 in Appendix A.
The performance of our algorithm is comparable to FW and Garber for large θ, but as demonstrated
for instance by Figure 1(c), when θ is small so rank(X ∗ ) is small, it is beneficial to use blockFW.
7
Conclusion
In this paper, we develop a rank-k variant of Frank-Wolfe for Problem (1.1) and show that: (1) it converges
in log(1/ε) rate for smooth and strongly convex functions, and (2) its per-iteration complexity scales with
7 We perform exact line search for all algorithms. For Garber [6], we tune the best η = c and exactly search for the
t
t
optimal η̃t . For blockFW, we let k be chosen automatically and choose η = 0.01 for all the six experiments.
8 We perform exact line search for all algorithms. For Garber [6], we tune the best η = c and exactly search for the
t
t
optimal η̃t . For blockFW, we let k be chosen automatically and choose η = 0.0005 for all the three experiments.
9
polylog(1/ε). Preliminary experiments suggest that the value k can also be automatically selected, and
our algorithm outperforms FW and Garber [6] when X ∗ is of relatively smaller rank.
We hope more rank-k variants of Frank-Wolfe can be developed in the future.
Acknowledgments
Elad Hazan is supported by NSF grant 1523815 and a Google research award.
References
[1] Zeyuan Allen-Zhu and Yuanzhi Li. LazySVD: Even faster SVD decomposition yet without agonizing
pain. In NIPS, pages 974–982, 2016.
[2] Emmanuel Candes and Benjamin Recht. Exact matrix completion via convex optimization. Communications of the ACM, 55(6):111–119, 2012.
[3] Emmanuel J Candes, Yonina C Eldar, Thomas Strohmer, and Vladislav Voroninski. Phase retrieval
via matrix completion. SIAM review, 57(2):225–251, 2015.
[4] Miroslav Dudik, Zaid Harchaoui, and Jérôme Malick. Lifted coordinate descent for learning with
trace-norm regularization. In AISTATS, pages 327–336, 2012.
[5] Marguerite Frank and Philip Wolfe. An algorithm for quadratic programming. Naval research logistics
quarterly, 3(1-2):95–110, 1956.
[6] Dan Garber. Faster projection-free convex optimization over the spectrahedron. In NIPS, pages
874–882, 2016.
[7] Dan Garber and Elad Hazan. A linearly convergent conditional gradient algorithm with applications
to online and stochastic optimization. arXiv preprint arXiv:1301.4666, 2013.
[8] Dan Garber and Elad Hazan. Faster rates for the frank-wolfe method over strongly-convex sets. In
ICML, pages 541–549, 2015.
[9] Elad Hazan. Sparse approximate solutions to semidefinite programs. In Latin American Symposium
on Theoretical Informatics, pages 306–316. Springer, 2008.
[10] Martin Jaggi and Marek Sulovský. A simple algorithm for nuclear norm regularized problems. In
ICML, pages 471–478, 2010.
[11] Simon Lacoste-Julien and Martin Jaggi. An affine invariant linear convergence analysis for frank-wolfe
algorithms. arXiv preprint arXiv:1312.7864, 2013.
[12] Roi Livni, Shai Shalev-Shwartz, and Ohad Shamir. On the computational efficiency of training neural
networks. In NIPS, pages 855–863, 2014.
[13] Cameron Musco and Christopher Musco. Randomized block krylov methods for stronger and faster
approximate singular value decomposition. In NIPS, pages 1396–1404, 2015.
[14] Yurii Nesterov. Introductory Lectures on Convex Programming Volume: A Basic course, volume I.
Kluwer Academic Publishers, 2004.
[15] Yurii Nesterov. Smooth minimization of non-smooth functions.
103(1):127–152, December 2005.
Mathematical Programming,
[16] Shai Shalev-Shwartz, Alon Gonen, and Ohad Shamir. Large-scale convex minimization with a
low-rank constraint. arXiv preprint arXiv:1106.1622, 2011.
10
Appendix
A
Additional Plots
6
6
FW
Garber
This paper
5.5
5
5.5
FW
Garber
This paper
5.5
4
3.5
3
Log(error)
5
Log(error)
4.5
Log(error)
6
FW
Garber
This paper
4.5
5
4.5
4
2.5
4
3.5
2
1.5
0
50
100
150
200
3
250
0
50
(a) MovieLens100K, θ = 3000
7
FW
Garber
This paper
200
250
3.5
300
0
50
100
150
200
# 1-SVD computations
(b) MovieLens100K, θ = 5000
(c) MovieLens100K, θ = 10000
7
7
FW
Garber
This paper
FW
Garber
This paper
6.5
6
4
Log(error)
6
5
5.5
5
5.5
5
3
4.5
2
1
150
6.5
Log(error)
Log(error)
6
100
# 1-SVD computations
# 1-SVD computations
4.5
4
0
50
100
150
3.5
200
# 1-SVD computations
0
50
100
150
200
250
4
300
0
50
# 1-SVD computations
(d) MovieLens1M, θ = 10000
100
150
200
250
300
# 1-SVD computations
(e) MovieLens1M, θ = 15000
(f) MovieLens1M, θ = 30000
Figure 2: Log objective error vs. the number of 1-SVD computations, for matrix completion.
4
4
FW
Garber
This paper
3.5
3
3.5
FW
Garber
This paper
3.5
FW
Garber
This paper
3
3
2
1.5
1
2.5
Log(error)
Log(error)
Log(error)
2.5
2
1.5
2.5
2
0.5
1
1.5
0
0.5
-0.5
-1
0
100
200
300
# 1-SVD computations
(a) Mnist, θ = 0.03
400
500
0
0
100
200
300
# 1-SVD computations
(b) Mnist, θ = 0.05
400
500
1
0
100
200
300
400
# 1-SVD computations
(c) Mnist, θ = 0.1
Figure 3: Log objective error vs. the number of 1-SVD computations, for polynomial neural network.
11
500
B
Missing Proofs for Section 3
The proof of Lemma 3.1 relies on the following folklore lemma.
Lemma B.1. Let A ∈ Rm×n and r = min{m, n}. Let g : Rr → R be a twice-differentiable convex
function. For any two sets of orthonormal vectors {u1 , . . . , ur } ⊂ Rm and {v1 , . . . , vr } ⊂ Rn , there exist
a permutation π over [r] and ξ1 , . . . , ξr ∈ {−1, 1} such that
>
>
g(u>
1 Av1 , u2 Av2 , . . . , ur Avr ) ≤ g(ξ1 σπ(1) (A), ξ2 σπ(2) (A), . . . , ξr σπ(r) (A)) .
We first prove Lemma 3.1 using Lemma B.1, and then give the proof of Lemma B.1.
Pk
Lemma 3.1 (restated). The minimizer Vt of (3.1) can be written as Vt = i=1 ai ui vi> , where a1 , . . . , ak
are nonnegative reals, and (ui , vi ) is the pair of the left and right singular vectors of At := βηXt − ∇f (Xt )
corresponding to its i-th largest singular value.
Proof. The quadratic objective in (3.1) can be written as h∇f (Xt ) − βηXt , V i + β2 ηkV k2F − h∇f (Xt ), Xt i +
β
2
2 ηkXt kF , where the last two terms do not depend on V . Hence, (3.1) is equivalent to
β
Vt ←
argmin
−hAt , V i + ηkV k2F .
(B.1)
2
V ∈Bm,n , rank(V )≤k
Pk
Since we have rank(V ) ≤ k, we can write its SVD as V = i=1 ai ui vi> , where {u1 , . . . , uk } ⊂ Rm and
{v1 , . . . , vk } ⊂ Rn are two sets of orthonormal vectors. To prove Lemma 3.1, it suffices to show for every
fixed set of singular values {ai }ki=1 where a1 ≥ · · · ≥ ak ≥ 0, the optimal choices of ui ’s and vi ’s must
coincide with the top k left and right singular vectors of At .
Pk
a1 , . . . , E
ak are fixed, kV k2F = i=1 a2i is also fixed, so (B.1) becomes maximizing hAt , V i =
D When
Pk
Pk
At , i=1 ai ui vi> = i=1 ai u>
i At vi . Now, applying Lemma B.1 (whose proof is provided later) with
Pk
Pk
Pk
A = At and g(x1 , . . . , xr ) = i=1 ai xi , we get i=1 ai u>
i At v i ≤
i=1 ai ξi σπ(i) (At ) for some permutation
π over [r] and some ξi ’s in {−1, 1}. Then by the non-negativity of singular values and the rearrangement
Pk
Pk
Pk
inequality we have i=1 ai ξi σπ(i) (At ) ≤ i=1 ai σπ(i) (At ) ≤ i=1 ai σi (At ).
Pk
Pk
In sum, we have shown hAt , V i = i=1 ai u>
i At v i ≤
i=1 ai σi (At ) for any ui ’s and vi ’s. It is easy to
see that the equality can be attained if u>
i At vi = σi (At ) for all i ∈ [k], or equivalently, if ui and vi are the
left and right singular vectors of At corresponding to its i-th largest singular value.
In order to prove Lemma B.1 (which we believe is a folklore result), we first prove its special case
where m = n.
Lemma B.2. Let A ∈ Rn×n and g : Rn → R be a twice-differentiable convex function. For any
two orthonormal bases {u1 , . . . , un } and {v1 , . . . , vn } of Rn , there exists a permutation π over [n] and
ξ1 , . . . , ξn ∈ {−1, 1} such that
>
>
g(u>
1 Av1 , u2 Av2 , . . . , un Avn ) ≤ g(ξ1 σπ(1) (A), ξ2 σπ(2) (A), . . . , ξn σπ(n) (A)) .
00
Proof. We denote by gi0 (x) the i-th partial derivative of g at x ∈ Rn , and by gi,j
(x) the (i, j)-th second-order
00
00
00
partial derivative. WLOG we assume that g is strictly convex, which implies that gi,i
(x)+gj,j
(x) > 2|gi,j
(x)|
holds for all i, j and x. In fact, if g is not strictly convex, we can add εkxk2 to g(x) to make it strictly
convex, which will add small perturbations to both sides of the desired inequality. Then we can let ε → 0
to make the perturbations arbitrarily small, and the desired inequality follows.
>
We now fix a pair of orthonormal bases {u1 , . . . , un } and {v1 , . . . , vn } that maximize g(u>
1 Av1 , u2 Av2 ,
>
n
. . . , un Avn ) over all orthonormal bases
of R , and consider a pair of indices i 6= j. For ϕ ∈ R, define
therocos ϕ − sin ϕ
u (ϕ)
tation matrix Rϕ =
. We apply the rotation Rϕ to vectors ui and uj , i.e., we let i
=
sin ϕ cos ϕ
uj (ϕ)
u
(cos ϕ)ui − (sin ϕ)uj
v (ϕ)
v
(cos ϕ)vi − (sin ϕ)vj
Rϕ i =
. Similarly define i
= Rϕ i =
. Note
uj
(sin ϕ)ui + (cos ϕ)uj
vj (ϕ)
vj
(sin ϕ)vi + (cos ϕ)vj
that if we replace ui and uj by ui (ϕ) and uj (ϕ), {u1 , . . . , ui (ϕ), . . . , uj (ϕ), . . . , un } is still an orthonormal
basis. Similarly, {v1 , . . . , vi (ϕ), . . . , vj (ϕ), . . . , vn } is also an orthonormal basis.
12
>
>
>
n
Let x(ϕ) = (u>
1 Av1 , . . . , ui (ϕ) Avi (ϕ), . . . , uj (ϕ) Avj (ϕ), . . . , un Avn ) ∈ R and consider the following function h defined on R:
>
>
>
h(ϕ) := g(x(ϕ)) = g(u>
1 Av1 , . . . , ui (ϕ) Avi (ϕ), . . . , uj (ϕ) Avj (ϕ), . . . , un Avn ) .
By the optimality of {ui } and {vi }, we know that h(ϕ) achieves its maximum at ϕ = 0. Since h is
twice-differentiable, this means h0 (0) = 0 and h00 (0) ≤ 0. We can directly calculate h0 (0) and h00 (0):
>
h0 (0) = gi0 − gj0 · u>
,
i Avj + uj Avi
>
2
>
00
00
00
h00 (0) =2 gi0 − gj0 u>
ui Avj + u>
,
j Avj − ui Avi + gi,i + gj,j − 2gi,j
j Avi
where all the partial derivatives of g are at point x(0).
>
0
0
0
Assume that u>
i Avj + uj Avi 6= 0. Then from h (0) = 0 we know that gi − gj = 0, which implies 0 ≥
>
2
00
00
00
00
00
00
>
> 0), a contradiction.
h (0) = 0 + gi,i + gj,j − 2gi,j ui Avj + uj Avi > 0 (recall that gi,i + gj,j − 2gi,j
>
>
Therefore we must have ui Avj + uj Avi = 0.
u
v
Next, we apply the rotation Rϕ on i and the rotation R−ϕ on i instead. Repeating the same
uj
vj
>
>
>
analysis as above, we can obtain u>
i Avj − uj Avi = 0. Combining this with ui Avj + uj Avi = 0, we
know that u>
6 j. Since {u1 , . . . , un } is an orthonormal basis and Avj is
i Avj = 0. This holds for all i =
orthogonal to ui for every i =
6 j, we must have Avj = λj uj (λj ∈ R) for all j. Hence ui ’s and vi ’s are (left
and right) singular vectors of A (up to sign flips), and ui and vi correspond to the same singular value.
This completes the proof of Lemma B.2.
00
Now we prove Lemma B.1.
Proof of Lemma B.1. The case where m = n is proved in Lemma B.2. Now
WLOG we assume m < n.
A
e
We pad (n − m) × n zeros to the bottom of A to make it n × n: A =
, and we also pad each ui
0
u
with (n − m) zeros: u
ei = i ∈ Rn . We extend {e
u1 , . . . , u
em } and {v1 , . . . , vm } to two orthonormal bases
0
{e
u1 , . . . , u
en } and {v1 , . . . , vn } of Rn .
Consider a function h(x1 , . . . , xn ) := g(x1 , . . . , xm ) defined on Rn . Using Lemma B.2, we have
>
>
e
e
g(u>
u>
e>
u>
e>
1 Av1 , u2 Av2 , . . . , um Avm ) = g(e
1 Av1 , . . . , u
m Avm ) = h(e
1 Av1 , . . . , u
n Avn )
e . . . , ξn σπ(n) (A))
e = g(ξ1 σπ(1) (A),
e . . . , ξm σπ(m) (A))
e
≤ h(ξ1 σπ(1) (A),
(B.2)
e
for some permutation π over [n] and ξi ’s in {−1, 1}. Note
( that the n singular values of A include all the
e = σi (A), i ∈ [m]
m singular values of A and (n − m) zeros, i.e., σi (A)
. Thus the remaining
0,
i ∈ {m + 1, . . . , n}
thing to prove is that we can choose the permutation π such that (B.2) still holds and that π(1), . . . , π(m)
is a permutation of [m]. Suppose that π(1), . . . , π(m) is not a permutation of [m]. Then there exist i ∈ [m]
and j ∈ {m + 1, . . . , n} such that π(i) > m and π(j) ≤ m. Now, if we swap π(i) and π(j) to obtain a new
permutation π 0 over [n], by the convexity of g we have
e . . . , ξi σπ(i) (A),
e . . . , ξm σπ(m) (A))
e = g(ξ1 σπ(1) (A),
e . . . , 0, . . . , ξm σπ(m) (A))
e
g(ξ1 σπ(1) (A),
1
e . . . , σπ(j) (A),
e . . . , ξm σπ(m) (A))
e + g(ξ1 σπ(1) (A),
e . . . , −σπ(j) (A),
e . . . , ξm σπ(m) (A))
e
≤
g(ξ1 σπ(1) (A),
2
e . . . , ξi0 σπ(j) (A),
e . . . , ξm σπ(m) (A))
e
≤ 0 max g(ξ1 σπ(1) (A),
ξi ∈{−1,1}
=
max
ξi0 ∈{−1,1}
e . . . , ξi0 σπ0 (i) (A),
e . . . , ξm σπ0 (m) (A)).
e
g(ξ1 σπ0 (1) (A),
Repeatedly doing this procedure, we can eventually make σ(1), . . . , σ(m) a permutation of [m] without
decreasing the right hand side of (B.2). (The values of ξi ’s may change.) This completes the proof of
Lemma B.1.
13
Theorem 3.4 (restated). Suppose rank(X ∗ ) ≤ k and ε > 0. If each Vt computed in blockFW is a
( 12 , 8ε )-approximate solution to (3.1), then for every t, the error ht = f (Xt ) − f (X ∗ ) satisfies
1 t−1
h1 + 2ε .
ht ≤ 1 − 8κ
As a consequence, it takes O(κ log
h1
ε )
iterations to achieve the target error ht ≤ ε.
Proof. The proof only requires a simple modification to our previous analysis (2.3). Using that Vt is a
( 12 , 8ε )-approximate solution to (3.1) we have
β 2
η kVt − Xt k2F = ht + ηgt (Vt )
2
ε
1
∗
ht +
.
≤ ht + η(gt /2 + ε/8) ≤ ht + η(−ht /4 + ε/8) = 1 −
8κ
16κ
Repeatedly applying the above inequality for t = 1, 2, . . ., we get
"
t−1
t−2 #
1
ε
1
1
ht ≤ 1 −
h1 +
1+ 1−
+ ··· + 1 −
8κ
16κ
8κ
8κ
t−1
t−1
1
1
ε
ε
1
·
= 1−
h1 +
h1 + .
≤ 1−
8κ
16κ 1 − (1 − 1/(8κ))
8κ
2
ht+1 ≤ ht + ηh∇f (Xt ), Vt − Xt i +
C
Missing Proofs for Section 4
Recall that in Section 4 we let Gt = k∇f (Xt )k2 + α and k ∗ = rank(X ∗ ) ≤ k.
Lemma 4.2 (restated). Suppose γ ∈ [0, 1] and ε ≥ 0. In each iteration t of blockFW, if the vectors
∗
u1 , v1 , . . . , uk , vk returned by k-SVD in Line 4 satisfy u>
i At vi ≥ (1 − γ)σi (At ) − ε for all i ∈ [k ], then
Pk
6G
t
Vt = i=1 ai ui vi> obtained in Line 6 is
ht + 2 γ, ε -approximate to (3.1).
Proof. We rewrite the objective gt (V ) in (3.1) as gt (V ) = −h∇At , V i + β2 ηkV k2F − h∇f (Xt ), Xt i +
β
β
β
2
2
2
2 ηkXt kF = pt (V ) + st , where pt (V ) = −h∇At , V i + 2 ηkV kF and st = −h∇f (Xt ), Xt i + 2 ηkXt kF . Note
that st does not depend on V , and that we can upper bound st as
β
α
st ≤ k∇f (Xt )k2 kXt k∗ + ηkXt k2F ≤ k∇f (Xt )k2 · 1 + · 12 ≤ Gt ,
2
4
α
where we have used η = 2β
. Since the rank of X ∗ is k ∗ ≤ k, let us define a rank k ∗ version of the minimum
as
β
gt0 =
min
h∇f (Xt ), V − Xt i + ηkV − Xt k2F .
2
V ∈Bm,n , rank(V )≤k∗
0
∗
∗
We know that gt ≤ gt (X ) = gt . Moreover, According to Lemma 3.1 and the discussion thereafter, we
know that
( k∗
)
k∗
X
β X 2
0
gt = min −
σi (At )ai + η
a + st ,
(C.1)
a∈∆k∗
2 i=1 i
i=1
Pk∗
∗
where ∆k∗ = {a ∈ Rk : a1 , . . . , ak∗ ≥ 0, i=1 ai ≤ 1}. Denote by a∗ ∈ ∆k∗ the minimizer in (C.1), i.e.,
we have
k∗
k∗
X
β X ∗ 2
0
∗
gt = −
σi (At )ai + η
(a ) + st .
(C.2)
2 i=1 i
i=1
Now our algorithm uses approximate singular vectors ui ’s and vi ’s which satisfy σi = u>
i At v i ≥
(1 − γ)σi (At ) − ε for all i ∈ [k ∗ ]. Then the Vt produced by the algorithm should satisfy for ∆k = {a ∈
Pk
Rk : a1 , . . . , ak ≥ 0, i=1 ai ≤ 1}:
( k
)
k
X
β X 2
gt (Vt ) = min −
σi ai + η
a + st
a∈∆k
2 i=1 i
i=1
14
(
)
k∗
β X 2
≤ min −
σi ai + η
a + st
a∈∆k∗
2 i=1 i
i=1
( k∗
)
k∗
X
β X 2
a + st
≤ min −
((1 − γ)σi (At ) − ε)ai + η
a∈∆k∗
2 i=1 i
i=1
(
)
k∗
k∗
X
β X 2
≤ min −(1 − γ)
a + st + ε .
σi (At )ai + η
a∈∆k∗
2 i=1 i
i=1
∗
k
X
Since (1 − γ)a∗ ∈ ∆k∗ , we can choose a = (1 − γ)a∗ on the right hand side and obtain
∗
gt (Vt ) ≤ −(1 − γ)
= (1 −
≤ (1 −
∗
k
X
γ)a∗i
σi (At )(1 −
i=1
γ)2 (gt0 − st ) + st + ε
γ)2 (gt∗ − st ) + st + ε
k
β X
+ η
((1 − γ)a∗i )2 + st + ε
2 i=1
(C.3)
.
Above, the only equality is due to (C.2). Finally, from st ≤ Gt and gt∗ ≤ − h2t ≤ 0 (see Fact 3.3), we have
t ∗
st ≤ − 3G
ht gt . Then it follows from (C.3) that
3Gt
3Gt
3Gt
2
∗
2
gt (Vt ) ≤ (1 − γ) 1 +
−
gt + ε = 1 − (2γ − γ ) 1 +
gt∗ + ε
ht
ht
ht
3Gt
≤ 1 − 2γ 1 +
gt∗ + ε ,
ht
6Gt
By definition, we know that Vt is
ht + 2 γ, ε -approximate to (3.1).
The following lemma is used to prove Theorem 4.7.
Lemma C.1. The matrix A∗ :=
α ∗
2X
− ∇f (X ∗ ) satisfies σk∗ (A∗ ) − σk∗ +1 (A∗ ) ≥
α
∗
2 σmin (X ).
Proof. Since h∇f (X ∗ ), X − X ∗ i ≥ 0 for all X ∈ Bm,n , we have
h∇f (X ∗ ), X ∗ i = min h∇f (X ∗ ), Xi = −σ1 (∇f (X ∗ )) .
(C.4)
X∈Bm,n
Let the SVD of X ∗ be X ∗ =
Pk∗
i=1
σi (X ∗ )ui vi> . We first show that for all i ∈ [k ∗ ], we must have
h∇f (X ∗ ), ui vi> i = h∇f (X ∗ ), X ∗ i .
Assume (C.5) is false, then there exists i ∈ [k ] such that h∇f (X
∗
0
X =
X ∗ − 21 σi (X ∗ )ui vi>
.
1− 12 σi (X ∗ )
We have
h∇f (X ∗ ), X 0 i =
<
Since kX 0 k∗ =
proved (C.5).
kX ∗ k∗ − 12 σi (X ∗ )
1− 12 σi (X ∗ )
∗
), ui vi> i
(C.5)
> h∇f (X ), X i. Consider
∗
∗
h∇f (X ∗ ), X ∗ i − 12 σi (X ∗ )h∇f (X ∗ ), ui vi> i
1 − 12 σi (X ∗ )
h∇f (X ∗ ), X ∗ i − 12 σi (X ∗ )h∇f (X ∗ ), X ∗ i
1 − 21 σi (X ∗ )
= h∇f (X ∗ ), X ∗ i .
≤ 1, we have X 0 ∈ Bm,n and this contradicts (C.4). Thus, we have
From (C.4) and (C.5) we know that for each i ∈ [k ∗ ] we have h−∇f (X ∗ ), ui vi> i = σ1 (∇f (X ∗ )) =
σ1 (−∇f (X ∗ )). This implies that the largest k ∗ singular values of −∇f (X ∗ ) are all equal to the same
value, and all pairs (ui , vi ) for i ∈ [k ∗ ] are left and right singular vectors corresponding to this sinPk∗
gular value. Therefore we can write the SVD of −∇f (X ∗ ) as −∇f (X ∗ ) = σ1 (∇f (X ∗ )) i=1 ui vi> +
15
Pr
i=k∗ +1
σi (∇f (X ∗ ))ui vi> , where r = min{m, n}. It follows that
∗
k
r
X
X
α
α
A = X ∗ − ∇f (X ∗ ) =
σi (X ∗ ) + σ1 (∇f (X ∗ )) ui vi> +
σi (∇f (X ∗ ))ui vi> ,
2
2
∗
i=1
∗
i=k +1
which is the SVD of A∗ . Therefore, we have
(
α
σi (X ∗ ) + σ1 (∇f (X ∗ )), i = 1, 2, . . . , k ∗ ,
∗
σi (A ) = 2
σi (∇f (X ∗ )),
i = k ∗ + 1, . . . , r .
which implies σk∗ (A∗ ) − σk∗ +1 (A∗ ) =
α
∗
∗
2 σk (X )
+ σ1 (∇f (X ∗ )) − σk∗ +1 (∇f (X ∗ )) ≥
α
∗
∗
2 σk (X ).
Theorem 4.7 (restated). The k-SVD comuputation
in the t-th iteration of blockFW can be implemented
√
κ Gt /α
2
in Õ k · nnz(∇f (Xt )) + k (m + n)t σmin (X ∗ ) time.
Proof of Theorem 4.7. From Theorem 3.4 and Lemma 4.2, we only need to show that the stated running
ε
1
∗
time is enough for LazySVD to ensure u>
. We
i At vi ≥ (1 − γ)σi (At ) − 8 for all i ∈ [k ], where γ = 12Gt
ht
consider two cases.
α3 (σmin (X ∗ ))2
. From Fact 2.2
162β 2
α
Then, At = 2 Xt − ∇f (Xt ) and
• Case 1: ht ≤
σmin (X ∗ )
.
9κ
we know that kXt − X ∗ kF ≤
q
2
α ht
≤
q
2
α
·
+4
α3 (σmin (X ∗ ))2
162β 2
=
A =
− ∇f (X ) satisfy
α
(Xt − X ∗ ) − (∇f (Xt ) − ∇f (X ∗ ))
kAt − A∗ kF =
2
F
α
∗
∗
≤ kXt − X kF + k∇f (Xt ) − ∇f (X )kF
2
α
3β
≤ kXt − X ∗ kF + βkXt − X ∗ kF ≤
kXt − X ∗ kF
2
2
α
3β σmin (X ∗ )
·
= σmin (X ∗ ) .
≤
2
9κ
6
By Weyl’s inequality we know that |σi (At ) − σi (A∗ )| ≤ kAt − A∗ k2 ≤ kAt − A∗ kF ≤ α6 σmin (X ∗ ).
Since we have σk∗ (A∗ )− σk∗ +1 (A∗ ) ≥ α2 σmin (X ∗ )by Lemma C.1, we have σk∗ (At ) − σk∗ +1 (At ) ≥
σk∗ (A∗ ) − α6 σmin (X ∗ ) − σk∗ +1 (A∗ ) + α6 σmin (X ∗ ) = σk∗ (A∗ )−σk∗ +1 (A∗ )− α3 σmin (X ∗ ) ≥ α6 σmin (X ∗ ).
2
(m+n)
We can now apply the second running time of Theorem 4.3: Õ k·nnz(At√)+k
, where gap =
gap
σk∗ (At )−σk∗ +1 (At )
;
σk∗ (At )
Gt
=
Õ
∗
ασmin (X )
.
6Gt
α ∗
2X
∗
∗
in such running time LazySVD can guarantee u>
i At vi ≥ σi (At ) − ε for all i ∈ [k ].
Note that σk∗ (At ) ≤ kAt k2 ≤
∗
α
6 σmin (X )
∗
α
2 kXt k2
+ k∇f (Xt )k2 ≤ Gt . which implies gap =
σk∗ (At )−σk∗ +1 (At )
σk∗ (At )
Thus the running time is
s
!
k · nnz(At ) + k 2 (m + n)
G
t
= Õ (k · nnz(At ) + k 2 (m + n))
.
√
gap
ασmin (X ∗ )
≥
α3 (σmin (X ∗ ))2
. Lemma 4.2 implies that it suffices to ensure u>
i At vi ≥ (1 − γ)σi (At ) for
162β 2
3
(X ∗ ))2
α3 (σmin (X ∗ ))2
ht
1
∗
all i ∈ [k ], where γ = 12Gt
= Ω min{1, Gt } = Ω(min{1,
}) = Ω( α (σmin
) – the
β 2 Gt
β 2 Gt
+4
ht
last step here is because we have α3 (σmin (X ∗ ))2 ≤ β 2 · Gt · 12 = β 2 Gt .
• Case 2: ht >
We apply the first running time of Theorem 4.3:
s
!
k · nnz(At ) + k 2 (m + n)
β 2 Gt
2
Õ
= Õ (k · nnz(At ) + k (m + n))
√
γ
α3 (σmin (X ∗ ))2
!
p
κ
G
/α
t
= Õ (k · nnz(At ) + k 2 (m + n))
.
σmin (X ∗ )
16
Note that the running time for Case 2 is always no smaller than the running time for Case 1 (ignoring
logarithmic factors). Combining the two cases and plugging the bound of nnz(At ) from Fact 4.4, we have
the desired running time statement.
17
| 8cs.DS
|
Microsimulation of the Port of Dover
Chris Roadknight, Uwe Aickelin and Galina Sherman
Intelligent Modelling & Analysis Research Group (IMA)
School of Computer Science
The University of Nottingham
Jubilee Campus
Wollaton Road
Nottingham NG8 1BB
{cmr,uxa}@cs.nott.ac.uk
[email protected]
Abstract— Modelling and simulating the traffic of heavily used but secure environments such as seaports and airports is of
increasing importance. Errors made when simulating these environments can have long standing economic, social and
environmental implications. This paper discusses issues and problems that may arise when designing a simulation strategy. Data
for the Port is presented, methods for lightweight vehicle assessment that can be used to calibrate and validate simulations are also
discussed along with a diagnosis of overcalibration issues. We show that decisions about where the intelligence lies in a system has
important repercussions for the reliability of system statistics. Finally, conclusions are drawn about how microsimulations can be
moved forward as a robust planning tool for the 21st century.
Keywords-Simulation, Transport, Operations Research, Agents
I.
INTRODUCTION
Traffic microsimulation software is becoming increasingly complex, parameterized and configurable. Regardless of how
graphically realistic the end product may appear, the core statistics generated by the simulation still needs to be validated and
verified. Events and statistics that show up when simulations are tested must also appear in real life and vice versa. Real world
validation of simulation results can be an expensive, time consuming, subjective and erroneous process and deciding exactly
how much validation to commission is usually an imprecise art. Existing methods for micro validation such as number plate
recognition and manual sampling are expensive and error prone. Weighing up the cost/reward ratio of validation is an
important but non-trivial process.
Traffic Microsimulations use a discrete event [17] approach to the movement of vehicles over time where the behavior of a
system is represented as a chronological sequence of events. Each event occurs at a unique instant in time, with each new
instance of the system viewed as new state. They combine this with some degree of agent based behavior where elements
within the simulation have a set of parameters and policies that they use to come to decisions. Agent based approaches are
successfully used in traffic and transport management [7]. However, despite their suitability research in this area is not mature
enough and moreover “some problem areas seem under-studied, e.g., the applicability of agent technology to strategic
decision-making within transportation logistics” [7].
Microscopic traffic simulation models have unique characteristics because of their representation of interaction between
drivers, vehicles, and roads. The increasing availability of powerful desktop computers has allowed sophisticated computer
software to be used to model the behaviour of individual vehicles and their drivers in real time. Microsimulation can be applied
to any scenario involving complex vehicle interactions and has been used to model roads, rail, air and sea ports [1, 2]. If
validation is not properly performed, a traffic simulation model may not provide accurate results and shouldn’t be used to
make important decisions with financial, environmental and social impacts.
Microsimulation breaks a simulation down into the smallest sensible connected components. In the case of the simulation of a
traffic scenario that would be vehicles and the smallest sensible stations (i.e. Toll booths, roundabouts, junctions, stop signs).
Each micro-component needs to be accurately modeled but it is also important to correctly define dependencies and flows. It
has been shown that questionable simulation predictions can result from a lack of dependencies that result from independently
microsimulating elements of a larger simulation [14], this brings into question how best to validate a simulation made up of
large numbers of subcomponents and how to ensure the simulation does not contain small but significant errors.
Many scientific search and optimization approaches have analysed the subject of overtraining and the resulting lack of
generalization. For instance, neural networks usually have a stopping condition which when reached signals the end of
training, beyond this point the representation model continues to improve on the training subset of instances but decays when
tested on an unseen dataset [15]. A similar situation occurs in statistics when a statistical model starts to describe random error
or noise instead of the underlying relationship, here it is called overfitting [8]. Overfitting or overtraining is more likely to
occur when a model is unnecessarily complex, such as having too many parameters relative to the number of observations.
While less well researched, similar situations may arise in simulations where a simulation is constructed to such an accuracy as
to completely mimic the situation used as an example. It is easier to create overcalibration errors using modern, componantised
simulation software where each individual element can be highly configured to be representative of the isolated sub-system
without requiring any system wide validity.
The research in this paper involves simulations and real worth data from the Port of Dover. It was chosen for this research as it
is the most important trading route between the UK and mainland Europe, has an intricate and multilevel layout (Figure 1) and
has a substantial amount of existing data and simulations. Over the past 20 years, the number of road haulage vehicles (RHVs)
using the Port of Dover has more than doubled to over 2.3 million [4]. Looking ahead over the next 30 years, both the Port and
UK Government have forecast substantial growth in RHV freight traffic. Approximately 3 million tourist vehicles also pass
through the ferry port annually making it a key European and global tourist gateway.
This paper sets out to identify the performance and characteristics of a microsimulation approach to closed system vehicle
simulation with particular reference to the stability and reproducibility of the simulations. The next section of this paper
outlines existing data, statistics and graphs for the Port of Dover, section three discusses the simulation software package
VISSIM, section four introduces a novel validation procedure which is tested at the Port of Dover, Section 5 discusses an
extension of the validated simulation and Section 6 offers some conclusions.
Figure 1: Plan of the Port of Dover
II.
DOVER EXISTING DATA
Looking at the RHV traffic for the 24 hour cycle over a full year it is apparent that systematic flow variations occur, Figure
2 highlights some key facts about this flow for 2009. For instance the maximum RHV flow is approximately four times the
minimum flow at between one and four vehicles per minute. The lowest flow was between 2:00 to 2:15am and the highest
flow between 3:15 to 3:30pm. These measurements were taken at the weighbridge, here all RHVs are weighed, timestamped
and the driver side of the vehicle noted. This showed that 1,194,973 RHVs exited the UK via the port in 2009, of these
960,878 were left hand drive. The arrival statistics at the weighstations is related to the arrival at the port in general but
modified at peak periods due to the first bottleneck at the port, the Customs check, where passports are inspected and also
by the queueing at the weighbridges themselves. This initial check has the effect of smoothing the flow to the weighbridge
and ticketing kiosks because at peak times queues build up, this effectively reduces the bursty-ness of the traffic flow. This
doesn’t change the overall numbers going through the port, just the arrival dynamics.
Figure 2: Aggregated RHV flows for 15 minute time windows for the whole of 2009
One way to measure actual arrival rates at the port is to use CCTV camera images to capture individual arrival of vehicles
(Figure 3). The flow at these points can be automated or manually assessed. An example of the arrival process at the first
bottleneck (passport check) is shown in Figure 4. When these values are plotted as a cumulative curve (Figure 5) the slope of
the curve is a good indicator of arrival rates. We can also see the arrivals at one of the tourist check-in kiosks, Figure 6 shows
these in 10 second bins.
Figure 3: CCTV of vehicles arriving at the port
Figure 4: Sample of interarrival times of vehicles at the Port of Dover
6000
Cumulative Arrival Time (s)
5000
Off Peak
4000
3000
2000
Peak Time
1000
0
0
100
200
300
400
500
600
700
Arrival
Figure 5: Cumulative arrival times at peak and off-peak times
Figure 6: Interarrival times at tourist check-in in 10 second bins for two days. Note: 10 interarrival times of greater
than 30 minutes not displayed on graph
III.
SIMULATION
Traffic simulations of transport networks traditionally use a discrete event, cellular automata style approach. Examples of this
include TRANSIMS [9], PARAMICS [10], CORSIM [11] and more recently VISSIM [12]. Figure 7 shows an example of
what a VISSIM simulation of the Port may look like. VISSIM [3] is a leading microscopic simulation program for multi-modal
traffic flow modelling. It has a high level of vehicle behaviour detail that can be used to simulate urban and highway traffic,
including pedestrians, cyclists and motorised vehicles. It is a highly parameterised design system that allows a lot of flexibility.
Figure 7: VISSIM simulation of the Port of Dover
VISSIM models provide detailed estimates of evolving network conditions by modelling time-varying demand patterns and
individual drivers' detailed behavioural decisions [3]. Several model inputs (such as origin flows) and parameters (carfollowing and lane-changing coefficients) must be specified before these simulation tools can be applied, and their values must
be determined so that the simulation output accurately replicates the reality reflected in traffic measurements.
There are several significant choke points around the port, places where queues appear and significant delays can arise, namely
the passport checking area, the RHV weighbridge and the ticketing booths. Delays can also be introduced with additional
security checks to a randomly selected percentage of vehicles. There are five weighbridges that all RHVs must stop at, RHVs
are guided into the left two lanes coming off the Eastern Docks roundabout feeding into the three left most customs channels as
to not impede the flow of other vehicles into the port. The wait time at the weighbridge is modeled as a normal distribution
with a mean of twenty seconds and standard deviation of two seconds.
VISSIM allows the specification of an initial random number seed, this allows for the same simulation to be repeatedly
stressed with a different sequence of random numbers but also allows direct comparisons of different scenarios using the same
random numbers. Variability between runs with different seeds is a good metric for how robust the system is. Large differences
in run statistics when using different random numbers suggests one of two things.
Some kind of chaotic data/environment interaction.
Some kind of illogical and pathological fault in the simulation design.
Each section of road, link, junction etc has to be accurately modelled. The simulation might develop an inbuilt fault whereby
a small design aspect that appears (on some levels) to be sensible produces considerable variation in validation statistics just
by modifying the random number seed. For instance, an integral component of traffic simulations is the decisions made by
drivers as they navigate the desired road sections. Lane selection, overtaking, acceleration, deceleration, follow gaps are all
examples of driver behaviour parameters. The port of Dover has many lane selection points, and the number possible of lanes
changes repeatedly. By monitoring the lane usage we can see the desired occupancy rates of lanes but configuring the system
to correctly reflect this is non-trivial. For instance, we know the occupancy of the five weighbridges over the whole of 2009
was for bays one to five were 22.3%, 25.4%, 22.8%, 15.6% and 13.9% respectively.
One way to enforce this ratio is to use a probabilistic, ‘roulette wheel’ style lane selection policy. VISSIM, along with most
simulation toolkits, offers methods to specify probabilistic routing whereby a defined percentage of vehicles are sent down
unique routes. This is a piecewise technique that can be reapplied at various locations around a simulation. While these
methods are attractive from a calibration perspective as exact representations of existing statistics can be ensured, the process
is an unrealistic one as it assumes that drivers make probabilistic decisions at precise locations. So in this case when a vehicle
arrives at a point prior to the weighbridges it is allocated one of the lanes based on the respective probabilities. It turns out
that this method leads to significant variations in trip times depending on the initial random number seed (Figure 8). One of
the benefits of graphical microsimulation is that the 2D and 3D simulations help the researcher to visualise a new scheme and
its potential benefits but also to highlight unrealistic behaviour. Figure 8 shows the congestion at the decision point for two
different runs. Using probabilistic routing to enforce correct routing percentages is a clear case of overcalibration affecting
simulation brittleness.
Figure 8: VISSIM simulations at identical times, with identical traffic flows but different random number seeds
showing how this can effect congestion if probabilistic routing is used
These runs have identical and realistic inbound traffic flow rates that have been constructed based on observations of flows at
peak rates, yet considerable difference is behaviour. The flows were generated by recording arrival at the port using CCTV
camera footage and constructing an arrival process based on two minute segments. Each two minute segment produced
exactly the number of vehicles required (Figure 9), vehicles enter the link during that two minute period according to a
Poisson distribution, if the defined traffic volume exceeds the link capacity the vehicles are stacked outside the network until
space is available again but this was not required in the Dover simulation. Figure 8 shows how one busy 90 minute period
was represented. Figure 10a shows the queue lengths at the weighbridge for the two simulation runs where the only
difference is the random number seed, differences in queue lengths of up to 140 meters are apparent. Figure 10b shows the
associated trip time differences between these two runs with average trip times for all vehicles varying by 100% between the
two simulations during some points of the simulation. The probabilistic approach has a lack of flexibility where drivers will
stick to their random number allocated lane even if it is congested. We found much more repeatable results could be gained
by allowing VISSIMs inbuilt driver behaviour features to select the best lane. Even though the desired percentage
occupancies were not enforced, similar weighbridge ratios were generated as simulated drivers avoided the congested centre
lanes at peak times and effect of random number seeds was much less.
30
Vehicles per interval
25
20
15
10
5
0
1
3
5
7
9 11 13 15 17 19 21 23 25 27 29 31 33 35 37 39 41 43 45
2 Minute Intervals
Figure 9: Typical arrival process in two minute intervals
Figure 10a: Queue lengths for identical flows but different random number seeds
Figure 10b: Effect of random seed on trip time
IV.
DECISION MAKING IN SIMULATIONS
All simulations require decisions to be made, frequently and at many locations. For vehicle simulations these decisions
relate to the behaviour of drivers and their vehicles in response to environmental stimulus and their own goals. An agent
based solution would migrate these decisions to each individual driver and vehicle. This approach is sensible and realistic
but extremely difficult to manage and configure, there are also computation issues when decisions are not aggregated.
Decisions can be made on a higher level and most modern simulation software systems allow for methods such as
conditional, probabilistic and deterministic routing. The benefit of aggregated high level decision making is that calibration
and configuration is much easier to manage and enforce, for instance if we know that 30% of vehicles take one route and
70% take another route then a simple random or round robin selection procedure would ensure a near exact reflection of this
in the simulation. Achieving precise behaviour statistics using a low level agent approach is much more difficult.
The Port of Dover is an ideal example of the difficult trade-offs for where the decision making processes should be located.
Firstly there is the driver specific decisions that affect how their vehicle behaves, how much space do they leave to the
vehicle in front, when do they overtake, how hard do they brake/accelerate etc. There are also lane choice decisions that are
made, how quickly these decisions are made, where they are made, how aggressive etc. In this section we contrast two
methods of decision making and how these methods differ in terms of accuracy, particularly under different loads.
The weighbridge is a key point for the flow of vehicles through the Port of Dover. Here every RHV must come to a halt to
be weighed, and every driver chooses from five lanes. The lane selection data for the whole of 2009 was made available so
we had excellent statistics for how which drivers chose which lanes at which times. So when making the decision on which
lane to choose we assessed two options:
1.
Probabilistic Routing. At a set point prior to the weighbridge a random number is generated and a biased roulette wheel
selection approach [16] is used to tell the driver which lane to head for. So if we knew 10% of drivers used lane four we
could generate a random number between zero and one and if the number was greater than zero and less than or equal
to 0.1 then the driver would select lane four. The ratios of vehicles to lanes is different at different flow rates, so a
coarse approach would be to use the average lane section ratios for the whole one year period for all flow rates (Nonflow specific probabilistic routing) and a more precise method would use flow specific lane occupancy ratios (Flow
specific probabilistic routing)
2.
Agent based Routing [7]. Each driver has a set of configuration parameters that decides when they should overtake,
change lanes, slow down etc. These configurations, along with the road layout and volume of traffic would be allowed
to dictate the flow of vehicles through the weighbridge
We use five different flow rates in these experiments, these are based on real world data extracted from video images of
vehicles entering the Port and the timestamped weighbridge data. The rates range from 30 to 800 vehicle an hour. All graphs
show average lane selection or trip time statistics captured over 21 replications and error bars based on standard deviation or
quartiles where appropriate. Figure 11 shows an overhead plan of the weighbridges with lane numbering, with lane five
being nearest farthest from the sea and lane four the nearest.
Figure 11: Weighbridge Plan with Numbering
Figure 12a shows the lane occupancy observed for the agent based simulation approach. At low flow levels lanes one and
two make up nearly 100% of lane usage, this is a result of the low occupancy of lanes meaning the drivers seldom need to
change lanes, this is shown in more detail in Figure 12b. As the flow increases all five lanes are used more, as queues for
lanes one and two develop. Figure 13 show what was actually observed at the weighbridge, again use of lanes one and two
predominate but to a less marked degree. Figure 14 shows when the agent based simulation error for each of the five lanes is
aggregated and compared with the error for two flow based approaches. It shows that at very low flows the agent based
approach has a significant error but at very high flow rates is more accurate than the probabilistic routing using average
flows. While in this instance it could be surmised that the performance of the simulation at high flow is more important this
will not always be the case for all simulations. Correcting the agent based approach would involve introducing more
complex decision making for the drivers and capturing the reasons that they choose the lanes they choose. Correcting the
flow based decision making process is just a case of using flow specific statistics when parameterising the probabilistic
route selection
Figure 12a: Lane occupancy ratios at various arrival rates for the agent based simulation approach
Weighbridge Lane Number
Weighbridge Lane Number
Figure 12b: Box and Whisker plots of the weighbridge lane selections made in the agent-based simulation for a low flow of
60 vehicle per hour (left) and a high flow of 800 vehicles per hour (right).
Figure 13a: Real World lane occupancy rates at various arrival rates
Figure 13b: Simulated lane occupancy rates at various arrival rates using probabilistic routing
Figure 14: Errors in lane occupancy prediction for various simulation methods
While the correct reflection of real world occupancy is an important goal for this simulation, it is also important to check
other metrics of evaluation, one such metric is the trip time or how long it takes for RHVs to travel between two points of
the Port. While the last section shows how closely the weighbridge occupancy statistics can be engineered, this experiment
will show how well the different simulation methods reflect the real point to point trip times, regardless of how well the
weighbridge selection is performed. For this experiment we use a much shorter trip distance that starts after the customs
check and continues until shortly after the weighbridge, this captures the delay caused by congestion at and approaching the
weighbridge and discounts the effects of the ticketing and customs stop points. Figure 15 shows how both the agent based
and probabilistic routing approaches are accurate reflectors of trip time at low and medium flow rates but also shows how
the probabilistic routing approach gives much higher predicted trip times than those actually observed at high flow rates.
This is due to an inflexibility of the probabilistic routing approach whereby once the desired lane has been selected the
driver has no ability to over-rule this dictate and may cause considerable and unnecessary congestion by pursuing the
require lane. Table 1 gives an overview of how much error is associated which each simulation method when measured
again trip time and lane occupancy. The two largest errors (X) are the agent based approach performing poorly at low flow
rate lane selection and the probabilistic routing approach performing poorly when trip times are measured at very busy
times. The performance of both of these scenarios could be improved by introducing more driver intelligence but not
without a significant increase in complexity and simulation run times.
Figure 15: Trip times through the weighbridge complex at different arrival rates for real data and simulation methods
Arrival Rates
Agent (Lane
Selection)
Probabilistic
Routing (Lane
Selection)
Agent (Trip
Time)
Prob Routing
(Trip Time)
Low (30-90
vehicle/hour)
X
Medium (300
vehicles/hour)
✓
High (600
vehicles/hour)
✓✓
V. High (800
vehicles/hour)
✓✓
✓✓✓
✓✓✓
✓✓✓
✓✓✓
✓✓✓✓
✓✓✓✓
✓✓✓
✓✓✓
✓✓✓✓
✓✓✓✓
✓
X
Table 1: Performance of two simulation methods, using two metrics and four arrival rates
V.
VALIDATION
Validating microscopic traffic simulation models incorporates several challenges because of the incompleteness and
rareness of validation data. Validation data is also usually measured in aggregate forms and not at the level of the individual
vehicle. The cost-performance relationship of validation is an important function that should be well understood and used
when deciding what extent any validation should be taken too [13]. Most of the model validation research uses average link
measurements of traffic characteristics [6]. However, these approaches have limitations including possible non-obvious
inconsistency between the observed and simulation estimated variables. Here we decided to use passive Bluetooth [5]
monitoring to sample a vehicles location by uniquely identifying it using their Bluetooth signal id.
Two sampling locations were chosen near the beginning and end of the drivers’ trip through the Port of Dover, from entry to
the Port to waiting to embark on the ferry. These sampling locations are labelled 1 and 2 on the diagram of the Port (Figure
1). The first sampling location used was the security check area close to the beginning of the route through the Port of
Dover. Here some of the drivers have their passports checked and all drivers slow down to find out if they are to be checked.
Bluetooth can be detected at ranges up to 100 meters [3] without sophisticated equipment but it is heavily dependent on
conditions. This area is a good location for Bluetooth monitoring because it is under cover (allowing some degree of signal
reflection), drivers are driving slowly, the capture point is close to the vehicles and drivers usually have their windows open
for passport checks. The only downside to capturing here is that there are 6 possible lanes for drivers to take, but most take
the three central ones. During the capture period of four hours approximately 1200 vehicles passed through this area and 796
Bluetooth devices were registered. Some vehicles may have more than one Bluetooth device so the exact percentage of
vehicles sampled is not trivial to ascertain but discarding all but one of multiple, time synchronised Bluetooth signals largely
removes replication.
The second sampling location used was after the ticketing area when sixteen lanes funnel into three. This area captured
much fewer Bluetooth signals than location 1 (125 vs 796) for a similar period, even though there are fewer lanes. The
reasons for this included much fewer open windows, faster speeds, and a propensity to use the middle lane of three. When
the unique Bluetooth addresses were compared we found a total of 104 examples of a Bluetooth device being found at both
locations. Figure 16 shows the trip times registered for these 104 Bluetooth emitters that were captured at the start and finish
of the trip through the Port of Dover. Visually it is apparent that there are general trends of increasing and decreasing trip
time associated with queues and congestion interspersed with longer or shorter individual trips that could be explained by
events such as security checks, vehicles parking to retrieve items from their boot or motor bikes/fast moving vehicles
negotiating the circuit quicker.
Figure 16: Trip times captured by Bluetooth monitoring
Arrival of anonymous vehicles at various locations was also captured using logging at the entry camera, the weighbridge
and the ticketing kiosks. These captures were all partial, the video capture missed some vehicles because of obscuring, the
weighbridge only captured RHVs and the ticketing was only available for a subset of the ferry operators. Using these real
flows the current VISSIM simulation was stressed with a traffic flow that was an accurate representation of the real flow.
Video cameras were also used to capture vehicle details at the same locations at the Bluetooth capture, enabling the
derivation of trip times that can be compared with the Bluetooth derived trip times. Figure 17 shows how the Simulation
compared with the Bluetooth capture and the visual inspection capture. Over 100 vehicles were simultaneously sampled for
each example. The Bluetooth and visual capture are very closely correlated. The correlation between the simulation results
and the capture results is less strong. Looking at the statistics for the three datasets (Table 2) it appears that the simulation
produces slightly lower trip times but not to a statistically significant degree.
Trip time data (seconds)
Simulation
Bluetooth
Camera
Mean
319
358
343
Median
291
301
306
Standard Deviation
110
181
137
729
1250
860
Max
Table 2: Means and Standard Deviations for Trip time measurements
30
% of measurements per bin
25
Bluetooth
20
Camera
Simulation
15
10
5
0
1
2
3
4
5
6
7
8
9
10 11 12 13
Time (1 minute bins)
14
15
16
17
18
19
20
21
Figure 17: Probability Density Function of Trip times as measured by Bluetooth and Visual Sampling
The more complex a microsimulation is the more validation options there are. For the Dover simulation we could validate the
trip times at the start and finish of the Dover route. It is important that the simulation shows similar trip times and trip statistics
to real observations as shown in Figure 17 and Table 2. With complex microsimulations it is also important to validate
individual components of the simulation. It is important to do this to ensure that correct gross statistics are not being achieved
by aggregation of incorrect modelling of sub components. For instance, in some circumstances the summation of two
incorrectly modelled sections may give accurate overall trip times but when the systems is stressed in different ways there are
no assurances that the overall statistics will remain accurate.
VI.
CONCLUSIONS
A procedure for validation of microscopic traffic simulation models is tested, and its application to the simulation toolbox
VISSIM is demonstrated. The validation efforts are performed at the microscopic level using Bluetooth monitoring and visual
counting of vehicles. Analysis of variance of the simulation results versus the field data shows Bluetooth capture to be a useful
approach for micro-monitoring of vehicles but care must be taken when choosing a collection site. While Bluetooth may not
remain popular forever it is entirely plausible that some kind of detectible wireless protocol will continue to be available for the
foreseeable future. The process detailed here may be considered a step towards the development of a systematic approach for
validation of traffic simulation models. We have taken this work forward by using the refined model to assess the performance
of modifications to the Port. We also plan to investigate how accurately a Microsimulation model can capture rare but
important events, the key to this will be assuring that anomalous behaviour is not due to simulation construction.
When building a microsimulation great care must be taken to ensure each component is as accurate as possible as small errors
in design can lead to disproportionately large errors. This is especially the case if actual behaviour is replaced with
probabilistic approaches, while these can ensure representative statistics they can also introduce gross errors when coupled
with strict lane discipline and can also be an example of overcalibration. Unlike areas such as Neural Networks and Statistics,
overfitting or overcalibration of a simulation is much less well understood and therefore methods to avoid it are less well
known.
There is a requirement in an agent based simulation to have appropriately intelligent agents that best reflect actual behaviour
without introducing significant overheads in terms of complexity and hardware requirements. Having agents with
representative behaviour reduces the need to overcalibrate the system by using popular methods such as probabilistic routing.
This is evidenced when probabilistic routing is replaced by allowing drivers to make natural decisions on lane selection
which results in much more consistent simulation performance at high flow levels.
Alternative methods for validation of simulation results are shown, with Bluetooth capture proving to be a viable, low
maintenance method of trip time sampling. A more comprehensive Bluetooth capture may generate enough trip times to test
the existing simulation at a more conclusive statistical level. Bluetooth is only one of many wireless protocols and may not be a
long term standard for vehicle wireless comes but the approach should be equally effective with future wireless communication
techniques.
ACKNOWLEDGMENTS
We would like to thank the Port of Dover, particularly Daniel Gillett and Paul Simmons, for their valuable contributions,
for supplying existing VISSIM models and for allowing access to locations within the Port. This research has been funded by
the EPSRC.
REFERENCES
[1]
Charypar, D., K.W. Axhausen and K. Nagel (2007a) An event-driven parallel queue-based microsimulation for large scale traffic scenarios, paper
presented at the 11th World Conference on Transportation Research, Berkeley, June 2007.
[2] Li, Ting, Eric van Heck, Peter Vervest, Jasper Voskuilen, Freek Hofker, and Fred Jansma. 2006. Passenger Travel Behavior Model in Railway
Network Simulation. In Proceedings of the 2006 Winter Simulation Conference, eds. L. F. Perrone, F. P. Wieland, J. Liu, and B. G. Lawson, 13801387. CD ISBN 1-4244-0501-7, IEEE Catalog Number 06CH37826
[3] M. Fellendorf and P. Vortisch. "Validation of the microscopic traffic flow model VISSIM in different real-world situations". 80th Meeting of the
Transportation Research Board. Washington, D.C., January, 2001.
[4] http://www.doverport.co.uk/?page=PortDevelopment
[5] Bluetooth SIG, Specification of the Bluetooth System, http://www.bluetooth.com, 2003.
[6] Do H. Nam, Donald R. Drew, Automatic measurement of traffic variables for intelligent transportation systems applications, Transportation Research
Part B: Methodological, Volume 33, Issue 6, August 1999, Pages 437-457, ISSN 0191-2615,
[7] Davidsson, P., Henesey, L., Ramstedt, L., Tornquist, J. and Wernstedt, F. 2005. An analysis of agent based approaches to transport logistics.
Transportation Research: Part C: Emerging Technologies, 13(4), 255–271.Hawkins, D. M. The problem of overfitting. J. Chem. Inf. Comput. Sci.
2004, 44, 1-12
[9] Smith, L., R. Beckman, D. Anson, K. Nagel, and M. E. Williams. 1995. Transims: Transportation analysis and simulationsystem. In Proceedings of the
Fifth National Conference on Transportation Planning Methods. Seattle, Washington.
[10] Cameron, G., and G. Duncan. 1996. Paramics: Parallel microscopic simulation of road traffic. In Journal of SuperComputing
[11] Prevedouros, P. D., and Y. Wang. 1999. Simulation of a large freeway/arterial network with integration, tsis/corsim and watsim. In Transportation
Research Record 1678, 197–207
[12] Concepts, I. T. 2001. Vissim simulation tool. In http://www.itc-world.com/VISSIMinfo.htm
[13] Sargent RG. Verifying and validating simulation models. 1996 Winter Simulation Conference, Coronado, CA,1996. p. 55–64.
[15] Tetko, I.V.; Livingstone, D.J.; Luik, A.I. Neural network studies. 1. Comparison of Overfitting and Overtraining, J. Chem. Inf. Comput. Sci., 1995, 35,
826-833
[16] D.E. Goldberg, K. Deb, A comparative analysis of selection schemes used in genetic algorithms, in: G.J.E. Rawlins (Ed.), Foundations of Genetic
Algorithms, Morgan Kaufmann, San Mateo, CA, 1991, pp. 69–93.
[17] Stewart Robinson (2004). Simulation - The practice of model development and use. Wiley.
| 5cs.CE
|
Some Investigations about the Properties of Maximum Likelihood
Estimations Based on Lower Record Values for a Sub-Family of the
Exponential Family
S. D. Gore1, P. Nasiri2, S. Hosseini3, and M. F. Naeini4
Abstract:
Here, in this paper it has been considered a sub family of exponential family. Maximum likelihood
estimations (MLE) for the parameter of this family, probability density function, and cumulative density
function based on a sample and based on lower record values have been obtained. It has been considered
Mean Square Error (MSE) as a criterion for determining which is better in different situations.
Additionally, it has been proved some theories about the relations between MLE based on lower record
values and based on a random sample. Also, some interesting asymptotically properties for these
estimations have been shown during some theories.
Keywords: Exponential Family, Lower Record Values, ML Estimation, Asymptotically Un-biasedness.
1- Introduction:
For the first time, Chandler in 1952 introduced the concept of record values and record times. After him
and in a short period of time there were done a large amount of researches about this new topic by
statisticians. Shorrock in 1973 obtained some asymptotically conclusion in the field of record values, in
the same time Ahsanullah during a paper named “an introduction to record values” clarified basic
concepts of record theory (Ahsanullah 1973). Record theory is useful topic that is used in any branch of
science from economy to any competitive phenomena. For the recent researches in this field it can be
mentioned works of Ahmadi & Doostparast (2006), Eric Marchand & Parsian (2009), Nasiri and Hosseini
(2013).
Before beginning the main topics it is needed to introduce a sub-family of exponential family. Generally
f X ( x; ) determines exponential family if
f ( x; ) h( x) g ( )exp{( )T ( x)},
Here, it is considered an important sub family of exponential family as bellow
F ( x; ) exp{ B( ) A( x)},
It is called second type of exponential family during this paper. In which A is a decreasing function and
1
Sharad Damodar Gore is a professor at department of statistics, Savirtibai Phule Pune University, India
Parviz Nasiri is an associate professor at Payam Noor University of Tehran, Iran
3
Correspondent author, he is a lecturer at Chihan University of Erbil, Kurdistan Independent Region, Iraq. Email
them at [email protected]
4
She has been graduated from Amir Kabir University of Technology, Tehran, Iran
2
1
B( ) 0; x [a, b]; A(a) , A(b) 0; a, b
It is possible to investigate a lot of distributions by considering the second type of exponential family.
The following table clarifies the subject.
Table 1.1: examples of the second type exponential family
Distribution Name
B( )
Frechet
A( x ) x
Gumbel
A( x) exp( x)
B( ) exp( )
Power function
A( x) Ln( x)
B( )
A( x)
In this section it is tried to study theorems about the relation between ML estimations based on a random
sample and ML estimation based on lower record values.
2- Maximum Likelihood Estimations Based on Lower Record Values and Based on a Random
Sample for Unknown Parameter of Type II Exponential Family
In this section it is tried to study some theorems about the relation between ML estimations based on a
random sample and ML estimation based on lower record values. Also, considering MSE as medium of
assessment, it is shown that in some cases estimations based on lower record values have less value of
error. Remember the general form for this family is as follows.
F ( x; ) exp{ A( x) B( )}; B( ) 0, x [a, b], A(a) , A(b) 0, a , b .
F ( x; ) exp{ A( x) B( )}, that’s why, probability density function is obtained as :
f ( x; ) A' ( x) B( ) exp{ A( x) B( )}.
Considering a random sample of size n, the joint probability density function for a random sample is
obtained as following.
n
n
n
i 1
i 1
i 1
f ( x1 ,..., xn ; ) f ( xi ; ) ( 1) n B n ( ) A' ( xi ) exp{ B( ) A( xi )}
n
n
i 1
i 1
B n ( ) { A' ( xi )}exp{ B( ) A( xi )}.
Hence,
n
L log f x1 ,, xn ; nlogB log{
i 1
n
1
}
B
A X i .
A' ( X i )
i 1
2
As previously said, the maximum likelihood estimation is calculated by solving the equation
L
0,
therefore
n
L
nB ' ( )
n
0
B ' A X i 0 B n
,
B( )
A
X
i 1
i1 i
(1)
assuming that B( ) is an one-to-one function, the above equation will be solved like this
B 1 (
n
i1A X i
n
).
Since the cumulative distribution function F ( x; ) exp{ A( x) B( )} with the properties and the
conditions described for A and B is a continuous function, it is possible to derive the first and the
second order derivatives. So,
n
2 L nB '' ( ) B( ) n[ B ' ( )]2
''
B
(
)
t
,
(
t
A( xi ))
2
B 2 ( )
i 1
(2)
In addition, substituting (1) in (2), the below result is reached.
L
2 B
2
n
i 1A X i
n
n
nB '' ( )( ) n[ B ' ( )]2
[ B ' ( )t ]2
''
t
B ( )t
0.
n
n
( )2
t
(3)
The value of second derivative is less than zero therefore; the relations (1) and (3) tell us the following
estimation is ML estimation.
ˆMLE ,( X ,..., X ) B 1 (
1
n
n
n
A Xi
i 1
n
), T X i .
i 1
On the other hand A( X i ) is distributed as Exponential distribution with parameter
1
.
B ( )
P( A( X i ) t ) P( A( X i ) t ) 1 (1 F (t; )) F (t; ) 1 Exp( B( )t ) .
On the other hand, A( X i ) ’s are independent variable that’s why it is possible to find the distribution of
A X easily like follow:
n
i 1
i
n
T A( X i )
i 1
distribution
1
Gamma (n,
).
B( )
3
As a result
ˆ
distribution
n
B 1 ( )
T
(4)
distribution
in which T
Gamma(n,
1
).
B( )
On the other hand, with respect to the definition, the lower record values of the joint probability density
function ( R1' ,..., Rn' ) are resulted as follows (Arnold, Balakrishnan and Nagaraja, 1998).
n 1
f r , , r ; f r ;
'
1
'
n
'
n
i 1
f ri' ;
F ri' ;
n 1
( A' ( rn' ) B( )) exp{ B( ) A( rn' )}{ A' ( ri' ) B( )}
i 1
n
B n ( ){ A' ( ri' )}exp{ B( ) A( rn' )}.
i 1
Considering the above probability density function, the maximum likelihood function based on lower
record values is obtained like below:
n
L' log( f ( r1' ,..., rn' )) nLog{B( )} Log{
R
i 1
1
} A( rn' ) B( ),
'
A ( ri )
'
the maximum likelihood estimation based on lower record values is obtained by solving the equation
L'
R
0 . Therefore,
ˆMLE BASED ON LOWER RECORDS ˆMLE ,( R ,...,R ) B 1 (
'
1
'
m
n
).
A( Rn' )
On the other hand, Rn' is distributed as the following probability density function (Arnold, Balakrishnan
and Nagaraja, 1998).
f R' ( rn' )
n
( A' ( rn' )) B n ( )
( A( rn' ))n 1 exp{ A( rn' ) B( )},
(n)
therefore
ˆMLE ,( R ,..., R
'
1
'
m)
Distribution
n
B 1 ( )
T
(5)
in which
4
T A( Rn' )
Distribution
Gamma(n,
1
)
B( )
Considering the above descriptions as well as the relations (4) and (5), the following theorem is readily
concluded.
Theorem: considering a random sample of size n from the exponential family of the second type and also
considering n values from the lower records for this family, the below relations are always true.
a) ˆMLE ,( X1 ,..., X n )
Distribution
b) (ˆMLE ,( X1 ,..., X n ) )
c)
ˆMLE ,( R ,..., R )
'
1
Distribution
'
n
(ˆMLE ,( R ,..., R ) ) in which is a real function.
'
1
'
n
MSE ( (ˆMLE ,( X1 ,..., X n ) )) MSE ( (ˆMLE ,( R' ,..., R' ) ))
1
d)
n
n
T
n
T
' (n) MSE ( (ˆMLE ,( X ,..., X ) )) MSE ( (ˆMLE ,( R ,...,R ) )) E[ 2{B 1 ( )}] 2 ( ) E[ {B 1 ( )}]
1
'
1
n
'
n
2 ( ),
Distribution
in which T
Gamma(n,
1
).
B( )
Proof:
Proofs for all parts are obvious considering the relations (4) and (5).
Note: From part d it is simply found out if a random of size n from exponential family of the second type
includes m values of the lower records ( m n ), then
MSE (ˆMLE ,( R' ,...,R' ) ) MSE (ˆMLE ,( X1 ,..., X m ) ) MSE (ˆMLE ,( X1 ,..., X m ) ) ' (m).
1
m
Simply speaking, the Mean Square Error of the estimations based on m values of lower records is equal
with the MSE’s of the estimations based on a random sample of size m. In fact, this is a practically useful
note because by considering this note it can be shown that in some cases estimations abased on lower
records have less error in comparison with the ones based on random sample. The following examples
clarify this
Example 1) Let A( x) Log ( x), B( ) , and (t ) t . It is wanted to estimate parameter B( )
in this case ' (n ) is obtained as bellow
5
n
T
n
T
' (n ) MSE ( (ˆMLE ,( X ,..., X ) )) MSE ( (ˆMLE ,( R ,...,R ) )) E[ 2{B 1 ( )}] 2 ( ) E[ {B 1 ( )}] 2 ( )
1
n
'
1
'
n
2 n 1 n
n t
n
n
n
n
exp( t )
E[{B 1 ( )}2 ] 2 E[ B 1 ( )] 2 E[( ) 2 ] 2 E[ ] 2
dt
2
0
T
T
T
T
t
( n )
2
0
n t n 1 n exp( t )
dt 2 .
t
( n )
After some mathematical conclusion it is easily obtained as:
n2
2n
(n) (
1) 2
(n 2)(n 1) (n 2)
'
Figure 1.2. Values of MSE for different values of n.
Considering the related figure of the above function, it is seen in each sample of size n with m values of
lower records (m<n), the estimation based on sample has less value of error in comparison with
estimation based on lower record values . It may instill the idea of uselessness of the estimation based on
lower record values in the mind, but next instance shows it is not generally true. In some cases the
estimation based on lower record values has less value of error.
Example 2) If in the previous example (t ) et , in this case
6
n
T
n
T
n
T
n
T
' (n ) E[ 2{B 1 ( )}] 2 ( ) E[ {B 1 ( )}] 2 ( ) E[exp(2 )] 2 exp( ) E[exp( )] exp(2 )
(n)
'
0
2n
t n 1
n t
e t e
( n )
dt 2e
2n
t
Considering the expansions e
0
i 0
' (n)
0
(
i 0
n
t n 1
e t n e t
dt e2 .
( n )
n
(2n)i
ni
t
and
, the above relation is obtained as follow
e
i
t ii !
i 0 t i !
(2n)i n 1 n t
ni n 1 n t
)
t
e
(
)t e
i
t ii !
i 0 t i !
dt 2e
dt e2 .
0
( n )
( n )
(2n )i
The value of i is greater than zero, that’s why it is possible to interchange the integration and
t i!
summation. Therefore
' (n )
i 0
n
i 0
(2n )i (n i ) t n i 1 n i e t
(n )i (n i ) t n i 1 n i e t
dt
2
e
dt e 2
0
0
(i 1)(n )
( n i )
( n i )
i 0 (i 1) ( n )
n
n
(2n )i (n i )
(n )i (n i ) 2
(n )i (n i ) i
2e
e
(2 2e ) e2 .
(i 1)(n )
i 0 (i 1) ( n )
i 0 (i 1) ( n )
Considering the graph of obtained ' (n ) , it can be said if in a sample of size 5, 6, 7, or 8 there is just 2
value of records the MSE of ML estimation based on lower record values has less value than ML
estimation based on a random sample. Also, it is clear for samples with large size the MSE of Estimators
based on lower record values and based on a random sample are approximately the same. For example for
a sample of size 500 includes 15 value of records, we can say
approximately
MSE based on lower records
MSE based on a random of size 500.
7
Figure 2.2. Values of MSE considering different values of n in estimation of the parameter e
3. Maximum Likelihood Estimations for Cumulative Distribution Function (CDF) and Probability
Density Function (PDF) from Exponential Family of the Second Type Based on Lower Record
Values
Regarding stability property of the maximum likelihood estimations and equation (5), hereunder results
are readily obtained for PDF and CDF of the type II exponential family.
ˆ
MLE ,( R'1 ,..., R'm )
B 1 (
m
)
A( R ' m )
( mA ( x ))
mA( x )
fˆMLE ,( R' ,...,R' )
exp(
)
'
1
m
A( Rm' )
A Rm
'
Distribution
( nA' ( x ))
nA( x )
exp(
)
T
T
and
mA( x )
FˆMLE ,( R' ,...,R' ) exp(
)
1
m
A( Rm' )
Distribution
Distribution
in which T
Gamma (m,
exp(
mA( x)
),
T
1
).
B( )
Theorem 1. If a random sample of size n, includes m values from the lower record values (
Ri' , i 1,..., m ), then the maximum likelihood estimations based on lower record values for PDF (
f ( x; ) ) and CDF ( F ( x; ) ) are biased estimators and
8
(m i 1)
ˆ
i E[ f MLE ,( R' ,...,R' ) ] [
( 1)i {mB( )}i 1 ( A' ( x ))]{ A( x )}i
1
m
i 0 ( m ) (i 1)
ii E[ FˆMLE ,( R' ,...,R' ) ] [
1
m
i 0
( m i )
{mB( ) A( x )}i ]
(m)(i 1)
Here, the second part is proven. The other part is proven similarly.
E[ FˆMLE ,( R' ,...,R' ) ] E[exp(
1
m
Considering the expansion e
mA( x )
e
)]
'
0
A( Rm )
mA( x )
t
i 0
mA( x )
m 1
t
B m ( )e B ( )t
dt
( m )
t
(6)
( mA( x ))i
and (6)
t ii !
( mA( x ))i m 1 m
t B ( )e B ( ) t
i
t
i
!
dt i 0
dt
0
( m )
E[ FˆMLE ,( R' ,...,R' ) ]
1
e
0
m
mA( x )
m 1
t
t
B m ( )e B ( ) t
( m )
(mA( x ))i m 1 m
( mA( x ))i m 1 m
B ( ) t
t
B
(
)
e
t B ( )e B ( ) t
i
i
t i!
t i!
dt odd indices
dt
0 even indices
0
( m )
( m )
A
B
In above expression
(mA( x ))i
and
t ii !
even indices
(mA( x ))i
in A and B are positive quantities and
t ii !
odd indices
that’s why it is possible to interchange the integration and summation.
E[ FˆMLE ,( R' ,...,R' ) ]
1
m
(m i )(mA( x ) B( ))i
(i 1)( m)
even indices
0
t mi 1B m i ( )e B ( ) t
dt
( m i )
(m i )(mA( x ) B( ))i
(i 1)( m)
odd indices
(m i )(mA( x ) B( ))i ( m i )( mA( x ) B( ))i
.
(i 1)( m)
(i 1)( m)
odd indices
i 0
0
t m i 1B m i ( )e B ( ) t
( m i )( mA( x ) B( ))i
dt
( m i )
(i 1)(m)
even indices
And finally,
E[ FˆMLE ,( R' ,...,R' ) ]
1
m
i 0
(m i )
( mA( x) B( ))i .
(i 1)(m)
Theorem 2. Assuming that a sample of size n from distribution of the exponential family of the second
'
type includes m values of the lower records ( Ri ), the hereunder results are always true.
i 1,..., m
9
( m i 2)
m2
i MSE[ fˆMLE ,( R ,...,R ) ] (mB( ) A' ( x )) 2 [
{2mB( ) A( x )}i ]
( m)(i 1)
'
1
'
m
i 0
2
( m i 1)
me A( x ) B ( ) (mA' ( x ) B( ))2 [
{mB( ) A( x )}i ] { A' ( x ) B( )}2 e 2 A( x ) B ( )
m
( m)(i 1)
i 0
m 1
m 1
ii MSE[ FˆMLE ,( R ,...,R ) ]
'
1
'
m
i 0
( mA( x ) B( ) ( m i ))i i
(2 2e A( x ) B ( ) ) e 2 A( x ) B ( )
(i 1) ( m)
Proof:
The second part has been proven,
MSE ( FˆMLE ,( R' ,...,R' ) ) E ( FˆMLE ,( R' ,...,R' ) F )2 E ( Fˆ 2MLE ,( R' ,...,R' ) ) 2 F ( x; ) E ( FˆMLE ,( R' ,...,R' ) ) F 2 ( x; )
1
e
n
1
n
1
2 mA( x )
m 1
t
e
B m ( )e B ( ) t
dt 2e A( x ) B ( )
0
( m )
t
0
Considering the expansion e
mA( x )
t
i 0
n
1
n
mA( x )
m 1
t
t
B m ( )e B ( ) t
dt e 2 A( x ) B ( )
( m )
( mA( x ))i
,
t ii !
MSE ( FˆMLE ,( R' ,...,R' ) )
1
n
( 2mA( x ))i m 1 m
( mA( x ))i m1 m
B ( ) t
t
B
(
)
e
t B ( )e B ( ) t
i
i
t i!
t i!
i 0
dt 2e A( x ) B ( ) i 0
dt e 2 A( x ) B ( )
0
0
( m )
( m )
0
(
(2mA( x ))i
(2mA( x ))i m 1 m
)t B ( )e B ( ) t
i
i
t i!
t i!
even i
odd i
dt
( m )
2 e A ( x ) B ( )
(
0
( mA( x ))i
( mA( x ))i m 1 m
)t B ( )e B ( ) t
i
i
t
i
!
t
i
!
even i
odd i
dt e 2 A( x ) B ( )
( m )
In above expression
(2mA( x ))i
(2mA( x ))i
,
t ii !
t ii !
even indices
odd indices
(mA( x ))i
, and
t ii !
even indices
(mA( x ))i
are all positive quantities and that’s why it is possible to interchange the integration
t ii !
odd indices
and summation.
10
MSE ( FˆMLE ,( R' ,...,R' ) )
1
n
(2mA( x ))i
even i (i 1) ( m )
2e
A ( x ) B ( )
m 1
i 0
m 1
i 0
m i 1 m
B ( ) t
t B ( )e dt
0
i
( mA( x ))
(
even i (i 1) ( m )
t
m i 1
(2mA( x ))i
odd i (i 1) ( m )
B ( )e
m
0
B ( ) t
t
m i 1
B m ( )e B ( ) t dt
0
(2mA( x ))i
dt
odd i (i 1) ( m )
t
m i 1
B m ( )e B ( ) t dt ) e 2 A( x ) B ( )
0
(2mA( x )) ( m i ) B ( )
( mA( x )) ( m i ) B ( )
2 e A ( x ) B ( )
e 2 A( x ) B ( )
(i 1) ( m)
(i 1) ( m)
i 0
i
m 1
i
i
i
( mA( x ) B( ) ( m i ))i i
(2 2e A( x ) B ( ) ) e 2 A( x ) B ( )
(i 1) ( m)
Theorem 3. Assuming that a sample of size n from the exponential distribution of the second type
( X1 ,..., X n ) includes m values of the lower records ( Ri' ), the estimators fˆMLE ,( R' ,..., R' ) and
1
i 1,..., m
m
FˆMLE ,( R' ,..., R' ) are always asymptotically unbiased estimators:
1
m
i Lim E[ fˆMLE ,( R' ,...,R' ) ] A( x ) B( ) exp( A( x ) B( )) f ( x; )
m
1
m
ii Lim E[ FˆMLE ,( R' ,...,R' ) ] exp( A( x ) B( )) F ( x; )
m
1
m
The following lemma is needed in order to prove the above theorem.
Lemma 1. Lim
n
Lim
n
(n i 1)ni 1
1 for the reason that
( n)
(n i 1)n i 1
Lim
n
( n )
ni
i 1
(n k )
Lim
n
ni
1
ni
k 1
It must be proven that
(n i 1)ni 1
0, N , s.t , n N :
1 .
(n)
Now, assuming that is arbitrary, N is chosen as N
i 1
( 1) i i 1 ( 1)
i 1
( 1) 1
. Now
(n i 1)ni 1 (n i 2)!ni 1
ni 1
n
n
n
n
...
...
( n)
(n 1)!
(n 1)(n 2)...(n i 1) n 1
n i 1 n i 1
n i 1
( i 1) times
11
( i 1) times
On the other hand, it can be simply shown that
n N , and , i 1,2,3,..., N 2 :
n
N
,
n i 1 N i 1
thus
n
n
N
N
...
...
n i 1
n i 1 N i 1
N i 1,
(i 1)times
(i 1)times
and consequently
n N :
(n i 1)ni 1
N
N
.
...
( n)
N i 1
N i 1
( i 1) times
Finally by substituting N
i 1
( 1) i i 1 ( 1)
i 1
( 1) 1
in the last relation, the following result is reached:
(n i 1)ni 1
1
( n)
Proof:
Only the first part is proven; the other parts have similar proofs. First notice that
(m i 1)
Lim[ fˆMLE ,( R' ,...,R' ) ] [ Lim
( 1)i mi 1{B( )}i 1 A' ( x )]{ A( x )}i ]
1
m
m
(
m
)
(
i
1)
m
i 0
Lim(
i 0
n
(m i 1)mi 1 { B( )}i B( ) A' ( x ){ A( x )}i ]
]
( m )
(i 1)
Considering lemma 1
{ A( x ) B( )}
Lim[ fˆMLE ,( R' ,...,R' ) ]
B( ) A' ( x ) A '( x ) B( ) exp{ A( x ) B( )}.
1
n
m
i
!
i 0
i
Theorem 4. In the exponential family of the second type, if the sample ( X1 ,..., X n ) has m values from
'
the lower records ( Ri ), then the maximum likelihood estimations for the distribution function and the
i 1,..., m
probability density function based on random sample and on lower records are consistent estimators for
the probability density function and for the distribution function.
12
P
i fˆMLE ,( R' ,...,R' )
f ( x )
1
m
P
ii FˆMLE ,( R' ,...,R' )
F ( x ).
1
m
Proof. Here it is just proven the first part; the second part has a similar proof. It is considered the
following abbreviation just for easiness.
fˆMLE ,( R' ,...,R' ) fˆMLE ,R' .
1
m
Based on the Markov’s theorem
P{ fˆMLE ,R' E[ fˆMLE ]}
Var ( fˆMLE ,R' )
.
On the other hand, it is concluded by some mathematical calculations that
(m i 2)
E[{ fˆMLE ,R' }2 ]
{ A '( x )}2 {B( )}2 ( 2)i {mB( )}i { A( x )}i ,
(
i
1)
(
m
)
i 0
and consequently:
(m i 2)
Var( fˆMLE ,R' )
{ A '( x )}2{B( )}2 ( 2)i {mB( )}i { A( x)}i {E[ fˆMLE ,R' ]}2
i 0 (i 1) ( m)
thus
LimVar ( fˆMLE ,R' ) Lim[
n
i 0
n
mi (m i 2) {2 B( )}i { A( x )}i
]
{ A '( x )}2 {B( )}2 { f ( x )}2
( m )
(i 1)
{ A '( x )}2 {B( )}2 exp{2 A( x ) B( )} { f ( x )}2 0.
Therefore
Lim P{ fˆMLE ,R' E[ fˆMLE ,R' ] } 0
n
so
P
fˆMLE ,R'
E[ fˆMLE ,R' ]
P
fˆMLE ,R' E[ fˆMLE ,R' ]
0
P
On the other hand, it is reminded that if Lim am a and Y
b then (Billingsley (1995))
m
P
Y am
b a
(7)
Considering (7) and respecting the first part of theorem 3 ( Lim E[ fˆMLE ,R' ] f ( x ) ), it is obtained that
n
P
fˆMLE ,R' E[ fˆMLE ,R' ] E[ fˆMLE ,R' ] f
0
,
13
or
or equally
P
fˆ( MLE ,R' ,...,R' )
f ( x; )
1
m
Theorem 5. If the sample ( X1 ,..., X n ) has m values from the lower records ( Ri' ), then the maximum
i 1,..., m
likelihood estimations based on lower records are consistent estimations for
P
ˆMLE ,( R ,...,R )
'
1
'
m
Proof:
With respect to the continuity of functions B(t ), Log (1 t ), F (t ) 1 exp{ A(t ) B( )}, and the
continuity theorem of composite functions, it is concluded that the function D(t ) B(
is always continuous on
Log (1 F (t ))
)
A( x)
. On the other hand, it is known from the second part of theorem 4 that
P
FˆMLE ,( R1 ,...,Rm )
F ( x ),
or equivalently
P
A( x) B(ˆMLE ,( R1 ,..., Rm ) ) exp{ A( x) B(ˆMLE ,( R1 ,..., Rm ) )}
F ( x) .
Now
P
D( FˆMLE ,( R1 ,..., Rm ) )
D( F ( x))
which is equivalent to
P
ˆMLE ,( R ,..., R )
1
m
4. Conclusions
In the beginning, a general subfamily from the exponential family was considered, it was tried to obtain
the ML estimations based on a random sample and based on lower record values for unknown parameter
of this family; then, relations between these estimations were shown theoretically. It was engaged via
some theorems to description of this problem that the record estimations can be acceptable in general
state. Additionally, the ML estimations based on a random sample and based on lower record values were
calculated for the CDF and PDF of this sub-family, and their corresponding errors were described and
studied by using some theorems; in addition, it was proven that at asymptotical states, these estimations
are unbiased. Subsequently, several theorems associated with convergence in probability were expressed
for the resulted estimators.
14
References
1- Chandler, K. N. (1952). The distribution and frequency of record values. Journal of the Royal
Statistical Society. Series B (Methodological), 220-228.
2- Shorrock, R. W. (1972). A limit theorem for inter-record times. Journal of Applied
Probability, 9(1), 219-223.
3- Ahsanullah, M. (1988). Introduction to record values.
4- Ahmadi, J., Doostparast, M., & Parsian, A. (2005). Estimation and prediction in a two-parameter
exponential distribution based on k-record values under LINEX loss function. Communications in
Statistics–Theory and Methods, 34(4), 795-805.
5- Ahmadi, J., & Doostparast, M. (2006). Bayesian estimation and prediction for some life
distributions based on record values. Statistical Papers, 47(3), 373-392.
6- Ahmadi, J., Jozani, M. J., Marchand, É., & Parsian, A. (2009). Bayes estimation based on krecord data from a general class of distributions under balanced type loss functions. Journal of
Statistical Planning and Inference, 139(3), 1180-1189.
7- Baklizi, A. (2008). Likelihood and Bayesian estimation of using lower record values from the
generalized exponential distribution. Computational Statistics & Data Analysis, 52(7), 34683473.
8- Raqab, M. Z., Ahmadi, J., & Doostparast, M. (2007). Statistical inference based on record data
from Pareto model. Statistics, 41(2), 105-118.
9- Nasiri, P., & Hosseini, S. (2012). Statistical inferences for Lomax distribution based on record
values (Bayesian and classical). Journal of Modern Applied Statistical Methods, 11(1), 15.
10- Ahmadi, J., & Balakrishnan, N. (2004). Confidence intervals for quantiles in terms of record
range. Statistics & probability letters, 68(4), 395-405.
11- Nasiri, P., Hosseini, S., Yarmohammadi, M., & Hatami, F. (2013). Bayesian inference for
exponential distribution based on upper record range. Arabian Journal of Mathematics, 2(4), 349364.
12- B. C. Arnold, N. Balakrishnan, H.N. Nagaraja,"records,"John Wiley and Sons,Canada1998.
13- Billingsley, P. (1995). Probability and Measure, ser. Probability and Mathematical Statistics.
New York: Wiley.
15
| 10math.ST
|
Using Convolutional Codes for Key Extraction in
SRAM Physical Unclonable Functions
Sven Müelich∗ , Sven Puchinger∗, Martin Bossert∗
arXiv:1704.01306v2 [cs.IT] 10 Jan 2018
∗
Ulm University, Institute of Communications Engineering, 89081 Ulm, Germany
Email: {sven.mueelich, sven.puchinger, martin.bossert}@uni-ulm.de
Abstract—Physical Unclonable Functions (PUFs) exploit variations in the manufacturing process to derive bit sequences from
integrated circuits, which can be used as secure cryptographic
keys. Instead of storing the keys in an insecure, non-volatile memory, they can be reproduced when needed. Since the reproduced
sequences are not stable due to physical reasons, error correction
must be applied. Recently, convolutional codes were shown to be
suitable for key reproduction in PUFs based on SRAM. This
work shows how to further decrease the reconstruction failure
probability and PUF implementation size using codes with larger
memory length and decoding concepts such as soft-information
and list decoding.
Index Terms—Physical Unclonable Functions, Key Generation
and Storage, Convolutional Codes, Sequential Decoding
I. M OTIVATION
AND
P RELIMINARY W ORK
Generating cryptographic keys by using Pseudo Random
Number Generators (PRNGs) and storing them in non-volatile
memory has several drawbacks. First, PRNGs do not provide
true randomness and hence violate the demand for unique
and unpredictable keys. Second, storing keys makes them
vulnerable to physical attacks and hence requires expensive
protecting mechanisms which need additional chip area and
power consumption. Physical Unclonable Functions (PUFs)
are devices which extract unique cryptographic keys from intrinsic randomness like delay characteristics of digital circuits
or initialization behavior of memory cells. All devices (even
with the same functionality) differ regarding this intrinsic
randomness, which is unique for each device and cannot be
controlled during the manufacturing process due to technical
and physical limitations. A PUF uses this randomness to
extract a random, unique and unpredictable sequence of bits
which can be used as cryptographic key. Since the randomness
is reproducible, the key can simply be regenerated when
needed instead of storing it permanently. Hence, PUFs implement secure key generation and secure key storage.
Since derived bit sequences suffer from environmental conditions like temperature or voltage supply, responses may
vary slightly. To circumvent this problem, error-correcting
codes have to be used. Therefore, the initial PUF response
is used to obtain helper data, which is needed for reliable
key reproduction. Since enough uncertainty remains when
knowing the helper data, a protected memory is not required.
Extraction of helper data and using them to reproduce a key
can be done by different helper data schemes, cf. [1], [2], [3],
[4].
Previous work most often used concatenated schemes for
error correction. [5] suggested different ordinary concatenated
schemes consisting of a repetition code as inner code and
BCH, Reed–Muller and Golay Codes respectively as outer
code. [6] applied an ordinary code concatenation consisting
of a BCH and a repetition code. Generalized concatenated
codes were applied for the first time to PUFs in [7] using
Reed–Muller codes, while a similar construction using BCH
and Reed–Solomon codes was given in [8].
In [3], convolutional codes were used the first time for
error correction in PUFs. Decoding was performed using a
hard-decision Viterbi algorithm. As channel model, a binary
symmetric channel (BSC) with variable crossover probabilities
pi was assumed, where pi corresponds to the probability of
a bit flip in the i-th SRAM cell. In order to increase the key
reproduction reliability, unreliable SRAM cells were ignored,
i.e., for a given threshold pT , only cells with pi < pT were
used, resulting in an implementation overhead.
In [9] the decoding failure rate was further decreased using
an estimator for the decoding output reliability (simplified
ROVA) in combination with multiple readouts.
In this paper, we show that using full soft information at
the Viterbi decoder’s input, the decoding failure rate can be
reduced significantly in the considered scenario.1 Furthermore,
more unreliable bit positions can be included without losing
error correction capability compared to [3], [9]. Thus, the
decoding overhead vanishes. We show, that the use of list
decoding further improves the scheme in terms of failure
probability and provides decoding reliability for free as an
alternative to simplified ROVA. Also, the combination of list
decoding and multiple readouts can improve the results of [9].
Our simulations indicate that convolutional codes with
larger memory length result in smaller failure probabilities,
which is a well-known fact in coding theory. In order to make
decoding of codes with a memory length up to 25 feasible,
we apply sequential decoding. To our knowledge, sequential
decoding is applied to the PUF scenario for the first time.
1 Full soft information at the decoder’s input was used for the first time
in [10]. However, instead of convolutional codes and Viterbi algorithm, a
concatenation of short block codes was used. The use of convolutional codes
was labeled inefficient without further study, due to apparently required very
long code sequences. In contrast to that statement, [3], [9] as well as our
analysis show that sufficiently short length can be achieved.
II. P HYSICAL U NCLONABLE F UNCTIONS (PUF S )
There is a variety of ways to realize PUFs. In this work
we mainly focus on SRAM PUFs which were introduced
in [11]. Dependent on the mismatch of two inverters of an
SRAM cell (which is caused by manufacturing variations) and
electrical noise, every SRAM cell converges to either 0 or 1
when powering-up the device. Since the mismatch is static
and dominates over the noise, most cells always power-up
with either 0 or 1. However, some of the cells do not have
a preferred behavior and hence sometimes initialize with 0
and sometimes with 1. It is possible to produce reliability
information of the individual cells by obtaining statistics
during an enrollment phase. Based on such statistics, [12]
derived a cumulative density function (cdf) and a probability
density function (pdf) which approximate the measured oneprobability Pone of SRAM cells:
cdf Pone (x) = Φ(λ1 · Φ−1 (x) − λ2 )
pdf Pone (x) =
λ1 · ϕ(λ2 − λ1 · Φ
ϕ(Φ−1 (x))
−1
(x))
(1)
(2)
In Equations (1)–(2), Φ denotes the standard normal cumulative distribution function. For these functions we use
λ1 = 0.51 and λ2 = 0 according to [12], in order to obtain a
model which is equivalent to a BSC with crossover probability
p = 0.15, which is often used in the PUF scenario. [12] also
derived a cumulative density function and a probability density
function of the error probability Pe of an SRAM cell:
cdf Pe (x) = cdf Pone (x) + 1 − cdf Pone (1 − x)
(
pdf Pone (x) + pdf Pone (1 − x) , if x < 0.5,
pdf Pe (x) =
0
, otherwise.
cf. Figure 1 for a (2, 1, [2]) encoder with G0 = (1, 1),
G1 = (1, 0), G2 = (1, 1). Initially, all memory elements
contain zeros. Information block ir is mapped to codeword
cr,1 cr,2 by applying linear operations. Then all bits move one
memory element to the right. Bit ir enters the leftmost memory
element, bit ir−2 is dropped. Encoding continues with block
ir+1 .
cr,1
ir
Fig. 1. Encoder of a (2, 1, [2]) convolutional code.
B. Viterbi Algorithm
The Viterbi algorithm realizes efficient maximum-likelihood
decoding [17]. We can visualize decoding using a graph
structure called trellis (cf. Figure 2 for the trellis of the
(2, 1, [2]) code introduced in Section III-A). The nodes in the
trellis represent the 2µ states of the shift register. Each node
has two outgoing edges for encoding a 0 (dashed lines) or a 1
(solid lines). Since the register is initialized with zeros, not all
states occur in the beginning. The edges are labeled with the
corresponding codeword and represent the transitions to the
next memory state. Hence, each path in the trellis represents a
possible code sequence c. The decoder aims to find the code
sequence which was transmitted with highest probability when
the sequence y was received, hence it maximizes p(y|c).
00
00
00
00
00
00
11
10
11
10
11
10
11
10
01
11
01
11
01
11
01
00
01
00
01
10
10
00
01
Convolutional codes, introduced by Elias [13], are widely
used in applications. Efficient methods for implementations
in hardware make them suited for PUFs. [3] applied convolutional codes to PUFs for the first time. An optimized
FPGA Viterbi decoder adapted to PUFs was designed in [14].
We explain encoding and decoding briefly, for a detailed
description of convolutional codes we refer to [15], [16].
10
Convolutional codes can be described using the parameters
(n, k, [µ]). Using a code of rate nk , we partition an information
sequence into blocks of length k. Each information block is
mapped to a codeword of length n > k. In contrast to block
codes, the codeword cr depends on information block ir and
on the µ previous information blocks ir−1 , . . . , ir−µ .
Encoding of an information block can be done by calculating cr = ir G0 + ir−1 G1 + . . . ir−µ Gµ , or equivalently
cr = (ir , ir−1 , . . . ir−µ ) · (GT0 , GT1 , . . . , GTµ )T , using (k × n)
generator matrices G0 , G1 , . . . , Gµ . This process can be implemented by using a linear shift register for each of the k
inputs, n outputs and µ memory elements per shift register,
ir-2
cr,2
III. C ONVOLUTIONAL C ODES
A. Description and Encoding
ir-1
11
11
Fig. 2. Trellis diagram for decoding a (2, 1, [2]) code.
Let c = (c1 , c2 , . . . ) denote the code sequence which is sent
over a noisy channel, ci denotes a length-n block. Analog to
this notation, y denotes the received sequence. We define a
submetric and a metric.
Definition 1. For every edge, the submetric is defined as
(
pi
, yi 6= cj
.
λi = p(yi |ci ) =
1 − p i , y i = ci
where distH (ci , yi ) denotes the Hamming distance of codeword ci and received word yi . The metric of a P
node is the
η
sum over all η submetrics λi on a path, i.e. Λ = i=1 λi .
Viterbi calculates the metric for all nodes in the trellis. To
calculate the metric for a specific node, for each incoming
edge Viterbi adds the incoming submetric to the metric of the
corresponding node on the left side of the segment. Viterbi
C. Sequential Decoding
Viterbi’s algorithm provides maximum-likelihood decoding.
However, its complexity grows exponentially when increasing µ. In this section we briefly discuss sequential decoding, introduced in [19]. Since for this class of algorithms,
complexity does not depend on the constraint length, larger
constraint lengths can be used. Sequential decoding is a almost
maximum-likelihood decoding and can easily be implemented.
A further advantage in comparison to Viterbi is a runtime
depending on the number of errors in the received sequence.
There exist different algorithms to realize sequential decoding,
e.g. ZJ [20], [21] and Fano [22]. We decided to use the Fano
algorithm which is suitable for hardware implementations,
since it requires essentially no memory [16, Chapter 6] and
hence is a reasonable choice for the PUF scenario. Compared
to ZJ, the complexity of the Fano algorithm is slightly higher.
However, the higher complexity is only relevant in cases where
many errors occur, since Fano backtracks more often than
in cases with fewer errors. For a small number of errors,
ZJ is practically not faster, since it uses a stack which must
be maintained during the decoding is performed. The Fano
algorithm is said to be the most practical sequential decoding
algorithm [16, Section 6.10].
Sequential Decoding belongs to the class of backtracking
algorithms which perform a depth first search in a tree. The
Fano algorithm traverses the code tree from the root until a
leaf node is reached. At each node, the algorithm can move
either to one of its child nodes, or to its root. The decision
is based on the Fano metric: The bit metric for a BSC with
crossover probability pi and rate R is given by
(
log2 2pi − R
if yi 6= ci ,
M (yi |ci ) =
(3)
log2 2(1 − pi ) − R if yi = ci .
The metric of a new node is calculated by adding the metric
of the corresponding edge to the metric of the current node.
1
relative number of ignored bits
stores the maximum reached metric for the node as its metric
and adds the edge which contributes to this metric to its
survivor path. To circumvent error propagation, often m tail
bits are inserted after L information bits to reset the register
(cf. Figure 2, segments 5 and 6). The bold path in Figure 2
is the survivor, which gives us c with p(y|c) maximized. The
complexity of Viterbi increases exponentially in µ and k, and
linearly in the number of info bits.
There is also a list decoding variant by [18]. Using this
modification of the Viterbi algorithm, each node stores the
difference between the metric of the survivor and the second
best path in addition to the metric. When the maximum
likelihood path was determined, the minimum of the metric
differences of all nodes on the ML path is selected, since the
most unreliable decision was taken at the corresponding node.
At that node, we take the non-surviving edge to differ from the
surviving path, and follow the second path backwards until it
merges again with the survivor path. By recursively applying
this decision process, we can generate the L most likely code
sequences.
0.8
0.6
0.4
0.2
0
0
0.1
0.2
0.3
0.4
0.5
pT
Fig. 3. Relative number of ignored bits for a given threshold pT
In each iteration, the algorithm examines which of the child
nodes of the current node maximizes the metric (3), moves to
that node and updates the metric. This process continues, until
the metric falls below a threshold T , what happens when it
becomes unlikely that the correct path was chosen. In that case,
the next best path starting at the current node is examined.
If all paths beginning at the current node are examined and
the metric becomes lower than the threshold in all cases,
the algorithm backtracks to the previous level. If the metric
value of that preceding node also is below the threshold, the
threshold is decreased by a threshold increment ∆. Every time
a node is visited the first time, the threshold is increased by
the maximum multiple of ∆ such that the metric value is still
larger than the threshold.
More information about sequential decoding with full length
examples can be found in [15, Chapter 13].
IV. S ETUP
AND
R ESULTS
In a first approach we extract t bits from a PUF. Using the
model described in Section IV-B, each of these t bits can be
modeled as a BSC with crossover probability pi (i = 1, ..., t).
We select s ≤ t Bits ri (i = 1, ..., s) which belong to BSCs
with crossover probabilities pi < pT , where pT is a chosen
threshold. Thus, unreliable PUF bits are ignored.
In a second step, we replace Viterbi by Fano algorithm,
which is a sequential decoding approach, in order to gain
performance for codes with large constraint lengths.
Figure 3 shows the amount of wasted bits in our first
approach. In a third step we use all extracted bits without
refusing the unreliable ones. Our goal is to study, whether
the good channels compensate the bad ones when using soft
information at the input of the decoder. For a fair comparison
we have to adjust the code rate for this scenario.
A. Coding Theoretical Concepts
Instead of using the SRAM reliability information only
for choosing reliable bits from the PUF response, we can
additionally translate them into soft information which can
to further improve the results. For our experiments we used
information sequences of length 256 and responses of length
512 + µ. We want to emphasize, that in contrast to [9] we do
not use ROVA.
TABLE I
C ONVOLUTIONAL CODES USED IN THIS WORK .
k
1
1
1
1
1
1
1
1
1
1
1
n
2
2
2
2
2
2
3
3
3
3
3
µ
2
6
7
10
14
16
6
7
8
9
10
polynomials used (in octal)
g2
g1
g0
5
7
133
171
247
371
3645
2671
63057
44735
313327
231721
133
165
171
225
331
367
557
663
711
1117
1365
1633
2353
2671
3175
1) Full Soft Information: In a first experiment, we use the
same scenario as in [9], using thresholds pT ∈ {0.3, 0.2, 0.1}.
Instead of hard decision decoding, we use full soft information
at the input of the Viterbi algorithm.
be used at the input of Viterbi’s algorithm to improve the
decoding results. We calculate soft information
(j)
si
(j)
= (−1)hi ⊕ri · (log(1 − pei ) − log(pei ))
(4)
for cell i at time instance j, where ri is the i-th response
bit and pei is the error probability of SRAM cell i. This
formula is already given in [12], however the authors use
short linear block codes with soft decision decoding instead of
convolutional codes. We obtain pei for all cells i by using the
model described in Section II. In a practical scenario, the soft
information can be obtained during the enrollment phase by
doing multiple measurements for every individual PUF[12],
[10]. A technique how to obtain soft information from one
single measurement was described in [23].
The simplified ROVA used in [9], has the drawback, that it
calculates error probabilities for each symbol in order to compute the total error probability for the whole code sequence.
As alternative to the calculation of error probabilities, we use a
list decoding variant of the Viterbi algorithm according to [18],
which determines the L most reliable paths. After computing
a list, we can obtain reliability values by calculating
n
Y
p(ci |r)
p(r|ci )
qi
=
, where p(ci |r) =
p(r|cj )
p(cj |r)
i=1
(
pi
, ri 6= ci
and qi =
.
1 − pi , ri = ci
(5)
(6)
for all candidate solutions ci , cj and hence decide for the most
likely code sequence.
Table I lists all convolutional codes used within this work.
Length of the used information sequence is 256 Bits with
termination.
B. Simulation Results
The results presented in this section are based on the
SRAM model introduced in [12] and discussed in Section II.
We perform simulations to compare our ideas to the results
in [9]. According to [9], we aim for an error probability
in the order of 10−6 . For comparability we use (2, 1, [6])
and (2, 1, [7]) convolutional codes. Additionally, we use rate
1
2 convolutional codes with memory length 10, 14, and 16
The results in Table II show, that using soft information as
input results in smaller error probabilities compared to [9].
This becomes clear when comparing Column 4 and Column 7
in Table II. E.g., using a (2,1,[7]) code with pT = 0.1, [9]
achieved Perr = 4.86·10−3, using full soft information results
in Perr = 7.63 · 10−4 .
2) Full Soft Information and List Decoding: Applying list
decoding allows to increase pT without loss. This can be seen
when studying for example the (2,1,[6]) code in Table II. In
[9] a key error probability of 1.08 · 10−2 was obtained using
pT = 0.1. Using list decoding with list size L = 4 results
in a key error probability of the same order, however we can
increase pT to 0.2. Since more PUF bits can be used for the
key, we can save 13 of the extracted bits, as can be seen in
Figure 3.
3) Additional Use of Multiple Readouts: The goal in this
section is to further decrease the error probabilities gained in
Section IV-B2, by additionally using multiple readouts as done
in [9]. Table III presents the error probabilities obtained from
simulations with list size L and using m readouts. Applying
multiple readouts results in an essential improvement. For
example, considering a (2,1,[6]) code with p = 0.1, using full
soft information with list size 3 and m readouts is as good as
using m+ 1 readouts in [9]. We can also use the same number
of multiple readouts as in [9] and do not use list decoding to
result in the desired error probability 10−6 . For convolutional
codes with larger memory lengths, pT can be larger in order
to achieve Perr ≈ 10−6 .
Figure 4 compares the key error probability of the (2,1,[6])
code from [9] with our soft information input and list decoding
approach for different values of pT . Note that the threshold
measure used in [9], pmax , is specific for the Differential Sequence Coding (DSC) scheme implemented in that reference,
and hence is translated to pT in our tables. Figure 5 compares
convolutional codes with different memory length, using soft
information input, list size 3 and multiple readouts.
4) Sequential Decoding: Our results using the Viterbi algorithm show that the error probability depends on the memory
length, which is a well-known fact in coding theory. As
indicated by the tables, to reach an error probability of 10−6 ,
a comparably large memory length is needed. However, the
complexity of the Viterbi decoder depends exponentially on
the memory and becomes infeasible for µ ≈ 20 in case of
k = 1 and even lower values for k > 1. This drawback can
be circumvented by using sequential decoding as discussed in
Section III-C. Table IV shows our results.
U SING RATE
TABLE II
C OMPARISON OF [9] WITH HARD DECISION DECODING (HD) AND SOFT DECISION DECODING (SD), USED WITH LIST SIZE L.
E XTR . MEANS THE NUMBER OF SIMULATED KEY EXTRACTIONS . ∗ ONLY 500.000 EXTRACTIONS WERE USED IN [9].
1
CODES .
2
Code
(2,1,[6])
pT
0.3
0.2
0.1
0.3
0.2
0.1
0.3
0.2
0.1
0.3
0.2
0.1
0.3
0.2
0.1
(2,1,[7])
(2,1,[10])
(2,1,[14])
(2,1,[16])
Extractions
10.000.000
10.000.000
10.000.000
10.000.000
10.000.000
10.000.000
10.000.000
10.000.000
10.000.000
500.000
500.000
500.000
500.000
500.000
500.000
Ref.[9]
4.11e-01∗
6.98e-02∗
≤1.98e-02
–
–
≤4.86e-03
–
–
–
–
–
–
–
–
–
HD, L = 3
6.72e-01
1.83e-01
1.04e-02
4.98e-01
8.37e-02
2.68e-03
3.05e-01
2.66e-02
3.20e-04
1.20e-01
3.37e-03
1.60e-05
1.56e-03
<2e-06
<2e-06
HD, L = 4
6.66e-01
1.80e-01
1.02e-02
4.91e-01
8.18e-02
2.62e-03
2.98e-01
2.57e-02
3.09e-04
1.16e-01
3.24e-03
1.60e-05
1.5e-03
1.5e-03
<2e-06
SD, L = 1
1.86e-01
3.62e-02
2.15e-03
1.18e-01
1.80e-02
7.63e-04
2.97e-02
2.28e-03
3.90e-05
4.17e-03
8.60e-05
<2e-06
2.00e-05
<2e-06
<2e-06
SD, L = 3
1.29e-01
2.09e-02
1.09e-03
7.92e-02
1.01e-02
3.80e-04
1.85e-02
1.21e-03
1.80e-05
2.54e-03
4.20e-05
<2e-06
<2e-06
<2e-06
<2e-06
SD, L = 4
1.27e-01
2.05e-02
1.07e-03
7.74e-02
9.90e-03
3.73e-04
1.79e-02
1.17e-03
1.80e-05
2.45e-03
4.00e-05
<2e-06
<2e-06
<2e-06
<2e-06
TABLE III
A DDING THE CONCEPT OF USING m READOUTS TO THE TECHNIQUES OF SOFT INFORMATION AND LIST DECODING . U SING RATE
Code
pT
Extractions
(2,1,[6])
0.3
500.000
10.000.000
500.000
10.000.000
10.000.000
10.000.000
10.000.000
10.000.000
10.000.000
10.000.000
10.000.000
0.2
0.1
0.3
0.2
0.1
0.3
0.2
0.1
(2,1,[7])
(2,1,[10])
Ref.[9]
m=1
4.11e-01
–
4.11e-01
–
≤1.98e-02
–
–
≤4.86e-03
–
–
–
Ref.[9]
m=3
–
–
–
–
7.6e-07
–
–
<1e-07
–
–
–
HD, L = 3
m=3
6.32e-01
6.32e-01
1.84e-01
1.84e-01
3.27e-02
4.89e-01
1.05e-01
2.66e-02
3.25e-01
6.43e-02
3.98e-02
SD, L = 1,
m=3
7.81e-03
7.86e-03
5.40e-05
6.10e-05
<1e-07
2.19e-03
8.00e-06
<1e-07
4.70e-05
<1e-07
<1e-07
SD, L = 3,
m=2
3.58e-02
3.59e-02
1.35e-03
1.36e-03
6.00e-06
1.49e-02
3.44e-04
1.00e-06
1.02e-03
5.00e-06
<1e-07
1
CODES .
2
SD, L = 3,
m=3
4.93e-03
4.91e-03
3.40e-05
3.20e-05
<1e-07
1.32e-03
4.00e-06
<1e-07
2.85e-04
<1e-07
<1e-07
100
100
10-1
10-1
key error probability
key error probability
10-2
10-2
10
-3
10-4
10-4
10-5
10-6
10-5
10-6
0.3
10-3
0.28
0.26
0.24
0.22
0.2
0.18
0.16
0.14
0.12
0.1
10-7
0.3
0.28
0.26
0.24
0.22
0.2
0.18
0.16
0.14
0.12
0.1
Fig. 4. (2, 1, [6]) code with soft information at the input, list decoding, and
m readouts.
Fig. 5. Comparison of convolutional codes with different memory lengths,
soft input information, list decoding with list size L, and m readouts.
TABLE IV
P ERFORMANCE OF SEQUENTIAL DECODING USING FANO ’ S A LGORITHM
WITH SOFT INFORMATION INPUT, L = 1, AND pT = 0.3
5) Without Refusing Unreliable Bits: A drawback of the
approaches described so far is, that not all response bits are
used. Hence more bits than needed have to be extracted from
the PUF, what can be time- and area-consuming (cf. Figure 3).
Code
(2,1,[16])
(2,1,[20])
(2,1,[24])
Extractions
10.000.000
10.000.000
10.000.000
SD2
∼ 10−5
∼ 10−6
∼ 10−7
2 Due to time restrictions, the numbers in the table are estimations based
on simulations with similar parameters. Simulations will have finished until
the final deadline and numbers in the table will be replaced.
A PPLYING A
TABLE V
1
RATE 3 CONVOLUTIONAL CODE , USING ALL RESPONSE BITS , EVEN THE UNRELIABLE ONES . T HE PARAMETERS µ, L, m DENOTE MEMORY
LENGTH , LIST SIZE AND NUMBER OF READOUTS RESPECTIVELY.
Code
(3,1,[6])
(3,1,[7])
(3,1,[8])
(3,1,[9])
(3,1,[10])
Extractions
10.000.000
10.000.000
10.000.000
10.000.000
10.000.000
SD, L = 1
3.98e-02
1.73e-02
9.72e-03
5.07e-03
2.30e-03
HD, L = 3
6.59e-01
5.93e-01
5.09e-01
4.28e-01
3.39e-01
Instead of refusing response bits with an error probability
higher than a certain threshold pT , we can use all response bits
and use their reliability information as soft input to a Viterbi
decoder. The idea is, that the soft information of reliable
bits compensates unreliable bits. As seen in Figure 3, using
pT ≈ 0.2 results in ≈ 13 of ignored bits. For a fair comparison
with [9], we therefore use rate 31 instead of rate 12 codes.
The results in Table V show that this method significantly
improves the previous approaches. This is obvious by comparing the columns “SD, L = 1” in Tables II and V. For
example we can realize, that the results for (2,1,[6]), (2,1,[7]),
and (2,1,[10]) for pT = 0.2 in Table II coincide with the
corresponding results in Table V. We want to emphasize, that
this implies advantages to the previous results. First, we do
not ignore extracted PUF bits, second, ROVA can be avoided.
V. C ONCLUSION
This work studied how previous approaches that use convolutional codes for PUFs can be improved by using soft
information, list decoding and sequential decoding. Our results
show that the selection of reliable bits from a PUF response
is not needed. Reliability values of the individual response
bits are translated into soft information which is used at
the input of Viterbi. Using these soft information, reliable
channels compensate unreliable ones. A rate 31 convolutional
code of memory length 7 is sufficient in order to result in
a key error probability of order 10−6 . The obtained error
probabilities can further be decreased by increasing memory
length or using a larger list size for list decoding. Using a
convolutional code with more than 8 memory elements, the
error probability becomes smaller than 10−7 . Using sequential
decoding decreases decoding complexity when using codes
with large memory length. To summarize, our strategy has
two advantages compared to previous approaches. First, all
response bits are used. Second, we avoid overhead since our
work does not require ROVA.
R EFERENCES
[1] J.-P. Linnartz and P. Tuyls, “New Shielding Functions to Enhance
Privacy and Prevent Misuse of Biometric Templates,” in Audio- and
Video-based Biometric Person Auth. Springer, 2003, pp. 393–402.
[2] Y. Dodis, L. Reyzin, and A. Smith, “Fuzzy Extractors: How to Generate
Strong Keys from Biometrics and other Noisy Data,” in Advances in
Cryptology-Eurocrypt. Springer, 2004, pp. 523–540.
SD, L = 3
2.24e-02
9.43e-03
5.14e-03
2.65e-03
1.17e-03
SD, L = 3, m = 3
5.70e-05
6.00e-06
1.00e-06
<1e-07
<1e-07
[3] M. Hiller, M. Weiner, L. Rodrigues Lima, M. Birkner, and G. Sigl,
“Breaking Through Fixed PUF Block Limitations with Differential
Sequence Coding and Convolutional Codes,” in TrustED. ACM, 2013,
pp. 43–54.
[4] S. Müelich and M. Bossert, “A New Error Correction Scheme for
Physical Unclonable Functions,” in ITG SCC, 2017.
[5] C. Bösch, J. Guajardo, A.-R. Sadeghi, J. Shokrollahi, and P. Tuyls,
“Efficient helper data key extractor on fpgas,” Cryptographic Hardware
and Embedded Systems–CHES 2008, pp. 181–197, 2008.
[6] R. Maes, A. Van Herrewege, and I. Verbauwhede, “Pufky: A fully functional puf-based cryptographic key generator,” Cryptographic Hardware
and Embedded Systems–CHES 2012, pp. 302–319, 2012.
[7] S. Müelich, S. Puchinger, M. Bossert, M. Hiller, and G. Sigl, “Error
correction for physical unclonable functions using generalized concatenated codes,” Int. Workshop on Algebraic and Combinatorial Coding
Theory, 2014.
[8] S. Puchinger, S. Müelich, M. Bossert, M. Hiller, and G. Sigl, “On
error correction for physical unclonable functions,” in SCC 2015; 10th
International ITG Conference on Systems, Communications and Coding;
Proceedings of. VDE, 2015, pp. 1–6.
[9] M. Hiller, A. G. Önalan, G. Sigl, and M. Bossert, “Online Reliability
Testing for PUF Key Derivation,” in Proc. of the 6th Int. Workshop on
Trustworthy Embedded Devices. ACM, 2016, pp. 15–22.
[10] R. Maes, P. Tuyls, and I. Verbauwhede, “Low-overhead implementation
of a soft decision helper data algorithm for sram pufs.” in CHES, vol. 9.
Springer, 2009, pp. 332–347.
[11] J. Guajardo, S. S. Kumar, G.-J. Schrijen, and P. Tuyls, “FPGA Intrinsic
PUFs and their Use for IP Protection,” in CHES. Springer, 2007, pp.
63–80.
[12] R. Maes, P. Tuyls, and I. Verbauwhede, “A Soft Decision Helper Data
Algorithm for SRAM PUFs,” in IEEE ISIT. IEEE, 2009, pp. 2101–
2105.
[13] P. Elias, “Coding for Noisy Channels,” in Proceedings of the Institute
of Radio Engineers, no. 3, 1955, pp. 37–46.
[14] M. Hiller, L. R. Lima, and G. Sigl, “Seesaw: An Area-Optimized FPGA
Viterbi Decoder for PUFs,” in 17th Euromicro Conference on Digital
System Design. IEEE, 2014, pp. 387–393.
[15] D. Costello and S. Lin, “Error Control Coding,” New Jersey, 2004.
[16] R. Johannesson and K. Zigangirov, “Fundamentals of Convolutional
Codes,” 1999.
[17] A. Viterbi, “Error Bounds for Convolutional Codes and an Asymptotically Optimum Decoding Algorithm,” IEEE Transactions on Information
Theory, vol. 13, no. 2, pp. 260–269, 1967.
[18] G. Schmidt, V. Sidorenko, V. V. Zyablov, and M. Bossert, “Finding a
List of Best Paths in a Trellis,” in IEEE ISIT, 2004.
[19] J. M. Wozencraft, “Sequential Decoding for Reliable Communication,”
1957.
[20] K. Zigangirov, “Some Sequential Decoding Procedures,” Problemy
Peredachi Informatsii, vol. 2, no. 4, pp. 13–25, 1966.
[21] F. Jelinek, “A Fast Sequential Decoding Algorithm Using a Stack,” IBM
Journal of Research and Development, vol. 13, no. 6, pp. 675–685, 1969.
[22] R. Fano, “A Heuristic Discussion of Probabilistic Decoding,” IEEE
Transactions on Information Theory, vol. 9, no. 2, pp. 64–74, 1963.
[23] V. Van der Leest, B. Preneel, and E. Van der Sluis, “Soft decision error
correction for compact memory-based pufs using a single enrollment.”
in CHES. Springer, 2012, pp. 268–282.
| 7cs.IT
|
Inverse monoids associated with the complexity class NP
J.C. Birget
arXiv:1703.02519v1 [math.GR] 7 Mar 2017
6 March 2017
Abstract
We study the P versus NP problem through properties of functions and monoids, continuing
the work of [3]. Here we consider inverse monoids whose properties and relationships determine
whether P is different from NP, or whether injective one-way functions (with respect to worst-case
complexity) exist.
1
Introduction
We give a few definitions before motivating the monoid approach to the P versus NP problem. Some
of these notions appeared already in [3].
By function we mean a partial function A∗ → A∗ , where A is a finite alphabet (usually, A = {0, 1}),
and A∗ denotes the set of all finite strings over A. Let Dom(f ) (⊆ A∗ ) denote the domain of f , i.e.,
{x ∈ A∗ : f (x) is defined}; and let Im(f ) (⊆ A∗ ) denote the image (or range) of f , i.e., {f (x) : x ∈
Dom(f )}. The length of x ∈ A∗ is denoted by |x|. The restriction of f to X ⊆ A∗ is denoted by f |X ,
and the identity function on X is denoted by idX or id|X .
Definition 1.1 (inverse, co-inverse, mutual inverse) A function f ′ is an inverse of a function
f iff f ◦ f ′ ◦ f = f . In that case we also say that f is a co-inverse of f ′ . If f ′ is both an inverse and
a co-inverse of f , we say that f ′ is a mutual inverse of f , or that f ′ and f are mutual inverses (of
each other).1
It is easy to see that f ′ is an inverse of f iff for every y ∈ Im(f ): f ′ (y) is defined and f ′ (y) ∈ f −1 (y).
Hence, in particular, Im(f ) ⊆ Dom(f ′ ). Also, f ′ is an inverse of f iff f ◦ f ′ |Im(f ) = idIm(f ) .
An element f of a semigroup S is called regular (or more precisely, regular in S) iff there exists
f ′ ∈ S such that f f ′ f = f ; in that case, f ′ is called an inverse of f (more precisely, an inverse in S of
f ), and f is called a co-inverse of f ′ (in S). A semigroup S is called regular iff all the elements of S
are regular in S; see e.g. [6, 5]. An inverse semigroup is, by definition, a semigroup S such that every
element of S has exactly one mutual inverse in S; see e.g. [9, 6, 5].
The connection between the P versus NP problem and inverses comes from the following well
known characterization: P 6= NP iff there exists a function that is computable in polynomial time
(by a deterministic Turing machine) and that is polynomially balanced, but that does not have an
inverse that is computable in polynomial time; see e.g. [8] p. 33, and Section 1 of [3]. A function f :
A∗ → A∗ is called polynomially balanced iff there exists a polynomial p such that for all x ∈ Dom(f ):
|f (x)| ≤ p(|x|) and |x| ≤ p(|f (x)|). Functions that are polynomial-time computable and polynomially
balanced, but that have no inverse of that type, are one-way functions in the sense of worst-case
complexity.2 In [3] this characterization of the P versus NP problem was reformulated in terms of
regularity of semigroups, by introducing the monoid fP, defined as follows.
1
The terminology about inverses varies. In semigroup theory, f ′ such that f f ′ f = f is called a semi-inverse or a
pseudo-inverse of f , in numerical mathematics f ′ is called a generalized inverse, in ring theory and in category theory it
is called a weak inverse or a von Neumann inverse. In semigroup theory the term “inverse” of f is only used if f ′ f f ′ = f ′
holds in addition to f f ′ f = f . Co-inverses don’t seem to have a name in the literature.
2
One-way functions in this sense are not necessarily useful for cryptography, but they are relevant for P vs. NP.
1
Definition 1.2.
fP = {f : f is a function A∗ → A∗ that is polynomially balanced, and
computable by a deterministic polynomial-time Turing machine}.
In particular, when f ∈ fP then Dom(f ) is in P. The set fP is closed under composition of functions,
and the identity function is in fP, so fP is a monoid. The above characterization of P 6= NP now
becomes:
P 6= NP iff fP is a non-regular monoid.
We will use some more definitions that apply to functions.
For any function f : X → Y , the equivalence relation modf on Dom(f ) is defined by: x1 modf x2
iff f (x1 ) = f (x2 ). The set of equivalence classes of modf is {f −1 (y) : y ∈ Im(f )}.
A choice set for f is, by definition, a subset of Dom(f ) that contains exactly one element of each
modf -class. A choice function for f is, by definition, any inverse f ′ of f such that Dom(f ′ ) = Im(f ).
A choice function f ′ maps each element y ∈ Im(f ) injectively to an element of f −1 (y), and every
modf -class contains exactly one element of Im(f ′ ); so Im(f ′ ) is a choice set. For a choice function f ′
we have f ◦ f ′ = f ◦ f ′ |Im(f ) = idIm(f ) , hence f f ′ f = f ; we also have f ′ f f ′ = f ′ ; more generally, if
f1′ , f2′ are two choice functions for the same function f then f2′ f f1′ = f2′ . A choice function for f is
uniquely determined by f and a choice set.
A representative choice function for f is, by definition, any function r such that Im(r) ⊆ Dom(r) =
Dom(f ), Im(r) is a choice set for f , modr = modf , and r|Im(r) = idIm(r) . Thus, r maps each modf -class
[x]f to one and the same chosen element of [x]f . It follows that r is an idempotent; and if r1 , r2 are
two representative choice functions for the same function f then r2 ◦ r1 (.) = r2 (.). If f ∈ fP and if in
addition r ∈ fP, then f ≡L r (by Prop. 2.1 in [3]). So in that case f is a regular element of fP, since it
is L-equivalent to an idempotent. Hence, if every function in fP had a representative choice function
in fP, then fP would be regular, which would imply P = NP.
There is a one-to-one correspondence between the choice functions of f and the representative
choice functions: If c is a choice function then ρc (.) = c ◦ f (.) is a representative choice function; and
if r is a representative choice function of f then vr (.) = r ◦ f −1 (.) is a choice function (where f −1 is
the inverse relation of f , and ◦ is composition of relations). Moreover, if r is a representative choice
function of f then ρvr = r; and if c is a choice function of f then vρc = c.
For functions in general, the existence of a choice function, the existence of a representative choice
function, and the existence of a choice set, are equivalent to the axiom of choice. Because of this
connection we are especially interested in inverses f ′ of f that are choice functions, i.e., that satisfy
Dom(f ′ ) = Im(f ).
These three formulations of the axiom of choice can also be considered for fP: (1) “Every f ∈ fP
has an inverse in fP.” Equivalently, “f has a choice function in fP”, and also equivalently, “f has a
mutual inverse in fP”. (2) “Every f ∈ fP has a representative choice function in fP.” (3) “Every f ∈ fP
has a choice set in P.” It is an open problem problem whether the three statements are true.
We saw that (1) is true iff P = NP. Moreover, (1) implies (2); indeed, if f has an inverse f ′ ∈ fP then
′
f f is a representative choice function in fP. And (2) implies (3); indeed, if r ∈ fP is a representative
choice function of f , then r is regular (being an idempotent), hence Im(r) belongs to P by Prop. 1.9 in
[3]; and Im(r) is a choice set for r and for f . It remains an open problem whether other implications
between (1), (2), (3) hold. Nevertheless, we have the following.
Proposition 1.3 There exists an injective one-way function (for worst-case complexity) iff there
exists a one-way function that has a choice set in P.
Proof. If g ∈ fP is injective then Dom(g) is a choice set for g, and Dom(g) ∈ P. If g is also one-way,
then it is therefore a one-way function with a choice set in P.
The converse follows immediately from the next Lemma. ✷
2
Lemma 1.4 If f ∈ fP has a choice set C ∈ P, but f is not regular (i.e., (3) holds but (1) does not),
then the restriction f |C is an injective one-way function.
Proof. Since f ∈ fP and C ∈ P then we have f |C ∈ fP. Moreover, if C is a choice set for f then f |C is
injective. And if C is a choice set, any inverse of f |C is also an inverse of f . Hence f |C has no inverse
in fP (since f is not regular). ✷
The motivation for this paper is based on the following simple observation: If f : A∗ → A∗ is any
partial function and if f ′ is an inverse of f , then the restriction f ′ |Im(f ) is an injective inverse of f .
Moreover, if f ∈ fP and f ′ ∈ fP, then Im(f ) ∈ P (by Prop. 1.9 in [3]); hence f ′ |Im(f ) ∈ fP. Thus we
have: P 6= NP iff there exists f ∈ fP such that f has no injective inverse in fP.
The classes P and NP are defined in terms of sets of strings (“formal languages”). To add more
structure we use functions, and we characterized P and NP by properties of monoids of functions [3].
Our next step, in the present paper, is to characterize P versus NP by properties of inverse monoids,
and of groups. However, whether this approach will help solve the P versus NP problem, remains to
be seen.
Overview: In Section 2 we introduce the monoid invfP consisting of all injective regular functions
in fP. We prove that every regular function in fP has an inverse in invfP, and that invfP is a maximal
inverse submonoid of fP.
In Section 3 we show that the polynomial-time polynomially balanced injective Turing machines
form a machine model for invfP, i.e., that a function f belongs to invfP iff f is computed by such a
Turing machine. We conclude from this that invfP is finitely generated. We also consider polynomialtime polynomially balanced injective Turing machines with an NP-oracle, and we show that the set
invfP(NP) of functions computed by such Turing machines is a finitely generated inverse monoid. To
prove the latter, we show that there exist languages that are NP-complete with respect to one-one
reductions in invfP. We show that every function in fP has an inverse in invfP(NP) .
In Section 4 we show that invfP 6= invfP(NP) iff P 6= NP. We introduce the set cofP of all functions
in invfP(NP) that have an inverse in fP, i.e., that are a co-inverse of a function in fP. We show that this
is a finitely generated monoid. We prove that P 6= NP iff invfP 6= cofP, iff cofP 6= invfP(NP) , iff cofP is
not regular.
We also introduce the monoid injfP of all injective functions in fP; injfP is also equal to invfP(NP) ∩
fP. We show that the inverse monoid invfP is the set of the regular elements of injfP. Hence, injective
one-way functions (for worst-case complexity) exist iff invfP 6= injfP. We do not know whether injfP is
finitely generated; if it turns out that injfP is not finitely generated then P 6= NP.
In Section 5 we show that every element of fP is equivalent (with respect to inversive-reductions)
to an element of fP that has an inverse in a subgroup of invfP(NP) .
Notation for the monoids used, and their definition:
fP
invfP
invfP(NP)
cofP
injfP
Definition
Definition
Definition
Definition
Definition
1.2
2.2
3.6
4.10
4.12
(polynomially balanced polynomial-time computable functions)
(injective functions in fP that have an inverse in fP)
(functions computed by injective Turing machines with NP-oracle)
(functions in invfP(NP) that have an inverse in fP)
(injective functions in fP)
The relation between these monoids is shown in Figure 1 (after Prop. 4.13).
2
Injective inverses of regular elements
For any function f : A∗ → A∗ , we will be interested in inverses f ′ of f that satisfy Dom(f ′ ) = Im(f );
such inverses are necessarily injective. Moreover, when Dom(f ′ ) = Im(f ) we not only have f f ′ f = f ,
but also f ′ f f ′ = f ′ . So f and f ′ are mutual inverses.
3
An injective inverse f ′ , as above, has exactly one inverse f ′′ that satisfies Dom(f ′′ ) = Im(f ′ ), namely
= f ′−1 (the set-theoretic inverse function of f ′ ); moreover, Im(f ′−1 ) = Im(f ). And f ′−1 = f |Im(f ′ ) ,
i.e., the restriction of f to the choice set Im(f ′ ) of f ′ .
Here are some more simple facts about inverses: If f ′ is a mutual inverse of f then Im(f ′ ) ⊆
Dom(f ). Moreover, f ′ is a mutual inverse of f iff Im(f ′ ) = Im(f ′ |Im(f ) ). Note that f ′ |Im(f ) is the
choice function determined by f ′ , so Dom(f ′ |Im(f ) ) = Im(f ). Thus an inverse f ′ is a mutual inverse of
f iff Im(f ′ ) is the choice set determined by f ′ in Dom(f ). As a consequence, a mutual inverse f ′ of
f is injective iff Dom(f ′ ) = Im(f ).
f ′′
Proposition 2.1 If g1 and g2 are injective regular elements of fP, then g2 ◦ g1 is also injective and
regular in fP.
Proof. The composite of injective functions is obviously injective. Since gi is regular (for i = 1, 2),
it has the injective regular function gi−1 as an inverse. Indeed, Im(gi ) ∈ P, and gi−1 = gi′ |Im(gi ) for any
inverse gi′ ∈ fP of gi ; hence, gi−1 ∈ fP. Thus, the injective function g1−1 ◦ g2−1 is a mutual inverse of
g2 ◦ g1 , and Dom(g1−1 ◦ g2−1 ) = Im(g2 ◦ g1 ). ✷
Note that in general, the product of regular elements in fP need not be regular (unless P = NP); but
by the above proposition, the product of injective regular elements is regular.
Definition 2.2 Let invfP denote the set of injective regular elements of fP.
Corollary 2.3 The set invfP of injective regular elements of fP is an inverse monoid. Every regular
element of fP has an inverse in invfP.
✷
Corollary 2.3 implies that P = NP iff every element of fP has an inverse in invfP.
Another motivation for invfP will be seen in Prop. 3.11, where is is shown that there exist languages
that are NP-complete with respect to one-one reductions in invfP.
A monoid M1 is called an inverse submonoid of a monoid M iff M1 is submonoid of M , and M1 is
an inverse monoid by itself. So, an element of M1 has exactly one mutual inverse in M1 , but it could
have additional mutual inverses in M .
Proposition 2.4 The inverse submonoid invfP is a maximal inverse submonoid of fP; i.e., if M is
an inverse monoid such that invfP ⊆ M ⊆ fP, then invfP = M .
Proof. Let M be an inverse submonoid such that invfP ⊆ M ⊆ fP. For a contradiction, let us assume
that invfP 6= M . Then M contains a non-injective element f , which is regular (since M is inverse).
Since f is regular, f has a choice function f1′ ; since f is non-injective, f has at least one other choice
function f2′ .
Moreover, since f1′ ∈ invfP, we can pick f2′ in such a way that f2′ ∈ invfP too. Indeed, let x1 , x2 ∈
Dom(f ) be such that f (x1 ) = f (x2 ), and x1 is in the choice set Im(f1′ ), and x2 6∈ Im(f1′ ). Let τ be the
transposition of x1 and x2 (and τ is the identity elsewhere). Then f2′ = f1′ ◦ τ belongs to invfP, and is
the same as f1′ , except that x2 has replaced x1 in the choice set.
Since the choice functions f1′ , f2′ belong to invfP (⊆ M ), f has two mutual inverses in M , which
implies that M is not an inverse monoid. ✷
Remarks. (1) Prop. 2.4 holds, whether fP is regular or not.
(2) An inverse submonoid of fP need not consist of injective functions only, and the submonoid invfP is
not the only maximal inverse submonoid of fP. E.g., fP contains non-injective idempotents, and these
are contained in maximal inverse monoids that are different from invfP (since they contain non-injective
elements).
Proposition 2.5 A maximal subgroup of fP, and more generally, any L-class of fP, is either disjoint
from invfP or entirely contained in invfP.
4
Proof. An L-class of fP that intersects invfP contains injective elements, hence it consist entirely of
injective elements (by Prop. 2.1 in [3]). Also, an L-class of fP that intersects invfP contains regular
elements, hence it consist entirely of regular elements (it is a well-known fact from semigroup theory
that if a D-class contains regular elements then it consists entirely of regular elements; see e.g. [5, 6]).
Hence, this L-class consists entirely of regular injective elements, hence it is contained in invfP. ✷
On the other hand, every regular R-class of fP intersects both invfP and fP − invfP. Indeed, for
every regular element f ∈ fP we have f ≡R idIm(f ) (and Im(f ) is in P when f is regular). But a
regular R-class always contains some non-injective elements (by Prop. 2.1 in [3]).
3
Computing inverses
There is a machine model that exactly characterizes the injective inverses of the regular elements of
fP, namely the polynomially balanced polynomial-time injective Turing machines. As a consequence
we will prove that invfP is finitely generated.
Turing machines of the above type, with an NP-oracle added, form a model of computation for
another inverse monoid, invfP(NP) (given in Def. 3.6), which contains some injective inverses for every
element of fP.
3.1
Injective Turing machines
There is a very simple machine model for the elements of invfP, namely the polynomially balanced
polynomial-time injective Turing machines. A deterministic Turing machine is called injective iff the
transition table of the Turing machine describes a (finite) injective function; for details, see [1, 2].
The reverse of a Turing machine M is the machine obtained by reversing every transition of M
(and also switching start and accept states). The reverse of a deterministic Turing machine M is is
not deterministic, unless M is injective.
Remark. In the literature, injective Turing machines are called “reversible”, because of historic
connections with the study of computation as a “reversible process” (in the sense of thermodynamics).
However, calling an injective Turing machine “reversible” can be misleading, since the transitions
obtained by reversing an injective Turing machine M are not part of M (but of a different machine).
Proposition 3.1.
(1) The reverse of a polynomially balanced polynomial-time injective Turing machine is also a polynomially balanced polynomial-time injective Turing machine. If a function f is polynomially balanced
and it is computed by a polynomial-time injective Turing machine, then f −1 is also polynomially balanced and it is computed by a polynomial-time injective Turing machine.
(2) There exists a polynomial-time injective Turing machine (that is not polynomially balanced)
whose reverse does not have polynomial time-complexity. There exists an injective function g that is
computed by a polynomial-time injective Turing machine, whose inverse g−1 is not computable by a
polynomial-time Turing machine.
Proof. (1) An injective Turing machine M , computing an injective function f , can be run in reverse.
This yields a new injective Turing machine M ′ , computing f −1 . If M runs in polynomial time, with
polynomial pT (.), and it is polynomially balanced, with polynomial pB (.), then M ′ is also polynomially
balanced, with polynomial pB (.) too, and runs in polynomial time, with polynomial bound pT ◦ pB (.).
(2) An example of such a function is
m
g : a(2 ) 7−→ am , for all m ∈ N,
where a is a fixed letter; g is undefined on any input that is not of the form ak with k a power of 2.
Obviously, g is injective and polynomial-time computable, but not polynomially balanced. Hence,
g−1 is not computable in polynomial time, since its output is exponentially longer than its input.
However, g is computable in linear time by the following injective Turing machine: We use a Turing
5
machine with a rubber tape for the input, and an ordinary tape for the output. A rubber tape is a tape
on which one can not only replace one letter by another one (as on an ordinary tape), but where one
also can insert a letter or remove a letter (in one transition). A rubber tape can easily be simulated
by two stacks; a stack is a special case of an ordinary tape. The machine that computes g works as
follows, in outline. It has a main loop (“while . . .”); it has an inner loop that is executed at the
beginning of the body of the main loop. For this inner loop, the machine uses two states to count the
input positions modulo 2, and erases every second a (which is possible on a rubber tape). When the
right end of the input tape is reached, the state must correspond to an even number of letters a on
the input tape (otherwise, the machine rejects and has no output for this input). At this moment,
one a is printed on the output tape.
while the input tape is not empty:
// main loop
{
in a loop, erase every second a on the input tape; // inner loop
when the right end of the input tape is reached, // after the inner loop
if the number of letters a read in the inner loop was even:
then add an a on the output tape;
else reject;
in a loop, move the head of the input tape back to the left end; // 2nd inner loop
}
This program runs in linear time, and every step is injective.
This injective Turing machine for g can be reversed, which yields an injective Turing machine for
g−1 with exponential time-complexity. ✷
Proposition 3.2 Let h: A∗ → A∗ be an injective function. Then h belongs to invfP iff h is
polynomially balanced, and is computable by some polynomial-time injective Turing machine.
Proof. For any h ∈ invfP, h is injective and computable in polynomial time; moreover, h−1 ∈ invfP,
i.e., h−1 is also injective and computable in polynomial time. Hence, Bennett’s theorem is applicable
(see [1, 2]), so h can be computed by a polynomial-time injective deterministic Turing machine. Also,
since both h and h−1 are computable in polynomial time, h is polynomially balanced. Conversely, if
h is polynomially balanced and is computed by a polynomial-time injective Turing machine, then h−1
can also be computed by such a machine, by Prop. 3.1. ✷
We saw in the Introduction that f ∈ fP is regular iff f has an inverse in invfP. Hence by Prop. 3.2,
we have for all f ∈ fP:
f is regular iff f has an inverse that is computable by a polynomially balanced polynomial-time
injective Turing machine. And P = NP iff every f ∈ fP has an inverse that can be computed by a
polynomially balanced polynomial-time injective Turing machine.
3.2
Evaluation functions and finite generation
Based on the transition table of a Turing machine, one can easily check whether this Turing is deterministic, and whether it is injective. Moreover, we saw in [3] and [4] that one can add a built-in
polynomial time bound and balance bound into the transition table. This can be done for injective
Turing machines too, without destroying injectiveness; indeed, a time bound p(|x|) can be computed
injectively on input x (where p is a stored polynomial, described by its degree and coefficients). So
we can design a set of strings (programs) that describe all polynomially balanced polynomial-time
injective Turing machines. And just as in [3], for every polynomial q(n) = a · (nk + 1), there exists an
evaluation function evq that evaluates all injective Turing machine programs with built-in polynomial
less than q. The details are the same as in [3], and injectiveness doesn’t change any reasoning. We
call such a Turing machine description an invfP-program.
A Turing machine with program w will be denoted by Mw ; we denote the injective input-output
function of Mw by φw .
6
An evaluation function maps a pair (w, x), consisting of a program w and an input x (for Mw ), to
(w, φw (x)), i.e., to that same program and the program-output. In order to represent a pair of strings
by one string we use the prefix code {00, 01, 11}, and the function code(.) defined by
code(0) = 00,
code(1) = 01.
The pair of strings (w, x) is represented unambiguously by code(w) 11 x, where 11 acts as a separator
(since {00, 01, 11} is a prefix code). The evaluation function injEvq is defined by
injEvq code(w) 11 x = code(w) 11 φw (x)
for any invfP-program w with built-in polynomial less than q, and any x ∈ Dom(φw ). The function
injEvq is itself injective and polynomially balanced and polynomial-time computable (with a larger
polynomial than q however). Moreover, injEvq is regular; indeed, the unique mutual inverse of injEvq
−1
is injEv−1
q , and this belongs to fP since φw (x) is injective and regular, i.e., φw ∈ fP for every invfPprogram w. So, injEvq ∈ invfP.
The proof of Prop. 4.5 in [3], showing that fP is finitely generated, goes through without much
change. We use the relation
(⋆)
φw (x) = π2′ |w′ |+2 ◦ contr ◦ recontr2 m ◦ injEvq ◦ reexpand2m ◦ expand ◦ πcode(w) 11 (x),
where w is a program of a polynomially balanced polynomial-time injective Turing machine with
built-in polynomial < q, and w′ = co2m+1 ◦ ex2m+1 (w). See Section 4 of [3] for the definition of contr,
recontr, reexpand, expand, πn′ , πv , co, and ex. The role of expand and reexpand is to increase the input
length, so as to reduce complexity by a padding argument; when complexity is below q, injEvq can be
applied; after that, recontr and contr remove the padding.
The functions contr, recontr, injEvq , reexpand, expand, and πcode(w) 11 are injective and regular (i.e.,
they belong to invfP). In the relation (⋆) we can replace π2′ |w′ |+2 by the injective regular function
′
πcode(w)
: code(w) 11 x 7−→ x. Moreover, πcode(w) is generated by {π0 , π1 }, and π2′ |w′ |+2 is generated
11
by {π0′ , π1′ }.
Thus, φw is generated by the finite set of functions {contr, recontr, injEvq , reexpand, expand, π0 ,
π1 , π0′ , π1′ } ⊂ invfP, and we have:
Proposition 3.3 The inverse monoid invfP is finitely generated, as a monoid.
✷
For every invfP-program w, we can easily obtain an invfP-program (let’s call it w′ ) for φ−1
w ; for this
purpose we simply reverse the injective Turing machine described by w. Thus the function w 7→ w′ is
polynomially balanced and polynomial-time computable; moreover, w 7→ w′ is injective and involutive,
so the function w 7→ w′ is in invfP. Thus we proved:
Proposition 3.4 There exists a function proginv ∈ invfP such that proginv (w) is an invfP-program for
φ−1
✷
w , for every invfP-program w.
We would like to extend the above function proginv to all “regular” fP-programs, i.e., the fPprograms w for which φw is regular in fP. We can do this by using universal search, a.k.a. Levin
search; this is described in [10, 11], but without much detail; for a detailed exposition, see for example
[12] (Theorem 7.21 in the 1993 edition). In the general regular case, proginv (w) is just a program such
that φproginv (w) has polynomial time-complexity, but proginv (w) does not have a built-in polynomial for
its time-complexity; Levin search is not able to explicitly find such a polynomial, although it exists if
φw is regular.
Proposition 3.5 (inversion by Levin search). There exists a function proginv ∈ fP such that for
every fP-program w, proginv (w) is a program for a mutual inverse of φw satisfying Dom(φproginv (w) ) =
Im(φw ).
7
The time-complexity of φproginv (w) is Θ(Tφw ) + Θ(Tw′ ), where Θ(Tφw ) is (up to big-Θ) the optimal
time-complexity of all Turing machines for φw , and Θ(Tw′ ) is (up to big-Θ) the optimal time-complexity
of all Turing machines for all mutual inverses of φw .
In particular:
If φw is regular then φprog inv (w) has polynomial time-complexity. But the program proginv (w) does
not have a built-in polynomial for its time-complexity, i.e., proginv (w) is not an fP-program.
If φw is not regular, proginv (w) is a non-polynomial-time program. The time-complexity of φproginv (w)
has nevertheless an exponential upper bound.
Proof. In our version of universal search the input has the form (w, y), where w is an fP-program,
and y is a possible output of the Turing machine Mw . Remark: For convenience we write (w, y) as a
pair of words, but our universal search will actually use the single word code(w) 11 y.
The output of universal search is (w, x) such that x ∈ φ−1
w (y), if w is an fP-program, and y ∈ Im(φw )
(i.e., if φ−1
(y)
=
6
∅);
there
is
no
output
if
w
is
not
an
fP-program,
or if y 6∈ Im(φw ). Thus universal
w
′
search computes a mutual inverse ev of the general evaluation function ev for fP. Both ev′ and ev
are partial recursive (but they do not belong to fP since no polynomial bound is prescribed, as for
evq ). Since we restrict universal search to fP-programs (hence, with built-in polynomial complexity
bound), the domain of ev is decidable; so universal search is an algorithm (that always halts). Indeed,
by rejecting programs with no polynomial time bound, the search algorithm also rejects programs
that do not halt. And for a polynomial program with known polynomial complexity, inversion has a
predictable exponential-time upper-bound.
For a fixed w in the input, universal search computes ev′ (w, ·) = (w, φ′w (.)), where φ′w (.) is a mutual
inverse of φw . For the details of the universal search algorithm we refer to [12] (proof of Theorem 7.21,
p. 412 in the 1993 edition). From the specification of universal search and w, we immediately derive
a program for ev′ (w, ·); this program is called proginv (w). Since proginv (w) is easily obtained from w,
the function proginv (.) belongs to fP.
The program proginv (w) has minimum time-complexity Θ(Tφw ) + Θ(Tw′ ); this is proved in [12].
It follows that the time-complexity of proginv (w) has an exponential upper bound (since w is an fPprogram). It also follows that proginv (w) is a polynomial-time program if φw is regular, i.e., if φw has
a polynomial-time inverse. But universal search does not find out explicitly what that polynomial is,
so proginv (w) cannot have a built-in polynomial for its time-complexity bound; i.e., proginv (w) is not
an fP-program. ✷
3.3
Injective Turing machines with NP-oracle
In preparation for the next Section, we generalize polynomially balanced polynomial-time injective
Turing machines by adding an oracle from NP.
Definition 3.6 By invfP(NP) we denote the set of all functions computed by polynomially balanced
polynomial-time injective Turing machines, with an oracle belonging to NP.
By Prop. 3.2 we have: invfP ⊆ invfP(NP) .
An injective Turing machine with an oracle is a deterministic and injective machine. Indeed, it
computes deterministically and injectively inbetween oracle transitions. And in an oracle transition
the only change is in the state, which goes from the query state qqu to either qyes or qno ; the contents
and head positions of the tapes (including the query tape) remain unchanged. The next transition
will go to a different state than qyes or qno , so the state qyes (or qno ) uniquely determines the previous
state qqu . Thus, an oracle transition is deterministic and injective.
The reverse of an injective Turing machine with an oracle has a slightly different format than an
an injective Turing machine with an oracle, as described above. Indeed, in an oracle call the following
transition happens: The machine is in the query state qqu and then it enters an answer state, either
8
qyes or qno , according as the word on the query tape belongs to the oracle set or not. (Before a query,
a word is written on the query tape; this happens in the course of a possibly long computation. After
a query, the computation continues, and during this computation the query tape content may be
gradually erased or changed.)
When this sequence is reversed, an answer state occurs before the query state. From the answer
state, the reverse computation goes to the query state, provided that the answer state is qyes and the
query word is in the oracle language, or if the answer state is qno and the query word is not in the oracle
language; the reverse computation rejects otherwise. We call this a reverse oracle call. Although a
reverse oracle call does not have the format of an oracle call, it can easily be simulated by an oracle
call and a few more transitions. Hence we have:
Proposition 3.7 Let M be a polynomially balanced polynomial-time injective Turing machine with
NP-oracle. Then M computes a polynomially balanced injective function.
The machine M ′ , obtained by running M in reverse, is equivalent to a polynomial-time injective
Turing machines with NP-oracle, and computes f −1 (where f is the injective function computed by
M ).
✷
Conversely, every injective oracle Turing machine M can be simulated by a reverse injective Turing
machine with the same oracle. Indeed, let M ′ be the reverse of M , and let M1′ be an ordinary injective
oracle machine that simulates M ′ . Finally, let (M1′ )′ be the reverse of M1′ . Then (M1′ )′ is a reverse
injective Turing machine, with the same oracle as M , simulating M .
So, there is no intrinsic difference between injective oracle Turing machines and reverse injective
oracle Turing machines. We could generalize injective Turing machines with oracle so as to allow oracle
calls and reverse oracle calls in the same machine; but by the above discussion, it doesn’t matter much
whether we use oracle calls, reverse oracle calls, or both.
Corollary 3.8
invfP(NP) is an inverse monoid.
Proof. The composite of two polynomially balanced injective functions, computed by polynomial-time
injective Turing machine with NP-oracle, is also polynomially balanced and injective. It is computed
by a polynomial-time injective Turing machine with NP-oracle, obtained by chaining the two machines.
The disjoint union of two sets in NP is in NP; thus the combined machine can use a single NP-oracle
language. Hence, invfP(NP) is a monoid. By Prop. 3.7, invfP(NP) is an inverse monoid. ✷
Remark. Not all functions in fP(NP) have an inverse in fP(NP) (unless the polynomial hierarchy PH
collapses, see Sect. 6 of [3]). Here, fP(NP) denotes the set of polynomially balanced functions computed
by deterministic polynomial-time Turing machines with NP-oracle. But all functions in invfP(NP) have
an inverse in invfP(NP) (by Cor. 3.8).
Proposition 3.9 Suppose f : A∗ → A∗ is an injective and polynomially balanced function such that
both f and f −1 can be computed by deterministic polynomial-time Turing machines with NP-oracle.
Then f can be computed by an injective polynomial-time Turing machine with NP-oracle.
Proof. Bennett’s proof in [1, 2] applies to Turing machines with NP-oracle. An injective Turing
machine for f (with NP-oracle) is obtained by first using the Turing machine for f (on input x), but
with a history tape added (which makes the machine injective). The input x is still present. Once
f (x) has been computed, a copy of it is made on the output tape. The history tape is then used to run
the previous computation in reverse, thus erasing the history tape and the work-tape copy of f (x).
The input x is still on the input tape at this moment, and a copy of f (x) is on the output tape. See
Lemma 1 of [2] (except that now the machine also has an NP-oracle).
To erase x injectively, the Turing machine for f −1 is used on input f (x) (copied from the output
tape). A history tape is added to make the computation injective; x is (re-)computed (while f (x) is
kept on the output tape). Once x has been (re-)computed, the history tape is used to run the previous
9
computation of the f −1 -machine in reverse, thus erasing the history tape, as well as x. A copy of
f (x) is kept as the output tape. See Theorem 2(b) of [2] (except that now the machine also has an
NP-oracle). ✷
We will prove next that invfP(NP) is finitely generated. For this we want to replace all NP-oracles
by one NP-complete set. Moreover, we want the reduction functions to be in invfP.
Recall that for sets L1 , L2 ⊆ A∗ , a many-one reduction from L1 to L2 is a polynomial-time computable function f : A∗ → A∗ such that L1 = f −1 (L2 ); equivalently, (∀x ∈ A∗ )[ x ∈ L1 ⇔ f (x) ∈ L2 ].
A one-one reduction is a many-one reduction f that is injective.
Definition 3.10 For sets L1 , L2 ⊆ A∗ , an invfP-reduction from L1 to L2 is any element f ∈ invfP
such that L1 = f −1 (L2 ).
This definition generalizes the usual many-one reductions in the sense that the elements of invfP are
partial functions (with domain in P). Note that if L1 = f −1 (L2 ) then L1 ⊆ Dom(f ).
By the definition of invfP, if f ∈ invfP then f −1 ∈ invfP, and Im(f ) ∈ P (by Prop. 1.9 in [3], since
f is regular). If f reduces L1 to L2 then f −1 reduces L2 ∩ Im(f ) to L1 . So, unless L2 ⊆ Im(f ), f −1
does not reduce L2 to L1 , and invfP-reducibility is not a symmetric relation.
As we saw in Prop. 3.2, an invfP-reduction can be computed by a polynomially balanced polynomialtime injective Turing machine.
Proposition 3.11 There exists languages that are NP-complete with respect to invfP-reductions.
Proof. An example is the “universal NP-complete language” of Hartmanis [7], defined as follows
(slightly reformulated):
|w|·pw (|x|) :
LNP
univ = {code(w) 11 x 11 0
w is a nondeterministic polynomial-time Turing machine program,
pw is the built-in polynomial of w, and x ∈ Lw }.
In [7] only many-one reductions were considered, but invfP-reductions can be used too. Indeed, for
any language Lv ∈ NP accepted by a nondeterministic polynomial-time Turing machine with program
v, we define the function gv for all x ∈ A∗ by
x 7−→ code(v) 11 x 11 0|v|·pv (|x|) .
Then gv belongs to invfP, since a fixed program v is chosen for the language Lv . And gv is a invfPNP
reduction reduction from Lv to LNP
✷
univ . So, Luniv is complete with respect to invfP-reductions.
Proposition 3.12 The inverse monoid invfP(NP) is finitely generated, as a monoid.
Proof. By Prop. 3.11 there exist sets that are NP-complete with respect to invfP-reductions. Hence,
for all polynomially balanced polynomial-time injective Turing machines with NP-oracle, we can use
a fixed set N as the oracle, where N is NP-complete with respect to invfP-reductions. This changes
the time-complexity of each function in invfP(NP) by a polynomial amount.
Now the proof of Prop. 3.3 goes through if we replace invfP-programs by invfP(NP) -programs. The
latter programs use the three additional states qqu , qyes , and qno , and one instruction for implementing
(N )
a call to the oracle N . We replace injEvq by injEvq , where the latter is computed by an injective
Turing machine that is similar to the one for injEvq , but with oracle calls to N added. ✷
In the next section we will consider inverses and co-inverses that are computed by polynomially
balanced polynomial-time injective Turing machines with NP-oracle.
10
4
Inverses and co-inverses of any element of fP
If P 6= NP then there exist non-regular functions f ∈ fP. In that case we can nevertheless consider the
injective inverses and co-inverses of f in invfP(NP) , as we shall see in Prop. 4.9. First we derive some
general results about injective inverses and co-inverses.
4.1
Inverses, co-inverses, sub-inverses
Here we consider arbitrary functions f, f ′ : A∗ → A∗ , unless more precise conditions are stated. A
subfunction of a function f is, by definition, any function g such that g ⊆ f ; equivalently, Dom(g) ⊆
Dom(f ) and for all x ∈ Dom(g): g(x) = f (x).
Lemma 4.1 If f f ′ f = f then Im(f ) ⊆ Dom(f ′ ).
Proof. If y = f (x) ∈ Im(f ) then f f ′ (y) = f f ′ f (x) = f (x) = y, so f f ′ (y) is defined, hence f ′ (y) is
defined, so y ∈ Dom(f ′ ). ✷
Lemma 4.2.
(1) If f ′ f f ′ = f ′ then Im(f ′ ) ⊆ Dom(f ).
(2) If in addition f ′ is injective, then Dom(f ′ ) ⊆ Im(f ). This does not hold in general when f ′ is
not injective.
Proof. (1) This is equivalent to Lemma 4.1, up to notation.
(2) If z ∈ Dom(f ′ ) then f ′ (z) = f ′ f f ′ (z) is defined, hence f f ′ (z) is defined and f f ′ (z) ∈ Im(f ).
When f ′ is injective we can consider its set-theoretic inverse f ′−1 , and we have f ′−1 f ′ = idDom(f ′ ) .
Applying f ′−1 on the left to f ′ (z) = f ′ f f ′ (z) yields idDom(f ′ ) ◦ f ′ f (z) = z (if z ∈ Dom(f ′ )). Hence
idDom(f ′ ) ◦ f f ′ (z) is defined, and idDom(f ′ ) ◦ f f ′ (z) = f f ′ (z) = z, so z ∈ Im(f ).
When f ′ is not injective, the result might not hold. Consider for example f = {(a, b)} and
′
f = {(a, a), (b, a)}. Then f ′ f f ′ = f ′ (and f f ′f = f as well, so f and f ′ are mutual inverses); but
Dom(f ′ ) = {a, b} 6⊆ {b} = Im(f ). ✷
Corollary 4.3 If f, f ′ are mutual inverses and f ′ is injective then Dom(f ′ ) = Im(f ). This does not
hold in general when f ′ is not injective.
✷
Lemma 4.4 If f ′ f f ′ = f ′ and f ′ is injective, then f ′−1 = f |Im(f ′ ) ; equivalently,
−1
f ′ = f |Im(f ′ )
= f −1 ∩ Dom(f ′ ) × Im(f ′ ).
Hence, f ′−1 ⊆ f .
Proof. Applying f ′−1 on the left and the right to f ′ = f ′ f f ′ yields idDom(f ′ ) ◦ f ◦ idIm(f ′ ) = f ′−1 . This
is equivalent to f |Im(f ′ ) ∩ f −1 (Dom(f ′ )) = f ′−1 . Hence, f ′−1 ⊆ f , so f ′ = (f ′−1 )−1 ⊆ f −1 , where f −1 =
{(y, x) : y = f (x)}. Thus, Im(f ′ ) = f ′ (Dom(f ′ )) ⊆ f −1 (Dom(f ′ )), hence Im(f ′ ) ∩ f −1 (Dom(f ′ )) =
Im(f ′ ); so f |Im(f ′ ) ∩ f −1 (Dom(f ′ )) = f |Im(f ′ ) = f ′−1 .
The equality f ′ = f −1 ∩ Dom(f ′ ) × Im(f ′ ) is just a logical reformulation of f ′ = (f |Im(f ′ ) )−1 . ✷
Corollary 4.5 If f ′ is an injective co-inverse of f , then f ′ is a mutual inverse of f |Im(f ′ ) (which is an
injective subfunction of f ). Moreover, f ′ is uniquely determined by f and Im(f ′ ), as f ′ = (f |Im(f ′ ) )−1 .
Proof. Since f ′ = idIm(f ′ ) ◦ f ′ , we have f ′ = f ′ f f ′ = f ′ ◦ f ◦ idIm(f ′ ) ◦ f ′ = f ′ ◦ f |Im(f ′ ) ◦ f ′ . Moreover,
since f |Im(f ′ ) = f ′−1 (by Lemma 4.4), we have f |Im(f ′ ) = f ′−1 = f ′−1 f ′ f ′−1 = f |Im(f ′ ) ◦ f ′ ◦ f |Im(f ′ ) .
✷
Corollary 4.6 If f ′ f f ′ = f ′ and Dom(f ′ ) = Im(f ), then f ′ is an injective mutual inverse of f .
11
Proof. When Dom(f ′ ) = Im(f ) then Im(f ′ ) is a choice set of f . Indeed, for every y ∈ Im(f ), f ′ (y)
is defined and f ′ (y) ∈ f −1 (y) (since f ′ ⊂ f −1 by the second equality in Lemma 4.4). Hence f ′ is an
inverse of f . Injectiveness follows from the observations made in the Introduction and at the beginning
of Section 2. ✷
Theorem 4.7 (anti-homomorphic property of co-inverses).
Suppose fi′ is an injective co-inverse of fi for i = 1, 2. Then f1′ f2′ is an injective co-inverse of f2 f1 .
Proof. We first observe that for subidentities idX , idY , we have idX idY = idX∩Y = idY idX . Since fi′
is injective, fi′ fi′−1 = idIm(fi′ ) , and fi′−1 fi′ = idDom(fi′ ) . Hence
f1′ f2′ f2′−1 f1′−1 f1′ f2′ = f1′ idIm(f2′ ) idDom(f1′ ) f2′ = f1′ idDom(f1′ ) idIm(f2′ ) f2′ = f1′ f2′ .
So, f1′ f2′ = f1′ f2′ f2′−1 f1′−1 f1′ f2′ ⊆ f1′ f2′ f2 f1 f1′ f2′ (by Lemma 4.4). Now, since f1′ f2′ ⊆ f1′ f2′ f2 f1 f1′ f2′ ,
we have f1′ f2′ (x) = f1′ f2′ f2 f1 f1′ f2′ (x) when f1′ f2′ (x) is defined. And when f1′ f2′ (x) is not defined,
f1′ f2′ f2 f1 f1′ f2′ (x) is not defined either. Thus, f1′ f2′ = f1′ f2′ f2 f1 f1′ f2′ . ✷
Remark. Theorem 4.7, though straightforward, is remarkable as it applies to injective co-inverses
but not to injective inverses or mutual inverses.
The latter can be illustrated by the following example: Let f1 = {(0, 0), (1, 0)} = f2 , and let
f1′ = {(0, 1)} = f2′ , so fi and fi′ are mutual inverses (i = 1, 2), and fi′ is injective. But f1′ ◦ f2′ = θ (the
empty map), which is not an inverse of f2 f1 = {(0, 0), (1, 0)}.
In [3] (Prop. 6.1) we showed that every element f ∈ fP has an inverse in fP(NP) ; one such inverse
′ , defined by
is fmin
min(f −1 (y))
if y ∈ Im(f ),
′
fmin (y) =
undefined
otherwise,
where the min operation is taken with respect to the length-lexicographic order, ≤llex , of {0, 1}∗ . Recall
that by definition, u ≤llex v iff |u| < |v| or [ |u| = |v| and u ≤dict v ]; here, ≤dict is the dictionary order
of {0, 1}∗ determined by 0 <dict 1.
Proposition 4.8 For every f ∈ fP we have:
′
′
′
′
′ ;
(1) f ◦ fmin
= idIm(f ) ; hence, f ◦ fmin
◦ f = f ; also fmin
◦ f ◦ fmin
= fmin
′ ) ∈ coNP;
(3) Im(fmin
′ ) = Im(f ) ∈ NP;
(4) Dom(fmin
′
(2) fmin
is injective;
′−1
∈ fP(NP) .
(5) fmin
′ .
Proof. (1) is obvious from the definition of fmin
′
(2) Injectiveness of fmin
follows from the fact that the sets f −1 (y) (for y ∈ Im(f )) are two-by-two
disjoint.
′ ) iff x ∈ Dom(f ) and (∀z)[ z <
(3) We have: x ∈ Im(fmin
llex x ⇒ f (z) 6= f (x) ] (the latter sentence
expresses that x is the ≤llex -minimum element in its modf -class). Since Dom(f ) ∈ P, since the relation
{(z, x) : z <llex x ⇒ f (z) 6= f (x)} is in P, and since this relation is universally quantified (by
′ ) ∈ coNP.
(∀z, z <llex x)), it follows that Im(fmin
′
′ ) = Im(f ). And Im(f ) ∈ NP for every f ∈ fP.
(4) By the definition of fmin
we have Dom(fmin
′−1
′−1
(NP)
′ ) (by Lemma 4.4), and f ∈ fP,
= f |Im(fmin
. Indeed, fmin
(5) As a consequence of (2), fmin ∈ fP
′−1
′
and Im(fmin ) ∈ coNP, so fmin is computed by any Turing machine for f with an added oracle call to
′ ).
Im(fmin
✷
′
In the proof of Prop. 6.1 in [3] we showed that fmin
∈ fP(NP) . Next we strengthen this by showing
′
that the oracle Turing machine that computes fmin is an injective Turing machine with an NP-oracle. In
Def. 3.6 we introduced the set invfP(NP) of polynomially balanced functions computed by polynomialtime injective Turing machines with an NP-oracle.
Proposition 4.9 Every f ∈ fP has an inverse f ′ in invfP(NP) (hence f ′−1 ∈ invfP(NP) ), with the
additional property that Im(f ′ ) ∈ coNP.
12
′ , which is injective as we just saw, hence f ′−1 is an injective function. Both f ′
Proof. Let f ′ be fmin
(NP)
and f ′−1 are in fP
; this holds for f ′ by Prop. 6.1 in [3], and for f ′−1 by Prop. 4.8. Hence by Prop.
(NP)
′
3.9, f ∈ invfP
.
Since f ′ is computed by an injective Turing machine (with oracle), f ′−1 can be computed by the
same machine run backwards; so f ′−1 ∈ invfP(NP) .
′ ) ∈ coNP.
We proved in Prop. 4.9 that Im(fmin
✷
′ (y) in
Remark. As an alternative proof of Prop. 4.9, observe that the algorithm for computing fmin
the proof of Prop. 6.1 in [3] actually describes an injective Turing machine; it makes oracle calls to
the set {(z, u) : z ∈ f (uA∗ )} (which is in NP for any fixed f ∈ fP).
Based on the anti-homomorphic property of co-inverses (Theorem 4.7) we define:
Definition 4.10 The monoid of co-inverses of fP is
cofP = {f ′ ∈ invfP(NP) : f ′ is a co-inverse of some element of fP}.
Equivalently, an element f ′ of invfP(NP) belongs to cofP iff f ′ has an inverse in fP.
By Theorem 4.7, cofP is indeed a monoid.
′
′
Also, fmin
∈ cofP for every f ∈ fP, since f is a mutual inverse of fmin
(Prop. 4.8). Hence, for every
f ∈ fP, cofP contains a mutual inverse of f .
One could also consider the monoid of inverses of fP, namely {g ∈ invfP(NP) : g is an inverse of
some element of fP}. However, this is just all of invfP(NP) , since every element of invfP(NP) is an inverse
of the empty map θ ∈ fP. So, being an inverse of an element in fP is a trivial notion; but having an
inverse in fP (i.e., being a co-inverse of an element in fP) is non-trivial. The question whether cofP is
equal to invfP(NP) has an interesting “answer”, which is presented in Prop. 4.20 below.
Proposition 4.11
cofP ∩ fP = invfP.
Proof. [⊇] If f ′ ∈ invfP then f ′−1 ∈ invfP ⊆ fP. Moreover, f ′ f ′−1 f ′ = f ′ , so f ′ is a co-inverse of an
element in fP. Hence f ′ ∈ cofP.
[⊆] If f ′ ∈ cofP ∩ fP then there exists f ∈ fP such that f ′ f f ′ = f ′ , hence f ′ is a regular element of
fP. Since we assume f ′ ∈ cofP, it follows that f ′ is injective. So, f ′ is an injective regular element of
fP, i.e., f ′ ∈ invfP. ✷
On the other hand, invfP(NP) ∩ fP contains all injective one-way functions. In fact, invfP(NP) ∩ fP =
invfP iff injective one-way functions do not exist. The latter is the case iff P = UP (see e.g. section
12.1 in [13]). But cofP contains no one-way function, since every element of cofP has an inverse in fP.
Definition 4.12 Let injfP denote the monoid of all injective functions in fP.
The set of regular elements of injfP is exactly the submonoid invfP (by the definition of invfP), hence
the set of injective one-way functions (for worst-case complexity, see Prop. 1.3) is injfP − invfP. We
saw that invfP and invfP(NP) are regular monoids (in fact, inverse monoids). On the other hand, injfP
is regular iff injective one-way functions do not exist. Thus we have:
injective one-way functions exist iff invfP 6= injfP 6= invfP(NP) .
Since invfP consists of the regular elements of injfP, we also have:
injective one-way functions exist iff invfP 6= injfP.
Since invfP(NP) is regular we have,
if injective one-way functions exist, then injfP 6= invfP(NP) .
We will also see (Theorem 4.18) that one-way functions exist iff invfP 6= invfP(NP) .
We saw that cofP contains no one-way functions (see the observations after Prop. 4.11), so we
have: injfP ∩ cofP = invfP. Moreover:
13
Proposition 4.13 invfP(NP) ∩ fP = injfP.
Proof. [⊆] Every f ∈ invfP(NP) is injective, so the inclusion ⊆ holds.
′
[⊇] If f ∈ injfP then fmin
= f −1 ∈ invfP(NP) , by Prop. 4.9. It follows that (f −1 )−1 = f ∈ invfP(NP) ,
by Prop. 3.7. Hence f ∈ invfP(NP) ∩ injfP ⊆ invfP(NP) ∩ fP. ✷
✬
✬
invfP(NP)
injfP
✩
✩
✬
cofP
✫
✩ fP
✩
✫
invfP
✫
✪
✪
✪
✪
Figure 1: Relations between fP, invfP(NP) , cofP, invfP, and injfP.
There is an interesting connection between co-inverses and sub-inverses of f .
Definition 4.14 A function g′ is a sub-inverse of a function f iff there exists g ⊆ f such that
gg′ g = g and g ′ gg′ = g ′ . I.e., the sub-inverses of f are the mutual inverses of the subfunctions of f .
Proposition 4.15 For any function f : A∗ → A∗ , the injective co-inverses of f are the same as the
injective sub-inverses of f .
Proof. [⇒] Let f ′ be an injective co-inverse of f ; then Dom(f ′ ) ⊆ Im(f ) and Im(f ′ ) ⊆ Dom(f ), by
Lemmas 4.1 and 4.2. Let us consider the following restriction of f : g = f |Im(f ′ ) . We claim that f ′ is
a mutual inverse of g, hence f ′ is a sub-inverse of f . Recall that f ′ = (f |Im(f ′ ) )−1 , by Lemma 4.4. It
follows from this that gf ′ g = f |Im(f ′ ) ◦ (f |Im(f ′ ) )−1 ◦ f |Im(f ′ ) = f |Im(f ′ ) = g; similarly it follows that
f ′ gf ′ = (f |Im(f ′ ) )−1 ◦ f |Im(f ′ ) ◦ (f |Im(f ′ ) )−1 = (f |Im(f ′ ) )−1 = f ′ .
So, f ′ is an injective sub-inverses of f .
[⇐] Let g ⊂ f be any restriction of f , and let g′ be any injective mutual inverse of g; in other words,
g′ is an injective sub-inverse of f . We claim that g ′ is a co-inverse of f .
By Corollary 4.3, and Lemma 4.4, we have: Dom(f ′ ) = Im(f ), and g′ = (g|Im(g′ ) )−1 . Since g ⊂ f
the latter implies g ′ = (f |Im(f ′ ) )−1 . Hence, g′ f g′ = (f |Im(f ′ ) )−1 ◦ f ◦ (f |Im(f ′ ) )−1 = (f |Im(f ′ ) )−1 = g′ .
So, g ′ is a co-inverse of f . ✷
Note that in part [⇐] of the above proof, g need not be injective, but g|Im(g′ ) = f |Im(g′ ) is injective.
Thus we have:
14
If g′ is an injective sub-inverse of f and it is a mutual inverse of a subfunction of f , then g ′ is
also a mutual inverse of an injective subfunction of f .
The following gives a relation between two notions of sub-inverse. We defined a sub-inverse of f to be
a mutual inverse of a subfunction of f (Def. 4.14). We could, instead, have defined a sub-inverse of f
to be a subfunction of a mutual inverse of f .
Proposition 4.16.
(1) If g′ is a subfunction of a mutual inverse of f , then g ′ is a sub-inverse of f . (For this fact, g′
need not be injective, and f need not be in fP.)
(2) If g′ ∈ invfP(NP) is a sub-inverse of f ∈ fP, then g ′ is a subfunction of some mutual inverse
f ′ ∈ invfP(NP) of f .
Proof. (1) Let g ′ ⊆ f ′ with f f ′ f = f and f ′ f f ′ = f ′ . Let g = f g′ f ; then g ⊆ f f ′ f = f . And for
all x ∈ Dom(g) we have gg ′ g(x) = f f ′ f (x) = f (x) = g(x) (the latter since x ∈ Dom(g)). So gg ′ g = g.
Moreover, for all y ∈ Dom(g′ ) we have g′ gg ′ (y) = f ′ f f ′ (y), since g′ ⊆ f ′ and g ⊆ f ; and f ′ f f ′ (y) = y.
So, g ′ gg ′ = g′ .
(2) Since g ′ ∈ invfP(NP) , g ′ is injective, hence it is an injective co-inverse of f (by Prop. 4.15).
Hence, g′ = (f |Im(g′ ) )−1 , and g ′ is a mutual inverse of g = f |Im(g′ ) ⊆ f . We can extend g ′ to the
following mutual inverse of f : For all y ∈ Im(f ) let
′
g (y)
if y ∈ Dom(g′ ),
′
f (y) =
′
fmin (y)
otherwise.
′
′ (y ) for
It follows that f ′ is injective. Indeed, both g′ and fmin
are injective. Moreover, g′ (y1 ) 6= fmin
2
′ (y ); but
all y1 ∈ Dom(g′ ) and all y2 ∈ Im(f ) − Dom(g ′ ), since if we had equality then f g′ (y1 ) = f fmin
2
′ (y ) = y ; so we would have y = y , but y ∈ Dom(g ′ )
f g′ (y1 ) = y1 since g′ = (f |Im(g′ ) )−1 , and f fmin
2
2
1
2
1
′
and y2 6∈ Dom(g ).
Also, f ′ is clearly a mutual inverse of f .
′ , there is an
Since g′ is computed by a injective Turing machine with NP-oracle, and likewise fmin
injective Turing machine with NP-oracle for f ′ ; the Turing machine for g′ , being deterministic, can
check whether g′ (y) is defined, i.e., whether y ∈ Dom(g′ ). Hence, f ′ ∈ invfP(NP) . ✷
4.2
Connections with NP
The next theorem motivates the study of the monoid cofP (and of invfP) in the context of the P versus
NP problem. We prove an easy Lemma first.
Lemma 4.17 Let f, h be functions A∗ → A∗ and let Z ⊂ A∗ . If f ◦ h|Z ◦ f = f , then f hf = f .
Proof. For x 6∈ Dom(f ), f hf (x) and f (x) are both undefined. For x ∈ Dom(f ), h|Z (f (x)) is defined,
otherwise f ◦ h|Z ◦ f (x) would be undefined, and thus would not be equal to f (x). Hence we have
f (x) ∈ Z, so Im(f ) ⊆ Z. It follows that h|Z (f (x)) = h(f (x)), for all x ∈ Dom(f ). Hence h|Z ◦f = h◦f ,
thus f ◦ h|Z ◦ f = f h f = f . ✷
Theorem 4.18 The following are equivalent:
P = NP;
cofP = invfP;
cofP ⊂ fP;
cofP is regular;
invfP(NP) = invfP.
15
Proof. We use the following logical fact: If X ⇒ Yi and X ⇒ Y i , then X ⇔ Yi (for i = 1, 2, . . . ).
Here, “P = NP” plays the role of X.
[⇒] We assume P = NP. We obviously have invfP ⊆ cofP. If P = NP then fP(NP) = fP, hence
cofP ⊆ fP. Also, fP is regular when P = NP. Since all elements of cofP are injective, all elements of
cofP are injective regular elements of fP, i.e., cofP ⊆ invfP. In conclusion, cofP = invfP.
If cofP = invfP then cofP ⊂ fP. The inclusion is strict because fP has some non-injective elements.
And if cofP = invfP then, since invfP is regular, cofP is regular.
[⇐] If P 6= NP then there exists f ∈ fP such that f is not regular. For f there exists a mutual inverse
f ′ ∈ invfP(NP) such that Dom(f ′ ) = Im(f ); hence, f ′ ∈ cofP. Moreover, Dom(f ′ ) = Im(f ) ∈ NP; and f ′
′
can be chosen so that in addition, Im(f ′ ) ∈ coNP (e.g., by choosing f ′ = fmin
as in Prop. 4.9). Hence,
(NP)
′−1
′−1
′
′
f
∈ invfP
(since f
= f |Im(f ′ ) and Im(f ) ∈ coNP). However, f 6∈ fP, since f is not regular
in fP.
Let us assume by contradiction that the monoid cofP is regular. Then f ′ (as above) is a regular element of cofP, hence (by Lemma 4.9), f ′−1 ∈ cofP. This implies (by the definition of cofP)
that f ′−1 h f ′−1 = f ′−1 for some h ∈ fP. Composing on the left and on the right with f ′ yields
idIm(f ′ ) h idDom(f ′ ) = f ′ ; equivalently, h|Dom(f ′ ) ∩ h−1 (Im(f ′ )) = f ′ . Hence, f f ′ f = f h|Dom(f ′ ) ∩ h−1 (Im(f ′ )) f
= f . It follows (by Lemma 4.17) that h is an inverse of f , since the restriction h|Dom(f ′ ) ∩ h−1 (Im(f ′ )) is
an inverse of f . Hence (since f is not regular), h 6∈ fP; but this contradicts the previous assumption
that h ∈ fP. Thus cofP is not regular.
Since the monoid invfP is regular, we conclude that cofP 6= invfP.
And cofP 6= invfP implies cofP 6⊂ fP; indeed, if we had cofP ⊂ fP then we would have cofP =
cofP ∩ fP = invfP.
Finally, if invfP(NP) = invfP then cofP = invfP, hence by what we proved above, P = NP. Conversely, if P = NP then every polynomial-time injective Turing machine with NP-oracle is equivalent
to a polynomial-time deterministic Turing machine (without oracle). Thus if f ∈ invfP(NP) (hence by
Prop. 3.7, f −1 ∈ invfP(NP) ), then f, f −1 ∈ fP. By Bennett’s results [1, 2] (especially Lemma 1 and
Theorem 2 (b) in [2]), this implies f ∈ invfP. ✷
Although every element of cofP has an inverse in fP (by the definition of cofP), not every element
of cofP has an inverse in cofP, unless P = NP; the latter follows from the fact that cofP is regular iff
P = NP.
By definition, cofP ⊆ invfP(NP) . We will now address the question whether this inclusion is strict;
it turns out, interestingly, that this is equivalent to the question whether P 6= NP.
Lemma 4.19 Let f ∈ fP, and let f ′ ∈ cofP be a co-inverse of f such that Im(f ′ ) ∈ coNP. Then
f ′−1 ∈ fP(NP) .
For every f ∈ fP there exists a co-inverse f ′ ∈ cofP of f such that Im(f ′ ) ∈ coNP and Dom(f ′ ) ∈
NP.
Proof. Since f ∈ fP and Im(f ′ ) ∈ coNP, f ′−1 = f |Im(f ′ ) can be computed by a Turing machine for f
with NP-oracle Im(f ′ ).
We saw above (and in [3]) that for every f ∈ fP there exists a mutual inverse (hence a co-inverse)
′ .
f ′ for which Im(f ′ ) ∈ coNP and Dom(f ′ ) ∈ NP; e.g., let f ′ = fmin
✷
Proposition 4.20 P 6= NP iff cofP 6= invfP(NP) .
In any case (i.e., whether P 6= NP or not), cofP and invfP(NP) have the same set of idempotents.
Proof. [⇐] If cofP = invfP(NP) , then for every f ′ ∈ invfP(NP) there exists f ∈ fP such that f ′ f f ′ = f ′ .
By Lemma 4.4, f ′−1 ⊆ f , and by Lemma 4.19, f ′−1 ∈ invfP(NP) . It follows that f ′ has an injective
inverse f ′−1 ∈ invfP(NP) , i.e., f ′ is regular in cofP = invfP(NP) . This holds for all f ′ ∈ cofP, hence cofP
is regular. Thus by Theorem 4.18, P = NP.
16
[⇒] If P = NP then cofP = invfP, which in this case is also equal to invfP = invfP(NP) .
If e is an idempotent of invfP(NP) then e is also a co-inverse of an element of fP; indeed, e ◦ idA∗ ◦ e =
e. Hence e ∈ cofP. ✷
Proposition 4.21
P 6= NP iff cofP contains regular elements that do not belong to invfP.
Proof. [⇐] If P = NP then cofP = invfP (by Theorem 4.18), so there are no (regular) elements in
cofP that are not in invfP.
[⇒] If P 6= NP then we construct the following example of a regular element in cofP that does not
belong to invfP. Let L ⊂ A∗ be a coNP-complete set. Then idL ∈ invfP(NP) and it is regular (being an
idempotent). Also, idL = idL ◦idA∗ ◦idL , hence idL is a co-inverse of an element of fP. Thus, idL ∈ cofP.
But if P 6= NP then idL 6∈ invfP, since L is coNP-complete. ✷
We saw that invfP and invfP(NP) are finitely generated. Hence, if cofP were not finitely generated,
this would imply that cofP 6= invfP, which would imply that P 6= NP (by Theorem 4.18). However, we
will prove next that cofP is finitely generated.
We will first show that cofP has a machine (or program) model, and that there is a corresponding
evaluation function for all bounded-complexity functions in cofP. This is similar to the situation in fP,
invfP, and invfP(NP) .
A cofP-program is of the form (v ′ , w), where v ′ is any invfP(NP) -program (with built-in timecomplexity and balance function), and w is any fP-program (with built-in time-complexity and balance
function). An invfP(NP) -program describes a polynomially balanced polynomial-time injective Turing
machine, with a fixed NP-oracle N (where N is NP-complete). The functions in fP or invfP(NP) specified
by programs w or v ′ are denoted by φw , respectively ψv′ . On input y ∈ A∗ , the program (v ′ , w) is
evaluated as follows:
(1)
(2)
(3)
x = ψv′ (y) is computed;
φw (x) is computed;
if φw ◦ ψv′ ◦ φw (x) = φw (x) and ψv′ ◦ φw ◦ ψv′ (y) = ψv′ (y) then the output is ψv′ (y);
there is no output otherwise.
The function specified by program (v ′ , w) on input y is denoted by Φ(v′ ,w) . It is easy to see that Φ(v′ ,w)
is a subinverse of φw , and a subfunction of ψv′ . Thus, Φ(v′ ,w) ∈ cofP.
Conversely, every function h ∈ cofP has an invfP(NP) -program, say v ′ ; and h has an inverse φw ∈ fP
for some fP-program w. Then (v ′ , w) is a cofP-program for h.
Based on the cofP-programs and a polynomial bound q, we can construct an evaluation function
(N )
(N )
evCoq such that evCoq (code(v ′ ) 11 code(w) 11 y) = code(v ′ ) 11 code(w) 11 Φ(v′ ,w) (y), for y and
(v ′ , w) as above. This is similar to Section 3.2.
Theorem 4.22 The monoid cofP is finitely generated.
Proof. The proof is the same as for invfP(NP) (Prop. 3.12), except that evCoq (with programs (v ′ , w)
(N )
as above) is used instead of injEvq . ✷
We saw that P 6= NP iff cofP is not regular. The set of elements of cofP that are regular in cofP will
be denoted by RegcofP. We will see that it has interesting properties.
Proposition 4.23 (1) The set RegcofP is a finitely generated inverse submonoid of cofP.
(2) If P 6= NP then invfP $ RegcofP $ cofP $ invfP(NP) . If P = NP then invfP = invfP(NP) .
(3) An element f ∈ fP has an inverse in RegcofP iff f is regular in fP.
17
Proof. (1) For any g ∈ RegcofP, let g′ ∈ RegcofP be a mutual inverse of g. Multiplying gg′ g = g
on the left and the right by g−1 we obtain: idDom(g) ◦ g′ ◦ idIm(g) = g−1 . Hence g−1 ∈ cofP, being a
product of elements of cofP (since the idempotents idDom(g) , idIm(g) belong to cofP). Thus we proved:
For all g ∈ RegcofP : g −1 ∈ cofP.
For any g1 , g2 ∈ RegcofP we have therefore, g1−1 , g2−1 ∈ cofP, hence g2 g1 has g1−1 g2−1 ∈ cofP as an
inverse, so g2 g1 is regular in cofP. This proves that RegcofP is closed under composition. Also, since
all elements of RegcofP are injective, RegcofP is an inverse monoid.
The proof of finite generation of RegcofP is similar to the proof of finite generation of cofP. We
construct an evaluation function for the elements of RegcofP, based on the following machine (or
program) model for the elements of RegcofP. A RegcofP-program is any pair (u, v ′ ) of cofP-programs,
and the function Φ(u,v′ ) computed by this program is defined by
Φ(u,v′ ) (x) = ψu (x)
if
ψu ψv′ ψu (x) = ψu (x);
and Φ(u,v′ ) (x) is undefined otherwise. Since the relation ψu ψv′ ψu (x) = ψu (x) can be checked in P(NP) ,
Φ(u,v′ ) belongs to invfP(NP) . And since ψu ∈ cofP has some inverse f ∈ fP, and Φ(u,v′ ) is a subfunction
of ψu , Φ(u,v′ ) also has f as an inverse; hence, Φ(u,v′ ) ∈ cofP. Moreover, since ψu ψv′ ψu (x) = ψu (x) for
every x ∈ Dom(Φ(u,v′ ) ), and Φ(u,v′ ) ⊆ ψu , we have Φ(u,v′ ) ψv′ Φ(u,v′ ) = Φ(u,v′ ) . So Φ(u,v′ ) is regular in
cofP. Finally, every regular element of cofP obviously has a program of the form (u, v ′ ) as above.
The rest of the proof of finite generation is very similar to the one for cofP.
(2) The second part of (2) was already proved in Theorem 4.18. On the other hand, if P 6= NP then
cofP is not regular (by Theorem 4.18), so RegcofP 6= cofP 6= invfP(NP) . Also, cofP contains regular
elements that are not in invfP, by Prop. 4.21, hence invfP 6= RegcofP.
(3) If f is a regular element of fP then f has an inverse in invfP, by Coroll. 2.3.
In general, let f be an element in fP that has an inverse f ′ ∈ RegcofP; we want to show that f is
regular in fP. By what we proved in (1) of the present proof, f ′−1 ∈ cofP. We have f ′−1 gf ′−1 = f ′−1 for
some g ∈ fP, since f ′−1 ∈ cofP. Multiplying this on the left and the right by f ′ yields idI ◦ g ◦ idD = f ′ ,
where D = Dom(f ′ ) and I = Im(f ′ ). Since f ′ is an inverse of f , Lemma 4.17 implies that g is an
inverse of f . Since g ∈ fP, it follows that f is regular in fP. ✷
We saw that the monoids fP, invfP, invfP(NP) , and cofP are finitely generated. For injfP we ask
similarly:
Question: Is injfP finitely generated?
We do not know the answer. If we could show that injfP is not finitely generated, then this would
prove that injfP 6= invfP, i.e., injective one-way functions exist, hence P 6= UP, and hence P 6= NP.
When we proved that fP, invfP, invfP(NP) , and cofP are finitely generated, we used machine (or
program) models, and evaluation functions. It seems that injfP does not have a program model, since
injectiveness is a for-all property, that does not have finite witnesses in general. To illustrate the
difficulty of finding a program model for injfP, here is an idea that does not work. We saw that
injfP = fP ∩ invfP(NP) , and this suggests that a function in injfP can be specified by a pair (u, w), where
w is an fP-program (for φw ∈ fP), and u is an invfP(NP) -program (for ψu ∈ invfP(NP) ); the program
(u, w) specifies the injective function ψu ∩ φw . But ψu ∩ φw ranges over all of invfP(NP) and does
not necessarily belong to injfP (unless injfP = invfP(NP) , which would imply that P = NP). So, this
approach towards proving finite generation does not work (unless one also proves that P = NP). It
seems that injfP is not finitely generated (but this will probably be very difficult to prove).
5
Subgroups and group-inverses
We first characterize the maximal subgroups of invfP, cofP, and invfP(NP) . Then we consider elements
of fP that have an inverse in such a subgroup (i.e., a group-inverse). Typically, elements of fP do not
18
have a group-inverse, but we will see that all elements of fP are reduction-equivalent to elements that
have group-inverses.
5.1
Maximal subgroups
We first characterize the idempotents of injfP.
Proposition 5.1 For any f ∈ injfP we have: f is an idempotent iff f = idZ for some set Z ⊆ A∗
with Z ∈ P. Hence injfP and invfP have the same idempotents.
Proof. If f ∈ injfP and f = f ◦ f , then we compose on the left with f −1 (which exists, since f is
injective, but f −1 might not belong to injfP); this yields idD = idD ◦ f , where D = Dom(f ). Hence for
all x ∈ D = Dom(f ) we have x = f (x). Thus, f = idD . Since f ∈ fP, D = Dom(f ) ∈ P.
Conversely, if Z ∈ P then idZ can be computed by a deterministic polynomial-time Turing machine.
Moreover, idZ ∈ invfP since idZ is injective, and it is regular (being an idempotent). ✷
Proposition 5.2 The maximal subgroup of invfP and injfP are the same.
Let idZ be an idempotent of invfP. The maximal subgroup of invfP with identity idZ consists of
the permutations of Z that belong to invfP. In particular, the group of units of invfP consists of all
permutations of A∗ that belong to invfP.
Proof. If f belongs to a subgroup of injfP then f is regular, hence f ∈ invfP.
Obviously, the permutations in invfP with domain and image Z form a subgroup of invfP. Conversely, if f ∈ invfP belongs to the maximal subgroup with identity idZ , then f −1 f = f f −1 = idZ ; so
the domain and the image of f are both Z. Since f ∈ invfP, f is injective, and f −1 ∈ invfP. Hence f
permutes Z. ✷
Proposition 5.3 Every maximal subgroup of invfP is also a maximal subgroup of fP. The group of
units of fP is the same as the group of units of invfP.
Proof. A maximal subgroup of fP is a regular H-class. For fP the characterization of the L- and the
R-relation (see Prop. 2.1 in [3]) implies that every element f of a maximal subgroup of fP with unit
idZ satisfies Im(f ) = Z = Dom(f ) and f is injective (since f has the same partition as idZ ). Thus f
belongs to invfP. ✷
The converse of Prop. 5.3 is of course not true; e.g., fP contains non-injective idempotents, hence
subgroups that are not contained in invfP.
Let us now look at the idempotents and subgroups of invfP(NP) and of cofP. We saw that invfP(NP)
and cofP have the same idempotents (Prop. 4.20). And we saw that invfP(NP) is an inverse monoid
(Cor. 3.8); on the other hand, cofP is not regular, unless P = NP (by Theorem 4.18).
Lemma 5.4 For every L ⊂ A∗ we have: L ∈ P(NP) iff idL ∈ invfP(NP) .
Proof. [⇒] If L ∈ P(NP) then idL can be computed by deterministic polynomial-time Turing machine
with NP-oracle. Moreover, the inverse of idL (which is just idL itself) also has that property, hence
by Prop. 3.9, idL can be computed by an injective polynomial-time Turing machine with NP-oracle.
Hence, by the definition of invfP(NP) we have idL ∈ invfP(NP) .
[⇐] From a polynomial-time injective Turing machine with NP-oracle, computing idL , one immediately obtains a deterministic polynomial-time Turing machine with NP-oracle, accepting L. ✷
Proposition 5.5 Every idempotent of invfP(NP) (and cofP) is of the form idL , where L ∈ P(NP) ,
L ⊆ A∗ . Moreover, every L ∈ P(NP) is accepted by an injective polynomial-time Turing machine with
NP-oracle.
19
Proof. Let f = f ◦f be an idempotent of invfP(NP) . Multiplying in the left by f −1 yields idD = idD ◦f ,
where D = Dom(f ). Hence, f = idD . Since f ∈ invfP(NP) we have D ∈ P(NP) (by Lemma 5.4).
Conversely, if Z ∈ P(NP) then idZ ∈ invfP(NP) (by Lemma 5.4). From this we obtain an injective
polynomial-time Turing machine with NP-oracle, accepting Z. ✷
Proposition 5.6 (1) The maximal subgroup of invfP(NP) with idempotent idL (where L ∈ P(NP) )
consists of all the permutations of L that belong to invfP(NP) . In particular, the group of units of
invfP(NP) consists of all permutations of A∗ that belong to invfP(NP) .
(2) The maximal subgroup of cofP with idempotent idL (where L ∈ P(NP) ) consists of all permutations
g of L such that g = f ∩ (L × L) and g −1 = h ∩ (L × L) for some f, h ∈ fP.
Hence, if L ∈ P then the maximal subgroup of cofP with idempotent idL is a subgroup of invfP. In
particular, the group of units of cofP is the same as the group of units of invfP.
Proof. (1) By injectiveness of the elements of invfP(NP) , every element of the maximal subgroup
of invfP(NP) with idempotent idL is a permutation of L. Moreover, for every g ∈ invfP(NP) we have
g−1 ∈ invfP(NP) (by Prop. 3.7); hence the permutations of L in invfP(NP) form a subgroup of invfP(NP) .
(2) If g ∈ invfP(NP) belongs to a maximal subgroup of cofP with idempotent idL , then both g and
g−1 are permutations of L. If g has an inverse h ∈ fP then multiplying ghg = g on the left and on the
right by g−1 yields idL ◦ h ◦ idL = g −1 ; moreover, idL ◦ h ◦ idL = h ∩ (L × L). Similarly, for g−1 ∈ coP
we obtain f ∩ (L × L) = g for some f ∈ fP.
Conversely, let g be a permutation of L ∈ P(NP) such that g = f ∩ (L × L) and g−1 = h ∩ (L × L)
for some f, h ∈ fP. Then g, g −1 belong to invfP(NP) since L ∈ P(NP) and f, h ∈ fP. Moreover, g and
g−1 belong to cofP; indeed, the inverses f ∩ (L × L) and h ∩ (L × L) can be extended to inverses h,
respectively f , in fP (by Lemma 4.17). Since g is a permutation of L, it follows now that g belongs to
a subgroup of cofP. ✷
We saw in Prop. 4.21 that unless P = NP, cofP contains regular elements that are not in invfP.
The next Proposition describes the regular elements of cofP.
Proposition 5.7 Every element of cofP that is regular in cofP has the form f ∩ (K × H), for some
f ∈ fP and K, H ∈ P(NP) .
Conversely, suppose f, h ∈ fP and K, H ∈ P(NP) are such that f ∩ (K × H) and h ∩ (H × K) are
injective and mutual inverses. Then f ∩ (K × H), h ∩ (H × K) ∈ cofP, and they are regular in cofP.
Proof. If g is regular in cofP then g −1 ∈ cofP; hence, g −1 f g−1 = g−1 for some f ∈ fP. Hence
by multiplying on the left and the right by g we obtain: f ∩ (K × H) = g, where K = Im(g) and
H = Dom(g).
For the converse, f ∩ (K × H) = (f ∩ (K × H)) ◦ (h ∩ (H × K)) ◦ (f ∩ (K × H)) =
(f ∩ (K × H)) ◦ h ◦ (f ∩ (K × H)); the latter holds by Lemma 4.17. Hence, f ∩ (K × H) has an
inverse in fP, so f ∩ (K × H) belongs to cofP. Similarly, h ∩ (H × K) belongs to cofP. Since they are
mutual inverses, they are regular in cofP. ✷
5.2
Group inverses of elements of fP
After noticing that every regular element of fP has an injective inverse, belonging to an inverse submonoid of fP, we wonder whether we can go even further: Does every regular element in fP have an
inverse in a subgroup of fP? This is of course not the case; for immediate counter-examples consider the
functions that are total but not surjective, or surjective but not total. Even within the subsemigroup
of non-total non-surjective functions of fP there are counter-examples (due to the polynomial balance
and time requirements). Nevertheless, we will find that every regular element of fP is equivalent, with
respect to inversive reduction, to a regular element of fP that has an inverse in a subgroup of fP.
We will also investigate elements of fP that are possibly non-regular, but that have an inverse in
a subgroup of invfP(NP) . Again, we will find that every element of fP is equivalent, with respect to
20
inversive reduction, to an element of fP that has an inverse in a subgroup of invfP(NP) . In particular,
there are elements of fP that are complete with respect to inversive reduction and that have an inverse
in a subgroup of invfP(NP) .
Some definitions about inversive reductions (from [3]) between functions f1 , f2 :
• f1 is simulated by f2 (denoted by f1 4 f2 ) iff there exist β, α ∈ fP such that f1 = β ◦ f2 ◦ α.
• f1 reduces inversively to f2 (notation, f1 6inv f2 ) iff
(1) f1 4 f2 and
(2) for every inverse f2′ of f2 there exists an inverse f1′ of f1 such that f1′ 4 f2′ .
• f, g ∈ fP are equivalent via inversive reduction iff f 6inv g and g 6inv f .
Theorem 5.8 Every function f ∈ fP is equivalent, via inversive reduction, to a function f0 ∈ fP
such that f0 has an inverse in a subgroup of invfP(NP) .
In particular, f0 has an inverse in the group of permutations of 0 Dom(f ) ∪ 1 Im(f ), in invfP(NP) .
Moreover, f0 has an inverse in the group of units of invfP(NP) .
If f (and hence f0 ) is regular, then f0 has an inverse in the group of permutations of 0 Dom(f ) ∪
1 Im(f ), in invfP. Moreover, f0 has an inverse in the group of units of invfP.
Remarks. Since f and f0 are equivalent via reduction it follows immediately that f0 ∈ fP iff f ∈ fP;
and f is regular iff f0 is regular. And by Theorem 4.18, if f is regular then f0 has an inverse in the
inverse monoid invfP.
Proof of Theorem 5.8. (1) With f we associate f0 defined by
Dom(f0 ) = 0 Dom(f ), and
f0 (0 x) = 1 f (x), for all x ∈ Dom(f ).
Similarly, we define f1 by f1 (1 x) = 0 f (x).
To show that f and f0 simulate each other (and similarly for f1 ), we introduce the functions πa
and πa′ for each a ∈ A; they are defined for all z ∈ A∗ by
πa (z) = az, and
πa′ (az) = z, with Dom(πa′ ) = aA∗ .
Then we have:
f0 = π1 ◦ f ◦ π0′ , and f = π1′ ◦ f0 ◦ π0 .
Hence, f and f0 simulate each other.
Let us show that we have an inversive reduction of f0 to f . If f has an inverse f ′ , let us define f1′
by f1′ (1 y) = 0 f ′ (y) for all y ∈ Dom(f ′ ). Then f1′ is an inverse of f0 . And f1′ = π0 ◦ f ′ ◦ π1′ , so f1′ is
simulated by f ′ .
Conversely, let us show that there is an inversive reduction of f to f0 . Let g be any inverse
of f0 , i.e., f0 gf0 (x) = f0 (0 x), all x ∈ Dom(f ). Then h = g ∩ 1A∗ × 0A∗ can be simulated by
g (indeed, h(.) = id0A∗ ◦ g ◦ id1A∗ (.)), and h also satisfies f0 hf0 = f0 . Moreover, h ⊂ 1A∗ × 0A∗
implies that h = k1 for some function k. Then we have f kf = f ; indeed, for all x ∈ Dom(f )
we have f kf (x) = π1′ f0 k1 f0 (0x) = π1′ f0 (0x) = f (x). We saw that k is simulated by k1 (indeed,
k = π0′ ◦ k1 ◦ π1 ). Thus there exists an inverse (namely k) of f that is simulated by g.
This completes the proof that f is equivalent to f0 via inversive reduction. It follows immediately
that f0 ∈ fP iff f ∈ fP, and that f is regular iff f0 is regular.
(2) Let f ′ be any mutual inverse of f with f ′ ∈ cofP, or with f ′ ∈ invfP if f is regular. Let us
now extend f1′ to a group element. Recall that by definition, f1′ (1 y) = 0 f ′ (y) for all y ∈ Dom(f ′ ),
Dom(f1′ ) = 1 Dom(f ′ ) = 1 Im(f ), and Im(f1′ ) = 0 Im(f ). So, Dom(f1′ ) ∩ Im(f1′ ) = ∅, and f1′ only has
orbits of length 1. We extend f1′ to a permutation F ′ of Dom(f1′ ) ∪ Im(f1′ ), defined by
F ′ |Dom(f1′ ) = f1′
and
F ′ |Im(f1′ ) = f1′−1 .
Hence, F ′ = f1′ ∪ f1′−1 . It follows that F ′ ∈ invfP(NP) , since both f1′ and f1′−1 belong to invfP(NP) .
21
Also, F ′ is injective and F ′ ◦ F ′ = idZ , where Z = Dom(f1′ ) ∪ Dom(f1′−1 ); so F ′ = F ′−1 belongs to
a two-element group. Clearly, Dom(F ′ ) = Im(F ′ ) = Z.
Moreover, F ′ is an inverse of f0 . Indeed, for every x ∈ Dom(f ) we have f0 F ′ f0 (0x) = f0 F ′ (1f (x)) =
f0 f1′ (1f (x)), since F ′ = f1′ ∪ f1′−1 and f1′−1 is undefined on 1A∗ ; hence, f0 F ′ f0 (0x) = f0 (0x).
Finally, let us show that if f ′ ∈ invfP then F ′ ∈ invfP. Indeed, when f ′ ∈ invfP then f ′−1 ∈ invfP,
from which it follows (by disjointness of the domains and disjointness of the images) that f1′ ∪ f1′−1 ∈
invfP.
(3) We can extend F ′ to a permutation of all of A∗ (belonging to the group of units of invfP(NP) ): We
simply define F ′ as a permutation of Dom(f1′ ) ∪ Im(f1′ ), as above, and then extend F ′ to the identity
function on A∗ − (Dom(f1′ ) ∪ Im(f1′ )). Since f1′ , idDom(f1′ ) , and idIm(f1′ ) belong to invfP(NP) , it follows
that this extension of F ′ to all of A∗ belongs to invfP(NP) .
Finally, if f is regular then this extension of F ′ belongs to invfP; indeed, in that case, Dom(f1′ ) and
Im(f1′ ) belong to P. ✷
Corollary 5.9 There exists a function that is complete in fP for inversive reduction, and that has
an inverse in the group of units of invfP(NP) .
Proof. In [3] we saw examples of functions F ∈ fP that are complete with respect to inversive
reduction. Then F0 , defined by F0 (0x) = 1 F (x), is also complete (since F and F0 are equivalent via
inversive reduction); and F0 has an inverse in the group of units of invfP(NP) (as we saw in the proof
of Theorem 5.8). ✷
Remark. The group-inverses constructed above belong to invfP(NP) , but not necessarily to cofP. If
there were a complete function in fP that has an inverse in the group of units of cofP then P = NP
(since the group of units of cofP is the same as the group of units of invfP).
6
Appendix: Simple facts about inverses (related to Section 2)
We present some additional properties of invfP that are not used in the rest of the paper.
For f ∈ fP we define the right fixator of f by RFix(f ) = {α ∈ fP : f ◦ α = f }; in other words,
α ∈ RFix(f ) iff α is a right-identity of f . Similarly, the left fixator of f is LFix(f ) = {α ∈ fP : α◦f = f }.
We observe that RFix(f ) ∩ invfP is an inverse monoid; i.e., α ∈ RFix(f ) ∩ invfP implies α−1 ∈
RFix(f ) ∩ invfP. Indeed, if α ∈ invfP satisfies f α = f then f αα−1 = f α−1 ; moreover, f α = f implies
that Dom(f ) ⊆ Im(α), hence αα−1 = idIm(α) acts as a right-identity on f ; hence f = f αα−1 . Thus,
f = f α−1 .
For every α ∈ RFix(f ) we have: α(Dom(f )) ⊆ Dom(f ), α|Dom(f ) ∈ RFix(f ), and α(A∗ −Dom(f )) ⊆
∗
A − Dom(f ).
Every representative choice function of f belongs to RFix(f ), since any right fixator α ∈ RFix(f )
maps every modf -class into itself. But the converse is not true, since a right fixator α does not
necessarily map a modf -class to a single element of the modf -class (as a representative choice function
does).
Proposition 6.1 Let f1′ , f2′ ∈ invfP be inverses of the same element f ∈ fP, such that Dom(fi′ ) =
Im(f ) (i = 1, 2). Then:
(1)
f1′−1 ≡R f2′−1 ≡R idIm(f ) (for the R-relation of invfP), and
idIm(f ) ≡R f (for the R-relation of fP); and
f1′ ≡L f2′ (for the L-relation of invfP).
If f1′ ≡L f2′ and f1′ ≡R f2′ (for the L- and R-relations of fP), then f1′ = f2′ .
(2)
fi′ f ∈ RFix(f ),
f2′ f1′−1 is a bijection from the choice set Im(f1′ ) onto the choice set Im(f2′ );
22
fi′ fi′−1 = idIm(fi′ ) .
fi′−1 fi′ = idIm(f ) .
(3) f2′ f f1′ = f2′ .
(4) If α ∈ RFix(f ), and if f ′ ∈ fP is an inverse of f ∈ fP such that Dom(f ′ ) = Im(f ), then α ◦ f ′ is
also an inverse of f such that Dom(αf ′ ) = Im(f ). So, the monoid RFix(f ) acts on the left on the set
{g′ ∈ invfP : g ′ is an inverse of f and Dom(g′ ) = Im(f )}.
Moreover, this action is transitive, and the action of {α|Dom(f ) : α ∈ RFix(f )} is faithful.
Proof. (1) We saw that fi′−1 = f |Im(fi′ ) , hence fi′−1 = f ◦ idIm(fi′ ) , where Im(fi′ ) ∈ P (since fi′ is
regular); hence idIm(fi′ ) ∈ fP and fi′−1 ≤R f .
Also, fi′ = fi′ f fi′ ≡L f fi′ = idIm(f ) . We also have fi′−1 ≡R idIm(f ) ; indeed, fi′−1 maps Im(fi′ )
bijectively onto Im(f ), so fi′−1 = idIm(f ) ◦ fi′−1 ≤R idIm(f ) ; and idIm(f ) = fi′−1 fi′ ≤R fi′−1 . Moreover,
f ≡R idIm(f ) , hence fi′−1 ≡R f .
The ≡L -equivalence then follows, since in any inverse monoid, x′ ≡R y ′ implies x ≡L y (where z ′
is the inverse of z for any element z).
The fact that f1′ ≡L f2′ and f1′ ≡R f2′ imply f1′ = f2′ is well-known in semigroup theory (see e.g. p.
26 in [5]).
(2) Obviously, f fi′ f = f . The rest is straightforward.
(3) We know that f1′ and f2′ have the same domain, namely Im(f ). And f f1′ |Im(f ) = idIm(f ) , hence
′
f2 f f1′ = f2′ idIm(f ) = f2′ (the latter again holds since Dom(fi′ ) = Im(f )).
(4) If f α = f then f αf ′ f = f f ′ f = f , so αf ′ is also an inverse of f . And Dom(αf ′ ) ⊆ Dom(f ′ ) =
Im(f ); since αf ′ is an inverse of f we also have Im(f ) ⊆ Dom(αf ′ ); hence Dom(αf ′ ) = Im(f ). It
follows that αf ′ ∈ invfP, so RFix(f ) acts on the given set on the left.
Transitivity follows from (f2′ f ) f1′ = f2′ (proved in (3)), where f2′ f ∈ RFix(f ) by (2).
Proof of faithfulness: Note that Dom(f ) ⊆ Dom(α) for all α ∈ RFix(f ). If α1 |Dom(f ) 6= α2 |Dom(f )
then there exists x0 ∈ Dom(f ) such that α1 (x0 ) 6= α2 (x0 ). There exists an inverse f0′ ∈ invfP of
f with Dom(f0′ ) = Im(f ) such that f0′ (f (x0 )) = x0 ; indeed, we can start with any inverse f0′ with
Dom(f0′ ) = Im(f ), and if f0′ (f (x0 )) 6= x0 we can redefine f0′ on f (x0 ); this does not change Dom(f0′ ) (it
just changes the choice set of f0′ ). Now, α1 f0′ f (x0 ) = α1 (x0 ) 6= α2 (x0 ) = α2 f0′ f (x0 ), so α1 f0′ 6= α2 f0′ .
✷
Proposition 6.2 If f ∈ fP is regular, and if Dom(f ◦ α) = Dom(f ), then we have:
α ∈ RFix(f ) iff for every mutual inverse f ′ of f : α ◦ f ′ is also a mutual inverse of f .
(If f is not regular, the Proposition holds for inverses in invfP(NP) .)
Proof. [⇒] If α ∈ RFix(f ) and f ′ is a mutual inverse of f then f αf ′ f = f f ′ f = f ; the first equality
holds because α ∈ RFix(f ). And αf ′ f αf ′ = αf ′ f f ′ = αf ′ (again, using f α = f ). So αf ′ is a mutual
inverse of f .
[⇐] If α 6∈ RFix(f ) then f α(x0 ) 6= f (x0 ) for some x0 ∈ A∗ . If x0 ∈ Dom(f ), let f ′ ∈ fP be a mutual
inverse of f such that f ′ f (x0 ) = x0 ; such a mutual inverse exists (if f ′ f (x0 ) 6= x0 we can redefine f ′
on the one element f (x0 )). Then f αf ′ f (x0 ) = f α(x0 ) 6= f (x0 ), so f αf ′ f 6= f . So there exists a
mutual inverse f ′ such that αf ′ is not an inverse of f . ✷
From the above we conclude that finding an inverse f ′ for f with Dom(f ′ ) = Im(f ) can be broken
up into two steps: (1) find a maximal injective subfunction f ′−1 of f (being a maximal injective
subfunction is equivalent to Im(f ′−1 ) = Im(f )); (2) find the inverse f ′ of f ′−1 . (If f is injective, step
(1) is skipped.)
Proposition 6.3 When f ∈ fP is regular and if Im(β ◦ f ) = Im(f ) we have:
β ∈ LFix(f ) iff for every mutual inverse f ′ of f : f ′ β is also a mutual inverse of f .
(If f is not regular, the Proposition holds for inverses in invfP(NP) .)
23
Proof. [⇒] If β ∈ LFix(f ) and f ′ is a mutual inverse of f then f f ′ β f = f f ′ f = f ; and f ′ β f f ′ β =
f ′ f f ′ β = f ′ β (using βf = f ).
[⇐] If β 6∈ LFix(f ) then βf (x0 ) 6= f (x0 ) for some x0 ∈ A∗ . If βf (x0 ) ∈ Im(f ) then f f ′ β f (x0 ) = β f (x0 )
since f f ′ |Im(f ) = idIm(f ) ; and β f (x0 ) 6= f (x0 ), hence f f ′ β f 6= f . ✷
References
[1] Ch.H. Bennett, “Logical reversibility of computation”, IBM Journal of Research and Development 17.6
(1973) 525-532.
[2] Ch.H. Bennett, “Time/space trade-offs for reversible computation”, SIAM J. on Computing 18(4) (1989)
766-776.
[3] J.C. Birget, “Semigroups and one-way functions”, International J. of Algebra and Computation 25.1-2
(2015) 3-36. – Preprint: http://arXiv.org/abs/1306.1447
[4] J.C. Birget, “Infinitely generated semigroups and polynomial complexity”, International J. of Algebra and
Computation 26.04 (2016) 727-750.
[5] A.H. Clifford, G.B. Preston, The Algebraic Theory of Semigroups, Vol. 1 (Mathematical Survey, No 7 (I))
American Mathematical Society, Providence (1961).
[6] P.A. Grillet, Semigroups, an introduction to the structure theory, Marcel Dekker (1995).
[7] J. Hartmanis, Feasible Computations and Provable Complexity Properties, SIAM/CBMS-NSF Book 30
(1978).
[8] L.H. Hemaspaandra, M. Ogihara, The Complexity Theory Companion, Springer (2002).
[9] M.V. Lawson, Inverse Semigroups: The Theory of Partial Symmetries, World Scientific Publishing (1998).
[10] L.A. Levin, “Universal sequential search problems”, Problemy Peredatshi Informatsii, 9.3 (1973) 115-116.
[11] L.A. Levin, “Randomness conservation inequalities; information and independence in mathematical theories”, Information and Control, 61.1 (1984) 15-37.
[12] M. Li, P. Vitányi, An Introduction to Kolmogorov Complexity and its Applications, Springer (1993, 1997,
2008).
[13] Ch. Papadimitriou, Computational Complexity, Addison-Wesley (1994).
24
| 4math.GR
|
1
Coordination in Distributed Networks via
Coded Actions with Application to Power
Control
arXiv:1501.03685v4 [cs.IT] 14 Aug 2017
Benjamin Larrousse, Samson Lasaulce, and Matthieu R. Bloch
{larrousse, lasaulce}@lss.supelec.fr, [email protected]
Abstract
This paper investigates the problem of coordinating several agents through their actions, focusing
on an asymmetric observation structure with two agents. Specifically, one agent knows the past, present,
and future realizations of a state that affects a common payoff function, while the other agent either
knows the past realizations of nothing about the state. In both cases, the second agent is assumed to
have strictly causal observations of the first agent’s actions, which enables the two agents to coordinate.
These scenarios are applied to distributed power control; the key idea is that a transmitter may embed
information about the wireless channel state into its transmit power levels so that an observation of
these levels, e.g., the signal-to-interference plus noise ratio, allows the other transmitter to coordinate its
power levels. The main contributions of this paper are twofold. First, we provide a characterization of
the set of feasible average payoffs when the agents repeatedly take long sequences of actions and the
realizations of the system state are i.i.d.. Second, we exploit these results in the context of distributed
power control and introduce the concept of coded power control. We carry out an extensive numerical
analysis of the benefits of coded power control over alternative power control policies, and highlight a
simple yet non-trivial example of a power control code.
Index Terms
This paper was presented in part at the 2013 IEEE International Symposium on Information Theory [1], the 2013 International
Workshop on Stochastic Methods in Game Theory [2], and the 12th IEEE International Symposium on Modeling and Optimization
in Mobile, Ad Hoc, and Wireless Networks in 2014 [3]. The work of B. Larrousse and S. Lasaulce was supported in part by
the French agency ANR through the project LIMICOS - ANR-12-BS03-0005. The work of M. Bloch was supported in part by
NSF under award CCF-1320304.
August 15, 2017
DRAFT
2
Channels with state; coding theorems; coordination; distributed power control; distributed resource
allocation; game theory; information constraints; interference channel; optimization.
I. I NTRODUCTION
The main technical problem studied in this paper is the following. Given an integer N ≥ 1, three
discrete alphabets X0 , X1 , X2 , and a stage payoff function w : X0 × X1 × X2 → R, one wants to
maximize the average payoff
N
N
WN (xN
0 , x1 , x2 ) ,
N
1X
w(x0,n , x1,n , x2,n )
T
(1)
n=1
N
N
N
with respect to (w.r.t.) the sequences xN
1 , (x1,1 , . . . , x1,N ) ∈ X1 and x2 , (x2,1 , . . . , x2,N ) ∈ X2
N
given the knowledge of xN
0 , (x0,1 , . . . , x0,N ) ∈ X0 . Without further restrictions and with instantaneous
knowledge of x0,n , solving this optimization problem consists in finding one of the optimal pairs of
variables (x?1,n , x?2,n ) for every n. The corresponding maximum value1 of WN is then
WN?
N
1 X
max w(x0,n , x1 , x2 ).
=
x1 ,x2
N
(2)
n=1
We assume here that the variable x2 cannot be controlled or optimized directly. As formally described
in Section II, the variable x2 results from imperfect observations of x0 through x1 , which induces an
information constraint in the aforementioned optimization problem. One contribution in Section III is to
precisely characterize this constraint for large N when xN
0 consists of independent identically distributed
(i.i.d.) realizations of a given random variable X0 .
This setting is a special case of distributed optimization, in which K agents2 connected via a given
observation structure have the common objective of maximizing the average payoff WN for large N .
The variable xk with k ∈ {1, . . . , K} is the action of Agent k and represents the only variable under
its control. The variable x0 is outside of the agents’ control and typically represents the realization of a
random system state. The observation structure defines how the agents interact through observations of
the random state and of each other’s actions. The average payoff then measures the degree of coordination
between the agents, under the observation constraints of the actions imposed by the observation structure.
As a concrete example, we apply this framework to power control in Section V, in which x0 represents
the global wireless channel state information and xk the power level of Transmitter k .
1
This ideal situation is referred to as the “costless communication” case. In Section V, the corresponding power control
scenario is called costless communication power control (CCPC).
2
In other disciplines such as computer science, control, or economics, agents are sometimes called nodes, controllers, or
decision-makers.
August 15, 2017
DRAFT
3
A central question is to characterize the possible values of the average payoff WN when the agents
interact many times, i.e., when N is large. Answering this question in its full generality still appears out
of reach, and the present paper settles for a special case with K = 2 agents. Specifically, we assume
that Agent 1 has perfect knowledge of the past, current, and future realizations of the state sequence xN
0 ,
while Agent 2 obtains imperfect and strictly causal observations of Agent 1’s actions and possesses either
strictly causal or no knowledge of the realizations of the state. Despite these restricting assumptions, one
may extract valuable concepts and insights of practical relevance from the present work, which can be
extended to the general case of K ≥ 2 agents and arbitrary observation structures.
A. Related work
In most of the literature on agent coordination, including classical team decision problems [4], agents
coordinate their actions through dedicated channels, which allow them to signal or communicate with
each other without affecting the payoff function. The works most closely related to the present one are [5],
[6], in which the authors introduce the notions of empirical and strong coordination to measure agents’
ability to coordinate their actions in a network with noiseless dedicated channels. Empirical coordination
measures an average coordination over time and requires the joint empirical distribution of the actions to
approach a target distribution asymptotically in variational distance; empirical coordination relates to the
communication of probability distributions [7] and tools from rate-distortion theory. Strong coordination
is more stringent and asks the distribution of sequences of actions to be asymptotically indistinguishable
from sequences of actions drawn according to a target distribution, again in terms of variational distance;
strong coordination relates to the notion of channel resolvability [8]. The goal is then to establish the
coordination capacity [5], which relates the achievable joint distributions of actions to the fixed rate
constraints on the noiseless dedicated channels. The results of [5], [6] have been extended to a variety of
networks with dedicated channels [9], [10], [11], [12], and optimal codes have been designed for specific
settings [13], [14], [15].
Much less is known about the coordination via the actions of agents in the absence of dedicated
channels, which is the main focus of the present work. The most closely related work is [16], in which the
authors characterize the set of possible average payoffs for two agents, assuming that each agent perfectly
monitors the other agent’s actions; the authors establish the set of implementable distributions, which
are the achievable empirical joint distributions of the actions under the assumed observation structure.
In particular, this set is characterized by an information constraint that captures the observation structure
between the agents. While [16] largely relies on combinatorial arguments, [17] provides an informationtheoretic approach of coordination via actions under the name of implicit communication. Coordination
August 15, 2017
DRAFT
4
via actions also relates to earlier works on encoders with cribbing [18], [19], [20]; in such models,
encoders observe the output signals of other encoders, which effectively creates indirect communication
channels to coordinate. Another class of relevant models in which agent actions influence communication
are channels with action-dependent states [21], in which the signals emitted by an agent influence the
state of a communication channel.
To the best of our knowledge, the present work is the first to exploit coordination via actions for
distributed resource allocation in wireless networks, and specifically here for distributed power control
over an interference channel and multiple-access channels. Much of the distributed power control literature
studies the performance of power control schemes using game-theoretic tools. One example is the iterative
water-filling algorithm [22], which is an instance of best-response dynamics (BRD), and is applied over a
time horizon over which the wireless channel state is constant. One of the main drawbacks of the various
implementations of the BRD for power control problems, see e.g., [23], [24], [25], is that they tend to
converge to Nash-equilibrium power control (NPC) policies. The latter are typically Pareto-inefficient,
meaning that there exist some schemes that would allow all the agents to improve their individual utility
w.r.t. the NPC policies. Another drawback is that such iterative schemes do not always converge. Only
restrictive sufficient conditions for convergence are known, see e.g., [26] for the case of multiple input
multiple output (MIMO) interference channels, and are met with probability zero for some special cases
such as the parallel multiple-access channels [27]. In contrast, one of the main benefits of coded power
control developed in Section V is precisely to obtain efficient operating points for the network. This is
made possible by having the transmitters exchange information about the quality of the communication
links through observed quantities, such as the signal-to-interference plus noise ratio (SINR). The SINRs
of the different users effectively act as the outputs of a channel over which transmitters communicate
to coordinate their actions. A transmitter codes several realizations of the wireless channel state into a
sequence of power levels, which then allows other transmitters to exploit their corresponding sequence
of SINRs to select their power levels. No iterative procedure is required and convergence issues are
therefore avoided. We focus our study on efficiency, and NPC is therefore compared to coded power
control in terms of average sum-rate; other aspects such as wireless channel state information availability
and complexity should also be considered but are deferred to future work.
B. Contributions
The contributions of the present work are as follows.
•
The results in Section III extend [16] by relaxing assumptions about the observation structure.
While [16] assumes that Agent 2 perfectly monitors the actions of Agent 1, we consider the case
August 15, 2017
DRAFT
5
of imperfect monitoring and analyze situations in which Agent 2 has a strictly causal knowledge
(Theorem 4 and Corollary 12) or no knowledge (Theorem 5) of the state.
•
We clarify the connections between the game-theoretic formulation of [16] and information-theoretic
considerations from the literature on state-dependent channels [28], [29], [30], [31], [32], separation
theorems, and empirical coordination [5], [33]. We also formulate the determination of the long-run
average payoff as an optimization problem, which we study in detail in Section IV and exploit for
power control in Section V.
•
We establish a bridge between the coordination via actions and power control in wireless networks.
We develop a new perspective on resource allocation and control, in which designing a resource
allocation with high average common payoff amounts to designing a code. Such a code has to strike
a balance between sending information about the upcoming realizations of the state, to obtain high
payoff in the future, and achieving a good value of the current payoff. As an illustration, we provide
a complete description of a power control code for the multiple-access channel in Section V-E.
II. P ROBLEM STATEMENT
For convenience, we provide a summary of the notation used throughout this paper in Table I.
We now formally introduce the problem of interest. We consider K = 2 agents that have to select
their actions repeatedly over N ≥ 1 stages and wish to coordinate via their actions in the presence
of a random state and with an observation structure detailed next. At each stage n ∈ [1 : N ], the
action of Agent k ∈ {1, 2} is xk,n ∈ Xk with |Xk | < ∞, while the realization of the random state is
x0,n ∈ X0 with |X0 | < ∞. The realizations of the state are i.i.d. according to a random variable X0 with
distribution ρ0 ∈ ∆(X0 ). The random state does not depend on the agents’ actions but affects a common
payoff function3 w : X0 × X1 × X2 → R. Coordination is measured in terms of the average payoff
N
N
WN (xN
0 , x1 , x2 ) as defined in (1). At every stage n, Agent 2 only has access to imperfect observations
yn ∈ Y of Agent 1’s actions with |Y| < ∞, which are the output of channel without memory and with
transition probability
P(yn |xn0 , xn1 , xn2 , y n−1 ) = Γ(yn |x0,n , x1,n , x2,n )
(3)
for some fixed conditional probability Γ. We consider two observation structures defined by the strategies
(σn )1≤n≤N and (τn )1≤n≤N of Agents 1 and 2, respectively, which restrict how agents observe the state
3
The function w can be any function such that the asymptotic average payoffs defined in the paper exist.
August 15, 2017
DRAFT
6
and each other’s actions at all stage n ∈ [1 : N ] as follows:
σI :
X0N
→ X1
n
case I:
τ I : X n−1 × Y n−1 → X
2
n
0
σ II : X N → X
1
n
0
case II:
.
τ II : Y n−1 → X
(4)
(5)
2
n
Note that the strategies differ from conventional block channel coding, since an agent acts at every
stage; they may rather be viewed as joint source-channel codes with online coding and decoding.
These strategies are also asymmetric since Agent 1 does not observe Agent 2’s actions. Symmetric
strategies, in which agents would interact, are much more involved and partial results have been recently
developed in [34]. There exist, however, many scenarios, such as cognitive radio, heterogeneous networks,
interference alignment, and master-slave communications [35] in which asymmetric strategies are relevant.
Our objective is to characterize the set of average payoffs that are asymptotically feasible, i.e., the possible
P
values for limN →∞ N1 N
n=1 w(x0,n , x1,n , x2,n ) under the observation structures defined through (4)
and (5). The definition of the two corresponding feasible sets is as follows.
Definition 1 (Feasible sets of payoffs). The feasible set of payoffs in case I is defined as
N
n
o
1 X
I
I
Ω = ω ∈ R : ∃ (σn , τn )1≤n≤N , ω = lim
E w X0,n , σnI (X0N ), τnI (X0n−1 , Y n−1 )
. (6)
N →∞ N
I
n=1
The feasible set of payoffs in case II is defined as
N
n
o
1 X
E w X0,n , σnII (X0N ), τnII (Y n−1 )
.
ΩII = ω ∈ R : ∃ (σnII , τnII )1≤n≤N , ω = lim
N →∞ N
(7)
n=1
The feasible sets of payoffs are directly related to the set of empirical coordinations over the alphabet
X , X0 × X1 × X2 , defined as follows.
Definition 2 (Type [36]). Let N ≥ 1. For any sequence of realizations z N of the generic random variable
Z , the type of z N , denoted by Tz N , is the probability distribution on Z defined by
4
Tz N (z) =
N
1 X
1{zn =z} .
N
(8)
n=1
Definition 3 (Empirical coordination [5]). For ` ∈ {I, II}, Q ∈ ∆(X ) is an achievable empirical
coordination if there exists a sequence of strategies (σn` , τn` )1≤n≤N that generates, together with X0N , a
sequence X N ∈ X such that
∀ > 0,
August 15, 2017
lim P(||TX N − Q||1 > ) = 0,
N →∞
(9)
DRAFT
7
i.e., the distance between the histogram of a sequence of actions and Q converges in probability to 0.
Each feasible set of payoffs is the linear image of the corresponding set of empirical distributions
under the expectation operator. A value ω is asymptotically feasible if there exists an achievable emP
pirical coordination Q such that ω = EQ [w] = x0 ,x1 ,x2 Q(x0 , x1 , x2 )w(x0 , x1 , x2 ). We focus on the
characterization of achievable empirical coordinations rather than the direct characterization of the feasible
sets of payoffs.
Remark 1. The notion of empirical coordination relates to the game-theoretic notion of implementability [16]. For ` ∈ {I, II}, Q ∈ ∆(X ) is implementable if there exists a sequence of strategies (σn` , τn` )1≤n≤N ,
` ∈ {I, II}, that induce at each stage n a joint distribution
PX0,n ,X1,n ,X2,n ,Yn (x0 , x1 , x2 , y) , Γ(y|x0 , x1 , x2 )PX1,n ,X2,n |X0,n (x1 , x2 |x0 )ρ0 (x0 ),
(10)
and that generate, together with the sequence X0N , the sequence X N ∈ X such that
lim ||E(TX N ) − Q||1 = 0,
N →∞
(11)
i.e., the average histogram of a sequence of actions is arbitrarily close to Q. As shown in Appendix A,
if Q ∈ ∆(X ) is an achievable empirical coordination, then it is implementable.
We conclude this section by a brief discussion of the model, especially Agent 1’s strategy in (4)
and (5) that exploits non-causal knowledge of an i.i.d. state. This assumption has been often used since
the work of Gel’fand and Pinsker [28], but we provide here additional justifications motivated by the
application to power control in Section V. First, even if Agent 1 only knows future realizations over a
limited time horizon, coordination may be significantly improved compared to conventional approaches,
such as implementing single-stage game Nash equilibrium-type distributed policies [22], [23], [26], [37].
For instance, power control is typically based on a training phase and an action phase, assuming that a
single channel state is known in advance; this corresponds to N = 2 in our model and, as illustrated in
Fig. 1, a simple coordination strategy is for Agent 1 to inform Agent 2 about the upcoming channels state
during odd stages4 and coordinate their actions during even ones. In that context, assuming that Agent 1
knows the state non-causally is a way to establish an upper bound on the performance all strategies with
limited time horizon. Second, predicting the wireless channel state over a long time horizon has recently
become realistic. For instance, the trajectory of a mobile user can be forecast [38], [39], [40], which
4
For example, Transmitter 1 might use a high (resp. low) power level on an odd stages to inform Transmitter 2 that the
channel is good (resp. bad) in the next even stage.
August 15, 2017
DRAFT
8
makes our approach relevant when the wireless channel state is interpreted as the variation of path loss
and shadowing. References [38], [39], [40] also suggest that, by sampling the channel at the appropriate
rate, the state is nearly i.i.d.. Finally, note that the proposed approach also applies if the state is only i.i.d.
from block to block, where a block consists of several stages, and suggests that gains can be obtained
10
by varying the power level from stage to stage, even if the channel is constant over a block.
Stage
index |
1
|
2
|
3
|
4
|
5
|
good for good for good for good for good for
Channel user 1 user 2 user 1 user 1 user 2
|
|
|
|
|
state |
Transmit
High
High
Low
Low
power Low
|
|
|
|
|
|
level 1
Random
Random
Transmit Random
power choice High choice Low choice
|
|
|
|
|
level 2 |
Fig. 1. This figure illustrates a simple coordination scheme between two transmitters (which are the agents) in a simplified
Fig. 1. This figure illustrates a simple coordination scheme between two transmitters (which are the agents) in a simplified
scenario inspired by [16] where the alphabets are binary: X0 = {good for user 1, good for user 2}, X1 = {low, high}, X2 =
scenario inspired by [16] where the alphabets are binary: X0 = {good for user 1, good for user 2}, X1 = {low, high}, X2 ′ =
{low, high}. The informed transmitter (i.e., 1) chooses the lowest (resp. highest) transmit power on the current stage 2t + 1
{low, high}. The informed transmitter (i.e., 1) chooses
the lowest (resp. highest) transmit power on the current stage 2t0 + 1
if the upcoming wireless channel state on stage 2t′ + 2 is good for user 2 (resp. 1). If Transmitter 2 can perfectly retrieve the
if the upcoming wireless channel state on stage 2t0 + 2 is good for user 2 (resp. 1). If Transmitter 2 can perfectly retrieve the
power levels of Transmitter 1, it therefore knows the realization of the wireless channel state on stages whose index is even.
power levels of Transmitter 1, it therefore knows the realization of the wireless channel state on stages whose index is even.
It transmits at low (resp. high) power if the channel is good for user 1 (resp. 2). For stages whose index is odd, it chooses its
It transmits at low (resp. high) power if the channel is good for user 1 (resp. 2). For stages whose index is odd, it chooses its
power at random.
power at random.
for achievability, while Theorem 5 and Theorem 11 in Section III-B provide sufficient conditions for
III. I NFORMATION CONSTRAINTS ON ACHIEVABLE EMPIRICAL COORDINATION
achievability in case I and case II, respectively.
We first characterize the sets of achievable empirical coordinations Q ∈ ∆(X ) for the strategies (4)
and
We show condition
that thesefor
sets
consist of distributions in ∆(X ) subject to an information constraint
A. (5).
A necessary
achievability
!
that
captures4.the
imposedinby
the0 ×X
observation
structure. We provide aQ(x
necessary
condition
Theorem
Letrestrictions
Q be a distribution
∆(X
1 ×X2 ) such that ∀ x0 ∈ X0 ,
0 , x1 , x2 ) = ρ0 (x0 ).
x1 ,x(5)
2
for achievability in Theorem 4 and sufficient conditions for strategies (4) and
in Theorem 5 and
In both case I and case II, a distribution Q is an achievable empirical coordination only if it is the
Corollary 12, respectively.
marginal of a distribution Q ∈ ∆(X0 × X1 × X2 × Y) factorizing as
Q(x0 , x1 , x2 , y) = Γ(y|x0 , x1 , x2 )Q(x0 , x1 , x2 ),
∀(x0 , x1 , x2 , y) ∈ X0 × X1 × X2 × Y
(13)
and satisfying the information constraint
August 15, 2017
IQ (X0 ; X2 ) ≤ IQ (X1 ; Y |X0 , X2 )
DRAFT
(14)
9
A. A necessary condition for achievability
Theorem 4. In both cases I and II, if Q ∈ ∆(X ) is an achievable empirical coordination then it must
be the marginal of Q ∈ ∆(X × Y) factorizing as
Q(x0 , x1 , x2 , y) = Γ(y|x0 , x1 , x2 )Q(x1 , x2 |x0 , )ρ0 (x0 ),
(12)
and satisfying the information constraint
(13)
IQ (X0 ; X2 ) ≤ IQ (X1 ; Y |X0 , X2 ).
Proof: Since the strategies of case II are special cases of strategies for case I, we derive the necessary
conditions by considering strategies for case I, in which Agent 2 has causal knowledge of the state X0 .
Let Q ∈ ∆(X ) be an achievable empirical coordination. Note that
N
1 X
E
E TX0N X1N X2N (x0 , x1 , x2 ) =
N
n=1
1{X
0,n
,X1,n ,X2,n =(x0 ,x1 ,x2 )}
N
1 X
PX0,n ,X1,n ,X2,n (x0 , x1 , x2 ),
=
N
n=1
where PX0,n ,X1,n ,X2,n ,Yn is defined in (10). It follows from Appendix A that for ` ∈ {I, II}, there exists
a pair (σn` , τn` )1≤n≤N such that for all (x0 , x1 , x2 ) ∈ X
N
1 XX
lim
PX0,n ,X1,n ,X2,n ,Yn (x0 , x1 , x2 , y) = Q(x0 , x1 , x2 ).
N →∞ N
(14)
n=1 y∈Y
Because of the specific form of (10), this also implies that
N
1 X
lim
PX0,n ,X1,n ,X2,n ,Yn (x0 , x1 , x2 , y) = Q(x0 , x1 , x2 , y)
N →∞ N
(15)
n=1
with Q as in (12). The core of the proof consists in establishing an information constraint on the generic
N
X
1
joint distribution N
PX0,n ,X1,n ,X2,n ,Yn . We start by expressing the quantity H(X0N ) in two different
n=1
manners. On one hand we have that
H(X0N ) = I(X0N ; X0N , Y N )
=
N
X
n=1
=
N
I(X0,n ; X0N , Y N |X0,n+1
)
N
X
n=1
(a)
=
N
X
n=1
August 15, 2017
N
N
N
I(X0,n ; X0n−1 , Y n−1 |X0,n+1
) + I(X0,n ; X0,n
, YnN |X0,n+1
, X0n−1 , Y n−1 )
N
N
N
I(X0,n ; X0n−1 , Y n−1 , X0,n+1
) + I(X0,n ; X0,n
, YnN |X0,n+1
, X0n−1 , Y n−1 )
(16)
(17)
(18)
(19)
DRAFT
10
where (a) follows from the fact that the sequence X0N is i.i.d.. On the other hand we have that
H(X0N ) = I(X0N ; X0N , Y N )
(20)
N
N
= I(X0,n+1
; X0N , Y N ) + I(X0n ; X0N , Y N |X0,n+1
)
=
N
X
n=1
Since 0 ≤
N
N
I(X0,n+1
; X0,n , Yn |X0n−1 , Y n−1 ) + I(X0n ; X0,n , Yn |X0,n+1
, X0n−1 , Y n−1 ) .
(21)
(22)
N
N
N
N
I(X0,n ; X0,n
, YnN |X0,n+1
, X0n−1 , Y n−1 ) = I(X0n ; X0,n , Yn , Yn+1
|X0,n+1
, X0n−1 , Y n−1 )
N
= I(X0n ; X0,n , Yn |X0,n+1
, X0n−1 , Y n−1 ),
we obtain
N
X
n=1
N
I(X0,n ; X0n−1 , Y n−1 , X0,n+1
)=
N
X
n=1
N
I(X0,n+1
; X0,n , Yn |X0n−1 , Y n−1 ).
(23)
Introducing the uniform random variable Z ∈ {1, · · · , N } independent of all others, we rewrite (23) as
N
N
I(X0,Z ; X0,Z+1
, X0Z−1 , Y Z−1 |Z) = I(X0,Z+1
; X0,Z , YZ |X0Z−1 , Y Z−1 , Z).
(24)
We first lower bound the left hand side of (24). Since X0,Z is independent of Z we have that
N
N
I(X0,Z ; X0,Z+1
, X0Z−1 , Y Z−1 |Z) = I(X0,Z ; X0,Z+1
, X0Z−1 , Y Z−1 , Z),
(25)
which expands as
N
I(X0,Z ; X0,Z+1
, X0Z−1 , Y Z−1 , Z) = I(X0,Z ; X0Z−1 , Y Z−1 , Z)
N
+ I(X0,Z ; X0,Z+1
, |X0Z−1 , Y Z−1 , Z). (26)
By definition, X2,Z is a function of (Z, X0Z−1 , Y Z−1 ). Consequently,
I(X0,Z ; X0Z−1 , Y Z−1 , Z) = I(X0,Z ; X2,Z , X0Z−1 , Y Z−1 , Z) ≥ I(X0,Z ; X2,Z ).
(27)
This gives us the desired lower bound for the left term of (24). We now upper bound for the right hand
side of (24) as follows. Using a chain rule, we have
N
N
I(X0,Z+1
; X0,Z , YZ |X0Z−1 , Y Z−1 , Z) = I(X0,Z+1
; X0,Z |X0Z−1 , Y Z−1 , Z)+
N
I(X0,Z+1
; YZ |X0Z−1 , X0,Z , Y Z−1 , Z). (28)
August 15, 2017
DRAFT
11
The last term of (28) is upper bounded as
N
; YZ |X0Z−1 , X0,Z , Y Z−1 , Z)
I(X0,Z+1
N
= H(YZ |X0Z−1 , X0,Z , Y Z−1 , Z) − H(YZ |X0Z−1 , X0,Z , X0,Z+1
Y Z−1 , Z)
(b)
N
= H(YZ |X2,Z , X0Z−1 , X0,Z , Y Z−1 , Z) − H(YZ |X1,Z , X2,Z , X0Z−1 , X0,Z , X0,Z+1
Y Z−1 , Z)
(29)
≤ H(YZ |X0,Z , X2,Z ) − H(YZ |X0,Z , X1,Z , X2,Z )
= I(X1,Z ; YZ |X0,Z , X2,Z ).
where (b) holds because X2,Z is a function of (Z, X0Z−1 , Y Z−1 ) and X1,Z is a function of (Z, X0N ); the inN
equality follows because conditioning reduces entropy and the Markov chain (Z, X0Z−1 , X0,Z+1
, Y Z−1 )−
(X0,Z , X1,Z , X2,Z ) − YZ deduced from from (3). By combining (24)-(29), we find that
I(X0,Z ; X2,Z ) ≤ I(X1,Z ; YZ |X0,Z X2,Z ).
To conclude the proof, note that the joint distribution of X0,Z , X1,Z , X2,Z , and YQ , is exactly the
N
X
1
PX0,n ,X1,n ,X2,n ,Yn and let us introduce function ΦI
distribution N
n=1
ΦI : ∆(X × Y) →
Q
R
,
7→ IQ (X0 ; X2 ) − IQ (X1 ; Y |X0 , X2 )
which is continuous. Because of (15), ∀ε0 > 0 there exists N 0 such that ∀N ≥ N 0 ,
!
N
1 X
I
I
Φ (Q) ≤ Φ
PX0,n ,X1,n ,X2,n ,Yn + ε0 .
N
(30)
(31)
n=1
Theorem 4 has the following interpretation. Agent 2’s actions are represented by X2 and correspond
to a joint source-channel decoding operation with distortion on the information source represented
by X0 . To be achievable, the distortion rate must not exceed the transmission rate allowed by the
channel, whose input and output are represented by Agent 1’s action X1 and the signal Y observed
by Agent 2. Therefore, the pair S = (X0 , X2 ) plays the same role as the side information in statedependent channels [41]. Although we exploit this interpretation when establishing sufficient conditions
for achievability in Section III-B, the argument seems inappropriate to show that the sufficient conditions
are also necessary. In contrast to classical arguments in converse proofs for state-dependent channels [28],
[42], in which the transmitted “message” is independent of the channel state, here the role of the message
is played by the quantity X0N , which is not independent of S N = (X0N , X2N ).
August 15, 2017
DRAFT
12
B. Sufficient conditions for achievability
We start by addressing the special case of distributions Q ∈ ∆(X ) with marginal ρ0 ∈ ∆(X0 ), for
which the distribution Q(x0 , x1 , x2 , y) = Γ(y|x0 , x1 , x2 )Q(x0 , x1 , x2 ) satisfies IQ (X1 ; Y |X0 X2 ) = 0. By
Theorem 4, such Q is an achievable empirical coordination only if IQ (X0 ; X2 ) = 0, so that Q factorizes
as Q(x1 |x0 , x2 )ρ0 (x0 )Q(x2 ). This distribution is trivially achievable by time-sharing between strategies
in which: i) Agent 2 plays a fixed action x2 ; ii) Agent 1 generates actions according to Q(x1 |x0 , x2 ); iii)
playing each strategy with fixed x2 a fraction Q(x2 ) of the time. Hence, we now focus on distributions
Q for which IQ (X1 ; Y |X0 X2 ) > 0
We now characterize achievable empirical coordination for the observation structure of case II in (5).
Theorem 5. Consider the observation structure in case II. Let U be a random variable whose realizations
lie in the alphabet U , |U| < ∞. Let Q ∈ ∆(X ) be with marginal ρ0 ∈ ∆(X0 ). If Q ∈ ∆(X × Y × U)
defined as
Q(x0 , x1 , x2 , y, u) = P(u|x0 , x1 , x2 )Γ(y|x0 , x1 , x2 )Q(x0 , x1 , x2 )
(32)
verifies the constraint
IQ (X0 ; X2 ) < IQ (U ; Y, X2 ) − IQ (U ; X0 , X2 ),
(33)
then Q is an achievable empirical coordination..
Proof: Consider a distribution QX0 X1 X2 Y U ∈ ∆(X × Y × U) that satisfies (32) and (33). We denote
by QU X0 X2 , QU , QX0 X1 X2 , QX0 X2 , and QX2 the resulting marginal distributions.
The crux of the proof is to design strategies from a block-Markov coding scheme that operates over
B blocks of m ≥ 1 actions each. As illustrated in Table II, in every block b ∈ [1 : B − 1], Agent 1
communicates to Agent 2 the actions that Agent 2 should play in block b+1. This is possible by restricting
the actions played by Agent 2 in each block b to a codebook of actions {x2 (ib ) : ib ∈ [1 : 2mR ]}, so
that Agent 1 only has to communicate the index ib to be played in the next block. The problem then
essentially reduces to a joint source-channel coding problem over a state-dependent channel, for which
in every block b:
•
the state is known non-causally by Agent 1, as per the observation structure in (5);
•
Agent 1 communicates with Agent 2 over a state-dependent discrete channel without memory and
with transition probability Γ(y|x0 , x1 , x2 );
•
(b)
the channel state consists of state sequence x0
and action sequence x2 (bib−1 ), where bib−1 is the
index decoded by Agent 2 at the end of block b − 1. Agent 1 only knows x2 (ib−1 ) but the effect of
using bib−1 in place of ib−1 is later proved to be asymptotically negligible;
August 15, 2017
DRAFT
13
•
the action sequence communicated is x2 (ib ), chosen to be empirically coordinated with the state
(b+1)
sequence x0
•
of block b + 1;
(b)
ib is encoded through Gel’fand-Pinsker coding into an action sequence x1 , chosen to be empirically
(b)
coordinated with (x0 , x2 (ib−1 )).
Intuitively, R must be sufficiently large so that one may find a codeword x2 (ib ) coordinated with any
(b+1)
state sequence x0
; simultaneously, R must be small enough to ensure that the index ib is reliably
decoded by Agent 2 after transmission over the channel Γ(y|x0 , x1 , x2 ). The formal analysis of these
conditions, which we develop next, establishes the result.
Unlike the block-Markov schemes used, for instance, in relay channels, in which all nodes may agree
on a fixed message in the first block at the expense of a small rate loss, the first block must be dealt
with more carefully. In fact, we may have to account for an “uncoordinated” transmission in the first
block, in which Agent 1 may not know the actions x2 of Agent 2 and is forced to communicate at rate
b that differs from the rate R used in subsequent blocks. To characterize R
b, we introduce another joint
R
b that factorizes as
distribution Q
b 0 , x1 , x2 , y, u) = Γ(y|x0 , x1 , x2 )P(u|x0 x1 x2 )Q(x1 |x0 x2 )ρ0 (x0 )1{x =x∗ }
Q(x
2
2
(34)
and differs from Q in that X0 is independent of X2 , which is a constant. Assume that IQb (U ; Y, X2 ) −
IQb (U ; X0 , X2 ) = 0 for all P and x∗2 . In particular, for U = X0 , we obtain IQb (X0 ; Y, X2 )−IQb (X0 ; X0 , X2 ) =
0; this is equivalent to HQb (X0 |Y X2 ) = 0, so that x0 must be a function of y and x∗2 . For U = X1 ,
we also obtain IQb (X1 ; Y, X2 ) − IQb (X1 ; X0 , X2 ) = 0, which using the previously established fact leads
to IQb (X1 ; Y |X0 X2 ) = 0. Then, for all (x0 , x2 ) ∈ X0 × X2 , it must be that IQb (X1 ; Y |X0 = x0 , X2 =
x2 ) = 0 and therefore IQ (X1 ; Y |X0 = x0 , X2 = x2 ) = 0. Consequently, IQ (X1 ; Y |X0 X2 ) = 0,
which we have excluded from the analysis. Hence, we can assume that there exist P and x∗2 such that
IQb (U ; Y, X2 ) − IQb (U ; X0 , X2 ) > 0.
b > 0, R
b 0 > 0,
Now, let > 0. Let R > 0, R0 > 0, R
to be specified later. Define
2
> 3 > 2 > 1 > 0 be real numbers and m ≥ 1
R
α , max
,1
b
R
4
B , 1+α
−1 .
(35)
(36)
b in the
Intuitively, α measures the rate penalty suffered from the uncoordinated transmission at rate R
first block. The choice of B merely ensures that
2α
B−1+α
≤ 2 , as exploited later.
Source codebook generation for b = 1. Choose x∗2 such that IQb (U ; Y, X2 ) − IQb (U ; X0 , X2 ) > 0. The
actions of Agent 2 are x∗2 consisting of m repetitions of x∗2 and revealed to both agents.
August 15, 2017
DRAFT
14
Source codebooks generation for b ∈ [2 : B +1]. Randomly and independently generate 2mR sequences
mR ] and reveal them to both agents.
according to Πm
n=1 QX2 (x2,n ), label them x2 (ib ) with ib ∈ [1 : 2
b0
Channel codebook generation for b = 1. Randomly and independently generate 2αm(R +R) sequences
b
αmR ] and j ∈ [1 : 2αmR ], and reveal
b
according to Πm
1
n=1 QU (un ), label them u(i1 , j1 ) with i1 ∈ [1 : 2
b0
b
them to both agents.
0
Channel codebook generation for b ∈ [2 : B]. Randomly and independently generate 2m(R +R)
0
mR ] and j ∈ [1 : 2mR ],
sequences according to Πm
b
n=1 QU (un ), label them u(ib , jb ) with ib ∈ [1 : 2
and reveal them to both agents.
Source encoding at Agent 1 in block b ∈ [1 : B]. At the beginning of block b, Agent 1 uses its
(b+1)
non-causal knowledge of the state x0
(b+1)
(x0
in the next block b + 1 to look for an index ib such that
, x2 (ib )) ∈ Tm
(QX0 X2 ). If there is more than one such index, it chooses the smallest among
1
them, otherwise it chooses ib = 1.
(1)
Channel encoding at Agent 1 in block b = 1. Agent 1 uses its knowledge of (x0 , x∗2 ) to look for an
index j1 such that
(1)
b U X0 X2 )
(Q
u(i1 , j1 ), x0 , x∗2 ∈ Tm
2
(37)
If there is more than one such index, it chooses the smallest among them, otherwise it chooses j1 = 1.
(1)
Finally, Agent 1 generates a sequence x1
(1)
by passing the sequences u(i1 , j1 ), x0 , and x∗2 through a
b X |U X X , and transmits it.
channel without memory and with transition probability Q
1
0 2
(b)
Channel encoding at Agent in block b ∈ [2 : B]. Agent 1 uses its knowledge of (x0 , x2 (ib−1 )) to
look for an index jb such that
(b)
u(ib , jb ), x0 , x2 (ib−1 ) ∈ Tm
(QU X0 X2 )
2
(38)
If there is more than one such index, it chooses the smallest among them, otherwise it chooses jb = 1.
(b)
(b)
Finally, Agent 1 generates a sequence x1 by passing the sequences u(ib , jb ), x0 , and x2 (ib−1 ) through
a channel without memory and with transition probability QX1 |U X0 X2 , and transmits it.
Decoding at Agent 2 in block b = 1. At the end of block 1, Agent 2 observes the sequence of channel
outputs y (1) and knows its sequence of actions x∗2 in block 1. Agent 2 then looks for a pair of indices
(bi1 , b
j1 ) such that
b U Y X2 ).
u(bi1 , b
j1 ), y (1) , x∗2 ∈ Tm
(Q
3
(39)
If there is none or more than one such index, Agent 2 sets bi1 = b
j1 = 1.
August 15, 2017
DRAFT
15
Channel decoding at Agent 2 in block b ∈ [2 : B]. At the end of block b, Agent 2 observes the sequence
of channel outputs y (b) and knows its sequence of actions x2 (bib−1 ) in block b. Agent 2 then looks for a
pair of indices (bib , b
jb ) such that
u(bib , b
jb ), y (b) , x2 (bib−1 ) ∈ Tm
(QU Y X2 ).
3
(40)
If there is none or more than one such index, Agent 2 sets bib = b
jb = 1.
Source decoding at Agent 2 in block b ∈ [1 : B]. Agent 2 transmits x2 (bib−1 ), where bib−1 is its estimate
of the message transmitted by Agent 1 in the previous block b − 1, with the convention that x2 (bi0 ) = x∗2 .
Analysis. We prove that Q is an achievable empirical coordination. We therefore introduce the event
N
N
E , {(X N
/ TN (Q)}
0 , X1 , X2 ) ∈
(41)
with N = mB and we proceed to show that P(E) can be made arbitrarily small for n and B sufficiently
b, and R
b0 . We start by introducing the following events..
large and a proper choice of the rates R, R0 , R
∀b ∈ [1 : N ]
E0 , {(I1 , J1 ) 6= (Ib1 , Jb1 )}
(b+1)
(b)
E1 , {(X 0
, x2 (i0b )) ∈
/ Tm
(QX0 X2 ) ∀ i0b ∈ [1 : 2mR ]}
1
(b)
(b)
0
E2 , {(u(b) (Ib , jb0 ), X 0 , x2 (Ib−1 )) ∈
/ Tm
(QU X0 X2 ) ∀ jb0 ∈ [1 : 2mR ]}
2
n
o
(b)
(b)
(b) (b)
E3 , (u(Ib , Jb ), X 0 , X1 , x2 (Ibb−1 ), Y (b) ) ∈
/ Tm
(Q
)
U
X
X
X
Y
0
1
2
3
n
o
(b)
0 0
E4 , (u(i0b , jb0 ), x2 (Ibb−1 ), Y (b) ) ∈ Tm
(Q)
for
some
(i
,
j
)
=
6
(I
,
J
)
.
b b
b b
3
We start by developing an upper bound for kTxN0 xN1 xN2 − Qk1 , whose proof can be found in Appendix B.
Lemma 6. We have that
B
kTxN0 xN1 xN2 − Qk1 ≤
2α
1 X
+
kTx(b) x(b) x(b) − Qk1 .
0
1
2
B−1+α B−1
b=2
Recalling the choice of B in (36), we therefore have
P(E) = P kTX N0 X N1 X N2 − Qk1 ≥
B
≤P
(42)
1 X
kTX (b) X (b) X (b) − Qk1 ≥
0
1
2
B−1
2
b=2
(43)
!
≤ P kTX (b) X (b) X (b) − Qk1 ≥ for some b ∈ [2 : B]
0
1
2
2
(b)
[
(1)
(b)
(b)
(b)
(b)
≤ P E0 ∪ E1
E1 ∪ E2 ∪ E3 ∪ E4
(44)
(45)
(46)
b=2
August 15, 2017
DRAFT
16
≤ P(E0 ) +
+
B
X
b=2
+
B
X
b=2
B
X
(b)
P(E1 )
b=1
+
B
X
b=2
(b)
)
(b−1)c
∩ E3
(b−1)c
∩ E4
(b−1)c
∩ E4
(b−1)c
∩ E0c )
P(E3 ∩ E1
(b)
(b−1)c
∩ E2
(b)
(b−1)c
∩ E3
P(E4 ∩ E2
(b−1)c
P(E2 |E1
(b−1)c
∩ E0c )
(47)
As proved in Appendix B, the following lemmas show that all the averages over the random codebooks
of the terms above vanish as n → ∞.
b > I b (U ; X0 X2 ) + δ(2 ) and R
b+R
b0 < I b (U ; Y X2 ) − δ(3 ), then
Lemma 7. If R
Q
Q
lim E (P(E0 )) = 0.
n→∞
Lemma 8. If R > IQ (X0 ; X2 ) + δ(1 ), then for any b ∈ [1 : B]
(b)
lim E P(E1 ) = 0.
n→∞
Lemma 9. If R0 > IQ (U ; X0 , X2 ) + δ(2 ), then for any b ∈ [2 : B]
(b))
(b−1)
lim E P(E2 |E1
) = 0.
n→∞
Lemma 10. For any b ∈ [2 : B]
(b)
(b)c
(b−1)c
(b−1)c
(b−1)c
lim E P(E3 |E2 ∩ E2
∩ E3
∩ E4
∩ E0c ) = 0.
n→∞
Lemma 11. If R + R0 < IQ (U ; Y, X2 ) − δ(3 ), then for any b ∈ [2 : B]
(b)
(b−1)c
(b−1)c
(b−1)c
lim E P(E4 ∩ E2
∩ E3
∩ E4
∩ E0c ) = 0.
n→∞
(48)
(49)
(50)
(51)
(52)
Hence, we can find 1 , 2 , and 3 small enough such that limn→∞ E(P(E)) = 0. In particular, there
must exists at least one sequence of codes such that limn→∞ P(E) = 0. Since > 0 can be chosen
arbitrarily small, Q is an achievable empirical coordination.
A few comments are in order regarding the result in Theorem 5. The condition IQ (X0 ; X2 ) <
IQ (U ; Y, X2 ) − IQ (U ; X0 X2 ) ensures bib−1 = ib−1 with high probability as m → ∞, so that the “side
information” x2 (ib−1 ) used by Agent 1 to correlate its actions is identical to the true actions x2 (bib−1 )
of Agent 2; hence, Agent 1 effectively knows the actions of Agent 2 without directly observing them.
(b+1)
Furthermore, the state sequence x0
, which plays the role of the message in block b, is independent
(b)
of the “side information” (x0 , x2 (ib−1 )); this allows us to reuse classical coding schemes for the
transmission of messages over state-dependent channels. However, the proof of Theorem 5 exhibits a
August 15, 2017
DRAFT
17
key difference with the usual Gel’fand-Pinsker coding scheme [28] and its extensions [42]. While using
the channel decoder’s past outputs does not improve the channel capacity, it helps for coordination.
Specifically, a classical Gel’fand-Pinsker coding results would lead to an information constraint
IQ (X0 ; X2 ) < IQ (U ; Y ) − IQ (U ; X0 , X2 )
(53)
which is more restrictive than (33).
Corollary 12. Consider the observation structure in case I. Let Q ∈ ∆(X ) be with marginal ρ0 ∈ ∆(X0 ).
If Q ∈ ∆(X × Y) defined as
Q(x0 , x1 , x2 , y) = Γ(y|x0 , x1 , x2 )Q(x0 , x1 , x2 )
(54)
IQ (X0 ; X2 ) < IQ (X1 ; Y |X0 , X2 ),
(55)
satisfies the constraint
then Q is an achievable empirical coordination.
Proof: Case I differs from Case II by having the state available strictly causally at Agent 2; we
can therefore apply the results of Theorem 5 by providing X0 as a second output to Agent 2. Applying
Theorem 5 with (Y, X0 ) in place of Y , we find that if Q defined as in (32) satisfies IQ (X0 ; X2 ) <
IQ (U ; Y X0 X2 )−IQ (U ; X0 X2 ), then Q is an achievable empirical coordination. Since, IQ (U ; Y X0 X2 )−
IQ (U ; X0 X2 ) = IQ (U ; Y |X0 X2 ), setting U = X1 yields the desired result.
Setting aside the already discussed case of equality in (13), the information constraints of Theorem 4
and Corollary 12 coincide, hence establishing a necessary and sufficient condition for a joint distribution
Q ∈ ∆(X ) to be implementable in case I and a complete characterization of the associated set of
achievable payoffs. This also shows that having Agent 1 select the actions played by Agent 2 and
separating source and channel encoding operations do not incur any loss of optimality. We apply this result
in Section V to an interference network with two transmitters and two receivers, in which Transmitter
1 may represent the most informed agent, such as a primary transmitter [35], [43], Γ may represent an
SINR feedback channel from Receiver 2 to Transmitter 2.
Our results hold under the assumption of perfect monitoring [16] in which Agent 2 perfectly monitors
the actions of Agent 1, i.e., Y = X1 . Equations (13), (55), (33), and (53) then coincide with the
information constraint IQ (X0 ; X2 ) ≤ HQ (X1 |X0 , X2 ) [16], confirming as noted in [16], [17] that
allowing Agent 1 to observe the action of the other agent or providing Agent 2 with the past realizations
of the state X0n−1 does not improve the set of feasible payoffs under perfect monitoring. However, this
observation regarding the set of feasible payoffs may not hold for the set of Nash equilibrium payoffs,
August 15, 2017
DRAFT
18
which are relevant when agents have diverging interests and in which case it matters whether an agent
observes the actions of the others or not. In a power control setting, if the transmitters implement
a cooperation plan that consists in transmitting at low power as long as no transmitter uses a high
power level, see e.g., [44], it matters if the transmitters are able to check whether the others effectively
use a low power level. We focus here on a cooperative setting in which a designer has a precise
objective (maximizing the network throughput, minimizing the total network energy consumption, etc.)
and wants the terminals to implement a power control algorithm with only local knowledge and reasonable
complexity. This setting can be seen as a first step toward analyzing the more general situation in which
agents may have diverging interests; this would happen in power control in the presence of several
operators.
Note that we have not proved whether the information constraint of Theorem 5 is a necessary condition
for implementability in case II. One might be tempted to adopt a side information interpretation of the
problem to derive the converse, since (33) resembles the situation of [45]; however, finding the appropriate
auxiliary variables does not seem straightforward and is left as a refinement of the present analysis.
While Theorem 5 and Corollary 12 have been derived for an i.i.d. random state, the results generalize
to a situation in which the state is constant over L ≥ 1 consecutive stages and i.i.d. from one block of
L stages to the next. For strategies as in case I, the information constraint becomes
1
IQ (X0 ; X2 ) < IQ (X1 ; Y |X0 , X2 ).
L
(56)
In fact, one can reproduce the argument in the proof of Corollary 12 and remark that one can communicate
over the channel at a rate L times larger than the rate required for the covering of the source. Specifically,
to encode m realizations of the random state, the source codebooks must contain 2mR codewords with R >
IQ (X0 ; X2 ); however, the channel codebooks can contain 2mLR codewords with R < IQ (X1 ; Y |X0 X2 ).
The source and channel codes are compatible if mIQ (X0 ; X2 ) < mLIQ (X1 ; Y |X0 X2 ), which is the
desired result in (56). This modified constraint is useful in some wireless communication settings for
which channel states are often block i.i.d.. When L → ∞, which correspond to a single realization of
the random state, the information constraint is always satisfied and any Q ∈ ∆(X ) is implementable.
Finally, we emphasize that the information constraint obtained when coordinating via actions differs
from what would be obtained when coordinating using classical communication [46] with a dedicated
channel. If Agent 1 could communicate with Agent 2 through a channel with capacity C , then all
e ∈ ∆(X ) subject to the information constraint
Q
IQe (X0 ; X2 ) ≤ C
August 15, 2017
(57)
DRAFT
19
would be implementable. In contrast, the constraint IQ (X0 ; X2 ) < IQ (X1 ; Y |X0 X2 ) reflects the following
two distinctive characteristics of communication via actions.
1) The input distribution X1 to the “implicit channel” used for communication between Agent 1 and
Agent 2 cannot be optimized independently of the actions and of the state.
2) The output Y of the implicit channel depends not only on X1 but also on (X0 , X2 ); essentially,
the state X0 and the actions X2 of Agent 2 act as a state for the implicit channel.
Under specific conditions, the coordination via actions may reduce to coordination with a dedicated
channel. For instance, if the payoff function factorizes as w(x0 , x1 , x2 ) , w1 (x1 )w2 (x0 , x2 ) and if the obe 0 , x1 , x2 ) , Q̈(x0 , x2 )Q̇(x1 )
servation structure satisfies (X0 , X2 )−X1 −Y , then any joint distribution Q(x
satisfying the information constraint
IQe (X0 ; X2 ) < IQ̇ (X1 ; Y )
(58)
would be an achievable empirical coordination; in particular, one may optimize Q̇ independently. In
addition, if w1 (x1 ) is independent of x1 , the information constraint further simplifies as
IQe (X0 ; X2 ) < max IQ̇ (X1 ; Y ),
(59)
Q̇
and the implicit communication channel effectively becomes a dedicated channel.
IV. E XPECTED PAYOFF OPTIMIZATION
We now study the problem of determining Q ∈ ∆(X ) that leads to the maximal payoff in case I
and case II. We establish two formulations of the problem: one that involves Q viewed as a function,
and one that explicitly involves the vector of probability masses of Q. Although the latter is seemingly
more complex, it is better suited to numerically determine the maximum expected payoff and turns out
particularly useful in Section V. While we study the general optimization problem in Section IV-A, we
focus on the case of perfect monitoring in Section IV-B, for which we are able to gain more insight into
the structure of the optimal solutions.
August 15, 2017
DRAFT
20
A. General optimization problem
From the results of Section III, the determination of the largest average payoff requires solving the
following optimization problem, with ` ∈ {I, II}:
minimize
−EQ [w(X0 , X1 , X2 )] = −
s.t.
−1 +
∀(x0 , x1 , x2 , y, u) ∈ X × Y × U,
Q(x0 , x1 , x2 , y, u)w(x0 , x1 , x2 )
(x0 ,x1 ,x2 ,y,u)
X
Q(x0 , x1 , x2 , y, u)
(c)
=
0
(x0 ,x1 ,x2 ,y,u)
Q(x0 , x1 , x2 , y, u)
X
− Γ(y|x0 , x1 , x2 )
Q(x0 , x1 , x2 , y, u)
(y,u)
∀x0 ∈ X0 ,
X
−ρ0 (x0 ) +
X
Q(x0 , x1 , x2 , y, u)
(d)
=
(e)
=
0
0
(x1 ,x2 ,y,u)
∀(x0 , x1 , x2 , y, u) ∈ X × Y × U, −Q(x0 , x1 , x2 , y, u)
Φ` (Q)
(f )
≤
0
≤
0
(g)
(60)
where in case I ΦI (·) is defined in (30) while in case II
ΦII (Q) , IQ (X0 ; X2 ) − IQ (U ; Y, X2 ) + IQ (U ; X0 , X2 ).
(61)
We start by addressing the potential convexity of the optimization problem [47]. The objective function
to minimize is linear in Q and the constraints (c), (d), (e), and (f ) restrict the domain to a convex
subset of the unit simplex. Therefore, it suffices to show that the domain resulting from the additional
constraint (g) is convex for the optimization problem to be convex. In case I, for which the set U reduces
to a singleton, the following lemma proves that ΦI is a convex function of Q, which implies that the
additional constraint (g) defines a convex domain.
Lemma 13. The function ΦI is strictly convex over the set of distributions Q ∈ ∆(X × Y) with marginal
ρ0 ∈ ∆(X0 ) that factorize as
Q(x0 , x1 , x2 , y) = Γ(y|x0 , x1 , x2 )ρ0 (x0 )Q(x1 , x2 |x0 ),
(62)
with ρ0 and Γ fixed.
Proof. See Appendix C.
For case II, we have not proved that ΦII is a convex function but, by using a time-sharing argument, it is
always possible to make the domain convex. In the remaining of the paper, we assume this convexification
is always performed, so that the optimization problem is again convex.
August 15, 2017
DRAFT
21
We then investigate whether Slater’s condition holds, so that the Karush-Kuhn-Tucker (KKT) conditions
become necessary conditions for optimality. Since the problem is convex, the KKT conditions would also
be sufficient.
Proposition 14. Slater’s condition holds in cases I and II for irreducible channel transition probabilities
i.e., such that ∀(x0 , x1 , x2 , y) ∈ X0 × X1 × X2 × Y, Γ(y|x0 , x1 , x2 ) > 0.
Proof: We establish the existence of a strictly feasible point in case II, from which the existence for
case I follows as a special case. Consider a distribution Q ∈ ∆(X × Y × U) such that X0 , X1 , and X2
are independent, and U = X1 . We assume without loss of generality that the support of the marginals
QXi , i ∈ {0, 1, 2} is full, i.e., ∀xi ∈ Xi , QXi (xi ) > 0. If the channel transition probability is irreducible,
note that Q(x0 , x1 , x2 , y, u) is then strictly positive, making the constraint (f ) inactive. As for inequality
constraint (g), notice that
IQ (X0 ; X2 ) − IQ (U ; Y, X2 ) + IQ (U ; X0 , X2 ) = 0 − IQ (X1 ; Y ) − I(X1 ; X2 |Y ) + IQ (X1 ; X0 , X2 )
(63)
= −IQ (X1 ; Y ) − IQ (X1 ; X2 |Y )
(64)
= −HQ (X1 ) + HQ (X1 |Y, X2 )
(65)
< 0.
(66)
Hence, the chosen distribution constitutes a strictly feasible point for the domain defined by constraints
(c)-(g), and remains a strictly feasible point after convexification of the domain.
Our objective is now to rewrite the above optimization problem more explicitly in terms of the vector
of probability masses that describes Q. This is useful not only to exploit standard numerical solvers in
Section V, but also to apply the KKT conditions in Section IV-B. We introduce the following notation.
Without loss of generality, the finite sets Xk for k ∈ {0, 1, 2} are written in the present section as set of
indices Xk = [1; nk ]; similarly, we write U = [1 : nu ] and Y = [1 : ny ]. With this convention, we define
a bijective mapping ψ ` : X × Y × U → [1 : n` ] as
ψ ` (i0 , j 0 , k 0 , l0 , m0 ) , m0 + nu (l0 − 1) + nu ny (k 0 − 1) + nu ny n2 (j 0 − 1) + nu ny n2 n1 (i0 − 1),
(67)
which maps a realization (i0 , j 0 , k 0 , l0 , m0 ) ∈ X × Y × U to a unique index ψ ` (i0 , j 0 , k 0 , l0 , m0 ) ∈ [1 : n` ].
We also set nI , n0 n1 n2 ny and nII , n0 n1 n2 ny nu . This allows us to introduce the vector of probability
`
masses q n = (q1 , q2 , . . . , qn` ) for ` ∈ {I, II}, in which each component qi , i ∈ [1 : n` ], is equal to
`
`
Q((ψ ` )−1 (i)), and the vector of payoff values wn = (w1 , w2 , . . . , wn` ) ∈ Rn , in which each component
August 15, 2017
DRAFT
22
`
wi is the payoff of (ψ ` )−1 (i). The relation between the mapping Q (resp. w) and the vector q n (resp.
`
wn ) is summarized in Table III.
Using the proposed indexing scheme, the optimization problem is written in standard form as follows.
n
X̀
−EQ [w(X0 , X1 , X2 )] = − qi wi
minimize
i=1
∀i ∈
n
X̀
s.t.
−1 +
[1; n` ],
qi
− Γi
Θi
(h)
= 0
qi
i=1
∀i ∈ [1 : n0 ], −ρ0 (i) +
(i)
=0
in1 n2 ny nu
X
(j)
=0
qj
j=1+(i−1)n1 n2 ny nu
(k)
∀i ∈ [1 : n` ], −qi
≤ 0
(`)
`
φ` (q n )
where
X
M
Θi =
j∈{1,...,ny nu }
k∈{1,...,n0 n1 n2 }
(68)
≤0
q(k−1)ny nu +j .1{(k−1)ny nu ≤i≤kny nu −1}
(69)
and ∀i ∈ [1 : n0 ], ρ0 (i) = P(X0 = i) and ∀i ∈ [1 : n` ], Γi corresponds to the value of Γ(y|x0 , x1 , x2 ),
according to Table III. As for the function associated with inequality constraint (`), it writes in case II
(case I follows by specialization with |U| = 1) as follows:
II
φII (q n ) = IqnII (X0 ; X2 ) − IqnII (U ; Y, X2 ) + IqnII (U ; X0 , X2 )
= HqnII (X0 ) − HqnII (U, X0 |X2 ) + HqnII (U |Y, X2 )
= HqnII (X0 ) + HqnII (X2 ) − HqnII (X0 , X2 , U ) + HqnII (X2 , Y, U ) − HqnII (X2 , Y )
with
n0
X
HqnII (X0 ) = −
i=1
in1 n2 ny nu
X
j=1+(i−1)n1 n2 ny nu
qj log
in1 n2 ny nu
X
j=1+(i−1)n1 n2 ny nu
qj ,
(70)
(71)
n2
n
n
y nu
y nu
0 n1 n
0 n1 n
X
X
X
X
X
HqnII (X2 ) = −
q(i−1)nu ny +(j−1)n2 ny nu +k log
q(i−1)nu ny +(j−1)n2 ny nu +k ,
i=1
j=1 k=1
j=1 k=1
(72)
n2 ny nu
HqnII (X2 , Y, U ) = −
August 15, 2017
X
i=1
n
0 n1
X
j=1
q(j−1)n2 ny nu +i log
n
0 n1
X
j=1
q(j−1)n2 ny nu +i ,
(73)
DRAFT
23
" n n ny
n0 X
n2 X
nu
0 1X
X
X
q(i−1)n1 n2 ny nu +(j−1)ny nu +k+(l−1)n2 ny nu +(m−1)nu
HqnII (X0 , X2 , U ) = −
l=1 m=1
i=1 j=1 k=1
ny
log
n
0 n1 X
X
q(i−1)n1 n2 ny nu +(j−1)ny nu +k+(l−1)n2 ny nu +(m−1)nu
l=1 m=1
and
n2 ny
HqnII (X2 , Y ) = −
X
i=1
(
#
(74)
,
nu
n
0 n1 X
X
nu
n
0 n1 X
X
q(j−1)n2 ny nu +(i−1)nu +k ) log(
j=1 k=1
j=1 k=1
q(j−1)n2 ny nu +(i−1)nu +k ) . (75)
This formulation is directly exploited in Section IV-B and in Section V.
B. Optimization problem for perfect monitoring
In the case of perfect monitoring, for which Agent 2 perfectly monitors Agent 1’s actions and Y = X1 ,
the information constraints (55) and (33) coincide and
M
I
II
φ(q n ) = φI (q n ) = φII (q n ) = Hqn (X2 ) − Hqn (X2 |X0 ) − Hqn (X1 |X0 , X2 )
(76)
with q n = (q1 , . . . , qn ), n = n0 n1 n2 . To further analyze the relationship between the vector of payoff
values wn and an optimal joint distribution q n , we explicitly express the KKT conditions. The Lagrangian
is
L(q n , λn , µ0 , µn0 , λIC ) = −
n
X
i=1
"
wi qi + λi qi + µ0 −1 +
n
X
i=1
#
qi +
n0
X
j=1
µj −ρ0i +
jn
1 n2
X
i=1+(j−1)n1 n2
qi
+ λIC φ(q n ) (77)
where λn = (λ1 , . . . , λn ), µn0 = (µ1 , . . . , µn0 ), and the subscript IC stands for information constraint.
A necessary and sufficient condition for a distribution q n to be an optimum point is that it is a solution
of the following system:
n
∀ i ∈ [1 : n],
0
X
∂L
∂φ n
= −wi − λi + µ0 +
(q ) = 0
µj 1{1+n1 n2 (j−1)≤i≤jn1 n2 } + λIC
∂qi
∂qi
(78)
j=1
q n verifies (h), (i), (j)
(79)
∀ i ∈ [1 : n], λi ≥ 0
(80)
λIC ≥ 0
(81)
∀ i ∈ [1 : n], λi qi = 0
(82)
λIC φ(q n ) = 0
(83)
August 15, 2017
DRAFT
24
where
∂φ n
(q ) =
∀i ∈ [1 : n],
∂qi
"
−
n0
X
k=1
−
1{1+(k−1)n n ≤i≤kn n } log
n2
X
k=1
1
2
1
2
kn
1 n2
X
j=1+(k−1)n1 n2
1{i∈{k,k+n ,...,k+(n n −1)n }} log
2
0
1
2
qj
n0X
n1 −1
j=0
#
qk+jn2 − 1 + log qi . (84)
In the following, we assume that there exists a permutation of [1 : n] such that the vector of payoff
values wn after permutation of the components is strictly ordered. A couple of observations can then
be made by inspecting the KKT conditions above. First, if the expected payoff were only maximized
under the constraints (h) and (k), the best joint distribution would be to only assign probability to the
greatest element of the vector wn ; in other words the best q n would correspond to a vertex of the unit
simplex ∆(X ). However, as the distribution of the random state fixed by constraint (j), at least n0
components of q n have to be positive. It is readily verified that under constraints (h), (j), and (k), the
optimal solution is that for each x0 the optimal pair (x1 , x2 ) is chosen; therefore, q n possesses exactly n0
positive components. This corresponds to the costless communication scenario. Now, in the presence of
the additional information constraint (`), the optimal solutions contain in general more than n0 positive
components because optimal communication between the two agents requires several symbols of X1 to
be associated with a given realization of the state. In fact, as shown in the following proposition, there
is a unique optimal solution under mild assumptions.
Proposition 15. If there exists a permutation such that the payoff vector wn is strictly ordered, then the
optimization problem (68) has a unique solution.
Proof: Assume λIC = 0 in the Lagrangian. Further assume that a candidate solution of the optimization problem q n has two or more positive components in a block of size n1 n2 associated with a
given realization x0 (see Table III). Then, there exist two indices (i1 , i2 ) such that λi1 = 0 and λi2 = 0.
Consequently, the conditions on the gradient
∂L
∂qi
= 0 for i ∈ {i1 , i2 } imply that wi1 = wi2 , which
contradicts the assumption of wn being strictly ordered under permutation. Therefore, a candidate solution
only possesses a single positive component per block associated with a given realization x0 , which means
that X1 and X2 are deterministic functions of X0 . Hence, Hqn (X2 |X0 ) = Hqn (X1 |X0 X2 ) = 0 and the
information constraint reads Hqn (X2 ) < 0, which is impossible. Hence, λIC > 0.
From Lemma 13, we know that φ(q n ) is strictly convex. Since λIC > 0, the Lagrangian is the sum of
linear functions and a strictly convex function. Since it is also continuous and optimized over a compact
and convex set, there exists a maximum point and it is unique.
August 15, 2017
DRAFT
25
Apart from assuming that wn can be strictly ordered, Proposition 15 does not assume anything on the
values of the components of wn . In practice, for a specific problem it will be relevant to exploit the special
features of the problem of interest to better characterize the relationship between the payoff function
(which is represented by wn ) and the optimal joint probability distributions (which are represented by
the vector q n ). This is one of the purposes of the next section.
V. C ODED POWER CONTROL
We now exploit the framework previously developed to study power control in interference networks. In
this context, the agents are the transmitters and the random state corresponds to the global wireless channel
state, i.e., all the channel gains associated with the different links between transmitters and receivers.
Coded power control (CPC) consists in embedding information about the global wireless channel state
into transmit power levels themselves rather than using a dedicated signaling channel. Provided that the
power levels of a given transmitter can be observed by the other transmitters, the sequence of power
levels can be used to coordinate with the other transmitters. Typical mechanisms through which agents
may observe power levels include sensing, as in cognitive radio settings, or feedback, as often assumed in
interference networks. One of the salient features of coded power control is that interference is directly
managed in the radio-frequency domain and does not require baseband detection or decoding, which
is useful in systems such as heterogeneous networks. The main goal of this section is to assess the
limiting performance of coded power control and its potential performance gains over other approaches,
such as the Nash equilibrium power control policies of a given single-stage non-cooperative game. This
comparison is relevant since conventional distributed power control algorithms, such as the iterative waterfilling algorithm, do not exploit the opportunity to exchange information through power levels or vectors
to implement a better solution, e.g., that would Pareto-dominate the Nash equilibrium power control
policies.
A. Coded power control over interference channels
We first consider an interference channel with two transmitters and two receivers, which we then
specialize to the multiple-access channel in Section V-E to develop and analyze an explicit non-trivial
power control code.
By denoting gij the channel gain between Transmitter i and Receiver j , each realization of the global
wireless channel state is given by
x0 = (g11 , g12 , g21 , g22 ),
August 15, 2017
(85)
DRAFT
26
where gij ∈ G , |G| < ∞; it is further assumed that the channel gains gij are independent and we
set X0 = G 4 . Each alphabet Xi , |Xi | < ∞, i ∈ {1, 2}, represents the set of possible power levels
for Transmitter i. Assuming that the sets are discrete is of practical interest, as there exist wireless
communication standards in which the power can only be decreased or increased by step and in which
quantized wireless channel state information is used. In addition, the use of discrete power levels may
not induce any loss of optimality [48] w.r.t. the continuous case, as further discussed in Section V-B.
We consider three stage payoff functions wrate , wSINR , and wenergy , which respectively represent the
sum-rate, the sum-SINR, and the sum-energy efficiency. Specifically,
wrate :
X
→
(x0 , x1 , x2 ) 7→
wSINR :
2
X
log2
i=1
X
X
→
→
(x0 , x1 , x2 ) 7→
R+
gii xi
1+ 2
σ + g−ii x−i
|
{z
}
,
(86)
SINRi
(x0 , x1 , x2 ) 7→
wenergy :
R+
2
X
i=1
R+
2 F 1+
X
i=1
(87)
,
gii xi
σ 2 + g−ii x−i
gii xi
σ +g−ii x−i
2
xi
.
(88)
The notation −i stands for the transmitter other than i; σ 2 corresponds to the reception noise level;
F : R+ → [0, 1] is a sigmoidal and increasing function that typically represents the block success rate,
see e.g., [49], [50]. The function F is chosen so that wenergy is continuous and has a limit when xi → 0.
The motivation for choosing these three payoff functions is as follows.
•
The sum-rate is a common measure of performance for distributed power control in wireless networks.
•
The sum-SINR is not only a linear approximation of the sum-rate but also an instance of sum-payoff
function that is more sensitive to coordination, since the dependency with respect to the SINR is
linear and not logarithmic.
•
The sum-energy efficiency has recently gathered more attention as a way to study a tradeoff between
the transmission benefit (namely, the net data rate which is represented by the numerator of the
individual payoff) and the transmission cost (namely, the transmit power which is represented by the
denominator of the individual payoff). As pointed out in [51], energy-efficiency can even represent
the energy consumed in a context with packet re-transmissions, indicating its relevance for green
August 15, 2017
DRAFT
27
communications. Finally, as simulations reveal next, total energy-efficiency may be very sensitive
to coordination.
Finally, we consider the following three possible observation structures.
•
Perfect monitoring, in which Agent 2 directly observes the actions of Agent 1, i.e., Y = X1 ;
•
BSC monitoring, in which Agent 2 observes the actions of Agent 1 through a binary symmetric
channel (BSC). The channel is given by the alphabets X1 = {Pmin , Pmax }, Y = {Pmin , Pmax }, and
the transition probability: P(Y = y|X1 = x1 ) = 1 − p if y = x1 and P(Y = y|X1 = x1 ) = p if
y 6= x1 , 0 ≤ p ≤ 1;
•
Noisy SINR feedback monitoring, in which Agent 2 observes a noisy version of the SINR of
Agent 1 as illustrated in Fig. 2; this corresponds to a scenario in which a feedback channel exists
between Receiver 2 and Transmitter 2. The channel is given by the alphabets X0 = G 4 , X1 =
{0, Pmax }, X2 = {0, Pmax }, Y = {SINR(1) , ..., SINR(N ) }, and the transition probability P(Y =
SINR(n) |(X0 , X1 , X2 ) = (x0 , x1 , x2 )) = P(Y = SINR(n) |Y0 = SINR(m) )δSINR(m) −γ(x0 ,x1 ,x2 ) ,
where γ is the function given by the SINR definition (86) and
1 − e if m = n,
P(Y = SINR(n) |Y0 = SINR(m) ) =
e if m = 1 and n = 2, or m = N and n = N − 1,
e else.
2
The performance of coded power control will be assessed against that of the following three benchmark
power control policies.
•
Nash-equilibrium power control (NPC) policy. In such a policy, each transmitter aims at maximizing
an individual stage payoff function ui (x0 , x1 , x2 ). In the sum-rate, sum-SINR, and sum-energy
efficiency cases, these individual stage payoff-functions are respectively given by ui (x0 , x1 , x2 ) =
log2 (1 + SINRi ), ui (x0 , x1 , x2 ) = SINRi , and ui (x0 , x1 , x2 ) =
F (SINRi )
.
xi
In the sum-rate and sum-
NE
SINR cases, the unique Nash equilibrium is (xNE
1 , x2 ) = (Pmax , Pmax ), irrespectively of the value
of x0 , Pmax being the maximal power level for the transmitters. In the sum-energy efficiency case,
the unique non-trivial Nash equilibrium may be determined numerically and generally requires some
knowledge of x0 , depending on how it is implemented (see e.g., [23]).
•
Semi-coordinated power control (SPC) policy. This policy corresponds to a basic coordination
scheme in which Transmitter 1 optimizes its power knowing that Transmitter 2 transmits at full
power; SPC requires the knowledge of the current wireless channel state realization at Transmitter
1. Specifically, x2 = Pmax , x†1 ∈ arg max wr (x0 , x1 , Pmax ), r ∈ {rate, SINR, energy}. SPC is a
x1
rather intuitive scheme, which also corresponds to the situation in which Transmitter 1 only knows
August 15, 2017
DRAFT
33
1−e
SINR(1)
28
y1 = SINR(1)
e
2
e
1−e
SINR(2)
y2 = SINR(2)
e
2
e
2
SINR(3)
1−e
SINR(N −1)
1−e
e
2
y3 = SINR(3)
e
SINR(N )
1−e
yN −1 = SINR(N −1)
yN = SINR(N )
Fig. 4.Fig.
The
observed
by Transmitter
2 lies 2inlies
an in
N −symbol
alphabet
i.e., |Y|
N .=The
correspond
to possible
2. signal
The signal
observed
by Transmitter
an N −symbol
alphabet
i.e.,=|Y|
N .symbols
The symbols
correspond
to possible
valuesvalues
for theforSINR
at
the
receiver
which
is
associated
with
Transmitter
2.
Transmitter
2
observes
the
actual
value
of
its
SINR
the SINR at the receiver which is associated with Transmitter 2. Transmitter 2 observes the actual value of its SINR
with probability
1 − e1while
there there
is probability
e ≥ 0ethat
neighboring
SINRSINR
is observed.
In theInsimulations,
e = 0.1.
with probability
− e while
is probability
≥ 0athat
a neighboring
is observed.
the simulations,
e = 0.1.
i)
log2 (1
SINR
ui (x0 , realizations
x1 , x2 ) = SINR
and uithis
(x0 ,scheme
x1 , x2 ) can
= Fin(SINR
. In
the sum-rate
sumthe+past
and
of the
fact
be
optimal
if Agentand
1 and
Agent
i ),current
i , state;
x
i
NE , xNE ) = (P
SINR2 cases,
the unique
Nash
equilibrium
is (xwith
irrespectively
of the Therefore,
value
only know
X0n and
Y n−1
, respectively,
Agent
2 choosing
the),best
constant action.
max , Pmax
1
2
of x0comparisons
, Pmax beingwith
the SPC
maximal
forthe
thepotential
transmitters.
In an
theadvanced
sum-energy
efficiency
allowpower
us tolevel
assess
gain of
coding
schemecase,
and the
the unique
non-trivial
equilibrium may be determined numerically and generally requires some
value of
knowingNash
the future.
knowledge
of x0 , depending on
howcontrol
it is implemented
(see e.g.,
• Costless-communication
power
(CCPC) policy.
This [23]).
policy corresponds to the situation in
•
Semi-coordinated
powermay
control
(SPC) policy.
policy
to aoptimize
basic coordination
which transmitters
communicate
at not This
cost, so
that corresponds
they may jointly
their powers to
scheme
in which
Transmitter
1 optimizes
its power
thatInTransmitter
transmits
at full
achieve
the maximum
of the
payoff function
wr at knowing
every stage.
such a case 2there
is no information
power;
SPC requires
theperformance
knowledge of CCPC
the current
wireless
channel
stateforrealization
at Transmitter
constraint,
and the
provides
an upper
bound
the performance
of all other
†
1. Specifically,
wr (x0 , x1 , Pmax ), r ∈ {rate, SINR, energy}. SPC is a
policies. x2 = Pmax , x1 ∈ arg max
x
1
rather
schemesignal-to-noise
an engineer might
fact, SPC
The intuitive
communication
ratio think
(SNR)of.
is In
defined
as also corresponds to the situation
where Transmitter 1 only knows the past and current realizations
Pmax of the state; this scheme can in
(89)
SNR(dB) , 10 log10 2 .
σ ). Optimality is obtained when Agent
fact be optimal if Agent 1 (resp. 2) only knows X0t (resp. Y t−1
2 chooses
constant
action. Therefore, the comparisons we made allow one both to assess
B.
Influencetheofbest
the payoff
function
the potential gain an advanced coding scheme might bring over a quite simple and natural scheme
The objective of this subsection is to numerically assess the relative performance gain of CPC over SPC
and to assess the value of knowing the future.
in the case of perfect monitoring. We assume that the channel gains gij ∈ {gmin , gmax } are Bernoulli
August
July 28,
2017 15, 2017
DRAFT
DRAFT
29
distributed gij ∼ B(pij ) with pij , P(gij = gmin ); with our definition of X0 in (85), this implies
that |X0 | = 16. All numerical results in this subsection are obtained for gmin = 0.1, gmax = 2 and
(p11 , p12 , p21 , p22 ) = (0.5, 0.1, 0.1, 0.5). The sets of transmit powers X1 , X2 are both assumed to be the
same alphabet of size four {P1 , P2 , P3 , P4 }, with P1 = 0, P2 =
Pmax
3 , P3
=
2Pmax
3 , P4
= Pmax . The
quantity Pmax is given by the operating SNR and σ 2 = 1. The function F is chosen as a typical instance
of the efficiency function used in [50], i.e.,
20.9 − 1
F (x) = exp −
x
(90)
.
For all r ∈ {rate, SINR, energy}, the relative performance gain with respect to the SPC policy is
r
?
EQ (w )
Relative gain (%) =
− 1 × 100
(91)
Eρ0 (max w(x0 , x1 , Pmax ))
x1
?
where Q is obtained by solving optimization problem (68) under perfect monitoring. This optimization
is numerically performed using the Matlab function fmincon. Fig. 3 illustrates the relative performance
gain in % w.r.t. the SPC policy for the sum-energy efficiency, while Fig. 4 illustrates it for the sum-SINR
and sum-rate.
As shown in Fig. 3, our simulation results suggest that CPC provides significant performance gains
for the sum-energy efficiency. This may not be surprising, as the payoff function (88) is particularly
sensitive to the lack of coordination; in fact, as the transmit power becomes high,
F (SINRi )
xi
→
1
xi ,
which
means that energy efficiency decreases rapidly. As shown in Fig. 4, the performance gains of CPC for
the sum-SINR and the sum-rate are more moderate, with gains as high as 43% for the sum-SINR and
25% for the sum-rate; nevertheless, such gains are still significant, and would be larger if we used NPC
instead of SPC as the reference case, as often done in the literature of distributed power control. The
shape of the sum-rate curve in Fig. 4 can be explained intuitively. At low SNR, interference is negligible
and the sum-rate is maximized when both transmitters use full power, which is also what SPC does in
this regime. At high SNR, SPC is not optimal but still provides a large sum-rate, which is comparable
to that provided by the best CPC scheme. Between these regimes, advanced coordination schemes are
particularly useful, which explains the peak at intermediate SNR.
X ?
?
?
Q (x0 , x1 , x2 ), QX2 (x2 ) =
We conclude this subsection by providing the marginals QX1 (x1 ) =
x0 ,x2
X ?
X ?
?
Q (x0 , x1 , x2 ), and joint distribution QX1 X2 (x1 , x2 ) =
Q (x0 , x1 , x2 ) of the optimal joint distri-
x0 ,x1
x0
bution for CPC and CCPC in Table IV and Table V, respectively. In both cases, the results correspond
to the maximization of the sum-rate payoff function wrate and SNR = 10 dB. Table IV shows that,
without information constraint, the sum-rate is maximized when the transmitters correlate their power
August 15, 2017
DRAFT
30
800
Expected sum energy-efficiency
Relative gain (%)
700
600
500
400
300
200
−10
0
10
20
30
40
Signal-to-Noise Ratio (dB)
Fig. 3. Relative sum-energy gain of coded power control (CPC) with perfect monitoring over semi-coordinated power control
(SPC).
Relative gain (%)
40
30
Expected sum-SINR
Expected sum-rate
20
10
0
−10
0
10
20
30
40
Signal-to-Noise Ratio (dB)
Fig. 4. Relative sum-SINR gain and sum-rate gain of coded power control (CPC) with perfect monitoring over semi-coordinated
power control (SPC).
August 15, 2017
DRAFT
31
levels so that only three pairs of transmit power levels are used out of 16. This result is consistent
with [48], which proves that, for interference channels with two transmitter-receiver pairs, there is no
loss of optimality in terms of wrate by operating over a binary set {0, Pmax } instead of a continuous
interval [0, Pmax ]. Interestingly, as seen in Table V, the three best configurations of the CCPC policy are
exploited 44.3+42.9+2.1 = 89.3% of the time in the CPC policy, despite the presence of communication
constraints between the two transmitters.
C. Influence of the observation structure
In this subsection, we focus on the observation structure defined by case I in (4) and we restrict
our attention to the sum-rate payoff function wrate . The set of powers is restricted to a binary set
X1 = X2 = {0, Pmax }, but unlike the study in Section V-B, we do not limit ourselves to perfect
monitoring. Fig. 5 shows the relative performance gain w.r.t. the SPC policy as a function of SNR for
three different observation structures. The performance of CPC for BSC monitoring is obtained assuming
a probability of error of 5%, i.e., Z1 ∼ B(0.05), P(Z1 = 1) = 0.05. The performance of CPC for noisy
SINR feedback monitoring is obtained assuming e = 0.1; in this case, it can be checked that the SINR
can take one of N = 7 distinct values.
Fig. 5 suggests that CPC provides a significant performance gain over SPC over a wide range of
operating SNRs irrespective of the observation structure. Interestingly, for SNR = 10 dB, the relative
gain of CPC only drops from 22% with perfect monitoring to 18% with BSC monitoring, which suggest
that for observation structures with typical noise levels the benefits of CPC are somewhat robust to
observation noise. Similar observations can be made for SINR feedback monitoring. Note again that one
would obtain higher performance gains by considering NPC as the reference policy or by considering
scenarios with stronger interference.
D. Influence of the wireless channel state knowledge
In this subsection, we restrict our attention to CPC with BSC monitoring with the same parameters as
in Section V-C, but we consider both Case I and Case II defined in (4) and (5), respectively. The results
for Case II are obtained assuming that |U| = 10. While we already know that the performance of CPC
is the same in Case I and Case II with perfect monitoring, the results in Fig. 6 suggest that, for typical
values of the observation noise, not knowing the past realizations of the global wireless channel state at
Transmitter 2 only induces a small performance loss.
August 15, 2017
DRAFT
32
CCPC
CPC w/ perfect monitoring
CPC w/ SINR feedback monitoring
CPC w/ BSC monitoring
30
Relative gain (%)
25
20
15
10
5
0
−10
0
10
20
30
40
Signal-to-Noise Ratio (dB)
Fig. 5. Relative sum-rate gain of costless communication power control (CCPC) and coded power control (CPC) over semicoordinated power control (SPC) under various monitoring assumptions in the observation structure of Case I.
CCPC
CPC in Case I
CPC in Case II
30
Relative gain (%)
25
20
15
10
5
0
−10
0
10
20
30
40
Signal-to-Noise Ratio (dB)
Fig. 6. Relative sum-rate gain of coded power control (CPC) and costless communication power control (CCPC) over semicoordinated power control (SPC) for the binary symmetric channel (BSC) monitoring in the observation structure of Case I and
Case II.
August 15, 2017
DRAFT
CPC in Case II
Relative gain (%)
25
20
33
15
E. Influence of the coordination scheme
10
In this last subsection, we assess the benefits of CPC for an explicit code that operates over blocks
5
of length n = 3. To simplify the analysis and clarify the interpretation, several assumptions are made.
0
First, we consider a multiple-access
channel, which is a special case of the interference channel studied
−10
0
10
20
30
40
earlier with two transmitters and a single receiver,
so that the
global
Signal-to-Noise
Ratio
(dB)wireless channel state comprises only
two components (g1 , g2 ). Second, we assume that the global wireless channel state X0 takes values in
Fig. 8. Relative sum-rate gain of coded power control (CPC) and costless communication power control (CCPC) over semi-
the binary alphabet X0 ∈ {(gmin , gmax ), (gmax , gmin )}, and is distributed according to Bernoulli random
coordinated power control (SPC) for the binary symmetric channel (BSC) monitoring in the observation structure of Case I and
variable
Case
II.
B(p) with p = P X0 = (gmin , gmax ) . In the remaining of this subsection, we identify the
realization (gmin , gmax ) with “0” and (gmax , gmin ) wth “1,” so that we may write X0 = {0, 1}. Third, we
assume that the transmitters may only choose power values in {Pmin , Pmax }, and we identify power Pmin
assume that the transmitters may only choose power values in {Pmin , Pmax }, and we identify power Pmin
with “0” and power Pmax with “1”, so that we may also write X1 = X2 = {0, 1}. Finally, we consider
with “0” and power Pmax with “1”, so that we may also write X1 = X2 = {0, 1}. Finally, we consider
the case of perfect monitoring and we restrict our attention to the sum-SINR payoff function wSINR .
the case of perfect monitoring and we restrict our attention to the sum-SINR payoff function wSINR .
The values of the payoff function used in numerical simulations are provided in Fig. 7 as the entries in
The values of the payoff function used in numerical simulations are provided in Fig. 9 as the entries in
a matrix. Each matrix corresponds to a different choice of the wireless channel state x0 ; in each matrix,
a matrix. Each matrix corresponds to a different choice of the wireless channel state x0 ; in each matrix,
the choice of the row corresponds to the action x1 of Transmitter 1, which the choice of the column
the choice of the row corresponds to the action x1 of Transmitter 1, which the choice of the column
corresponds to the action x2 of Transmitter 2.
corresponds to the action x2 of Transmitter 2.
Pmin Pmax
Pmin
0
20
Pmax
1
≃ 10
Pmin Pmax
Pmin
0
Pmax 20
x0 = (gmin , gmax )
1
≃ 10
x0 = (gmax , gmin )
Fig. 9. Payoff matrix of wSINR for power control over multiple-access channel. Numerical values of the payoff correspond
Fig. 7. Payoff matrix of wSINR for power control over multiple-access channel. Numerical values of the payoff correspond
to gmin = 0.1, gmax = 2, σ 2 = 1, Pmin = 0, Pmax = 10. For example, for the left matrix, when the power profile is
to gmin = 0.1, gmax = 2, σ 2 = 1, Pmin = 0, Pmax = 10. For example, for! the
leftgmin
matrix,Pmax
when
the" power profile is
Pmax
gmax
.
(p1 , p2 ) = (Pmax , Pmax ) the individual SINRs are given by: (SINR1 , SINR2 ) = 1+P
, 1+Pmax
g
g
max
max
min
Pmax gmin
Pmax gmax
(p1 , p2 ) = (Pmax , Pmax ) the individual SINRs are given by: (SINR1 , SINR2 ) = 1+P
,
.
g
1+P
g
max max
max min
The coordination code of length 3 that we develop next can be seen as a separate source channel code,
The coordination code of length 3 that we develop next can be seen as a separate source channel code,
which consists of a source code with distortion and a channel code with side information. The source
July 28, 2017
DRAFT
encoder and decoder are defined by the mappings
fS : X03 → {m0 , m1 }
x0
August 15, 2017
7→
,
(92)
i
DRAFT
34
gS : {m0 , m1 } → X23
7→
i
(93)
.
x2
Note that the chosen source code only uses 2 messages {m0 , m1 } to represent the 8 possible sequences
x0 . One of the benefits of this restriction is that it becomes computationally feasible to find the best two-
message code by brute force enumeration. Finding a more systematic low-compelxity design approach
to coordinatinon codes goes beyond the scope of the present work. The exact choice of fS and gS is
provided after we describe the channel code.
In each block b, Transmitter 1’s channel encoder implements the mapping
(b)
(b)
(b)
(94)
x1 = fC (x0 , x2 , ib+1 )
(b+1)
where ib+1 = fS (x0
(b+1)
) is the index associated with the sequence x2
. The idea behind the design
of the channel encoder fC is the following. If Transmitter 1 did not have to transmit the index ib+1 ,
its optimal encoding would be to exploit its knowledge of (x30 (b), x32 (b)) to choose the sequence x31 (b)
resulting in the highest average payoff in block b. However, to communicate the index ib+1 , Transmitter
1 will instead choose to transmit the sequence x31 (b) with the highest average payoff in block b if
(b)
ib+1 = m0 , or the sequence x1 with the second highest average payoff in block b if ib+1 = m1 . Note
(b)
(b)
(b)
that Transmitter 2 is able to perfectly decode this encoding given its knowledge of x0 , x2 , and x1 at
the end of block b. Formally, fC is defined as follows. The sequence x1 is chosen as
x1 = x01 ⊕ d
(95)
where the modulo-two addition is performed component-wise,
x01 ∈ arg max3
x1 ∈X1
3
X
wSINR (x0,n , x1,n , x2,n ),
n=1
d = (0, 0, 0) if ib+1 = m0 ,
3
d ∈ arg max
(96)
3
X
d s.t. Ω(d)=1 n=1
wSINR (x0,n , x01,n ⊕ dn , x2,n ) if ib+1 = m1
(97)
(98)
where Ω is the Hamming weight function that is, the number of ones in the sequence d ∈ {0, 1}3 . If the
argmax set is not a singleton set, we choose the sequence with the smallest Hamming weight.
To complete the construction, we must specify how the source code is designed. Here, we choose the
mappings fS and gS that maximize the expected payoff E(wSINR ) knowing the operation of the channel
code. The source code resulting from an exhaustive search is given in Table VI, and the corresponding
channel code is given in Table VII. The detailed expression of the expected payoff required for the search
is provided in Appendix D.
August 15, 2017
DRAFT
35
The proposed codes admit to an intuitive interpretation. For instance, the first line of Table VII indicates
that if the channel is bad for Transmitter 1 for the three stages of block b, then Transmitter 1 remains
silent over the three stages of the block while Transmitter 2 transmits at all three stages. In contrast,
the last line of Table VII shows that if the channel is good for Transmitter 1 for the three stages of
block b, then Transmitter 1 transmit at all stages while Transmitter 2 remains silent two thirds of the
time. While this is suboptimal for this specific global wireless channel state realization, this is required
to allow coordination and average optimality of the code.
To conclude this section, we compare the performance of this short code with the best possible
performance that would be obtained with infinitely long codes. As illustrated in Fig. 8, while the
performance of the short code suffers from a small penalty compared to that of ideal codes with infinite
block length, it still offers a significant gain w.r.t. the SPC policy and it outperforms the NPC policy.
100
CCPC
CPC length n → ∞
CPC length n = 3
SPC
NPC
Expected payoff (a.u.)
80
60
40
20
0
0
10
20
30
40
50
Signal-to-Noise Ratio (dB)
Fig. 8. Expected payoff versus SNR for different power control policies.
VI. C ONCLUSION
In this paper, we adopted the view that distributed control policies or resource allocation policies in
a network are joint source-channel codes. Essentially, an agent of a distributed network may convey its
knowledge of the network state by encoding it into a sequence of actions, which can then be decoded by
the agents observing that sequence. As explicitly shown in Section V-E, the purpose of such “coordination
codes” is neither to convey information reliably nor to meet a requirement in terms of maximal distortion
level, but to provide a high expected payoff. Consequently, coordination codes must implement a trade-off
August 15, 2017
DRAFT
36
between sending information about the future realizations of the network state, which plays the role of
an information source and is required to coordinate future actions, and achieving an acceptable payoff for
the current state of the network. Considering the large variety of payoff functions in control and resource
allocation problems, an interesting issue is whether universal codes performing well within classes of
payoff functions can be designed.
Remarkably, since a distributed control policy or resource allocation policy is interpreted as a code,
Shannon theory naturally appears to measure the efficiency of such policies. While the focus of this paper
was limited to a small network of two agents, the proposed methodology to derive the best coordination
performance in a distributed network is much more general. The assumptions made in this paper are
likely to be unsuited to some application scenarios, but provide encouraging preliminary results to further
research in this direction. For example, as mentioned in Section I, a detailed comparison between coded
power control and iterative water-filling like algorithms would lead to consider a symmetric observation
structure while only an asymmetric structure is studied in this paper. The methodology to assess the
performance of good coded policies consists in deriving the right information constraint(s) by building the
proof on Shannon theory for the problem of multi-source coding with distortion over multi-user channels
wide side information and then to use this constraint to find an information-constrained maximum of
the payoff (common payoff case) or the set of Nash equilibrium points which are compatible with the
constraint (non-cooperative game case). As a key observation of this paper, the observation structure of
a multi-person decision-making problem corresponds in fact to a multiuser channel. Therefore, multiterminal Shannon theory is not only relevant for pure communication problems but also for any multiperson decision-making problem. The above observation also opens new challenges for Shannon-theorists
since decision-making problems define new communication scenarios.
A PPENDIX A
ACHIEVABLE EMPIRICAL COORDINATION AND IMPLEMENTABILITY
Assume Q is an achievable empirical coordination. Then, for any > 0,
||E(TX N ) − Q||1 ≤ E(||TX N − Q||1 )
(99)
= E(||TX N − Q||1 | ||TX N − Q||1 ≥ )P(||TX N − Q||1 ≥ )
+ E(||TX N − Q||1 | ||TX N − Q||1 < )P(||TX N − Q||1 < )
≤ 2P(||TX N − Q||1 ≥ ) + .
Hence, ∀ > 0
August 15, 2017
(100)
(101)
limN →∞ ||E(TX N ) − Q||1 ≤ , which means that Q is implementable.
DRAFT
37
A PPENDIX B
L EMMAS USED IN PROOF OF T HEOREM 5
A. Proof of Lemma 6
Recall that N = αm + (B − 1)m with our coding scheme. Note that
kTxN0 xN1 xN2 − Qk1
=
N
X
1
1
− Q(x0 , x1 , x2 )
N {(x0,n ,x1,n ,x2,n )=(x0 ,x1 ,x2 )}
X
(102)
x0 ,x1 ,x2 n=1
=
B X
m
αm
X
X
1
1
1 (1) (1) (1)
+
1 (b) (b) (b)
− Q(x0 , x1 , x2 )
N {(x0,n ,x1,n ,x2,n )=(x0 ,x1 ,x2 )}
N {(x0,n ,x1,n ,x2,n )=(x0 ,x1 ,x2 )}
X
x0 ,x1 ,x2 n=1
≤
b=2 n=1
(103)
αm
X
αm
1
1 (1) (1) (1)
−
Q(x0 , x1 , x2 )
N {(x0,n ,x1,n ,x2,n )=(x0 ,x1 ,x2 )}
N
X
x0 ,x1 ,x2 n=1
+
B
X
X
m
X
1
N − αm
Q(x0 , x1 , x2 )
1 (b) (b) (b)
−
N {(x0,n ,x1,n ,x2,n )=(x0 ,x1 ,x2 )} N (B − 1)
(104)
b=2 x0 ,x1 ,x2 n=1
=
αm
αm X X 1
1 (1) (1) (1)
− Q(x0 , x1 , x2 )
N x ,x ,x
αm {(x0,n ,x1,n ,x2,n )=(x0 ,x1 ,x2 )}
0
1
2
n=1
B
m
mX X X 1
+
1 (b) (b) (b)
− Q(x0 , x1 , x2 )
N
m {(x0,n ,x1,n ,x2,n )=(x0 ,x1 ,x2 )}
x ,x ,x
b=2
0
1
2
(105)
n=1
B
2α
1 X
≤
+
kTx(b) x(b) x(b) − Qk1 ,
0
1
2
B−1+α B−1
(106)
b=2
where (104) follows from the triangle inequality and (106) follows from
for (P, Q) ∈ ∆2 (X ).
m
N
≤
1
B−1
and kP − Qk1 ≤ 2
B. Proof of Lemma 7
The proof of this result is similar to that of the following Lemmas, using the “uncoordinated” distrib instead of Q. For brevity, we omit the proof.
bution Q
C. Proof of Lemma 8
Note that
(b+1)
(b)
, X 2 (i0b )) ∈
/ Tn1 (QX0 X2 ) for all i0b ∈ [1 : 2nR ]
E P(E1 ) = P (X 0
August 15, 2017
(107)
DRAFT
38
Q
(b)
with X 0 distributed according to nn=1 QX0 , and X 2 (ib ) independent of each other distributed according
Q
to m
n=1 QX2 . Hence, the result directly follows from the covering lemma [41, Lemma 3.3], with the
following choice of parameters.
(b+1)
X̂ n (m) with m ∈ A ← X 2 (i0b ) with i0b ∈ [1 : 2nR ].
Xn ← X0
U ←∅
D. Proof of Lemma 9
Note that
0
(b)
(b−1)c
(b)
/ Tn2 (QU X0 X2 ) for all jb0 ∈ [1 : 2nR ] , (108)
E P(E2 |E1
) = P (U (Ib , jb0 ), X 0 , X 2 (Ib−1 )) ∈
(b)
where (X 0 , X 2 (Ib−1 )) ∈ Tn1 (QX0 X2 ), and U (Ib , j) are generated independently of each other according
Q
to nn=1 QU . Hence, the result follows directly from the covering lemma [41, Lemma 3.3] with the
following choice of parameters.
(b)
X n ← X 0 , X 2 (Ib−1 )
U ←∅
0
X̂ n (m) with m ∈ A ← U (Ib , jb0 ) with jb0 ∈ [1 : 2nR ].
E. Proof of Lemma 10
The result follows from a careful application of the conditional typicality lemma. Note that condition(b)c
ing on E2
(b−1)c
(b)
(b−1)c
ensures that (u(Ib , Jb ), X 0 , x2 (Ib−1 )) ∈ Tn2 (QU X0 X2 ), while conditioning on E2
∩
(109)
(b−1)c
∩ E0c guarantees that Ibb−1 = Ib−1 . Consequently,
(b)
(b)c
(b−1)c
(b−1)c
(b−1)c
P E3 ∩ E2 ∩ E2
∩ E3
∩ E4
∩ E0c
(b)
(b)
/ Tn3 (QX0 X2 X1 Y )
≤ P (u(Ib , Jb ), X 0 , X 1 , x2 (Ibb−1 ), Y (b) ) ∈
E3
=
∩ E4
X
(b)
|(u(Ib , Jb ), X 0 , x2 (Ib−1 )) ∈ Tn2 (QU X0 X2 ) ∩ Ibb−1 = Ib−1
XX
(b) (b)
pIbb−1 ,Ib ,Jb ,X (b) (ib−1 , ib , jb , x0 )
Γ(y|x0 , x1 , x2 (ib−1 ))
0
(b)
ib−1 ,ib ,jb ,x0
y
(b)
x1
Q(x1 |u(ib , jb ), x0 , x2 (ib−1 ))1{(u(ib ,jb ),x(b) ,x(b) ,x(b) (ib−1 ),y)∈T
/ n (QU X
(b)
(b)
0
1
2
3
0 X1 X2 Y
)} ,
(110)
(b)
(b)
where pIbb−1 ,Ib ,Jb ,X (b) denotes the joint distribution of Ibb−1 , Ib , Jb , X 0 given (u(Ib , Jb ), X 0 , x2 (Ib−1 )) ∈
0
Tn2 (QU X0 X2 )
and Ibb−1 = Ib−1 . Upon taking the average over the random codebooks, we obtain
(b)
(b−1)c
(b−1)c
(b−1)c
(b−1)c
E P E3 |E1
∩ E2
∩ E3
∩ E4
∩ E0c
XX
X
(b)
(b) (b)
=
E pIbb−1 ,Ib ,Jb ,X (b) (ib−1 , ib , jb , x0 )E
Γ(y|x0 , x1 , X 2 (ib−1 ))
0
y
(b)
ib−1 ,ib ,jb ,x0
(b)
x1
U
(i
,
j
),
X
(i
))
.
b b
2 b−1
0 X1 X2 Y }
Q(x1 |U (ib , jb ), x0 , X 2 (ib−1 ))1{(u(ib ,jb ),x(b) ,x(b) ,X (b) (ib−1 ),y)∈T
/ n (QU X
(b)
(b)
0
August 15, 2017
1
2
3
(111)
DRAFT
39
The inner expectation is therefore
P (u, x0 , X 1 , x2 , Y ) ∈
/ Tn3 (QU X0 X1 X2 Y ) ,
where (X 1 , Y ) is distributed according to
Qm
n=1 QX1 |U X0 X2 (x1,n |u1,n , x2,n x0,n )Γ(yn |x0,n x1,n x2,n )
(112)
given
(u, x0 , x2 ) ∈ Tn2 (QX0 X2 ). The conditional typicality lemma [41, p. 27] guarantees that (112) vanishes
as n → ∞.
F. Proof of Lemma 11
The result is a consequence of the packing lemma. Note that
(b)
(b−1)c
(b−1)c
(b−1)c
E P E4 ∩ E2
∩ E3
∩ E4
∩ E0c
(b)
≤ E P E4 |Ibb−1 = Ib−1
= P (U (i0b , jb0 ), X 2 (Ibb−1 ), Y (b) ) ∈ Tn2 (Q) for some (i0b , jb0 ) with i0b 6= Ib |Ibb−1 = Ib−1
(b−1)c
(113)
(114)
(b−1)c
∩ E4
∩ E0c guarantees that Ibb−1 = Ib−1 . Since every U (i0b , jb0 )
Q
with i0b =
6 Ib is generated according to m
i=1 pU (ui ) independently of (Y , X 2 (Ib−1 )), byt the packing
since conditioning on E2
(b−1)c
∩ E3
lemma [41, Lemma 3.1] we know that if R + R0 < IQ (U ; Y, X2 ) − δ(3 ) then
P (U (i0b , jb0 ), X 2 (Ibb−1 ), Y (b) ) ∈ Tn2 (Q) for some (i0b , jb0 ) with i0b 6= Ib |Ibb−1 = Ib−1
vanishes as n → ∞.
A PPENDIX C
P ROOF OF L EMMA 13
The function Φ = ΦI can be rewritten as Φ(Q) = HQ (X0 ) − HQ (Y, X0 |X2 ) + HQ (Y |X0 , X2 , X1 ).
P
The first term HQ (X0 ) = − x0 ρ0 (x0 ) log ρ0 (x0 ) is a constant w.r.t. Q. The third term is linear w.r.t.
Q since, with Γ fixed,
HQ (Y |X0 , X2 , X1 ) = −
X
Q(x0 , x1 , x2 , y) log Γ(y|x0 , x1 , x2 ).
(115)
x0 ,x1 ,x2 ,y
It is therefore sufficient to prove that HQ (Y, X0 |X2 ) is concave. Let λ1 ∈ [0, 1], λ2 = 1 − λ1 , (Q1 , Q2 ) ∈
∆2 (X0 × X1 × X2 × Y) and Q = λ1 Q1 + λ2 Q2 . We have that:
P
X X
x1 ,i λi Qi (x0 , x1 , x2 , y)
P
HQ (Y, X0 |X2 ) = −
λi Qi (x0 , x1 , x2 , y) log
i λi Qi (x2 )
x0 ,x2 ,y
x1 ,i
P
P
X X X
i λi
x1 Qi (x0 , x1 , x2 , y)
P
=−
λi
Qi (x0 , x1 , x2 , y) log
i λi Qi (x2 )
x ,x ,y
x
0
August 15, 2017
2
i
(116)
(117)
1
DRAFT
40
X X
X
P
Qi (x0 , x1 , x2 , y)
>−
λi
Qi (x0 , x1 , x2 , y) log
λi Qi (x2 )
x0 ,x2 ,y
x1
i
P
X
X
X
x1 Qi (x0 , x1 , x2 , y)
=−
λi
Qi (x0 , x1 , x2 , y) log
Qi (x2 )
x ,x ,y
x
i
0
2
λi
x1
(118)
(119)
1
(120)
= λ1 HQ1 (Y, X0 |X2 ) + λ2 HQ2 (Y, X0 |X2 )
where the strict inequality comes from the log-sum inequality [36], with:
ai = λi Qi (x0 , x1 , x2 )
(121)
bi = λi Qi (x2 )
(122)
and
for i ∈ {1, 2} and for all (x0 , x1 , x2 ) such that Qi (x2 ) > 0.
A PPENDIX D
E XPRESSION OF THE EXPECTED PAYOFF W3
WHICH ALLOWS THE BEST MAPPINGS fS AND gS TO BE
SELECTED
We introduce the composite mapping χS = gS ◦ fS . For the channel code defined in Section V-E,
the expected payoff only depends on the mappings fS and χS , and we denote it by W3 (fS , χS ). The
following notation is used below: χS (x30 ) = (χ1 (x30 ), χ2 (x30 ), χ3 (x30 )) to stand for the three components
of χS .
It can be checked that
X
W3 (fS , χS ) =
(123)
Wijk (fS , χS )
(i,j,k)∈{0,1}3
where:
3
W000 (fS , χS ) =
3
0
p
|{z}
P[x [b]=(0,0,0)]
P0 (fS ) ×
| {z }
P[ib+1 =m0 ]
+ (1 − P0 (fS )) ×
max
|
x31 ∈{0,1}3 \X1000
X1000 = arg 3 max {
x1 ∈{0,1}3
August 15, 2017
|
max {
x31 ∈{0,1}3
{
3
X
n=1
3
X
w(0, x1,n , χn ((0, 0, 0)))}
n=1
{z
Best payoff
w(0, x1,n , χn ((0, 0, 0)))}
{z
Second best payoff
3
X
w(0, x1,n , χn ((0, 0, 0)))},
!
,
!
}
(124)
}
(125)
n=1
DRAFT
41
W001 (fS , χS ) = p (1 − p) P0 (fS ) ×
2
+ (1 − P0 (fS )) ×
max
3
x1 ∈{0,1}3
n
w(0, x1,1 , χ1 ((0, 0, 1))) + w(0, x1,2 , χ2 ((0, 0, 1)))
!
o
+ w(1, x1,3 , χ3 ((0, 0, 1)))
n
w(0, x1,1 , χ1 ((0, 0, 1))) + w(0, x1,2 , χ2 ((0, 0, 1)))
x31 ∈{0,1}3 \X1001
o
,
(126)
+ w(1, x1,3 , χ3 ((0, 0, 1)))
max
X1001 = arg 3 max {w(0, x1,1 , χ1 ((0, 0, 1))) + w(0, x1,2 , χ2 ((0, 0, 1))) + w(1, x1,3 , χ3 ((0, 0, 1)))},
x1 ∈{0,1}3
(127)
W010 (fS , χS ) = p (1 − p) P0 (fS ) ×
2
max
3
x1 ∈{0,1}3
n
w(0, x1,1 , χ1 (0, 1, 0)) + w(1, x1,2 , χ2 (0, 1, 0))
!
o
+ w(0, x1,3 , χ3 (0, 1, 0))
n
w(0, x1,1 , χ1 (0, 1, 0)) + w(1, x1,2 , χ2 (0, 1, 0))+
x31 ∈{0,1}3 \X1010
o
w(0, x1,3 , χ3 (0, 1, 0))
,
(128)
+ (1 − P0 (fS )) ×
max
X1010 = arg 3 max {w(0, x1,1 , χ1 (0, 1, 0)) + w(1, x1,2 , χ2 (0, 1, 0)) + w(0, x1,3 , χ3 (0, 1, 0))},
x1 ∈{0,1}3
W100 (fS , χS ) = p (1 − p) P0 (fS ) ×
2
max
3
x1 ∈{0,1}3
(129)
n
w(1, x1,1 , χ1 ((1, 0, 0))) + w(0, x1,2 , χ2 ((1, 0, 0)))
!
+ w(0, x1,3 , χ3 ((1, 0, 0)))}
+ (1 − P0 (fS )) ×
{w(1, x1,1 , χ1 ((1, 0, 0))) + w(0, x1,2 , χ2 ((1, 0, 0)))
+ w(0, x1,3 , χ3 ((1, 0, 0)))} ,
(130)
max
x31 ∈{0,1}3 \X1100
X1100 = arg 3 max {w(1, x1,1 , χ1 ((1, 0, 0))) + w(0, x1,2 , χ2 ((1, 0, 0))) + w(0, x1,3 , χ3 ((1, 0, 0)))},
x1 ∈{0,1}3
(131)
3
W111 (fS , χS ) = (1 − p) P0 (fS ) ×
+ (1 − P0 (fS )) ×
August 15, 2017
max
x31 ∈{0,1}3
max
3
nX
x31 ∈{0,1}3 \X1111
w(1, x1,n , χn ((1, 1, 1)))
n=1
3
nX
n=1
!
!
o
w(1, x1,n , χn ((1, 1, 1)))
,
(132)
DRAFT
42
X1111
W011 (fS χS ) = p(1 − p)
2
"
= arg 3 max {
x1 ∈{0,1}3
P0 (fS ) ×
3
X
w(1, x1,n , χn ((1, 1, 1)))},
(133)
n=1
max
3
x1 ∈{0,1}3
n
w(0, x1,1 , χ1 ((0, 1, 1))) + w(1, x1,2 , χ2 ((0, 1, 1)))
o
+ w(1, x1,3 , χ3 ((0, 1, 1)))
n
+ (1 − P0 ((fS ))) ×
max
w(0, x1,1 , χ1 ((0, 1, 1))) + w(1, x1,2 , χ2 ((0, 1, 1)))+
x31 ∈{0,1}3 \X1011
#
o
,
(134)
w(1, x1,3 , χ3 ((0, 1, 1)))
X1011 = arg 3 max {w(0, x1,1 , χ1 ((0, 1, 1))) + w(1, x1,2 , χ2 ((0, 1, 1))) + w(1, x1,3 , χ3 ((0, 1, 1)))},
x1 ∈{0,1}3
W101 (fS , χS ) = p(1 − p)
(135)
2
"
P0 (fS ) ×
max
3
x1 ∈{0,1}3
n
w(1, x1,1 , χ1 ((1, 0, 1))) + w(0, x1,2 , χ2 ((1, 0, 1)))
o
+ w(1, x1,3 , χ3 ((1, 0, 1)))
n
w(1, x1,1 , χ1 ((1, 0, 1))) + w(0, x1,2 , χ2 ((1, 0, 1)))+
+ (1 − P0 (fS )) ×
max
x31 ∈{0,1}3 \X1101
#
o
w(1, x1,3 , χ3 ((1, 0, 1)))
,
(136)
X1101 = arg 3 max
x1 ∈{0,1}3
n
o
w(1, x1,1 , χ1 ((1, 0, 1))) + w(0, x1,2 , χ2 ((1, 0, 1))) + w(1, x1,3 , χ3 ((1, 0, 1))) ,
W110 (fS , χS ) = p(1 − p)
(137)
2
"
P0 (fS ) ×
+ (1 − P0 (fS )) ×
max
3
x1 ∈{0,1}3
n
w(1, x1,1 , χ1 ((1, 1, 0))) + w(1, x1,2 , χ2 ((1, 1, 0)))
o
+ w(0, x1,3 , χ3 ((1, 1, 0)))
n
w(1, x1,1 , χ1 ((1, 1, 0))) + w(1, x1,2 , χ2 ((1, 1, 0)))
x31 ∈{0,1}3 \X1110
#
o
+ w(0, x1,3 , χ3 ((1, 1, 0)))
,
(138)
max
X1110 = arg 3 max {w(1, x1,1 , χ1 ((1, 1, 0))) + w(1, x1,2 , χ2 ((1, 1, 0))) + w(0, x1,3 , χ3 ((1, 1, 0)))}.
x1 ∈{0,1}3
August 15, 2017
(139)
DRAFT
43
In the case of Table VI, P0 (fS ) is given by
P0 (fS ) = P x30 = (0, 0, 0) + P x30 = (0, 0, 1) + P x30 = (0, 1, 0) + P x30 = (0, 1, 1)
+ P x30 = (1, 0, 0) + P x30 = (1, 0, 1)
(140)
= p(2 − p)
(141)
3
.
4
(142)
p= 21
=
R EFERENCES
[1] B. Larrousse and S. Lasaulce, “Coded power control: Performance analysis,” in Proc. of IEEE International Symposium
on Information Theory, Istanbul, Turkey, July 2013.
[2] S. Lasaulce and B. Larrousse, “More about optimal use of communication resources,” in Proc. of International Workshop
on Stochastic Methods in Game Theory, International School of Mathematics G. Stampacchia, Erice, Italy, September
2013.
[3] B. Larrousse, A. Agrawal, and S. Lasaulce, “Implicit coordination in two-agent team problems; application to distributed
power allocation,” in Proc. of 12th IEEE International Symposium on Modeling and Optimization in Mobile, Ad Hoc, and
Wireless Networks, 2014, pp. 579–584.
[4] T. Başar and S. Yüksel, Stochastic Networked Control Systems, Birhäuser, Ed.
Springer, 2013, vol. XVIII.
[5] P. Cuff, H. H. Permuter, and T. M. Cover, “Coordination capacity.” IEEE Transactions on Information Theory, vol. 56,
no. 9, pp. 4181–4206, 2010.
[6] P. Cuff, “Distributed channel synthesis,” IEEE Transactions on Information Theory, vol. 59, no. 11, pp. 7071–7096, 2013.
[7] G. Kramer and S. A. Savari, “Communicating probability distributions,” IEEE Transactions on Information Theory, vol. 53,
no. 2, pp. 518–525, February 2007.
[8] T. Han and S. Verdú, “Approximation theory of output statistics,” IEEE Transactions on Information Theory, vol. 39, no. 3,
pp. 752–772, May 1993.
[9] A. Bereyhi, M. Bahrami, M. Mirmohseni, and M. Aref, “Empirical coordination in a triangular multiterminal network,” in
Proc. of IEEE International Symposium on Information Theory, Istanbul, Turkey, July 2013, pp. 2149–2153.
[10] F. Haddadpour, M. H. Yassaee, A. Gohari, and M. R. Aref, “Coordination via a relay,” in Proc of IEEE International
Symposium on Information Theory, Boston, MA, July 2012, pp. 3048–3052.
[11] M. R. Bloch and J. Kliewer, “Strong coordination over a line network,” in Proc. IEEE International Symposium on
Information Theory, Istanbul, Turkey, July 2013, pp. 2319–2323.
[12] ——, “Strong coordination over a three-terminal relay network,” in Proc. of IEEE Information Theory Workshop, Hobart,
Tasmania, November 2014, pp. 646–650.
[13] R. Blasco-Serrano, R. Thobaben, and M. Skoglund, “Polar codes for coordination in cascade networks,” in Proc. of
International Zurich Seminar on Communications, Zurich, Switzerland, March 2012, pp. 55–58.
[14] M. R. Bloch, L. Luzzi, and J. Kliewer, “Strong coordination with polar codes,” in Proc. of 50th Allerton Conference on
Communication, Control, and Computing, Monticello, IL, 2012, pp. 565–571.
[15] R. A. Chou, M. R. Bloch, and J. Kliewer, “Polar coding for empirical and strong coordination via distribution
approximation,” in Proc. of IEEE International Symposium on Information Theory, Hong Kong, June 2015, pp. 1512–
1516.
August 15, 2017
DRAFT
44
[16] O. Gossner, P. Hernandez, and A. Neyman, “Optimal use of communication resources,” Econometrica, vol. 74, no. 6, pp.
1603–1636, 2006.
[17] P. Cuff and L. Zhao, “Coordination using implicit communication,” in Proc. of IEEE Information Theory Workshop, Paraty,
Brazil, 2011, pp. 467–471.
[18] E. van der Meulen, “A survey of multi-way channels in information theory: 1961-1976,” IEEE Transactions on Information
Theory, vol. 23, no. 1, pp. 1–37, Jan 1977.
[19] F. Willems and E. van der Meulen, “The discrete memoryless multiple-access channel with cribbing encoders,” IEEE
Transactions on Information Theory, vol. 31, no. 3, pp. 313–327, May 1985.
[20] H. Asnani and H. Permuter, “Multiple-access channel with partial and controlled cribbing encoders,” IEEE Transactions
on Information Theory, vol. 59, no. 4, pp. 2252–2266, April 2013.
[21] T. Weissman, “Capacity of channels with action-dependent states,” IEEE Transactions on Information Theory, vol. 56,
no. 11, pp. 5396–5411, Nov 2010.
[22] W. Yu, G. Ginis, and J. M. Cioffi, “Distributed multiuser power control for digital subscriber lines,” IEEE Journal on
Selected Areas in Communications, vol. 20, no. 5, pp. 1105–1115, Jun. 2002.
[23] S. Lasaulce and H. Tembine, Game Theory and Learning for Wireless Networks : Fundamentals and Applications.
Academic Press, 2011.
[24] A. Zappone, S. Buzzi, and E. Jorswieck, “Energy-efficient power control and receiver design in relay-assisted DS/CDMA
wireless networks via game theory,” IEEE Communications Letters, vol. 15, no. 7, pp. 701–703, July 2011.
[25] G. Bacci, L. Sanguinetti, M. Luise, and H. Poor, “A game-theoretic approach for energy-efficient contention-based
synchronization in OFDMA systems,” IEEE Transactions on Signal Processing, vol. 61, no. 5, pp. 1258–1271, 2013.
[26] G. Scutari, D. Palomar, and S. Barbarossa, “The MIMO iterative waterfilling algorithm,” IEEE Transactions on Signal
Processing, vol. 57, no. 5, pp. 1917–1935, May 2009.
[27] P. Mertikopoulos, E. V. Belmega, A. L. Moustakas, and S. Lasaulce, “Distributed learning policies for power allocation in
multiple access channels,” IEEE Journal on Selected Areas in Communications, vol. 30, no. 1, pp. 96–106, January 2012.
[28] S. Gel’Fand and M. S. Pinsker, “Coding for channel with random parameters,” Probl. Contr. Inform. Theory, vol. 9, no. 1,
pp. 19–31, 1980.
[29] Y.-H. Kim, A. Sutivong, and T. Cover, “State amplification,” IEEE Transactions on Information Theory, vol. 54, no. 5,
pp. 1850–1859, 2008.
[30] C. Choudhuri, Y.-H. Kim, and U. Mitra, “Capacity-distortion trade-off in channels with state,” in Proc. of 48th Annual
Allerton Conference on Communication, Control, and Computing, 2010, pp. 1311–1318.
[31] ——, “Causal state amplification,” in Proc. of IEEE International Symposium on Information Theory Proceedings, Saint
Petersburg, Russia, August 2011, pp. 2110–2114.
[32] C. Choudhuri and U. Mitra, “Action dependent strictly causal state communication,” in Proc. of IEEE International
Symposium on Information Theory, 2012, pp. 3058–3062.
[33] P. Cuff and C. Schieler, “Hybrid codes needed for coordination over the point-to-point channel,” in Proc. of 49th Annual
Allerton Conference on Communication, Control, and Computing, Monticello, IL, 2011, pp. 235–239.
[34] B. Larrousse, S. Lasaulce, and M. Wigger, “Coordinating partially-informed agents over state-dependent networks,” in
Proc. of Information Theory Workshop, Jerusalem, Israel, April 2015, pp. 1–5.
[35] Q. Li, D. Gesbert, and N. Gresset, “Joint precoding over a master-slave coordination link,” in Proc. of the IEEE International
Conference on Acoustics Speech and Signal Processing (ICASSP), Florence, Italy, May 2014.
[36] T. M. Cover and J. A. Thomas, Elements of Information Theory.
August 15, 2017
Wiley-Interscience, 2006.
DRAFT
45
[37] D. Goodman and N. Mandayam, “Power control for wireless data,” IEEE Personal Communications, vol. 7, no. 2, pp.
45–54, April 2000.
[38] B. Fourestié, “Communication switching method, access point, network controller and associated computer programs,”
France Telecom, Patent WO 2008/081137, Dec. 2007.
[39] M. Olama, S. Djouadi, and C. Charalambous, “Stochastic power control for time-varying long-term fading wireless
networks,” EURASIP Journal on Applied Signal Processing, 2006.
[40] M. Malmirchgini and Y. Mostofi, “On the spatial predictability of commucation channels,” IEEE Transactions on Vehicular
Technology, 2012.
[41] A. El Gamal and Y. Kim, Network Information Theory.
Cambridge University Press, 2011.
[42] N. Merhav and S. Shamai, “On joint source-channel coding for the Wyner-Ziv source and the Gel’fand-Pinsker channel,”
IEEE Transactions on Information Theory, vol. 49, no. 11, pp. 2844–2855, 2003.
[43] S. Haykin, “Cognitive radio: brain-empowered wireless communications,” IEEE Journal on Selected Areas in Communications, vol. 23, no. 2, pp. 201–220, 2005.
[44] M. L. Treust and S. Lasaulce, “A repeated game formulation of energy-efficient decentralized power control,” IEEE
Transactions on Wireless Communications, vol. 9, no. 9, pp. 2860–2869, September 2010.
[45] T. M. Cover and M. Chiang, “Duality between channel capacity and rate distortion with two-sided state information,” IEEE
Transactions on Information Theory, vol. 48, no. 6, pp. 1629–1638, 2002.
[46] C. E. Shannon, “A mathematical theory of communication,” Bell System Technical Journal, vol. 27, pp. 379–423, 1948.
[47] S. P. Boyd and L. Vandenberghe, Convex optimization.
Cambridge University Press, 2004.
[48] A. Gjendemsj, D. Gesbert, G. E. Oien, and S. G. Kiani, “Binary power control for sum rate maximization over multiple
interfering links,” IEEE Transactions on Wireless Communications, vol. 7, no. 8, pp. 3164–3173, 2008.
[49] F. Meshkati, H. Poor, and S. Schwartz, “Energy-efficient resource allocation in wireless networks: An overview of game
theoretic approaches,” IEEE Signal Processing Magazine, vol. 58, pp. 58–68, May 2007.
[50] E. V. Belmega and S. Lasaulce, “Energy-efficient precoding for multiple-antenna terminals,” IEEE Transactions on Signal
Processing, vol. 59, no. 1, pp. 329–340, January 2011.
[51] V. S. Varma, S. Lasaulce, Y. Hayel, and S. E. Elayoubi, “A cross-layer approach for distributed energy-efficient power
control in interference networks,” IEEE Transactions on Vehicular Technology, vol. 75, no. 7, pp. 3218 – 3232, 2015.
August 15, 2017
DRAFT
46
TABLE I
S UMMARY OF NOTATION USED THROUGHOUT THE PAPER .
Symbol
Meaning
Z
A generic random variable
Zij
Sequence of random variables (Zi , . . . , Zj ), j ≥ i
Z n or Z
Z
|Z|
∆(Z)
z
z n or z
EP
H(Z)
Zij when i = 1 and j = n
Alphabet of Z
Cardinality of Z
Unit simplex over Z
Realization of Z
Sequence or vector (z1 , . . . , zn )
Expectation operator under the probability P
Entropy of Z
I(Y ; Z)
Mutual information between Y and Z
Z1 − Z2 − Z3
Markov chain P(z1 |z2 , z3 ) = P(z1 |z2 )
1{.}
Indicator function
⊕
Modulo−2 addition
R+
[0, +∞)
Tn (Q)
{z n ∈ Z n : kTz n − Qk1 < }
δ()
A function of such that lim→0 δ() = 0
Tz n
Type of the sequence z n
TABLE II
E NCODING AND DECODING USED IN THE PROOF OF T HEOREM 5
August 15, 2017
Block
1
2
b
···
B
Message
···
i1
i2
State
···
iB
(2)
x0
···
ib
(1)
x0
Agent 2 action
x∗2
Agent 1 codeword
u1 (i1 , j1 )
u(i2 , j2 )
Agent 1 action
(1)
x1
x2 (bi1 )
(2)
x1
Agent 2 decoding
bi1
bi2
···
···
···
···
···
(b)
x0
x2 (bib−1 )
u(ib , jb )
(b)
x1
bib
···
···
···
···
···
(B)
x0
x2 (biB−1 )
u(iB jB )
(B)
x1
biB
DRAFT
47
TABLE III
C HOSEN INDEXATION FOR THE PAYOFF VECTOR w
BLOCKS OF SIZE
n1 n2 ny nu
AND PROBABILITY DISTRIBUTION VECTOR
q. B OLD LINES DELINEATE
AND EACH BLOCK CORRESPONDS TO A GIVEN VALUE OF THE RANDOM STATE
X0 . T HE
5−UPLETS ARE SORTED ACCORDING TO A LEXICOGRAPHIC ORDER .
Index of qi
X0
X1
X2
Y
U
1
1
1
1
1
1
2
..
.
1
..
.
1
..
.
1
..
.
1
..
.
2
nu
1
1
1
1
nu
nu + 1
..
.
1
..
.
1
..
.
1
..
.
2
..
.
1
..
.
2nu
..
.
1
..
.
1
..
.
1
..
.
2
..
.
nu
..
.
n1 n2 ny nu − nu + 1
..
.
1
..
.
n1
..
.
n2
..
.
ny
..
.
1
..
.
n1 n2 ny nu
1
n1
n2
ny
nu
..
.
..
.
..
.
..
.
..
.
..
.
..
.
..
.
..
.
..
.
..
.
..
.
(n0 − 1)n1 n2 ny nu + 1
..
.
n0
..
.
1
..
.
1
..
.
1
..
.
1
..
.
n0 n1 n2 ny nu
n0
n1
n2
ny
nu
TABLE IV
O PTIMAL MARGINAL AND JOINT DISTRIBUTIONS ( EXPRESSED IN %) FOR THE SUM - RATE PAYOFF FUNCTION OF THE CCPC
10 20
POLICY, WITH SNR = 10 D B AND WITH FOUR POSSIBLE TRANSMIT POWER LEVELS 0, 3 , 3 , 10 .
?
?
?
(QX1 (x1 ), QX2 (x2 ), QX1 X2 (x1 , x2 ))
x1 = 0
x1 =
10
3
x1 =
20
3
x1 = 10
in %
August 15, 2017
x2 = 00
(47.5,47.5,00.0)
(00.0,47.5,00.0)
(00.0,47.5,00.0)
(52.5,47.5,47.5)
x2 =
10
3
(47.5,00.0,00.0)
(00.0,00.0,00.0)
(00.0,00.0,00.0)
(52.5,00.0,00.0)
x2 =
20
3
(47.5,00.0,00.0)
(00.0,00.0,00.0)
(00.0,00.0,00.0)
(52.5,00.0,00.0)
x2 = 10
(47.5,52.5,47.5)
(00.0,52.5,00.0)
(00.0,52.5,00.0)
(52.5,52.5,05.5)
DRAFT
48
TABLE V
O PTIMAL MARGINAL AND JOINT DISTRIBUTIONS ( EXPRESSED IN %) FOR THE SUM - RATE PAYOFF FUNCTION OF THE CPC
10 20
POLICY, WITH SNR = 10 D B AND WITH FOUR POSSIBLE TRANSMIT POWER LEVELS 0, 3 , 3 , 10 .
?
?
?
(QX1 (x1 ), QX2 (x2 ), QX1 X2 (x1 , x2 ))
x1 = 0
x1 =
10
3
x1 =
20
3
x1 = 10
in %
x2 = 00
(44.4,50.4,00.1)
(02.6,50.4,00.9)
(08.0,50.4,06.5)
(45.0,50.4,42.9)
x2 =
10
3
(44.4,00.0,00.0)
(02.6,00.0,00.0)
(08.0,00.0,00.0)
(45.0,00.0,00.0)
x2 =
20
3
(44.4,00.0,00.0)
(02.6,00.0,00.0)
(08.0,00.0,00.0)
(45.0,00.0,00.0)
x2 = 10
(44.4,49.6,44.3)
(02.6,49.6,01.7)
(08.0,49.6,01.5)
(45.0,49.6,02.1)
TABLE VI
P ROPOSED SOURCE CODING AND DECODING FOR p = 21 .
x30
Index i =
fS (x30 )
gS (i)
000
001
010
011
100
101
110
111
m0
m0
m0
m0
m0
m0
m1
m1
111
111
111
111
111
111
001
001
TABLE VII
P ROPOSED CHANNEL CODING FOR p = 12 .
x30 (b)
000
001
010
011
100
101
110
111
x32 (b)
111
111
111
111
111
111
001
001
ib+1
m0
m1
m0
m1
m0
m1
m0
m1
m0
m1
m0
m1
m0
m1
m0
m1
x31 (b)
000
001
001
000
010
000
011
001
100
000
101
001
110
111
111
110
August 15, 2017
DRAFT
| 7cs.IT
|
SPRINGER’S WEYL GROUP REPRESENTATION VIA LOCALIZATION
arXiv:1505.06404v3 [math.AG] 12 Feb 2017
JIM CARRELL AND KIUMARS KAVEH
Abstract. Let G denote a reductive algebraic group over C and x a nilpotent element
of its Lie algebra g. The Springer variety Bx is the closed subvariety of the flag variety
B of G parameterizing the Borel subalgebras of g containing x. It has the remarkable
property that the Weyl group W of G admits a representation on the cohomology of
Bx even though W rarely acts on Bx itself. Well-known constructions of this action
due to Springer et al use technical machinery from algebraic geometry. The purpose of
this note is to describe an elementary approach that gives this action when x is what
we call parabolic-surjective. The idea is to use localization to construct an action of W
on the equivariant cohomology algebra HS∗ (Bx ), where S is a certain algebraic subtorus
of G. This action descends to H ∗ (Bx ) via the forgetful map and gives the desired
representation. The parabolic-surjective case includes all nilpotents of type A and, more
generally, all nilpotents for which it is known that W acts on HS∗ (Bx ) for some torus S
(see [K-P] and [G-McP]). Our result is deduced from a general theorem describing when
a group action on the cohomology of the fixed point set of a torus action on a space lifts
to the full cohomology algebra of the space.
1. Introduction
Let G be a reductive linear algebraic group over C with Lie algebra g, and fix a maximal
torus T and a Borel subgroup B of G such that T ⊂ B. The flag variety B = G/B of
G will be viewed as the variety of all Borel subgroups of G or, equivalently, as the variety
of all Borel subalgebras of g. Let W = NG (T )/T be the Weyl group of the pair (G, T ),
and recall that W acts topologically on B. Thus the cohomology algebra H ∗ (B) admits
a representation as a graded W -module, which is well-known to be isomorphic with the
+
graded W -algebra C[t]/IW
, the coinvariant algebra of W . Here t is the Lie algebra of T
+
and IW is the ideal in C[t] generated by the nonconstant homogeneous W -invariants. Note:
throughout this paper, H ∗ (Y ) will denote the standard cohomology algebra of a space Y
with complex coefficients.
A celebrated theorem of T. A. Springer [Spr1, Spr2] says that if x is a nilpotent element
of g and Bx is the Springer variety associated to x, namely the closed subvariety of B
consisting of all Borel subalgebras of g containing x, then there is a graded C-algebra
representation of W on H ∗ (Bx ) so that the cohomology restriction map i∗x : H ∗ (B) →
H ∗ (Bx ) associated to the inclusion ix : Bx ֒→ B is W -equivariant (see [H-S] for the proof of
W -equivariance). As remarked often, the existence of this representation is quite surprising
because W itself does not usually act on Bx , exceptions being when x = 0 or x is regular
in g, and the usual definitions require a lot of technical machinery. Subsequent definitions
involve either replacing H ∗ (Bx ) with an isomorphic algebra on which W is known to act (cf.
[Kraft, D-P, JC1]) or replacing Bx with a space having both a W -action and isomorphic
cohomology algebra (cf. [Ross, Slo, Treu]).
Date: February 14, 2017
AMS Subject Classification: 14N45, 14NF43, 55N91.
1
Our plan is to give a simple new construction of Springer’s representation when x is a
nilpotent element of g which is what we call parabolic-surjective. We call a nilpotent x ∈ g
parabolic if it is regular in a Levi subalgebra l of g, and we say x is parabolic-surjective
if in addition the cohomology restriction map i∗x is surjective. The key idea is to exploit
the fact that in the parabolic-surjective case, there exists subtorus S of G acting on Bx so
that (Bx )S is stable under W so localization and the parabolic-surjective condition can be
used to obtain an C[s]-module action of W on the torus equivariant cohomology HS∗ (Bx )
descending to an action on H ∗ (Bx ) which turns out to coincide with Springer’s action.
Recently, the existence of this W -action on HS∗ (Bx ) was established by finding a geometric realization of Spec(HS∗ (Bn )): see [K-P] for the general parabolic-surjective setting and
[G-McP] for x of type A. We note that although [G-McP] only treats the type A case, their
argument is valid for all parabolic-surjective x ∈ g fo all g. The paper [A-H] establishes this
action by employing Tanisaki’s presentation of H ∗ (Bx ) in type A.
We now state the main result which will give Springer’s action for parabolic-surjective x
by first obtaining it for HS∗ (Bx ).
Theorem 1.1. Let Y be a projective variety with vanishing odd cohomology. Also suppose
that we have actions of an algebraic torus S and a finite group W on Y and these two actions
commute with each other. Let X be an S-stable subvariety of Y such that the cohomology
restriction map i∗ : H ∗ (Y ) → H ∗ (X) is surjective, where i : X ֒→ Y is the inclusion. Then
if W acts on H ∗ (X S ) so that the cohomology restriction map (iS )∗ : H ∗ (Y S ) → H ∗ (X S )
induced by the inclusion iS : X S ֒→ Y S is W-equivariant, then W also acts on HS∗ (X) by
graded W-algebra C[s]-module isomorphisms. Moreover, the natural map HS∗ (X) → H ∗ (X)
induces a representation of W on the graded algebra H ∗ (X) compatible with its action on
H ∗ (Y ) such that all the maps in the following commutative diagram are W-equivariant:
(1)
HS∗ (Y )
H ∗ (Y )
ι∗
i∗
/ H ∗ (X)
S
/ H ∗ (X).
In the above diagram, ι∗ is the map on equivariant cohomology induced by i : X ֒→ Y .
The proof, given in the next section, is an application of the localization theorem. The
reader can easily reformulate this result as a statement involving topological torus actions.
In the final section, we will verify the above assertions about parabolic-surjective Springer
varieties.
Acknowledgement We would like to thank Shrawan Kumar for useful comments and
Megumi Harada for pointing out the paper [A-H].
2. proof of the main theorem
We will begin by reviewing some facts about equivariant cohomology. Excellent references
for the facts below are [A-B, Brion]. Recall that all cohomology is over C. Let Y be a complex
projective variety with vanishing odd cohomology admitting a nontrivial action (S, Y ) by
an algebraic torus S ∼
= (C∗ )ℓ . In particular, the fixed point set Y S is nontrivial. Recall that
the S-equivariant cohomology algebra HS∗ (Y ) of Y is defined as the cohomology algebra
H ∗ (YS ) of the Borel space YS = (Y × E)/S, where E is a contractible space with a free
S-action and S acts diagonally on the product. The projection Y × E → E induces a map
YS = (Y × E)/S → E/S which in turn induces an H ∗ (E/S)-module structure on HS∗ (Y ).
2
On the other hand, the inclusion νY : Y ֒→ YS along a fibre gives a map from HS∗ (Y ) to the
ordinary cohomology H ∗ (Y ). Moreover, there is a natural identification H ∗ (E/S) ∼
= C[s],
where s = Lie(S). Note that by the Kunneth formula, HS∗ (Y S ) = C[s] ⊗ H ∗ (Y S ). When
HS∗ (Y ) is a free C[s]-module, the action (S, Y ) is said to be equivariantly formal. It is wellknown that equivariant formality is implied by the vanishing of odd cohomology of Y . The
proof of our main result is based on the following well-known result, the first assertion of
which is a special case of the localization theorem.
Theorem 2.1. If (S, Y ) is equivariantly formal, then the inclusion mapping jY : Y S ֒→ Y
induces an injection (jY )∗S : HS∗ (Y ) → HS∗ (Y S ). Moreover, the map (νY )∗ fits into an exact
sequence
(2)
(νY )∗
0 → C[s]+ HS∗ (Y ) → HS∗ (Y ) −−−−→ H ∗ (Y ) → 0,
where C[s]+ is the augmentation ideal, generated by all the nonconstant homogeneous polynomials.
We now prove Theorem 1.1. Assume that W is a finite group acting on Y such that the
action (W, Y ) commutes with (S, Y ). Then W acts linearly on both H ∗ (Y ) and H ∗ (Y S ).
Furthermore, it acts on HS∗ (Y ) and HS∗ (Y S ) as C[s]-module isomorphisms so that the map
jY∗ : HS∗ (Y ) → HS∗ (Y S ) induced by the inclusion jY : Y S ֒→ Y is a W-equivariant C[s]module injection. Let X be an S-stable subvariety of Y such that the cohomology restriction
map i∗ : H ∗ (Y ) → H ∗ (X) is surjective.
Since W acts on H ∗ (X S ) and the restriction map H ∗ (Y S ) → H ∗ (X S ) is W-equivariant,
W also acts on HS∗ (X S ) as a group of C[s]-module isomorphisms so that the natural map
µ : HS∗ (Y S ) → HS∗ (X S ) is a W-equivariant C[s]-module homomorphism. Now consider the
commutative diagram
(3)
HS∗ (Y )
ι∗
∗
jY
HS∗ (Y S )
/ H ∗ (X)
S
∗
jX
µ
/ H ∗ (X S )
S
We will define the action of W on HS∗ (X) by imposing the requirement that ι∗ be a W-module
homomorphism. To show this action is well-defined, it suffices to show that the kernel of
ι∗ is a W-submodule. Suppose then that ι∗ (a) = 0. By assumption, H ∗ (Y ) → H ∗ (X)
is surjective and thus X also has vanishing odd cohomology. It follows that (S, X) is
∗
equivariantly formal and hence jX
is injective. Thus to show ι∗ (w · a) = 0 for any w ∈ W,
∗ ∗
it suffices to show that jX
ι (w · a) = 0. But
∗ ∗
∗ ∗
jX
ι (w · a) = µjY∗ (w · a) = w · µjY∗ (a) = wjX
ι (a) = 0,
since µjY∗ is a W-module homomorphism. Thus, W acts on HS∗ (X) as claimed. It follows
∗
from this argument that jX
is W-equivariant. To show that W acts on H ∗ (X), consider the
exact sequence (2) for X. As above, we may define the W-action by requiring that (νX )∗ be
equivariant. It suffices to show its kernel is W-stable. But if (νX )∗ (a) = 0, then a = f b for
some f ∈ C[s]+ and b ∈ HS∗ (X). Thus,
w · a = w · f b = f (w · b) ∈ C[s]+ HS∗ (X) = ker(νX )∗ .
Finally, we remark that the above definitions make the diagram (1) commutative.
3
If one omits the assumption that cohomology restriction map i∗ : H ∗ (Y ) → H ∗ (X) is
surjective, the best one can hope for is that W acts on the image i∗ (H ∗ (Y )). The following
result gives a sufficient condition for W to act in this case.
Theorem 2.2. Assume that the setup in Theorem 1.1 holds except for the assumption
that i∗ : H ∗ (Y ) → H ∗ (X) is surjective, and, in addition, assume also that H ∗ (X) has
vanishing odd cohomology. Then there exists an action of W on ι∗ (HS∗ (Y )) by C[s]-module
isomorphisms. Moreover, if ι∗ (HS∗ (Y )) is free of rank dim i∗ (H ∗ (Y )), then the action of W
on ι∗ (HS∗ (Y )) descends to i∗ (H ∗ (Y )) so that the cohomology restriction map i∗ : H ∗ (Y ) →
i∗ (H ∗ (Y )) is W-equivariant.
Proof. For the first assertion, we have to show that the kernel of ι∗ is W-invariant. Since
∗
H ∗ (X) has vanishing odd cohomology, jX
is injective, so this follows from the argument
above. Next, note that if N denotes a free C[s]-module of finite rank, then the C-vector
space dimension of N /C[s]+ N is equal to the rank of N . Thus it follows by assumption
that the sequence
ν
0 → C[s]+ ι∗ (HS∗ (Y )) → ι∗ (HS∗ (Y )) −
→ i∗ (H ∗ (Y )) → 0
is exact, where ν is the restriction of (νX )∗ . Hence, as above, the kernel of ι∗ (HS∗ (Y )) →
i∗ (H ∗ (Y )) is W-stable, so W acts on i∗ (H ∗ (Y )). Moreover, the map H ∗ (Y ) → i∗ (H ∗ (Y ))
is W-equivariant.
Remark 2.3. In the case when S = C∗ , since (S, X) is assumed to be equivariantly formal,
the module ι∗ (HS∗ (Y )) is always free. This is because C[s] is a principal ideal domain and
HS∗ (X) is free.
3. The Weyl group action on HS∗ (Bx )
We now return to the parabolic-surjective setting. First, recall that W acts as a group of
homeomorphisms of B which commute with T . Let K be a maximal compact subgroup in
G such that H = K ∩ T is a maximal torus in K. Then the natural mapping K/H → B is
a homeomorphism; but W = NK (H)/H acts on K/H (from the left) by w · kH = k ẇ−1 H,
where ẇ ∈ NK (H) is a representative of w. Thus W acts on B as asserted. Since this action
∗
commutes with the action of H on K/H, the group W acts on both HT∗ (B) = HH
(K/H)
∗
∗
∗
and H (B) and the natural mapping HT (B) → H (B) is W -equivariant.
In order to apply Theorem 1.1, we need the following lemma from [JC1].
Lemma 3.1. Let x ∈ g be nilpotent, and suppose x is a regular element in the Lie algebra of
the Levi L = CG (S) for a subtorus S of T . Then S acts on Bx with exactly [W : WL ] fixed
points. Moreover, every component of B S contains exactly one point of (Bx )S , so W acts on
H ∗ ((Bx )S ) so that the cohomology restriction map H ∗ (B S ) → H ∗ ((Bx )S ) is W -equivariant
and surjective.
Proof. By assumption, Bx is the variety of Borel subalgebras of g containing x, so Bx is
stable under the action of S on g. Each irreducible component of B S is isomorphic to the
flag variety of L, so each component contains a unique fixed point of the one parameter
group exp(tx), t ∈ C, since x is regular in l. It follows that the cohomology restriction map
H ∗ (B S ) → H ∗ ((Bx )S ) is surjective. Moreover, since W permutes the components of B S ,
and each component contains a unique point of Bx , we can uniquely define an action of W on
H ∗ ((Bx )S ) = H 0 ((Bx )S ) by requiring that the cohomology restriction map be equivariant.
Finally, it is well-known that the number of components of B S is the index [W : WL ] of the
Weyl group of L in W .
4
Remark 3.2. In fact, W actually acts on (Bx )S itself (see [JC1] or [JC2, Lemma 6.3 and
p. 137]).
Let us now return to the problem considered in the introduction. As above, G is reductive
linear algebraic group over C and B is its flag variety. By the main result (Theorem 1.1),
we have
Corollary 3.3. Let x be a parabolic-surjective nilpotent in g, say x is regular in the Lie
algebra of the Levi subgroup CG (S). Then W acts on HS∗ (Bx ) and this action descends to
to H ∗ (Bx ) so that the diagram (1) is commutative for Y = B and X = Bx . Consequently,
this W -action is Springer’s representation.
Proof. The only thing to show is that this action of W on H ∗ (Bx ) coincides with Springer’s
representation. But this follows since i∗ : H ∗ (B) → H ∗ (Bx ) is W -equivariant by [H-S].
This seems to give the most elementary construction of Springer’s action in type A. It was
originally conjectured in [Kraft] that for any x ∈ sl(n, C), the action of W = Sn on H ∗ (Bx )
is equivalent to the action of W on the coordinate ring A(t∩Cy ) of the schematic intersection
of the diagonal matrices in sl(n, C) and the closure Cy in sl(n, C) of the conjugacy class of
of the nilpotent y dual to x. This was immediately verified in [D-P] where it was shown
that H ∗ (Bx ) ∼
= A(t ∩ Cy ) as graded W -algebras. This isomorphism was extended in [JC1]
to the case of parabolic-surjective nilpotents in an arbitrary g which satisfy some additional
conditions. Here the nilpotent y dual to x turns out to be a Richardson element in the
nilradical of the parabolic subalgebra of g associated to the Levi l in which x is a regular
nilpotent. We refer to [JC1] for more details.
Finally, let us mention that by a well-known of result DeConcini, Lusztig and Procesi
[D-L-P], Bx has vanishing odd cohomology for any nilpotent x ∈ g. Moreover, the JacobsonMorosov lemma guarantees that every Springer variety Bx has a torus action (S, Bx ). This
suggests that HS∗ (Bx ) should be studied in the general case. For example, when does W act
on any of H ∗ ((Bx )S ), HS∗ (Bx ) or even ι∗ (HS∗ (B))?
References
[A-H] Abe, H.; Horiguchi, T. The torus equivariant cohomology rings of Springer varieties. Topology Appl.
208 (2016), 143–159.
[A-B] Atiyah, M.; Bott, R. The moment map and equivariant cohomology. Topology 23 (1984), 1–28.
[Brion] Brion, M. Equivariant cohomology and equivariant intersection theory. Notes by Alvaro Rittatore.
NATO Adv. Sci. Inst. Ser. C Math. Phys. Sci., 514, Representation theories and algebraic geometry
(Montreal, PX, 1997), 1–37, Kluwer Acad. Publ., Dordrecht, 1998.
[JC1] Carrell, J. B. Orbits of the Weyl group and a theorem of DeConcini and Procesi. Compositio Math.
60 (1986), 45–52.
[JC2] Carrell, J. B. Torus actions and cohomology. The adjoint representation and the adjoint action,
83–158, Encyclopaedia Math. Sci., 131, Springer, Berlin, 2002.
[D-L-P] De Concini, C.; Lusztig, G.; Procesi, C. Homology of the zero-set of a nilpotent vector field on a
flag manifold. J. Amer. Math. Soc. 1 (1988), 15–34.
[D-P] De Concini, C.; Procesi, C. Symmetric functions, conjugacy classes, and the flag variety. Invent.
Math., 64 (1981), 203–219.
[G-McP] Goresky, M.; MacPherson, R. On the spectrum of the equivariant cohomology ring. Canad. J.
Math. 62 (2010), 262–283.
[H-S] Hotta, R.; Springer, T.A. A specialization theorem for certain Weyl group representations and an
application to the Green polynomials of unitary groups, Invent. Math. 41 (1977) 113–127.
[Kraft] Kraft, H. P. Conjugacy classes and Weyl group representations, Young tableaux and Schur functors
in algebra and geometry (Torun,1980), 191205, Astrisque, 8788, Soc. Math. France, Paris, 1981.
[K-P] Kumar, S.; Procesi, C. An algebro-geometric realization of equivariant cohomology of some Springer
fibers. Journal of Algebra 368 (2012), 70–74.
5
[Ross] Rossmann, W. Picard-Lefschetz theory for the coadjoint quotient of a semisimple Lie algebra. (English summary) Invent. Math. 121 (1995), 531–578.
[Slo] Slodowy, P. Four lectures on simple groups and singularities. Communications of the Math. Inst.,
Rijksuniversiteit Utrecht, v. 11, ( 1 ) 1980.
[Spa] Spaltenstein, N. The fixed point set of a unipotent transformation on the flag manifold. Indag. Math.
38 (1976), 452–456.
[Spr1] Springer, T. A. Trigonometric sums, Green functions of finite groups and representations of Weyl
groups. Invent. Math. 36 (1976), 173–207.
[Spr2] Springer, T. A. A construction of representations of Weyl groups. Invent. Math. 44 (1978), 279–293.
[Treu] Treumann, D. A topological approach to induction theorems in Springer theory. Represent. Theory
13 (2009), 8–18.
Department of Mathematics, University of British Columbia, Vancouver, B.C., Canada
E-mail address: [email protected]
Department of Mathematics, University of Pittsburgh, Pittsburgh, PA, USA.
E-mail address: [email protected]
6
| 4math.GR
|
1
A New Backpressure Algorithm for Joint Rate
Control and Routing with Vanishing Utility
Optimality Gaps and Finite Queue Lengths
Hao Yu and Michael J. Neely
Department of Electrical Engineering
University of Southern California
arXiv:1701.04519v1 [cs.NI] 17 Jan 2017
Abstract
The backpressure algorithm has been widely used as a distributed solution to the problem of joint rate control and routing
in multi-hop data networks. By controlling a parameter V in the algorithm, the backpressure algorithm can achieve an arbitrarily
small utility optimality gap. However, this in turn brings in a large queue length at each node and hence causes large network delay.
This phenomenon is known as the fundamental utility-delay tradeoff. The best known utility-delay tradeoff for general networks
is [O(1/V ), O(V )] and is attained by a backpressure algorithm based on a drift-plus-penalty technique. This may suggest that
to achieve an arbitrarily small utility optimality gap, the existing backpressure algorithms necessarily yield an arbitrarily large
queue length. However, this paper proposes a new backpressure algorithm that has a vanishing utility optimality gap, so utility
converges to exact optimality as the algorithm keeps running, while queue lengths are bounded throughout by a finite constant.
The technique uses backpressure and drift concepts with a new method for convex programming.
I. I NTRODUCTION
In multi-hop data networks, the problem of joint rate control and routing is to accept data into the network to maximize
certain utilities and to make routing decisions at each node such that all accepted data are delivered to intended destinations
without overflowing any queue in intermediate nodes. The original backpressure algorithm proposed in the seminal work [1]
by Tassiulas and Ephremides addresses this problem by assuming that incoming data are given and are inside the network
stability region and develops a routing strategy to deliver all incoming data without overflowing any queue. In the context of
[1], there is essentially no utility maximization consideration in the network. The backpressure algorithm is further extended
by a drift-plus-penalty technique to deal with data network with both utility maximization and queue stability considerations
[2], [3], [4]. Alternative extensions for both utility maximization and queue stabilization are developed in [5], [6], [7], [8].
The above extended backpressure algorithms have different dynamics and/or may yield different utility-delay tradeoff results.
However, all of them rely on “backpressure” quantities, which are the differential backlogs between neighboring nodes.
It has been observed in [9], [5], [7], [10] that the drift-plus-penalty and other alternative algorithms can be interpreted as
first order Lagrangian dual type methods for constrained optimization. In addition, these backpressure algorithms follow certain
fundamental utility-delay tradeoffs. For instance, the primal-dual type backpressure algorithm in [5] achieves an O(1/V ) utility
optimality gap with an O(V 2 ) queue length, where V is an algorithm parameter. By controlling parameter V , a small utility
optimality gap is available only at the cost of a large queue length. The drift-plus-penalty backpressure algorithm [4], which
has the best utility-delay tradeoff among all existing first order Lagrangian dual type methods for general networks, can only
achieve an O(1/V ) utility optimality gap with an O(V ) queue length. Under certain restrictive assumptions over the network,
a better [O(1/V ), O(log(V ))] tradeoff is achieved via an exponential Lyapunov function in [11], and an [O(1/V ), O(log2 (V ))]
tradeoff is achieved via a LIFO-backpressure algorithm in [12]. The existing utility-delay tradeoff results seem to suggest that
a large queueing delay is unavoidable if a small utility optimality gap is demanded.
Recently, there have been many attempts in obtaining new variations of backpressure algorithms by applying Newton’s
method to the Lagrangian dual function. In the recent work [10], the authors develop a Newton’s method for joint rate control
and routing. However, the utility-delay tradeoff in [10] is still [O(1/V ), O(V 2 )]; and the algorithm requires a centralized
projection step (although Newton directions can be approximated in a distributed manner). Work [13] considers a network
flow control problem where the path of each flow is given (and hence there is no routing part in the problem), and proposes
a decentralized Newton based algorithm for rate control. Work [14] considers network routing without an end-to-end utility
and only shows the stability of the proposed Newton based backpressure algorithm. All of the above Netwon’s method based
algorithms rely on distributed approximations for the inverse of Hessians, whose computations still require certain coordinations
for the local information updates and propagations and do not scale well with the network size. In contrast, the first order
Lagrangian dual type methods do not need global network topology information. Rather, each node only needs the queue
length information of its neighbors.
This paper proposes a new first order Lagrangian dual type backpressure algorithm that is as simple as the existing algorithms
in [4], [5], [7] but has a better utility-delay tradeoff. The new backpressue algorithm achieves a vanishing utility optimality gap
that decays like O(1/t), where t is the number of iterations. It also guarantees that the queue length at each node is always
2
bounded by a fixed constant of the same order as the optimal Lagrange multiplier of the network optimization problem. This
improves on the utility-delay tradeoffs of prior work. In particular, it improves the [O(1/V ), O(V 2 )] utility-delay tradeoff in
[5] and the [O(1/V ), O(V )] utility-delay tradeoff of the drift-plus-penalty algorithm in [4], both of which yield an unbounded
queue length to have a vanishing utility optimality gap. The new backpressure algorithm differs from existing first order
backpressure algorithms in the following aspects:
1) The “backpressure” quantities in this paper are with respect to newly introduced weights. These are different from queues
used in other backpressure algorithms, but can still be locally tracked and updated.
2) The rate control and routing decision rule involves a quadratic term that is similar to a term used in proximal algorithms
[15].
Note that the benefit of introducing a quadratic term in network optimization has been observed in [16]. Work [16] considers
a network utility maximization problem with given routing paths that is a special case of the problem treated in this paper.
The algorithm of [16] considers a fixed set of predetermined paths for each session and does not scale well when treating all
(typically exponentially many) possible paths of a general network. The algorithm proposed in [16] is not a backpressure type
and hence is fundamentally different from ours. For example, the algorithm in [16] needs to update the primal variables (source
session rates for each path) at least twice per iteration, while our algorithm only updates the primal variables (source session
rates and link session rates) once per iteration. The prior work [16] shows that the utility optimality gap is asymptotically zero
without analyzing the decay rate, while this paper shows the utility optimality gap decays like O(1/t).
II. S YSTEM M ODEL AND P ROBLEM F ORMULATION
Consider a slotted data network with normalized time slots t ∈ {0, 1, 2, . . .}. This network is represented by a graph
G = (N , L), where N is the set of nodes and L ⊆ N × N is the set of directed links. Let |N | = N and |L| = L. This network
is shared by F end-to-end sessions denoted by a set F. For each end-to-end session f ∈ F, the source node Src(f ) and
destination node Dst(f ) are given but the routes are not specified. Each session f has a continuous and concave utility function
Uf (xf ) that represents the “satisfaction” received by accepting xf amount of data for session f into the network at each
slot. Unlike [5], [10] where Uf (·) is assumed to be differentiable and strongly concave, this paper considers general concave
utility functions Uf (·), including those that are neither differentiable nor strongly concave. Formally, each utility function Uf
is defined over an interval dom(Uf ), called the domain of the function. It is assumed throughout that either dom(Uf ) = [0, ∞)
or dom(Uf ) = (0, ∞), the latter being important for proportionally fair utilities [17] Uf (x) = log(x) that have singularities
at x = 0 .
(f )
Denote the capacity of link l as Cl and assume it is a fixed and positive constant.1 Define µl as the amount of session
f ’s data routed at link l that is to be determined by our algorithm. Note that in general, the network may be configured such
that some session f is forbidden to use link l. For each link l, define Sl ⊆ F as the set of sessions that are allowed to use
link l. The case of unrestricted routing is treated by defining Sl = F for all links l.
(f )
(f )
Note that if l = (n, m) with n, m ∈ N , then µl and Cl can also be respectively written as µ(n,m) and C(n,m) . For each
node n ∈ N , denote the sets of its incoming links and outgoing links as I(n) and O(n), respectively. Note that xf , ∀f ∈ F
(f )
and µl , ∀l ∈ L, ∀f ∈ F are the decision variables of a joint rate control and routing algorithm. If the global network topology
information is available, the optimal joint rate control and routing can be formulated as the following multi-commodity network
flow problem:
X
max
Uf (xf )
(1)
(f )
xf ,µl
s.t.
f ∈F
xf 1{n=Src(f )} +
X
(f )
µl
l∈I(n)
X
(f )
µl
X
≤
(f )
µl , ∀f ∈ F, ∀n ∈ N \ {Dst(f )}
(2)
l∈O(n)
≤ Cl , ∀l ∈ L,
(3)
f ∈F
(f )
µl
≥ 0, ∀l ∈ L, ∀f ∈ Sl ,
(4)
(f )
µl
= 0, ∀l ∈ L, ∀f ∈ F \ Sl ,
(5)
xf ∈ dom(Uf ), ∀f ∈ F
(6)
where 1{·} is an indicator function; (2) represents the node flow conservation constraints relaxed by replacing the equality
with an inequality, meaning that the total rate of flow f into node n is less than or equal to the total rate of flow f out of the
node (since, in principle, we can always send fake data for departure links when the inequality is loose); and (3) represents
link capacity constraints. Note that for each flow f , there is no constraint (2) at its destination node Dst(f ) since all incoming
data are consumed by this node.
1 As
stated in [10], this is a suitable model for wireline networks and wireless networks with fixed transmission power and orthogonal channels.
3
The above formulation includes network utility maximization with fixed paths as special cases. In the case when each
session only has one single given path, e.g., the network utility maximization problem considered in [18], we could modify
the sets Sl used in constraints (4) and (5) to reflect this fact. For example, if link l1 is only used for sessions f1 and f2 , then
Sl1 = {f1 , f2 }. Similarly, the case [16] where each flow is restricted to using links from a set of predefined paths can be
treated by modifying the sets Sl accordingly. See Appendix A for more discussions.
The solution to problem (1)-(6) corresponds to the optimal joint rate control and routing. However, to solve this convex
program at a single computer, we need to know the global network topology and the solution is a centralized one, which is
not practical for large data networks. As observed in [9], [5], [7], [10], various versions of backpressure algorithms can be
interpreted as distributed solutions to problem (1)-(6) from first order Lagrangian dual type methods.
(f ),∗
Assumption 1: (Feasibility) Problem (1)-(6) has at least one optimal solution vector [x∗f ; µl
]f ∈F ,l∈L .
Assumption 2: (Existence of Lagrange multipliers) Assume the convex program (1)-(6) has Lagrange multipliers attaining
(f )
the strong duality. Specifically, define convex set C = {[xf ; µl ]f ∈F ,l∈L : (3)-(6) hold}. Assume there exists a Lagrange
(f ),∗
∗
multiplier vector λ = [λn ]f ∈F ,n∈N \{Dst(f )} ≥ 0 such that
q(λ∗ ) = sup{(1) : (2)-(6)}
P
P
P
P
P
(f )
(f )
(f )
xf 1{n=Src(f )} + l∈I(n) µl − l∈O(n) µl
where q(λ) = sup[x ;µ(f ) ]∈C
f ∈F Uf (xf ) −
f ∈F
n∈N \{Dst(f )} λn
f
l
is the Lagrangian dual function of problem (1)-(6) by treating (3)-(6) as a convex set constraint.
Assumptions 1 and 2 hold in most cases of interest. For example, Slater’s condition guarantees Assumption 2. Since the
constraints (2)-(6) are linear, Proposition 6.4.2 in [19] ensures that Lagrange multipliers exist whenever constraints (2)-(6) are
feasible and when the utility functions Uf are either defined over open sets (such as Uf (x) = log(x) with dom(Uf ) = (0, ∞))
ef : (−, ∞) → R such
or can be concavely extended to open sets, meaning that there is an > 0 and a concave function U
ef (x) = Uf (x) whenever x ≥ 0.2
that U
(f ),∗
Fact 1: (Replacing inequality with equality) If Assumption 1 holds, problem (1)-(6) has an optimal solution vector [x∗f ; µl
]f ∈F ,l∈L
such that all constraints (2) take equalities.
(f )
Proof: Note that each µl can appear on the left side in at most one constraint (2) and appear on the right side in at
(f ),∗
most one constraint (2). Let [x∗f ; µl
]f ∈F ,l∈L be an optimal solution vector such that at least one inequality constraint (2)
(f ),∗
is loose. Note that we can reduce the value of µl
on the right side of a loose (2) until either that constraint holds with
(f ),∗
equality, or until µl
reduces to 0. The objective function value does not change, and no constraints are violated. We can
repeat the process until all inequality constraints (2) are tight.
III. T HE N EW BACKPRESURE A LGORITHM
A. Discussion of Various Queueing Models
At each node, an independent queue backlog is maintained for each session. At each slot t, let xf [t] be the source session
(f )
(f )
rates; and let µl [t] be the link session rates. Some prior work enforces the constraint (2) via virtual queues Yn [t] of the
following form:
n
o
X (f )
X (f )
Yn(f ) [t + 1] = max Yn(f ) [t] + xf [t]1{n=Src(f )} +
µl [t] −
µl [t], 0 .
(7)
l∈I(n)
l∈O(n)
While this virtual equation is a meaningful approximation, it differs from reality in that new injected data are allowed to
be transmitted immediately, or equivalently, a single packet is allowed to enter and leave many nodes within the same slot.
(f )
Further, there is no clear connection between the virtual queues Yn [t] in (7) and the actual queues in the network. Indeed,
(f )
it is easy to construct examples that show there can be an arbitrarily large difference between the Yn [t] value in (7) and the
physical queue size in actual networks (see Appendix B).
(f )
An actual queueing network has queues Zn [t] with the following dynamics:
n
o
X (f )
X (f )
Zn(f ) [t + 1] ≤ max Zn(f ) [t] −
µl [t], 0 + xf [t]1{n=Src(f )} +
µl [t].
(8)
l∈O(n)
l∈I(n)
This is faithful to actual queue dynamics and does not allow data to be retransmitted over multiple hops in one slot. Note that
P
(f )
(8) is an inequality because the new arrivals from other nodes may be strictly less than l∈I(n) µl [t] because those other
nodes may not have enough backlog to send. The model (8) allows for any decisions to be made to fill the transmission values
P
(f )
(f )
(f )
µl [t] in the case that Zn [t] ≤ l∈O(n) µl [t], provided that (8) holds.
2 If dom(U ) = [0, ∞), such concave extension is possible if the right-derivative of U at x = 0 is finite (such as for U (x) = log(1 + x) or
f
f
f
√
Uf (x) = min[x, 3]). Such an extension is impossible for the example Uf (x) = x because the slope is infinite at x = 0. Nevertheless, Lagrange multipliers
often exist even for these utility functions, such as when Slater’s condition holds [19].
4
This paper develops an algorithm that converges to the optimal utility defined by problem (1)-(6), and that produces worstcase bounded queues on the actual queueing network, that is, with actual queues that evolve as given in (8). To begin, it is
convenient to introduce the following virtual queue equation
X (f )
X (f )
)
(f )
Q(f
µl [t] + xf [t]1{n=Src(f )} +
µl [t],
(9)
n [t + 1] =Qn [t] −
l∈O(n)
l∈I(n)
(f )
where Qn [t] represents a virtual queue value associated with session f at node n. At first glance, this model (9) appears to
(f )
be only an approximation, perhaps even a worse approximation than (7), because it allows the Qn [t] values to be negative.
(f )
Indeed, we use Qn [t] only as virtual queues to inform the algorithm and do not treat them as actual queues. However, this
paper shows that using these virtual queues to choose the µ[t] decisions ensures not only that the desired constraints (2) are
(f )
satisfied, but that the resulting µ[t] decisions create bounded queues Zn [t] in the actual network, where the actual queues
evolve according to (8). In short, our algorithm can be faithfully implemented with respect to actual queueing networks, and
converges to exact optimality on those networks.
(f )
The next lemma shows that if an algorithm can guarantee virtual queues Qn [t] defined in (9) are bounded, then actual
physical queues satisfying (8) are also bounded.
(f )
Lemma 1: Consider a network flow problem described by problem (1)-(6). For all l ∈ L and f ∈ F, let µl [t], xf [t]
(f )
(f )
(f )
be decisions yielded by a dynamic algorithm. Suppose Yn [t], Zn [t], Qn [t] evolve by (7)-(9) with initial conditions
(f )
(f )
(f )
(f )
Vn [0] = Zn [0] = Qn [0] = 0. If there exists a constant B > 0 such that |Qn [t]| ≤ B, ∀t, then
P
(f )
1) Zn [t] ≤ 2B + l∈O(n) Cl for all t ∈ {0, 1, 2, . . .}.
P
(f )
2) Yn [t] ≤ 2B + l∈O(n) Cl for all t ∈ {0, 1, 2, . . .}.
Proof:
P
)
b (f )
b (f
1) Fix f ∈ F, n ∈ N \ {Dst(f )}. Define an auxiliary virtual queue Q
n [t] that is initialized by Qn [0] = B +
l∈O(n) Cl
P
(f
)
(f
)
(f
)
b n [t] = Qn [t] + B +
and evolves according to (9). It follows that Q
l∈O(n) Cl , ∀t. Since Qn [t] ≥ −B, ∀t by
P
P
(f )
(f )
)
b
b (f
assumption, we have Qn [t] ≥
Cl ≥
µ [t], ∀t. This implies that Q
n [t] also satisfies:
l∈O(n)
l∈O(n)
l
n
o
X (f )
X (f )
)
(f )
b (f
b
Q
[t
+
1]
=
max
Q
[t]
−
µ
[t],
0
+ xf [t]1{n=Src(f )} +
µl [t], ∀t
n
n
l
l∈O(n)
(10)
l∈I(n)
)
(f )
b (f )
b (f
which is identical to (8) except the inequality is replaced by an equality. Since Zn [0] = 0 < Q
n [0]; and Qn [t]
)
(f )
b (f
satisfies (10), by inductions, Zn [t] ≤ Q
n [t], ∀t.
P
P
)
(f )
(f )
(f )
b (f
b
Since Qn [t] = Qn [t] + B + l∈O(n) Cl , ∀t and Qn [t] ≤ B, ∀t, we have Q
n [t] ≤ 2B +
l∈O(n) Cl , ∀t. It follows
P
(f )
that Zn [t] ≤ 2B + l∈O(n) Cl , ∀t.
2) The proof of part (2) is similar and is in Appendix C.
B. The New Backpressure Algorithm
In this subsection, we propose a new backpressure algorithm that yields source session rates xf [t] and link session rates
at each slot such that the physical queues for each session at each node are bounded by a constant and the time average
utility satisfies
t−1
X
1XX
Uf (xf [t]) ≥
Uf (x∗f ) − O(1/t), ∀t
t τ =0
(f )
µl [t]
f ∈F
f ∈F
where x∗f are from the optimal solution to (1)-(6). Note that Jensen’s inequality further implies that
X
f ∈F
Uf
t−1
X
1X
xf [τ ] ≥
Uf (x∗f ) − O(1/t), ∀t
t τ =0
f ∈F
The new backpressure algorithm is described in Algorithm 1. Similar to existing backpressure algorithms, the updates in
Algorithm 1 at each node n are fully distributed and only depend on weights at itself and its neighbor nodes. Unlike existing
(f )
(f )
backpressure algorithms, the weights used to update decision variables xf [t] and µl [t] are not the virtual queues Qn [t]
(f )
themselves, rather, they are augmented values Wn [t] equal to the sum of the virtual queues and the amount of net injected
data in the previous slot t − 1. In addition, the updates involve an additional quadratic term, which is similar to a term used
in proximal algorithms [15].
5
Algorithm 1 The New Backpressure Algorithm
(f )
(f )
Let αn > 0, ∀n ∈ N be constant parameters. Initialize xf [−1] = 0, µl [−1] = 0, ∀f ∈ F, ∀l ∈ L and Qn [0] = 0, ∀n ∈
N , ∀f ∈ F. At each time t ∈ {0, 1, 2, . . .}, each node n does the following:
(f )
• For each f ∈ F, if node n is not the destination node of session f , i.e., n 6= Dst(f ), then define weight Wn [t]:
X (f )
X (f )
)
Wn(f ) [t] =Q(f
µl [t − 1] −
µl [t − 1],
(11)
n [t] + xf [t − 1]1{n=Src(f )} +
l∈I(n)
l∈O(n)
(f )
•
If node n is the destination node, i.e., n = Dst(f ), then define Wn [t] = 0. Notify neighbor nodes (nodes k that can
(f )
send session f to node n, i.e., ∀k such that f ∈ S(k,n) ) about this new Wn [t] value.
For each f ∈ F, if node n is the source node of session f , i.e., n = Src(f ), choose xf [t] as the solution to
2
max Uf (xf ) − Wn(f ) [t]xf − αn xf − xf [t − 1]
(12)
xf
s.t.
•
xf ∈ dom(Uf )
(13)
(f )
For all (n, m) ∈ O(n), choose {µ(n,m) [t], ∀f ∈ F} as the solution to the following convex program:
X
(f )
X (f )
2
(f )
(f )
max
Wn(f ) [t] − Wm
[t] µ(n,m) − αn + αm
µ(n,m) − µ(n,m) [t − 1]
(14)
(f )
µ(n,m) f ∈F
s.t.
X
f ∈F
(f )
µ(n,m) ≤ C(n,m)
(15)
f ∈F
(f )
•
µ(n,m) ≥ 0, ∀f ∈ S(n,m)
(16)
(f )
µ(n,m)
(17)
= 0, ∀f 6∈ S(n,m)
(f )
For each f ∈ F, if node n is not the destination of f , i.e., n 6= Dst(f ), update virtual queue Qn [t + 1] by (9).
C. Almost Closed-Form Updates in Algorithm 1
(f )
This subsection shows the decisions xf [t] and µl [t] in Algorithm 1 have either closed-form solutions or “almost” closedform solutions at each iteration t.
Lemma 2: Let x̂f ≡ xf [t] denote the solution to (12)-(13).
(f )
1) Suppose dom(Uf ) = [0, ∞) and Uf (xf ) is differentiable. Let h(xf ) = Uf0 (xf ) − 2αn xf + 2αn xf [t − 1] − Wn [t]. If
h(0) < 0, then x̂f = 0; otherwise x̂f is the root to the equation h(xf ) = 0 and can be found by a bisection search.
2) Suppose dom(Uf ) = (0, ∞) and Uf (xf ) = wf log(xf ) for some weight wf > 0. Then:
q
(f )
(f )
(Wn [t] − 2αn xf [t − 1])2 + 8αn wf
2αn xf [t − 1] − Wn [t]
x̂f =
+
4αn
4αn
Proof: Omitted for brevity.
(f )
The problem (14)-(17) can be represented as follows by eliminating µ(n,m) , f 6∈ S(n,m) , completing the square and replacing
maximization with minimization. (Note that K = |S(n,m) | ≤ |F|.)
K
min
s.t.
1X
(zk − ak )2
2
(18)
zk ≤ b
(19)
k=1
K
X
k=1
zk ≥ 0, ∀k ∈ {1, 2, . . . , K}
(20)
Lemma 3: The solution to problem (18)-(20) is given by zk∗ = max{0, ak − θ∗ }, ∀k ∈ {1, 2, . . . , K} where θ∗ ≥ 0 can be
found either by a bisection search (See Appendix D) or by Algorithm 2 with complexity O(K log K).
Proof: A similar problem where (19) is replaced with an equality constraint in considered in [20]. The optimal solution
to this quadratic program is characterized by its KKT condition and a corresponding algorithm can be developed to obtain its
KKT point. A complete proof is presented in Appendix D.
Note that step (3) in Algorithm 2 has complexity O(K) and hence the overall complexity of Algorithm 2 is dominated by
the sorting step (2) with complexity O(K log(K)).
6
Algorithm 2 Algorithm to solve problem (18)-(20)
PK
1) Check if k=1 max{0, ak } ≤ b holds. If yes, let θ∗ = 0 and zk∗ = max{0, ak }, ∀k ∈ {1, 2, . . . , K} and terminate the
algorithm; else, continue to the next step.
2) Sort all ak , ∈ {1, 2, . . . , K} in a decreasing order π such that aπ(1) ≥ aπ(2) ≥ · · · ≥ aπ(K) . Define S0 = 0.
3) For k = 1 to K
S −b
∗
• Let Sk = Sk−1 + ak . Let θ = kk .
∗
∗
∗
• If θ ≥ 0, aπ(k) − θ > 0 and aπ(k+1) − θ ≤ 0, then terminate the loop; else, continue to the next iteration in the
loop.
4) Let zk∗ = max{0, ak − θ∗ }, ∀k ∈ {1, 2, . . . , K} and terminate the algorithm.
IV. P ERFORMANCE A NALYSIS OF A LGORITHM 1
A. Basic Facts from Convex Analysis
Definition 1 (Lipschitz Continuity): Let Z ⊆ Rn be a convex set. Function h : Z → Rm is said to be Lipschitz continuous
on Z with modulus β if there exists β > 0 such that kh(z1 ) − h(z2 )k ≤ βkz1 − z2 k for all z1 , z2 ∈ Z.
Definition 2 (Strongly Concave Functions): Let Z ⊆ Rn be a convex set. Function h is said to be strongly concave on Z
with modulus α if there exists a constant α > 0 such that h(z) + 21 αkzk2 is concave on Z.
By the definition of strongly concave functions, it is easy to show that if h(z) is concave and α > 0, then h(z) − αkz − z0 k2
is strongly concave with modulus 2α for any constant z0 .
Lemma 4: Let Z ⊆ Rn be a convex set. Let function h be strongly concave on Z with modulus α and zopt be a global
maximum of h on Z. Then, h(zopt ) ≥ h(z) + α2 kzopt − zk2 for all z ∈ Z.
B. Preliminaries
(f )
Define column vector y = [xf ; µl ]f ∈F ,l∈L . For each f ∈ F, n ∈ N \ {Dst(f )}, define column vector
(
(f )
[xf ; µl ]l∈I(n)∪O(n) if n = Src(f ),
(f )
yn =
(f )
[µl ]l∈I(n)∪O(n)
else,
(21)
(f )
which is composed by the control actions appearing in each constraint (2); and introduce a function with respect to yn as
X (f )
X (f )
gn(f ) (yn(f ) ) = xf 1{n=Src(f )} +
µl −
µl
(22)
l∈I(n)
l∈O(n)
Thus, constraint (2) can be rewritten as
gn(f ) (yn(f ) ) ≤ 0, ∀f ∈ F, ∀n ∈ N \ {Dst(f )}.
(f )
Note that each vector yn is a subvector of y and has length dn + 1 where dn is the degree of node n (the total number
of outgoing links and incoming links) if node n is the source of session f ; and has length dn if node n is not the source of
session f .
(f )
(f )
Fact 2: Each function gn (·) defined in (22) is Lipschitz continuous with respect to vector yn with modulus
p
βn ≤ dn + 1.
where dn is the degree of node n.
(f )
(f )
(f )
Proof: This fact can be easily shown by noting that each gn (yn ) is a linear function with respect to vector yn and
has at most dn + 1 non-zero coefficients that are equal to ±1.
Note that virtual queue update equation (9) can be rewritten as:
)
(f )
(f )
(f )
Q(f
n [t + 1] = Qn [t] + gn (yn [t]),
(23)
and weight update equation (11) can be rewritten as:
)
(f )
(f )
Wn(f ) [t] = Q(f
n [t] + gn (yn [t − 1]).
(24)
Define
L(t) =
1X
2
X
f ∈F n∈N \Dst(f )
2
)
Q(f
n [t]
(25)
7
and call it a Lyapunov function. In the remainder of this paper, double summations are often written compactly as a single
summation, e.g.,
X
X
X
∆
· =
· .
f ∈F n∈N \Dst(f )
f ∈F ,
n∈N \Dst(f )
Define the Lyapunov drift as
∆[t] = L(t + 1) − L(t).
The following lemma follows directly from equation (23).
Lemma 5: At each iteration t ∈ {0, 1, . . .} in Algorithm 1, the Lyapunov drift is given by
X
1 (f ) f 2
)
(f )
f
Q(f
g (yn [t])
∆[t] =
.
n [t]gn (yn [t]) +
2 n
(26)
f ∈F ,
n∈N \Dst(f )
Proof: Fix f ∈ F and n ∈ N \ Dst(f ), we have
2 1
2
1 (f )
)
Qn [t + 1] − Q(f
n [t]
2
2
2 1
2
(a) 1
(f )
(f )
(f )
)
= Qn [t] + gn (yn [t]) − Q(f
n [t]
2
2
1 (f ) f 2
(f )
(f )
f
=Qn [t]gn (yn [t]) + gn (yn [t])
2
(27)
where (a) follows from (23).
By the definition of ∆[t], we have
∆[t] =
(a)
=
1
2
X
)
Q(f
n [t + 1]
2
2
)
− Q(f
n [t]
f ∈F ,
n∈N \Dst(f )
X
f ∈F ,
n∈N \Dst(f )
)
(f )
f
Q(f
n [t]gn (yn [t]) +
1 (f ) f 2
g (yn [t])
2 n
where (a) follows from
P (27).
Define f (y) = f ∈F Uf (xf ). At each time t, consider choosing a decision vector y[t] that includes elements in each
(f )
subvector yn [t] to solve the following problem:
X
X
X (f )
(f )
max f (y) −
Wn(f ) [t]gn(f ) (yn(f ) ) + αn kyn(f ) − yn(f ) [t − 1]k2 −
αn
(µl − µl [t − 1])2
(28)
y
f ∈F ,
n∈N \Dst(f )
f ∈F ,
n=Dst(f )
l∈I(n)
s.t. (3)-(6)
(29)
The expression (28) is a modified drift-plus-penalty expression. Unlike the standard drift-plus-penalty expressions from [4],
(f )
(f )
(f )
(f )
the above expression uses weights Wn [t], which arguments each Qn [t] by gn (yn [t − 1]), rather than virtual queues
(f )
Qn [t]. It also includes a “prox”-like term that penalizes deviation from the previous y[t − 1] vector. This results in the novel
backpressure-type algorithm of Algorithm 1. Indeed, the decisions in Algorithm 1 were derived as the solution to the above
problem (28)-(29). This is formalized in the next lemma.
Lemma 6: At each iteration t ∈ {0, 1, . . .}, the action y[t] jointly chosen in Algorithm 1 is the solution to problem (28)-(29).
(f )
Proof: The proof involves collecting terms associated with the xf [t] and µl [t] decisions. See Appendix E for details.
Furthermore, the next lemma summarizes that the action y[t] jointly chosen in Algorithm 1 provides a lower bound for the
drift-plus-penalty expression at each iteration t ∈ {0, 1, . . .}.
(f ),∗
(f )
(f ),∗
Lemma 7: Let y∗ = [x∗f ; µl
]f ∈F ,l∈L be an optimal solution to problem (1)-(6) given in Fact 1, i.e., gn (yn ) = 0, ∀f ∈
(f )
F, ∀n ∈ N \Dst(f ). If αn ≥ 12 (dn +1), ∀n ∈ N , where dn is the degree of node n, then the action y[t] = [xf [t]; µl [t]]f ∈F ,l∈L
jointly chosen in Algorithm 1 at each iteration t ∈ {0, 1, . . .} satisfies
−∆[t] + f (y[t]) ≥ f (y∗ ) + Φ[t] − Φ[t − 1]
P
P
(f ),∗
(f )
(f ),∗
(f )
where Φ[t] = f ∈F ,n∈N αn 1{n6=Dst(f )} kyn
− yn [t]k2 + αn 1{n=Dst(f )} l∈I(n) (µl
− µl [t])2 .
Proof: See Appendix F.
It remains to show that this modified backpressure algorithm leads to fundamentally improved performance.
8
C. Utility Optimality Gap Analysis
(f )
(f )
Define column vector Q[t] = Qn [t] f ∈F ,n∈N \{Dst(f )} as the stacked vector of all virtual queues Qn [t] defined in (9).
(f )
(f )
Note that (25) can be rewritten as L(t) = 21 kQ[t]k2 . Define vectorized constraints (2) as g(y) = [gn (yn )]f ∈F ,n∈N \Dst(f ) .
(f ),∗
(f )
(f ),∗
Lemma 8: Let y∗ = [x∗f ; µl
]f ∈F ,l∈L be an optimal solution to problem (1)-(6) given in Fact 1, i.e., gn (yn ) =
1
0, ∀f ∈ F, ∀n ∈ N \ Dst(f ). If αn ≥ 2 (dn + 1), ∀n ∈ N in Algorithm 1, where dn is the degree of node n, then for all t ≥ 1,
t−1
X
1
f (y[τ ]) ≥ tf (y∗ ) − ζ + kQ[t]k2 .
2
τ =0
P
P
(f ),∗
(f ),∗ 2
where ζ = Φ[−1] = f ∈F ,n∈N αn 1{n6=Dst(f )} kyn k2 + αn 1{n=Dst(f )} l∈I(n) (µl
) is a constant.
Proof: By Lemma 7, we have −∆[τ ] + f (y[τ ]) ≥ f (y∗ ) + Φ[t] − Φ[t − 1], ∀τ ∈ {0, 1, . . . , t − 1}. Summing over
τ ∈ {0, 1, . . . , t − 1} yields
t−1
X
f (y[τ ]) −
τ =0
t−1
X
∆[τ ]
τ =0
≥tf (y∗ ) +
t−1
X
Φ[τ ] − Φ[τ − 1]
τ =0
=tf (y∗ ) + Φ[t] − Φ[−1]
(a)
≥ tf (y∗ ) − Φ[−1]
where (a) follows from the fact that Φ[t] ≥ 0, ∀t.
Recall ∆[τ ] = L[τ + 1] − L[τ ], simplifying summations and rearranging terms yields
t−1
X
f (y[τ ]) ≥tf (y∗ ) − Φ[−1] + L[t] − L[0]
τ =0
1
= tf (y∗ ) − Φ[−1] + kQ[t]k2
2
(a)
where (a) follows from the fact that L[0] = 0 and L[t] = 12 kQ[t]k2 .
The next theorem summarizes that Algorithm 1 yields a vanishing utility optimality gap that approaches zero like O(1/t).
(f ),∗
(f ),∗
(f )
Theorem 1: Let y∗ = [x∗f ; µl
]f ∈F ,l∈L be an optimal solution to problem (1)-(6) given in Fact 1, i.e., gn (yn ) =
1
0, ∀f ∈ F, ∀n ∈ N \ Dst(f ). If αn ≥ 2 (dn + 1), ∀n ∈ N in Algorithm 1, where dn is the degree of node n, then for all
t ≥ 1, we have
t−1
X
1XX
1
Uf (xf [τ ]) ≥
Uf (x∗f ) − ζ,
t τ =0
t
f ∈F
f ∈F
where ζ is a constant defined in Lemma 8. Moreover, if we define xf [t] =
X
Uf (xf [t]) ≥
f ∈F
Proof: Recall that f (y) =
P
f ∈F
X
f ∈F
1
t
Pt−1
τ =0
xf [τ ], ∀f ∈ F, then
1
Uf (x∗f ) − ζ.
t
Uf (xf ). By Lemma 8, we have
t−1 X
X
Uf (xf [τ ]) ≥t
τ =0 f ∈F
X
f ∈F
(a)
≥t
X
1
Uf (x∗f ) − ζ + kQ[t]k2
2
Uf (x∗f ) − ζ.
f ∈F
where (a) follows from the trivial fact that kQ[t]k2 ≥ 0.
Dividing both sides by a factor t yields the first inequality in this theorem. The second inequality follows from the concavity
of Uf (·) and Jensen’s inequality.
9
D. Queue Stability Analysis
Lemma 9: Let Q[t], t ∈ {0, 1, . . .} be the virtual queues in Algorithm 1. For any t ≥ 1,
Q[t] =
t−1
X
g(y[τ ])
τ =0
Proof: This lemma follows directly from the fact that Q[0] = 0 and queue update equation (9) can be written as
Q[t + 1] = Q[t] + g(y[t − 1]).
(f )
The next theorem shows the boundedness of all virtual queues Qn [t] in Algorithm 1.
(f ),∗
(f )
(f ),∗
∗
∗
Theorem 2: Let y = [xf ; µl
]f ∈F ,l∈L be an optimal solution to problem (1)-(6) given in Fact 1, i.e., gn (yn ) =
∗
0, ∀f ∈ F, ∀n ∈ N \ Dst(f ), and λ be a Lagrange multiplier vector given in Assumption 2. If αn ≥ 21 (dn + 1)2 , ∀n ∈ N in
Algorithm 1, where dn is the degree of node n, then for all t ≥ 1,
p
∗
)
|Q(f
2ζ, ∀f ∈ F, ∀n ∈ N \ {Dst(f )}.
n [t]| ≤ 2kλ k +
where ζ is a constant defined in Lemma
8.
Proof: Let q(λ) = supy∈C f (y) − λT g(y)
{0, 1, . . . , }, by Assumption 2, we have
be the Lagrangian dual function defined in Assumption 2. For all τ ∈
(a)
f (y∗ ) = q(λ∗ ) ≥ f (y[τ ]) − λ∗,T g(y[τ ])
where (a) follows from the definition of q(λ∗ ). Rearranging terms yields
f (y[τ ]) ≤ f (y∗ ) + λ∗,T g(y[τ ]), ∀τ ∈ {0, 1, . . .}.
Fix t > 0. Summing over τ ∈ {0, 1, . . . , t − 1} yields
t−1
X
f (y[τ ]) ≤tf (y∗ ) +
τ =0
t−1
X
λ∗,T g(y[τ ])
τ =0
=tf (y∗ ) + λ∗,T
t−1
X
g(y[τ ])
τ =0
(a)
= tf (y∗ ) + λ∗,T Q[t]
(b)
≤tf (y∗ ) + kλ∗ kkQ[t]k
where (a) follows form Lemma 9 and (b) follows from Cauchy-Schwarz inequality.
On the other hand, by Lemma 8, we have
t−1
X
1
f (y[τ ]) ≥ tf (y∗ ) − ζ + kQ[t]k2 .
2
τ =0
Combining the last two inequalities and cancelling the common terms yields
1
kQ[t]k2 − ζ ≤ kλ∗ kkQ[t]k
2
2
⇒ kQ[t]k − kλ∗ k ≤ kλ∗ k2 + 2ζ
q
⇒kQ[t]k ≤ kλ∗ k + kλ∗ k2 + 2ζ
p
(a)
⇒kQ[t]k ≤ 2kλ∗ k + 2ζ
√
√
√
where (a) follows from the basic inequality a + b ≤ a + b for any a, b ≥ 0.
Thus, for any f ∈ F and n ∈ N \ {Dst(f )}, we have
p
∗
)
|Q(f
2ζ.
n [t]| ≤ kQ[t]k ≤ 2kλ k +
√
(f )
This theorem shows that the absolute values of all virtual queues Qn [t] are bounded by a constant B = 2kλ∗ k + 2ζ from
(f )
(f )
above.P
By Lemma 1 and discussions in Section III-A, the actual physical queues Zn [t] evolving via (8) satisfy Zn [t] ≤
2B + l∈O(n) Cl , ∀t. This is summarized in the next corollary.
(f ),∗
(f )
(f ),∗
Corollary 1: Let y∗ = [x∗f ; µl
]f ∈F ,l∈L be an optimal solution to problem (1)-(6) given in Fact 1, i.e., gn (yn ) =
0, ∀f ∈ F, ∀n ∈ N \ Dst(f ), and λ∗ be a Lagrange multiplier vector given in Assumption 2. If αn ≥ 21 (dn + 1)2 , ∀n ∈ N in
10
(f )
Algorithm 1, where dn is the degree of node n, then all actual physical queues Zn [t], ∀f ∈ F, ∀n ∈ N \ {Dst(f )} in the
network evolving via (8) satisfy
X
p
Cl , ∀t.
Zn(f ) [t] ≤4kλ∗ k + 2 2ζ +
l∈O(n)
where ζ is a constant defined in Lemma 8.
E. Performance of Algorithm 1
Theorems 1 and 2 together imply that Algorithm 1 with αn ≥ 12 (dn + 1), ∀n ∈ N can achieve a vanishing utility optimality
gap that decays like O(1/t), where t is number of iterations, and guarantees the physical queues at each node for each session
are always bounded by a constant that is independent of the utility optimality gap.
This is superior to existing backpressure algorithms from [5], [4], [10] that can achieve an O(1/V ) utility gap only at
the cost of an O(V 2 ) or O(V ) queue length, where V is an algorithm parameter. To obtain a vanishing utility gap, existing
backpressure algorithms in [5], [4], [10] necessarily yield unbounded queues. To obtain a vanishing utility gap, existing
backpressure algorithms in [5], [4] yield unbounded queues. We also comment that O(V 2 ) queue bound in the primal-dual
type backpressure algorithm [5] is actually of the order V 2 kλ∗ k + B1 where λ∗ is the Lagrangian multiplier vector attaining
strong duality and B1 is a constant determined by the problem parameters. A recent work [21] also shows that the O(V ) queue
bound in the backpressure algorithm from drift-plus-penalty is of the order V kλ∗ k+B2 where B2 is also a constant determined
by the problem parameters. Since λ∗ is a constant vector independent of V , both algorithms are claimed to have O(V 2 ) or
O(V ) queue bounds. By Corollary 1, Algorithm 1 guarantees physical queues at each node are bounded by 4kλ∗ k + B3 ,
where B3 is constant given a problem. Thus, the constant queue bound guaranteed by Algorithm 1 is typically smaller than
the O(V 2 ) or O(V ) queue bounds from [5] and [21] even for a small V . (A small V can yield a poor utility performance in
the backpressure algorithms in [5], [4].)
V. N UMERICAL E XPERIMENT
In this section, we consider a simple network with 6 nodes and 8 links and 2 sessions as described in Figure 1. This network
has two sessions: session 1 from node 1 to node 6 has utility function log(x1 ) and session 2 from node 3 to node 4 has utility
function 1.5 log(x2 ). (The log utilities are widely used as metrics of proportional fairness in the network [17].) The routing
path of each session is arbitrary as long as data can be delivered from the source node to the destination node. For simplicity,
assume that each link has capacity 1. The optimal source session rate to problem (1)-(6) is x∗1 = 1.2 and x∗2 = 1.8 and link
session rates, i.e., static routing for each session, is drawn in Figure 2.
2
l3
4
l7
l1
11
l4
l2
Session 1: 1->6
3
l5
6
l6
5
l8
Session 2: 3->4
Fig. 1.
A simple network with 6 nodes, 8 links and 2 sessions.
To compare the convergence performance of Algorithm 1 and the backpressure algorithm in [4] (with the best utility-delay
tradeoff among all existing backpressure algorithms), we run both Algorithm 1 with αn = 12 dn + 1), ∀n ∈ N and the
backpressure algorithm in [4] with V = 500 to plot Figure 3. It can be observed from Figure 3 that Algorithm 1 converges to
the optimal source session rates faster than the backpressure algorithm in [4]. The backpressure algorithm in [4] with V = 400
takes around 2500 iterations to converges to source rates close to (1.2, 1.8) while Algorithm 1 only takes around 800 iterations
to converges to (1.2, 1.8) (as shown in the zoom-in subfigure at the top right corner.) In fact, the backpressure algorithm in [4]
with V = 500 can not converge to the exact optimal source session rate (1.2, 1.8) but can only converge to its neighborhood
with a distance gap determined by the value of V . This is an effect from the fundamental [O(1/V ), O(V )] utility-delay tradeoff
11
2
1
4
11
6
0.2
Session 1: 1->6
4
1
1
0.8
1
0.2
3
0.2
5
0.8
3
5
Session 2: 3->4
Fig. 2.
The optimal routing for the network in Figure 1.
of the the backpressure algorithm in [4]. In contrast, Algorithm 1 can eventually converge to the the exact optimal source
session rate (1.2, 1.8). A zoom-in subfigure at the bottom right corner in Figure 1 verifies this and shows that the source
rate for Session 1 in Algorithm 1 converges to 1.2 while the source rate in the backpressure algorithm in [4] with V = 500
oscillates around a point slightly larger than 1.2.
5
4.5
Backpressure Alg in [4]
New Backpressure Alg: Algorithm 1
2
4
1.5
1
3.5
Source Rates
0.5
3
0
Session 2: 3->4
0
200
400
600
800
2.5
2
1.5
1
1.205
1.2
1.195
3000
0.5
Session 1: 1->6
0
0
500
1000
1500
2000
2500
3500
3000
4000
3500
4000
Iterations
Fig. 3.
Convergence performance comparison between Algorithm 1 and the backpressure algorithm in [4].
√
∗
∗
PCorollary 1 shows that Algorithm 1 guarantees each actual queue in the network is bounded by constant 4kλ k+2 2ξky k+
l∈O(n) Cl . Recall that the backpressure algorithm in [4] can guarantee the actual queues in the network are bounded by a
constant of order V kλ∗ k. Figure 4 plots the sum of actual queue length at each node for Algorithm 1 and the backpressure
algorithm in [4] with V = 10, 100 and 500. (Recall a larger V in the backpressure algorithm in [4] yields a smaller utility
gap but a larger queue length.) It can be observed that Algorithm 1 has the smallest actual queue length (see the zoom-in
subfigure) and the actual queue length of the backpressure algorithm in [4] scales linearly with respect to V .
VI. C ONCLUSION
This paper develops a new first-order Lagrangian dual type backpressure algorithm for joint rate control and routing in multihop data networks. The new backpressure algorithm can achieve vanishing utility optimality gaps and finite queue lengths.
This improves the state-of-art [O(1/V ), O(V 2 )] or [O(1/V ), O(V )] utility-delay tradeoff attained by existing backpressure
algorithms [5], [9], [7], [10].
12
1400
Network Sum Physical Queue Length
1200
Backpressure Alg in [4] (V=500)
1000
800
40
30
600
20
10
400
0
0
50
100
150
200
250
300
Backpressure Alg in [4] (V=100)
200
Backpressure Alg in [4] (V=10)
0
New Backpressure Alg: Algorithm 1
0
500
1000
1500
2000
2500
3000
Iterations
Fig. 4.
Actual Queue length comparison between Algorithm 1 and the backpressure algorithm in [4].
A PPENDIX A
N ETWORK U TILITY M AXIMIZATION WITH P REDETERMINED M ULTI -PATH
Consider multi-path network utility maximization in [16] where each session has multiple given paths, then the source
session rate xf in problem (1)-(6) becomes a vector xf = [xf,j ]j∈Pf where Pf is the set of paths for session f and the
(f )
(f )
(f,j)
(f )
link session rate µl becomes a vector µl = [µl ]j∈Pl . Define Sl as the set of paths for session f that are allowed
(f )
to use link l. Note that if all paths of session f are forbidden to use link l, then Sl = ∅. The multi-path network utility
maximization problem can be formulated as follows:
X
X
max
Uf (
xf,j )
f ∈F
s.t.
X
j∈Pf
xf,j 1{n=Src(f )} +
j∈Pl
X X
(f )
µl
j∈Pl l∈I(n)
X X
(f,j)
µl
≤
X X
(f )
µl , ∀f ∈ F, ∀n ∈ N \ {Dst(f )}
j∈Pl l∈O(n)
≤ Cl , ∀l ∈ L,
f ∈F j∈Pl
(f,j)
µl
(f )
≥ 0, ∀l ∈ L, ∀f ∈ F, ∀j ∈ Sl ,
(f,j)
(f )
µl
= 0, ∀l ∈ L, ∀f ∈ F, ∀j ∈ Pf \ Sl ,
X
xf,j ∈ dom(Uf ), ∀f ∈ F,
j∈Pl
xf,j ≥ 0, ∀f ∈ F, j ∈ Pf
The above formulation is in the form of problem (1)-(6) except that the variable dimension is extended.
A PPENDIX B
A N E XAMPLE I LLUSTRATING THE P OSSIBLY L ARGE G AP BETWEEN M ODEL (7) AND M ODEL (8)
Consider a network example shown in Figure 5. The network has 3k + 1 nodes where only node 0 is a destination; and
ai , i ∈ {1, 2, . . . , k} and bi , i ∈ {1, 2, . . . , k} can have exogenous arrivals. Assume all link capacities are equal to 1; and the
exogenous arrivals are periodic with period 2k, as follows:
• Time slot 1: One packet arrives at node a1 .
• Time slot 2: One packet arrives at node a2 .
13
···
Time slot k: One packet arrives at node ak .
• Time slot k + 1: One packet arrives at node b1 .
• Time slot k + 2: One packet arrives at node b2 .
• ···
• Time slot 2k: One packet arrives at node bk .
Under dynamics (7), each packet arrives on its own slot and traverses all links of its path to exit on the same slot it arrived.
The queue backlog in each node is 0 for all time.
Under dynamics (8), the first packet arrives at time slot 1 to node a1 . This packet visits node a2 at time slot 2, when the
second packet also arrives at a2 . One of these packets is delivered to node a3 at time slot 3, and another packet also arrives to
node 3. The nodes {1, . . . , k} do not have any exogenous arrivals and act only to delay the delivery of all packets from the ai
nodes. It follows that the link from node k to node 0 will send exactly one packet over each slots t ∈ {2k+1, 2k+2, . . . , 2k+k}.
Similarly, the link from bk to 0 sends exactly one packet to node 0 over each of these same slots. Thus, node 0 receives 2
packets on each slot t ∈ {2k + 1, 2k + 2, . . . , 2k + k}, but can only output 1 packet per slot. The queue backlog in this node
grows linearly and reaches k + 1 at time 2k + k. Thus, the backlog in node 0 can be arbitrarily large when k is large. This
example demonstrates that, even when there is only one destination, the deviation between virtual queues under dynamics (7)
and actual queues under dynamics (8) can be arbitrarily large, even when the in-degree and out-degree of 1 and an in-degree
of at most 2.
•
•
a1
ak-1
a2
ak
1
2
k
b1
Fig. 5.
b2
bk-1
bk
0
An example illustrating the possibly large gap between queue model (7) and queue model (8)
A PPENDIX C
P ROOF OF PART (2) IN L EMMA 1
Fix f ∈ F, n ∈ N \ {Dst(f )}. By (10),
)
b (f
Q
n [t + 1]
X (f )
X (f )
(f )
b n [t] −
= max Q
µl [t], 0 + xf [t]1{n=Src(f )} +
µl [t]
l∈O(n)
l∈I(n)
n
X (f )
X (f )
X (f ) o
)
b (f
= max Q
µl [t] −
µl [t], xf [t]1{n=Src(f )} +
µl [t]
n [t] + xf [t]1{n=Src(f )} +
l∈I(n)
l∈O(n)
l∈I(n)
(a)
n
o
X (f )
X (f )
)
b (f
≥ max Q
µl [t] −
µl [t], 0
n [t] + xf [t]1{n=Src(f )} +
l∈I(n)
l∈O(n)
(f )
where (a) follows from the fact that µl [t], xf [t], ∀f, l, t are non-negative.
(f )
)
b (f
is identical to the right side of (7) and recall that Yn [0] = 0 < Q
n [0].
P
(f
)
(f
)
(f
)
b n [t] = Qn [t] + B +
Since Q
l∈O(n) Cl , ∀t and Qn [t] ≤ B, ∀t, we have
P
(f )
Yn [t] ≤ 2B + l∈O(n) Cl , ∀t.
Note that the right side of the above equation
(f )
)
b (f
By inductions, we have Yn [t] ≤ Q
n [t], ∀t.
P
(f
)
b n [t] ≤ 2B +
Q
l∈O(n) Cl , ∀t. It follows that
A PPENDIX D
P ROOF OF L EMMA 3
Note that problem (18)-(20) satisfies Slater’s condition. So the optimal solution toP
problem (18)-(20) is characterized by KKT
K
conditions [22]. Introducing Lagrange multipliers θ ∈ R+ for inequality constraint k=1 zk ≤ b and ν = [ν1 , . . . , νK ]T ∈ RK
+
∗
∗
∗ T
for inequality constraints zk ≥ 0, k ∈ {1, 2, . . . , K}. Let z = [z1 , . . . , zK ] and (θ∗ , ν ∗ ) be any primal and dual pair
PK
with the zero duality
gap. By KKT conditions, we have zk∗ − ak + θ∗ − νk∗ = 0, ∀k ∈ {1, 2, . . . , K}; k=1 zk∗ ≤ b; θ∗ ≥
P
K
∗
∗
∗
∗ ∗
0; θ∗
k=1 zk − b = 0; zk ≥ 0, ∀k ∈ {1, 2, . . . , K}; νk ≥ 0, ∀k ∈ {1, 2, . . . , K}; νk zk = 0, ∀k ∈ {1, 2, . . . , K}.
14
PK
PK
∗
νk∗ , ∀k ∈ {1, 2, . . . , K} in all equations yields θ∗ ≥ ak −zk∗ , k ∈ {1, 2, . . . , K}; k=1 zk∗ ≤ b; θ∗ ≥ 0; θ∗
k=1 zk −
Eliminating
∗
∗
∗ ∗
b = 0; zk ≥ 0, ∀k ∈ {1, 2, . . . , K}; (zk − ak + θ )zk = 0, ∀k ∈ {1, 2, . . . , K}.
For all k ∈ {1, 2, . . . , K}, we consider θ∗ < ak and θ∗ ≥ ak separately:
1) If θ∗ < ak , then θ∗ ≥ ak − zk∗ holds only when zk∗ > 0, which by (zk∗ − ak + θ∗ )zk∗ = 0 implies that zk∗ = ak − θ∗ .
2) If θ∗ ≥ ak , then zk∗ > 0 is impossible, because zk∗ > 0 implies that zk∗ − ak + θ∗ > 0, which together with zk∗ > 0
contradicts the slackness condition (zk∗ − ak + θ∗ )zk∗ = 0. Thus, if θ∗ ≥ ak , we must have zk∗ = 0.
PK
Summarizing both cases, wehave zk∗ = max{0, ak − θ∗ }, ∀k ∈ {1, 2, . . . , K}, where θ∗ is chosen such that k=1 zk∗ ≤ b,
P
K
∗
θ∗ ≥ 0 and θ∗
k=1 zk − b = 0.
PK
∗
zk∗ − b is guaranteed to hold
To find such θ , we first check if θ∗ = 0. If θ∗ = 0 is true, the slackness condition θ∗
k=1P
PK
P
K
K
and we need to further require k=1 zk∗ = k=1 max{0, ak } ≤ b. Thus θ∗ = 0 if and only if k=1 max{0, ak } ≤ b. Thus,
PK
Algorithm 2 check if k=1 max{0, ak } ≤ b holds at the first step and if this is true, then we conclude θ∗ = 0 and we are
done!
PK
PK
PK
Otherwise, we know θ∗ > 0. By the slackness condition θ∗
zk∗ −b = 0, we must have k=1 zk∗ = k=1 max{0, ak −
k=1
PK
θ∗ } = b. To find θ∗ > 0 such that k=1 max{0, ak − θ∗ } = b, we could apply a bisection search by noting that all zk∗ are
decreasing with respect to θ∗ .
Another algorithm of finding θ∗ is inspired by the observation that if aj ≥ ai , ∀i, j ∈ {1, 2, . . . , K}, then zj∗ ≥ zi∗ . Thus, we
first sort all ak in a decreasing order, say π is the permutation such that aπ(1) ≥ aπ(2) ≥ · · · ≥ aπ(K) ; and then sequentially
check if k ∈ {1, 2, . . . , K} is the index such that aπ(k) −θ∗ ≥ 0 and aπ(k+1) −θ∗ < 0. To check this, we first assume k is indeed
Pk
such an index and solve the equation j=1 (aπ(j) − θ∗ ) = b to obtain θ∗ ; (Note that in Algorithm 2, to avoid recalculating the
Pk
Pk
partial sum j=1 aπ(j) for each k, we introduce the parameter Sk = j=1 aπ(j) and update Sk incrementally. By doing this,
the complexity of each iteration in the loop is only O(1).) then verify the assumption by checking if θ∗ ≥ 0, aπ(k) − θ∗ ≥ 0
and aπ(k+1) − θ∗ ≤ 0. The algorithm is described in Algorithm 2 and has complexity O(K log(K)). The overall complexity
is dominated by the step of sorting all ak .
A PPENDIX E
P ROOF OF L EMMA 6
The objective function (28) can be rewritten as
X
f (y) −
Wn(f ) [t]gn(f ) (yn(f ) ) + αn kyn(f ) − yn(f ) [t − 1]k2 −
f ∈F ,
n∈N \Dst(f )
(a)
=
X
f ∈F
−
l∈I(n)
f ∈F ,
n∈N \Dst(f )
−
f ∈F ,n=Dst(f )
(b)
=
X
X
αn (xf − xf [t − 1])2 1{n=Src(f )} +
X
X
(f )
(µl
l∈I(n)
(f )
(µl
(f )
− µl [t − 1])2 +
X
(f )
(µl
(f )
− µl [t − 1])2
l∈O(n)
(f )
− µl [t − 1])2
l∈I(n)
(f )
Uf (xf ) − WSrc(f ) [t]xf − αSrc(f ) (xf − xf [t − 1])2 +
f ∈F
−
(f )
− µl [t − 1])2
l∈O(n)
l∈I(n)
αn
(f )
(µl
X (f )
X (f )
Wn(f ) [t] xf 1{n=Src(f )} +
µl −
µl
f ∈F ,
n∈N \Dst(f )
X
X
αn
f ∈F ,n=Dst(f )
X
Uf (xf ) −
X
X
X
(f )
(f )
Wn(f ) [t] − Wm
[t] µ(n,m)
(n,m)∈L f ∈F
X
X
(αn + αm )
(n,m)∈L
(f )
(µ(n,m)
−
(f )
µ(n,m) [t
2
− 1])
(30)
f ∈F
P
P
(f )
(f )
(f )
(f )
(f )
(f )
where (a) follows from the fact that gn (yn ) = xf 1{n=Src(f )} + l∈I(n) µl − l∈O(n) µl and kyn − yn [t − 1]k2 =
P
P
(f )
(f )
(f )
(f )
(xf − xf [t − 1])2 1{n=Src(f )} + l∈I(n) (µl − µl [t − 1])2 + l∈O(n) (µl − µl [t − 1])2 ; and (b) follows by collecting
(f )
(f )
(f )
(f )
each linear term µl and each quadratic term (µl − µl [t − 1])2 . Note that each link session rate µl appears twice with
P
P
P
(f )
(f )
(f )
opposite signs in the summation term f ∈F ,n∈N \{Dst(f )} Wn [t] xf 1{n=Src(f )} + l∈I(n) µl − l∈O(n) µl
unless link
(f )
l flows into Dst(f ) and recall that WDst(f ) = 0, ∀f ∈ F. The quadratic terms are collected in a similar way. Note that the
P
P
(f )
(f )
term f ∈F ,n=Dst(f ) αn l∈I(n) (µl − µl [t − 1])2 introduced to the objective function (28) is necessary to guarantee each
(f )
(f )
quadratic term (µ(m,n) − µ(m,n) [t − 1])2 with the same link index (n, m) but different flow indices f ∈ F have the same
coefficient αn + αm in the last line of (30).
(f )
Note that equation (30) is now separable for each scalar xf and vector [µ(n,m) ]f ∈F . Thus, problem (28)-(29) can be
decomposed into independent smaller optimization problems in the form of problem (12)-(13) with respect to each scalar xf ,
(f )
and in the form of problem (14)-(17) with respect to each vector [µ(n,m) ]f ∈F .
15
A PPENDIX F
P ROOF OF L EMMA 7
(f )
(f )
Note that Wn [t] appears as a known constant in (12). Since Uf (xf ) is concave and Wn [t]xf is linear, it follows that
(12) is strongly concave with respect to xf with modulus 2αn . Since xf [t] is chosen to solve (12)-(13), by Lemma 4, ∀f ∈ F,
we have
(f )
(f )
Uf (xf [t]) − WSrc(f ) [t]xf [t] − αn (xf [t] − xf [t − 1])2 ≥ Uf (x∗f ) − WSrc(f ) [t]x∗f − αn (x∗f − xf [t − 1])2 + αn (x∗f − xf [t])2 .
{z
} |
{z
}
|
(31)-I
(31)-II
(31)
Similarly, we know (14) is strongly concave with respect to vector [µf(n,m) ]f ∈F with modulus 2(αn + αm ). By Lemma 4,
∀(n, m) ∈ O(n), we have
X
(f )
X (f )
2
(f )
(f )
Wn(f ) [t] − Wm
[t] µ(n,m) [t] − αn + αm
µ(n,m) [t] − µ(n,m) [t − 1]
f ∈F
f ∈F
|
≥
{z
}
(32)-I
X
(f ),∗
X (f ),∗
2
X (f ),∗
2
(f )
(f )
(f )
Wn(f ) [t] − Wm
[t] µ(n,m) − αn + αm
µ(n,m) − µ(n,m) [t − 1] + αn + αm
µ(n,m) − µ(n,m) [t] .
f ∈F
f ∈F
|
f ∈F
{z
}
(32)-II
(32)
(f )
Recall that each column vector yn defined in (21) is composed by control actions that appear in each constraint (2); column
P
(f )
vector y = [xf ; µl ]f ∈F ,l∈L is the collection of all control actions; and f (y) = f ∈F Uf (xf ). Summing term (31)-I over
all f ∈ F and term (32)-I over all (n, m) ∈ L and using an argument similar to the proof of Lemma 6 (Recall that y[t] is
jointly chosen to minimize (28) by Lemma 6.) yields
X
X
(31)-I +
(32)-I
f ∈F
(n,m)∈N
X
=f (y[t]) −
Wn(f ) [t]gn(f ) (yn(f ) [t]) + αn kyn(f ) [t] − yn(f ) [t − 1]k2 −
X
f ∈F ,
n=Dst(f )
f ∈F ,
n∈N \Dst(f )
(f )
X
αn
(f )
(µl [t] − µl [t − 1])2 . (33)
l∈I(n)
P
P
(f ),∗
(f )
(f )
(f ),∗
− µl [t])2 . Summing
− yn [t]k2 + αn 1{n=Dst(f )} l∈I(n) (µl
Recall that Φ[t] = f ∈F ,n∈N αn 1{n6=Dst(f )} kyn
term (31)-II over all f ∈ F and term (32)-II over all (n, m) ∈ L yields
X
X
X
(31)-II +
(32)-II = f (y∗ ) + Φ[t] − Φ[t − 1] −
Wn(f ) [t]gn(f ) (yn(f ),∗ ),
(34)
f ∈F
f ∈F ,
n∈N \Dst(f )
(n,m)∈N
Combining (31)-(34) and rearranging terms yields
f (y[t])
X
≥f (y∗ ) + Φ[t] − Φ[t − 1] −
f ∈F ,
n∈N \Dst(f )
+
X
f ∈F ,
n=Dst(f )
αn
X
(f )
X
Wn(f ) [t]gn(f ) (yn(f ),∗ ) +
Wn(f ) [t]gn(f ) (yn(f ) [t]) + αn kyn(f ) [t] − yn(f ) [t − 1]k2
f ∈F ,
n∈N \Dst(f )
(f )
(µl [t] − µl [t − 1])2
l∈I(n)
(a)
X
≥ f (y∗ ) + Φ[t] − Φ[t − 1] +
Wn(f ) [t]gn(f ) (yn(f ) [t]) + αn kyn(f ) [t] − yn(f ) [t − 1]k2
f ∈F ,
n∈N \Dst(f )
(b)
X
=f (y∗ ) + Φ[t] − Φ[t − 1] +
)
(f )
(f )
(f )
(f )
(f )
(f )
(f )
(f )
2
Q(f
n [t]gn (yn [t]) + gn (yn [t − 1])gn (yn [t]) + αn kyn [t] − yn [t − 1]k
f ∈F ,
n∈N \Dst(f )
(35)
(f )
(f ),∗
where (a) follows because gn (yn
0; (b) follows from the fact that
) = 0, ∀f ∈ F, ∀n ∈ N \Dst(f ), and
(f )
Wn [t]
=
(f )
Qn [t]
+
(f )
(f )
gn (yn [t
− 1]).
P
f ∈F ,n=Dst(f )
αn
(f )
(f )
2
l∈I(n) (µl [t]−µl [t−1])
P
≥
16
Recall that uT1 u2 = 12 u21 + 21 u22 − 12 (u1 − u2 )2 for any u1 , u2 ∈ R. Thus, for all f ∈ F, n ∈ N \ Dst(f ), we have
gn(f ) (yn(f ) [t − 1])gn(f ) (yn(f ) [t]) =
2 1
2 1
2
1 (f ) (f )
g (yn [t − 1]) + gn(f ) (yn(f ) [t]) − gn(f ) (yn(f ) [t − 1]) − gn(f ) (yn(f ) [t]) .
2 n
2
2
(36)
Substituting (36) into (35) yields
f (y[t]) ≥f (y∗ ) + Φ[t] − Φ[t − 1] +
X
)
(f )
(f )
Q(f
n [t]gn (yn [t]) +
f ∈F ,
n∈N \Dst(f )
2 1
2
1 (f ) (f )
gn (yn [t − 1]) + gn(f ) (yn(f ) [t])
2
2
2
1 (f ) (f )
gn (yn [t − 1]) − gn(f ) (yn(f ) [t]) + αn kyn(f ) [t] − yn(f ) [t − 1]k2
2
X
(a)
2 1
2
1 (f ) (f )
)
(f )
(f )
gn (yn [t − 1]) + gn(f ) (yn(f ) [t])
Q(f
≥ f (y∗ ) + Φ[t] − Φ[t − 1] +
n [t]gn (yn [t]) +
2
2
−
f ∈F ,
n∈N \Dst(f )
1
+ αn − βn2 kyn(f ) [t] − yn(f ) [t − 1]k2
2
X
(b)
1 (f ) (f ) 2
)
(f )
(f )
Q(f
≥f (y∗ ) + Φ[t] − Φ[t − 1] +
[t]g
(y
[t])
+
g (yn [t])
n
n
n
2 n
(37)
f ∈F ,
n∈N \Dst(f )
(f )
where (a) follows from the Fact 2, i.e., each gn (·) is Lipschitz with modulus βn and (b) follows because αn ≥ 12 (dn + 1),
2
√
(f )
(f )
βn ≤ dn + 1 and 12 gn (yn [t − 1]) ≥ 0.
Subtracting (26) from (37) and cancelling the common terms on both sides yields
−∆[t] + f (y[t]) ≥ f (y∗ ) + Φ[t] − Φ[t − 1].
R EFERENCES
[1] L. Tassiulas and A. Ephremides, “Stability properties of constrained queueing systems and scheduling policies for maximum throughput in multihop
radio networks,” IEEE Transactions on Automatic Control, vol. 37, no. 12, pp. 1936–1948, 1992.
[2] M. J. Neely, “Dynamic power allocation and routing for satellite and wireless networks with time varying channels,” Ph.D. dissertation, Massachusetts
Institute of Technology, 2003.
[3] L. Georgiadis, M. J. Neely, and L. Tassiulas, “Resource allocation and cross-layer control in wireless networks,” Foundations and Trends in Networking,
2006.
[4] M. J. Neely, Stochastic network optimization with application to communication and queueing systems. Morgan & Claypool Publishers, 2010.
[5] A. Eryilmaz and R. Srikant, “Joint congestion control, routing, and mac for stability and fairness in wireless networks,” IEEE Journal on Selected Areas
in Communications, vol. 24, no. 8, pp. 1514–1524, 2006.
[6] A. L. Stolyar, “Maximizing queueing network utility subject to stability: Greedy primal-dual algorithm,” Queueing Systems, vol. 50, no. 4, pp. 401–457,
2005.
[7] X. Lin and N. B. Shroff, “Joint rate control and scheduling in multihop wireless networks,” in Proceedings of IEEE Conference on Decision and Control
(CDC), 2004.
[8] J.-W. Lee, R. R. Mazumdar, and N. B. Shroff, “Opportunistic power scheduling for dynamic multi-server wireless systems,” IEEE Transactions on
Wireless Communications, vol. 5, no. 6, pp. 1506–1515, 2006.
[9] M. J. Neely, E. Modiano, and C. E. Rohrs, “Dynamic power allocation and routing for time-varying wireless networks,” IEEE Journal on Selected Areas
in Communications, vol. 23, no. 1, pp. 89–103, 2005.
[10] J. Liu, N. B. Shroff, C. H. Xia, and H. D. Sherali, “Joint congestion control and routing optimization: An efficient second-order distributed approach,”
IEEE/ACM Transactions on Networking, vol. 24, no. 3, pp. 1404–1420, 2015.
[11] M. J. Neely, “Super-fast delay tradeoffs for utility optimal fair scheduling in wireless networks,” IEEE Journal on Selected Areas in Communications,
vol. 24, no. 8, pp. 1489–1501, 2006.
[12] L. Huang, S. Moeller, M. J. Neely, and B. Krishnamachari, “LIFO-backpressure achieves near-optimal utility-delay tradeoff,” IEEE/ACM Transactions
on Networking, vol. 21, no. 3, pp. 831–844, 2013.
[13] E. Wei, A. Ozdaglar, and A. Jadbabaie, “A distributed Newton method for network utility maximization–I: algorithm,” IEEE Transactions on Automatic
Control, vol. 58, no. 9, pp. 2162–2175, 2013.
[14] M. Zargham, A. Ribeiro, and A. Jadbabaie, “Accelerated backpressure algorithm,” in Proceedings of IEEE Global Communications Conference
(GLOBECOM), 2013.
[15] N. Parikh and S. Boyd, “Proximal algorithms,” Foundations and Trends in Optimization, vol. 1, no. 3, pp. 123–231, 2013.
[16] X. Lin and N. B. Shroff, “Utility maximization for communication networks with multipath routing,” IEEE Transactions on Automatic Control, vol. 51,
no. 5, pp. 766–781, 2006.
[17] F. P. Kelly, A. K. Maulloo, and D. K. Tan, “Rate control for communication networks: Shadow prices, proportional fairness and stability,” Journal of
the Operational Research Society, pp. 237–252, 1998.
[18] S. H. Low and D. E. Lapsley, “Optimization flow control—I: basic algorithm and convergence,” IEEE/ACM Transactions on Networking, vol. 7, no. 6,
pp. 861–874, 1999.
[19] D. P. Bertsekas, A. Nedić, and A. E. Ozdaglar, Convex Analysis and Optimization. Athena Scientific, 2003.
[20] J. Duchi, S. Shalev-Shwartz, Y. Singer, and T. Chandra, “Efficient projections onto the l1 -ball for learning in high dimensions,” in Proceedings of
International Conference on Machine learning (ICML), 2008.
[21] M. J. Neely, “A simple convergence time analysis of drift-plus-penalty for stochastic optimization and convex programs,” arXiv preprint arXiv:1412.079,
2014.
[22] S. Boyd and L. Vandenberghe, Convex Optimization. Cambridge University Press, 2004.
| 3cs.SY
|
High-dimensional Bayesian inference via the Unadjusted
Langevin Algorithm
arXiv:1605.01559v3 [math.ST] 21 Feb 2018
Alain Durmus1 and Éric Moulines2
1
CMLA - École normale supérieure Paris-Saclay, CNRS, Université Paris-Saclay,
94235 Cachan, France.
2
Centre de Mathématiques Appliquées, UMR 7641, Ecole Polytechnique, France.
February 22, 2018
Keywords: total variation distance, Langevin diffusion, Markov Chain Monte Carlo,
Metropolis Adjusted Langevin Algorithm, Rate of convergence
AMS subject classification (2010): primary 65C05, 60F05, 62L10; secondary 65C40,
60J05,93E35
Abstract
: We consider in this paper the problem of sampling a high-dimensional probability distribution π having a density w.r.t. theR Lebesgue measure on Rd , known up
to a normalization factor x 7→ π(x) = e−U(x) / Rd e−U(y) dy. Such problem naturally
occurs for example in Bayesian inference and machine learning. Under the assumption that U is continuously differentiable, ∇U is globally Lipschitz and U is strongly
convex, we obtain non-asymptotic bounds for the convergence to stationarity in
Wasserstein distance of order 2 and total variation distance of the sampling method
based on the Euler discretization of the Langevin stochastic differential equation,
for both constant and decreasing step sizes. The dependence on the dimension of
the state space of the obtained bounds is studied to demonstrate the applicability
of this method. The convergence of an appropriately weighted empirical measure
is also investigated and bounds for the mean square error and exponential deviation inequality are reported for functions which are measurable and bounded. An
illustration to Bayesian inference for binary regression is presented.
1
Introduction
There has been recently an increasing interest in Bayesian inference applied to highdimensional models often motivated by machine learning applications. Rather than
1
2
Email: [email protected]
[email protected]
1
obtaining a point estimate, Bayesian methods attempt to sample the full posterior distribution over the parameters and possibly latent variables which provides a way to
assert uncertainty in the model and prevents from overfitting [28], [37].
The problem can be formulated as follows. RWe aim at sampling a posterior distribution π on Rd , d ≥ 1, with density x 7→ e−U (x) / Rd e−U (y) dy w.r.t. the Lebesgue measure,
where U is continuously differentiable. The Langevin stochastic differential equation
associated with π is defined by:
√
(1)
dYt = −∇U (Yt )dt + 2dBt ,
where (Bt )t≥0 is a d-dimensional Brownian motion defined on the filtered probability
space (Ω, F, (Ft )t≥0 , P), satisfying the usual conditions. Under mild technical conditions,
the Langevin diffusion admits π as its unique invariant distribution.
We study the sampling method based on the Euler-Maruyama discretization of
(1). This scheme defines the (possibly) non-homogeneous, discrete-time Markov chain
(Xk )k≥0 given by
p
Xk+1 = Xk − γk+1 ∇U (Xk ) + 2γk+1 Zk+1 ,
(2)
where (Zk )k≥1 is an i.i.d. sequence of d-dimensional standard Gaussian random variables
and (γk )k≥1 is a sequence of step sizes, which can either be held constant or be chosen
to decrease to 0. This algorithm has been first proposed by [14] and [30] for molecular
dynamics applications. Then it has been popularized in machine learning by [17], [18]
and computational statistics by [28] and [32]. Following [32], in the sequel this method
will be refer to as the unadjusted Langevin algorithm (ULA). When the step sizes are held
constant, under appropriate conditions on U , the homogeneous Markov chain (Xk )k≥0
has a unique stationary distribution πγ , which in most of the cases differs from the
distribution π. It has been proposed in [33] and [32] to use a Metropolis-Hastings step
at each iteration to enforce reversibility w.r.t. π. This algorithm is referred to as the
Metropolis adjusted Langevin algorithm (MALA).
The ULA algorithm has already been studied in depth for constant step sizes in
[35], [32] and [27]. In particular, [35, Theorem 4] gives an asymptotic expansion for the
weakPerror between π and πγ . For sequence of step sizes such that limk→+∞ γk = 0
and ∞
k=1 γk = ∞, weak convergence of the weighted empirical distribution of the ULA
algorithm has been established in [23], [24] and [25].
Contrary to these previously reported works, we focus in this paper on non-asymptotic
results. More precisely, we obtain explicit bounds between the distribution of the n-th
iterate of the Markov chain defined in (2) and the target distribution π in Wasserstein
and total variation distance for nonincreasing step sizes. When the sequence of step sizes
is held constant γk = γ for all k ≥ 0, both fixed horizon (the total computational budget
is fixed and the step size is chosen to minimize the upper bound on the Wasserstein or
total variation distance) and fixed precision (for a fixed target precision, the number of
iterations and the step size are optimized simultaneously to meet this constraint) strategies are studied. In addition,
Pquantitative estimates between π and πγ are obtained.
When limk→+∞ γk = 0 and ∞
k=1 γk = ∞, we show that the marginal distribution of
2
the non-homogeneous Markov chain (Xk )k≥0 converges to the target distribution π and
provide explicit convergence bounds. A special attention is paid to the dependency of
the proposed upper bounds on the dimension of the state space, since we are particularly
interested in the application of these methods to sampling in high-dimension.
Compared to [10] and [11], our contribution is twofold. First, we report sharper
convergence bounds for constant and non-increasing step sizes. In particular, it is shown
that the number of iterations required to reach the target distribution π with a precision
ǫ > 0 in total variation is upper bounded (up to logarithmic factors) by (ǫd)−1 , where
d is the dimension, under appropriate regularity assumption (in [11], the upper bound
was shown to be d−1 ǫ−2 ). We show that our result is optimal (up to logarithmic factors
again) in the case of the standard d-dimensional Gaussian distribution, for which explicit
computation can be done. Besides, we establish computable bound for the mean square
error associated with ULA for measurable bounded functions, which is an important
result in Bayesian statistics to estimate credibility regions. For that purpose, we study
the convergence of the Euler-Maruyama discretization towards its stationary distribution
in total variation using a discrete time version of reflection coupling introduced in [4].
The paper is organized as follows. In Section 2, we study the convergence in the
Wasserstein distance of order 2 of the Euler discretization for constant and decreasing
step sizes. In Section 3, we give non asymptotic bounds in total variation distance
between the Euler discretization and π. This study is completed in Section 4 by nonasymptotic bounds of convergence of the weighted empirical measure applied to bounded
and measurable functions. Our claims are supported in a Bayesian inference for a binary
regression model in Section 5. The proofs are given in Section 6. Finally in Section 8,
some results of independent interest, used in the proofs, on functional autoregressive
models are gathered. Some technical proofs and derivations are postponed and carried
out in a supplementary paper [12].
Notations and conventions
Denote by B(Rd ) the Borel σ-field of Rd , F(Rd ) the set of all Borel measurable functions
on Rd and for f ∈ F(Rd ), kf k∞ = supx∈Rd |f (x)|. For µ a probability measure on
(Rd , B(Rd )) and f ∈ F(Rd ) a µ-integrable function, denote by µ(f ) the integral of f
w.r.t. µ. We say that ζ is a transference plan of µ and ν if it is a probability measure
on (Rd × Rd , B(Rd × Rd )) such that for all measurable set A of Rd , ζ(A × Rd ) = µ(A)
and ζ(Rd × A) = ν(A). We denote by Π(µ, ν) the set of transference plans of µ and ν.
Furthermore, we say that a couple of Rd -random variables (X, Y ) is a coupling of µ and
ν if there exists ζ ∈ Π(µ, ν) such that (X, Y ) are distributed according to ζ. For two
probability measures µ and ν, we define the Wasserstein distance of order p ≥ 1 as
def
Wp (µ, ν) =
inf
Z
1/p
.
kx − yk dζ(x, y)
p
ζ∈Π(µ,ν) Rd ×Rd
By [36, Theorem 4.1], for all µ, ν probability measures on Rd , there exists a transference plan ζ ⋆ ∈ Π(µ, ν) such that for any coupling (X, Y ) distributed according to ζ ⋆ ,
3
Wp (µ, ν) = E[kX − Y kp ]1/p . This kind of transference plan (respectively coupling) will
be called an optimal transference plan (respectively optimal coupling) associated with
Wp . We denote
by Pp (Rd ) the set of probability measures with finite p-moment: for all
R
µ ∈ Pp (Rd ), Rd kxkp dµ(x) < +∞. By [36, Theorem 6.16], Pp (Rd ) equipped with the
Wasserstein distance Wp of order p is a complete separable metric space.
Let f : Rd → R be a Lipschitz function, namely there exists C ≥ 0 such that for all
x, y ∈ Rd , |f (x) − f (y)| ≤ C kx − yk. Then we denote
kf kLip = inf{|f (x) − f (y)| kx − yk−1 | x, y ∈ Rd , x 6= y} .
The Monge-Kantorovich theorem (see [36, Theorem 5.9]) implies that for all µ, ν probabilities measure on Rd ,
Z
Z
d
f (x)dν(x) | f : R → R ; kf kLip ≤ 1 .
f (x)dµ(x) −
W1 (µ, ν) = sup
Rd
Rd
Denote by Fb (Rd ) the set of all bounded Borel measurable functions on Rd . For f ∈
Fb (Rd ) set osc(f ) = supx,y∈Rd |f (x) − f (y)|. For two probability measures µ and ν on
Rd , the total variation distance distance between µ and ν is defined by kµ − νkTV =
supA∈B(Rd ) |µ(A) − ν(A)|. By the Monge-Kantorovich theorem the total variation distance
between µ and ν can be written on the form:
Z
1Dc (x, y)dζ(x, y) ,
kµ − νkTV = inf
ζ∈Π(µ,ν) Rd ×Rd
where D = {(x, y) ∈ Rd × Rd |x = y}. For all x ∈ Rd and M > 0, we denote by B(x, M ),
the ball centered at x of radius M . For a subset A ⊂ Rd , denote by Ac the complementary
of A. Let n ∈ N∗ and M be a n × n-matrix, then denote by M T the transpose of M
and kM k the operator norm associated with M defined by kM k = supkxk=1 kM xk.
Define the Frobenius norm associated with M by kM k2F = Tr(M T M ). Let n, m ∈ N∗
and F : Rn → Rm be a twice continuously differentiable function. Denote by ∇F and
~ the vector
∇2 F the Jacobian and the Hessian of F respectively. Denote also by ∆F
d
m
~ (x) is the vector of R such that for all
Laplacian of F defined by: for all x ∈ R , ∆F
P
~
i ∈ {1, · · · , m}, the i-th component P
of ∆F (x) is Q
equals to dj=1 (∂ 2 Fi /∂x2j )(x). In the
sequel, we take the convention that np = 0 and np = 1 for n, p ∈ N, n < p.
2
Non-asymptotic bounds in Wasserstein distance of order
2 for ULA
Consider the following assumption on the potential U :
H 1. The function U is continuously differentiable on Rd and gradient Lipschitz: there
exists L ≥ 0 such that for all x, y ∈ Rd , k∇U (x) − ∇U (y)k ≤ L kx − yk.
4
Under H1, for all x ∈ Rd by [21, Theorem 2.5, Theorem 2.9 Chapter 5] there exists
a unique strong solution (Yt )t≥0 to (1) with Y0 = x. Denote by (Pt )t≥0 the semi-group
associated with (1). It is well-known that π is its (unique) invariant probability. To get
geometric convergence of (Pt )t≥0 to π in Wasserstein distance of order 2, we make the
following additional assumption on the potential U .
H 2. U is strongly convex, i.e. there exists m > 0 such that for all x, y ∈ Rd ,
U (y) ≥ U (x) + h∇U (x), y − xi + (m/2) kx − yk2 .
Under H2, [29, Theorem 2.1.8] shows that U has a unique minimizer x⋆ ∈ Rd . If in
addition H1 holds, then [29, Theorem 2.1.12, Theorem 2.1.9] show that for all x, y ∈ Rd :
h∇U (y) − ∇U (x), y − xi ≥
1
κ
ky − xk2 +
k∇U (y) − ∇U (x)k2 ,
2
m+L
(3)
where
2mL
.
(4)
m+L
We briefly summarize some background material on the stability and the convergence in
W2 of the overdamped Langevin diffusion under H1 and H2. Most of the statements in
Proposition 1 are known and are recalled here for ease of references; see e.g. [6].
κ=
Proposition 1. Assume H1 and H2.
(i) For all t ≥ 0 and x ∈ Rd ,
Z
ky − x⋆ k2 Pt (x, dy) ≤ kx − x⋆ k2 e−2mt + (d/m)(1 − e−2mt ) .
Rd
(ii) The stationary distribution π satisfies
R
Rd
kx − x⋆ k2 π(dx) ≤ d/m.
(iii) For any x, y ∈ Rd and t > 0, W2 (δx Pt , δy Pt ) ≤ e−mt kx − yk.
(iv) For any x ∈ Rd and t > 0, W2 (δx Pt , π) ≤ e−mt kx − x⋆ k + (d/m)1/2 .
Proof. The proof is given in the supplementary document [12, Section 1.1].
Note that the convergence rate in Proposition 1-(iv) is independent from the dimension. Let (γk )k≥1 be a sequence of positive and non-increasing step sizes and for n, ℓ ∈ N,
denote by
ℓ
X
def
γk ,
Γn = Γ1,n .
(5)
Γn,ℓ =
k=n
For γ > 0, consider the Markov kernel Rγ given for all A ∈ B(Rd ) and x ∈ Rd by
Z
Rγ (x, A) = (4πγ)−d/2 exp −(4γ)−1 ky − x + γ∇U (x)k2 dy .
A
5
(6)
The process (Xk )k≥0 given in (2) is an inhomogeneous Markov chain with respect to the
family of Markov kernels (Rγk )k≥1 . For ℓ, n ∈ N∗ , ℓ ≥ n, define
Qn,ℓ
γ = Rγn · · · Rγℓ ,
Qnγ = Q1,n
γ
(7)
with the convention that for n, ℓ ∈ N, n < ℓ, Qℓ,n
γ is the identity operator.
∗
We first derive a Foster-Lyapunov drift condition for Qn,ℓ
γ , ℓ, n ∈ N , ℓ ≥ n.
Proposition 2. Assume H1 and H2.
(i) Let (γk )k≥1 be a nonincreasing sequence with γ1 ≤ 2/(m + L). Let x⋆ be the unique
minimizer of U . Then for all x ∈ Rd and n, ℓ ∈ N∗ ,
Z
ky − x⋆ k2 Qn,ℓ
γ (x, dy) ≤ ̺n,ℓ (x) ,
Rd
where ̺n,ℓ (x) is given by
̺n,ℓ (x) =
ℓ
Y
(1 − κγk ) kx − x⋆ k2 + 2dκ−1
k=n
(
1 − κ−1
ℓ
Y
i=n
)
(1 − κγi )
,
(8)
and κ is defined in (4).
(ii) For any γ ∈ (0, 2/(m + L)], Rγ has a unique stationary distribution πγ and
Z
kx − x⋆ k2 πγ (dx) ≤ 2dκ−1 .
Rd
Proof. The proof is postponed to [12, Section 1.2].
We now proceed to establish the contraction property of the sequence (Qnγ )n≥1 in
W2 . This result implies the geometric convergence of the sequence (δx Rγn )n≥1 to πγ in
W2 for all x ∈ Rd . Note that the convergence rate is again independent of the dimension.
Proposition 3. Assume H1 and H2. Then,
(i) Let (γk )k≥1 be a nonincreasing sequence with γ1 ≤ 2/(m + L). For all x, y ∈ Rd
and ℓ ≥ n ≥ 1,
n,ℓ
W2 (δx Qn,ℓ
γ , δy Qγ ) ≤
(
ℓ
Y
(1 − κγk ) kx − yk2
k=n
(ii) For any γ ∈ (0, 2/(m + L)), for all x ∈ Rd and n ≥ 1,
)1/2
.
n
o1/2
W2 (δx Rγn , πγ ) ≤ (1 − κγ)n/2 kx − x⋆ k2 + 2κ−1 d
.
6
Proof. The proof is postponed to [12, Section 1.3]
Corollary 4. Assume H 1 and H 2. Let (γk )k≥1 be a nonincreasing sequence with
γ1 ≤ 2/(m + L). Then for all Lipschitz functions f : Rd → R and ℓ ≥ n ≥ 1, Qn,ℓ
γ f is a
Qℓ
1/2 kf k
Lipschitz function with kQn,ℓ
(1
−
κγ
)
.
f
k
≤
γ
Lip
Lip
k
k=n
Proof. The proof follows from Proposition 3-(i) using
n,ℓ
n,ℓ
n,ℓ
Qn,ℓ
γ f (y) − Qγ f (z) ≤ kf kLip W2 (δy Qγ , δz Qγ ) .
We now proceed to establish explicit bounds for W2 (δx Qnγ , π), with x ∈ Rd . Since π
is invariant for Pt for all t ≥ 0, it suffices to get some bounds on W2 (δx Qnγ , ν0 PΓn ), with
ν0 ∈ P2 (Rd ) and take ν0 = π. To do so, we construct a coupling between the diffusion
and the linear interpolation of the Euler discretization. An obvious candidate is the
synchronous coupling (Yt , Y t )t≥0 defined for all n ≥ 0 and t ∈ [Γn , Γn+1 ) by
(
√
Rt
Yt = YΓn − Γn ∇U (Ys )ds + 2(Bt − BΓn )
√
(9)
Ȳt = ȲΓn − ∇U (ȲΓn )(t − Γn ) + 2(Bt − BΓn ) ,
with Y0 is distributed according to ν0 , Ȳ0 = x and (Γn )n≥1 is given in (5). Therefore since
for all n ≥ 0, W22 (δx Qnγ , ν0 PΓn ) ≤ E[kYΓn − ȲΓn k2 ], taking ν0 = π, we derive an explicit
bound on the Wasserstein distance between the sequence of distributions (δx Qkγ )k≥0 and
the stationary measure π of the Langevin diffusion (1).
Theorem 5. Assume H1 and H2. Let (γk )k≥1 be a nonincreasing sequence with γ1 ≤
1/(m + L). Then for all x ∈ Rd and n ≥ 1,
n
o
⋆ 2
(2)
W22 (δx Qnγ , π) ≤ u(1)
n (γ) kx − x k + d/m + un (γ) ,
where
u(1)
n (γ)
=2
n
Y
(1 − κγk /2)
(10)
k=1
κ is defined in (4) and
"
#
Y
n
n
X
−1
dL2 γi dL2 γi2
2
(2)
2
γi κ + γi
un (γ) = L
2d +
(1 − κγk /2) . (11)
+
m
6
i=1
k=i+1
Proof. Let x ∈ Rd , n ≥ 1 and ζ0 = π ⊗ δx . Let (Yt , Y t )t≥0 with (Y0 , Y 0 ) distributed
according to ζ0 and defined by (9). Byhdefinition of W
i 2 and since for all t ≥ 0, π
2
n
2
is invariant for Pt , W2 (µ0 Q , π) ≤ Eζ0 YΓn − Y Γn
. Lemma 20 with ǫ = κ/4,
Proposition 1-(ii) imply, using a straightforward induction, that for all n ≥ 0
Z
i
h
2
(1)
ky − xk2 π(dy) + An (γ) ,
(12)
≤ un (γ)
Eζ0 YΓn − Y Γn
Rd
7
(1)
where (un (γ))n≥1 is given by (10), and
def
An (γ) = L
2
n
X
i=1
γi2
−1
κ
+ γi (2d +
dL2 γi2 /6)
k=i+1
+ L4
n
X
i=1
with
n
Y
(1 − κγk /2)
δ̃i γi3 κ−1 + γi
n
Y
(1 − κγk /2) (13)
k=i+1
i
h
δ̃i = e−2mΓi−1 Eζ0 kY0 − x⋆ k2 + (1 − e−2mΓi−1 )(d/m) ≤ d/m .
Since Y0 is distributed according to π, Proposition 1-(ii) shows that for all i ∈ {1, · · · , n},
δ̃i ≤ d/m .
2
(14)
x⋆ k2 +kx⋆
2
In addition since
for all y ∈ Rd , kx − yk ≤ 2(kx −
− yk ), using Proposition 1R
(ii), we get Rd ky − xk2 π(dy) ≤ kx − x⋆ k2 + d/m. Plugging this result, (14) and (13)
in (12) completes the proof.
Corollary 6. Assume H 1 and H 2. Let (γk )k≥1 be a nonincreasing sequence with
γ1 ≤ 1/(m + L). Assume that limk→∞ γk = 0 and limn→+∞ Γn = +∞. Then for all
x ∈ Rd , limn→∞ W2 (δx Qnγ , π) = 0.
We preface the proof by a technical lemma which is established in the supplementary
document Section 2.1.
Lemma 7. Let (γk )k≥1 be a sequence of nonincreasing real numbers, ̟ > 0 and γ1 <
̟−1 . Then for all n ≥ 0, j ≥ 1 and ℓ ∈ {1, . . . , n + 1},
n+1
Y
X n+1
i=1 k=i+1
(1 −
̟γk ) γij
≤
n+1
Y
k=ℓ
(1 − ̟γk )
ℓ−1
X
γij +
i=1
(1)
γℓj−1
.
̟
(2)
Proof of Corollary 6. By Theorem 5, it suffices to show that un (γ) and un (γ), defined
by (10) and (11) respectively, goes to 0 as n → +∞. Using the bound 1 + t ≤ et
(1)
for t ∈ R, and limn→+∞ Γn = +∞, we have limn→+∞ un (γ) = 0. Since (γk )k≥0
(2)
is nonincreasing,
Pn Qnnote that to show2 that limn→+∞ un (γ) = 0, it suffices to prove
limn→+∞ i=1 k=i+1 (1 − κγk /2)γi = 0. But since (γk )k≥1 is nonincreasing, there
exists c ≥ 0 such that cΓn ≤ n − 1 and by Lemma 7 applied with ℓ = ⌊cΓn ⌋ the integer
part of cΓn :
n
n
Y
X
i=1 k=i+1
(1 −
κγk /2) γi2
−1
≤ 2κ
−1
γ⌊cΓn ⌋ + exp −2
⌊cΓn ⌋−1
X
γi .
κ(Γn − Γ⌊cΓn ⌋ )
(15)
i=1
Since limk→+∞ γk = 0, by the Cesáro theorem, limn→+∞ n−1 Γn = 0. Therefore since
limn→+∞ Γn = +∞, limn→+∞ (Γn )−1 Γ⌊cΓn ⌋ = 0, and the conclusion follows from comP⌊cΓn ⌋−1
γi ≤
bining in (15), this limit, limk→+∞ γk = 0, limn→+∞ Γn = +∞ and
i=1
cγ1 Γn .
8
In the case of constant step sizes γk = γ for all k ≥ 1, we can deduce from Theorem 5,
a bound between π and the stationary distribution πγ of Rγ .
Corollary 8. Assume H 1 and H 2. Let (γk )k≥1 be a constant sequence γk = γ for all
k ≥ 1 with γ ≤ 1/(m + L). Then
W22 (π, πγ ) ≤ 2κ−1 L2 γ κ−1 + γ (2d + dL2 γ/m + dL2 γ 2 /6) .
Proof. Since by Proposition 3, for all x ∈ Rd , (δx Rγn )n≥0 converges to πγ as n → ∞
in (P2 (Rd ), W2 ), the proof then follows from Theorem 5 and Lemma 7 applied with
ℓ = 1.
We can improve the bound provided by Theorem 5 under additional regularity assumptions on the potential U .
H 3. The potential U is three times continuously differentiable and there exists L̃ such
that for all x, y ∈ Rd , ∇2 U (x) − ∇2 U (y) ≤ L̃ kx − yk.
Note that under H1 and H3, we have that for all x, y ∈ Rd ,
~
∇2 U (x)y ≤ L kyk , ∆(∇U
)(x)
2
≤ d2 L̃2 .
(16)
Theorem 9. Assume H 1, H 2 and H 3. Let (γk )k≥1 be a nonincreasing sequence with
γ1 ≤ 1/(m + L). Then for all x ∈ Rd and n ≥ 1,
n
o
⋆ 2
W22 (δx Qnγ , π) ≤ u(1)
(γ)
kx
−
x
k
+
d/m
+ u(3)
n
n (γ) ,
(1)
where un is given by (10), κ in (4) and
(
"
!
)
n
X
dL̃2
4L4
2
−1
3
(3)
4 γi
4
−1
dγi 2L + κ
un (γ) =
+ γi L
+ γi L +
+m
3
3m
6
i=1
#
n
Y
κγk
1−
×
.
2
(17)
k=i+1
Proof. The proof of is along the same lines than Theorem 5, using Lemma 21 in place
of Lemma 20.
In the case of constant step sizes γk = γ for all k ≥ 1, we can deduce from Theorem 9,
a sharper bound between π and the stationary distribution πγ of Rγ .
Corollary 10. Assume H 1, H 2 and H 3. Let (γk )k≥1 be a constant sequence γk = γ
for all k ≥ 1 with γ ≤ 1/(m + L). Then
!
)
(
4
2
4L
d
L̃
+ γL4 (γ/6 + m−1 ) .
+ γL4 +
W22 (π, πγ ) ≤ 2κ−1 dγ 2 2L2 + κ−1
3
3m
9
Proof. The proof follows the same line as the proof of Corollary 8 and is omitted.
Using Proposition 3-(ii) and Corollary 6 or Corollary 10, given ε > 0, we determine the smallest number of iterations nε and an associated step-size γε satisfying
W2 (δx⋆ Rγnε , π) ≤ ε for all n ≥ nε . The dependencies on the dimension d and the
precision ε of nε based on Theorem 5 and Theorem 9 are reported in Table 1. Under
H 1 and H 2, the complexity we get is the same as the one established in [11] for the
total variation distance. However if in addition H 3 holds, we improve this complexity
with respect to the precision parameter ε as well. Finally, in the case where L̃ = 0 (for
example for non-degenerate d-dimensional Gaussian distributions), then the dependency
on the dimension of the number of iterations nε given by Theorem 9 turns out to be
only of order O(d1/2 log(d)).
Details and further discussions are included in the supplementary paper [12, Section 6-Section 7-Section 8]. In particular, the dependencies of the obtained bounds with
respect to the constants m and L which appear in H1, H2 are evidenced.
In a recent work [9] (based on a previous version of this paper), an improvement of
the proof of Theorem 5 has been proposed for constant step size. Whereas the constants
are sharper, the dependence with respect to the dimension and the precision parameter
ε > 0 is the same (first line of Table 1).
Parameter
Theorem 5 and Proposition 3-(ii)
Theorem 9 and Proposition 3-(ii)
d
O(d log(d))
O(d log(d))
ε
O(ε−2 |log(ε)|)
O(ε−1 |log(ε)|)
Table 1: Dependencies of the number of iterations nε to get W2 (δx⋆ Rγnεε , π) ≤ ε
For simplicity, consider sequences (γk )k≥1 defined for all k ≥ 1 by γk = γ1 k−α , for
(1)
(2)
γ1 < 1/(m + L) and α ∈ (0, 1]. Then for n ≥ 1, un ≤ O(e−κΓn /2 ), un ≤ dO(n−α ) and
(3)
un ≤ d2 O(n−2α ) (see [12, Sections 6 to 8] for details). Based on these upper bounds,
we can obtained the convergence rates provided by Theorem 5 and Theorem 9 for this
particular setting, see Table 2.
Theorem 5
Theorem 9
α ∈ (0, 1)
d O(n−α )
d2 O(n−2α )
α=1
d O(n−1 ) for γ1 > 2κ−1 see [12, Section 8]
d2 O(n−2 )
Table 2: Order of convergence of W22 (δx⋆ Qnγ , π) for γk = γ1 k−α
3
Quantitative bounds in total variation distance
We deal in this section with quantitative bounds in total variation distance. For Bayesian
inference application, total variation bounds are of importance for computing highest
posterior density (HPD) credible regions and intervals. For computing such bounds we
10
will use the results of Section 2 combined with the regularizing property of the semigroup
(Pt )t≥0 . Under H2, define χm for all t ≥ 0 by
p
χm (t) = (4/m)(e2mt − 1) .
The first key result consists in upper-bounding the total variation distance kµPt −νPt kTV
for µ, ν ∈ P1 (Rd ). To that purpose, we use the coupling by reflection; see [26] and [6,
Example 3.7]. This coupling is defined as the unique solution (Xt , Yt )t≥0 of the SDE:
(
√
dXt = −∇U (Xt )dt + 2dBtd
√
where et = e(Xt − Yt )
dYt = −∇U (Yt )dt + 2(Id −2et eTt )dBtd ,
with X0 = x, Y0 = y, e(z) = z/ kzk for z 6= 0 and e(0) = 0 otherwise. Define the
coupling time Tc = inf{s ≥ 0 R| Xs = Ys }. By construction Xt = Yt for t ≥ Tc . By
t
Levy’s characterization, B̃td = 0 (Id −2es eTs )dBsd is a d-dimensional Brownian motion,
therefore (Xt )t≥0 and (Yt )t≥0 are weak solutions to (1) started at x and y respectively.
Then by Lindvall’s inequality, for all t > 0 we have kPt (x, ·) − Pt (y, ·)kTV ≤ P (Xt 6= Yt ).
Theorem 11. Assume H1 and H2.
(i) For any x, y ∈ Rd and t > 0, it holds
kPt (x, ·) − Pt (y, ·)kTV ≤ 1 − 2Φ{− kx − yk /χm (t)} ,
where Φ is the cumulative distribution function of the standard normal distribution.
(ii) For any µ, ν ∈ P1 (Rd ) and t > 0,
kµPt − νPt kTV ≤ p
W1 (µ, ν)
(2π/m)(e2mt − 1)
.
(iii) For any x ∈ Rd and t ≥ 0,
(d/m)1/2 + kx − x⋆ k
kπ − δx Pt kTV ≤ p
.
(2π/m)(e2mt − 1)
(i) We now bound for t > 0, P (Xt 6= Yt ). For t < Tc , denoting by B1t =
T
d
0 1{s<Tc } es dBs , we get
√
d{Xt − Yt } = − {∇U (Xt ) − ∇U (Yt )} dt + 2 2et dB1t .
Proof.
Rt
By Itô’s formula and H2, we have for t < Tc ,
Z t
√
h∇U (Xs ) − ∇U (Ys ), es i ds + 2 2B1t
kXt − Yt k = kx − yk −
0
Z t
√
kXs − Ys k ds + 2 2B1t .
≤ kx − yk − m
0
11
By Grönwall’s inequality, we have
−mt
kXt − Yt k ≤ e
√ 1
√ Z t 1 −m(t−s)
kx − yk + 2 2Bt − m2 2
Bs e
ds .
0
Therefore by integration by part, kXt − Yt k ≤ Ut where (Ut )t∈(0,Tc ) is the one-dimensional
Ornstein-Uhlenbeck process defined by
Z 8t
√ Z t m(s−t) 1
−mt
−mt
Ut = e
kx − yk + 2 2
e
dBs = e
kx − yk +
em(s−t) dB̃s1
0
0
Therefore, for all x, y ∈ Rd and t ≥ 0, we get
P(Tc > t) ≤ P min Ut > 0 .
0≤s≤t
Finally the proof follows from [2, Formula 2.0.2, page 542].
(ii) Let µ, ν ∈ P1 (Rd ) and ξ ∈ Π(µ, ν) be an optimal transference plan for (µ, ν)
w.r.t. W1 . Since for all s > 0, 1/2 − Φ(−s) ≤ (2π)−1/2 s, (i) implies that for all x, y ∈ Rd
and t > 0,
Z
kx − yk
p
kµPt − νPt kTV ≤
dξ(x, y) ,
(2π/m)(e2mt − 1)
Rd ×Rd
which is the desired result.
(iii) The proof is a straightforward consequence of (ii) and Proposition 1-(iv).
Since for all s > 0, s ≤ es − 1, note that Theorem 11-(ii) implies that for all t > 0
and µ, ν ∈ P1 (Rd ),
kµPt − νPt kTV ≤ (4πt)−1/2 W1 (µ, ν) .
(18)
Therefore for all bounded measurable function f , Pt f is a Lipschitz function for all t > 0
with Lipshitz constant
kPt f kLip ≤ (4πt)−1/2 osc(f ) .
(19)
For all n, ℓ ≥ 1, n < ℓ and (γk )k≥1 a nonincreasing sequence denote by
ℓ
Y
−1
−1
Λn,ℓ (γ) = κ
(1 − κγj ) − 1 ,
Λℓ (γ) = Λ1,ℓ (γ) .
(20)
j=n
We will now study the contraction of Qn,ℓ
γ in total variation for non-increasing sequences
(γk )k≥1 . Strikingly, we are able to derive results which closely parallel Theorem 11. The
proof is nevertheless completely different because the reflection coupling is no longer
applicable in discrete time. We use a coupling construction inspired by the method of
[4, Section 3.3] for Gaussian random walks. This construction has been used in [13]
to establish convergence of homogeneous Markov chain in Wasserstein distances using
different method of proof. To avoid disrupting the stream of the proof, this construction
is introduced in Section 8.
12
Theorem 12. Assume H1 and H2.
(i) Let (γk )k≥1 be a nonincreasing sequence satisfying γ1 ≤ 2/(m + L). Then for all
x, y ∈ Rd and n, ℓ ∈ N∗ , n < ℓ, we have
n,ℓ
1/2
kδx Qn,ℓ
}.
γ − δy Qγ kTV ≤ 1 − 2Φ{− kx − yk /{8 Λn,ℓ (γ)}
(ii) Let (γk )k≥1 be a nonincreasing sequence satisfying γ1 ≤ 2/(m + L). For all µ, ν ∈
P1 (Rd ) and ℓ, n ∈ N∗ , n < ℓ,
n,ℓ
−1/2
kµQn,ℓ
W1 (µ, ν) .
γ − νQγ kTV ≤ {4πΛn,ℓ (γ)}
(iii) Let γ ∈ (0, 2/(m + L)]. Then for any x ∈ Rd and n ≥ 1,
o
n
kπγ − δx Rγn kTV ≤ {4πκ(1− (1− κγ)n/2 )}−1/2 (1− κγ)n/2 kx − x⋆ k + (2κ−1 d)1/2 .
Proof.
(i) By (3) for all x, y and k ≥ 1, we have
kx − γk ∇U (x) − y + γk ∇U (y)k ≤ (1 − κγk )1/2 kx − yk .
Let n, ℓ ≥ 1, n < ℓ, then applying Theorem 23 in Section 8, we get
n,ℓ
1/2
,
kδx Qn,ℓ
γ − δy Qγ kTV ≤ 1 − 2Φ − kx − yk /{8 Λn,ℓ (γ)}
(ii) Let f ∈ Fb (Rd ) and ℓ > n ≥ 1. For all x, y ∈ Rd by definition of the total
variation distance and (i), we have
n,ℓ
n,ℓ
n,ℓ
Qn,ℓ
γ f (x) − Qγ f (y) ≤ osc(f )kδx Qγ − δy Qγ kTV
o
n
,
≤ osc(f ) 1 − 2Φ − kx − yk /{8 Λn,ℓ (γ)}1/2
Using that for all s > 0, 1/2 − Φ(−s) ≤ (2π)−1/2 s concludes the proof.
(iii) The proof follows from (iii), the bound for all s > 0, 1/2 − Φ(−s) ≤ (2π)−1/2 s
and Proposition 2-(ii).
We can combine this result and Theorem 5 or Theorem 9 to get explicit bounds in
total variation between the Euler-Maruyama discretization and the target distribution
π. To that purpose, we use the following decomposition, for all nonincreasing sequence
(γk )k≥1 , initial point x ∈ Rd and ℓ ≥ 0:
kπ − δx Qℓγ kTV ≤ kπ − δx PΓℓ kTV + kδx PΓℓ − δx Qℓγ kTV .
13
(21)
The first term is dealt with Theorem 11-(iii). It remains to bound the second term
in (21). Since we will use Theorem 5 and Theorem 9, we have two different results
depending on the assumptions on U again. Define for all x ∈ Rd and n, p ∈ N,
2
ϑ(1)
n,p (x) = L
n
X
γi2
i=1
ϑ(2)
n,p (x)
=
n
X
γi3
i=1
n
Y
(1 − κγk /2)
k=i+1
n
Y
κ−1 + γi (2d + dL2 γi2 /6)
(22)
+L2 γi δi,n,p (x) κ−1 + γi
(1 − κγk /2) L4 δi,n,p (x)(4κ−1 /3 + γn+1 )
(23)
k=i+1
n
oi
2
+d 2L2 + 4κ−1 (dL̃2 /12 + γn+1 L4 /4) + γn+1
L4 /6
,
where ̺n,p (x) is given by (8) and
δi,n,p (x) = e−2mΓi−1 ̺n,p (x) + (1 − e−2mΓi−1 )(d/m) ,
Theorem 13. Assume H 1 and H 2. Let (γk )k≥1 be a nonincreasing sequence with
γ1 ≤ 1/(m + L). Then for all x ∈ Rd and ℓ, n ∈ N∗ , ℓ > n,
kδx PΓℓ − δx Qℓγ kTV ≤ (ϑn (x)/(4πΓn+1,ℓ ))1/2
+ 2−3/2 L
ℓ
X
(γk3 L2 /3)̺1,k−1 (x) + dγk2
k=n+1
!1/2
. (24)
(2)
where ̺1,n (x) is defined by (8), ϑn (x) is equal to ϑn,0 (x) given by (23), if H3 holds, and
(1)
to ϑn,0 (x) given by (22) otherwise.
Proof. Applying Lemma 20 or Lemma 21, we get that for all x ∈ Rd
( (1)
ϑn,0 (x) (H1, H2) ,
1/2
n
W1 (δx Qγ , δx PΓn ) ≤ {ϑn (x)}
, ϑn (x) =
(2)
ϑn,0 (x) (H1, H2, H3) ,
(25)
By the triangle inequality, we get
δx PΓℓ − δx Qℓγ
TV
≤
δx PΓn − δx Qnγ PΓn+1,ℓ
Using (18) and (25), we have
δx PΓn − δx Q1,n
PΓn+1,ℓ
γ
TV
n
o
+ δx Qnγ Qγn+1,ℓ − PΓn+1,ℓ
TV
14
≤ (ϑn (x)/(4πΓn+1,ℓ ))1/2 .
TV
. (26)
(27)
For the second term, by [10, Equation 11] (note that we have a different convention for
the total variation distance) and the Pinsker inequality, we have
n
o 2
n+1,ℓ
Q
−
P
δx Q1,n
Γ
γ
γ
n+1,ℓ
TV
−3
≤2
L
2
Z
ℓ
X
3
(γk /3)
k=n+1
⋆
Rd
2
+
dγk2
(γk3 L2 /3)̺1,k−1 (x) + dγk2
.
k∇U (z) − ∇U (x )k
Qγk−1 (x, dz)
.
By H1 and Proposition 2, we get
n
o
n+1,ℓ
Q
−
P
δx Q1,n
Γ
γ
γ
n+1,ℓ
2
TV
≤ 2−3 L2
ℓ
X
k=n+1
Combining the last inequality and (27) in (26) concludes the proof.
Consider the case of decreasing step sizes defined: γk = γ1 k−α for k ≥ 1 and α ∈
(0, 1]. Under H 1 and H 2 , choosing n = ℓ − ⌊ℓα ⌋ in the bound given by Theorem 13
and using Table 2 implies that kδx⋆ Qℓγ − πkTV = d1/2 O(ℓ−α/2 ). Note that for α = 1,
this rate is true only for γ1 > 2κ−1 . If in addition H 3 holds, choosing n = ℓ − ℓα/2
in the bound given by (24) and using Table 2, (21) and Theorem 11-(iii) implies that
kδx⋆ Qℓγ − πkTV = dO(ℓ−(3/4)α ). Note that these rates are explicit compared to the ones
obtained in [11, Proposition 4]. These conclusions and the dependency on the dimension
are summarized in Table 3.
H1, H2
γk = γ1
k−α ,
α ∈ (0, 1]
d1/2 O(ℓ−α/2 )
H1, H2 and H3
dO(ℓ−3α/4 )
Table 3: Order of convergence of kδx⋆ Qℓγ − πkTV for γk = γ1 k−α based on Theorem 13
−1
When γk = γ ∈ (0,
1/(m
+
L))
for
all
k
≥
1,
under
H
1
and
H
2,
for
ℓ
>
γ
−1
choosing n = ℓ − γ
implies that (see the supplementary document [12, Section 2.2])
1/2
kδx Rγℓ − δx Pℓγ kTV ≤ (4π)−1/2 γD1 (γ, d) + γ 3 D2 (γ)D3 (γ, d, x)
+ D4 (γ, d, x) , (28)
where
D1 (γ, d) = 2L2 κ−1 κ−1 + γ 2d + L2 γ 2 /6 , D2 (γ) = L4 κ−1 + γ
o
n
−1
D3 (γ, d, x) = (ℓ − γ −1 )e−mγ(ℓ−⌈γ ⌉−1) kx − x⋆ k2 + 2d(κγm)−1
D4 (γ, d, x) = 2−3/2 L [dγ(1 + γ)
n
oi1/2
−1
+(L2 γ 3 /3) (1 + γ −1 )(1 − κγ)ℓ−⌈γ ⌉ kx − x⋆ k2 + 2(1 + γ)κ−1 d
.
(29)
Using this bound and Theorem 11-(iii), the minimal number of iterations ℓε > 0 for a
target precision ε > 0 to get kδx⋆ Rγℓεε − πkTV ≤ ε is of order d log(d)O(|log(ε)| ε−2 ) (the
proper choice of the step size γε is given in Table 5). In addition, letting ℓ go to infinity
in (28) we get the following result.
15
Corollary 14. Assume H1 and H2. Let γ ∈ (0, 1/(m + L)]. Then it holds
1/2
kπγ − πkTV ≤ 2−3/2 L dγ(1 + γ) + 2(L2 γ 3 /3)(1 + γ)κ−1 d
1/2
+ (4π)−1/2 γD1 (γ, d) + 2dγ 2 D2 (γ)(κm)−1
,
where D1 (γ) and D2 (γ) are given in (29).
Note that the bound provided by Corollary 14 is of order d1/2 O(γ 1/2 ).
If in addition H3 holds, for constant step sizes, we can improve with respect to the
step size γ, the bounds given by Corollary 14.
Theorem 15. Assume H1, H2 and H3. Let γ ∈ (0, 1/(m + L)]. Then it holds
1/2
kπγ − πkTV ≤ (4π)−1/2 γ 2 E1 (γ, d) + 2dγ 2 E2 (γ)/(κm)
1/2
+ (4π)−1/2 log γ −1 / log(2) γ 2 E1 (γ, d) + γ 2 E2 (γ)(2κ−1 d + d/m)
+ 2−3/2 L 2dγ 3 L2 /(3κ) + dγ 2
1/2
,
where E1 (γ, d) and E2 (γ) are defined by
n
o
E1 (γ, d) = 2dκ−1 2L2 + 4κ−1 (dL̃2 /12 + γL4 /4) + γ 2 L4 /6
E2 (γ) = L4 (4κ−1 /3 + γ) .
Proof. The proof is postponed to Section 7.
Note that the bound provided by Theorem 15 is of order dO(γ |log(γ)|), improving
the dependency given by Corollary 14, with respect to the step size γ. Furthermore when
L̃ = 0, this bound given by Theorem 15 is of order d1/2 O(γ |log(γ)|) and is sharp up
to a logarithmic factor. Indeed, assume that π is the d-dimensional standard Gaussian
distribution. In such case, the ULA sequence (Xk )k≥0 is the autoregressive process
√
given for all k ≥ 0 by Xk+1 = (1 − γ)Xk + 2γZk+1 . For γ ∈ (0, 1), this sequence
has a stationary distribution πγ , which is a d-dimensional Gaussian distribution with
zero-mean and covariance matrix σγ2 Id , with σγ2 = (1 − γ/2)−1 . Therefore, using [22,
Lemma 4.9] (or the Pinsker inequality), we get the following upper bound: kπ −πγ kTV ≤
Cd1/2 |σγ2 − 1| = Cd1/2 γ/2, where C is a universal constant.
We can also for a precision target ε > 0 choose γε > 0 and the number of iterations
nε > 0 to get kδx Rγnεε −πkTV ≤ ε. By Theorem 11-(iii), Theorem 12-(iii) and Theorem 15,
the minimal number of iterations ℓε is of order d log2 (d)O(ε−1 log2 (ε)) for a well chosen
step size γε . This result improves the conclusion of [11] and 14 with respect to the
precision parameter ε, which provides an upper bound of the number of iterations of order
d log(d)O(ε−2 log2 (ε)). We can also compare our reported upper bound with the one
obtained for the d-dimensional standard Gaussian distribution. If the initial distribution
is the Dirac mass at zero (the minimum of the potential U (x) = kxk2 /2) and γ ∈ (0, 1),
the distribution of the ULA sequence after n iterations is zero-mean Gaussian with
16
covariance (1 − (1 − γ)2(n+1) )/(1 − γ/2) Id . If we use [22, Lemma 4.9] again, we get for
γ ∈ (0, 1),
kδ0 Rγn − πkTV ≤ Cd1/2 γ|1 − 2γ −1 (1 − γ)2(n+1) | ,
where C is a universal constant. To get an ε precision we need to choose γε = d−1/2 ε/(2C)
and then nε = ⌈(1/2) log(γε /4)/ log(1 − γε )⌉ = d1/2 log(d)O(ε−1 | log(ε)|). On the other
hand since L̃ = 0, based on the bound given by Theorem 15, the minimal number of
iterations to get kδx Rγnεε − πkTV ≤ ε is of order d1/2 log2 (d)O(ε−1 log2 (ε)). It follows that
our upper bound for the stepsize and the optimal number of iterations is again sharp
up to a logarithmic factor in the dimension and the precision. The discussions on the
bounds for constant sequences of step sizes are summarized in Table 4 and Table 5.
H1, H2
kπ − πγ kTV
d1/2 O(γ 1/2 )
H1, H2 and H3
dO(γ |log(γ)|)
Table 4: Order of the bound between π and πγ in total variation function of the step
size γ > 0 and the dimension d.
γε
nε
H1, H2 and H3
d−1 log−1 (d)O(ε log−1 (ε) )
d log2 (d)O(ε−1 log2 (ε))
H1, H2
d−1 O(ε2 )
d log(d)O(ε−2 |log(ε)|)
Table 5: Order of the step size γε > 0 and the number of iterations nε ∈ N∗ to get
kδx⋆ Rγnεε − πkTV ≤ ε for ε > 0.
4
Mean square error and concentration for bounded measurable functions
R
The result of the previous section allows us to study the approximation of Rd f (y)π(dy)
by the weighted average estimator π̂nN (f ) defined, for a measurable and bounded function
f : Rd → R, N, n ∈ N, n ≥ 1 by
π̂nN (f ) =
N
+n
X
N
f (Xk ) ,
ωk,n
N
= γk+1 Γ−1
ωk,n
N +2,N +n+1 .
(30)
k=N +1
We derive a bound on the mean-square error, defined as MSEN,n
= Ex
f
The
MSEN,n
f
h
π̂nN (f ) − π(f )
can be decomposed as the sum of the squared bias and variance:
MSEN,n
= Ex [π̂nN (f )] − π(f )
f
17
2
+ Varx π̂nN (f )
.
2
i
.
We first obtain a bound for the bias. By the Jensen inequality and because f is bounded,
we have:
N
+n
X
2
N
Ex [π̂nN (f )] − π(f ) ≤ osc(f )2
kδx Qkγ − πk2TV ;
ωk,n
k=N +1
Using the results of Section 3, we can deduce different bounds for the bias, depending
on the assumptions on U and the sequence of step sizes (γk )k≥1 . It is now required to
derive a bound for the variance. Our main tool is the Gaussian Poincaré inequality [3,
Theorem 3.20] which can be applied to Rγ (y, ·) defined by (6), noticing that Rγ (y, ·)
is a Gaussian distribution with mean y − γ∇U (y) and covariance matrix 2γ Id : for all
Lipschitz function g : Rd → R
Rγ {g(·) − Rγ g(y)}2 (y) ≤ 2γ kgk2Lip .
(31)
Lemma 16. Assume H 1 and H 2. Let (γk )k≥1 be a nonincreasing sequence with γ1 ≤
2/(m + L). Let g : Rd → R be a Lipschitz function. Then for all n, p ≥ 1, n ≤ p and
y ∈ Rd
Z
2
n,p
Qn,p
≤ 2κ−1 kgk2Lip ,
0≤
γ (y, dz) g(z) − Qγ g(y)
Rd
where
Qn,p
γ
is given by (7).
Proof. By decomposing g(Xp ) − EGy n [g(Xp )] =
and using EGy k [g(Xp )] = Qγk+1,p g(Xk ), we get
VarGy n
{g(Xp )} =
=
p
X
k=n+1
p
X
EGy n
EGy n
k=n+1
G
Ey k−1
Rγk
n
Pp
Gk
k=n+1 {Ey
G
[g(Xp )] − Ey k−1 [g(Xp )]},
2
Gk−1
Gk
Ey [g(Xp )] − Ey
[g(Xp )]
Qγk+1,p g(·)
−
o2
Rγk Qγk+1,p g(Xk−1 )
(Xk−1 ) .
P
Equation (31) implies VarGy n {g(Xp )} ≤ 2 pk=n+1 γk kQγk+1,p gk2Lip . The proof follows
from Corollary 4 and Lemma 7, using the bound (1 − t)1/2 ≤ 1 − t/2 for t ∈ [0, 1].
Theorem 17. Assume H 1 and H 2. Let (γk )k≥1 be a nonincreasing sequence with
γ1 ≤ 2/(m + L). Then for all N ≥ 0, n ≥ 1, x ∈ Rd and f ∈ Fb (Rd ), we get
(4)
Varx {π̂nN (f )} ≤ osc(f )2 {2γ1 Γ−1
N +2,N +n+1 + uN,n (γ)}
(4)
uN,n (γ) =
NX
+n−1
k=N
γk+1
( N +n
X
N
ωi,n
(πΛk+2,i (γ))1/2
i=k+2
for n1 , n2 ∈ N, Λn1 ,n2 (γ) is given by (20).
18
)2
+ κ−1
(
N
+n
X
i=N +1
N
ωi,n
(4πΛN +1,i (γ))1/2
)2
,
Proof. Let N ≥ 0, n ≥ 1, x ∈ Rd and f ∈ Fb (Rd ). We decompose π̂nN (f ) − Ex [π̂nN (f )] as
the sum of martingale increments, w.r.t. (Gn )n≥0 , the natural filtration associated with
Euler approximation (Xn )n≥0 , and we get
NX
+n−1
N
N 2
Gk+1 N
Gk
Varx π̂n (f ) =
Ex Ex
π̂n (f ) − Ex π̂n (f )
k=N
+ Ex
h
EGxN
π̂nN (f )
−
2
Ex [π̂nN (f )]
i
. (32)
G
The martingale increment Exk+1 π̂nN (f ) − EGxk π̂nN (f ) has a simple expression. For
k = N + n − 1, . . . , N + 1, define (backward in time) the function
N
N
ΦN
n,k : xk 7→ ωk,n f (xk ) + Rγk+1 Φn,k+1 (xk ) ,
(33)
N
N
where ΦN
n,N +n : xN +n 7→ Φn,N +n (xN +n ) = ωN +n,n f (xN +n ). Denote finally
N
ΨN
n : xN 7→ RγN+1 Φn,N +1 (xN ) .
(34)
Note that for k ∈ {N, . . . , N + n − 1}, by the Markov property,
Gk+1 N
N
π̂n (f ) − EGxk π̂nN (f ) ,
ΦN
(35)
n,k+1 (Xk+1 ) − Rγk+1 Φn,k+1 (Xk ) = Ex
GN π̂ N (f ) . With these notations, (32) may be equivalently expressed
and ΨN
n (XN ) = Ex
n
as
NX
+n−1
i
h
N
2
N
(X
)
(·)
−
R
Φ
(X
)
Ex Rγk+1 ΦN
Varx π̂n (f ) =
γ
k
k
n,k+1
k+1 n,k+1
k=N
+ Varx ΨN
n (XN )
.
(36)
Let k ∈ {N, . . . , N + n − 1}. We cannot directly apply the Poincaré inequality (31) since
the function ΦN
n,k , defined in (33), is not Lipschitz. However, Theorem 12-(ii) shows that
for all ℓ, n ∈ N∗ , n < ℓ, Qn,ℓ
γ f is a Lipschitz function with
Qn,ℓ
γ f
Lip
≤ osc(f )/{4πΛn,ℓ (γ)}1/2 .
N
N
N
Using (33), we may decompose ΦN
n,k = ωk+1,n f + Φ̃n,k , where Φ̃n,k =
which is Lipshitz with constant
Φ̃N
n,k
Lip
≤
N
+n
X
N
ωi,n
Qγk+2,i f
i=k+2
Using the inequality (a
+ b)2
≤
2a2
Lip
+
N
Rγk+1 ΦN
n,k+1 (·) − Rγk+1 Φn,k+1 (y)
≤ osc(f )
2b2 ,
N
+n
X
i=k+2
(37)
PN +n
k+2,i
N
f
i=k+2 ωi,n Qγ
N
ωi,n
/ {4πΛk+2,i (γ)}1/2 .
(38)
(31), we finally get for any y ∈ Rd
2
N
)2 osc(f )2
(y) ≤ 2(ωk+1,n
( N +n
)2
X
2
N
1/2
. (39)
+ γk+1 osc(f )
ωi,n /{πΛk+2,i (γ)}
i=k+2
19
3.5
3.0
3.0
Γn u0(4)
,n (γ)
Γn u0(4)
,n (γ)
3.5
(Γn u0(4)
,n (γ))n for γk =1/2 for all k ≥1
2.5
2.0
1.5
2.0
−α
(Γn u0(4)
,n (γ))n for γk =(1 + k)
1.5
1.00
1000
2000
3000
Iteration
n
4000
1.00
5000
3.5
3.5
3.0
3.0
2.5
2.0
−α
(Γn u0(4)
,n (γ))n for γk =(1 + k)
1.5
1.00
1000
2000
3000
Iteration
n
1000
2000
3000
Iteration
Γn u0(4)
,n (γ)
Γn u0(4)
,n (γ)
2.5
and
α =1/2
4000
α =1/4
4000
n
5000
2.5
2.0
−α
(Γn u0(4)
,n (γ))n for γk =(1 + k)
1.5
1.00
5000
and
1000
2000
3000
Iteration
n
and
α =3/4
4000
5000
(4)
Figure 1: Plots of (u0,n )n≥1 Γn for four sequences of step-sizes (γk )k≥1 , γk = (1 + k)−α
for α = 0, 1/4, 1/2, 3/4
N
N
It remains to control Varx ΨN
n (XN ) , where Ψn is defined in (34). Using (37), Ψn is
a Lipschitz function with Lipschitz constant bounded by:
ΨN
n Lip
≤
N
+n
X
i=N +1
N
ωi,n
QγN +1,i f Lip
≤ osc(f )
N
+n
X
N
ωi,n
/{4πΛN +1,i (γ)}1/2 .
(40)
i=N +1
By Lemma 16, we have the following result which is the counterpart of Corollary 3:
for all y ∈ Rd ,
−1
Vary ΨN
kf k2∞
n (XN ) ≤ 2κ
(
N
+n
X
N
ωi,n
/(πΛN +1,i )1/2
i=N +1
)2
.
(41)
Finally, the proof follows from combining (39) and (41) in (36).
(4)
To illustrate the result Theorem 17, we study numerically the behaviour (uN,n )n≥1
for κ = 1 N = 0, and four different non-increasing sequences of step-sizes (γk )k≥1 ,
γk = (1 + k)−α for α = 1/4, 1/2, 3/4 and γk = 1/2 for k ≥ 1. These results are
(4)
gathered in Figure 1, where it can be observed that (Γn u0,n (γ))n≥1 converges to a limit
as n → +∞.
We now establish an exponential deviation inequality for π̂nN (f ) − Ex [π̂nN (f )] given
by (30) for a bounded measurable function f .
20
Theorem 18. Assume H 1 and H 2. Let (γk )k≥1 be a nonincreasing sequence with
γ1 ≤ 2/(m + L). Let (Xn )n≥0 be given by (2) and started at x ∈ Rd . Then for all N ≥ 0,
n ≥ 1, r > 0, and functions f ∈ Fb (Rd ):
where
2 (5)
−1 2
Px π̂nN (f ) ≥ Ex [π̂nN (f )] + r ≤ e−{r−osc(f )(ΓN+2,N+n+1 ) } /{2osc(f ) uN,n (γ)} ,
(5)
uN,n (γ) =
NX
+n−1
k=N
γk+1
( N +n
X
N
ωi,n
(πΛk+2,i )1/2
i=k+2
)2
+ κ−1
(
N
+n
X
N
ωi,n
i=N +1
(πΛN +1,i )1/2
)2
.
Proof. The proof is postponed in the supplementary document to Section 3.
5
Numerical experiments
Consider a binary regression set-up in which the binary observations (responses) {Yi }pi=1
β T Xi )}pi=1 ,
are conditionally independent Bernoulli random variables with success probability {̺(β
p
where ̺ is the logistic function defined for z ∈ R by ̺(z) = ez /(1 + ez ) and {Xi }i=1 and
β are d dimensional vectors of known covariates and unknown regression coefficients,
respectively. The prior distribution for the parameter β is a zero-mean Gaussian distribution with covariance matrix Σβ . The density of the posterior distribution of β is up
to a proportionality constant given by
!
p n
o
X
p
−1 T −1
T
β T Xi
β |{(Xi , Yi )}i=1 ) ∝ exp
) − 2 β Σβ β .
Yiβ Xi − log(1 + e
πβ (β
i=1
Bayesian inference for the logistic regression model has long been recognized as a numerically involved problem, due to the analytically inconvenient form of the likelihood function. Several algorithms have been proposed, trying to mimick the data-augmentation
(DA) approach of [1] for probit regression; see [20], [15] and [16]. Recently, a very
promising DA algorithm has been proposed in [31], using the Polya-Gamma distribution
in the DA part. This algorithm has been shown to be uniformly ergodic for the total
variation by [7, Proposition 1], which provides an explicit expression for the ergodicity
constant. This constant is exponentially small in the dimension of the parameter space
and the number of samples. Moreover, the complexity of the augmentation step is cubic
in the dimension, which prevents from using this algorithm when the dimension of the
regressor is large.
We apply ULA to sample from the posterior distribution πβ (·|{(Xi , Yi )}pi=1 ). The
gradient of its log-density may be expressed as
β |{Xi , Yi }pi=1 )}
∇ log{πβ (β
p
X
=
Yi Xi −
i=1
21
Xi
T
1 + e−ββ Xi
− Σβ−1β ,
Therefore − log πβ (·|{XiP
, Yi }pi=1 ) is strongly convex H 2 with m = λ−1
max (Σβ ) and satisfies H 1 with L = (1/4) pi=1 XiT Xi + λ−1
(Σ
),
where
λ
(Σ
)
and
λmax (Σβ ) denote
min
β
β
min
the minimal and maximal eigenvalues of Σβ , respectively. We first compare the histograms produced by ULA and the Pòlya-Gamma Gibbs sampling from [31]. For that
purpose, we take d P
= 5, p = 100, generate synthetic data (Yi )1≤i≤p and (Xi )1≤i≤p , and
set Σβ−1 = (dp)−1 ( pi=1 XiT Xi ) Id . We produce 108 samples from the Pólya-Gamma
sampler using the R package BayesLogit [38]. Next, we make 103 runs of the Euler approximation scheme with n = 106 effective iterations, with a constant sequence (γk )k≥1 ,
γk = 10(κn1/2 )−1 for all k ≥ 0 and a burn-in period N = n1/2 . The histogram of
the Pólya-Gamma Gibbs sampler for first component, the corresponding mean of the
obtained histograms for ULA and the 0.95 quantiles are displayed in Figure 2. The
same procedure is also applied with the decreasing step size sequence (γk )k≥1 defined by
γk = γ1 k−1/2 , with γ1 = 10(κ log(n)1/2 )−1 and for the burn in period N = log(n), see also
Figure 2. In addition, we also compare MALA and ULA on five real data sets, which are
1.4
quantile at 95%
Mean of ULA
1.2
Polya−Gamma
1
quantile at 95%
Mean of ULA
Polya−Gamma
1.4
1.2
1
0.8
0.8
0.6
0.6
0.4
0.4
0.2
0.2
0.4
0.6
0.8
1
1.2
1.4
1.6
0.4
1.8
0.6
0.8
1
1.2
1.4
1.6
1.8
Figure 2: Empirical distribution comparison between the Polya-Gamma Gibbs Sampler
and ULA. Left panel: constant step size γk = γ1 for all k ≥ 1; right panel: decreasing
step size γk = γ1 k−1/2 for all k ≥ 1
summarized in Table 6. Note that for the Australian credit data set, the ordinal covariates have been stratified by dummy variables. Furthermore, we normalized the
Ppdata sets
−1
T
where
Σ
=
p
and consider the Zellner prior setting Σ−1 = (π2 d/3)Σ−1
X
i=1 Xi Xi
X
; see [34], [19] and the references therein. Also, we apply a pre-conditioned version of
1/2
MALA and ULA, targeting the probability density π̃β (·) ∝ πβ (ΣX ·). Then, we ob1/2
tain samples from πβ by post-multiplying the obtained draws by ΣX . We compare
MALA and ULA for each data sets by estimating for each component i ∈ {1, . . . , d} the
marginal accuracy between their d marginal empirical distributions and the d marginal
posterior distributions, where the marginal accuracy between two probability measure
µ, ν on (R, B(R)) is defined by
MA(µ, ν) = 1 − (1/2)kµ − νkTV .
This quantity has already been considered in [5] and [8] to compare approximate samplers. To estimate the d marginal posterior distributions, we run 2 · 107 iterations of
the Polya-Gamma Gibbs sampler. Then 100 runs of MALA and ULA (106 iterations
22
❵❵❵
❵❵❵
❵❵Dimensions
Observations p
❵❵❵
Data set
❵❵❵
German credit 1
Heart disease 2
Australian credit3
Pima indian diabetes4
Musk5
Covariates d
1000
270
690
768
476
25
14
35
9
167
Table 6: Dimension of the data sets
0.98
0.985
0.97
0.98
0.975
0.96
0.95
ULA
MALA
ULA
MALA
ULA
MALA
0.98
0.98
0.97
0.97
0.96
0.96
0.95
ULA
MALA
0.95
0.98
0.96
0.94
0.92
ULA
MALA
Figure 3: Marginal accuracy across all the dimensions.
Upper left: German credit data set. Upper right: Australian credit data set. Lower left:
Heart disease data set. Lower right: Pima Indian diabetes data set. At the bottom:
Musk data set
per run) have been performed. For MALA, the step-size is chosen so that the acceptance probability at stationarity is approximately equal to 0.5 for all the data sets. For
ULA, we choose the same constant step-size than MALA. We display the boxplots of the
mean of the estimated marginal accuracy across all the dimensions in Figure 3. These
results all imply that ULA is an alternative to the Polya-Gibbs sampler and the MALA
algorithm.
1
http://archive.ics.uci.edu/ml/datasets/Statlog+(German+Credit+Data)
http://archive.ics.uci.edu/ml/datasets/Statlog+(Heart)
3
http://archive.ics.uci.edu/ml/datasets/Statlog+(Australian+Credit+Approval)
4
http://archive.ics.uci.edu/ml/datasets/Pima+Indians+Diabetes
5
https://archive.ics.uci.edu/ml/datasets/Musk+(Version+1)
2
23
6
6.1
Proofs
Proof of Theorem 5
We preface the proof by two technical Lemmata.
Lemma 19. Let (Yt )t≥0 be the solution of (1) started at x ∈ Rd . For all t ≥ 0 and
x ∈ Rd ,
i
h
Ex kYt − xk2 ≤ dt(2 + L2 t2 /3) + (3/2)t2 L2 kx − x⋆ k2 .
Proof. Let A be the generator associated with (Pt )t≥0 defined by (S2).
R t Denote for all
x, y ∈ Rd , Ṽx (y) = ky − xk2 . Note that the process (Ṽx (Yt ) − Ṽx (x) − 0 A Ṽx (Ys )ds)t≥0 ,
is a (Ft )t≥0 -martingale under Px . Denote for all t ≥ 0 and x ∈ Rd by ṽ(t, x) = Pt Ṽx (x).
Then we get,
∂ṽ(t, x)
= Pt A Ṽx (x) .
(42)
∂t
By H2, we have for all y ∈ Rd , h∇U (y) − ∇U (x), y − xi ≥ m kx − yk2 , which implies
A Ṽx (y) = 2 (− h∇U (y), y − xi + d) ≤ 2 −mṼx (y) + d − h∇U (x), y − xi .
Using (42), this inequality and that Ṽx is positive, we get
Z
∂ṽ(t, x)
h∇U (x), y − xi Pt (x, dy) .
= Pt A Ṽx (x) ≤ 2 d −
∂t
Rd
(43)
By the Cauchy-Schwarz inequality, ∇U (x⋆ ) = 0, (1) and the Jensen inequality, we have,
|Ex [h∇U (x), Yt − xi]| ≤ k∇U (x)k kEx [Yt − x]k
Z t
⋆
≤ k∇U (x)k Ex
{∇U (Ys ) − ∇U (x )} ds
0
√
≤ t k∇U (x) − ∇U (x⋆ )k
Z
h
t
⋆
Ex k∇U (Ys ) − ∇U (x )k
0
2
i
ds
1/2
.
Furthermore, by H1 and Proposition 1-(i), we have
Z
Rd
h∇U (x), y − xi Pt (x, dy) ≤
≤
√
2
⋆
tL kx − x k
√
2
⋆
tL kx − x k
Z
0
t
h
⋆ 2
Ex kYs − x k
i
ds
1/2
1/2
1 − e−2mt
2tm + e−2mt − 1
⋆ 2
kx − x k +
(d/m)
2m
2m
≤ L2 kx − x⋆ k (t kx − x⋆ k + t3/2 d1/2 ) ,
where we used for the last line that by the Taylor theorem with remainder term, for all
s ≥ 0, (1 − e−2ms )/(2m) ≤ s and (2ms + e−2ms − 1)/(2m) ≤ ms2 , and the inequality
24
√
√
a + b ≤ a + b. Plugging this upper bound in (43), and since 2 kx − x⋆ k t3/2 d1/2 ≤
t kx − x⋆ k2 + t2 d, we get
√
∂ṽ(t, x)
≤ 2d + 3L2 t kx − x⋆ k2 + L2 t2 d
∂t
Since ṽ(0, x) = 0, the proof is completed by integrating this result.
Let (Ft )t≥0 be the filtration associated with (Bt )t≥0 and (Y0 , Y 0 ).
Lemma 20. Assume H 1 and H 2. Let (γk )k≥1 be a nonincreasing sequence with γ1 ≤
1/(m + L). Let ζ0 ∈ P2 (Rd × Rd ), (Yt , Y t )t≥0 such that (Y0 , Y 0 ) is distributed according
to ζ0 and given by (9). Then almost surely for all n ≥ 0 and ǫ > 0,
YΓn+1 − Y Γn+1
2
2
≤ {1 − γn+1 (κ − 2ǫ)} YΓn − Y Γn
Z Γn+1
−1
k∇U (Ys ) − ∇U (YΓn )k2 ds ,
+ (2γn+1 + (2ǫ) )
Γn
i
h
2
2
FΓn
≤ {1 − γn+1 (κ − 2ǫ)} YΓn − Y Γn
E
YΓn+1 − Y Γn+1
2
2
+ L2 γn+1
(1/(4ǫ) + γn+1 ) 2d + L2 γn+1 kYΓn − x⋆ k2 + dL2 γn+1
/6 .
(44)
(45)
Proof. Let n ≥ 0 and ǫ > 0, and set Θn = YΓn − Y Γn . We first show (44). By definition
we have:
2
2
kΘn+1 k = kΘn k +
Z
Γn+1
Γn
2
∇U (Ys ) − ∇U (Y Γn ) ds
− 2γn+1 Θn , ∇U (YΓn ) − ∇U (Y Γn ) − 2
Z
Γn+1
hΘn , {∇U (Ys ) − ∇U (YΓn )}i ds . (46)
Γn
Young’s inequality and Jensen’s inequality imply
Z
Γn+1
Γn
2
∇U (Ys ) − ∇U (Y Γn ) ds
2
≤ 2γn+1
∇U (YΓn ) − ∇U (Y Γn )
+ 2γn+1
Z
Γn+1
Γn
2
k∇U (Ys ) − ∇U (YΓn )k2 ds .
Using (3), γ1 ≤ 1/(m + L) and (γk )k≥1 is nonincreasing, (46) becomes
2
2
kΘn+1 k ≤ {1 − γn+1 κ} kΘn k + 2γn+1
−2
Z
Z
Γn+1
Γn
k∇U (Ys ) − ∇U (YΓn )k2 ds
Γn+1
hΘn , {∇U (Ys ) − ∇U (YΓn )}i ds .
Γn
Using the inequality | ha, bi | ≤ ǫkak2 + (4ǫ)−1 kbk2 concludes the proof of (44).
25
(47)
We now prove (45). Note that (44) implies that
i
h
EFΓn kΘn+1 k2 ≤ {1 − γn+1 (κ − 2ǫ)} kΘn k2
Z Γn+1
i
h
−1
EFΓn k∇U (Ys ) − ∇U (YΓn )k2 ds . (48)
+ (2γn+1 + (2ǫ) )
Γn
By H1, the Markov property of (Yt )t≥0 and Lemma 19, we have
Z Γn+1
i
h
EFΓn k∇U (Ys ) − ∇U (YΓn )k2 ds
Γn
2
4
3
+ dL2 γn+1
/12 + (1/2)L2 γn+1
kYΓn − x⋆ k2 .
≤ L2 dγn+1
The proof is then concluded plugging this bound in (48) .
6.2
Proofs of Theorem 9
Lemma 21. Assume H 1, H 2 and H 3. Let (γk )k≥1 be a nonincreasing sequence with
γ1 ≤ 1/(m + L). and ζ0 ∈ P2 (Rd × Rd ). Let (Yt , Y t )t≥0 be defined by (9) such that
(Y0 , Y 0 ) is distributed according to ζ0 . Then for all n ≥ 0 and ǫ > 0, almost surely
h
i
2
2
EFΓn YΓn+1 − Y Γn+1
≤ {1 − γn+1 (κ − 2ǫ)} YΓn − Y Γn
n
2
3
d(2L2 + ǫ−1 (dL̃2 /12 + γn+1 L4 /4) + γn+1
L4 /6)
+ γn+1
o
+L4 (ǫ−1 /3 + γn+1 ) kYΓn − x⋆ k2 .
Proof. Let n ≥ 0 and ǫ > 0, and set Θn = YΓn − Y Γn . Using Itô’s formula, we have for
all s ∈ [Γn , Γn+1 ),
Z sn
o
~
∇2 U (Yu )∇U (Yu ) + ∆(∇U
)(Yu ) du
∇U (Ys ) − ∇U (YΓn ) =
Γn
√ Z s 2
∇ U (Yu )dBu .
(49)
+ 2
Γn
Rs
Since Θn is FΓn -measurable and ( 0 ∇2 U (Yu )dBu )s∈[0,Γn+1 ] is a (Fs )s∈[0,Γn+1 ] -martingale
under H1, by (49) we have:
EFΓn [hΘn , ∇U (Ys ) − ∇U (YΓn )i]
Z
FΓn
= Θn , E
s
Γn
n
o
~
∇ U (Yu )∇U (Yu ) + ∆(∇U )(Yu ) du
2
Combining this equality and | ha, bi | ≤ ǫkak2 + (4ǫ)−1 kbk2 in (47) we have
i
h
EFΓn kΘn+1 k2 ≤ {1 − γn+1 (κ − 2ǫ)} kΘn k2 + +(2ǫ)−1 A
Z Γn+1
2γn+1 EFΓn
k∇U (Ys ) − ∇U (YΓn )k2 ds ,
Γn
26
(50)
where
A=
Z
Γn+1
E
Z
FΓn
Γn
s
~
∇ U (Yu )∇U (Yu ) + (1/2)∆(∇U )(Yu )du
2
Γn
2
ds .
We now separately bound the two last terms of the right hand side. By H1, the Markov
property of (Yt )t≥0 and Lemma 19, we have
Z Γn+1
i
h
EFΓn k∇U (Ys ) − ∇U (YΓn )k2 ds
Γn
2
4
3
+ dL2 γn+1
/12 + (1/2)L2 γn+1
kYΓn − x⋆ k2 . (51)
≤ L2 dγn+1
We now bound A. We get using Jensen’s inequality, Fubini’s theorem, ∇U (x⋆ ) = 0 and
(16)
Z s
Z Γn+1
h
i
2
du ds
EFΓn ∇2 U (Yu )∇U (Yu )
(s − Γn )
A≤2
Γn
−1
+2
≤2
Z
Γn+1
Γn
Z
Γn
Γn+1
Γn
(s − Γn )L
4
Z
(s − Γn )
s
Γn
Z
s
E
FΓn
Γn
~
∆(∇U
)(Yu )
2
du ds
h
i
3
EFΓn kYu − x⋆ k2 du ds + γn+1
d2 L̃2 /6 .
(52)
By Lemma 19-(i), the Markov property and for all t ≥ 0, 1 − e−t ≤ t, we have for all
s ∈ [Γn , Γn+1 ],
Z s
i
h
EFΓn kYu − x⋆ k2 du ≤ (2m)−1 (1 − e−2m(s−Γn ) ) kYΓn − x⋆ k2 + d(s − Γn )2 .
Γn
Using this inequality in (52) and for all t ≥ 0, 1 − e−t ≤ t , we get
3
4
3
A ≤ (2L4 γn+1
/3) kYΓn − x⋆ k2 + L4 dγn+1
/2 + γn+1
d2 L̃2 /6 .
Combining this bound and (51) in (50) concludes the proof.
7
Proof of Theorem 15
We preface the proof by a preliminary lemma. Define for all γ > 0, the function n :
R∗+ → N by
n(γ) = log γ −1 / log(2) .
(53)
Lemma 22. Assume H 1, H 2 and H 3. Let γ ∈ (0, 1/(m + L)). Then for all x ∈ Rd
and ℓ ∈ N∗ , ℓ > 2n(γ) ,
(2)
kδx Pℓγ − δx Rγℓ kTV ≤ (ϑℓ−2n(γ) ,0 (x)/(π2n(γ)+2 γ))1/2
−3/2
+2
3
2
L (γ L /3)̺1,ℓ−1 (x) + dγ
2 1/2
+
n(γ)
X
k=1
27
(2)
(ϑ2k−1 ,ℓ−2k (x)/(π2k+1 γ))1/2 .
(2)
where ̺1,ℓ−1 (x) is defined by (8) and for all n1 , n2 ∈ N, ϑn1 ,n2 is given by (23).
Proof. Let γ ∈ (0, 1/(m + L)) and ℓ ∈ N∗ . For ease of notation, let n = n(γ), and
assume that ℓ > 2n . Consider the following decomposition
o
n
n
δx Pℓγ − δx Rγℓ
δx P(ℓ−2n )γ − δx Rγℓ−2 P2n γ
≤
TV
TV
+ δx Rγℓ−1 {Pγ − Rγ }
TV
+
n
X
ℓ−2k
δx Rγ
k=1
n
k−1
P2k−1 γ − Rγ2
o
P2k−1 γ
TV
. (54)
We bound each term in the right hand side. First by (18) and Equation (25), we have
o
n
n
(2)
≤ (ϑℓ−2n ,0 (x)/(π2n+1 γ))1/2 ,
(55)
δx P(ℓ−2n )γ − δx Rγℓ−2 P2n γ
TV
(2)
where ϑn,0 (x) is given by (23). Similarly but using in addition Proposition 2, we have
for all k ∈ {1, · · · , n},
o
n
k−1
k
(2)
P2k−1 γ
≤ (ϑ2k−1 ,ℓ−2k (x)/(π2k+1 γ))1/2 ,
(56)
δx Rγℓ−2 P2k−1 γ − Rγ2
TV
(2)
where ϑ2k−1 ,ℓ−2k (x) is given by (23). For the last term, by [10, Equation 11] and the
Pinsker inequality, we have
Z
2
≤ 2−3 L2 (γ 3 /3)
k∇U (z)k2 Rγℓ−1 (x, dz) + dγ 2 .
δx Rγℓ−1 {Pγ − Rγ }
TV
Rd
By H1 and Proposition 2, we get
δx Rγℓ−1 {Rγ − Pγ }
2
TV
≤ 2−3 L2 (γ 3 L2 /3)̺1,ℓ−1 (x) + dγ 2
.
(57)
Combining (55), (56) and (57) in (54) concludes the proof.
Proof of Theorem 15. By Lemma 22, we get (see the supplementary document [12, Section 7] for the detailed calculation):
o1/2
n
n
o
kδx Pℓγ − δx Rγℓ kTV ≤ 2−3/2 L (γ 3 L2 /3) (1 − κγ)ℓ−1 kx − x⋆ k2 + 2κ−1 d + dγ 2
1/2
+ (4π)−1/2 γ 2 E1 (γ, d) + γ 3 E2 (γ)E3 (γ, d, x)
1/2
+ n(γ)(4π)−1/2 γ 2 E1 (γ, d) + γ 2 E2 (γ)E4 (γ, d, x)
, (58)
where
E3 (γ, d, x) = (ℓ − γ −1 )e−mγ(ℓ−2γ
E4 (γ, d, x) = e−mγ(ℓ−2γ
−1 −1)
−1 −1)
kx − x⋆ k2 + 2d/(κγm)
kx − x⋆ k2 + 2κ−1 d + d/m .
Letting ℓ go to infinity, using Theorem 11-(iii) and Theorem 12-(iii), we get the desired
conclusion.
28
8
Contraction in total variation for functional autoregressive models
In this section, we consider functional autoregressive models of the form: for all k ≥ 0
Xk+1 = hk+1 (Xk ) + σk+1 Zk+1 ,
(59)
where (Zk )k≥1 is a sequence of i.i.d. d dimensional standard Gaussian random variables,
(σk )k≥1 is a sequence of positive real numbers and (hk )k≥1 are a sequence of measurable
map from Rd to Rd which satisfies the following assumption:
AR 1. For all k ≥ 1, there exists ̟k ∈ R such that hk is 1 − ̟k -Lipschitz, i.e. for all
x, y ∈ Rd , khk (x) − hk (y)k ≤ (1 − ̟k ) kx − yk.
The sequence {Xk , k ∈ N} defines an inhomogeneous Markov chain associated with
the sequence of Markov kernel (Pk )k≥1 on (Rd , B(Rd )) given for all x ∈ Rd and A ∈ Rd
by
Z
1
2
2
Pk (x, A) =
)
dy .
(60)
exp
−
ky
−
h
(x)k
/(2σ
k
k
(2πσk2 )d/2 A
We denote for all n ≥ 1 by Qn the marginal laws of the sequence (Xk )k≥1 and given by
Q n = P1 · · · Pn .
(61)
In this section we compute an upper bound of {kδx Qn − δy Qn kTV , n ∈ 1} which does
not depend on the dimension d.
Let x, y ∈ Rd . We consider for all k ≥ 1 the following coupling (X1 , Y1 ) between
Pk (x, ·) and Pk (y, ·). Define the function E and e from Rd × Rd to Rd by
(
Ek (x, y)/ kEk (x, y)k if Ek (x, y) 6= 0
(62)
Ek (x, y) = hk (y) − hk (x) , ek (x, y) =
0
otherwise ,
For all x, y, z ∈ Rd , x 6= y, define
Fk (x, y, z) = hk (y) + Id −2ek (x, y)ek (x, y)T z
ϕσ2 (kEk (x, y)k − hek (x, y), zi)
,
αk (x, y, z) = k
ϕ σ2 (hek (x, y), zi)
(63)
(64)
k
where ϕ σ2 is the probability density of a zero-mean gaussian variable with variance σk2 .
k
Let Z1 be a standard d-dimensional Gaussian random variable. Set X1 = hk (x) + σk Z1
and
(
hk (y) + σk Z1
if Ek (x, y) = 0
Y1 =
B1 X1 + (1 − B1 ) Fk (x, y, Z1 ) if Ek (x, y) 6= 0 ,
where B1 is a Bernoulli random variable independent of Z1 with success probability
pk (x, y, z) = 1 ∧ αk (x, y, z) .
29
The construction above defines for all (x, y) ∈ Rd × Rd the Markov kernel Kk on (Rd ×
Rd , B(Rd ) ⊗ B(Rd )) given for all (x, y) ∈ Rd × Rd and A ∈ B(Rd ) ⊗ B(Rd ) by
Z
1D (hk (x), hk (y))
−kτk (x̃,x)k2 /(2σk2 )
Kk ((x, y), A) =
1
dx̃
(65)
(x̃,
x̃)e
A
(2πσk2 )d/2
Rd
Z
2
1Dc (hk (x), hk (y))
2
1A (x̃, x̃)pk (x, y, τk (x̃, x))e−kτk (x̃,x)k /(2σk ) dx̃
+
2
d/2
(2πσk )
Rd
Z
−kτk (x̃,x)k2 /(2σk2 )
+
1A (x̃, Fk (x, y, τk (x̃, x))) {1 − pk (x, y, τk (x̃, x))} e
dx̃ ,
Rd
where for all x̃ ∈ Rd , τk (x̃, x) = x̃ − hk (x) and D = (x̃, ỹ) ∈ Rd × Rd x̃ = ỹ . It is
shown in [4, Section 3.3] that for all x, y ∈ Rd and k ≥ 1, Kk ((x, y), ·) is a transference
plan of Pk (x, ·) and Pk (y, ·). Furthermore, we have for all x, y ∈ Rd and k ≥ 1
kEk (x, y)k
Kk ((x, y), D) = 2Φ −
.
(66)
2σk
eµ and E
e µ denote the
For all initial distribution µ0 on (Rd × Rd , B(Rd ) ⊗ B(Rd )), P
0
0
probability and the expectation respectively, associated with the sequence of Markov
kernels (Kk )k≥1 defined in (65) and µ0 on the canonical space ((Rd × Rd )N , (B(Rd ) ⊗
B(Rd ))⊗N ), {(Xi , Yi ), i ∈ N} denotes the canonical process and {Fei , i ∈ N} the corresponding filtration. Then if (X0 , Y0 ) = (x, y) ∈ Rd × Rd , for all k ≥ 1 (Xk , Yk ) is
coupling of δx Qk and δy Qk . Using Lindvall inequality, bounding kδx Qn − δy Qn kTV for
e(x,y) (Xn 6= Yn ). It is the content of the proof of the main
n ≥ 1 amounts to evaluate P
result of this section.
Theorem 23. Assume AR1. Then for all x, y ∈ Rd and n ≥ 1,
(
!)
kx
−
yk
kδx Qn − δy Qn kTV ≤ 1Dc ((x, y)) 1 − 2Φ −
,
1/2
2Ξn
where (Ξi )i≥1 is defined for all k ≥ 1 by Ξk =
Pk
Qi
2
i=1 σi {
We preface the proof by a technical Lemma.
j=1 (1
− ̟j )−2 }.
Lemma 24. For all ς, a > 0 and t ∈ R+ , the following identity holds
Z
ϕ 2 (t − y)
ϕ ς 2 (y) 1 − 1 ∧ ς
ϕ ς 2 (y)
R
|2y − t|
1 − 2Φ −
dy
2a
t
= 1 − 2Φ − 2
2(ς + a2 )1/2
Proof. The proof is postponed to [12, Section 4.2].
30
.
Proof of Theorem 23. Since for all k ≥ 1, (Xk , Yk ) is a coupling of δx Qk and δy Qk ,
e (x,y) (Xk 6= Yk ).
kδx Qk − δy Qk kTV ≤ P
P 2
Q
Define for all k1 , k2 ∈ N∗ , k1 ≤ k2 , Ξk1 ,k2 = ki=k
σ 2 { ij=k1 (1 − ̟j )−2 }. Let n ≥ 1.
1 i
We show by backward induction that for all k ∈ {0, · · · , n − 1},
e(x,y) (Xn 6= Yn )
P
e (x,y)
≤E
"
"
(
1Dc (Xk , Yk ) 1 − 2Φ −
kXk − Yk k
2 (Ξk+1,n )1/2
)##
, (67)
Note that the inequality for k = 0 will conclude the proof.
Since Xn 6= Yn implies that Xn−1 6= Yn−1 , the Markov property and (66) imply
e(x,y) (Xn 6= Yn ) = E
e (x,y)
P
h
i
e (X ,Y ) [1Dc (X1 , Y1 )]
1Dc (Xn−1 , Yn−1 )E
n−1
n−1
e (x,y)
≤E
kEn−1 (Xn−1 , Yn−1 )k
1Dc (Xn−1 , Yn−1 ) 1 − 2Φ −
2σn
Using AR1 and (62), kEn (Xn−1 , Yn−1 )k ≤ (1 − ̟n ) kXn−1 − Yn−1 k, showing (67) holds
for k = n − 1.
Assume that (67) holds for k ∈ {1, . . . , n − 1}. On {Xk 6= Yk }, we have
kXk − Yk k = − kEk (Xk−1 , Yk−1 )k + 2σk ek (Xk−1 , Yk−1 )T Zk ,
which implies
kX − Y k
1Dc (Xk , Yk ) 1 − 2Φ − k 1/2 k
2Ξk+1,n
2σ e (X , Y )T Z − kE (X , Y )k
k k
k−1
k−1
k
k
k−1
k−1
.
= 1Dc (Xk , Yk ) 1 − 2Φ −
1/2
2Ξ
k+1,n
Since Zk is independent of Fek−1 , σk ek (Xk−1 , Yk−1 )T Zk is a real Gaussian random variable
with zero mean and variance σk2 , therefore by Lemma 24, we get
e Fek−1 1Dc (Xk , Yk ) 1 − 2Φ − kXk − Yk k
E
(x,y)
1/2
2Ξk+1,n
"
(
kEk (Xk−1 , Yk−1 )k
≤ 1Dc (Xk−1 , Yk−1 ) 1 − 2Φ −
1/2
2 σk2 + Ξk+1,n
)#
.
Using by AR1 that kEk (Xk−1 , Yk−1 )k ≤ (1 − ̟k ) kXk−1 − Yk−1 k concludes the induction.
31
Acknowledgements
The authors would like to thank Arnak Dalalyan for helpful discussions. The work
of A.D. and E.M. is supported by the Agence Nationale de la Recherche, under grant
ANR-14-CE23-0012 (COSMOS).
References
[1] J. H. Albert and S. Chib. Bayesian analysis of binary and polychotomous response
data. Journal of the American Statistical Association, 88(422):669–679, 1993.
[2] A. N. Borodin and P. Salminen. Handbook of Brownian motion—facts and formulae.
Probability and its Applications. Birkhäuser Verlag, Basel, second edition, 2002.
[3] S. Boucheron, G. Lugosi, and P. Massart. Concentration inequalities. Oxford University Press, Oxford, 2013. A nonasymptotic theory of independence, With a
foreword by Michel Ledoux.
[4] R. Bubley, M. Dyer, and M. Jerrum. An elementary analysis of a procedure for
sampling points in a convex body. Random Structures Algorithms, 12(3):213–235,
1998.
[5] Faes C., Ormerod J. T., and Wand M. P. Variational bayesian inference for parametric and nonparametric regression with missing data. Journal of the American
Statistical Association, 106(495):959–971, 2011.
[6] M. F. Chen and S. F. Li. Coupling methods for multidimensional diffusion processes.
Ann. Probab., 17(1):151–177, 1989.
[7] H. M. Choi and J. P. Hobert. The Polya-Gamma Gibbs sampler for Bayesian logistic
regression is uniformly ergodic. Electron. J. Statist., 7:2054–2064, 2013.
[8] N. Chopin and Ridgway J. Leave pima indians alone: binary regression as a benchmark for bayesian computation. Technical Report 0864, arXiv, une 2015.
[9] A. S. Dalalyan. Further and stronger analogy between sampling and optimization:
Langevin monte carlo and gradient descent. In Proceedings of the 30th Annual
Conference on Learning Theory.
[10] A. S. Dalalyan. Theoretical guarantees for approximate sampling from smooth and
log-concave densities. Journal of the Royal Statistical Society: Series B (Statistical
Methodology), pages n/a–n/a, 2016.
[11] A. Durmus and É. Moulines. Non-asymptotic convergence analysis for the unadjusted langevin algorithm. Accepted for publication in Ann. Appl. Probab.
[12] A. Durmus and É. Moulines. Supplement to “high-dimensional bayesian inference
via the unadjusted langevin algorithm”, 2015. https://hal.inria.fr/hal-01176084/.
32
[13] A. Eberle. Quantitative contraction rates for Markov chains on continuous state
spaces. In preparation.
[14] D. L Ermak. A computer simulation of charged particles in solution. i. technique
and equilibrium properties. The Journal of Chemical Physics, 62(10):4189–4196,
1975.
[15] S. Frühwirth-Schnatter and R. Frühwirth. Data augmentation and MCMC for
binary and multinomial logit models statistical modelling and regression structures.
In Thomas Kneib and Gerhard Tutz, editors, Statistical Modelling and Regression
Structures, chapter 7, pages 111–132. Physica-Verlag HD, Heidelberg, 2010.
[16] R. B. Gramacy and N. G. Polson. Simulation-based regularized logistic regression.
Bayesian Anal., 7(3):567–590, 09 2012.
[17] U. Grenander. Tutorial in pattern theory. Division of Applied Mathematics, Brown
University, Providence, 1983.
[18] U. Grenander and M. I. Miller. Representations of knowledge in complex systems.
J. Roy. Statist. Soc. Ser. B, 56(4):549–603, 1994. With discussion and a reply by
the authors.
[19] T. E. Hanson, A J. Branscum, and W. O. Johnson. Informative g-priors for logistic
regression. Bayesian Anal., 9(3):597–611, 2014.
[20] C. C. Holmes and L. Held. Bayesian auxiliary variable models for binary and
multinomial regression. Bayesian Anal., 1(1):145–168, 03 2006.
[21] I. Karatzas and S.E. Shreve. Brownian Motion and Stochastic Calculus. Graduate
Texts in Mathematics. Springer New York, 1991.
[22] B. Klartag. A central limit theorem for convex sets. Invent. Math., 168(1):91–131,
2007.
[23] D. Lamberton and G. Pagès. Recursive computation of the invariant distribution
of a diffusion. Bernoulli, 8(3):367–405, 2002.
[24] D. Lamberton and G. Pagès. Recursive computation of the invariant distribution
of a diffusion: the case of a weakly mean reverting drift. Stoch. Dyn., 3(4):435–451,
2003.
[25] V. Lemaire. Estimation de la mesure invariante d’un processus de diffusion. PhD
thesis, Université Paris-Est, 2005.
[26] T. Lindvall and L. C. G. Rogers. Coupling of multidimensional diffusions by reflection. Ann. Probab., 14(3):860–872, 1986.
33
[27] J. C. Mattingly, A. M. Stuart, and D. J. Higham. Ergodicity for SDEs and approximations: locally Lipschitz vector fields and degenerate noise. Stochastic Process.
Appl., 101(2):185–232, 2002.
[28] R. M. Neal. Bayesian learning via stochastic dynamics. In Advances in Neural Information Processing Systems 5, [NIPS Conference], pages 475–482, San Francisco,
CA, USA, 1993. Morgan Kaufmann Publishers Inc.
[29] Y. Nesterov. Introductory Lectures on Convex Optimization: A Basic Course. Applied Optimization. Springer, 2004.
[30] G. Parisi. Correlation functions and computer simulations. Nuclear Physics B,
180:378–384, 1981.
[31] N. G. Polson, J. G. Scott, and J. Windle. Bayesian inference for logistic models using
Polya-Gamma latent variables. Journal of the American Statistical Association,
108(504):1339–1349, 2013.
[32] G. O. Roberts and R. L. Tweedie. Exponential convergence of Langevin distributions and their discrete approximations. Bernoulli, 2(4):341–363, 1996.
[33] P. J. Rossky, J. D. Doll, and H. L. Friedman. Brownian dynamics as smart Monte
Carlo simulation. The Journal of Chemical Physics, 69(10):4628–4633, 1978.
[34] D. Sabanés Bové and L. Held. Hyper-g priors for generalized linear models. Bayesian
Anal., 6(3):387–410, 2011.
[35] D. Talay and L. Tubaro. Expansion of the global error for numerical schemes solving
stochastic differential equations. Stochastic Anal. Appl., 8(4):483–509 (1991), 1990.
[36] C. Villani. Optimal transport : old and new. Grundlehren der mathematischen
Wissenschaften. Springer, Berlin, 2009.
[37] M. Welling and Y. W. Teh. Bayesian learning via stochastic gradient langevin
dynamics. In Proceedings of the 28th International Conference on Machine Learning
(ICML-11), pages 681–688, 2011.
[38] J. Windle, N. G. Polson, and J. G. Scott. Bayeslogit: Bayesian logistic regression, 2013. http://cran.r-project.org/web/packages/BayesLogit/index.html R package version 0.2.
34
This figure "var_tv_0_v2.png" is available in "png" format from:
http://arxiv.org/ps/1605.01559v3
This figure "var_tv_1_2_v2.png" is available in "png" format from:
http://arxiv.org/ps/1605.01559v3
This figure "var_tv_1_4_v2.png" is available in "png" format from:
http://arxiv.org/ps/1605.01559v3
This figure "var_tv_3_4_v2.png" is available in "png" format from:
http://arxiv.org/ps/1605.01559v3
| 10math.ST
|
1
Privacy-preserving Average Consensus: Privacy
Analysis and Optimal Algorithm Design
arXiv:1609.06368v2 [cs.SY] 9 Feb 2017
Jianping He1,2 , Lin Cai2 , Chengcheng Zhao3 , Peng Cheng3 , and Xinping Guan1
Abstract—Privacy-preserving average consensus aims to guarantee the privacy of initial states and asymptotic consensus on
the exact average of the initial value. In existing work, it is
achieved by adding and subtracting variance decaying and zerosum random noises to the consensus process. However, there
is lack of theoretical analysis to quantify the degree of the
privacy protection. In this paper, we introduce the maximum
disclosure probability that the other nodes can infer one node’s
initial state within a given small interval to quantify the privacy. We develop a novel privacy definition, named (, δ)-dataprivacy, to depict the relationship between maximum disclosure
probability and estimation accuracy. Then, we prove that the
general privacy-preserving average consensus (GPAC) provides
(, δ)-data-privacy, and provide the closed-form expression of
the relationship between and δ. Meanwhile, it is shown that
the added noise with uniform distribution is optimal in terms
of achieving the highest (, δ)-data-privacy. We also prove that
when all information used in the consensus process is available,
the privacy will be compromised. Finally, an optimal privacypreserving average consensus (OPAC) algorithm is proposed to
achieve the highest (, δ)-data-privacy and avoid the privacy
compromission. Simulations are conducted to verify the results.
Index Terms—Average consensus, Data privacy, Optimal algorithm, Disclosure probability.
I. I NTRODUCTION
Consensus has attracted extensive attention over the past
decades, since it is an efficient algorithm for distributed computing and control. A consensus algorithm refers to the action
that nodes in the network reach a global agreement regarding a
certain opinion using their local neighbors’ information only
[1]. Due to the strong robustness and scalability, consensus
has been applied in a variety of areas, e.g., coordination and
cooperation [2], [3], distributed estimation and optimization
[4], [5], sensor fusion [6], distributed energy management [7]
and sensing scheduling [8], and time synchronization [9]–[11].
Average consensus is the most commonly adopted consensus algorithm, where the agreement reached by the algorithm
equals the average of all nodes’ initial states. For traditional
average consensus algorithms, each node will broadcast its
real state to neighbor nodes during consensus process. Hence,
under traditional average consensus algorithms, the state information of each node is disclosed to its neighbor nodes.
1: Department of Automation, Shanghai Jiao Tong University, and Key Laboratory of System Control and Information Processing, Ministry of Education
of China, Shanghai, China [email protected]
2: The Dept. of Electrical & Computer Engineering at
the University of Victoria, BC, Canada [email protected];
[email protected]; [email protected]
3: The State Key Lab. of Industrial Control Technology and Innovation Joint
Research Center for Industrial Cyber Physical Systems, Zhejiang University,
China [email protected]; [email protected]
However, in some applications, the initial states of nodes are
private information, which means that nodes do not want to
release their real initial states to other nodes [16]. For example,
consensus algorithm is adopted in social networks for a group
of members to compute the common opinion on a subject
[17]. In this application, each member may want to keep his
personal opinion on the subject secret to other members. Also,
in the multi-agent rendezvous problem [18], a group of nodes
want to eventually rendezvous at a certain location, while the
participators may not want to release their initial locations to
others. This means that when the privacy is concerned, each
node’s real state may not be available to the other nodes, and
thus the traditional consensus algorithm becomes invalid.
Recently, researchers have investigated the privacypreserving average consensus problem, which aims to guarantee that the privacy of initial state is preserved while average
consensus can still be achieved [12]–[16]. The basic idea is to
add random noise to the real state value during the communication to protect the privacy, and then carefully design the
noise adding process such that average consensus is achieved.
For example, Huang et al. [13] designed a differentially private
iterative synchronous consensus algorithm by adding independent and exponentially decaying Laplacian noises to the
consensus process. Their algorithm can guarantee differential
privacy. As the algorithm may converge to a random value,
the exact average consensus may not be guaranteed. Nozari et
al. [14] pointed out and proved that it is impossible to achieve
average consensus and differential privacy simultaneously.
Hence, they design a novel linear Laplacian-based consensus
algorithm, which guarantees that an unbiased estimate of
the average consensus can be achieved almost surely with
differential privacy. Manitara and Hadjicostis [15] proposed
a privacy preserving average consensus algorithm by adding
correlated noises to the consensus process. The proposed
algorithm guarantees the initial state of each node cannot be
perfectly inferred by the other “malicious” nodes. A sufficient
condition is provided under which the privacy of benign
agents’ initial states are preserved. More recently, Mo and
Murray in [16] well addressed the privacy-preserving average
consensus problem by designing a novel PPAC algorithm,
where exponentially decaying and zero-sum normal noises
are added to traditional consensus process. They proved that
PPAC algorithm achieves the exact average consensus in
the mean-square sense, and also proved that the algorithm
achieves minimum privacy breach in the sense of disclosed
space. Braca et al. in [22] examined the interplay between
learning and privacy over multi-agent consensus networks.
They provided an analytical characterization of the interplay
2
between learning and privacy for the consensus perturbing and
preserving strategy, respectively.
However, there is lack of theoretical results to quantify the
degree of the privacy protection and what is the relationship
between the estimation accuracy and privacy. To fill this gap,
in this paper, we provide theoretical privacy analysis for the
GPAC algorithm (consider the general noise adding process)
in the sense of the maximum disclosure probability that other
nodes can infer one node’s initial state within a given small
interval (a given estimation accuracy). A privacy definition,
named (, δ)-data-privacy, which is first introduced in our
previous work [21], is exploited to depict the maximum disclosure probability. This privacy definition reveals the relationship
between privacy and estimation accuracy. We provide theoretical results to quantify the degree of the privacy preservation
and demonstrate the quantitative relationship of the estimation
accuracy and the privacy under the GPAC algorithm. Based on
the analysis, it is found that the noise with uniform distribution
is the optimal one in terms of achieving the highest (, δ)data-privacy, and the exact initial state of a node can be
perfectly inferred, i.e., privacy is compromised, when a node
has all information used in the consensus process. Hence, to
solve this problem, we design a novel OPAC algorithm to
achieve average consensus as well as data-privacy. The main
contributions of this paper are summarized as follows.
• We prove that the GPAC algorithm provides (, δ)-dataprivacy, and obtain a closed-form expression of the
relationship between the estimation accuracy and the
privacy (the relationship between and δ).
• We prove that for the added random noises, the uniform
distribution is optimal in the sense that a PPAC algorithm
can achieve the highest privacy when the mean and
variance of noises are fixed.
• We prove that when all the information used in the
consensus process are available for the estimation, the
maximum disclosure probability will converge to one,
i.e., the initial state of a node is perfectly inferred. This
result reveals how the exact initial state can be inferred.
• We design a novel OPAC algorithm to achieve average consensus while guarantees the highest (, δ)-dataprivacy. It is proved that OPAC algorithm converges to
the exact average consensus, and avoids the privacy to
be lost even if all the information used in the consensus
process is available for the estimation.
The remainder of this paper is organized as follows. Section
II introduces preliminary results and problem formulation. In
Section III, we provide theoretical results on the degree of
privacy pretection. The OPAC algorithm is proposed in Section
IV. Section V verifies the main results and conclusions are
given in Section VI.
II. P RELIMINARIES AND P ROBLEM F ORMULATION
The network is abstracted as an undirected and connected
graph, G = (V, E), where V is the set of nodes and E is
the set of the communication links (edges) between nodes.
(i, j) ∈ E if and only if (iff) nodes i and j can communicate
with each other. Let Ni be the neighbor set of node i, where
j ∈ Ni iff (i, j) ∈ E, i.e., Ni = {j|j ∈ V, (i, j) ∈ E, j 6= i}.
A. Average Consensus
Suppose that there are n (n ≥ 3) nodes in the network
(i.e., |V | = n), and each node i has an initial scalar state
xi (0), where xi (0) ∈ R. For an average consensus algorithm,
each node will communicate with its neighbor nodes and
update its state based on the received information to obtain
the average of all initial state’s values. Hence, the traditional
average consensus algorithm is given as follows,
X
xi (k + 1) = wii xi (k) +
wij xj (k),
(1)
j∈Ni
for ∀i ∈ V , which can be written in the matrix form as
x(k + 1) = W x(k),
(2)
where wii and wij are weights, and W is the weight matrix.
It is well known from [19] that if, 1) wii > 0, and wij > 0 for
(i, j) ∈ E and wij = 0 for otherwise; and 2) W is a doubly
stochastic matrix, then average consensus can be achieved by
(1), i.e.,
Pn
x` (0)
= x̄.
(3)
lim xi (k) = `=1
k→∞
n
When the privacy of nodes’ initial states are concerned, all
nodes are unwilling to release its real state to the neighbor
nodes at each iteration. It means that each xj (k) is unavailable
in (1). To preserve the privacy of nodes’ initial states, a widely
used approach is to add a random noise to the real state value
when a node needs to communicate with its neighbor nodes
at each iteration. We define a new state as
x+
i (k) = xi (k) + θi (k), i ∈ V,
(4)
where θi (k) is the added random noise for privacy preservation
at iteration k. With the noise adding process, the update
equation (1) is changed to,
X
xi (k + 1) = wii x+
wij x+
(5)
i (k) +
j (k)
j∈Ni
= wii [xi (k) + θi (k)] +
X
wij [xj (k) + θj (k)],
(6)
j∈Ni
for ∀i ∈ V . Therefore, a privacy-preserving average consensus algorithm is to design the added noises (including the
distribution and the correlations among them), such that the
goal of (1) is achieved under (5). Note that in (4), the noise
θi (k) is a general random noise (where its distribution is not
fixed), the algorithm (4)–(6) is thus named as the general
privacy-preserving average consensus (GPAC) algorithm in the
remainder part of this paper.
B. Privacy Definitions
Under (4), the broadcast information sequence of node i is
+
+
x+
i (0), xi (1), ..., xi (k), which will be received by its neighbor nodes. Hence, each neighbor node j can infer/estimate
the initial state xi (0) with the received information sequence
from node i. Note that each of the information output, x+
i (k),
equals the weighted sum of the received information in the
previous round plus a noise. Based on the information output,
node j will take the probability over the space of all noises
3
{θi (k)}∞
k=0 (where the space is denoted by Θ) to estimate
the values of the added noises. It then will infer xi (0) by
using the difference between each information output and the
estimated noises, i.e., x̂i (0) = x+
i (k) − θ̂i (k), where θ̂i (k)
is the estimation of random noise θik (θik = x+
i (k) − xi (0)).
Under this estimation, we have
n
o
Pr {|x̂i (0) − xi (0)| ≤ } = Pr |θ̂i (k) − θik | ≤ , (7)
where ≥ 0 is a small constant.
To investigate the relationship between the estimation accuracy and privacy, by referring to [21], we then introduce an
privacy definition, named (, δ)-data-privacy, where 0 ≤ and
0 ≤ δ ≤ 1, as follows.
Definition 2.1: A GPAC algorithm provides (, δ)-dataprivacy, if and only if (iff),
δ=
max
θ̂i (k)∈Θ,k≥0
θik
Pr{|θ̂i (k) −
θik |
≤ },
x+
i (k)
simplicity, we assume that θi (k) and θj (k) are independently
and identically distributed (i.i.d) ∀k ≥ 0 and i 6= j. Let X be
the output of a random variable whose distribution is unknown
and with domain R. Without the knowledge of the distribution,
according to the principle of maximum entropy, we should
take the same probability over all the possible values of the
random variable to estimate the values of X. Therefore, given
an estimation X̂, it is reasonable to assume that
Z ν+
Pr{|X̂ − X| ≤ } max
fθi (0) (y)dy.
(9)
ν∈Θ
ν−
Then, we define two information sets of node i up to
iteration k as follows,
+
Ii0 (k) = {x+
i (0), ..., xi (k)},
(10)
(8)
and
θik .
=
where
− xi (0) and θ̂i (k) is the estimation of
In the above definition, the estimation accuracy is denoted
by parameter and the privacy is expressed by parameter δ.
From (8), it follows that δ is the maximum probability that
each neighbor node j can successfully estimate the initial
state xi (0) in a given interval [xi (0) − , xi (0) + ] with the
information output of node i only. δ is thus named as the
maximum disclosure probability.
Definition 2.2: Given an , if algorithms A1 and A2 provide (, δ1 )-data-privacy and (, δ2 )-data-privacy, respectively,
where δ1 < δ2 , then we say that A1 achieves a higher (, δ)data-privacy than A2 .
+
Ii1 (k) ={Ni , wii , wij , x+
i (0), xj (0),
+
..., x+
i (k), xj (k)|j ∈ Ni }.
(11)
The information set Ii0 (k) only includes the states x+
i (`), ` =
0, 1, ..., k, which are used for communication at iteration `.
Thus, its neighbor nodes can easily obtain Ii0 (k) by storing
the information received from node i at each iteration. The
information set Ii1 (k) includes all information used in consensus process (6) for node i. Other nodes may obtain these
information by an eavesdropping attack.
C. Problem Formulation
A. Privacy of the Algorithm
In this paper, we will investigate the privacy of the GPAC
algorithm (4)–(6) based on the definition of (, δ)-data-privacy,
and then design an optimal privacy-preserving average consensus (OPAC) algorithm in terms of (, δ)-data-privacy protection. In summary, we will consider the following four critical
problems: i) how to quantify and analyze the privacy of the
GPAC algorithm; ii) how will the distribution and correlation
of the added random noises affect the privacy; iii) when and
how will a node’s exact initial state be inferred by the other
nodes; iv) how to achieve the optimal (, δ)-data-privacy and
the exact average consensus, and avoid the privacy of nodes’
initial states to be lost.
In this subsection, based on the definition of (, δ)-dataprivacy, we first analyze the privacy of the GPAC algorithm
and reveal the relationship between the privacy and estimation
accuracy, when Ii0 (k) is available only.
Theorem 3.1: If Ii0 (k) is the only information available to
node j to estimate the value of xi (0) at iteration k, then
δ(k) =
max
θ̂i (k)∈Θ,k∈N+
= max Pr{|θ̂i (0) − θi0 | ≤ |Ii0 (0) }
Before presenting the main results, we first give the basic
assumptions and the information set used for state estimation.
Assume that the distribution and the correlation of the random
variable θi (k), k = 0, 1, ..., and the update rule of the GPAC
algorithm are available to all nodes. The full topology information and n are assumed to be unknown to any node, which
means that each node cannot know the neighbor set of its
neighbor nodes and the number of nodes in the whole network.
The initial states of nodes are assumed to be independent
from each other. For estimation, if there is no information
of a variable, then the variable is viewed with domain R. For
(12)
θ̂i (0)∈Θ
Z
θ̂i (0)+
= max
θ̂i (0)∈Θ
III. P RIVACY A NALYSIS OF GPAC
Pr{|θ̂i (k) − θik | ≤ |Ii0 (k) }
fθi (0) (y)dy,
(13)
θ̂i (0)−
i.e., the relationship between the privacy and the estimation
accuracy always satisfies (13), and the maximum disclosure
probability is not increased with iteration.
Proof: We first prove that, under Ii0 (0), (13) holds. With
0
Ii (0), node j can estimate xi (0) based on the fact that
0
x+
i (0) = xi (0) + θi (0) = xi (0) + θi ,
(14)
and the corresponding estimation x̂i (0) satisfies
x̂i (0) = x+
i (0) − θ̂i (0).
(15)
4
Then, for any estimation θ̂i (0), we have
n
o
Pr |θ̂i (0) − θi0 | ≤ |Ii0 (0)
o
n
= Pr θi (0) ∈ [θ̂i (0) − , θ̂i (0) + ]|Ii0 (0)
Z θ̂i (0)+
=
fθi (0)|I0 (0) (y)dy
n
o
Pr |θ̂i (1) − θi1 | ≤ |Ii0 (1)
o
n
≤ Pr |θi1 (1) − θ̂i1 (1)| ≤ |Ii0 (1),wii
Z z+
≤ max
fθi (0) (y)dy,
z∈Θ
(21)
z−
i
θ̂i (0)−
Z
θ̂i (0)+
≤ max
θ̂i (0)∈Θ
that
fθi (0) (y)dy,
(16)
θ̂i (0)−
which means that (13) holds under information Ii0 (0) at
iteration k = 0.
Then, we prove that (13) holds under Ii0 (1). With Ii0 (1),
node j can estimate xi (0) by using the fact of both (14) and
the following equation for estimation,
xi (1) + θi (1)
x+
i (1)
=
wii
wii
X wil
1
=x+
x+ (0) +
θi (1)
i (0) +
wii l
wii
l∈Ni
X wil
1
=xi (0) + θi (0) +
θi (1) +
x+ (0). (17)
wii
wii l
l∈Ni
Using (14) only, we have
n
o
Pr |θ̂i (0) − θi0 | ≤ |Ii0 (1)
Z θ̂i (0)+
=
fθi (0)|I0 (1) (y)dy
i
θ̂i (0)−
Z
θ̂i (0)+
≤ max
θ̂i (0)∈Θ
fθi (0) (y)dy.
(18)
θ̂i (0)−
Then, we consider the estimation using (17) only. Let
θi1 = θi (0) +
X wil
1
θi (1) +
x+ (0)
wii
wii l
l∈Ni
1
θi (1) + θi1 (1)
= θi (0) +
wii
= θi1 (0) + θi1 (1).
(19)
For any estimation θ̂i (1) (the estimation of θi1 ), we have
n
o
Pr |θ̂i (1) − θi1 | ≤ |Ii0 (1)
n
o
≤ Pr |θi1 − θ̂i (1)| ≤ |Ii0 (1),wii ,θi (1),θi (0)
n
o
≤ Pr |θi1 − θi1 (0) − θ̂i (1) + θi1 (0)| ≤ |Ii0 (1),wii ,θi1 (0)
n
o
≤ Pr |θi1 (1) − θ̂i1 (1)| ≤ |Ii0 (1),wii ,
(20)
where θ̂i1 (1) = θ̂i (1) − θi1 (0) can be viewed as one of the
estimation of θi1 (1). Since the initial states of nodes are
independent from each other and the topology information
is not available for estimating/inferring, there is at least one
variable included in θi1 (1) which is unknown to the other
nodes. Hence, θi1 (1) is viewed as a random variable in (20)
and its distribution is not available to the estimation. It follows
where we have used the assumption (9). Meanwhile, note that
one node can combine (14) and (17) together for estimation.
In this case, we have
Pr{x̂i (0) ∈ [xi (0) − , xi (0) + ]|Ii0 (1) }
Z t1 + Z t2 +
fθi0 ,θi1 (y, z)dzdy
≤ max
t1 ,t2 ∈Θ t −
t2 −
1
Z t1 + Z t2 +
fθi1 |θi0 (z|y)fθi0 (y)dzdy
≤ max
t1 ,t2 ∈Θ t −
t2 −
1
Z t+
fθi (0) (y)dy.
≤ max
t∈Θ
(22)
t−
From (18), (21) and (22), one concludes that (13) holds under
information Ii0 (1) at iteration k = 1.
Following the similar analysis, we can prove that (13) holds
under information set Ii0 (k) at any iteration k. It means that
δ(k) is not an increased function of iteration k, although there
are more information of Ii0 (k) than Ii0 (0) for k > 0.
We thus have completed the proof.
From the above proof, it is observed that the privacy is not
decreased with iteration when only the information set Ii0 (=
{Ii0 (k)|k = 0, 1, ..., ∞}) is available for estimation. The main
reason is that based on Ii0 , node j cannot know the neighbor
set information of node i, so that after one iteration there are
unknown information embedded into x+
i (k) for k ≥ 1. Hence,
after one iteration, using x+
(k)
for
k
≥
1 cannot improve the
i
estimation accuracy. Also, one can see that the value of δ does
not depend on the estimation approaches. Hence, we state the
following theorem.
Theorem 3.2: If Ii0 is the only information available to the
other nodes to estimate the value of xi (0), the GPAC algorithm
achieves (, δ)-data-privacy, where and δ satisfy
Z θ̂i (0)+
δ = max
fθi (0) (y)dy
(23)
θ̂i (0)∈Θ
θ̂i (0)−
and lim→0 δ = 0.
Remark 3.3: It should be noticed that the results in the
above two theorems are obtained under the assumption that
the topology information is unknown to the nodes. If the
assumption is relaxed, the above results could not be true
for the GPAC algorithm in some cases. For example, if the
topology
is available and Ni ⊆ Nj , then x+
i (0)
P information
wij +
wil +
and
l∈Ni wii xl (0) = wii xj (0) in (17) are available to
node j. It leads to that the value of θi (1) is released, which
may decrease the uncertainty of θi (0) due to the correlation
between them. Then, fθi (0)|I0 (1) (y) in (18) will have a smaller
i
variance than fθi (0) (y), such that δ increases w.r.t. k in this
case. Therefore, (12) and (13) are no longer guaranteed.
From the above theorem, one sees that δ depends only on
fθi (0) (y) and since the estimation θ̂i (0) can be any value in
5
the domain of θi (0). Thus, δ is a function of fθi (0) (y) and ,
i.e, δ = δ(fθi (0) (y), ). Based on Definition 2.2, a smaller δ
can provide a higher (, δ)-data-privacy for any given . Then,
we aim to find the optimal distribution for θi (0) such that the
algorithm can achieve the highest (, δ)-data-privacy.
Since f1 (y) is a uniform distribution satisfying (26),
f1 (y) − f2 (y) > 0, y ∈ [−1, 1].
It directly follows that
Z 1
Z
f1 (y)dy −
−1
B. Optimal Noise Distribution
In this subsection, we find an optimal distribution for the
noise adding process in the sense of achieving the highest
(, δ)-data-privacy for the GPAC algorithm. Note that a smaller
means a higher accuracy estimation. It means that when
becomes smaller, the value of δ is more important for the
privacy preservation. Hence, we define the optimal distribution
for privacy concerns as follows.
Definition 3.4: Let fθ∗i (0) (y) be the optimal distribution of
θi (0), it means that for any given distribution fθ1i (0) (y), there
exists an 1 such that δ(fθ∗i (0) (y), ) < δ(fθ1i (0) (y), ) holds
for ∀ ∈ (0, 1 ].
To obtain the optimal distribution described in Definition
3.4, we define arg minfθi (0) (y) δ = fθ∗i (0) (y). Then, we formulate the following minimization problem,
min
(24)
Var{θi (0)} = σ 2 .
The solution of (24) is the optimal distribution for the added
noises with a given mean and variance in terms of (, δ)-dataprivacy for the GPAC algorithm.
Theorem 3.5: If Ii0 is the only information available to node
j to estimate the value of xi (0), then the optimal solution of
problem (24) is that
√ √
√1 , if y ∈ [− 3σ, 3σ],
∗
(25)
fθi (0) (y) = 2 3σ
0,
otherwise,
i.e., given the finite variance, the uniform distribution is
optimal.
Proof: We prove this theorem by contradiction. Without
loss of generality, we assume that σ 2 = 13 . Let f1 (y) and
f2 (y) be the PDF of two random variables with mean 0 and
variance σ 2 = 31 , and they follow a uniform and non-uniform
distribution, respectively. Clearly, we have
1 , if y ∈ [−1, 1],
f1 (y) = 2
(26)
0, otherwise.
Suppose that the non-uniform distribution f2 (y) is the
optimal distribution. From Definition 3.4, there exists an 2 ,
such that
Z t+
Z t+
max
f1 (y)dy > max
f2 (y)dy,
(27)
t∈R
t−
holds for ∀ ∈ (0, 2 ]. Since the above equation holds for
arbitrarily small value of , we infer that
max f1 (y) > max f2 (y).
y∈R
(28)
R1
From the definition of a PDF, we have −1 f1 (y)dy = 1. Then,
we infer from (28) that
Z 1
f2 (y)dy < 1.
(29)
−1
1
3,
Since both f1 (y) and f2 (y) have mean 0 and variance σ 2 =
we have
Z +∞
Z ∞
f1 (y)y 2 dy −
f2 (y)y 2 dy = 0,
(30)
−∞
−∞
which means that
Z
Z 1
(f1 (y) − f2 (y)) y 2 dy =
−1
−1
Z
+
−∞
+∞
f2 (y)y 2 dy.
1
(31)
s.t. E{θi (0)} = 0,
t−
f2 (y)dy > 0.
−1
δ
fθi (0) (y)
t∈R
1
y∈R
For the left hand side of (31), we have
Z 1
Z 1
(f1 (y) − f2 (y)) y 2 dy <
(f1 (y) − f2 (y)) dy
−1
−1
Z 1
=1−
f2 (y)dy.
(32)
−1
R +∞
For the right hand side of (31), since we have −∞ f2 (y)dy =
1 and (29), it holds that
Z −1 Z +∞
Z −1 Z +∞
+
f2 (y)y 2 dy >
+
f2 (y)dy
−∞
1
−∞
1
Z 1
=1−
f2 (y)dy.
(33)
−1
Combining (31), (32) and (33) renders a contradiction that
Z 1
Z 1
1−
f2 (y)dy <
(f1 (y) − f2 (y)) y 2 dy
−1
−1
Z 1
<1−
f2 (y)dy.
(34)
−1
Hence, we cannot find a non-uniform distribution f2 (y) such
that the value of δ is smaller than that under uniform distribution f1 (y). It means that, given the finite variance, the uniform
distribution is the optimal solution of (24). Then, based on the
definition of uniform distribution, it is not difficult to obtain
(25).
We thus have completed the proof.
We have known that for the existing PPAC algorithm
proposed in [16], the normal distribution noises is used in
the noise adding process. It follows from Theorem 3.2 that
the PPAC algorithm provides (, δ)-data-privacy with
Z
1
y2
δ= √
exp − 2 dy.
2σ
σ 2π −
6
If we use the uniform distribution noises to substitute the
normal distribution noises, it can still provide (, δ)-dataprivacy, where δ = √3σ . Clearly, given a small ( σ 2 ),
we have
Z
1
y2
√
< √
exp − 2 dy,
2σ
3σ
σ 2π −
which means that the privacy of PPAC is enhanced.
C. Privacy Compromission
In this subsection, we reveal that for the GPAC algorithm,
when Ii1 (k) (including more information than Ii0 (k), e.g.,
the topology information and information used in consensus
process) is available to other nodes for estimation, the exact
initial state of node i can be perfectly inferred, and thus the
privacy of the initial state is compromised.
Theorem 3.6: If the information set Ii1 (k) of node i is
available to the other nodes for estimation, then
Z θ̂i (0)+
δ(k) ≥ max
fθi (0)|θi (1),...,θi (k) (y)dy, ∀k ≥ 0,
θ̂i (0)∈Θ
θ̂i (0)−
(35)
where fθi (0)|θi (1),...,θi (k) (y) is the conditional
P∞PDF of θi (0)
given conditions θi (1), ..., θi (k). Then, if
`=0 θi (`) = 0,
we have δ = 1, i.e., xi (0) is disclosed and the privacy is
compromised.
Proof: Based on the information set Ii1 (k), the information of weights and states used in (5) is available. That is
the state sequence xi (1), xi (2), ..., xi (k) of node i is released
to other nodes. Then, with (4), one obtains the values of
θi (1), θi (2), ...., θi (k). Thus, when k > 0, all the adding noises
and the states of node i are available to other nodes, except
xi (0) and θi (0).
Then, under information set Ii1 (k), using (14), we have
n
o
Pr |θ̂i (0) − θi0 | ≤ |Ii0 (k)
Z θ̂i (0)+
=
fθi (0)|I0 (k) (y)dy
i
θ̂i (0)−
Z
θ̂i (0)+
=
fθi (0)|θi (1),...,θi (k) (y)dy.
(36)
i.e., θi (0) is fixed and no longer a random variable given
θi (1), θi (2), ...., θi (∞). It follows that
Z θ̂i (0)+
lim max
fθi (0)|θi (1),...,θi (k) (y)dy = 1,
k→∞ θ̂i (0)∈Θ
θ̂i (0)−
which implies that δ = 1. Actually, when both x+
i (0) and
θi (0) in (14) are disclosed, xi (0) is disclosed.
We thus have completed the proof.
Consider the existing PPAC algorithms, e.g., [16], [21]. One
obtain the correlation of the added noises satisfies
k
X
θi (`) = θi (0) +
`=0
k
X
%` νi (`) − %`−1 νi (` − 1)
`=1
= νi (0) − %0 νi (0) + %1 νi (1) − %1 νi (1) + %2 νi (2)
− ... − %k−1 νi (k − 1) + %k νi (k)
= %k νi (k) = φi (k),
(39)
where νi (k) is a random variable with fixed mean (= 0)
2
and variance (= σP
). Given θi (1), ..., θP
i (k), we obtains that
k
k
θi (0) = φi (k) − `=1 θi (`), where
`=1 θi (`) is known.
Then, we have
n
o
Pr |θ̂i (0) − θi (0)| ≤ |Ii1 (k)
n
o
= Pr |φ̂i (k) − φi (k)| ≤
Z φ̂i (k)+
=
fφi (k) (y)dy,
(40)
φ̂i (k)−
and
Z
θ̂i (0)+
fθi (0)|θi (1),...,θi (k) (y)dy
max
θ̂i (0)∈Θ
θ̂i (0)−
Z
φ̂i (k)+
fφi (k) (y)dy,
= max
θ̂i (0)∈Θ
φ̂i (k)−
which satisfies (35). When k → ∞, we have
Z φ̂i (k)+
lim max
fφi (k) (y)dy = 1,
k→∞ θ̂i (0)∈Θ
φ̂i (k)−
since the variance of φ̂i satisfies limk→∞ %2k σ 2 = 0, and thus
δ = 1. Therefore, it further verifies the result given in Theorem
3.6.
θ̂i (0)−
According the definition of δ, it follows that
n
o
δ(k) ≥ max Pr |θ̂i (0) − θi0 | ≤ |Ii0 (k)
D. Further Discussion on Privacy
θ̂i (0)∈Θ
Z
θ̂i (0)+
≥ max
θ̂i (0)∈Θ
fθi (0)|θi (1),...,θi (k) (y)dy,
(37)
θ̂i (0)−
which means
P∞ that (35) holds.
When `=0 θi (`) = 0, we have
θi (0) = −
∞
X
θi (`).
(38)
`=1
Since θi (1), θi (2), ...., θi (k) are available under Ii1 (k) for
any integer k, θi (0) is inferred with (38) when k → ∞,
Differential privacy is a well-known and widely used privacy concept in computer and communication area [20], and
it has been employed in control and network systems recently
[23]. A differentially private algorithm promises that any two
similar/close inputs will have approximately the same outputs,
so that an adversary cannot infer from the data output with a
high probability whether the data are associated with a single
user or not. It has been proved by Nozari et al. in [14] that
nodes in the network system cannot simultaneously converge
to the average of their initial states and preserve differential
privacy of their initial states. This motivated us to develop the
definition of the (, δ)-data-privacy. The proposed (, δ)-data
privacy can be used to reveal the relationship between the
7
the maximum data disclosure probability (δ) under a given
estimation accuracy range ().
Consider the general noise adding mechanism that added
a random noise to the initial data for data publishing. It is
well known that when the adding noise is Laplacian noise,
the mechanism ensures differential privacy, but if the noise
is Gaussian or Uniform distribution, the differential privacy
cannot be guaranteed. Hence, the uniform noise is not good
in the sense of differential privacy. However, in term of
(, δ)-data-privacy, it is shown in this paper that both the
Gaussian and Uniform noise are (, δ)-data-private, and using
the Uniform noise can achieve the highest privacy. Clearly,
the privacy of (, δ)-data-privacy is different from that of
differential privacy. It is worth to investigate the relationship
between these two kinds of privacy definition in theory, which
beckons further investigation.
IV. OPAC A LGORITHM
In this section, we design an OPAC algorithm to achieve
the highest (, δ)-data-privacy, and at the same time to avoid
privacy to be compromised even if the information Ii1 (∞) of
each node i is available to other nodes.
A. Algorithm Design
From the privacy analysis in the above section, we note
that the uniform distribution is optimal for the added noise
in terms of achieving the highest (, δ)-data-privacy with δ =
√ (given variance σ). Hence, in each iteration of the OPAC
3σ
algorithm, we will use uniform distribution noise. We also
note that privacy is compromised when Ii1 (∞) is available.
It is because that the nodes can use Ii1 (∞) to obtain the real
values
P∞ of θi (1), θi (2), ...., θi (∞), and then use the correlation
k=0 θi (k) = 0 to infer θi (0), and thus the value of xi (0)
is revealed. To avoid the privacy compromission in this case,
we introduce a secret continuous function Fij (z) : R → R for
node i with respect to its neighbor node j. Suppose that Fij (z)
and Fji (z) are only available to nodes i and j, and Fij (z)
may or may not equal to Fji (z). Then, the OPAC algorithm
is described as follows.
Algorithm 1 : OPAC Algorithm
1: Initialization: Each node i selects
uniform distribution random
√ a√
variable νi (0) from interval [− 3σ, 3σ], and arbitrarily selects
a constant sequence zij (∈ R) for j ∈ Ni .
2: Let θi (0) = νi (0) and x+
i (0) = xi (0) + θi (0). Then, each node
i transmits x+
i (0) and zij to its neighbor node j.
3: Each node i calculates ν̃i (0) by
X
[Fij (zij ) − Fji (zji )] , ∀i ∈ V.
(41)
ν̃i (0) = νi (0) −
j∈Ni
4: Iteration: Each node updates its state with (5).
5: Each node generates
distribution random variable νi (k)
√ a uniform
√
(
θi (k) =
%νi (1) − ν̃i (0),
if k = 1;
%k νi (k) − %k−1 νi (k − 1),
if k ≥ 2,
(42)
where % ∈ (0, 1) is a constant for all nodes.
7: Each node i communicates with its neighbors with x+
i (k).
8: Let k = k + 1 and go to step 4.
difficult to infer that each θi (k) is bounded. Meanwhile, it
follows from (42) that
lim |θi (k)| ≤ lim |%k νi (k) − %k−1 νi (k − 1)|
k→∞
√
√
≤ lim [%k 3σ + %k−1 3σ] = 0,
k→∞
k→∞
which means that the noises are decaying and converge to 0.
Next, we prove that the sum of all nodes’ added noises
equals to zero. Note that
n X
∞
X
n
X
θi (k) =
i=1 k=0
+
θi (0) +
i=1
n X
∞
X
=
n
X
(%νi (1) − ν̃i (0))
i=1
i=1
+
θi (1)
(%k νi (k) − %k−1 νi (k − 1))
νi (0) +
n
X
n
X
i=1
i=1 k=2
n
X
=
B. Convergence and Privacy Analysis
In this subsection, we analyze the convergence and the
privacy of the OPAC algorithm.
Theorem 4.1: Using the OPAC algorithm, we have (3) holds
for ∀i ∈ V , i.e., an exact average consensus is achieved.
Proof: From Theorem 4.1 of [21], we know that if the
added noises in (4) are bounded and decaying, and the sum of
all nodes’ added noises equals zero, then average consensus
can be achieved. In the following, we prove that the added
noises used for the OPAC algorithm satisfy these conditions.
We first prove that the added noises are bounded
√ and
√ exponentially decaying. Clearly, θi (0) = νi (0) ∈ [− 3σ, 3σ] is
bounded. Since each Fij (z) is continuous function, its value
is bounded for any given z. Then, it follows from (41) that
ν̃i (0) is bounded.
√ √ For k ≥ 1, because νi (k) is selected from
interval [− 3σ, 3σ] and θi (k) is generated by (42), it is not
3σ] for k ≥ 1.
from interval [− 3σ,
6: Each node i uses θi (k) in (4) to get x+
i (k), where
(%∞ νi (∞) − %1 νi (1))
i=1
n
X
νi (0) −
i=1
n
X
ν̃i (0),
i=1
where we have used the fact that %∞ νi (∞) = 0. Substituting
(41) into the above equation yields that
n X
∞
X
θi (k) =
i=1 k=0
n
X
νi (0)
i=1
−
n
X
νi (0) −
i=1
=
n X
X
X
(Fij (zij ) − Fji (zji ))
j∈Ni
[Fji (zji ) − Fij (zij )] .
i=1 j∈Ni
Since for each pair of Fji (zji )−Fij (zij ) using in node i, there
exists a pair of Fij (zij ) − Fji (zji ) with negative value using
8
in node j, it follows that
n
X
X
V. P ERFORMANCE E VALUATION
[Fji (zji ) − Fij (zij )] = 0.
i=1 j∈Ni
Pn P∞
Hence, we have i=1 k=0 θi (k) = 0.
Thus, the proof is completed.
The following theorem can be obtained from Theorem 3.2
directly, since OPAC is one of the GPAC algorithm.
Theorem 4.2: If Ii0 is the only information available to the
other nodes to estimate the value of xi (0), then the OPAC
algorithm achieves (, δ)-data-privacy, where δ = √3σ and
lim→0 δ = 0.
Then, the following theorem shows that under Ii1 , the
privacy compromission can be avoided by OPAC.
Theorem 4.3: Suppose that the information set Ii1 of node
i is available to the other nodes and each node has at least
two neighbors (i.e., |Ni | ≥ 2 for all i ∈ V ). Then, the privacy
compromission can be avoided by OPAC.
Proof: It has been known that when Ii1 of node i is
available to other nodes, its neighbor node j can obtain
the
P∞real values of θi (1), θi (2), ...., θi (∞). Then, the value of
k=1 θi (k) is released. Note that
∞
X
θi (k) = (%1 νi (1) − ν̃i (0)) +
k=1
∞
X
θi (k)
k=2
∞
= (%1 νi (1) − ν̃i (0)) + (% νi (∞) − %1 νi (1))
= ν̃i (0).
It means that the value of ν̃i (0) is released and available to
node j. From (41), one sees that ν̃i (0) 6= θi (0)1 and
X
ν̃i (0) = θi (0) −
[Fij (zij ) − Fji (zji )] .
(43)
j∈Ni
Since |Ni | ≥ 2 and only Fij and Fji is known to node j, there
exists Fijo (zijo ) − Fjo i (zjo i ) for j0 ∈ Ni in (43) is not known
by node j. Meanwhile, Fijo (zijo ) −√Fjo i (z
√jo i ) has domain R,
thus one infers that for any c ∈ [− 3σ, 3σ],
Pr{θi (0) = c|ν̃i (0)} = Pr{θi (0) = c}.
Hence, even if the value of ν̃i (0) is released, node j cannot
increase the estimation accuracy of θi (0) with (43). One thus
concludes that based on the OPAC algorithm, the privacy
compromission is avoided.
We thus have completed the proof.
If node i has only one neighbor node j, node j can infer
the value of Fij (zij ) − Fji (zji ). Then, from (43), node j can
obtain the value of θi (0) and xi (0) when ν̃i (0) is known.
Remark 4.4: From the above two theorems, one sees that
using OPAC algorithm, we have δ = √3σ , which is the
optimal privacy that can be achieved from solving problem
(24). Furthermore, δ = √3σ can be guaranteed by OPAC
algorithm under Ii1 (∞). Thus, OPAC algorithm can achieve
much higher (, δ)-data-privacy than the existing PPAC.
1 This is the main difference between OPAC and PPAC algorithm, and the
main reason why OPAC can avoid privacy compromission.
In this section, we conduct simulations to verify the obtained theoretical results and evaluate the performance of the
proposed OPAC algorithm.
A. Simulation Scenario
Consider the network with 50 nodes which are randomly
deployed in a 100m × 100m area, and the maximum communication range of each node is 30m. We consider the
normal distribution and uniform distribution of the added
noises, respectively, where the mean and variance of them
are set 0 and σ 2 = 1. We set % = 0.9. The initial states
of the nodes are randomly selected from [0, 10]. The function,
d(t) = max |xi (t) − x̄|, is defined as the maximum difference
i∈V
between the nodes’ states and the average value.
B. Verification
Fig. 1(a) compares the convergence speed of the PPAC
algorithm using normal and uniform distribution noises, in
which the basic design is the same as PPAC proposed in [16].
It is observed that under the two different distributions, the
PPAC algorithm has the same convergence speed. This justifies
that the convergence speed only depends on the eigenvalues
of the weighted matrix W and the value of % as proved in
[16].
Fig. 1(b) compares the (, δ)-data-privacy under Ii0 with
normal and uniform distribution noises. In simulation, we
conduct 10, 000 simulation runs. For each run, one node first
generates a state θi (0) randomly with the given distribution,
and the other node generates 10, 000 random numbers with the
same distribution and use them as the estimation of θi (0) (i.e.,
θ̂i (0)). Then, one get the probability of |θ̂i (0) − θi (0)| ≤ in
each run, and we use the maximum probability among these in
all runs simulation as the value of δ. For the theoretical results,
we use (23) to calculate the value of δ under two different
distributions. Clearly, one can observe from Fig. 1(b) that
uniform distribution is much better than normal distribution
in the sense of (, δ)-data-privacy. It is also observed that δ in
simulation matches its value in theory.
Fig. 1(c) compares the (, δ)-data-privacy under Ii1 using
normal and uniform distribution noises. The simulations here
are conducted similarly as those in Fig. 1(b), except that
when the iteration increases, the variance of the noises will
be changed to σ 2 = %2k since (40) will be used for estimation
at iteration k. We use (23) to calculate the value of δ, and
the corresponding results are denoted as theoretical results.
Both in simulation and theory, we set = 0.2. As shown in
Fig. 1(c), the maximum disclosure probability increases with
iteration and will converge to 1, i.e., the privacy decays with
iteration and will eventually be compromised.
C. Evaluation
In this subsection, we will evaluate the performance of the
OPAC algorithm. Using the same setting as the above subsection, the OPAC algorithm can guarantee the similar privacy
as the blue line shown in Fig. 1(b) under Ii1 . This is because
9
uniform distribution:simulation
normal distribution:simulation
uniform distribution:theorem
normal distribution:theorem
0.6
Disclosed Probability
Maximum Difference
4
3
2
1
0
1
0.7
Uniform Distribution
Normal Distribution
0.5
0.4
0.3
0.2
0.1
0
20
40
60
80
Iteration
100
120
0.2
140
0.4
0.6
0.8
ε
(b) privacy under Ii0
0.4
uniform distribution:simulation
normal distribution:simulation
uniform distribution:theorem
normal distribution:theorem
1
0
5
10
15
Iteration
20
25
30
(c) privacy under Ii1
The convergence and privacy comparison under different random noise distribution.
uniform distribution noise is used in OPAC and the secret
function makes the subsequent (k ≥ 1) information cannot
increase the disclosure probability. Therefore, the OPAC can
guarantee much stronger privacy than the GPAC, since it can
achieve the same data-privacy under Ii1 as the GPAC under
Ii0 .
12
x1(k)
x2(k)
States of Nodes
10
x3(k)
8
x4(k)
average
6
4
2
0
20
40
60
80
100
120
140
Iteration
(a) converge to average consensus
5
OPAC
PPAC
Maximum Difference
4
3
2
1
0
20
40
60
80
100
120
140
Iteration
(b) same covergence speed
Fig. 2.
0.6
0.2
(a) convergence
Fig. 1.
0.8
Disclosed Probability
5
The performance evaluation of the OPAC algorithm.
Then, we test the convergence of the OPAC algorithm. Set
Fij = i+2j
50 . As shown in Fig. 2(a), we find that the nodes’
states will converge to the exact average with the OPAC, which
means that an exact average consensus can be achieved by the
proposed algorithm. Fig. 2(b) compares the convergence speed
of the OPAC and PPAC, it is found that they almost have the
same convergence speed. Hence, added secret function will
not affect the convergence speed.
VI. C ONCLUSIONS
In this paper, we investigated the privacy of the GPAC algorithm. We proposed a novel privacy definition, named (, δ)data-privacy, to depict the relationship between privacy and
estimation accuracy, so that the degree of the privacy can be
well quantified. We proved that the GPAC algorithm achieves
(, δ)-data-privacy, and obtained the closed-form expression
of the relationship between and δ. We also proved that the
noise with uniform distribution guarantee a highest privacy
when is small enough. We revealed that the privacy will
be lost when the information used in each consensus iteration
is available to the other nodes. Then, to solve this problem
and achieve highest (, δ)-data-privacy, we proposed OPAC
algorithm, followed by the convergence and privacy analysis.
Lastly, simulations are conducted to demonstrate the efficiency
of the proposed algorithm.
R EFERENCES
[1] R. Olfati-Saber, J. A. Fax, and R. M. Murray, “Consensus and cooperation in networked multi-agent systems,” Proceedings of the IEEE, 95(1):
215–233, 2007.
[2] V. Blondel, J. M. Hendrickx, A. Olshevsky, and J. Tsitsiklis, “Convergence in multiagent coordination, consensus, and flocking,” in Proc. of
IEEE CDC, 2005.
[3] W. Ren, B. Randal., and A. Ella, “Information consensus in multivehicle
cooperative control: Collective group behavior through local interaction,”
IEEE Control Systems Magazine, 27(2): 71–82, 2007.
[4] F. Pasqualetti, R. Carli, A. Bicchi, and F. Bullo, “Distributed estimation
and detection under local information,” in Proc. of IFAC, 2010.
[5] G. Mateos, I. Schizas and G. Giannakis, “Distributed recursive leastsquares for consensus-based in-network adaptive estimation,” IEEE
Trans. Signal Processing, 57(11): 4583–4588, 2009.
[6] L. Xiao, S. Boyd, and S. Lall, “A scheme for robust distributed sensor
fusion based on average consensus,” in Proc. of ISIPSN, 2005.
[7] C. Zhao, J. He, P. Cheng and J. Chen, “Consensus-based energy
management in smart grid with transmission losses and directed communication,” IEEE Trans. Smart Grid, DOI: 10.1109/TSG.2015.2513772.
[8] J. He, L. Duan, F. Hou, P. Cheng, and J. Chen, “Multi-period scheduling
for wireless sensor networks: A distributed consensus approach,” IEEE
Trans. Signal Processing, 63(7): 1651-1663, 2015.
[9] L. Schenato and F. Fiorentin, “Average timesynch: A consensus-based
protocol for clock synchronization in wireless sensor networks,” Automatica, 47(9): 1878–1886, 2011.
10
[10] R. Carli, and S. Zampieri, “Network clock synchronization based on the
second order linear consensus algorithm,” IEEE Trans. Automat. Contr.,
59(2): 409–422, 2014.
[11] J. He, P. Cheng, L. Shi, and J. Chen, “Time synchronization in WSNs:
A maximum value based consensus approach,” IEEE Trans. Automat.
Contr., 59(3): 660–674, 2014.
[12] J. Le Ny and G. Pappas, “Differentially private filtering,” IEEE Trans.
Automat. Contr., 59(2): 341–354, 2014.
[13] Z. Huang, S. Mitra, and G. Dullerud, “Differentially private iterative
synchronous consensus.” in Proc. ACM workshop on Privacy in the
electronic society, 2012.
[14] E. Nozari, P. Tallapragada, and J. Cortes, “Differentially private average
consensus: Obstructions, trade-offs, and optimal algorithm design.”
arXiv preprint arXiv:1512.09039, 2016.
[15] N. Manitara and C. Hadjicostis, “Privacy-preserving asymptotic average
consensus.” in Proc. of IEEE ECC, 2013.
[16] Y. Mo, and R. Murray, “Privacy preserving average consensus,” IEEE
Trans. Automat Contr., accepted and to appear, 2016.
[17] M. DeGroot, “Reaching a consensus,” Journal of the American Statistical Association, 69(345), 118–121, 1974.
[18] J. Lin, A. S. Morse, and B. D. Anderson, “The multi-agent rendezvous
problem,” in Proc. of IEEE CDC, 2003.
[19] A. Olshevsky and J. Tsitsiklis, “Convergence speed in distributed
consensus and averaging,” SIAM Review, 53(4): 747–772, 2011.
[20] C. Dwork, “Differential privacy,” in Automata, languages and programming, Springer, 1-12, 2006.
[21] J. He, L. Cai, P. Cheng, M. Xing, J. Pan and L. Shi.“Private
and accurate data aggregation against dishonest nodes,”
https://arxiv.org/pdf/1609.06381v2.pdf, 2016.
[22] P. Braca, R. Lazzeretti, S. Marano, and V. Matta. “Learning with privacy
in consensus + obfuscation,” IEEE Signal Processing Letters, 23(9):
1174-1178, 2016.
[23] J. Cortes, G. Dullerud, S. Han, J. Le Ny, S. Mitra, and G. Pappas.
“Differential privacy in control and network systems,” in Proc. of IEEE
CDC, 2016.
| 3cs.SY
|
Lower Bounds for Symbolic Computation on Graphs:
Strongly Connected Components, Liveness, Safety, and Diameter
Krishnendu Chatterjee1 , Wolfgang Dvořák2 , Monika Henzinger3 , and
Veronika Loitzenbauer3,4
1
IST Austria
TU Wien, Institute of Information Systems, Vienna, Austria
3
University of Vienna, Faculty of Computer Science, Vienna, Austria
4
Bar-Ilan University
arXiv:1711.09148v1 [cs.DS] 24 Nov 2017
2
Abstract
A model of computation that is widely used in the formal analysis of reactive systems is
symbolic algorithms. In this model the access to the input graph is restricted to consist of
symbolic operations, which are expensive in comparison to the standard RAM operations. We
give lower bounds on the number of symbolic operations for basic graph problems such as the
computation of the strongly connected components and of the approximate diameter as well
as for fundamental problems in model checking such as safety, liveness, and co-liveness. Our
lower bounds are linear in the number of vertices of the graph, even for constant-diameter graphs.
For none of these problems lower bounds on the number of symbolic operations were known
before. The lower bounds show an interesting separation of these problems from the reachability
problem, which can be solved with O(D) symbolic operations, where D is the diameter of the
graph.
Additionally
we present an approximation algorithm for the graph diameter which requires
√
Õ(n D) symbolic steps to achieve a (1 + )-approximation for any constant > 0. This
compares to O(n · D) symbolic steps for the (naive) exact algorithm and O(D) symbolic
steps for a 2-approximation. Finally we also give a refined analysis of the strongly connected
components algorithms of [GPP08], showing that it uses an optimal number of symbolic steps
that is proportional to the sum of the diameters of the strongly connected components.
1
Introduction
Graph algorithms are central in the formal analysis of reactive systems. A reactive system consists
of a set of variables and a state of the system corresponds to a set of valuations, one for each of
these variables. This naturally induces a directed graph: Each vertex represents a state of the system
and each directed edge represents a state transition that is possible in the system. As the number of
vertices is exponential in the number of variables of the system, these graphs are huge and, thus, they
are usually not explicitly represented during their analysis. Instead they are implicitly represented
using e.g., binary-decision diagrams (BDDs) [Bry86, Bry92]. To avoid considering specifics of the
implicit representation and their manipulation, an elegant theoretical model for algorithms that work
on this implicit representation has been developed, called symbolic algorithms (see e.g. [Bur+ 90,
Cla+ 96, Som99, CGP99, Cla+ 03, GPP08, Cha+ 13]). In this paper we will give novel upper and
(unconditional) lower bounds on the number of operations required by a symbolic algorithm for
solving classic graph-algorithmic questions, such as computing the strongly connected components
and the (approximate) diameter, as well as for graph-algorithmic questions that are important in the
analysis of reactive systems, such as safety, liveness, and co-liveness objectives. Our lower bounds
are based on new reductions of problems from communication complexity to symbolic algorithms.
Symbolic algorithms. A symbolic algorithm is allowed to use the same mathematical, logical,
and memory access operations as a regular RAM algorithm, except for the access to the input graph:
It is not given access to the input graph through an adjacency list or adjacency matrix representation
but instead only through two types of symbolic operations:
1. One-step operations Pre and Post: Each predecessor (Pre) (resp., successor (Post)) operation
is given a set X of vertices and returns the set of vertices Y with an edge to (resp., edge from)
some vertex of X.
2. Basic set operations: Each basic set operation is given one or two sets of vertices and performs
a union, intersection, or complement on these sets.
An initial set of vertices is given as part of the input, often consisting of a single vertex.
Symbolic operations are more expensive than the non-symbolic operations and thus one is mainly
interested in the number of symbolic operations of such an algorithm (and the exact number of
non-symbolic operations is often neglected). Moreover, as the symbolic model is motivated by the
compact representation of huge graphs, we aim for symbolic algorithms that only store O(1) or
O(log n) many sets of vertices as otherwise algorithms become impractical due to the huge space
requirements. Additionally, every computable graph-algorithmic question can be solved with 2n
symbolic one-step operations when storing O(n) many sets (and allowing an unbounded number of
non-symbolic operations): For every vertex v perform a Pre and a Post operation, store the results,
which represent the full graph, and then compute the solution on this graph, using only non-symbolic
operations. Note, however, that our lower bounds do not depend on this requirement, i.e., they
also apply to symbolic algorithms that store an arbitrary number of sets. Furthermore the basic set
operations (that only require vertices, i.e., the current state variables) are computationally much less
expensive than the one-step operations (that involve both variables of the current and of the next
state). Thus, to simplify the analysis of symbolic algorithms, we only analyze the number of one-step
operations in the lower bounds that we present. For all upper bounds in prior work and in our work
the number of basic-set operations is at most linear in the number of one-step operations.
There is an interesting relationship between the two types of symbolic operations and Boolean
matrix-vector operations: Interpreting the edge relationship as a Boolean matrix and a vertex set
as a Boolean vector, the one-step operations correspond to (left- and right-sided) matrix-vector
multiplication, where the matrix is the adjacency matrix, and basic set operations correspond to basic
1
vector manipulations. Thus, an equivalent way of interpreting symbolic algorithms is by saying that
the access to the graph is only allowed by performing a Boolean matrix-vector multiplication, where
the vector represents a set of vertices and the matrix is the adjacency matrix.
Note also that there is a similarity to the CONGEST and the LOCAL model in synchronous
distributed computation, as in these models each vertex in a synchronous network knows all its
neighbors and can communicate with all of them in one round (exchanging O(log n) bits in the
CONGEST model), and the algorithmic complexity is measured in rounds of communication. While
in these models all neighbors of every individual vertex, i.e., all edges of the graph, can be determined
in one round of communication, in the symbolic model this might require n Pre and n Post operations,
each on an singleton set. Thus, determining (and storing) all edges of a symbolically represented
graph is expensive and we would ideally like to have algorithms that use sub-linear (in the number of
vertices) many symbolic one-step operations.
Objectives. First we formally introduce the most relevant graph-algorithmic questions from the
analysis of reactive systems [MP92]. Given a graph G = (V, E) and a starting vertex s ∈ V , let Ps
be the set of infinite paths in G starting from s. Each objective corresponds to a set of requirements
on an infinite path and the question that needs to be decided by the algorithm is whether there is a
path in Ps that satisfies these requirements, in which case we say the path satisfies the objective. An
objective A is the dual of an objective B if a path satisfies A iff it does not satisfy B.
Let T ⊆ V be a set of target vertices given as input. The most basic objective is reachability
where an infinite path satisfies the objective if the path visits a vertex of T at least once. The dual
safety objective is satisfied by infinite paths that only visit vertices of T . The next interesting objective
is the liveness (aka Büchi) objective that requires an infinite path to visit some vertex of T infinitely
often. The dual co-liveness (aka co-Büchi) objective requires an infinite path to eventually only visit
vertices in T . Verifying these objectives are the most fundamental graph-algorithmic questions in the
analysis of reactive systems.
Computing the strongly connected components (SCCs) is at the heart of the fastest algorithms for
liveness and co-liveness: For example, there is a reduction from liveness to the computation of SCCs
that takes symbolic steps in the order of the diameter of the graph. Thus, determining the symbolic
complexity of SCCs also settles the symbolic complexity of liveness.
Furthermore, the diameter computation plays a crucial role in applications such as bounded
model-checking [Bie+ 03], where the goal is to analyze the system for a bounded number of steps,
and it suffices to choose the diameter of the graph as bound. Second, in many scenarios, such as in
hardware verification, the graphs have small diameter, and hence algorithms that can detect if this is
the case and then exploit the small diameter are relevant [Bie+ 03]. Motivated by these applications,
we define the diameter of a graph as the largest finite distance in the graph, which coincides with the
usual graph-theoretic definition on strongly connected graphs and is more general otherwise.
Note that linear lower bounds for the number of symbolic operations are non-trivial, since a
one-step operation can involve all edges. For example, to determine all the neighbors of a given
vertex v takes one symbolic operation, while it takes O(deg(v)) many operations in the classic
setting. In the following we use n to denote the number of vertices of a graph G and D(G) = D to
denote its diameter.
Previous results. To the best of our knowledge, no previous work has established lower bounds for
symbolic computation.
There is some prior work on establishing upper bounds on the number of symbolic operations:
In [GPP08] a symbolic algorithm that computes the SCCs with O(n) symbolic operations is presented.
This algorithm leads to an algorithm for liveness and co-liveness with O(n) symbolic operations and
improves on earlier work by [BGS06], which requires O(n log n) symbolic operations.
2
Table 1: Bounds on the number of required symbolic operations for different tasks. Θ(n) bounds
hold even for graphs with constant diameter D.
Reach(T )
SCC
Safe(T )
Büchi(T )
coBüchi(T )
Θ(D)
Θ(n)
Θ(n)
Θ(n)
Θ(n)
Table 2: Bounds on the number of symbolic operations for approximating the diameter of a graph.
The lower bounds even hold for strongly connected graphs with constant diameter D.
approx.
upper bound
lower bound
exact
O(n · D)
Ω(n)
1+ε
√
e
O(n
D)
Ω(n)
3/2 − ε
√
e
O(n
D)
Ω(n)
2
O(D)
Note that for the reachability objective the straightforward algorithm requires O(D) symbolic
operations: Starting from the set containing only the start vertex s, repeatedly perform a Postoperation until T is reached. For safety the straightforward algorithm takes O(n) symbolic operations:
Iteratively remove from F vertices that do not have an outgoing edge to another vertex of F , i.e.,
vertices of F \ Pre(F ), until a fixed point is reached.
Finally, there is a trivial algorithm for computing the diameter of the graph: Simply determine the
depth of a breadth-first search from every vertex and output the maximum over all depths. Computing
the depth of a breadth-first search can be done with O(D) many symbolic steps, thus this requires
O(nD) many symbolic steps in total. In a strongly connected graph a 2-approximation of the
diameter of the graph can be obtained by computing one breadth-first search from and one to some
arbitrary vertex and output the sum of the depths. This takes O(D) symbolic steps.
Our contributions. Our main contributions are novel lower bounds for the number of symbolic
operations for many of the above graph-algorithmic questions, leading to an interesting separation
between seemingly similar problems.
1. For reachability objectives, the basic symbolic algorithm requires O(D) symbolic operations.
Quite surprisingly, we show that such diameter-based upper bounds are not possible for its
dual problem, namely safety, and are also not possible for liveness and co-liveness objectives.
Specifically, we present tight lower bounds to show that, even for constant-diameter graphs,
Ω(n) one-step symbolic operations are required for safety, liveness, and co-liveness objectives.
In addition we establish tight bounds for symbolic operations required for the computation
of SCCs, showing a lower bound of Ω(n) for constant-diameter graphs. See Table 1 for a
summary of these results.
2. We show that even for strongly-connected constant-diameter graphs approximating the diameter
requires Ω(n) symbolic steps. More precisely, any (3/2 − ε)-approximation algorithm requires
Ω(n) symbolic one-step operations, even for undirected and connected graphs with constant
diameter. We also give a novel upper√bound: We present a (1 + ε)-approximation algorithm
e
for any constant ε > 0 that takes O(n
D) symbolic steps. This can be compared to the trivial
O(D) 2-approximation algorithm and the O(nD) exact algorithm. Notice that for explicitly
represented graphs the approximation of the diameter is already hard for constant-diameter
graphs while in the symbolic model there exists a trivial O(n) upper bound in this case, thus
showing a lower bound of Ω(n) is non-trivial. See Table 2 for a summary of these results.
3. Finally we give a refined analysis of the number of symbolic steps required for computing
strongly connected components based on a different problem parameter. Let SCCs(G) be
3
the set of all SCCs of G and DC the diameter of the strongly connected component C.
WePgive matching upper and lower bounds showing P
that the SCCs can be computed with
Θ( C∈SCCs(G) (DC + 1)) symbolic steps. Note that C∈SCCs(G) (DC + 1) can be a factor
n larger than D(G).
Key technical contribution. Our key technical contribution is based on the novel insight that lower
bounds for communication complexity can be used to establish lower bounds for symbolic computation. We feel that this connection is of interest by itself and might lead to further lower bounds for
symbolic algorithms.
Our lower bounds are by two kinds of reductions, both from the communication complexity
problem of Set Disjointness with k elements. First, we give reductions that construct graphs such that
one-step operations can be computed with O(1) bits of communication between Alice and Bob and
thus allow for linear lower bounds √
on the number symbolic operations.
Second, we give a reduction
√
that constructs a graph with only k many vertices, i.e., n = k, but allows one-step operations
to require O(n) bits of communication. This again results in linear lower bounds on the number of
symbolic operations.
2
Preliminaries
Symbolic Computation. We consider symbolic computation on graphs. Given an input graph
G = (V, E) and a set of vertices S ⊆ V , the graph G can be accessed only by the following two
types of operations:
1. Basic set operations like ∪, ∩, \, ⊆, and =;
2. One-step operations to obtain the predecessors or successors of the vertices of S in G. In
particular we define the operations
Pre(S) = {v ∈ V | ∃s ∈ S : (v, s) ∈ E} and Post(S) = {v ∈ V | ∃s ∈ S : (s, v) ∈ E} .
In the applications the basic set operations are much cheaper as compared to the one-step operations.
Thus we aim for lower bounds on the number of one-step operations, while not accounting for
set operations. In all our upper bounds the number of set operations is at most of the same order
as the number of one-step operations. Note that there is a one-to-one correspondence between a
one-step operation and a Boolean matrix-vector multiplication with the adjacency matrix and that for
undirected graphs Pre and Post are equivalent.
Communication Complexity Lower Bound for Set Disjointness. Our lower bounds are based on
the known lower bounds for the communication complexity of the Set Disjointness problem. The
classical symmetric two-party communication complexity model is as follows [KN97]. There are
three finite sets X, Y, Z, the former two are possible inputs for a function f : X × Y → Z, where
the actual input x ∈ X is only known by Alice, and the actual input y ∈ Y is only known by Bob.
Alice and Bob want to evaluate a function f (x, y) while sending as few bits as possible to each other.
The communication happens according to a fixed protocol, known to both players beforehand, that
determines which player sends which bits when, and when to stop.
Set Disjointness. In the Set Disjointness problem we have a universe U = {0, 1, . . . , k − 1} of
k elements and both sets X, Y contain all bit vectors of length k, i.e., they represent all possible
subsets of U and are of size 2k . Alice has a vector x ∈ X and Bob has a vector y ∈ Y , and the
function f is defined as f (x, y) = 1 if for all 0 ≤ i ≤ k − 1 either xi = 0 or yi = 0, and f (x, y) = 0
otherwise. We will sometimes use Sx and Sy to denote the sets corresponding to the vectors x and
4
v0,0
v0,1
v0,2
v1,0
v1,1
v1,2
Figure 1: Reduction 3.1 for k = 4, ` = 2, Sx = {2, 3}, Sy = {0, 1, 3}
y, i.e., Sx = {i | xi = 1} and Sy = {i | yi = 1} and f (x, y) = 1 iff Sx ∩ Sy = ∅. We next state a
fundamental lower bound for the communication complexity of the Set Disjointness problem which
will serve as basis for our lower bounds on the number of symbolic operations.
Theorem 2.1 ([KS92, Raz92, Bar+ 04, HW07, KN97]). Any (probabilistic bounded error or deterministic) protocol for the Set Disjointness problem sends Ω(k) bits in the worst case over all
inputs.
3
Lower Bounds
In this section we present our lower bounds, which are the main results of the paper.
3.1
Lower Bounds for Computing Strongly Connected Components
We first consider the problem of computing the strongly connected components (SCCs) of a symbolically represented graph. The best known symbolic algorithm is by Gentilini et al. [GPP08] and
computes the SCCs of a Graph G with O(min(n, D · |SCCs(G)|)) symbolic one-step operations
and thus matches the linear running time of the famous Tarjan algorithm [Tar72] in the non-symbolic
world.
We provide lower bounds showing that the algorithm is essentially optimal, in particular we show
that O(D) algorithms are impossible. These lower bounds are by reductions from the communication
complexity problem of Set Disjointness to computing SCCs in a specific graph. In particular, we
show that any algorithm that computes SCCs with o(n) symbolic one-step operations would imply a
communication protocol for the Set Disjointness problem with o(k) communication.
Reduction 3.1. Let (x, y) be an instance of Set Disjointness and let w.l.o.g. k = ` · k̄ for some
integers `, k̄. We construct a directed graph G
= (V, E) with n = k + ` vertices and O(n2 ) edges as
S`−1
follows. (1) The vertices are given by V = i=0 Vi with Vi = {vi,0 , . . . , vi,k̄ }. (2) There is an edge
from vi,j to vi0 ,j 0 if either i < i0 or i = i0 and j < j 0 . (3) For 0 ≤ i < `, 0 ≤ j < k̄ there is an edge
from vi,j+1 to vi,j iff xi·k̄+j = 0 or yi·k̄+j = 0.
In our communication protocol both Alice and Bob compute the number of SCCs on the graph
from Reduction 3.1, according to a given algorithm. While both know all the vertices of the graph,
they do not know all the edges (some depend on both x and y) and thus whenever such an edge is
relevant for the algorithm, Alice and Bob have to communicate with each other. We show that the
graph is constructed such that for each subset S ⊆ V the operations Pre(S) and Post(S) can be
computed with only four bits of communication between Alice and Bob.
Theorem 3.2. Any (probabilistic bounded error or deterministic) symbolic algorithm that computes
the SCCs of graphs with n vertices needs Ω(n) symbolic one-step operations. Moreover, for a graph
with the set SCCs(G) of SCCs and diameter D any algorithm needs Ω(|SCCs(G)| · D) symbolic
one-step operations.
5
We first show that Reduction 3.1 is a valid reduction from the Set Disjointness problem to an
SCC problem. The missing proofs are given in Section 6.1.
Lemma 3.3. f (x, y) = 1 iff the graph constructed in Reduction 3.1 has exactly ` SCCs.
The critical observation for the proof of Theorem 3.2 is that any algorithm that computes SCCs
with N many symbolic one-step operations implies the existence of a communication protocol for
Set Disjointness that only requires O(N ) communication.
Lemma 3.4. For any algorithm that computes SCCs with N symbolic one-step operations there is a
communication protocol for Set Disjointness that requires O(N ) communication.
Proof. In our communication protocol both Alice and Bob consider the graph from Reduction 3.1.
We call edges of the graph that are present independently of x and y definite edges and edges whose
presence depends on x and y possible edges.
Both Alice and Bob execute the given symbolic algorithm to decide whether the graph has `
SCCs (cf. Lemma 3.3). As both know all the vertices, they can execute set operations without
communicating. Communication is only needed when executing symbolic one-step operations, since
for these some of the possible edges might affect the outcome.
We next argue that each symbolic one-step operations can be executed with a constant number
of bits of communication. First notice that as both Alice and Bob execute the same algorithm
simultaneously, they both know the input set to an operation and they only need communication
about the possible edges that can change the output. Both can independently identify these possible
edges and they can decide whether such an edge exists by sending one bit each. We next argue that
for each one-step operation we need to consider at most two possible edges. For this we consider the
vertices vi,j in their linear ordering given by i · (k̄ + 1) + j, e.g., v0,0 = v0 and v`−1,k̄ = vk+`−1 .
Post operation: Let S be the input set and let vmin the vertex with the minimum index in S. Then
we have {vmin +1 , . . . vk+`−1 } ⊆ Post(S) and potentially also vmin and vmin −1 can be in Post(S),
but no other vertices. That is, we have {vmin +1 , . . . vk+`−1 } ⊆ Post(S) ⊆ {vmin −1 , . . . vk+`−1 }.
To decide whether vmin is in Post(S), we first check whether vmin +1 ∈ S and if so we check
whether the edge (vmin +1 , vmin ) is present. To decide vmin −1 ∈ Post(S), we check whether the
edge (vmin , vmin −1 ) is present. That is, to compute Post(S) we only access two possible edges.
Pre operation: Let S be the input set and let vmax the vertex with the maximum index in S.
Then we have {v0 , . . . vmax −1 } ⊆ Pre(S) and potentially also vmax and vmax +1 can be in Pre(S),
but no other vertices. That is, we have {v0 , . . . vmax −1 } ⊆ Pre(S) ⊆ {v0 , . . . vmax +1 }. To decide
whether vmax is in Pre(S), we first check whether vmax −1 ∈ S and if so we check whether
the edge (vmax , vmax −1 ) is present. To decide if vmax +1 ∈ Pre(S), we check whether the edge
(vmax +1 , vmax ) is present. That is, we can compute Pre(S) with accessing only two possible edges.
By the above we have that a symbolic algorithm with N one-step operations gives rise to a
communication protocol for Set Disjointness with O(N ) bits of communication.
By Lemma 3.4 we have that any algorithm computing SCCs with o(n) symbolic one-step
operations would contradict Theorem 2.1. Now inspecting the graph of Reduction 3.1, we observe
that its diameter is equal to k̄, which leads to the following lower bounds. For ` = k/2 the graph has
diameter 2 and thus the Ω(n) holds even for graphs of constant diameter. On the other side, for ` = 1
disjoint sets Sx and Sy correspond to strongly connected graphs and thus the Ω(n) lower bounds
also holds for graphs with a√bounded number of SCCs, i.e., there are no O(|SCCs(G)|) symbolic
algorithms. Finally for ` = k we obtain a Ω(|SCCs(G)| · D) lower bound.
6
v0
v1
v2
v3
v4
Figure 2: Reduction 3.6 for k = 4, ` = 2, Sx = {2, 3}, Sy = {0, 1, 3}
Remark 3.5. By the above no algorithm can compute SCCs with f (D)·no(1) or f (|SCCs(G)|)·no(1)
symbolic one-step operations for any function f . In contrast, if we consider both parameters
simultaneously, there is an O(|SCCs(G)| · D) symbolic algorithm.
The above lower bounds for computing SCCs match the O(min(n, D · |SCCs(G)|)) bound by
the algorithm of Gentilini et al. [GPP08]. One way to further improve the algorithm is to not consider
the diameter of the whole graph but the P
diameter DC of each single SCC C. In that direction the
previous reduction
already
gives
us
an
Ω(
and we will next improve
C∈SCCs(G) (DC )) lower bound
P
P
it to an Ω( C∈SCCs(G) (DC + 1)) lower bound (i.e., it is Ω(n) even if C∈SCCs(G) (DC ) ∈ O(1)).
These two bounds differ if the graph has a large number of trivial SCCs. Thus we next give a
reduction that constructs a graph that has only trivial SCCs if Sx and Sy are disjoint.
Reduction 3.6. Given an instance (x, y) of Set Disjointness, we construct a directed graph G =
(V, E) with n = k + 1 vertices and O(n2 ) edges as follows. (1) The vertices are given by V =
{v0 , v1 , . . . , vk }. (2) There is an edge from vi to vj for i < j. (3) For 0 ≤ j ≤ k − 1 there is an edge
from vj+1 to vj iff xj = 1 and yj = 1.
error or deterministic) symbolic algorithm that computes
Theorem 3.7. Any (probabilistic bounded
P
the SCCs needs Ω(|SCCs(G)| + C∈SCCs(G) DC ) symbolic one-step operations.
3.2
Lower Bounds for Liveness, Reachability, and Safety Objectives
In this section we extend our lower bounds for SCC computation to Liveness, Reachability, and
Safety Objectives on graphs.
Lower Bounds for Reachability. The lower bounds for Reachability are an immediate consequence
from our lower bounds for SCC computation in Theorem 3.2. When setting ` = 1 in Reduction 3.1
then the vertex v0,0 is reachable from all vertices iff the graph is strongly connected iff the sets Sx
and Sy are disjoint.
Theorem 3.8. Any (probabilistic bounded error or deterministic) symbolic algorithm that solves
Reachability in graphs with diameter D requires Ω(D) symbolic one-step operations.
Lower Bounds for Liveness. To show an Ω(n) lower bound for Liveness objectives which holds even
for graphs of bounded diameter, we introduce another reduction. This reduction is again from the Set
Disjointness Problem and also constructs a graph such that Pre and Post operations can be executed
with a constant number of bits of communication between Alice and Bob.
Reduction 3.9. Given an instance (x, y) of Set Disjointness, we construct a directed graph G =
(V, E) with n = k + 1 vertices and O(n2 ) edges as follows. (1) The vertices are given by V =
{v0 , v1 , . . . , vk }. (2) There is an edge from vi to vj for i < j and there is a loop edge (vk , vk ). (3)
For 0 ≤ j ≤ k − 1 there is a loop edge (vj , vj ) iff xj = 1 and yj = 1.
7
v0
v1
v2
v3
v4
Figure 3: Reduction 3.9 for k = 4, ` = 2, Sx = {2, 3}, Sy = {0, 1, 3}
Notice that the graph in Reduction 3.9 has diameter D = 1 and thus allows to show the lower
bounds stated in Theorem 3.10 when considering T = {v0 , v1 , . . . , vk−1 }, with the exception of (2)
which is by Reduction 3.1 and T = {v0 }.
Theorem 3.10. For any (probabilistic bounded error or deterministic) symbolic algorithm that solves
Büchi(T ) the following lower bounds on the required number of symbolic one-step operations hold:
(1) Ω(n) even for instances with constant D; (2) Ω(D) even for instances with P
|T | = 1; (3) Ω(|T |)
even for instances with constant D; (4) Ω(|T | + D); and (5) Ω(|SCCs(G)| + C∈SCCs(G) DC ).
Lower Bounds for co-Liveness and Safety. The following lower bounds are by Reduction 3.9 (and
variations of it) and the set of safe vertices T = {v0 , v1 , . . . , vk−1 }.
Theorem 3.11. For any (probabilistic bounded error or deterministic) symbolic algorithm that solves
Safe(T ) or coBüchi(T ) the following lower bounds on the required number of symbolic one-step
operations hold: (1)PΩ(n) even for constant diameter graphs; (2) Ω(|T |) even for constant diameter
graphs; and (3) ω( C∈SCCs(G) (DC + 1)) even for constant diameter graphs.
Notice that the parameters diameter, number of SCCs, or diameters of SCCs do not help in
the case of Safety. This is because every graph can be reduced to a strongly connected graph with
diameter 2 without changing the winning set as follows: Add a new vertex v that has an edge to and
an edge from all original vertices but do not add v to the safe vertices T .
We complete this section with a Ω(D) lower bound for coBüchi(T ) which is by a variant of
Reduction 3.1.
Proposition 3.12. Any (probabilistic bounded error or deterministic) symbolic algorithm that solves
coBüchi(T ) on graphs with diameter D needs Ω(D) symbolic one-step operations.
3.3
Lower Bound for Approximate Diameter
The Approximate Diameter Problem. Let G = (V, E) be a directed graph with n vertices V and
m edges E. Let d(u, v) denote the shortest distance from u ∈ V to v ∈ V in G, i.e., the smallest
number of edges of any path from u to v in G. Recall that we define the diameter of G as the
maximum of d(u, v) over all pairs u, v for which u can reach v in G. 1 We consider the problem of
e
approximating the diameter D of a graph by a factor c, where the goal is to compute an estimate D
e
such that D/c ≤ D ≤ D. As undirected graphs are special cases of directed graphs, the lower bound
is presented for undirected graphs and the upper bound for directed graphs (see Section 4), i.e., both
hold for undirected and directed graphs.
1
Usually the diameter is defined over all pairs of vertices, not just the reachable ones, and is therefore ∞ if G is not
strongly connected. Our definition is more general since determining whether the graph is strongly connected takes only
O(D) symbolic steps and additionally our definition is more natural in the symbolic setting as it provides an upper bound
on the number of one-step operations needed until a fixed point is reached, which is an essential primitive in symbolic
graph algorithms.
8
u
t
a0
b0
c0
a1
b1
c1
a2
b2
c2
Figure 4: Reduction 3.13 for k = 9, s = 3, Sx = {2, 4, 8}, Sy = {1, 3, 5}.
Result. We show a lower bound of Ω(n) on the number of symbolic steps needed to distinguish
between a diameter of 2 and a diameter of 3, even in an undirected connected graph. The basic
symbolic algorithm for computing the diameter exactly takes O(n · D) many symbolic steps. Thus
our lower bound is tight for constant-diameter graphs.
Outline Lower Bound. We show how to encode an instance of the Set
√ Disjointness Problem with
a universe of size k in an (undirected, connected) graph with Θ( k) vertices and Θ(k) edges
such
√ that 1) in a communication protocol any symbolic one-step operation can be simulated with
Θ( k) bits and 2) the graph has diameter 2 if the two sets are disjoint and diameter 3 otherwise.
Thus the communication
complexity lower bound of Ω(k) for Set Disjointness implies a lower
√
bound of Ω( k) = Ω(n) for the number of symbolic one-step operations to compute a (3/2 − ε)approximation of the diameter of a graph with n vertices.
√
Reduction 3.13. Let (x, y) be an instance of the Set Disjointness problem of size k and let s = k.
We construct an undirected graph G = (V, E) with n = 3s + 2 vertices and O(k) edges as follows.
(1) There are three sets A, B, C with s vertices each and two auxiliary vertices u and t. We denote
the i-th vertex of each of A, B, C with a lowercase letter indicating the set and subscript i. (2) There
is an edge between u and t and between u and each vertex of A and B and between t and each vertex
of C. (3) For each 0 ≤ i < s there is an edge between ai ∈ A and bi ∈ B. (4) For 0 ≤ ` < k let
i, j < s be such that ` = i · s + j. There is an edge between ai ∈ A and cj ∈ C iff x` = 0 and there
is an edge between bi ∈ B and cj ∈ C iff y` = 0.
We first show that this graph has diameter 2 if Sx and Sy are disjoint and diameter 3 otherwise
and then show how Alice can obtain a communication protocol for the Set Disjointness problem from
any symbolic algorithm that can distinguish these two cases.
Lemma 3.14. Let G = (V, E) be the graph given by Reduction 3.13 and let D denote its diameter.
If Sx ∩ Sy = ∅, then D = 2, otherwise D = 3.
In the graph G = (V, E) given by Reduction 3.13 Alice knows all the vertices of the graph
and all the edges except those who are constructed based on y, i.e., Alice does not know the edges
between B and C. To take into account the edges between B and C, Alice has to communicate
with Bob. To show a lower bound on the number of symbolic steps, we show next an upper bound
on the number of bits of communication between Alice and Bob needed to simulate a symbolic
one-step operation on G. With the simulation of the one-step operations, the symbolic algorithm can
be used as a communication protocol for distinguishing whether G has diameter 2 or 3 and thus by
Lemma 3.14 to solve the Set Disjointness problem. Whenever the symbolic algorithm performs a Pre
9
or Post operation (which are equivalent on undirected graphs) for a set S that contains vertices of B
or C, then Alice can simulate this one-step operation by specifying the vertices of B and C that are in
S with a bit vector of size 2s, where Bob answers with a bit vector, again of size 2s, that indicates all
vertices that are adjacent to (B ∪ C) ∩ S. Thus the communication protocol can simulate a symbolic
algorithm that performs T one-step operations with at most 4sT bits of communication. Hence
we have by Theorem 2.1 that 4sT ≥ Ω(k) = Ω(s2 ) and thus T ≥ Ω(s) = Ω(n). Together with
Lemma 3.14, this proves the following theorem. Note that any (3/2 − ε)-approximation algorithm
for the diameter of a graph can distinguish between diameter 2 and 3.
Theorem 3.15. Any (probabilistic bounded error or deterministic) symbolic algorithm that computes
a (3/2 − ε)-approximation of the diameter of an undirected connected graph with n vertices needs
Ω(n) symbolic one-step operations.
4
Upper Bounds
In this work we present the following upper bounds.
4.1
Upper Bounds for Computing Strongly Connected Components
We revisit the symbolic algorithm of Gentilini et al.P
[GPP08] that computes the SCCs and present
a refined analysis to show that it only requires O( C∈SCCs(G) (DC + 1)) symbolic operations,
improving the previously known O(min(n, D · |SCCs(G)|)) bound and matching the lower bound
of Theorem 3.7 (details in Section 6.4).
Theorem 4.1. ThePalgorithm of Gentilini et al. [GPP08] computes the SCCs of a graph G with
O(|SCCs(G)| + C∈SCCs(G) DC ) symbolic operations.
4.2
Upper Bounds for Liveness, Reachability, and Safety Objectives
The upper bounds for Reachability, Safety, Liveness, and co-Liveness, are summarized in the
following proposition, which is straightforward to obtain as discussed below.
Proposition 4.2. Let SCC be the number of symbolic steps required to compute the SCCs of
the graph and let T be the set of target/safe vertices. Then Reach(T ) can be solved with O(D)
symbolic operations; Büchi(T ) can be solved with O(min(SCC + D, |T | · D)) symbolic operations;
coBüchi(T ) can be solved with O(|T | + D) symbolic operations; and Safe(T ) can be solved with
O(|T |) symbolic operations.
Algorithm for Reachability. Given a target set T , we can easily compute the vertices that can reach T
by iteratively applying Pre operations until a fixed-point is reached. By the definition of diameter,
this requires only O(D) symbolic operations.
Algorithms for Liveness. A simple algorithm for Liveness first starts an algorithm for computing
SCCs and whenever an SCC is reported it tests whether the SCC contains one of the target vertices
and if so adds all vertices of the SCC to the winning set. Finally, after all SCCs have been processed,
the algorithm computes all vertices that can reach the current winning set and adds them to the
winning set. That is, in total the algorithm only needs O(SCC + D) many symbolic operations
where SCC is the number of symbolic operations required by the SCC algorithm. An alternative
algorithm for Liveness with O(|T | · D) symbolic operations is as follows. For each v ∈ T check
10
with O(D) symbolic one-step operations whether the vertex can reach itself and if not remove the
vertex from T . Then do a standard reachability with the updated set T as target, again with O(D)
symbolic one-step operations.
Algorithm for co-Liveness. Given a set T of safe vertices, an algorithm for co-Liveness first restricts
the graph to the vertices of T ; in the symbolic model this can be done by intersecting the outcome of
each Pre and Post operation with T . One then uses an SCC algorithm and whenever a non-trivial
SCC is reported, all its vertices are added to the winning set. Finally, after all SCCs have been
processed, all vertices that can reach the current winning set in the original graph are added to the
winning set. That is, in total the algorithm only needs O(|T | + D) many symbolic operations, where
|T | comes from the linear number of symbolic operations required by the SCC algorithm for the
modified graph.
Algorithm for Safety. Given a set T of safe vertices, an algorithm for safety first restricts the graph
to the vertices of T . One then uses an SCC algorithm and whenever a non-trivial SCCs is reported,
all its vertices are added to the winning set. Finally, after all SCCs have been processed, all vertices
that can reach, within T , the current winning set are added to the winning set. That is, in total the
algorithm only needs O(|T |) symbolic operations as both the SCC algorithm for the modified graph
and reachability in the modified graph are in O(T ). Also notice that reachability is not bounded by
O(D) as restricting the graph to vertices of T can change the diameter.
Notice that none of the above algorithms stores all the SCCs, but processes one SCC at a time.
That is, the algorithms themselves only need a constant number of sets plus the sets stored in the
algorithm for computing SCCs (which can be done with O(log n) many sets).
4.3
Upper Bounds for Approximate Diameter
We present
a (1 + ε)-approximation algorithm for the diameter (for any constant ε > 0) that takes
√
e
e
O(n
D) symbolic operations (the O-notation
hides logarithmic factors).
Theorem 4.3. A (1 + )-approximation
of the diameter of a directed graph for any constant > 0
√
e
D) symbolic operations, using O(1) many sets.
can be obtained with O(n
Technical Overview (1 + ε)-Approximation Algorithm. The symbolic algorithm is based on the
3/2-approximation algorithm by Aingworth et al. [Ain+ 99] for explicit graphs, we give a high-level
overview of the differences here. An expensive step in the algorithm of [Ain+ 99] is the computation
of s-partial BFS trees that contain s vertices closest to a root vertex v and can be determined with
O(s2 ) explicit operations (this part was later replaced and improved upon by [RW13, Che+ 14]). In
the symbolic model computing s-partial BFS trees would be less efficient, however, we can compute
all vertices with distance at most x from v with only O(x) many symbolic operations. The limitation
of the approximation ratio c to 3/2 in the algorithm of [Ain+ 99] comes from having to deal with
vertices for which less than s vertices are within distance at most D/(2c). In the symbolic model we
do not have to consider this case since with a budget of O(x) operations we can always reach at least
x vertices (assuming for now that the graph is strongly connected and x < D). Thus the algorithm
simplifies to the second part of their algorithm, whose core part is to find a vertex within distance at
most x for each vertex of the graph by using a greedy approximation algorithm for dominating set.
However, in the symbolic model storing a linear number of sets is too costly, hence we inherently use
that we can recompute vertices at distance at most x efficiently when needed. Details are presented
in Section 6.5.
11
5
5.1
Discussion
SCCs and Verification Objectives
First, our results show that the symbolic SCC algorithm by Gentilini et al. [GPP08]
is essentially optiP
mal. That is, we have the three upper bounds of O(n), O(|SCCs(G)·D|) and O( C∈SCCs(G) (DC +
lower bounds of Ω(n) (Theorem 3.2), Ω(|SCCs(G)| · D) (Theorem 3.2), and
1)) and matching
P
Ω(|SCC| + C∈SCCs(G) DC ) (Theorem 3.7).
Table 3: Results
number of symbolic operations
SCC
Θ(n)
Θ(|SCCs(G)| · D)
P
Θ(
C∈SCCs(G) (DC
+ 1))
Our results for the different kinds of verification objectives are summarized in Table 4. We
have an interesting separation between the reachability objective and the other objectives in terms
of the diameter D. While reachability can be solved with O(D) symbolic operations, all the other
objectives require Ω(n) symbolic one-step operation on graphs of constant diameter.
When considering the diameters DC of the SCCs
we get another separation. There we have that
P
Liveness and Reachability can be solved with O( C∈SCCs(G) (DC + 1)) many symbolic operations,
while Safety and co-Liveness requires Ω(n) symbolic one-step operations on strongly connected
graphs with constant diameter. This reflects the fact that in the standard algorithm for Safety and
co-Liveness the SCC computation is performed on a modified graph.
5.2
Approximate Diameter
e √n)
For explicitly represented graphs a 3/2-approximation of the diameter can be computed in O(m
time [RW13, Che+ 14], while under the strong exponential time hypothesis no O(n2−o(1) ) time
algorithm exists to distinguish graphs of diameter 2 and 3 (i.e., no (3/2 − ε)-approximation can be
e
obtained) [RW13]. The fastest exact algorithms take O(mn)
time. While for explicitly represented
graphs small, constant diameters are a hard case, the current results suggest that in the symbolic
model the diameter of graphs with constant diameter can be determined more efficiently than for large
diameters, as both the upper bound for exact and approximate computation of the diameter depend
on the diameter of the graph and are linear when the diameter is constant. While the threshold of an
approximation ratio of 3/2 appears in our (linear) lower bound, the current symbolic upper bounds do
not show this behavior. Several interesting open questions remain: Is there a o(n) c-approximation
algorithm when c ∈ [3/2, 2)? Is there a linear (1 + ε)-approximation algorithm for graphs with
super-constant diameter? Or are there better lower bounds?
6
6.1
Detailed Proofs
Proofs of Section 3.1
Proof of Lemma 3.3. We have to show that f (x, y) = 1 iff the graph constructed in Reduction 3.1
has exactly ` SCCs.
First notice that there are no edges from a set Vj to a set Vi if i < j and thus there are at least `
SCCs, independently of the actual values of x and y.
12
Table 4: Results
Objective
n
Reach(T)
Safe(T)
Büchi(T)
coBüchi(T)
Θ(n)
Θ(n)
Θ(n)
Θ(n)
number of symbolic operations in terms of
D & |T |
DC
Θ(D)
Θ(|T |)
O(|T | · D) / Ω(|T | + D)
Θ(|T | + D)
Θ(D)
P
ω( C∈SCCs(G) (DC + 1))
P
Θ( C∈SCCs(G) (DC + 1))
P
ω(
C∈SCCs(G) (DC
+ 1))
⇒: If f (x, y) = 1 then all possible edges exists and it is easy to verify that the SCCs of the graphs
are exactly the sets Vi for 0 ≤ i < `.
⇐: If f (x, y) = 0 then there are 0 ≤ j ≤ k̄, 0 ≤ i < ` such that there is no edge from vi,j+1 to vi,j .
Now the set Vi splits up in at least two SCCs and thus there are at least ` + 1 SCCs.
Proof of Theorem 3.7. We have to show that any (probabilistic bounded
error or deterministic)
P
symbolic algorithm that computes the SCCs needs Ω(|SCCs(G)| + C∈SCCs(G) DC ) symbolic
one-step operations.
P
First consider Reduction 3.1 and notice P
that for the constructed graph C∈SCCs(G) DC ∈ Θ(n).
Then by Theorem 3.2 we already have a Ω( C∈SCCs(G) (DC )) bound. Now consider Reduction 3.6
and notice that the constructed graph has n SCCS iff x and y are disjoint. Now we can use the same
argument as in the proof of Theorem 3.2 that each symbolic one-step operations
just needs constant
P
communication. The instances where x and y are disjoint have n SCCs and C∈SCCs(G) DC = 0.
Hence an algorithm with o(|SCCs(G)|) symbolic one-step operations would imply a communication
protocol with o(k) communication, a contradiction
to Theorem 2.1. By combining the two lower
P
bounds we get the desired Ω(|SCCs(G)| + C∈SCCs(G) DC ) bound.
6.2
Proofs of Section 3.2
Proof of Theorem 3.8. We have to show that any (probabilistic bounded error or deterministic)
symbolic algorithm that solves Reachability in graphs with diameter D requires Ω(D) symbolic
one-step operations.
Consider the graph from Reduction 3.1 with parameter ` = 1. We have that v0 is reachable from
all vertices iff the graph is strongly connected. From the proof of Theorem 3.2 we have that testing
whether the graph is strongly connected requires Ω(k) symbolic one-step operations. Now notice that
(a) the graph is strongly connected iff vk can reach v0 and that (b) if the graph is strongly connected
then D = k.
Proof of Theorem 3.10. For (1) & (3) consider the graph constructed in Reduction 3.9 and the target
set T = {v0 , v1 , . . . , vk−1 }. We have a valid reduction from the Set Disjointness problem as the vertex
v0 is winning for Büchi(T ) iff there is a loop for one of the vertices in T iff Sx ∩ Sy 6= ∅. By the same
argument as in the proof of Theorem 3.2 we have that each symbolic one-step operations just needs
constant communication. Hence an algorithm with o(n), o(|T |) or o(|SCCs(G)|) symbolic one-step
operations would imply a communication protocol with o(k) communication, a contradiction.
For (2) consider the graph constructed in Reduction 3.1 with ` = 1 and the target set T = {v0 }.
It is easy to verify that the vertex vk is winning if it can reach
T . Thus the Ω(D) lower bound for
P
reachability also applies here. Notice that this also gives a Ω( C∈SCCs(G) DC ) lower bound.
13
The Ω(|T |+D) lower bound in (4) is a direct consequence of the Ω(D) lower bound for instances
with constant size target sets T , and the Ω(|T |) lower bound for instances with constant
diameter D.
P
Finally, (5) is by the Ω(|SCCs(G)|) bound from Reduction 3.9 and the Ω( C∈SCCs(G) DC )
bound by Reduction 3.1 with ` = 1.
Proof of Theorem 3.11. 1) & 2) Consider the graph constructed in Reduction 3.9 and the set of safe
vertices T = {v0 , v1 , . . . , vk−1 }. We have a valid reduction from the Set Disjointness problem as the
vertex v0 is winning for Safe(T ) iff there is a loop for one of the vertices in T iff Sx ∩ Sy 6= ∅. By the
same argument as in the proof of Theorem 3.2 we have that each symbolic one-step operations just
needs constant communication. Thus an algorithm with o(n) or o(|T |) symbolic one-step operations
would imply a communication protocol with o(k) communication, a contradiction.
3) Consider the graph constructed in Reduction 3.9 but replace the edge (vk , vk ) by the edge
(vk , v0 ). When considering the set of safe vertices T = {v0 , v1 , . . . , vk−1 } the same arguments as
above apply and thus we get a Ω(n)P
lower bound. However, the graph is strongly connected and has
diameter 2 and thus |SCCs(G)| + C∈SCCs(G) DC = O(1).
Proof of Proposition 3.12. Consider the graph constructed in Reduction 3.1 with ` = 1, add an
additional loop edge (v0 , v0 ), and consider the set T = {v0 } of safe vertices, i.e., v0 is the only safe
vertex. It is easy to verify that the vertex vk is winning in coBüchi(T ) if it can reach T . Thus the
Ω(D) lower bound for reachability also applies here.
6.3
Proofs of Section 3.3
Proof of Lemma 3.14. Let G = (V, E) be the graph given by Reduction 3.13 and let D denote its
diameter. We have to show that if Sx ∩ Sy = ∅, then D = 2, otherwise D = 3.
First note that through the edges adjacent to the auxiliary vertices the diameter of G is at most 3.
Furthermore we have for all 0 ≤ i, j < s that d(ai , bj ) ≤ 2, d(u, cj ) = 2, and d(t, cj ) = 1, for
all i 6= j additionally d(ai , aj ) = d(bi , bj ) = d(ci , cj ) = 2, and for all v ∈ A ∪ B it holds that
d(u, v) = 1 and d(t, v) = 2. Thus whether D is 2 or 3 depends only on the maximum over all
0 ≤ i, j < s of d(ai , cj ) and d(bi , cj ).
If Sx ∩ Sy = ∅, then for each pair of indices 0 ≤ i, j < s at least one of the edges (ai , cj ) and
(bi , cj ) exists. Since (ai , bi ) ∈ E, we have for all 0 ≤ j < s and all v ∈ A ∪ B that d(v, cj ) ≤ 2 and
hence D = 2.
If Sx ∩ Sy 6= ∅, let ` ∈ Sx ∩ Sy and let 0 ≤ i, j < s be such that ` = i · s + j. Then neither the
edge (ai , cj ) nor the edge (bi , cj ) exists. Thus the vertex ai , and analogously the vertex bi , has edges
to the following vertices only: the auxiliary vertex u, the vertex bi , and vertices cj 0 ∈ C with j 0 6= j.
The vertex cj has edges only to the auxiliary vertex t and to vertices ai0 ∈ A and bi0 ∈ B with i0 6= i.
Hence none of the vertices adjacent to ai , and respectively for bi , is adjacent to cj and thus we have
D = d(ai , cj ) = d(bi , cj ) = 3.
6.4
An Improved Upper Bound for Strongly Connected Components
Result. Gentilini et al. [GPP08] provide a symbolic algorithm for computing the strongly connected
components (SCCs) and show a bound of O(min(n, D · |SCCs(G)|)) on the number of its symbolic
operations for a directed graph with n vertices, diameter D, and |SCCs(G)| many SCCs. Let DC
be the diameter of an P
SCC C. We give a tighter analysis of the algorithm of [GPP08] that shows
an upper bound of O( C∈SCCs(G)
(DC + 1)) symbolic operations that matches
our lower bound
P
P
(Theorem 3.7). We have both C∈SCCs(G) (DC +1) ≤ (D+1)·|SCCs(G)| and C∈SCCs(G) (DC +
14
1) ≤ n + |SCCs(G)| ≤ 2n and thus our upper bound is always at most the previous one. We
additionally observe that the algorithm can be implemented with O(log n) many sets (when the
SCCs are output immediately and not stored). We first explain the intuition behind the algorithm of
[GPP08] and then present the improved analysis of its number of symbolic steps.
Symbolic Breadth-First Search. While explicit algorithms for SCCs are based on depth-first search
(DFS), DFS is impractical in the symbolic model. However, breadth-first search (BFS) from a set
U ⊆ V can be performed efficiently symbolically, namely proportional to its depth, as defined below.
Definition 6.1 (Symbolic BFS). A forward search from a set of vertices U = U0 is given by a
sequence of Post operations such that Ui = Ui−1 ∪ Post(Ui−1 ) for i > 0 until we have Ui = Ui−1 .
We call Ui \ Ui−1 the i-th level of the forward search and the index of the last non-empty level the
→
−
→
depth B (U ) of the forward search. Let F W (U ) = U−
be the forward set, which is equal to
B (U )
←
−
the vertices reachable from U . Analogously we define the backward search of depth B (U ) and the
backward set BW (U ) for Pre operations. We denote a singleton set U = {u} by u.
There is a simple algorithm for computing the SCCs symbolically with BFS that takes O(D ·
|SCCs(G)|) many symbolic steps: Start with an arbitrary vertex v. Compute the SCC containing
v by taking the intersection of F W (v) and BW (v), remove the obtained SCC from the graph, and
repeat.
Skeleton-based Ordering. The importance of DFS for SCCs lies in the order in which the SCCs are
computed. Starting from a vertex v that lies in an SCC without outgoing edges (i.e. a sink in the
DAG of SCCs of the graph), the forward search does not leave the SCC and for computing SCCs the
backward search can be restricted to the vertices of F W (v), i.e., the SCC of v can be determined
with proportional to the diameter of the SCC many symbolic steps. The DFS-based SCC algorithm of
[Tar72] finds such an SCC first. The algorithm of [GPP08] is based on an ordering obtained via BFS
that achieves a DFS-like ordering suitable for computing SCCs symbolically. Our tighter analysis
essentially shows that their approach achieves the best ordering we can hope for. The ordering is
given by so-called skeletons.
Definition 6.2. A pair (S, v) with v ∈ V and S ⊆ V is a skeleton of F W (u) for u ∈ V if v has
maximum distance from u and the vertices of S form a shortest path from u to v.
Let SCC(v) denote the SCC containing v ∈ V . The SCCs of the vertices of S will be computed in
the following order: First SCC(u) is computed by performing a backward search from u within
F W (u). The remaining SCCs of of the vertices of S are then computed in the reverse order of the
path induced by S, starting with SCC(v). We now describe the overall algorithm, where in addition
to SCC(S, v) the SCCs of V \ F W (u) are computed (potentially using a different, previously
computed skeleton).
The Algorithm. The pseudo-code of the algorithm is given in Algorithm SCC-Find, the pseudo-code
for the sub-procedure for computing a forward set including a skeleton in Algorithm Skel-Forward.
Processing a graph G, the algorithm proceeds as follows: it starts from some vertex v and computes
the set of reachable vertices, i.e, the forward set F W (v), including a skeleton that includes exactly
one vertex of each level of the forward search and forms a shortest path in G. It then starts a backward
search starting from v in the subgraph induced by F W (v). Clearly the SCC of v is given by the
vertices that are reached by the backward search. The algorithm returns this SCC as an SCC of G and
recurses on (a) the subgraph GV \F W (v) induced by the vertices of V \ F W (v) and (b) the subgraph
15
Algorithm SCC-Find: Symbolic SCC Algorithm
Input : Graph G = (V, E), Skeleton (S, v)
1
2
3
4
5
if V = ∅ then
return ∅;
if S = ∅ then
v ← P ick(V ) ;
vertex
// If there is no skeleton, pick arbitrary
(F W, S 0 , v 0 ) ← Skel_Forward(V, E, v) ;
skeleton
8
/* Compute the SCC containing v
SCC ← {v};
while (Pre(SCC) ∩ F W ) \ SCC 6= ∅ do
SCC ← SCC ∪ (Pre(SCC) ∩ F W )
9
output SCC as an SCC;
6
7
// Forward search incl.
11
/* Recursive calls
SCC-Find(GV \F W , (S \ SCC, (pre(SCC ∩ S) \ SCC) ∩ S));
SCC-Find(GF W \SCC , (S 0 \ SCC, v 0 ));
12
return SCCs;
10
*/
*/
GF W (v)\SCC(v) induced by the vertices of F W (v) \ SCC(v). For the recursion on (a) we update a
potentially already existing skeleton by removing all vertices that are in the current SCC (initially
we have an empty skeleton) while for the recursion on (b) we use the skeleton computed by the
forward search (but also remove vertices of the current SCC). The skeleton is then used to select the
starting vertex in the consecutive steps of the algorithm: when the algorithm is called with skeleton
(S, v), then the forward search is started from v; when the skeleton was computed in a forward search
from a vertex u, then this corresponds to the vertex of the skeleton that is furthest away from u and
contained in this recursive call.
A Refined Analysis. The correctness of the algorithm is by [GPP08]. Notice that the algorithm would
be correct even without the usage of skeletons but the skeletons are necessary to make it efficient, i.e.,
to avoid unnecessarily long forward searches. We show the following theorem.
Theorem 6.3 (RestatementP
of Theorem 4.1). With input (G, ∅) Algorithm SCC-Find computes the
SCCs of G and requires O( C∈SCCs(G) (DC + 1)) symbolic operations.
The analysis of [GPP08] of the number of symbolic steps of Algorithm SCC-Find uses that (1) each
vertex is added to at most two skeletons, (2) the steps of the forward searches can be charged to the
vertices in the skeletons, and (3) backward searches are only performed to immediately identify an
SCC and thus can be charged to the vertices of the SCC; hence both the steps of the forward and of
the backward searches can be bounded with O(n). For the backward searches it can easily be seen
that the number of Pre operations to identify the SCC C is also bounded by DC + 1. For the forward
searches we show that each part of the skeleton (that in turn is charged for the forward search) can be
charged to DC + 1 for some SCC C; this in particular exploits that skeletons are shortest paths (in
the graph in which they are computed).
16
Algorithm Skel-Forward: Skeleton Forward Search Algorithm
Input : Graph G = (V, E), Node v
Output : F W Set of vertices reachable from v;
(S 0 , v 0 ) Skeleton for F W
1
F W ← ∅; i ← 0; LEVEL[0] ← {v};
/* Forward Search
while LEVEL[i] 6= ∅ do
F W ← F W ∪ LEVEL[i];
i ← i + 1;
LEVEL[i] ← Post(LEVEL[i − 1]) \ F W ;
*/
*/
11
/* Compute Skeleton
i ← i − 1;
v 0 ← P ick(LEVEL[i]);
S 0 ← {v 0 };
while i ≥ 1 do
i ← i − 1;
S 0 ← S 0 ∪ {P ick(Pre(S 0 ) ∩ LEVEL[i])};
12
return (F W, S 0 , v 0 );
2
3
4
5
6
7
8
9
10
Lemma 6.4 ([GPP08]). For each recursive call of SCC-Find with input G, (S, v) we have that S is
a set of vertices that induces a shortest path in the graph G and v is the last vertex of this path.
We first recall the result from [GPP08] that shows that the number of symbolic operations in an
execution of Skel-Forward is proportional to the size of the computed skeleton.
Lemma 6.5 ([GPP08]). Skel-Forward only requires O(S 0 ) symbolic operations, i.e., is linear in the
output, and can be implemented using only constantly many sets.
Proof. For each level of the forward search we need one Post operation in the first while loop and
one Pre operation in the second while loop, and the number of set operations is in the order of
one-step operations. As for each level we add one vertex to S 0 , the result follows. Moreover, there is
no need to explicitly store all the levels as they can be easily recomputed from the next level when
needed, increasing the number of symbolic operations only by a constant factor.
Remark 6.6. Given Lemma 6.5 we can implement SCC-Find using O(log n) many sets at a time by
recursing on the smaller of the two sub-graphs GV \F W and GF W \SCC first.
We split the cost of O(S 0 ) symbolic operations for Skel-Forward into two parts: the part S 0 ∩
SCC(v) where SCC(v) is the SCC identified in this level of recursion and the part S 0 \ SCC(v)
that is passed to one of the recursive calls. The following lemma shows that the first part and the
subsequent backward search can be charged to DSCC(v) + 1, using that S 0 is a shortest path in
F W (v).
Lemma 6.7. Without accounting for the recursive calls, each call of SCC-Find takes O(DSCC(v) +
1 + |S 0 \ SCC(v)|) symbolic operations, where S 0 is the new skeleton computed by Skel-Forward.
17
Proof. By Lemma 6.5, the call to Skel-Forward takes O(|S 0 |) symbolic operations. In the i-th
iteration of the while loop (Line 7) we add those vertices of SCC(v) that can reach v in i steps.
That is, the loop terminates after DSCC(v) iterations and thus only requires O(DSCC(v) + 1) many
symbolic operations. All the other steps just need a constant number of symbolic operations. That
is, we have an upper bound of O(DSCC(v) + 1 + |S 0 |). Now as S 0 induces a path starting at
v in F W (v), we have that whenever a vertex u 6= v of S 0 can reach v, then also all vertices
on the path from v to u can reach v and are therefore in the same SCC as v. Since the path is
a shortest path, also every sub-path is a shortest path and thus we have that |S 0 ∩ SCC(v)| ≤
DSCC(v) + 1, i.e., |S 0 | ∈ O(DSCC(v) + 1 + |S 0 \ SCC(v)|). Hence we obtain the desired bound of
O(DSCC(v) + 1 + |S 0 \ SCC(v)|) for the number of symbolic operations.
Note that at each level of recursion we only charge the diameter of the SCC that is output and
the vertices of the newly computed skeleton. Thus we do not charge vertices of a skeleton again
until they are contained in an SCC that is identified. The following lemma shows that in this case we
can charge the symbolic steps that were charged to the vertices of the skeleton to DC + 1, where C
is the SCC the part of the skeleton belongs to. Notice that S \ SCC(v) is the skeleton for the first
recursive call and S 0 \ SCC(v) is the skeleton for the second recursive call, i.e., all vertices of a
skeleton
are finally assigned to an SCC. That is, we can bound the total number of symbolic steps by
P
O( C∈SCCs(G) (DC + 1)).
Lemma 6.8. Whenever SCC-Find is called for a graph H and a skeleton (S, v), then |S∩SCC(v)| ≤
DSCC(v) + 1 .
Proof. By Lemma 6.4 the set S induces a shortest path in H that ends at v. Thus if v can reach a
vertex u 6= v of S, then it can also reach all vertices of S that are on the path from u to v and all
vertices on this sub-path are in the same SCC as v. Furthermore, the sub-path is a shortest path as
well and thus the vertices |S ∩ SCC(v)| form a shortest path in SCC(v) and hence the diameter
DSCC(v) of SCC(v) is at least |S ∩ SCC(v)| − 1.
√
f
6.5 (1 + ε)-Approximation of Diameter with O(n
D) Symbolic Operations
→
−
Notation. Given a vertex u ∈ V , let N x (u) denote the vertices with distance at most x from u and let
←
−
→
−
N y (u) be the set of vertices with distance at most y to u. We have that {u} ∪ Post({u}) = N 1 (u)
→
−
→
−
→
−
and N x (u) ∪ Post( N x (u)) = N x+1 (u). The maximum distance from the vertex u to any other
→
−
→
−
→
−
vertex v ∈ V is given by the smallest x for which N x (u) ∪ Post( N x (u)) = N x (u). Note
that x is at most D and that computing x in this way corresponds to performing a breadth-firstsearch (BFS) from x on explicitly represented graphs; thus following [Ain+ 99], we denote the
→
−
→
−
→
−
→
−
smallest x for which N x (u) ∪ Post( N x (u)) = N x (u) with B (u) and the smallest y for which
←
−
←
−
←
−
←
−
N y (u) ∪ Pre( N y (u)) = N y (u) with B (u). The set of vertices reachable from u ∈ V is given by
→
−
→
N−
(u).
B (u)
The Basic Exact Algorithm. The maximum d(u, v) over all pairs u, v ∈ V for which u can reach v
→
−
→
−
can be computed by taking the maximum of B (u) over all u ∈ V . Computing B (u) for all u ∈ V
takes O(n · D) many Post operations. To obtain only the value of D, only a constant number of sets
have to be stored. Note that this basic algorithm only uses a linear number of symbolic operations for
graphs with constant diameter. See Section 3.3 for a matching lower bound for this case.
18
A Simple 2-Approximation Algorithm. If the graph G is strongly connected, then a 2-approximation
→
−
←
−
of D is given by ( B (u) + B (u))/2 for any vertex u ∈ V . This follows from the triangle inequality
and takes O(D) many symbolic steps to compute.
e
Result. We present an algorithm that computes an
√estimate D of the diameter D of the input graph G
e
such that D ∈ [D − x, D] for a parameter
log n) symbolic steps and
√x ≤ D and takes O(n · D/x √
uses a constant number of sets. For x = D this implies a bound of O(n D log n) on the number
of symbolic steps and an approximation
√ guarantee that is better than a (1 + ε)-approximation for
any constant ε > 0 (here we assume D ≥ (1 + ε)/ε; otherwise D is constant anyway and thus the
exact algorithm only takes O(n) symbolic steps). To pick the parameter x correctly, one can use the
2-approximation algorithm if the graph is strongly connected or use doubling search at the cost of an
additional factor of log n.
Searching from Neighborhood. Let a and b be two vertices with maximum distance in G, i.e.,
d(a, b) = D. We start with the simple observation that it is sufficient to determine the depth of a BFS
from a vertex with distance at most x from a to obtain an estimate that is at most x smaller than D.
→
−
Observation 6.9 (see also [Ain+ 99]). Let a, b ∈ V be such that d(a, b) = D. Then B (v) ≥ D − x
→
−
←
−
←
−
for v ∈ N x (a) and B (u) ≥ D − y for u ∈ N y (b).
→
−
Thus to obtain an estimate for the diameter, it is certainly sufficient to find a vertex u in N x (v) for
→
−
every vertex v ∈ V and compute B (u) for all these vertices. If the graph is not strongly connected,
→
−
it can happen that some vertices v can not reach x vertices and hence N x (v) might contain less than
→
−
→
−
x vertices. In this case we know that B (v) < D. Thus it also suffices to find a vertex u in N x (v) for
→
−
every vertex v ∈ V for which | N x (v)| ≥ x; we denote this set of vertices with Vx .
→
−
Corollary 6.10. Let S be a set of vertices such that S ∩ N x (v) 6= ∅ for all v ∈ Vx for x < D. Let
−
e = maxu∈S →
e ∈ [D − x, D]. Given S, computing D
e takes O(|S| · D) symbolic
D
B (u). Then D
operations and storing O(1) many sets.
Dominate each Neighborhood. An out-dominating set for a set of vertices A ⊆ V contains for each
→
−
vertex of A either the vertex or one of its successors. Finding a set S ⊆ V such that S ∩ N x (v) 6= ∅
for all v ∈ Vx is equivalent to find an out-dominating set for all vertices with degree at least x
in the following graph: Let Ĝ be the graph obtained from G by adding an edge from v to each
→
−
vertex of N x (v) \ {v} for all v ∈ V . In Ĝ every vertex of Vx has out-degree at least x. Thus
→
−
an out-dominating set for Vx in Ĝ contains a vertex of N x (v) for all v ∈ Vx , i.e., for all vertices
→
−
v ∈ V with | N x (v)| ≥ x. We adopt the classical greedy algorithm for dominating set to compute
an out-dominating set in Ĝ with the following guarantees. We prove Lemma 6.11 in the following
subsection.
Lemma 6.11. An out-dominating set S for the vertices of Vx in Ĝ with |S| ∈ O(n/x · log n) can be
found with O(n · x · log n) symbolic operations on G, storing O(1) many sets.
Overall Algorithm. Hence our algorithm is as follows. First we find a set S of size O(n/x · log n)
→
−
that contains a vertex of N x (v) for every v ∈ Vx in O(n · x · log n) symbolic steps (Lemma 6.11).
→
−
Then we compute B (u) for all u ∈ S with O(n · D/x · log n) many symbolic steps and return the
→
−
maximum value of B (u) that was found (Corollary 6.10). Together with the observations at the
e
beginning of this section we obtain the following theorem. The O-notation
hides the logarithmic
factors.
19
Algorithm Dominating Set: Algorithm for Out-Dominating Set in Ĝ
Input : Graph G = (V, E), parameter x
→
−
→
−
Output : Set S ⊆ V that contains a vertex of N x (v) for all v with | N x (v)| ≥ x
1
2
3
4
5
6
7
8
9
10
11
S ←∅;
/* dominating set */
C ←∅;
/* covered vertices */
j ← blog2 nc ;
/* size threshold */
for v ∈ V do
/* don’t have to cover vertices that reach < x
vertices */
→
−
if | N x (v)| < x then
C ← C ∪ {v};
while j ≥ 0 do
for v ∈ V \ S do
←
−
if | N x (v) \ C| ≥ 2j then
S ← S ∪ {v};
←
−
C ← C ∪ N x (v);
j ← j − 1;
12
13
return S;
Theorem 6.12 (Restatement of Theorem 4.3). A (1 + )-approximation
of the diameter of a directed
√
e
graph for any constant > 0 can be obtained with O(n
D) symbolic operations, using O(1) sets.
6.5.1
Proof of Lemma 6.11
A fractional out-dominating set of a set A ⊆ V is a function that assigns a weight wv ∈ [0, 1] to each
v ∈ V such that for every v ∈ A the sum of the weights over v and its successors is at least one.
The size of a fractional out-dominating set is the sum of all weights wv . For Lemma 6.11 we want
to obtain an out-dominating set of Vx . The vertices of Vx have out-degree at least x in Ĝ. Thus a
fractional out-dominating set of Vx in Ĝ is obtained by assigning each vertex a weight of 1/x. The
size of this fractional out-dominating set is O(n/x). We show a greedy algorithm that finds an outdominating set of Vx in Ĝ of size within a logarithmic factor of the optimal fractional out-dominating
set, i.e., of size O(n/x · log n). The greedy algorithm is given in Algorithm Dominating Set and
is a modification of the greedy algorithm by [Joh74, Lov75, Chv79] using an idea from [BV14].
We first describe the algorithm and show that it takes O(n · x · log n) symbolic steps to output an
out-dominating set of Vx and then prove that the size of the obtained out-dominating set for Vx is
within O(log n) of the optimal fractional solution.
Algorithm Dominating Set takes the graph G = (V, E) and the parameter x as input. Constructing
→
−
the graph Ĝ, i.e., storing the sets N x (v) for all v ∈ V , is too costly. Note that there is a one-to-one
correspondence between the edges of Ĝ and the paths of length ≤ x in G and that the union of a
→
−
vertex v with its successors in Ĝ is given by N x (v) and the union of a vertex v with its predecessors
←
−
←
−
in Ĝ is given by N x (v). We recompute the sets N x (v) in O(x) symbolic steps per set when needed
by the algorithm.
←
−
Observe that the set S is an out-dominating set of Vx in Ĝ if and only if ∪v∈S N x (v) ⊇ Vx . We
20
←
−
say that the vertices of ∪v∈S N x (v) are covered by the set S. In the algorithm the set S denotes
the set of vertices added to the out-dominating set so far and the set C denotes the vertices that are
covered by S; we additionally add the vertices of V \ Vx to C, as they do not have to be covered
(lines 4–6).
The main part of the algorithm consists of a while-loop with O(log n) many iterations. The
variable j is initialized with blog2 nc and is decreased after each iteration of the while-loop; in the
last iteration of the while-loop we have j = 0. In each iteration every vertex that is not yet in S
←
−
is considered one after the other. For each vertex v ∈ V \ S the set N x (v) is computed and the
←
−
vertex is added to S if the set N x (v) contains more than 2j vertices that are not yet covered. When
←
−
a vertex v is added to S, the vertices of N x (v) are marked as covered by adding them to C. In the
last iteration we have 2j = 1 and thus all vertices of Vx that were not covered yet are added to S.
Hence the returned set S is an out-dominating set of Vx . In each iteration of the while-loop O(n · x)
symbolic operations are used, thus the algorithm takes O(n · x log n) symbolic steps in total, storing
O(1) many sets at a time.
It remains to show that the size of S is within a factor of O(log n) of the size of an optimal
fractional out-dominating set S ∗ of Vx in Ĝ with weights wv ∈ [0, 1] for all v ∈ V . The outline of
the proof is as follows: For each vertex v that the greedy algorithm adds to S, we charge a total
cost ≥ 1 to the weights wu and show that the total cost charged to wu for each u ∈ V is at most
2 · H∆u · wu , where Hi ∈ O(log i) is the i-th harmonic number and ∆u is the in-degree of u in Ĝ.
This implies that the number of vertices in S is bounded by the sum of 2 · wu · Hn over all u ∈ V ,
which proves the claim.
We charge the weights when adding v to S in Algorithm Dominating Set as follows: For each
←
−
vertex of N x (v) \ C (i.e. the newly covered vertices) we consider all vertices u that contribute
←
−
to the cover of this vertex and charge them wu /| N x (v) \ C|. Note that all vertices of V \ Vx are
←
−
←
−
contained in C and thus N x (v) \ C ⊆ Vx , hence each vertex of N x (v) \ C is covered by the optimal
←
−
fractional out-dominating set with a weight of at least one. Thus this charges at least 1/| N x (v) \ C|
←
−
per vertex of N x (v) \ C and hence at least one per vertex added to S. We have that for a fractional
←
−
out-dominating set the set N x (u) is the set of vertices to whose covering the weight wu contributes.
Thus the charge for vertex u when adding v is given by
←
−
←
−
N x (v) \ C ∩ N x (u)
· wu .
←
−
N x (v) \ C
We finally show that each vertex v is charged at most wv · Hn+1 . Let j 0 be the value of j when v
←
−
0
is added to S, i.e., | N x (v) \ C| ≥ 2j . Note that if a vertex u is charged a non-zero amount, then it is
not contained in C and therefore not in S. Hence we have that u was not added to S in the previous
←
−
0
iteration of the while-loop and thus by the greedy condition | N x (u) \ C| ≤ 2j +1 . Hence whenever
u is charged for a vertex v, we have
←
−
−
1 ←
N x (v) \ C ≥
N x (u) \ C ,
2
and thus
←
−
←
←
←
−
−
−
N x (v) \ C ∩ N x (u)
2 · N x (v) \ C ∩ N x (u)
· wu ≤
· wu .
←
−
←
−
N x (v) \ C
N x (u) \ C
21
Now consider the vertex u over the whole algorithm. The vertex u is charged for each vertex in
←
−
←
−
N x (u) at most once. By the above it is charged at most 2wu /| N x (u)| for the first vertex it is charged
←
−
←
−
for, at most 2wu /(| N x (u)| − 1) for the second vertex, and at most 2wu /(| N x (u)| − i + 1) for the
←
−
i-th vertex. Thus a vertex u with | N x (u)| = ∆u is charged at most 2wu · H∆u ∈ O(wu log(∆u )),
and hence we obtain an O(log(n)) approximation of S ∗ .
Acknowledgements. All authors are partially supported by the Vienna Science and Technology
Fund (WWTF) through project ICT15-003. K. C. is partially supported by the Austrian Science Fund
(FWF) NFN Grant No S11407-N23 (RiSE/SHiNE) and an ERC Start grant (279307: Graph Games).
V. L. is partially supported by the ISF grant #1278/16 and an ERC Consolidator Grant (project MPM).
For W. D. and M. H. the research leading to these results has received funding from the European
Research Council under the European Union’s Seventh Framework Programme (FP/2007-2013) /
ERC Grant Agreement no. 340506.
References
[Ain+ 99]
D. Aingworth, C. Chekuri, P. Indyk, and R. Motwani. “Fast Estimation of Diameter and Shortest
Paths (Without Matrix Multiplication)”. In: SIAM J. Comput. 28.4 (1999). Announced at SODA’96,
pp. 1167–1181 (cit. on pp. 11, 18, 19).
[BV14]
A. Badanidiyuru and J. Vondrák. “Fast algorithms for maximizing submodular functions”. In:
SODA. 2014, pp. 1497–1514 (cit. on p. 20).
[Bar+ 04]
Z. Bar-Yossef, T. S. Jayram, R. Kumar, and D. Sivakumar. “An information statistics approach to
data stream and communication complexity”. In: J. Comput. Syst. Sci. 68.4 (2004). Announced at
FOCS’02, pp. 702–732 (cit. on p. 5).
[Bie+ 03]
A. Biere, A. Cimatti, E. M. Clarke, O. Strichman, and Y. Zhu. “Bounded model checking”. In:
Advances in Computers 58 (2003), pp. 117–148 (cit. on p. 2).
[BGS06]
R. Bloem, H. N. Gabow, and F. Somenzi. “An Algorithm for Strongly Connected Component
Analysis in n log n Symbolic Steps”. In: Form. Methods Syst. Des. 28.1 (2006). Announced at
FMCAD’00, pp. 37–56 (cit. on p. 2).
[Bry86]
R. E. Bryant. “Graph-Based Algorithms for Boolean Function Manipulation”. In: IEEE Trans.
Comput. C-35.8 (1986), pp. 677–691 (cit. on p. 1).
[Bry92]
R. E. Bryant. “Symbolic Boolean Manipulation with Ordered Binary-decision Diagrams”. In:
ACM Comput. Surv. 24.3 (1992), pp. 293–318 (cit. on p. 1).
[Bur+ 90]
J. R. Burch, E. M. Clarke, K. L. McMillan, D. L. Dill, and L. J. Hwang. “Symbolic Model
Checking: 10ˆ20 States and Beyond”. In: LICS. 1990, pp. 428–439 (cit. on p. 1).
[Cha+ 13]
K. Chatterjee, M. Henzinger, M. Joglekar, and N. Shah. “Symbolic algorithms for qualitative
analysis of Markov decision processes with Büchi objectives”. In: Form. Methods Syst. Des. 42.3
(2013). Announced at CAV’11, pp. 301–327 (cit. on p. 1).
[Che+ 14]
S. Chechik, D. H. Larkin, L. Roditty, G. Schoenebeck, R. E. Tarjan, and V. Vassilevska Williams.
“Better Approximation Algorithms for the Graph Diameter”. In: SODA. 2014, pp. 1041–1052
(cit. on pp. 11, 12).
[Chv79]
V. Chvatal. “A Greedy Heuristic for the Set-Covering Problem”. In: Mathematics of Operations
Research 4.3 (1979), pp. 233–235 (cit. on p. 20).
[Cla+ 96]
E. M. Clarke, K. L. McMillan, S. V. Aguiar Campos, and V. Hartonas-Garmhausen. “Symbolic
Model Checking”. In: CAV. 1996, pp. 419–427 (cit. on p. 1).
22
[Cla+ 03]
E. Clarke, O. Grumberg, S. Jha, Y. Lu, and H. Veith. “Counterexample-guided Abstraction
Refinement for Symbolic Model Checking”. In: J. ACM 50.5 (Sept. 2003). Announced at CAV’00,
pp. 752–794 (cit. on p. 1).
[CGP99]
E.M. Clarke, O. Grumberg, and D. Peled. “Symbolic Model Checking”. In: Model Checking. MIT
Press, 1999. ISBN: 9780262032704 (cit. on p. 1).
[GPP08]
R. Gentilini, C. Piazza, and A. Policriti. “Symbolic Graphs: Linear Solutions to Connectivity
Related Problems”. In: Algorithmica 50.1 (2008). Announced at SODA’03, pp. 120–158 (cit. on
pp. 1, 2, 5, 7, 10, 12, 14–17).
[HW07]
J. Håstad and A. Wigderson. “The Randomized Communication Complexity of Set Disjointness”.
In: Theory of Computing 3.1 (2007), pp. 211–219 (cit. on p. 5).
[Joh74]
D. S. Johnson. “Approximation algorithms for combinatorial problems”. In: J. Comput. System
Sci. 9 (1974), pp. 256–278 (cit. on p. 20).
[KS92]
B. Kalyanasundaram and G. Schnitger. “The Probabilistic Communication Complexity of Set
Intersection”. In: SIAM J. Discrete Math. 5.4 (1992), pp. 545–557 (cit. on p. 5).
[KN97]
E. Kushilevitz and N. Nisan. Communication Complexity. New York, NJ, USA: Cambridge
University Press, 1997 (cit. on pp. 4, 5).
[Lov75]
L. Lovász. “On the ratio of optimal integral and fractional covers”. In: Discrete Math. 13 (1975),
pp. 383–390 (cit. on p. 20).
[MP92]
Z. Manna and A. Pnueli. The Temporal Logic of Reactive and Concurrent Systems: Specification.
New York: Springer-Verlag, 1992 (cit. on p. 2).
[Raz92]
A. A. Razborov. “On the Distributional Complexity of Disjointness”. In: Theor. Comput. Sci. 106.2
(1992), pp. 385–390 (cit. on p. 5).
[RW13]
L. Roditty and V. Vassilevska Williams. “Fast approximation algorithms for the diameter and
radius of sparse graphs”. In: STOC. 2013, pp. 515–524 (cit. on pp. 11, 12).
[Som99]
F. Somenzi. “Binary Decision Diagrams”. In: Calculational System Design. Ed. by M. Broy and
R. Steinbrüggen. F]: [Nato ASI series. IOS Press, 1999, pp. 303–366 (cit. on p. 1).
[Tar72]
R. E. Tarjan. “Depth First Search and Linear Graph Algorithms”. In: SIAM Journal of Computing
1.2 (1972), pp. 146–160 (cit. on pp. 5, 15).
23
| 8cs.DS
|
Max-affine estimators for convex stochastic programming
arXiv:1609.06331v1 [math.OC] 18 Sep 2016
Gábor Balázs and Csaba Szepesvári
Department of Computing Science, University of Alberta,
Edmonton, Alberta, T6G 2E8, Canada
András György
Department of Electrical and Electronic Engineering,
Imperial College London, London, SW7 2BT, United Kingdom
January 11, 2018
Abstract
In this paper, we consider two sequential decision making problems with a convexity
structure, namely an energy storage optimization task and a multi-product assembly
example. We formulate these problems in the stochastic programming framework and
discuss an approximate dynamic programming technique for their solutions. As the
cost-to-go functions are convex in these cases, we use max-affine estimates for their approximations. To train such a max-affine estimate, we provide a new convex regression
algorithm, and evaluate it empirically for these planning scenarios.
1
Introdution
This paper considers multi-stage stochastic programming problems (e.g., Shapiro et al.,
2009; Birge and Louveaux, 2011) with restricting the model so that the cost-to-go functions remain convex (Section 2). To motivate this framework, we provide two realistic
benchmark planning problems (Section 5), namely solar energy production with storage,
and the operation of a beer brewery.
To address these problems, we consider an approximate dynamic programming approach
(e.g., Bertsekas, 2005; Powell, 2011). For this, we estimate the cost-to-go functions by
convex regression techniques using max-affine representations (formed by the maximum
of finitely many affine functions). To train a max-affine estimator, we propose a novel
algorithm (AMAP, Section 3), which combines ideas of the convex adaptive partitioning
method (CAP, Hannah and Dunson, 2013) and the least squares partition algorithm (LSPA,
Magnani and Boyd, 2009), while learns the model size by cross-validation.
We discuss a full approximate dynamic programming approach (Section 4) that estimates the cost-to-go functions globally. It uses a forward pass to generate a data set for
the cost-to-go estimation by uniformly sampling the reachable decision space. Then it performs a backward pass to recursively approximate the cost-to-go functions for all stages.
The technical details of this algorithm is provided for polyhedral decision sets and convex
piecewise-linear cost-to-go representations.
Finally, we evaluate max-affine estimators in the contexts of our benchmark planning
problems through the full approximate dynamic programming algorithm (Section 5).
1
2
Convex stochastic programming
Consider a T -stage stochastic programming (SP) problem (e.g., Ruszczyński and Shapiro,
2003; Shapiro et al., 2009; Birge and Louveaux, 2011), where the goal is to find a decision
x∗1 solving the following:
h
i
.
x∗1 ∈ argmin J1 (x1 ) , Jt (xt ) = E ct (xt , Z t ) +
min
Jt+1 (xt+1 ) ,
(1)
xt+1 ∈Xt+1 (xt ,Z t )
x1 ∈X1 (x0 ,z0 )
.
.
with t = 1, . . . , T , some fixed initial values x0 , z0 , XT +1 (xT , Z T ) = {⊥}, JT +1 (⊥) = 0, and
a sequence of independent random variables Z 1 , . . . , Z T .
Notice that (1) includes discrete-time finite-horizon Markov decision process formulations (e.g., Puterman, 1994; Sutton, 1998; Bertsekas, 2005; Szepesvári, 2010; Powell, 2011)
after the state and action variables are combined into a single decision variable xt , and
the environment dynamics along with the action constraints are described by the decision
constraint functions Xt .
In this text, we consider only a subset of SP problems (1) when the cost functions
c1 , . . . , cT are convex in xt , and graph(Xt (xt , Z t )) are convex sets for all t = 1, . . . , T and
all Z t realizations,
where the graph of a set-valued function C : X → 2Y is defined as
.
graph(C) = (x, y) ∈ X × Y : y ∈ C(x) . In this case, the cost-to-go functions Jt (·) are
convex for all t = 1, . . . , T (e.g., see Lemma 1), hence we call these SP problems convex.
One approach to deal with such SP problems (1) is to use approximate dynamic programming (ADP) methods (e.g., Bertsekas, 2005; Powell, 2011; Birge and Louveaux, 2011;
Hannah et al., 2014) which construct nested approximations to the cost-to-go functions,
ˆ
ˆ
Jt+1 (xt+1 ) ≈ Jt (xt ) ,
(2)
Jt (xt ) ≈ E ct (xt , Z t ) +
min
xt+1 ∈Xt+1 (xt ,Z t )
backwards for t = T, T −1, . . . , 1. Notice that for convex SP problems with convex cost-to-go
functions Jt , the estimates Jˆt can be restricted to be convex without loss of generality, and
so the minimization tasks in (2) can be solved efficiently.
3
Adaptive max-affine partitioning algorithm
To represent convex cost-to-go functions, we use max-affine maps formed by the maximum
of finitely many affine functions (hyperplanes). To train the parameters of such a maxaffine estimate, we present a new convex regression algorithm called Adaptive Max-Affine
Partitioning (AMAP), which combines the partitioning technique of the convex adaptive
partitioning algorithm (CAP, Hannah and Dunson, 2013), the least squared partition algorithm (LSPA, Magnani and Boyd, 2009), and learns the model size (number of hyperplanes)
by a cross-validation scheme.
Just as LSPA and CAP, AMAP also P
aims to reduce the empirical risk with respect to
. 1 n
the squared loss defined as Rn (f ) = n i=1 |f (xi ) − yi |2 for some function f and data
.
set Dn = {(xi , yi ) : i = 1, . . . , n} ⊆ (Rd × R)n with n samples and dimension d. For the
discussion of AMAP, denote the max-affine function of model M = {(ak , bk ) : k = 1, . . . , K}
.
d
by fM (x) = maxk=1,...,K a⊤
k x+bk for x ∈ R . Notice that each max-affine model M induces
a partition over the data set Dn as
.
P = {C1 , . . . , CK } ,
.
Ck = i ∈ {1, . . . , n} x⊤
i ak + bk = fM (xi ) ,
2
(3)
for some K ∈ N and all k = 1, . . . , K, where ties are broken arbitrarily so that Ck 6= ∅ and
Ck ∩ Cl 6= ∅ ⇐⇒ k = l for all k, l = 1, . . . , K. Furthermore, each partition P induces a
max-affine model by fitting each cell of P using the linear least squares algorithm:
1 X
.
xi ,
∆ik = xi −
|Ck |
i∈Ck
−1 X
X
1
. X
.
ak =
+
βI
∆ik ∆⊤
∆
y
,
b
=
yi − a⊤
i
d
ik
k
ik
k xi ,
|Ck |
.
M = {(ak , bk ) : k = 1, . . . , K} ,
i∈Ck
i∈Ck
(4)
i∈Ck
where Id is the d × d identity matrix, and β > 0 is set to some small value for stability (we
.
use β = 10−6 ).
Similar to CAP, AMAP builds the model incrementally by cell splitting, and improves
the partition using LSPA by alternating steps (3) and (4). The AMAP model improvement
step is given by Algorithm 1. Notice that AMAP performs coordinate-wise cell splitting
1.
2.
3.
4.
5.
6.
7.
8.
9.
10.
11.
12.
13.
14.
15.
16.
17.
18.
19.
20.
21.
22.
23.
24.
input: training set Dn , model M = {(ak , bk ) : k = 1, . . . , K},
partition P = {C1 , . . . , CK }, empirical risk E, minimum cell size s∗
{cell splitting}
M∗′ ← M, E∗′ ← E, P∗′ ← P
for all k = 1, . . . , K do
if |Ck | ≥ 2s∗ then
for all j = 1, . . . , d do
mj ← median({Xij : i ∈ Ck })
Cle ← {i ∈ Ck : Xij ≤ mj }, Cgt ← {i ∈ Ck : Xij ≥ mj }
(ale , ble ) ← least squares on {(X i , Yi ) : i ∈ Cle }
(agt , bgt ) ← least squares
on {(X i , Yi ) : i ∈ Cgt }
M ′ ← M \ {(ak, bk )} ∪ {(ale , ble ), (agt , bgt )}
E ′ ← Rn fM ′ (·)
if E ′ < E∗′ then
M∗′ ← M ′ , E∗′ ← E ′ , P∗′ ← (P \ {Ck }) ∪ {Cle , Cgt }
end if
end for
end if
end for
{running LSPA}
repeat
M∗ ← M∗′ , E∗ ← E∗′ , P∗ ← P∗′
P∗′ ← induced partition of M∗ by (3)
M∗′ ← fitting of partition
P∗′ by (4)
′
E∗ ← Rn fM∗′ (·)
until minC∈P∗′ |C| ≥ s∗ and E∗′ < E∗
output: model M∗ , partition P∗ , empirical risk E∗
Algorithm 1: AMAP model improvement step.
(steps 5 to 15), just as CAP, however, AMAP makes the split always at the median (steps 6
and 7) instead of checking multiple cut points. This saves computation time, but can also
create worse splits. To compensate for this loss in quality, AMAP runs a restricted version
of LSPA (steps 18 to 23) not just for a single step as CAP, but until the candidate model
3
improves the empirical risk and its induced partition satisfies the minimum cell requirement
(step 23). We also mention that indices {i ∈ Ck : Xij = mj } are assigned to Cle and Cgt
(step 7) in order to preserve the minimum cell requirement.
Notice that the difference between M ′ and M is only two hyperplanes (step 10), so the
number of arithmetic operations for computing E ′ (step 11) can be improved from O(nKd)
to O(nd). Further, the cost of least squares regressions (steps 8 and 9) is O(|Ck |d2 ). Hence,
the computational cost of the entire cell splitting process (steps 2 to 17) is bounded by
O(max{K, d}d2 n). For the LSPA part, the partition fitting (step 21) is O(nd2 ) and the
error calculation (step 22) is O(nKd). So, the cost of a single LSPA iteration (steps 19
to 22) is bounded by O(max{K, d}dn),
implying that the cost of Algorithm 1 is bounded
by O max{tLSPA , d} max{K, d}dn , where tLSPA denotes the number of LSPA iterations.
Undesirably, coordinate-wise cell splitting does not provide rotational invariance. To fix
this, we run AMAP after a pre-processing step, which uses thin singular value decomposition
(thin-SVD) to drop redundant coordinates and align the data along a rotation invariant
basis. Formally, let the raw (but already centered) data be organized into X ∈ Rn×d and
.
y ∈ Rn . Then, we scale the values [y1 . . . yn ]⊤ = y/ max{1, kyk∞ }, and decompose X
by thin-SVD as X = U SV ⊤ , where U ∈ Rn×d is semi-orthogonal, S ∈ Rd×d is diagonal
with singular values in decreasing order, and V ∈ Rd×d is orthogonal. Coordinates related
.
to zero singular values are dropped1 and the points are scaled by S as [x1 . . . xn ]⊤ =
U S/ max{1, S11 }, where S11 is the largest singular value. Now observe that rotating the
raw points X as XQ with some orthogonal matrix Q ∈ Rd×d only transforms V to Q⊤ V
and does not affect the pre-processed points x1 , . . . , xn . Finally, we note that thin-SVD
can be computed using O(nd2 ) arithmetic operations (with n ≥ d),2 which is even less than
the asymptotic cost of Algorithm 1.
AMAP is presented as Algorithm 2 and run using uniformly shuffled (and pre-processed)
data Dn , and a partition {F1 , . . . , Fu∗ } of {1, . . . , n} with equally sized cells (the last one
might be smaller) defining the cross-validation folds.
For model selection, AMAP uses a u∗ -fold cross-validation scheme (steps 8 to 19) to
find an appropriate model size that terminates when the cross-validation error (step 14) of
the best model set M∗ (steps 7 and 16) cannot be further improved for twait iterations. At
the end, the final model is chosen from the model set M∗ having the best cross-validation
error, and selected so to minimize the empirical risk on the entire data (step 20). For this
.
.
scheme, we use the parameters u∗ = 10 and twait = 5.
AMAP starts with models having a single hyperplane (steps 2 to 6) and increments each
model by at most one hyperplane in every iteration (step 11). Notice that if AMAP cannot
find a split for a model Mu to improve the empirical risk Eu , the update for model Mu
(steps 11 and 12) can be skipped in the subsequent iterations as Algorithm 1 is deterministic.
.
We also mention that for the minimum cell size, we use s∗ = max{2(d + 1), ⌈log 2 (n)⌉}
allowing model sizes up to O nd/(d+4) / ln(n) , which is enough for near-optimal worst-case
performance (Balázs et al., 2015, Theorems 4.1 and 4.2).
1
By removing columns of U and V , and columns and rows of S.
First decompose X by a thin-QR algorithm in O(nd2 ) time (Golub and Loan, 1996, Section 5.2.8) as
X = QR, where Q ∈ Rn×d has orthogonal columns and R ∈ Rd×d is upper triangular. Then apply SVD for
R in O(d3 ) time (Golub and Loan, 1996, Section 5.4.5).
2
4
1.
2.
3.
4.
5.
6.
7.
8.
9.
10.
11.
12.
13.
14.
15.
16.
17.
18.
19.
20.
21.
input: training set Dn , minimum cell size s∗ , folds F1 , . . . , Fu∗ , iterations twait
{initialization}
for u = 1, . . . , u∗ do
.
Pu ← {Fu } with Fu = {1, . . . , n} \ Fu
u )} ← least squares on {(x , y ) : i ∈ F }
Mu = {(au1 , bP
i i
1
Pu
Eu ← |Fu |−1 i∈Fu |fMu (xi ) − yi |2 , Eu ← |Fu |−1 i∈Fu |fMu (xi ) − yi |2
end for
P ∗
Eu
M∗ ← {M1 , . . . , Mu∗ }, E∗ ← u1∗ uu=1
{cross-validation training}
t ← 1, tmax ← twait
while t ≤ min{tmax , ⌈nd/(d+4) ⌉} do
for u = 1, . . . , u∗ do
(Mu , Pu , Eu ) P
← update by Algorithm 1 using {(xi , yi ) : i ∈ Fu }, Mu , Pu , Eu , s∗
2
−1
Eu ← |Fu |
i∈Fu |fMu (xi ) − yi |
end for P
∗
E ← u1∗ uu=1
Eu
if E < E∗ then
M∗ ← {M1 , . . . , Mu∗ }, E∗ ← E, tmax ← t + twait
end if
t←t+1
end while
{choosing the final model}
M∗ ← argminM ∈M∗ Rn fM (·)
output: model M∗
Algorithm 2: Adaptive max-affine partitioning (AMAP).
4
Approximate dynamic programming
Here we use max-affine estimators to approximate the cost-to-go functions of convex SP
problems. First, notice that solving (1) is equivalent to the computation of π1 (x0 , z0 ),
where
.
πt (xt−1 , zt−1 ) =
argmin
Jt (xt ) ,
xt ∈Xt (xt−1 ,zt−1 )
(5)
Jt (xt ) = E ct (xt , Z t ) + Jt+1 (πt+1 (xt , Z t )) ,
.
.
for all t = 1, . . . , T , and πT +1 (·, ·) = ⊥ with JT +1 (⊥) = 0. The sequence π = (π1 , . . . , πT )
represents an optimal policy.
We only consider SP problems with convex polyhedral decision constraints written as
Xt+1 (xt , Z t ) = {xt+1 : Qt+1 xt+1 + Wt+1 (Z t )xt ≤ ct+1 (Z t )} which are non-empty for all
possible realizations of xt and Z t . As the coefficient Qt+1 of the decision variable xt+1 is
independent of random disturbances Z t and the constraint xt ∈ Xt (xt−1 , zt−1 ) for policy
πt is feasible for any xt−1 and zt−1 , these SP problems are said to have a fixed, relatively
complete recourse (Shapiro et al., 2009, Section 2.1.3). We will exploit the fixed recourse
property for sampling (6), while relatively complete recourse allows us not to deal with
infeasibility issues which could make these problems very difficult to solve.3
3
Infeasible constraints can be equivalently modeled by cost-to-go functions assigning infinite value for
points outside of the feasible region. Then notice that the estimation of functions with infinite magnitude
and slope can be arbitrarily hard even for convex functions (Balázs et al., 2015, Section 4.3).
5
In order to construct an approximation Jˆt to the cost-to-go function Jt , we need “realizable” decision samples xt,i at stage t. We generate these incrementally during a forward pass
for t = 1, 2, . . . , T , where given n decisions xt,1 , . . . , xt,n and m disturbances zt,1 , . . . , zt,m
at stage t, we uniformly sample new decisions for stage t + 1 from the set
[
.
Xt+1 (xt,i , zt,j )
X̂t+1 = conv
i=1,...,n,
j=1,...,m
n
= xt+1 : Qt+1 xt+1
(6)
o
≤ max ct+1 (zt,j ) − Wt+1 (zt,j )xt,i ,
i=1,...,n,
j=1,...,m
where conv(·) refers to the convex hull of a set, and the maximum is taken component-wise.
To uniformly sample the convex polytope X̂t+1 , we use the Hit-and-run Markov-Chain
Monte-Carlo algorithm (Smith, 1984, or see Vempala, 2005) by generating 100 chains (to
reduce sample correlation) each started from the average of 10 randomly generated border
points, and discarding d2t+1 samples on each chain during the burn-in phase,4 where dt+1 is
the dimension of xt+1 .
Then, during a backward pass for t = T, T − 1, . . . , 1, we recursively use the cost-to-go
estimate of the previous stage Jˆt+1 (·) to approximate the values of the cost-to-go function Jt
at the decision samples xt,i generated during the forward pass, that is
m
. 1 X
Jˆt+1 (xt+1 )
ct (xt,i , zt,j ) +
min
Jt (xt,i ) ≈ yt,i =
m
xt+1 ∈Xt+1 (xt,i ,zt,j )
(7)
j=1
for all t = 1, . . . , T , and JˆT +1 (·) ≡ 0. This allows us to set up a regression problem with
data {(xt,i , yt,i ) : i = 1, . . . , n}, and to construct an estimate Jˆt (·) of the cost-to-go Jt (·).
We call the resulting method, shown as Algorithm 3, full approximate dynamic programming (fADP) because it constructs global approximations to the cost-to-go functions.
1.
2.
3.
4.
5.
6.
7.
8.
9.
10.
11.
12.
13.
input: SP problem, trajectory count n, evaluation count m, estimator REG
x0,i ← x0 for all i = 1, . . . , n
z0,j ← z0 for all j = 1, . . . , m
{forward pass}
for all t = 0, 1, . . . , T − 1 do
sample xt+1,1 , . . . , xt+1,n from X̂t+1 by (6) using xt,1 , . . . , xt,n , and zt,1 , . . . , zt,m
sample zt+1,1 , . . . , zt+1,m from the distribution of Z t+1
end for
{backward pass}
JˆT +1 (·) ← 0
for all t = T, T − 1, . . . , 1 do
compute yt,1 , . . . , yt,n by (7) using Jˆt+1 (·), xt,1 , . . . , xt,n , and zt,1 , . . . , zt,m
Jˆt (·) ← REG({(xt,i , yt,i ) : i = 1, . . . , n})
end for
output: cost-to-go functions Jˆt (·), t = 1, . . . , T
Algorithm 3: Full approximate dynamic programming (fADP).
When the cost-to-go functions Jt are approximated by “reasonably sized” convex piecewise linear representations Jˆt , the minimization problem in (7) can be solved efficiently by
4
The choice was inspired by the O(d2t+1 ) mixing result of the Hit-and-run algorithm (Lovász, 1999).
6
linear programming (LP). In the following sections, we exploit the speed of LP solvers for
fADP using either AMAP or CAP as the regression procedure. Then, the computational
cost to run Algorithm 3 is mostly realized by solving T nm LP tasks for (7) and training T
estimators using the regression algorithm REG.
Here we mention that max-affine estimators using as many hyperplanes as the sample
size n can be computed by convex optimization techniques efficiently up to a few thousand
samples (Mazumder et al., 2015). These estimators provide worst-case generalization error
guarantees, however, the provided max-affine models are large enough to turn the LP tasks
together too costly to solve (at least using our hardware and implementation).
The situation is even worse for nonconvex estimators REG for which LP has to be replaced for (7) by a much slower nonlinear constrained optimization method using perhaps
randomized restarts to minimize the chance of being trapped in a local minima. Furthermore, when the gradient information with respect to the input is not available (as
for many implementations), the minimization over these representations require an even
slower gradient-free nonlinear optimization technique. Hence, multivariate adaptive regression splines, support vector regression, and neural networks were impractical to use for
fADP on our test problems.
5
Experiments
To test fADP using max-affine estimators, we consider two SP planning problems, namely
solar energy production with storage management (Section 5.1), and the operation of a
beer brewery (Section 5.2).
For our numerical experiments, the hardware has a Dual-Core AMD Opteron(tm) Processor 250 (2.4GHz, 1KB L1 Cache, 1MB L2 Cache) with 8GB RAM. The software uses
MATLAB (R2010b), and the MOSEK Optimization Toolbox (v7.139).
To measure the performance of the fADP algorithm, we evaluate the greedy policy
with respect to the learned cost-to-go functions {Jˆt : t = 1, . . . , T }. More precisely, we
.
run π̂ = (π̂1 , . . . , π̂T ) with π̂t (xt−1 , zt−1 ) ∈ argminxt ∈Xt (xt−1 ,zt−1 ) Jˆt (xt ) on 1000 episodes,
.
(e) (e)
1 P1000 PT
and record the average revenue (negative cost) as REV = − 1000
e=1
t=1 ct xt , zt
(e) (e)
: t = 1, . . . , T , e = 1, . . . , 1000. We repeat
over the episodes’ trajectories
xt , zt
this experiment 100 times for each regression algorithm REG,5 and show the mean and
standard deviation of the resulting sample.
5.1
Energy storage optimization
Inspired by a similar example of Jiang and Powell (2015, Section 7.3), we consider an
energy storage optimization problem where a renewable energy company makes a decision
every hour and plans for two days (T = 48). The company owns an energy storage with
state s which can be charged with maximum rate rc , using the company’s renewable energy
source (E) or the electrical grid that the company can buy electricity from while paying the
retail price (p). The goal is to maximize profit by selling electricity to local clients on retail
price (p) according to their stochastic demand (D) or selling it back to the electrical grid
on wholesale price (w). Electricity can be sold directly from the renewable energy source
5
The random seeds are kept synchronized, so every algorithm is evaluated on the same set of trajectories.
Furthermore, fADP algorithms with the same n and m parameters use the same training data xt,1 , . . . , xt,n
and zt,1 , . . . , zt,m for all t = 0, . . . , T .
7
or from the battery with maximum discharge rate rd . The energy flow control variables,
fes , fed , feg , fsd , fsg , fgs , are depicted on Figure 1.
Figure 1: Flow diagram of a convex energy storage problem.
The SP model (1) of the energy storage problem can be formulated by using the decision
.
.
variable xt = [st fes,t fed,t feg,t fsd,t fsg,t fgs,t ]⊤ ∈ [0, ∞)7 and setting Z t = [Et+1 Dt+1 ]⊤ .
Furthermore, the cost function is defined as
.
ct (xt , Z t ) = pt (fgs,t − fed,t − fsd,t ) − wt (feg,t + fsg,t ) ,
for all t = 1, . . . , T , and the dynamics and control constraints are described by
s
f
s
=
s
+
f
−
f
−
f
+
f
,
es
t
es,t
sg,t
gs,t
sd,t
fes , fed , feg , fsd , fsg , fgs ≥ 0,
fed
.
,
0 ≤ s + fes − fsd − fsg + fgs ≤ smax ,
Xt+1 (xt , Z t ) = feg
f
f
+
f
≤
r
,
f
+
f
≤
r
,
es
gs
c
sg
sd
d
sd
fsg fes + fed + feg ≤ Et+1 , fed + fsd ≤ Dt+1
fgs
.
.
for all t = 0, . . . , T − 1. To initialize the system, define x0 = [s0 0 . . . 0]⊤ and z0 = [d1 e1 ]⊤ ,
where s1 = s0 ∈ [0, smax ] is the current storage level and d1 , e1 ≥ 0 are the currently
observed demand and energy production, respectively.
To set the parameters of this energy storage optimization problem, we consider a solar
energy source, a discounted nightly electricity pricing model (Economy 7 tariff), and plan.
ning for a two days horizon on hourly basis (T = 48). Retail and expected wholesale price
curves, along with the electricity demand and energy production distributions of this model
are shown on Figure 2 for the two consecutive sunny days. The distributions are truncated
normal with support D ∈ [0, 15] and E ∈ [0, 12] for demand and energy production, respec.
.
tively and the storage has capacity smax = 20 with charge and discharge rates rc = 4 and
.
.
.
.
rd = 10, respectively. The model is initialized by s0 = 0, d1 = E[D1 ], and e1 = E[E1 ].
To evaluate fADP on this problem, we use the CAP6 and AMAP convex regression
algorithms, and multiple configurations determined by the number of trajectories n generated for training (which is the sample size for the regression tasks as well), and the number
of evaluations m used to approximate the cost-to-go functions Jt at a single point (7). The
result is presented on Figure 3, which also includes a “heuristic” algorithm to provide a
6
As the implementation of CAP is not too reliable for highly-correlated features, we combined it with
the data preprocessing step of AMAP (see Section 3) which improved its stability.
8
retail price (p)
wholesale price (w)
demand (D)
energy (E)
25
20
15
10
5
0
1
12
24
36
48
hours
Figure 2: Parameters of the energy storage optimization problem. Retail (p) and expected
wholesale (w) price curves, energy demand (D) and production (E) distributions with mean
and standard deviation are shown for two-day long period.
fADP(CAP,m = 10)
fADP(AMAP,m = 10)
30
3700
minutes
revenue
3800
fADP(CAP,m = 100)
fADP(AMAP,m = 100)
heuristic
3600
3500
15
0
25
50
100
25
150
50
100
150
training samples per stage (nm, in 1000s)
Figure 3: Energy storage optimization results for the fADP algorithm using AMAP or CAP
as the inner convex regression procedure. Results show the total revenue and the training
time in minutes with mean and standard deviation for the number of trajectories n and
cost-to-go evaluations m.
baseline. The heuristic uses a fixed policy of immediately selling the solar energy preferrably
for demand (fed → max, feg ≥ 0, fes = 0), selling from the battery during the day when
demand still allows (fgs = 0, fsd ≥ 0), charging the battery overnight (fgs → max, fsd = 0),
and selling everything close to the last stage (fgs = 0, fsd → max, fsg → max). This policy
is much better than the optimal policy without storage 7 which scores 3227 ± 6.
The results of Figure 3 show that fADP using either CAP or AMAP significantly outperforms the heuristic baseline algorithm when the sample size is large enough. The regression
algorithms prefer larger sample sizes n to better sample quality m, although this significantly increases the computation time for CAP and provides only an insignificant revenue
increase compared to AMAP.
.
Because p ≥ w, the optimal policy for smax = 0 minimizes the immediate cost by fed = min{E, D},
.
.
.
.
.
= max{0, E − fed }, and fes = fsd = fgs = fsg = 0.
7
feg
9
5.2
Beer brewery operation
Inspired by Bisschop (2016, Chapter 17), we consider the multi-product assembly problem
of operating a beer brewery which makes a decision in every two weeks and plans for about
one year (48 weeks, T = 24). The factory has to order ingredients (stratch source, yeast,
hops) to produce two types of beers (ale and lager) which have to be fermented (for at least
2 weeks for ale and 6 weeks for lager) before selling. The states and actions of this process
are illustrated on Figure 4.
Figure 4: Diagram of the beer brewery problem. The state includes three types of ingredients and two types of bottled beer in the warehouse, and the four fermentation tanks.
Actions are formed by ingredient orders ur , brewing amounts ub , and beer sales us .
The decision variable xt is a 16 dimensional vector with the following components:
stratch source in storage
yeast in storage
hops in storage
ale beer fermented for less than 2 weeks
produced ale beer
lager beer fermented for less than 2 weeks
lager beer fermented for 2 to 4 weeks
lager
beer
fermented
for
4
to
6
weeks
.
∈ [0, ∞)16 .
xt =
produced lager beer
stratch source order
yeast order
hops order
ale beer brewing
lager beer brewing
ale beer sales
lager beer sales
The first 9 coordinates are state variables, while the last 7 coordinates represent actions.
.
⊤
⊤
The cost functions (which may take negative values) are ct (xt , Z t ) = [h⊤
t ct −rt ]xt , where
9
5
ht ∈ [0, ∞) is the storage cost, ct ∈ [0, ∞) is the market price of the ingredients with
the brewing costs (adjusted by the water price), and rt ∈ [0, ∞)2 is the selling price of the
10
beers, for stages t = 1, . . . , T . The constraint on the dynamics of the system is given by
F xt + Rur + Bub − Sus
ur , ub ≥ 0, us ∈ [0, Dt+1 ],
u
.
r
Xt+1 (xt , Z t ) =
F
x
+
Bu
−
Su
≥
0,
,
t
s
b
ub
F xt + Rur + Bub ≤ kt+1
us
for all t = 1, . . . , T −1, where Z t = Dt+1 ∈ [0, ∞)2 is the beer demand, kt+1 ∈ [0, ∞]9 is the
capacity bound, 0 denotes a zero vector of appropriate size, and the fermentation matrix
F ∈ {0, 1}9×16 , the brewing matrix B ∈ R9×2 , the storage loading matrix R ∈ {0, 1}9×3
and the selling matrix S ∈ {0, 1}9×2 are defined as
−ba −bl
03×2
03×13
I3
1 0
00
00
.
0 0
B=
,
02×4
10
1
1
0 1
.
.
,
S=
F =
0000
00
06×3
,
06×7
03×2
00
1
0
0
0
04×2
00
0100
I3
.
,
R=
0011
01
06×3
where 0m×n is a zero matrix of size m × n, and ba , bl ∈ [0, ∞)3 are the required ingredients
.
.
for brewing ales and lagers, respectively. To initialize the system, we use x0 = 0 and z0 = 0.
To test the fADP algorithm on this problem, we set the horizon to 48 weeks horizon
.
on a fortnight basis (T = 24). The demand distributions Z t for lager and ale beers are
shown on Figure 5 with mean and standard deviation. Both distributions are truncated
lager demand
ale demand
12
10
8
6
4
2
0
1
6
12
18
24
fortnights
Figure 5: Lager and ale beer demand distributions for the beer brewery optimization problem with mean and standard deviation are shown for a 48 weeks horizon.
normal with support [0.1, 12]. The cost vectors are set to fixed values for all t = 1, . . . , T
.
.
.
as ht = [1 0.2 0.2 1 2 1 1 1 2]⊤ , ct = [20 10 5 1 1]⊤ , and rt = [90 50]⊤ . Furthermore, the
.
.
ingredient requirement vectors for brewing are ba = [1 1 1]⊤ and bl = [0.5 0.9 0.8]⊤ for
.
ale and lager, respectively, and the capacity vector is k = [10 10 10 10 ∞ 10 ∞ ∞ ∞]⊤ to
ensure the feasibility (relatively complete recourse) requirements, as discussed in Section 4.
11
Similar to the energy optimization case, we use the CAP and AMAP estimators for
fADP with various trajectory set sizes n and cost-to-go evaluation numbers m. As a baseline, we use the optimal solution for the deterministic version of (1) where Z t is replaced by
its expectation EZ t for all t = 1, . . . , T . The results are presented on Figure 6, showing that
fADP(CAP,m = 10)
fADP(AMAP,m = 10)
5000
4000
3000
2000
1000
0
3
hours
revenue
fADP(CAP,m = 100)
fADP(AMAP,m = 100)
2
1
0
25 50
100
150
200
250
25 50
100
150
200
250
training samples per stage (nm, in 1000s)
Figure 6: Beer brewery optimization results for fADP algorithm using AMAP and CAP
convex regression to approximate the convex cost-to-go functions. Results show the revenue
with mean and standard deviation, and the training time in minutes for trajectories n and
cost-to-go evaluations m.
AMAP improves the performance significantly by collecting revenue over 4100 compared
to CAP which stays around 3600.
However, the result also shows that the running time of AMAP become significantly
larger than CAP. This indicates that larger max-affine models are trained by AMAP compared to CAP to improve the accuracy of the cost-to-go approximations, that increases the
computational cost of the LP tasks of (7), and eventually slows down the fADP algorithm
significantly.
Finally, notice that using larger trajectory sets n for AMAP provide better performance
for low sample levels nm, but the improved sample quality with m = 100 eventually achieves
the same using significantly less computational resources. So it seems to be still possible
for max-affine estimators to find better tradeoffs between accuracy and model size.
6
Conclusions and future work
In this paper, we considered solving convex SP problems by the combination of an approximate dynamic programming method (fADP) and convex regression techniques. For the
latter, we proposed a new state-of-the-art max-affine estimator (AMAP), and combined it
with an approximate dynamic programming algorithm to address two benchmark convex
SP problems of moderate size.
Clearly, scaling up the fADP algorithm for larger problems would require further improvements. One of these could be using more expressive convex piecewise-linear representations (e.g., sum-max-affine maps), which might compress the LP tasks enough for the
current solvers. For this, Hannah and Dunson (2012) considered various ensemble tehniques
(bagging, smearing, and random forests) to enhance the performance of the CAP estimator.
However, these techniques still seem to construct too large models to increase the accuracy
significantly, that makes the vast amount of LP tasks impractical to solve. Maybe, LP algo12
rithms which can more efficiently solve large number of similar LP problems with different
right hand sides (e.g., Gassmann and Wallace, 1996) could help with this issue.
But eventually, it would become inevitable to localize cost-to-go approximations to a
fraction of the decision space, perhaps by running fADP iteratively alternating between
sampling and estimation phases, and exploring at the boundary of the accurately approximated region of cost-to-go functions in order to find and avoid delayed rewards and costs.
However, this is left for future work.
A
Auxiliary tools
The following result is a slight generalization of Theorem 5.3 in Rockafellar (1972).
Lemma 1. Let X, Y be two convex sets and f : X × Y be a jointly–convex function in its
arguments. Additionally, let C : X → 2Y be a set–valued function for which graph(C) is
.
convex. Then g(x) = inf y∈C(x) f (x, y) is a convex function.
Proof. Let x1 , x2 ∈ X, y1 , y2 ∈ Y and λ ∈ (0, 1). As graph(C) is convex,
y1 ∈ C(x1 ), y2 ∈ C(x2 )
⇒
λy1 + (1 − λ)y2 ∈ C λx1 + (1 − λ)x2 .
Using this with the fact that the infimum on a subset becomes larger, and the joint–
convexity of f , we get
inf
f λx1 + (1 − λ)x2 , z
g λx1 + (1 − λ)x2 =
z∈C(λx1 +(1−λ)x2 )
≤ inf
inf f λx1 + (1 − λ)x2 , λy1 + (1 − λ)y2
y1 ∈C(x1 ) y2 ∈C(x2 )
≤
inf
inf
y1 ∈C(x1 ) y2 ∈C(x2 )
λf (x1 , y1 ) + (1 − λ)f (x2 , y2 )
= λg(x1 ) + (1 − λ)g(x2 ) ,
which proves the convexity of g.
References
G. Balázs, A. György, and C. Szepesvári. Near-optimal max-affine estimators for convex
regression. In G. Lebanon and S. Vishwanathan, editors, The 18th International Conference on Artificial Intelligence and Statistics (AISTATS), volume 38 of JMLR W&CP,
pages 56–64, 2015.
D. P. Bertsekas. Dynamic Programming and Optimal Control, Volume I. Athena Scientific,
3rd edition, 2005.
J. R. Birge and F. Louveaux. Introduction to Stochastic Programming. Springer, 2011.
J. Bisschop. AIMMS Optimization Modeling. 2016. http://www.aimms.com.
H. I. Gassmann and S. W. Wallace. Solving linear programs with multiple right-hand sides:
Pricing and ordering schemes. Annals of Operation Research, 64:237–259, 1996.
G. H. Golub and C. F. V. Loan. Matrix Computations. The Johns Hopkins University
Press, 3rd edition, 1996.
13
L. A. Hannah and D. B. Dunson. Ensemble methods for convex regression with applications
to geometric programming based circuit design. In J. Langford and J. Pineau, editors,
The 29th International Conference on Machine Learning (ICML), pages 369–376, 2012.
L. A. Hannah and D. B. Dunson. Multivariate convex regression with adaptive partitioning.
Journal of Machine Learning Research, 14:3261–3294, 2013.
L. A. Hannah, W. B. Powell, and D. B. Dunson. Semiconvex regression for metamodelingbased optimization. SIAM Journal on Optimization, 24(2):573–597, 2014.
D. R. Jiang and W. B. Powell. An approximate dynamic programming algorithm for
monotone value functions. CoRR, 2015. http://arxiv.org/abs/1401.1590v6.
L. Lovász. Hit-and-run mixes fast. Mathematical Programming, 86(3):443–461, 1999.
A. Magnani and S. P. Boyd. Convex piecewise-linear fitting. Optimization and Engineering,
10(1):1–17, 2009.
R. Mazumder, A. Choudhury, G. Iyengar, and B. Sen. A computational framework for multivariate convex regression and its variants. 2015. http://arxiv.org/abs/1509.08165.
W. B. Powell. Approximate Dynamic Programming: Solving the Curses of Dimensionality.
John Wiley & Sons, 2nd edition, 2011.
M. L. Puterman. Markov Decision Processes: Discrete Stochastic Dynamic Programming.
John Wiley & Sons, 1994.
R. T. Rockafellar. Convex Analysis. Princeton University Press, 1972.
A. P. Ruszczyński and A. Shapiro. Stochastic Programming. Elsevier, 2003.
A. Shapiro, D. Dentcheva, and A. Ruszczyński. Lectures on Stochastic Programming, Modeling and Theory. Society for Industrial and Applied Mathematics and the Mathematical
Programming Society, 2009.
R. L. Smith. Efficient monte carlo procedures for generating points uniformly distributed
over bounded regions. Operations Research, 32(6):1296–1308, 1984.
R. S. Sutton. Reinforcement Learning: An Introduction. MIT Press, 1998.
C. Szepesvári. Algorithms for Reinforcement Learning: Synthesis Lectures on Artificial
Intelligence and Machine Learning. Morgan & Claypool Publishers, 2010.
S. Vempala. Geometric random walk: A survey. Combinatorial and Computational Geometry, 52:573–612, 2005.
14
| 3cs.SY
|
Cooperative Localization for Mobile Agents
A recursive decentralized algorithm based on
arXiv:1505.05908v3 [cs.RO] 5 Oct 2015
Kalman filter decoupling
Solmaz S. Kia, Stephen Rounds and Sonia Martı́nez
Technological advances in ad-hoc networking and miniaturization of electro-mechanical
systems are making possible the use of large numbers of mobile agents (e.g., mobile robots,
human agents, unmanned vehicles) to perform surveillance, search and rescue, transport and
delivery tasks in aerial, underwater, space, and land environments. However, the successful
execution of such tasks often hinges upon accurate position information, which is needed in lower
level locomotion and path planning algorithms. Common techniques for localization of mobile
robots are the classical pre-installed beacon-based localization algorithms [1], fixed featurebased Simultaneous Localization and Mapping (SLAM) algorithms [2], and GPS navigation [3],
see Fig. 1 for further details. However, in some operations such as search and rescue [4],
[5], environment monitoring [6], [7], and oceanic exploration [8], the assumptions required by
the aforementioned localization techniques include the existence of distinct and static features
that can be revisited often, or clear line-of-sight to GPS satellites. Such conditions may not
be realizable in practice, and thus these localization techniques become unfeasible. Instead,
Cooperative localization (CL) is emerging as an alternative localization technique that can be
1
employed in such scenarios.
In CL, a group of mobile agents with processing and communication capabilities use
relative measurements with respect to each other (no reliance on external features) as a feedback
signal to jointly estimate the poses of all team members, which results in an increased accuracy
for the entire team. The particular appeal of CL relies on the fact that sporadic access to accurate
localization information by a particular robot results into a net benefit for the rest of the team. This
is possible thanks to the coupling that is created through the state estimation process. Another nice
feature of CL is its cost effectiveness, as it does not require extra hardware beyond the operational
components normally used in cooperative robotic tasks. In such situations, agents are normally
equipped with unique identifiers and sensors which enable them to identify and locate other group
members. To achieve coordination, these agents often broadcast their status information to one
another. In addition, given the wide and affordable availability of communication devices, CL
has also emerged as an augmentation system to compensate for poor odometric measurements,
noisy and distorted measurements from other sensor suites such as onboard IMU systems, see
e.g., [9].
The idea of exploiting relative robot-to-robot measurements for localization can be traced
back to [10], where members of a mobile robotic team were divided into two groups, which took
turns remaining stationary as landmarks for the others. In later developments in [11], where the
term cooperative localization was also introduced, the necessity for some robots to be stationary
was removed. Since then, many cooperative localization algorithms using various estimation
strategies such as Extended Kalman filters (EKF) [12], maximum likelihood [13], maximum
2
a posteriori (MAP) [14], and particle filters [15], [16], [17] have been developed. Cooperative
localization techniques to handle system and measurement models with non-Gaussian noises are
also discussed in [18], [19].
Although CL is a very attractive concept for multi-robot localization, which does not
require environmental features or GPS information, it also poses new challenges associated with
the implementation of such a policy with acceptable communication, memory, and processing
costs. Cooperative localization is a joint estimation process which results in highly coupled pose
estimation for the full robotic team. These couplings/cross-correlations are created due to the
relative measurement updates. Accounting for these coupling/cross-correlations is crucial for both
filter consistency and also for propagating the benefit of a robot-to-robot measurement update to
the entire group. In Section “Cooperative localization via EKF” we demonstrate these features
in detail both through technical and simulation demonstrations.
A centralized implementation of CL is the most straightforward mechanism to keep an
accurate account of these couplings and, as a result, obtain more accurate solutions. In a
centralized scheme, at every time-step, a single device, either a leader robot or a fusion center
(FC), gathers and processes information from the entire team. Then, it broadcasts back the
estimated location results to each robot (see e.g., [12], [20]). Such a central operation incurs into
a high processing cost on the FC and a high communication cost on both FC and each robotic
team member. Moreover, it lacks robustness that can be induced by single point failures. This
lack of robustness and energy inefficiency make the centralized implementation less preferable.
A major challenge in developing decentralized CL (D-CL) algorithms is how to maintain
3
a precise account of cross-correlations and couplings between the agents’ estimates without
invoking all-to-all communication at each time-step. The design and analysis of decentralized
CL algorithms, which maintain the consistency of the estimation process while maintaining
“reasonable” communication and computation costs have been the subject of extensive research
since the CL idea’s conception. In Section “Decentralized cooperative localization: how to
account for intrinsic correlations in cooperative localization,” we provide an overview of
some of the D-CL algorithms in the literature, with a special focus on how these algorithms
maintain/account for intrinsic correlations of CL strategy. We provide readers a more technical
example of a D-CL algorithm in the Section “The Interim Master D-CL algorithm: a tightly
coupled D-CL strategy based on Kalman filter decoupling,” which is a concise summary of the
solution in [21] developed by the authors.
The reader interested on technical analysis and details beyond decentralization for CL can
find a brief literature guide in “Further Reading.”
Notations: Before proceeding further, let us introduce our notations. We denote by Mn , 0n×m
(when m = 1, we use 0n ) and In , respectively, the set of real positive definite matrices of
dimension n × n, the zero matrix of dimension n × m, and the identity matrix of dimension
n × n. We represent the transpose of matrix A ∈ Rn×m by A> . The block diagonal matrix of
set of matrices A1 , . . . , AN is Diag(A1 , · · · , AN ). For finite sets V1 and V2 , V1 \V2 is the set of
elements in V1 , but not in V2 . For a finite set V we represent its cardinality by |V |. In a team
of N agents, the local variables associated with agent i are distinguished by the superscript i,
e.g., xi is the state of agent i, x̂i is its state estimate, and Pi is the covariance matrix of its
4
state estimate. We use the term cross-covariance to refer to the correlation terms between two
agents in the covariance matrix of the entire network. The cross-covariance of the state vectors
of agents i and j is Pij . We denote the propagated and updated variables, say x̂i , at time-step
k by x̂i- (k) and x̂i+ (k), respectively. We drop the time-step argument of the variables as well
as matrix dimensions whenever they are clear from the context. In a network of N agents,
p = (p1 , . . . , pN ) ∈ Rd , d =
PN
i=1
i
ni is the aggregated vector of local vectors pi ∈ Rn .
Cooperative localization via EKF
This section provides an overview of a CL strategy that employs an EKF following [22]. By
a close examination of this algorithm, it is possible to explain why accounting for the intrinsic
cross-correlations in CL is both crucial for filter consistency and key to transmit the benefit
of an update of a relative robot-to-robot measurement to the entire team. We also discuss the
computational cost of implementing this algorithm in a centralized manner.
First, we briefly describe the model considered for the mobile robots in the team. Consider
a group of N mobile agents with communication, processing and measurement capabilities.
Depending on the adopted CL algorithm, communication can be in bidirectional manner with a
fusion center, a single broadcast to the entire team or in multi-hop fashion as shown in Fig. 2,
i.e., every agent re-broadcasts every received message intended to reach the entire team. Each
agent has a detectable unique identifier (UID) which, without loss of generality, we assume to
be a unique integer belonging to the set V = {1, . . . , N }. Using a set of so-called proprioceptive
sensors every agent i ∈ V measures its self-motion, for example by compass readings and/or
5
wheel encoders, and uses it to propagate its equations of motion.
xi (k + 1) = f i (xi (k), ui (k)) + gi (xi (k))η i (k),
i
(1)
i
i
where xi ∈ Rn , ui ∈ Rm , and η i ∈ Rp are, respectively, the state vector, the input vector and the
process noise vector of agent i. Here, f i (xi , ui ) and gi (xi ), are, respectively, the system function
and process noise coefficient function of the agent i ∈ V. The state vector of each agent can be
composed of variables that describe the robots global pose in the world (e.g. latitude, longitude,
direction), as well as other variables potentially needed to model the robots dynamics (e.g.
steering angle, actuation dynamics). The team can consist of heterogeneous agents, nevertheless,
the collective motion equation of the team can be represented by
x(k + 1) = f (x(k), u(k)) + g(x(k))η(k),
where, x
=
(x1 , · · · , xN ), u
=
(u1 , · · · , uN ), η
=
(η 1 , · · · , η N ), f (x, u)
(2)
=
(f 1 (x1 , u1 ), · · · , f N (xN , uN )) and g(x) = Diag(g1 (x1 ), · · · , gN (xN )).
Obviously, if each agent only relies on propagating its equation of motion in (1) using
self-motion measurements, this state estimate grows unbounded due to the noise term η i (k). To
reduce the growth rate of this estimation error, a CL strategy can be employed. Thus, let every
agent i ∈ V also carry exteroceptive sensors to monitor the environment to detect, uniquely, the
other agents j ∈ V in the team and take relative measurements
zij (k + 1) = hij (xi (k), xj (k)) + ν i (k),
6
(3)
i
where zij ∈ Rnz from them, e.g., relative pose, relative range, relative bearing measurements, or
both. Here, hij (xi , xj ) is the measurement model and ν i is the measurement noise of agent i ∈ V.
Relative-measurement feedback, as shown below, can help the agents improve their localization
accuracy, though the overall uncertainty can not be bounded (c.f. [22]). The tracking performance
can be improved significantly if agents have occasional absolute positioning information, e.g., via
GPS or relative measurements taken from a fixed landmark with a priori known absolute location.
Any absolute pose measurement by an agent i ∈ V, e.g., through intermittent GPS access, is
modeled by zii (k + 1) = hii (xi (k)) + ν̄ i (k). The agents can obtain concurrent exteroceptive
absolute and relative measurements.
Let us assume all the process noises η i and the measurement noise ν i , i ∈ V, are independent zero-mean white Gaussian processes with, respectively, known positive definite variances
i
Qi (k) = E[η i (k)η i (k)> ], Ri (k) = E[ν i (k)ν i (k)> ] and R̄ (k) = E[ν̄ i (k)ν̄ i (k)> ]. Moreover,
let all the sensor noises be white and mutually uncorrelated and all sensor measurements be
synchronized. Then, the centralized EKF CL algorithm is a straightforward application of EKF
over the collective motion model of the robotic team (2) and measurement model (3). The
propagation stage of this algorithm is
x̂- (k + 1) = f (x̂+ (k), u(k)),
(4a)
P- (k + 1) = F(k)P+ (k)F(k)> + G(k)Q(k)G(k)> .
(4b)
where F = Diag(F1 , · · · , FN ), G = Diag(G1 , · · · , GN ) and Q = Diag(Q1 , · · · , QN ), with,
for all i ∈ V, Fi =
∂
f (x̂i+ (k), ui (k))
∂xi
and Gi =
∂
g(x̂i+ (k)).
∂xi
If there exists a relative measurement in the network at some given time k + 1, say robot
7
a takes relative measurement from robot b, the states are updated as follows. The innovation of
the relative measurement and its covariance are, respectively,
ra = zab − hab (x̂a- (k + 1), x̂b- (k + 1)),
(5a)
Sab = Hab (k+1)P- (k+1)Hab (k+1)>+Ra (k+1).
(5b)
where (without loss of generality we let a < b)
a
1 ···
Hab (k) = 0 · · · −H̃a (k)
a+1
0
···
b
· · · H̃b (k)
b+1
0
∂
hab (x̂a- (k), x̂b- (k)),
∂xa
∂
H̃b (k) =
hab (x̂a- (k), x̂b- (k)).
∂xb
H̃a (k) = −
···
··· ,
(6)
An absolute measurement by a robot a ∈ V can be processed similarly, except that in (6),
H̃b becomes zero, while in (5), the index b should be replaced by a and Ra (k +1) should be
a
replaced by R̄ (k+1). Then, the Kalman filter gain is given by
K(k + 1) = P- (k + 1)Hab (k + 1)> Sab −1 .
And, finally, the collective pose update and covariance update equations for the network are:
x̂+ (k+1) =x̂- (k+1)+K(k+1)ra ,
(7a)
P+ (k+1) =P- (k+1)−K(k+1)Sab K(k+1)> .
(7b)
Because K(k+1)Sab K(k+1)> is a positive semi-definite term, the update equation (7b) clearly
shows that any relative measurement update results in a reduction of the estimation uncertainty.
To explore the relationship among the estimation equations of each robot, we express the
aforementioned collective form of the EKF CL in terms of its agent-wise components, as shown
8
>
in Algorithm 1. Here, the Kalman filter gain is partitioned into K =
i
K>
1 ,···
, K>
N
, where
i
Ki ∈ Rn ×nz is the portion of the Kalman gain used to update the pose estimate of the agent
i ∈ V. To process multiple synchronized measurements, sequential updating (c.f. for example
[23, Ch. 3],[24]) is employed.
Algorithm 1 clearly showcases the role of past correlations in a CL strategy. First, observe
that, despite having decoupled equations of motion, the source of the coupling in the propagation
phase is the cross-covariance equation (16c). Upon an incidence of a relative measurement
between agents a and b, this term becomes non-zero and its evolution in time requires the
information of these two agents. Thus, these two agents have to either communicate with each
other all the time or a central operator has to take over the propagation stage. As the incidences
of relative measurements grow, more non-zero cross-covariance terms are created. As a result,
the communication cost to perform the propagation grows, requiring the data exchange all the
time with either a Fusion Center (FC) or all-to-all agent communications, even when there is
no relative measurement in the network. The update equations (18) are also coupled and their
calculations need, in principle, a FC. The next observation regarding the role of the crosscovariance terms can be deduced from studying the Kalman gain equation (19). As this equation
shows, when an agent a takes a relative measurement from agent b, any agent whose pose
estimation is correlated with either of agents a and b in the past, (i.e., P-ib (k + 1) and/or P-ia (k +
1) are non-zero) has a non-zero Kalman gain and, as a result, the agent benefits from this
measurement update. The same is true in the case of an absolute measurement taken by a
robot a.
9
The following simple simulation study demonstrates the significance of maintaining an
accurate account of cross-covariance terms between the state estimates of the team members.
We consider a team of 3 mobile robots moving on a flat terrain whose equations of motion in
a fixed reference frame, for i ∈ {1, 2, 3}, are modeled as
xi (k + 1) =xi (k) + V i (k) cos(φ(k)) δt,
y i (k + 1) =y i (k) + V i (k) sin(φ(k)) δt,
φi (k + 1) =φi (k) + ω(k) δt,
where V i (k) and ω i (k) are true linear and rotational velocities of robot i at time k and δt is the
stepsize. Here, the pose vector of each robot is xi = [xi , y i , φi ]> . Every robot uses odometric
i
(k) = ω i (k) + ηωi (k),
sensors to measure its linear Vmi (k) = V i (k) + ηVi (k) and rotational ωm
velocities, where ηVi and ηωi are their respective contaminating measurement noise. The
standard deviation of ηVi (k), i ∈ {1, 2, 3}, is 0.1V i (k), while the standard deviation of ηωi
is 1 deg/s, for robot 1 and robot 2, and 0.5 deg/s for robot 3. Robots {1, 2, 3} can take
relative pose measurements from one another. Here, we use standard deviations of, respectively
(0.05 m, 0.05 m, 1 deg/s), (0.05 m, 0.05 m, 2 deg/s), (0.07 m, 0.07 m, 1.5 deg/s) for measurement
noises. Assume robot 1 can obtain absolute position measurement with a standard deviation of
(0.1m, 0.1m) for the measurement noise. Figure 3 demonstrates the x-coordinate estimation error
(solid line) and the 3σ error bound (dashed lines) of these robots when they (a) only propagate
their equations of motion using self-motion measurements (black plots), (b) employ an EKF
CL ignoring past correlations between the estimations of the robots (blue plots), (c) employ
an EKF CL with an accurate account of past correlations (red plots). As this figure shows,
10
employing a CL strategy improves the localization accuracy by reducing both the estimation error
and its uncertainty. However, as plots in blue show, ignoring the past correlations (here crosscovariances) among the robots state estimates results in overly optimistic estimations (notice the
almost vanished 3σ error bound in blue plots while the solid blue line goes out of these bounds,
an indication of inconsistent estimation). In contrast, by taking into account the past correlations
(see red plots), one sees a more consistent estimation.
Figure 3 also showcases the role of past cross-covariances to expand the benefit of a
relative measurement between two robots, or of an absolute measurement by a robot to the
rest of the team. For example consider robot 2. In the time interval [10, 90] seconds, robot 1 is
taking a relative measurement from robot 2. As a result, the state estimation equation of robot 1
and robot 2 are correlated, i.e, the cross-covariance term between these two robots is non-zero.
Therefore, in the time interval [90, 110] seconds, when the estimation update is due to the relative
measurement taken by robot 3 from robot 1, the estimation of robot 2 is also improved (see red
plots.) In the time interval [190, 240] seconds, when the estimation update is due to the absolute
measurement taken by robot 1, robot 2 and 3 also benefit from this measurement update due
to past correlations (see the red plots.) Figure 4 shows the trajectories of the robots when they
apply EKF CL strategy. For more enlightening simulation studies, we refer the interested reader
to [22].
11
Decentralized cooperative localization: how to account for intrinsic
correlations in cooperative localization
Based on the observations that
(a) past correlations cannot be ignored,
(b) they are useful to increase the localization accuracy of the team,
(c) the coupling that the correlations create in the state estimation of team members is the main
challenge in developing a decentralized cooperative localization algorithm,
one can find, regardless of the technique, two distinct trends in the design methodology of decentralized cooperative localization algorithms in the literature. We term these as “loosely coupled”
and “tightly coupled” decentralized cooperative localization (D-CL) strategies respectively (see
Fig. 5).
In the loosely coupled D-CL methodology, only one or both of the agents involved in a
relative measurement update their estimates using that measurement. Here, an exact account of
the “network” of correlations (see Fig. 5) due to the past relative measurement updates is not
accounted for. However, in order to ensure estimation consistency, some steps are taken to fix
this problem. Examples of loosely coupled D-CL are given in [8], [25], [26], [27] and [28]. In
the algorithm of [8], only the agent obtaining the relative measurement updates its state. Here, in
order to produce consistent estimates, a bank of extended Kalman filters (EKFs) is maintained
at each agent. Using an accurate book-keeping of the identity of the agents involved in previous
updates and the age of such information, each of these filters is only updated when its propagated
12
state is not correlated to the state involved in the current update equation. Although this technique
does not impose a particular communication graph on the network, the computational complexity,
the large memory demand, and the growing size of information needed at each update time are
its main drawbacks. In the approach [25] it is assumed that the relative measurements are in
the form of relative pose. This enables the agent taking the relative measurement to use its
current pose estimation and the current relative pose measurement to obtain and broadcast a
pose and the associated covariance estimation of its landmark agent (the landmark agent is the
agent the relative measurement is taken from). Then, the landmark agent uses the Covariance
Intersection method (see [29], [30]) to fuse the newly acquired pose estimation with its own
current estimation to increase its estimation accuracy. Covariance Intersection for D-CL is also
used in [26] for the localization of a group of space vehicles communicating over a fixed ring
topology. Here, each vehicle propagates a model of the equation of motion of the entire team
and, at the time of relative pose measurements, it fuses its estimation of the team and of its
landmark vehicle via Covariance Intersection. Another example of the use of split Covariance
Intersection is given in [27], for intelligent transportation vehicles localization. Even though
the Covariance Intersection method produces consistent estimations for a loosely coupled DCL strategy, this method is known to produce overly conservative estimates. Another looselycoupled CL approach is proposed in [28], which uses a Common Past-Invariant Ensemble Kalman
pose estimation filter of intelligent vehicles. This algorithm is very similar to the decentralized
Covariance Intersection data fusion method described above, with the main difference that it
operates with ensembles instead of with means and covariances. Overall, the loosely coupled
algorithms have the advantage of not imposing any particular connectivity condition on the team.
13
However, they are conservative by nature, as they do not enable other agent in the network to
fully benefit from measurement updates.
In the tightly coupled D-CL methodology, the goal is to exploit the “network” of
correlations created across the team (see Fig. 5), so that the benefit of the update can be
extended beyond the agents involved in a given relative measurement. However, this advantage
comes at a potentially higher computational, storage and/or communication cost. The dominant
trend in developing decentralized cooperative localization algorithms in this way is to distribute
the computation of components of a centralized algorithm among team members. Some of the
examples for this class of D-CL is given in [31], [22], [14], [32], [33]. In a straightforward
fashion, decentralization can be conducted as a multi-centralized CL, wherein each agent
broadcasts its own information to the entire team. Then, every agent can calculate and reproduce
the centralized pose estimates acting as a fusion center [31]. Besides a high-processing cost
for each agent, this scheme requires all-to-all agent communication at the time of each
information exchange. A D-CL algorithm distributing computations of an EKF centralized CL
algorithm is proposed in [22]. To decentralize the cross-covariance propagation, [22] uses a
singular-value decomposition to split each cross-covariance term between the corresponding
two agents. Then, each agent propagates its portion. However, at update times, the separated
parts must be combined, requiring an all-to-all agent communication in the correction step.
Another D-CL algorithm based on decoupling the propagation stage of an EKF CL using
new intermediate variables is proposed in [21]. But here, unlike [22], at update stage, each
robot can locally reproduce the updated pose estimate and covariance of the centralized EKF
after receiving an update message only from the robot that has made the relative measurement.
14
Subsequently, [14], [33] present D-CL strategies using maximum-a-posteriori (MAP) estimation
procedure. In the former, computations of a centralized MAP is distributed among all the
team members. In the latter, the amount of data required to be passed between mobile agents
in order to obtain the benefits of cooperative trajectory estimation locally is reduced by
letting each agent to treat the others as moving beacons whose estimate of positions is only
required at communication/measurement times. The aforementioned techniques all assume that
communication messages are delivered, as prescribed, perfectly all the time. A D-CL approach
equivalent to a centralized CL, when possible, which handles both limited communication ranges
and time-varying communication graphs is proposed in [32]. This technique uses an information
transfer scheme wherein each agent broadcasts all its locally available information to every agent
within its communication radius at each time-step. The broadcasted information of each agent
includes the past and present measurements, as well as past measurements previously received
from other agents. The main drawback of this method is its high communication and memory
cost, which may not be affordable in applications with limited communication bandwidth and
storage resources.
The Interim Master D-CL algorithm: a tightly coupled D-CL strategy based
on Kalman filter decoupling
Because of its recursive and simple structure, the EKF is a very popular estimation strategy.
However, as discussed in Section “Cooperative localization via EKF,”a naive decentralized
implementation of EKF requires an all-to-all communication at every time-step of the algorithm.
15
In this section, we describe how by exploiting a special pattern in the propagation estimation
equations, [22] and [21] proposed tightly coupledexact decentralized implementations of EKF
for CL with reduced communication workload per agent. Here, what we mean by “exact” is
that if these decentralized implementations are initialized the same as a centralized EKF, they
produce the same state estimate and the associated state error covariance of the centralized filter.
Our special focus in this section is on the Interim Master D-CL of [21].
The Interim Master D-CL algorithm and the algorithm of [22] are developed based on
the observation that, in localization problems, we are normally only interested in the explicit
value of the pose estimate and the error covariance associated with it, while cross-covariance
terms are only required in the update equations. Such an observation promoted the proposal
of the implicit tracking of cross-covariance terms by splitting them into intermediate variables
that can be propagated locally by the agents. Then, cross-covariance terms can be recovered by
putting together these intermediate variables at any update incidence. Let the last measurement
update be in time-step k and assume that for m subsequent and consecutive steps no relative
measurement incidence takes place among the team members, i.e., no intermediate measurement
update is conducted in this time interval. In such a scenario, the propagated cross-covariance
terms for these m consecutive steps are given by
>
>
j
j
P-ij (k + l) = Fi (k + l − 1) · · · Fi (k) P+
ij (k) F (k) · · · F (k + l − 1) ,
l ∈ {1, · · · , m},
(8)
for i ∈ V and j ∈ V\{i}. That is, at each time step after k, the propagated cross-covariance
term is obtained by recursively multiplying its previous value by the Jacobian of the system
function of agent i on the left and by the transpose of the Jacobian of the system function
16
of agent j at that time step on the right. Based on this observation, Roumeliotis and Bekey
in [22] proposed to decompose the last updated cross-covariance term P+
ij (k) between any
agent i and any other agent j of the team into two parts (for example using the singular value
decomposition technique). Then, agent i will be responsible for propagating the left portion
while agent j propagates the right portion. Note that, as long as there is no relative measurement
among team members, each agent can propagate its portion of the cross-covariance term locally
without a need of communication with others. This was an important result, which led to a
fully decentralized estimation algorithm during the propagation cycle. However, in the update
stage, all the agents needed to communicate with one and other to put together the split crosscovariance terms and proceed with the update stage. The approach to obtain Interim Master DCL , which is outlined below, is also based on the special pattern that the cross-covariance
propagation equations have in (8). That is, we also remove the explicit calculation of the
propagated cross-covariance terms by decomposing them into the intermediate variables that
can be propagated by agents locally. However, this alternative decomposition allows every agent
to update its pose estimate and its associated covariance in a centralized equivalent manner, using
merely an scalable communication message that is received from the team member that takes
the relative measurement. As such, the Interim Master D-CL algorithm removes the necessity of
an all-to-all communication in the update stage and replaces it with propagating a constant size
communication message that holds the crucial piece of information needed in the update stage.
In particular, we observe that P-ij (k + l) in (8) is composed of the following 3 parts: (a)
Fi (k + l − 1) · · · Fi (k) which is local to agent i, (b) the P+
ij (k) that does not change unless
>
>
there is relative measurement among the team members, and (c) Fj (k) · · · Fj (k + l − 1)
17
which is local to agent j. Motivated by this observation, we propose to write the propagated
cross-covariances (16c) as:
P-ij (k + 1) = Φi (k + 1)Πij (k)Φj (k + 1)> ,
i
k ∈ {0, 1, 2, · · · },
(9)
i
where Φi ∈ Rn ×n , for all i ∈ V, is a time-varying variable that is initialized at Φi (0) = Ini
and evolves as:
Φi (k + 1) = Fi (k)Φi (k),
k ∈ {0, 1, 2, · · · },
(10)
(it is interesting to notice the resemblance of (10) and the transition matrix for discrete-time
i
j
systems) and Πij ∈ Rn ×n , for i, j ∈ V and i 6= j, which is also a time-varying variable that is
initialized at Πij (0) = 0ni ×nj . When there is no relative measurement at time k + 1, (9) results
into Πij (k + 1) = Πij (k). However, when there is a relative measurement among the team
members Πij must be updated. Next, we derive an expression for Πij (k + 1) when there is a
relative measurement among team members at time k + 1, such that at time k + 2 one can write
P-ij (k + 2) = Φi (k + 2)Πij (k + 1)Φj (k + 2)> . For this, notice that the update equations (17)
and (19) of the centralized CL algorithm can be rewritten by replacing the cross-covariance terms
by (9) (recall that in the update stage, we are assuming that robot a has taken measurement robot
robot b):
>
>
Sab = Ra + H̃a Pa- (k + 1)H̃a + H̃b Pb- (k + 1)H̃b −
>
>
H̃a Φa (k + 1)Πab (k)Φb (k + 1)> H̃b − H̃b Φb (k + 1)Πba (k)Φa (k + 1)> H̃a ,
|
{z
}
|
{z
}
P-ab (k+1)
P-ba (k+1)
and the Kalman gain is
1
Ki = Φi (k + 1) Γi Sab− 2 ,
18
i ∈ V,
(11)
where
>
>
>
1
>
>
>
1
>
1
Γi = (Πib (k)Φb H̃b −Πia (k)Φa> H̃a ) Sab− 2 , i ∈ V\{a,b},
Γa = (Πab (k)Φb H̃b −(Φa )−1 Pa- H̃a ) Sab− 2,
>
Γb = ((Φb )−1 Pb- H̃b −Πba (k)Φa> H̃a ) Sab− 2 .
(12a)
(12b)
(12c)
Generally, Fi (k) is invertible for all k ≥ 0 and i ∈ V. Therefore, Φi (k), for all k ≥ 0 and i ∈ V,
is invertible.
Next, for i 6= j and i, j ∈ V, we can write the cross-covariance terms (18c) as:
>
P+
ij (k + 1) = Pij (k + 1)−Ki Sab Kj
1
1 >
= Φi (k + 1)Πij (k)Φj (k + 1)>− Φi (k + 1)Γi Sab− 2 Sab Φj (k + 1)Γj Sab− 2
= Φi (k + 1) Πij (k) − Γi Γj > Φj (k + 1)> .
Let us propose
Πij (k + 1) = Πij (k) − Γi Γ>
j .
Then, the cross-covariance update (18c) can be rewritten as:
i
j
>
P+
ij (k + 1) = Φ (k + 1) Πij (k + 1) Φ (k + 1) .
(13)
Therefore, at time k + 2, the propagated cross-covariances terms for i 6= j and i, j ∈ V are:
j
>
P-ij (k + 2) = Fi (k + 1) P+
ij (k + 1) F (k + 1)
= Fi (k + 1)Φi (k + 1)Πij (k + 1)Φj (k + 1)> Fj (k + 1)>
= Φi (k + 2) Πij (k + 1) Φj (k + 2)> .
19
In short, we can rewrite the propagated and the updated cross-covariance terms of the centralized
EKF CL as, respectively, (9) and (13) for all k ∈ {0, 1, · · · } where the variables Φi (k)’s and
Πij ’s, evolve according to, respectively, (10) and
Πij (k),
no relative measurement at k + 1,
Πij (k + 1) =
Π (k) − Γ Γ> , otherwise,
ij
i
(14)
j
for i, j ∈ V and i 6= j.
Next, notice that we can write the updated state estimate and covariance matrix in the new
variables as follows, for i ∈ V,
x̂i+ (k + 1) = x̂i- (k + 1) + Φi (k + 1) Γi r̄a ,
(15)
i
>
Pi+ (k + 1) = Pi- (k + 1) − Φi (k + 1)Γi Γ>
i Φ (k + 1) ,
1
where r̄a = Sab− 2 ra .
Using the alternative representations (9), (13), and (15) of the EKF CL, the decentralized
implementation Interim Master D-CL is given in Algorithm 2. We develop the Interim Master DCL algorithm by keeping a local copy of Πlj ’s at each agent i ∈ V, i.e., Πijl for all j ∈ V\{N }
and l ∈ {j + 1, · · · , N }–because of the symmetry of the covariance matrix we only need
to save, e.g., the upper triangular part of this matrix. For example, for a group of N = 4
robots, every agent maintains a copy of {Πi12 , Πi13 , Πi14 , Πi23 , Πi24 , Πi34 }. During the algorithm
implementation, we assume that if Πijl is not explicitly maintained by agent i, the agent
substitutes the value of (Πilj )> for it.
In Interim Master D-CL , every agent i ∈ V initializes its own state estimate x̂i+ (0),
20
the error covariance matrix Pi+ (0), Φi (0) = Ini , and its local copies Πijl (0) = 0nj ×nl , for
j ∈ V\{N } and l ∈ {j + 1, · · · , N }; see (20). At propagation stage, every agent evolves its
local state estimation, error covariance and Φi , according to, respectively, (16a), (16b), (10);
see (21). At every time step, when, there is no exteroceptive measurement in the team, the
local updated state estimates and error covariance matrices are replaced by their respective
propagated counterparts, while Πijl ’s, to respect (14), are kept unchanged; see (22). When there
is a robot-to-robot measurement, examining (6), (5a), (11), (12b) and (12c) shows that agent
a, the robot that made the relative measurement, can calculate these terms using its local Πijl
and acquiring x̂b- (k +1) ∈ Rn , Φb (k +1) ∈ Rn
b
b ×nb
, and Pb- (k +1) ∈ Mnb ; see (23) and (24).
Then, agent a can assume the role of the interim master and issue the update terms for other
agents in the team; see (25). Using this update message and their local variables, then each agent
i ∈ V can compute (12a) and use it to obtain its local state updates of (15) and (14); see (27).
Figure 6 demonstrates the information flow direction between agent while implementing the
Interim Master D-CL algorithm.
The inclusion of absolute measurements in the Interim Master D-CL is straightforward.
The agent making an absolute measurement is an interim master that can calculate the updatemessage using only its own data and then broadcast it to the team. Next, observe that the
Interim Master D-CL algorithm is robust to permanent agent dropouts from the network. The
operation only suffers from a processing cost until all agents become aware of the dropout. Also,
notice that an external authority, e.g., a search-and-rescue chief, who needs to obtain the location
of any agent, can obtain this location update in any rate (s)he wishes to by communicating with
that agent. This reduces the communication cost of the operation.
21
The Interim Master D-CL algorithm works under the assumption that the message from
the agent taking the relative measurement, the interim master, is reached by the entire team.
Any communication failure results in a mismatch between the local copies of Πlj at the agents
receiving and missing the communication message. The readers are referred to [34] where the
authors present a variation of Interim Master D-CL which is robust to intermittent communication
message dropouts. Such guarantees in [34] are provided by replacing the fully decentralized
implementation with a partial decentralization where a shared memory stores and updates the
Πlj ’s.
Complexity analysis
For the sake of an objective performance evaluation, a study of the computational
complexity, the memory usage, as well as communication cost per agent per time-step of the
Interim Master D-CL algorithm in terms of the size of the mobile agent team N is provided
next. At the propagation state of the Interim Master D-CL algorithm, the computations per agent
are independent of the size of the team. However, at the update stage, for each measurement
update, the computation of every agent is of order N (N −1)/2 due to (33c). As multiple relative
measurements are processed sequentially, the computational cost per agent at the completion of
any update stage depends on the number of the relative measurements in the team, henceforth
denoted by Nz . Then, the computational cost per agent is O(Nz ×N 2 ), implying a computational
complexity of order O(N 4 ) for the worst case where all the agents take relative measurement
with respect to all the other agents in the team, i.e., Nz = N (N − 1). The memory cost per agent
is of order O(N 2 ) which, due to the recursive nature of the Interim Master D-CL algorithm,
22
is independent of Nz . This cost is caused by the initialization (20) and update equation (33c),
which are of order N (N − 1)/2.
For the analysis of the communication cost, let us consider the case of a multi-hop
communication strategy. The Interim Master D-CL requires communication only in its update
stage, where landmark robots should broadcast their landmark message to their respective master,
and every agent should re-broadcast any update-message it receives. Let Nr be the number of
the agents that have made a relative measurement at the current time, i.e., Nr ≤ N is the
number of current sequential interim masters. These robots should announce their identity and the
number of their landmark robots to the entire team for sequential update cuing purpose, incurring
a communication cost of order Nr per robots. Next, the team will proceed by sequentially
processing the relative measurements. Every agent can be a landmark of Na ≤ Nr agents and/or
a master of Nb ≤ (N − 1) agents. The updating procedure starts by a landmark robot sending
its landmark-message to its active interim master, resulting in a total communication cost of
O(Na ) per landmark robot at the end of update stage. Every active interim master should pass
an update message to the entire team, resulting in a total communication cost of O(Nb ) per robot.
Because there are Nr masters, at the end of the update stage, every robot incurs a communication
cost of O(Nr × Nb ) to pass the update messages. Because Na , Nr < Nr × Nb ≤ Nz , the total
communication cost at the end of the update stage is of order O(Nz ) per agent, implying a worst
case broadcast cost of O(N 2 ) per agent. If the communication range is unbounded, the broadcast
cost per agent is O(max{Nb , Na }), with the worst case cost of order O(N ). The communication
message size of each agent in both single or multiple relative measurements is independent of
the group size N . As such for the worst case scenario the communication message size is of
23
order O(1).
The results of the analysis above are summarized in Table I and are compared to those
of a trivial decentralized implementation of the EKF for CL (denoted by T-D-CL) in which
every agent i ∈ V at the propagation stage computes (16)–using the broadcasted Fj (k) from
every other team member j ∈ V\{i}–and at the update stage computes (19) and (18)–using
the broadcast (a, b, ra , Sab , H̃a , H̃b , Ra , Pa- , Pb- ) from agent a that has made relative
measurement from agent b. Agent a calculates Sab , H̃a , H̃b by requesting (x̂b- , Pb- ) from agent
b. We assume that multiple measurements are processed sequentially and that the communication
procedure is multi-hop. Although the overall cost of the T-D-CL algorithm is comparable with
the Interim Master D-CL algorithm, this implementation has a more stringent communication
connectivity condition, requiring a strongly connected digraph topology (i.e., all the nodes on
the communication graph can be reached by every other node on the graph) at each time-step,
regardless of whether there is a relative measurement incidence in the team. As an example,
notice that the communication graph of Fig. 2 is not strongly connected and as such the T-D-CL
algorithm can not be implemented whereas the Interim Master D-CL algorithm can be. Recall
that the Interim Master D-CL algorithm needs no communication at the propagation stage and it
only requires an existence of a spanning tree rooted at the agent making the relative measurement
at the update stage. Finally, the Interim Master D-CL algorithm incurs less computational cost
at the propagation stage.
Algorithm 3 presents an alternative Interim Master D-CL implementation where, instead of
storing and evolving Πlj ’s of the entire team, every agent only maintains the terms corresponding
24
to its own cross-covariances; see (28) and (29). For example in a team of N = 4, robot 1
maintains {Π112 , Π113 , Π114 }, robot 2 maintains {Π221 , Π223 , Π224 }, etc. However, now the interim
master a needs to acquire the Πbj ’s from the landmark robot b and calculate and broadcast Γi ,
i ∈ V to the entire team; see (30), (31) and (33). In this alternative implementation, the processing
and storage cost of every agent is reduced from O(N 2 ) to O(N ), however the communication
message size is increased from O(1) to O(N ).
Tightly coupled versus loosely coupled D-CL: a numerical comparison study
The Interim Master D-CL falls under the tightly coupled D-CL classification. Fig. 7
demonstrates the positioning accuracy (time history of the root mean square error (RMSE)
plot for 50 Monte Carlo simulation runs) of this algorithm versus the loosely coupled EKF and
Covariance-Intersection based algorithm of [25] in the following scenario. We consider the 3
mobile robots employed in the numerical example of Section “Cooperative localization via EKF”
with motion as described in that section. For the sensing scenario here, we assume that, starting
at t = 10 seconds, robot 3 takes persistent relative measurements alternating every 50 seconds
from robot 1 to robot 2 and vice versa. As expected, the tightly coupled Interim Master D-CL
algorithm produces more accurate position estimation results than those of the loosely coupled
D-CL algorithm of [25] (similar results can be observed for the heading estimation accuracy,
which is omitted here for brevity).
In the algorithm of [25], every robot keeps an EKF estimation of its own pose. When a
robot takes a relative pose measurement from another robot (let us call this robot the interim
25
master here as well), it acquires the current position estimation and the corresponding error
covariance of the landmark robot. Then, it uses these along with its own current estimation
and the current relative measurement to extract a new state estimation and the corresponding
error covariance for the landmark robot. After this, the interim master robot transmits these
new estimates to the landmark robot which uses the Covariance Intersection method to fuse
them consistently to its current pose estimate. It is interesting to notice that in this particular
scenario, even though robot 3 has been taking all the relative measurements, it receives no benefit
from such measurements, because only the landmark robots are updating their estimations. Even
though the positioning accuracy of algorithm [25] is lower, it only requires O(1) computational
cost per agent as compared to the O(N 2 ) cost of the Interim Master D-CL algorithm. However,
it also requires more complicated calculations to perform Covariance Intersection fusion. If we
assume that the communication range of each agent covers the entire team, then interestingly
the communication cost of these two algorithms is the same as both use an O(1) landmark
and update messages. However, if the communication range is bounded, the loosely coupled
algorithm of [25] offers a more flexible and cost effective communication policy.
Conclusions
Here, we presented a brief review on Cooperative Localization as an strategy to increase the
localization accuracy of team of mobile agents with communication capabilities. This strategy
relies on use of agent-to-agent relative measurements (no reliance on external features) as a
feedback signal to jointly estimate the poses of the team members. In particular, we discussed
challenges involved in designing decentralized Cooperative Localization algorithms. Moreover,
26
we presented a decentralized cooperative localization algorithm that is exactly equivalent to the
centralized EKF algorithm of [22]. In this decentralized algorithm, the propagation stage is fully
decoupled i.e., the propagation is a local calculation and no intra-network communication is
needed. The communication between agents is only required in the update stage when one agent
makes a relative measurement with respect to another agent. The algorithm declares the agent
made the measurement as interim master that can, by using the data acquired from the landmark
agent, calculate the update terms for the rest of the team and deliver it to them by broadcast.
Future extensions of this work includes concern handling message dropouts and asynchronous
measurement updates.
References
[1] J. Leonard and H. F. Durrant-Whyte. Mobile robot localization by tracking geometric
beacons. IEEE Transactions on Robotics and Automation, 7(3):376–382, June 1991.
[2] G. Dissanayake, P. Newman, H. F. Durrant-Whyte, , S. Clark, and M. Csorba. A solution
to the simultaneous localization and map building (SLAM) problem. IEEE Transactions
on Robotics and Automation, 17(3):229–241, 2001.
[3] S. Cooper and H. Durrant-Whyte. A Kalman filter model for GPS navigation of land
vehicles. In IEEE/RSJ Int. Conf. on Intelligent Robots & Systems, pages 157–163, Munich,
Germany, 1994.
[4] J. S. Jennings, G. Whelan, and W. F. Evans. Cooperative search and rescue with a team
of mobile robots. In Int. Conf. on Advanced Robotics, pages 193–200, 1997.
27
[5] A. Davids. Urban search and rescue robots: from tragedy to technology. IEEE Intelligent
Systems, 17(2):81–83, March-April 2002.
[6] N. Bulusu, J. Heidemann, and D. Estrin. GPS-less low-cost outdoor localization for very
small devices. IEEE Personal Communications, 7(5):28–34, 2000.
[7] M. Trincavelli, M. Reggente, S. Coradeschi, A. Loutfi, H. Ishida, and A. J. Lilienthal.
Towards environmental monitoring with mobile robots. In IEEE/RSJ Int. Conf. on Intelligent
Robots & Systems, pages 2210–2215, 2008.
[8] A. Bahr, M. R. Walter, and J. J. Leonard. Consistent cooperative localization. In IEEE Int.
Conf. on Robotics and Automation, pages 8908–8913, Kobe, Japan, May 2009.
[9] H. Mokhtarzadeh and D Gebre-Egziabher. Cooperative inertial navigation. Navigation:
Journal of the Institute of Navigation, 61(2):77–94, 2014.
[10] R. Kurazume, S. Nagata, and S. Hirose. Cooperative positioning with multiple robots. In
IEEE Int. Conf. on Robotics and Automation, pages 1250–1257, San Diego, CA, May 8–13
1994.
[11] I. Rekleitis, G. Dudek, and E. Milios. Multi-robot collaboration for robust exploration. In
IEEE Int. Conf. on Robotics and Automation, pages 3164–3169, 2000.
[12] S. I. Roumeliotis. Robust mobile robot localization: from single-robot uncertainties to
multi-robot interdependencies. PhD thesis, University of Southern California, 2000.
[13] A. Howard, M. Matark, and G. Sukhatme. Localization for mobile robot teams using
maximum likelihood estimation. In IEEE/RSJ Int. Conf. on Intelligent Robots & Systems,
volume 1, pages 434–439, 2002.
[14] E. D. Nerurkar, S. I. Roumeliotis, and A. Martinelli. Distributed maximum a posteriori
28
estimation for multi-robot cooperative localization. In IEEE Int. Conf. on Robotics and
Automation, pages 1402–1409, Kobe, Japan, May 2009.
[15] D. Fox, W. Burgard, H. Kruppa, and S. Thrun. A probabilistic approach to collaborative
multi-robot localization. Autonomous Robots, 8(3):325–344, 2000.
[16] A. Howard, M. J. Mataric, and G. S. Sukhatm. Putting the ‘I’ in ‘team’: An ego-centric
approach to cooperative localization. In IEEE Int. Conf. on Robotics and Automation,
volume 1, pages 868–874, 2003.
[17] A. Prorok and A. Martinoli. A reciprocal sampling algorithm for lightweight distributed
multi-robot localization. In IEEE/RSJ Int. Conf. on Intelligent Robots & Systems, pages
3241–3247, 2011.
[18] A. T. Ihler, J. W. Fisher, R. L. Moses, and A. S. Willsky. Nonparametric belief propagation
for self-localization of sensor networks. IEEE Journal of Selected Areas in Communications,
23(4):809–819, 2005.
[19] J. Nilsson, D. Zachariah, I. Skog, and P. Händel. Cooperative localization by dual footmounted inertial sensors and inter-agent ranging. EURASIP Journal on Advances in Signal
Processing, 2013(164), 2013.
[20] A. Howard, M. J. Matarić, and G. S. Sukhatme. Mobile sensor network deployment
using potential fields: A distributed scalable solution to the area coverage problem. In
Int. Conference on Distributed Autonomous Robotic Systems, pages 299–308, Fukuoka,
Japan, June 2002.
[21] S. S. Kia, S. Rounds, and S. Martı́nez. A centralized-equivalent decentralized implementation of extended Kalman filters for cooperative localization. In IEEE/RSJ Int. Conf. on
29
Intelligent Robots & Systems, pages 3761–3765, Chicago, IL, September 2014.
[22] S. I. Roumeliotis and G. A. Bekey. Distributed multirobot localization. IEEE Transactions
on Robotics and Automation, 18(5):781–795, 2002.
[23] C. T. Leondes, editor. Advances in Control Systems Theory and Application, volume 3.
Academic Press, New York, 1966.
[24] Y. Bar-Shalom, P. K. Willett, and X. Tian. Tracking and Data Fusion, a Handbook of
Algorithms. YBS Publishing, Storts, CT, USA, 2011.
[25] L. C. Carrillo-Arce, E. D. Nerurkar, J. L. Gordillo, and S. I. Roumeliotis. Decentralized
multi-robot cooperative localization using covariance intersection. In IEEE/RSJ Int. Conf.
on Intelligent Robots & Systems, pages 1412–1417, Tokyo, Japan, 2013.
[26] P. O. Arambel, C. Rago, and R. K. Mehra. Covariance intersection algorithm for distributed
spacecraft state estimation. In American Control Conference, pages 4398–4403, Arlington,
VA, 2001.
[27] H. Li and F. Nashashibi. Cooperative multi-vehicle localization using split covariance
intersection filter. IEEE Intelligent Transportation Systems Magazine, 5(2):33–44, 2013.
[28] D. Marinescu, N. O’Hara, and V. Cahill. Data incest in cooperative localisation with the
common past-invariant ensemble kalman filter. pages 68–76, Istanbul, Turkey, 2013.
[29] S. J. Julier and J. K. Uhlmann. A non-divergent estimation algorithm in the presence of
unknown correlations. In American Control Conference, pages 2369–2373, Albuquerque,
NM, 1997.
[30] S. J. Julier and J. K. Uhlmann. Simultaneous localisation and map building using split
covariance intersection. In IEEE/RSJ Int. Conf. on Intelligent Robots & Systems, pages
30
1257–1262, Maui, HI, 2001.
[31] N. Trawny, S. I. Roumeliotis, and G. B. Giannakis. Cooperative multi-robot localization
under communication constraints. In IEEE Int. Conf. on Robotics and Automation, pages
4394–4400, Kobe, Japan, May 2009.
[32] K. Y. K. Leung, T. D. Barfoot, and H. H. T. Liu. Decentralized localization of sparselycommunicating robot networks: A centralized-equivalent approach. IEEE Transactions on
Robotics, 26(1):62–77, 2010.
[33] L. Paull, M. Seto, and J. J. Leonard. Decentralized cooperative trajectory estimation for
autonomous underwater vehicles. In IEEE/RSJ Int. Conf. on Intelligent Robots & Systems,
pages 184–191, 2014.
[34] S. S. Kia, S. Rounds, and S. Mart/’/inez. Cooperative Localization under message dropouts
via a partially decentralized EKF scheme. In IEEE Int. Conf. on Robotics and Automation,
Seattle, WA, May 2015.
[35] H. Durrant-Whyte and T. Bailey. Simultaneous localization and mapping: Part i. 13(2):99–
110, 2006.
31
Algorithm 1 EKF CL (centralized)
Require: Initialization (k = 0): For i ∈ V, the algorithm is initialized at
i
x̂i+(0) ∈ Rn, Pi+(0) ∈ Mni , P+
ij (0) = 0ni ×nj , j ∈ V\{i}.
Iteration k
1: Propagation: for i ∈ V, the propagation equations are:
x̂i- (k+1) = f i (x̂i+ (k), ui (k)),
(16a)
Pi- (k+1) = Fi (k)Pi+ (k)Fi (k)>+Gi (k)Qi (k)Gi (k)>,
(16b)
j
>
P-ij (k+1) = Fi (k)P+
ij (k)F (k) , j ∈ V\{i}.
(16c)
2: Update: While there are no relative measurements no update happens, i.e.,
x̂+ (k + 1) = x̂- (k + 1),
P+ (k + 1) = P- (k + 1).
When there is a relative measurement at time-step k + 1, for example robot a makes a relative measurement of robot b, the update proceeds
as below. The innovation of the relative measurement and its covariance are, respectively,
ra = zab − hab (x̂a- (k + 1), x̂b- (k + 1)),
and
Sab = Ra (k + 1) + H̃a (k + 1)Pa- (k + 1)H̃a (k + 1)> + H̃b (k + 1)Pb- (k + 1)H̃b (k + 1)>
− H̃b (k + 1)P-ba (k + 1)H̃a (k + 1)> − H̃a (k + 1)P-ab (k + 1)H̃b (k + 1)> .
(17)
The estimation updates for the centralized EKF are:
x̂i+ (k+1) =x̂i- (k+1) + Ki (k+1)ra (k+1),
(18a)
Pi+ (k+1) =Pi-(k+1)−Ki (k+1)Sab (k+1)Ki (k+1)>,
(18b)
>
P+
ij (k+1) =Pij(k+1)−Ki (k+1)Sab (k+1)Kj (k+1) ,
(18c)
>
>
Ki = (P-ib (k + 1)H̃b − P-ia (k + 1)H̃a )Sab −1 .
(19)
where i ∈ V, j ∈ V\{i} and
3: k ← k + 1
32
Algorithm 2 Interim Master D-CL
Require: Initialization (k = 0): Every agent i ∈ V initializes its filter at
i
x̂i+ (0) ∈ Rn , Pi+ (0) ∈ Mni , Φi (0) = Ini ,
Πijl (0) = 0nl ×nj , j ∈ V\{N }, l ∈ {j + 1, · · · , N }.
(20)
Iteration k
1: Propagation: Every agent i ∈ V propagates the variables below
x̂i- (k+1) = f i (x̂i+ (k), ui (k)),
Pi- (k+1) = Fi(k)Pi+(k)Fi(k)>+Gi(k)Qi(k)Gi(k)>,
Φi (k+1) = Fi (k)Φi (k).
(21)
2: Update: while there are no relative measurements in the network, every agent i ∈ V updates its variables as:
x̂i+ (k + 1) = x̂i- (k + 1),
Pi+ (k + 1) = Pi- (k + 1),
Πijl (k + 1) = Πilj (k), j ∈ V\{N }, l ∈ {j + 1, · · · , N }.
(22)
If there is an agent a that makes a measurement with respect to another agent b, then agent a is declared as the interim master and acquires
the following information from agent b:
landmark-message = x̂b- (k + 1), Φb (k + 1), Pb- (k + 1) .
(23)
Agent a makes the following calculations upon receiving the landmark-message:
ra = zab − hab (x̂a- , x̂b- ),
>
(24a)
>
>
>
b>
a>
Sab = Ra + H̃a Pa- H̃a + H̃b Pb- H̃b − H̃a Φa Πa
H̃b − H̃b Φb Πa
H̃a ,
ab Φ
ba Φ
>
>
>
1
b
Γa = ((Φa )−1 Φa Πa
H̃b − (Φa )−1 Pa- H̃a )Sab − 2 ,
ab Φ
>
(24b)
>
1
a>
Γb = ((Φb )−1 Pb- H̃b − Πa
H̃a )Sab − 2 ,
ba Φ
(24c)
where H̃a (k + 1) = H̃a (x̂a- , x̂b- ) and H̃b (k + 1) = H̃b (x̂a- , x̂b- ) are obtained using (6).
The interim master passes the following data, either directly or indirectly (by message passing), to the rest of the agents in the network:
1
1
> >
>
update-message = a, b, r̄a , Γa , Γb , Φb H̃b Sab − 2 , Φa> H̃a Sab − 2 .
(25)
Every agent i ∈ V, upon receiving the update-message, first calculates, ∀j ∈ V\{a, b}, using information obtained at k:
>
>
1
>
1
Γj = Πijb Φb H̃b Sab − 2 − Πija Φa> H̃a Sab− 2 ,
(26)
and then updates the following variables:
x̂i+ (k+1) = x̂i-(k+1)+Φi (k+1) Γi r̄a ,
(27a)
i
>
Pi+ (k+1) = Pi-(k+1)−Φi (k+1)Γi Γ>
i Φ (k+1) ,
(27b)
Πijl (k+1) = Πijl(k)−Γj Γ>
l ,
(27c)
3: k ← k + 1
33
j ∈ V\{N }, l ∈ {j + 1, · · · , N }.
Algorithm 3 Alternative Interim Master D-CL (larger communication message size in favor of
lower computation and storage cost per agent)
Require: Initialization (k = 0): Every agent i ∈ V initializes its filter at
i
x̂i+ (0) ∈ Rn , Pi+ (0) ∈ Mni , Φi (0) = Ini ,
Πiij (0) = 0ni ×nj , j ∈ V\{i}.
(28)
Iteration k
1: Propagation: Every agent i ∈ V propagates the variables below
x̂i- (k+1) = f i (x̂i+ (k), ui (k)),
Pi- (k+1) = Fi(k)Pi+(k)Fi(k)>+Gi(k)Qi(k)Gi(k)>,
Φi (k+1) = Fi (k)Φi (k).
(29)
2: Update: while there are no relative measurements in the network, every agent i ∈ V updates its variables as:
x̂i+ (k + 1) = x̂i- (k + 1),
Pi+ (k + 1) = Pi- (k + 1),
Πiij (k + 1) = Πiij (k), j ∈ V\{i}.
If there is an agent a that makes a measurement with respect to another agent b, then agent a is declared as the interim master and acquires
the following information from agent b:
landmark-message = x̂b- (k + 1), Φb (k + 1), Pb- (k + 1), Πbbj (k) where j ∈ V\{a, b} .
(30)
Agent a makes the following calculations upon receiving the landmark-message:
ra = zab − hab (x̂a- , x̂b- ),
>
(31a)
>
>
>
>
b
> a>
Sab = Ra + H̃a Pa- H̃a + H̃b Pb- H̃b − H̃a Φa Πa
H̃b − H̃b Φb (Πa
H̃a ,
ab Φ
ab ) Φ
>
>
1
b>
Γa = ((Φa )−1 Φa Πa
H̃b − (Φa )−1 Pa- H̃a )Sab − 2 ,
ab Φ
>
>
1
>
1
Γj = (Πbbj )> Φb H̃b Sab − 2 − (Πiaj )> Φa> H̃a Sab − 2 ,
>
(31b)
>
1
> a>
Γb = ((Φb )−1 Pb- H̃b − (Πa
H̃a )Sab − 2 ,
ab ) Φ
j ∈ V\{a, b},
(31c)
(31d)
where H̃a (k + 1) = H̃a (x̂a- , x̂b- ) and H̃b (k + 1) = H̃b (x̂a- , x̂b- ) are obtained using (6).
The interim master passes the following data, either directly or indirectly (by message passing), to the rest of the agents in the network:
update-message = a, b, r̄a , Γ1 , · · · , ΓN .
(32)
Every agent i ∈ V, upon receiving the update-message, updates the following variables:
x̂i+ (k+1) = x̂i-(k+1)+Φi (k+1) Γi r̄a ,
(33a)
i
>
Pi+ (k+1) = Pi-(k+1)−Φi (k+1)Γi Γ>
i Φ (k+1) ,
(33b)
Πiij (k+1) = Πiij(k)−Γi Γ>
j ,
(33c)
3: k ← k + 1
34
j ∈ V\{i}.
TABLE I: Complexity analysis per agent of the Interim Master D-CL algorithm (denoted by
IM-D-CL) compared to that of the trivial decentralized implementation of EKF for CL (denoted
by T-D-CL) introduced in Subsection .
Computation
Algorithm
Propagation
Update per Nz
IM-D-CL
T-D-CL
O(1)
O(N 2 )
O(Nz ×N 2 ) O(Nz ×N 2 )
Broadcast?
Storage
Message Size
IM-D-CL T-D-CL IM-D-CL T-D-CL IM-D-CL T-D-CL
O(N 2 ) O(N 2 )
O(N 2 ) O(N 2 )
0
O(N )
O(Nz ) O(Nz )
relative measur.
0
O(1)
O(1)
O(1)
Connectivity
IM-D-CL
T-D-CL
None
strongly
interim mas-
connected
ter can reach
digraph
O(N 4 )
O(N 4 )
O(N 2 ) O(N 2 ) O(N 2 ) O(N 2 )
O(1)
O(1) all the agents
Overall worst case
cost is for multi-hop communication. If the communication range is unbounded, the broadcast cost per agent is O(max{Nb , Na })
∗ Broadcast
with the worst cost of O(N ).
35
Figure 1: Schematic representation of common probabilistic localization techniques for mobile platforms: In
beacon-based localization, the map of the area is known and there are pre-installed beacons or landmarks with
known locations and identities. By taking relative measurements with respect to these landmarks, the mobile agents
can improve their localization accuracy. For operations where a priori knowledge about the environment is not
available, but nevertheless, the environment contains fixed and distinguishable features that agents can measure,
SLAM is normally used to localize the mobile agents. SLAM is a process by which a mobile agent can build a map
of an environment and at the same time use this map to deduce its location. On the other hand, GPS navigation
provides location and time information in all weather conditions, anywhere on or near the earth but it requires an
clear line of sight to at least four GPS satellites.
36
Figure 2: This figure depicts a multi-hop communication scenario for the multi-robot team. Plot (a) shows the
communication and measurement ranges. Here, robots 1 and 6 make relative measurements, respectively, of robots
2 and 3. Plot (b) shows the communication graph generated using the communication ranges given in plot (a). Here
the robot at the head of an arrow can send information to the robot at the tip of the arrow. As this graph shows,
each of robots 1 and 6 can pass communication message to the entire team via a multi-hop strategy.
37
1→1
3→1 1→3, 3→2, 2→1
1→2
10
x1 error [m]
x1 error [m]
1→2
30
20
10
0
-10
-20
-30
0
50
100
150
200
250
0
50
t [sec]
100
150
200
250
150
200
250
150
200
250
t [sec]
30
20
10
0
-10
-20
-30
10
x2 error [m]
x2 error [m]
1→1
-10
0
0
-10
0
50
100
150
200
250
0
50
t [sec]
100
t [sec]
30
20
10
0
-10
-20
-30
10
x3 error [m]
x3 error [m]
3→1 1→3, 3→2, 2→1
0
-10
0
50
100
150
200
250
t [sec]
0
50
100
t [sec]
Figure 3: Estimation error (solid line) and 3σ error bounds (dashed lines) in the x−coordinate variable for 3 robots
moving on a flat terrain when they (a) only propagate their equations of motion using self-motion measurements
(black plots), (b) employ cooperative localization ignoring past correlations between the estimations of the robots
(blue plots), (c) employ cooperative localization with accurate account of past correlations (red plots). The figures
on the right column are the same figures as on the left where the localization case (a) is removed for clearer
demonstration of cases (b) and (c). Here, a → b over the time interval marked by two vertical blue lines indicates
that robot a has taken a relative measurement with respect to robot b at that time interval. The symbol a → a
means that robot a obtains an absolute measurement.
38
10
10
10
0
0
0
-20
y 3 , ŷ 3 [m]
y 2 , ŷ 2 [m]
y 1 , ŷ 1 [m]
-10
-10
-20
-30
-30
-10
-20
-30
-40
-40
-40
-50
-50
-50
-30
-20
-10
0
1
1
x , x̂ [m]
Robot 1
10
20
30
-30
-20
-10
0
2
2
x , x̂ [m]
Robot 2
10
20
30
-30
-20
-10
0
10
20
30
x3 , x̂3 [m]
Robot3
Figure 4: Trajectories of the robots for the simulation study of Fig. 3. Here, the gray curve is the ground truth.
The red curve is the estimation of the trajectory by implementing a EKF CL. The blue (resp. green) ellipses show
the 95% uncertainty regions for the estimations at 2 seconds before (resp. after) any change in the measurement
scenario (see Fig. 3)
39
Figure 5: Schematic representation of the D-CL classification based on how the past correlations are accounted
for.
40
Figure 6: The in-network information flow of the Interim Master D-CL algorithm. In the Interim Master D-CL
algorithm, communication is only needed in the update stage when the team members use a robot-to-robot relative
measurement feedback to correct their pose estimation. Here, we assume that all the team members are in the
communication range of the interim master robot.
41
position RMSE [m]
50
40
30
20
10
0
0
50
100
150
200
150
200
150
200
t [sec]
position RMSE [m]
Robot 1
50
40
30
20
10
0
0
50
100
t [sec]
position RMSE [m]
Robot 2
20
15
10
5
0
0
50
100
t [sec]
Robot3
Figure 7: A comparison study between the positioning accuracy of 3 robots employing the Interim Master D-CL
algorithm (red plots), with that from the EKF Covariance-Intersection based CL algorithm of [25] (dashed green
plot). The curves in black show the positioning accuracy when the robots do not use any CL. As expected, the
Interim Master D-CL algorithm by keeping an accurate account of the cross-covariances produces more accurate
localization results than the algorithm of [25] . However, this higher accuracy comes with higher communication
and processing cost per robot. Notice here that using algorithm of [25] robot 3 does not get to update its estimation
equations.
42
Sidebar 1
Further Reading
A performance analysis of an EKF CL for a team of homogeneous robots moving on a flat terrain,
with the same level of uncertainty in their proprioceptive measurements and exteroceptive sensors
that measure relative pose, is provided in [S1] and [S2]. Interestingly, [S1] shows that the rate
of uncertainty growth decreases as the size of the robot team increases, but is subject to the law
of diminishing returns. Moreover, [S2] shows that the upper bound on the rate of uncertainty
growth is independent of the accuracy or the frequency of the robot-to-robot measurements.
The consistency of EKF CL from the perspective of observability is studies in [S3]. Huang et
al. in [S3] analytically show that the error-state system model employed in the standard EKF
CL always has an observable subspace of higher dimension than that of the actual nonlinear CL
system. This results in an unjustified reduction of the EKF covariance estimates in directions of
the state space where no information is available, and thus leads to inconsistency. To address this
problem, Huang et al. in [S3] adopt an observability-based methodology for designing consistent
estimators in which the linearization points are selected to ensure a linearized system model with
an observable subspace of the correct dimension. More results on observability analysis of CL
can be found in [22], [S4], [S5]. The use of an observability analysis to explicitly design an
active local path planning algorithm for unmanned aerial vehicles implementing a bearing-only
CL is discussed in [S6]. The necessity for an initialization procedure for CL is discussed in [S7].
There it is shown that, because of system nonlinearities and the periodicity of the orientation,
initialization errors can lead to erroneous results in covariance-based filters. An initialization
43
procedure for the state estimation in a CL scenario based on ranging and dead reckoning is
studied in [S8].
References
[S1] S. I. Roumeliotis and A. I. Mourikis. Propagation of uncertainty in cooperative multirobot
localization: Analysis and experimental results. Autonomous Robots, 17(1):1573–7527,
2004.
[S2] A. I. Mourikis and S. I. Roumeliotis. Performance analysis of multirobot cooperative
localization. IEEE Transactions on Robotics, 22(4):666–681, 2006.
[S3] G. Huang, N. Trawny, A. I. Mourikis, and S. I. Roumeliotis. Observability-based consistent
EKF estimators for multi-robot cooperative localization. Autonomous Robots, 30(1):37–58,
2011.
[S4] A. Martinelli and R. Siegwart. Observability analysis for mobile robot localization. In
IEEE/RSJ Int. Conf. on Intelligent Robots & Systems, pages 1471–1476, 2005.
[S5] R. Sharma, R. W. Beard, C. N. Taylor, and S. Quebe. Graph-based observability analysis
of bearing-only cooperative localization. IEEE Transactions on Robotics, 28(2):522–529,
2012.
[S6] R. Sharma. Bearing-Only Cooperative-Localization and Path-Planning of Ground and
Aerial Robots. PhD thesis, Brigham Young University, 2011.
[S7] N. Trawny. Cooperative localization: On motion-induced initialization and joint state
estimation under communication constraint. PhD thesis, University of Minnesota, 2010.
44
[S8] J. Nilsson and P. Händel. Recursive bayesian initialization of localization based on ranging
and dead reckoning. In IEEE/RSJ Int. Conf. on Intelligent Robots & Systems, pages 1399–
1404, 2013.
45
Authors Information
Solmaz S. Kia is an Assistant Professor in the Department of Mechanical and Aerospace
Engineering, University of California, Irvine (UCI). She obtained her Ph.D. degree in Mechanical
and Aerospace Engineering from UCI, in 2009, and her M.Sc. and B.Sc. in Aerospace
Engineering from the Sharif University of Technology, Iran, in 2004 and 2001, respectively.
She was a senior research engineer at SySense Inc., El Segundo, CA from Jun. 2009-Sep. 2010.
She held postdoctoral positions in the Department of Mechanical and Aerospace Engineering at
the UC San Diego and UCI. Dr. Kia’s main research interests, in a broad sense, include distributed
optimization/coordination/estimation, nonlinear control theory and probabilistic robotics.
Stephen Rounds is a Research Engineer with NavCom Technology, a John Deere company.
He is responsible for identifying and supporting new navigation technologies for the company and
developing and maintaining the intellectual property portfolio of the company. Prior to working
with John Deere, Mr. Rounds worked with multiple defense contractors, with a special emphasis
on GPS-denied navigation, GPS anti-jamming protection, and other sensor fusion applications.
Mr. Rounds holds a B.S. degree in physics from Stevens Institute of Technology in Hoboken,
N.J., and an M.S. degree in nuclear physics from Yale University. He is the Chairman of the
Southern California section of the Institute Of Navigation, holds multiple patents in the navigation
field, and has numerous publications in various aerospace and navigation forums.
Sonia Martı́nez is a Professor with the department of Mechanical and Aerospace
Engineering at the University of California, San Diego. Dr. Martinez received her Ph.D. degree
46
in Engineering Mathematics from the Universidad Carlos III de Madrid, Spain, in May 2002.
Following a year as a Visiting Assistant Professor of Applied Mathematics at the Technical
University of Catalonia, Spain, she obtained a Postdoctoral Fulbright Fellowship and held
appointments at the Coordinated Science Laboratory of the University of Illinois, UrbanaChampaign during 2004, and at the Center for Control, Dynamical systems and Computation
(CCDC) of the University of California, Santa Barbara during 2005. In a broad sense, Dr.
Martı́nez’s main research interests include the control of networked systems, multi-agent systems,
nonlinear control theory, and robotics. For her work on the control of underactuated mechanical
systems she received the Best Student Paper award at the 2002 IEEE Conference on Decision
and Control. She was the recipient of a NSF CAREER Award in 2007. For the paper “Motion
coordination with Distributed Information,” co-authored with Jorge Cortés and Francesco Bullo,
she received the 2008 Control Systems Magazine Outstanding Paper Award. She has served on
the editorial boards of the European Journal of Control (2011-2013), and currently serves on
the editorial board of the Journal of Geometric Mechanics and IEEE Transactions on Control of
Networked Systems.
47
| 3cs.SY
|
Object category learning and retrieval with
weak supervision
arXiv:1801.08985v1 [cs.CV] 26 Jan 2018
Steven Hickson, Anelia Angelova, Irfan Essa, Rahul Sukthankar
Google Brain / Google Research
(shickson, anelia, irfanessa, sukthankar)@google.com
Abstract
We consider the problem of retrieving objects from image data and learning to
classify them into meaningful semantic categories with minimal supervision. To
that end, we propose a fully differentiable unsupervised deep clustering approach
to learn semantic classes in an end-to-end fashion without individual class labeling
using only unlabeled object proposals. The key contributions of our work are 1)
a kmeans clustering objective where the clusters are learned as parameters of the
network and are represented as memory units, and 2) simultaneously building a
feature representation, or embedding, while learning to cluster it. This approach
shows promising results on two popular computer vision datasets: on CIFAR10 for
clustering objects, and on the more complex and challenging Cityscapes dataset
for semantically discovering classes which visually correspond to cars, people, and
bicycles. Currently, the only supervision provided is segmentation objectness masks,
but this method can be extended to use an unsupervised objectness-based object
generation mechanism which will make the approach completely unsupervised.
1
Introduction
Unsupervised discovery of common patterns is a long standing task for artificial intelligence as
shown in Barlow (1989); Bengio, Courville, and Vincent (2012). Recent deep learning approaches
have offered major breakthroughs in classification into multiple categories with millions of labeled
examples (e.g. Krizhevsky (2009); Szegedy et al. (2015); He et al. (2016) and many others). These
methods rely on a lot of annotated data for training in order to perform well. Unfortunately, labeling is
an inefficient and expensive progress, so learning from unlabeled data is desirable for many complex
tasks. At the same time, much of human knowledge and learning is obtained by unsupervised
observations Grossberg (1994).
The goal of this work is to show that semantically meaningful classes can be learned with minimal
supervision. Given a set of objectness proposals, we use the activations of foreground objects in order
to learn deep features to cluster the available data while simultaneously learning the embedding in an
end-to-end manner. More specifically, we propose a differentiable clustering approach that learns
better separability of classes and embedding. The main idea is to store the potential cluster means
as weights in a neural network at the higher levels of feature representation. This allows them to be
learned jointly with the potential feature representation. This differentiable clustering approach is
integrated with Deep Neural Networks (e.g. Szegedy et al. (2015)) to learn semantic classes in an
end-to-end fashion without manual class labeling.
The idea of doing this ‘end-to-end’ is that gradient descent can not only learn good weights for
clustering, it can also change the embedding to allow for better clustering without the use of labels.
We see that this leads to better feature representation. Our results show also that different object
categories emerge and can later be retrieved from test images never before seen by the network,
resulting in clusters of meaningful categories, such as cars, persons, bicycles.
31st Conference on Neural Information Processing Systems (NIPS 2017), Long Beach, CA, USA.
In this work we use given segmentation objectness masks, which are candidate objects without labels.
This can be extended by using an independent objectness-based object generation mechanism Pathak
et al. (2017); Faktor and Irani (2014) or by using unsupervised motion segmentation in videos or
structure from motion Vijayanarasimhan et al. (2017).
2
Related Work
Unsupervised learning (Barlow (1989)) and unsupervised deep learning (Bengio, Courville, and
Vincent (2012), Bengio (2012), Bengio and others (2009)) are central topics to Machine Learning.
Unsupervised deep learning has been shown to improve results on classification tasks per Erhan et
al. (2010), especially given small datasets and complicated high dimensional data such as video.
This has been explored by many representations including sequence to sequence learning and textual
representations (Radford, Jozefowicz, and Sutskever (2017), Ramachandran, Liu, and Le (2016)).
Our work focuses on unsupervised deep learning for discovering visual object categories. This has
also been shown to improve results such as in Doersch and Zisserman (2017). Unsupervised discovery
of visual objects has been a large topic of interest in computer vision (Sivic et al. (2005); Russell et
al. (2006); Singh, Gupta, and Efros (2012); Bach and Jordan (2005); Kwak et al. (2015); Pathak et al.
(2017)).
Building specialized, deep embeddings to help computer vision tasks is also a popular approach such
as in Agrawal, Carreira, and Malik (2015). Transfer learning from supervised tasks has proven to be
very successful. Further, Agrawal, Carreira, and Malik (2015) propose learning the lower dimensional
embedding through unsupervised learning and show improved performance when transfered to other
supervised tasks.
Despite the popularity of building different embeddings, there is little work investigating the use of
clustering to modify the embedding in an end-to-end deep learning framework. Bottou and Bengio
(1995) investigate a differentiable version of the kmeans algorithm and examine its convergence
properties. Our work focuses on learnable feature representations (instead of fixed ones as in Bottou
and Bengio (1995)) and introduces memory units for the task.
3
Unsupervised deep clustering
Our unsupervised deep clustering is inspired by Bottou and Bengio (1995), who consider differentiable
clustering algorithms.
We differ from this approach because the features we cluster also change with backpropogation. In
our work, we add a kmeans-like loss that is integrated end-to-end. Our idea is to store the potential
cluster means as weights in the network and thus have them be learned.
The proposed clustering is done simultaneously while building an embedding. Given information of a
potential object vs background (binary labels), clustering in a differentiable way provides a better
embedding for the input data. We show that this method can be used for meaningful semantic retrieval
of related objects.
3.1
Embedding with clustering
We train a convolutional neural network (CNN) to predict foreground and background using oracle
labels of patches of objects and background images. Concurrently, we learn the clustering of objects
by imposing constraints that will force the embedding to be partitioned into multiple semantically
coherent clusters of objects without explicit labels for different objects.
For our experiments, we use random initialization on the fully-connected layers (the last two layers)
and we add the differentiable clustering module after the second to last layer. Note that we only
cluster the foreground labels as background activations are not of interest for clustering; the classifier
can predict foreground vs background with high accuracy (above 90%).
The objective function is shown in Equation 1.
N
1 X
Lk =
mink [(xn − wk )2 ]
2N n=1
2
(1)
In this equation, N is the number of samples, k is the number of defined clusters, w is the “weight”
(theoretically and typically the mean of the cluster) for each k, and x is the activations from the fully
connected layer before the classification fully connected layer. This is differentiable and the gradient
descent algorithm is shown in Equation 2.
δwk = wk0 − wk =
N
X
lr (xn − wk ) if k = s(xn , w)
0
otherwise
n=1
(2)
where s(xn , w) = argmin
Pk [xn ] and lr is the learning rate. We also add L2 regularization over the
weights to the loss L2 = j wj2 . Furthermore, we use a custom clustering regularization loss LC
that enforces that the clusters are evenly distributed as defined in Equation 3 and Equation 4.
K K
1 XX
LC =
|countk − countj |
NK
(3)
k=0 j=k
countk =
N
X
0
n=0
1
if argmink [xn ] = 0
if argmink [xn ] = 1
(4)
The final loss to be optimized is shown in (Equation 5)
L = Lk + αr L2 + αc LC
(5)
where αr and αc are hyperparameters which are tuned during the training. For our method, we use
αr = 0.25 and αc = 1. We apply this loss to every point that is labeled as potentially an object
and ignore the background ones when clustering. This way we learn foreground vs background and
then learn clustering of the foreground activations. Optimization was performed with a ‘RMSProp’
optimizer, with a learning rate of 0.045, momentum 0.9, decay factor 0.9, and of 1.0.
4
Experimental evaluation
We experiment with a toy example using CIFAR10 and a more challenging example using Cityscapes.
4.1
CIFAR10 dataset
We first test the proposed unsupervised clustering approach on the CIFAR10 Krizhevsky (2009)
dataset. The goal of this experiment is to test if clustering can uncover separate categories in a simple
toy problem with a two class setting.
Clusters Automobile
Dog
Cluster 0
68.5%
17.9%
Cluster 1
31.5%
82.1%
Table 1: Unsupervised clustering results on CIFAR10 for discovery of two classes. Per cluster
accuracy for each of the two given classes on the test set (class labels are unknown during training).
We selected as an example the dog and automobile classes to label as foreground. We then train a
network from scratch based on the Network in Network architecture (NiN) of Lin, Chen, and Yan
(2013) from scratch for our experiments. All other classes of CIFAR are considered background for
this experiment. By attaching our modified clustering objective function to the next to last layer, we
attempt to cluster dog and automobile without labels.
We can see in our simple experimental results that classes are naturally clustered with the majority
of examples correctly assigned. Table 1 shows quantitative results on the test set. As seen 68.5% of
the automobile classes and 82.1% of the dog examples are correctly assigned to separate clusters.
Note that in these cases, the concepts and classes of dog and automobile are unknown to the training
algorithm and we are just looking at them after clustering for evaluation.
3
Classes
Cluster 0 Cluster 1
Person
4320
138
Rider
676
138
Car
1491
4399
Truck
60
69
Bus
49
89
Train
17
16
Motorcycle
88
205
Bicycle
795
787
Table 2: Unsupervised clustering of objects from Cityscapes using our method. The table shows
number of examples assigned to each learned cluster (for K=2).
4.2
Cityscapes dataset
The Cityscapes dataset (Cordts et al. (2016)) is a large-scale dataset that is used for evaluating various
classification, detection, and segmentation algorithms related to autonomous driving. It contains 2975
training, 500 validation, and 1525 test images, where the test set is provided for the purposes of the
Cityscape competition only. In this work, we used the training set for training and the validation set
for testing and visualizing results (as the test set has no annotation results). Annotation is provided
for classes which represent moving agents in the scene, such as pedestrian, car, motorcycle, bicycle,
bus, truck, rider, train. In this work we only use foreground/background labels and intend to discover
semantic groups from among the moving objects.
4.3
Weakly supervised discovery of classes
In this experiment we considered the larger, real-life dataset, Cityscapes (Cordts et al. (2016)),
described above to see if important class categories, e.g. the moving objects in the scene can be
clustered into semantically meaningful classes. We extract the locations and extents of the moving
objects and use that as weak supervision. Note the classes are uneven and car and person dominate.
We show results clustering 8 categories into 2 and 3 clusters despite the rarity of some of them (such
as bicycle). All results below are presented on the validation set. We report the results in terms of
the number of object patches extracted from the available test images. For this dataset, the CNN
architecture is based on the Inception architecture proposed by Szegedy et al. (2015). Since there are
a small number of examples, we pre-train only the convolutional layers of the network.
Results on clustering the 8 classes of moving objects into 2 and 3 clusters are presented in Table 2 and
Table 3 respectively for the learned embedding by the proposed approach and the baseline embedding.
The baseline embedding is calculated by fine-tuning the same architecture in the same manner, but
without our loss (Equation 5); it uses the same amount of information as input as our embedding.
For this experiment, we apply standard kmeans on both activations after training is completed. We
see here that our method provides better clustering for the two dominant classes in the dataset (car
and person). On the other hand, the baseline embedding clusters on one class only, similar to the two
class case. We have consistently observed this behavior for different runs and hypothesize this is due
to the sparse nature of the baseline embedding and it’s activations.
Figure 1 visualizes the three retrieved clusters (color-coded) when clustering into 3 clusters with our
approach. We can see that people (in blue) and cars (in green) are often correctly retrieved. Bikes
are more rare and may be more often mistaken, for example in cases where a portion of the patch
contains part of a car, or since the bicycle very often has a person riding it. Still this is exciting result,
given that it is learned by not providing a single class label during training.
5
Conclusions
We propose a differentiable clustering objective which learns to separate classes during learning and
build a better embedding. The key idea is to be able to learn the clusters which are stored as weights,
and simultaneously learn the feature representation and the clustering of the data. Our results show
that the proposed approach is useful for extracting semantically related objects.
4
Our method
Baseline
Classes
Cluster 0 Cluster 1 Cluster 2 Cluster 0 Cluster 1 Cluster 2
Person
151
4315
17
4482
1
0
Rider
258
551
7
816
0
0
Car
5195
950
180
6312
13
0
Truck
89
39
5
131
2
0
Bus
127
20
5
152
0
0
Train
25
9
1
35
0
0
Motorcycle
127
76
4
207
0
0
Bicycle
1128
541
450
2119
0
0
Table 3: Unsupervised clustering on the Cityscapes dataset with 3 clusters. The table shows the
number of examples assigned to each learned cluster. Our method (left) and baseline (right). Our
method results in 69.98% accuracy.
Figure 1: Visualization of clusters learned by our method (for K=3). From the figure, the green class
is responsible for retrieving cars, the blue one persons, and the red one bicycles. We can see that both
cars and persons are discovered well but bicycles, a rarer class, can be confused with a person or with
a partially visible car in the background.
5
References
Agrawal, P.; Carreira, J.; and Malik, J. 2015. Learning to see by moving. CVPR.
Bach, F. R., and Jordan, M. I. 2005. Learning spectral clustering. NIPS.
Barlow, H. 1989. Unsupervised learning. Neural computation.
Bengio, Y., et al. 2009. Learning deep architectures for ai. Foundations and trends R in Machine Learning
2(1):1–127.
Bengio, Y.; Courville, A. C.; and Vincent, P. 2012. Unsupervised feature learning and deep learning: A review
and new perspectives. CoRR, abs/1206.5538.
Bengio, Y. 2012. Deep learning of representations for unsupervised and transfer learning. In Proceedings of
ICML Workshop on Unsupervised and Transfer Learning, 17–36.
Bottou, L., and Bengio, Y. 1995. Convergence properties of the k-means algorithms. In Advances in neural
information processing systems, 585–592.
Cordts, M.; Omran, M.; Ramos, S.; Rehfeld, T.; Enzweiler, M.; Benenson, R.; Franke, U.; Roth, S.; and Schiele,
B. 2016. The cityscapes dataset for semantic urban scene understanding. In Proceedings of the IEEE
Conference on Computer Vision and Pattern Recognition, 3213–3223.
Doersch, C., and Zisserman, A.
arXiv:1708.07860.
2017.
Multi-task self-supervised visual learning.
arXiv preprint
Erhan, D.; Bengio, Y.; Courville, A.; Manzagol, P.-A.; Vincent, P.; and Bengio, S. 2010. Why does unsupervised
pre-training help deep learning? Journal of Machine Learning Research 11(Feb):625–660.
Faktor, A., and Irani, M. 2014. Video segmentation by non-local consensus voting. BMVC.
Grossberg, S. 1994. 3-d vision and figure-ground separation by visual cortex. Perception and Psychophysics.
He, K.; Zhang, X.; Ren, S.; and Sun, J. 2016. Deep residual learning for image recognition. CVPR.
Krizhevsky, A. 2009. Learning multiple layers of features from tiny images.
Kwak, S.; Cho, M.; Laptev, I.; Ponce2, J.; and Schmid, C. 2015. Unsupervised object discovery and tracking in
video collections. ICCV.
Lin, M.; Chen, Q.; and Yan, S. 2013. Network in network. arXiv preprint arXiv:1312.4400.
Pathak, D.; Girshick, R.; Dollar, P.; Darrell, T.; and Hariharan, B. 2017. Learning features by watching objects
move. CVPR.
Radford, A.; Jozefowicz, R.; and Sutskever, I. 2017. Learning to generate reviews and discovering sentiment.
arXiv preprint arXiv:1704.01444.
Ramachandran, P.; Liu, P. J.; and Le, Q. V. 2016. Unsupervised pretraining for sequence to sequence learning.
arXiv preprint arXiv:1611.02683.
Russell, B. C.; Efros, A. A.; Sivic, J.; Freeman, W. T.; and Zisserman, A. 2006. Using multiple segmentations to
discover objects and their extent in image collections. CVPR.
Singh, S.; Gupta, A.; and Efros, A. A. 2012. Unsupervised discovery of mid-level discriminative patches. ECCV.
Sivic, J.; Russell, B. C.; Efros, A. A.; Zisserman, A.; and Freeman, W. T. 2005. Discovering objects and their
location in images. ICCV.
Szegedy, C.; Liu, W.; Jia, Y.; Sermanet, P.; Reed, S.; Anguelov, D.; Erhan, D.; Vanhoucke, V.; and Rabinovich,
A. 2015. Going deeper with convolutions. CVPR.
Vijayanarasimhan, S.; Ricco, S.; Schmid, C.; Sukthankar, R.; and Fragkiadaki, K. 2017. Sfm-net: Learning of
structure and motion from video.
6
| 1cs.CV
|
Distributed Kalman Filter in a Network of Linear Dynamical Systems
Damián Marellia,b , Mohsen Zamanic , Minyue Fuc,a
a Department
arXiv:1711.07625v1 [cs.SY] 21 Nov 2017
of Control Science and Engineering and State Key Laboratory of Industrial Control Technology, Zhejiang University, 388
Yuhangtang Road Hangzhou, Zhejiang Province, 310058, P. R. China.
b French-Argentinean International Center for Information and Systems Sciences, National Scientific and Technical Research Council, Ocampo
Esmeralda, Rosario 2000, Argentina.
c School of Electrical Engineering and Computer Science, University of Newcastle, University Drive, Callaghan, NSW 2308, Australia.
Abstract
This paper is concerned with the problem of distributed Kalman filtering in a network of interconnected subsystems
with distributed control protocols. We consider networks, which can be either homogeneous or heterogeneous, of
linear time-invariant subsystems, given in the state-space form. We propose a distributed Kalman filtering scheme for
this setup. The proposed method provides, at each node, an estimation of the state parameter, only based on locally
available measurements and those from the neighbor nodes. The special feature of this method is that it exploits the
particular structure of the considered network to obtain an estimate using only one prediction/update step at each time
step. We show that the estimate produced by the proposed method asymptotically approaches that of the centralized
Kalman filter, i.e., the optimal one with global knowledge of all network parameters, and we are able to bound the
convergence rate. Moreover, if the initial states of all subsystems are mutually uncorrelated, the estimates of these
two schemes are identical at each time step.
Keywords: Estimation, Kalman Filter, Distributed Systems.
1. Introduction
There has been an increasing activity in the study of
distributed estimation in a network environment. This
is due to its broad applications in many areas, including
formation control Subbotin and Smith [16], Lin et al.
[8], distributed sensor network Zhang et al. [23] and
cyber security Teixeira et al. [18], Zamani et al. [22].
This paper examines the problem of distributed estimation in a network of subsystems represented by a finite
dimensional state-space model. Our focus is on the scenario where each subsystem obtains some noisy measurements, and broadcasts them to its nearby subsystems, called neighbors. The neighbors exploit the received information, together with an estimate of their
internal states, to make a decision about their future
states. This sort of communication coupling arises in
different applications. For example, in control system
security problems Teixeira et al. [18], distributed state
estimation is required to calculate certain estimation er-
Email addresses: [email protected]
(Damián Marelli), [email protected] (Mohsen
Zamani), [email protected] (Minyue Fu)
Preprint submitted to Systems and Control Letters
ror residues for attack detection. Similarly, for formation control Lin et al. [10], Zheng et al. [24], Lin et al.
[9], each subsystem integrates measurements from its
nearby subsystems, and states of each subsystem need
to be estimated for distributed control design purposes.
The main objective of this paper is to collectively estimate the states of all subsystems within such a network.
We will propose a novel distributed version of the celebrated Kalman filter.
The current paper, in broad sense, belongs to the large
body of literature regarding distributed estimation. One
can refer to Lopes and Ali [11], Kar et al. [6], Conejo
et al. [4], Gómez-Expósito et al. [5], Marelli and Fu
[12], Olfati-Saber [13], Ugrinovskii [19, 20], Zamani
and Ugrinovskii [21], Khan and Moura [7], Olfati-Saber
[14] and the survey paper Ribeiro et al. [15], as well as
references listed therein, for different variations of distributed estimation methods among a group of subsystems within a network. A consensus based Kalman filter
was proposed in Olfati-Saber [13]. The author of Ugrinovskii [19] utilized a linear matrix inequality to minimize a H∞ index associated with a consensus based estimator, which can be implemented locally. Some of the
results there were then extended to the case of switching
November 22, 2017
topology in Ugrinovskii [20]. The same problem was
solved using the minimum energy filtering approach
in Zamani and Ugrinovskii [21]. A common drawback
of the state estimation methods described above is that,
being based on consensus, they require, in theory, an infinite number of consensus iterations at each time step.
This results in computational and communication overload. To avoid this, in this paper we exploit the network
structure to achieve a distributed Kalman filter method
which requires only one prediction/update step at each
time step. Moreover, it is worthwhile noting that there is
a major difference between the above-mentioned works
and the problem formulation in the current paper. More
precisely, in the former, the aim of each subsystem is
to estimate the aggregated state which is common to all
subsystems. In contrast, in the problem studied here,
each subsystem is dedicated to the estimation of its own
internal state, which in general is different from those
of other subsystems. This allows the distributed estimation algorithm to be scalable to networked systems with
a large number of subsystems where requiring each subsystem to estimate the aggregated state is both computationally infeasible and practically unnecessary.
To show the effectiveness of the proposed algorithm,
we compare our method with the classical (centralized)
Kalman filter, which is known to be optimal (in the minimum error covariance sense). The classical method requires the simultaneous knowledge of parameters and
measurements from all subsystems within the network
to carry out the estimation. In contrast, our proposed
distributed estimation algorithm runs a local Kalman filter at each subsystem, which only requires the knowledge of local measurements and parameters, as well as
measurements from neighbor subsystems. Hence, it can
be implemented in a fully distributed fashion. We show
that the state estimate, and its associated estimation error covariance matrix, produced by the proposed distributed method asymptotically converge to those produced by the centralized Kalman filter. We provide
bounds for the convergence of both the estimate and the
estimation error covariance matrix. A by-product of our
result is that, if the initial states of all subsystems are
uncoupled (i.e., they are mutually uncorrelated), the estimates produced by our method are identical to that of
the centralized Kalman filter.
The rest of the paper is structured as follows. In Section 2, we describe the network setup and its associated centralized Kalman filter. In Section 4, we describe
the proposed distributed Kalman filter scheme. In Section 5, we demonstrate the asymptotic equivalence between the proposed distributed filter and the centralized
one, and provide bounds for the convergence of the esti-
mates and their associated estimation error covariances.
Simulation results that support our theoretical claims
are presented in Section 6. Finally, concluding remarks
are given in Section 7.
2. System Description
In this paper we study networks of I linear timeinvariant subsystems. Subsystem i is represented by the
following state-space model
(i)
xk+1
=
(i)
A(i) xk(i) + z(i)
k + wk ,
y(i)
k
=
C (i) xk(i)
+
v(i)
k .
The subsystems are interconnected as follows
X
z(i)
=
L(i, j) y(kj) ,
k
(1)
(2)
(3)
j∈Ni
where xk(i) ∈ Rni is the state, y(i)
∈ R pi the outk
(i)
put, wk is an i.i.d Gaussian disturbance process with
(0, Qi ), and v(i)
w(i)
k ∼ N
k is an i.i.d. Gaussian measurement noise process with v(i)
∼ N (0, Ri ). We further
(i) ( j)> k
>
suppose that E wk wk
= 0 and E v(i)
v(kj) = 0,
k
>
>
∀i , j and E xk(i) wk( j) = 0, E xk(i) v(kj) = 0 ∀i, j.
We also
n denote theo neighbor set of the subsystem i by
Ni = j : L(i, j) , 0 .
Remark 1. We note in (1)-(2) that the coupling between neighboring subsystems is solely caused through
the zk(i) term in (3). The main motivation for considering such coupling comes from distributed control
where (1) represents the model of an autonomous subsystem (or agent) with the zk(i) being the control input
and that (3) represents a distributed control protocol
which employs feedback only from neighboring measurements. This type of distributed control is not only
common for control of multi-agent systems (see, for example, [8, 10, 9, 24]), but also realistic for large networked systems in the sense that only neighbouring information is both easily accessible and most useful for
each subsystem. We emphasize that the distributed state
estimation problem arises for the networked system (1)(3) because of our allowance for measurement noises v(i)
k
in (2). This consideration is very important for applications because measurement noises are unavoidable in
practice. This also sharply distinguishes our distributed
control formulation from most distributed control algorithms in the literature where perfect state measurement
is often implicitly assumed.
>
>
We define ξk> = ξk(1) , · · · , ξk(I)
and Ξk =
{ξ1 , · · · , ξk }, where (ξ, Ξ) stands for either (x, X), (y, Y),
2
(z, Z),n (w, W) or (v,
o V); moreover, we denote Υ =
diag Υ(1) , · · · , Υ(I) , where Υ stands for either A, B, C,
h
i
Q or R, and L = L(i, j) : i, j = 1, · · · , I .
Using the above notation, we let the initial state x1 of
all subsystems have the joint distribution x0 ∼ N (0, P).
We can also write the aggregated model of the whole
network as
xk+1
yk
=
Axk + LCxk + wk + BLvk
=
Ãxk + ek ,
(4)
= Cxk + vk ,
(5)
4. Distributed Kalman Filter
Consider the i-th subsystem (1)-(2). Notice that,
since the measurements y(kj) , j ∈ Ni , are known by the
i-th subsystem, they can be treated as external inputs.
This observation leads us to the following intuitive approach for a distributed Kalman filter scheme.
Let, for all i = 1, · · · , I and k, l ∈ N,
(i)
, E xk(i) |y(mj) ; j ∈ Ni ∪ {i}, m = 1, · · · , l ,
x̂k|l
h
i h (i)
i
(i)
(i)
(i) >
Σ(i)
,
E
x
−
x̂
x
−
x̂
,
k
k
k|l
k|l
k|l
with
à = A + LC
ek = wk + Lvk .
and
It then follows that
"
#
ek h >
ek
cov
vk
v>k
i!
"
=
Q̃
S̃ >
S̃
R
(6)
and Σ(i)
= P(i) . Then, the prediction and update steps
0|0
for the proposed distributed Kalman filter are:
#
,
(14)
(7)
1. Prediction:
(i)
x̂k+1|k
where Q̃ = Q + LRL> and S̃ = LR.
=
(i)
+
A(i) x̂k|k
X
Li,(i,j j) yk( j) ,
(15)
j∈Ni
Σ(i)
k+1|k
3. Centralized Kalman Filter
(i)
(i)
(i)
x̂k|k
= x̂k|k−1
+ Kk(i) yk(i) − C (i) x̂k|k−1
,
(i)
(i)
(i) (i)
Σk|k = I − Kk C Σk|k−1 ,
(8)
Since the distributed Kalman filter approach given in
Section 4 is motivated by intuition, the question naturally arises as to which extent it is optimal. In this
section
this question.
we address
h (i)>To this end, we
i de?
?
?>
fine x̂k|l , Σk|l , where x̂k|l = x̂k|l : i = 1, · · · , N and
(i)
Σ?k|l = diag Σk|l
: i = 1, · · · , N , to be the outcomes of
distributed filter and x̂k|l , Σk|l to be those of centralized
one. In Section 5.1, we show that the estimation error
covariance of the distributed filter Σ?k|k converges to that
of the centralized one Σk|k , and provide a bound for this
convergence. In Section 5.2, we do the same for the
?
convergence of x̂k|k
to x̂k|k .
(10)
= AΣk|k A + Q.
1. Update:
>
5.1. Convergence of Σ?k|k to Σk|k
(11)
In this section, we show that the covariance matrices
Σk|k and Σ?k|k exponentially converge to each other, and
(12)
with
−1
Kk = Σk|k−1C CΣk|k−1C > + R .
(18)
5. Optimality analysis
>
Σk+1|k = M − S̃ R−1C Σk|k M − S̃ R−1C
>
(17)
(i)>
(i)>
(i) (i)
(i) −1
C
C
Kk(i) = Σ(i)
C
Σ
+
R
. (19)
k|k−1
k|k−1
and
x̂k|k = x̂k|k−1 + Kk yk − C x̂k|k−1 ,
Σk|k = (I − Kk C) Σk|k−1 ,
(16)
with
and Σ0|0 = P. Our aim in this subsection is to compute
x̂k|k in a standard centralized way. Notice that equation (7) implies that, in the aggregated system formed
by (1)-(2), the process noise ek and the measurement
noise vk are mutually correlated. Taking this into account, it follows from [1, S 5.5] that the prediction and
update steps of the (centralized ) Kalman filter are given
by:
1. Prediction:
x̂k+1|k = Ã − S̃ R−1C x̂k|k + S̃ R−1 yk
(9)
= A x̂k|k + Lyk ,
+ Q̃ − S̃ R−1 S̃
>
(i) (i)
A(i) Σk|k
A + Q(i) .
2. Update:
Consider the standard (centralized) Kalman filter. For
all k, l ∈ N, let
x̂k|l , E (xk |Yl ) ,
Σk|l , E xk − x̂k|l xk − x̂k|l > ,
=
introduce a bound on Σk|k − Σ?k|k . To this end, we require the following definition from [3, Def 1.4].
(13)
3
Proof. Let Σ̄? = limk→∞ Σ?k|k and
n
o
σ̃ = max kPk , P? , Σ̄ , Σ̄? ,
(25)
−1
−1
. (26)
ω̃ = max P−1 , P? , Σ̄−1 , Σ̄?
Definition 2. For n × n matrices P, Q > 0, the Riemannian distance is defined by
v
t n
X
log2 σk PQ−1 ,
δ (P, Q) =
k=1
We can then appeal to the fact that the Riccati equation
is monotonic Bitmead et al. [2], to conclude that, for all
k ∈ N,
n
o
Σk|k ≤ max kPk , Σ̄ ≤ σ̃,
(27)
n
o
(28)
Σ?k|k ≤ max P? , Σ̄? ≤ σ̃,
n
o
−1
−1
Γk|k ≤ max P , Σ̄
≤ ω̃,
(29)
−1
−1
≤ ω̃.
(30)
Γ?k|k ≤ max P? , Σ̄?
where σ1 (X) ≥ · · · ≥ σn (X) denote the singular values
of matrix X.
Several properties of the above definition, which we
use to derive our results, are given in the following
proposition.
Proposition 3. [17, Proposition 6] For n × n matrices
P, Q > 0, the following holds true:
1. δ(P,
P) = 0.
2. δ P−1 , Q−1 = δ (Q, P) = δ (P, Q) .
3. For any m × m matrix W > 0 and m × n matrix B,
we have
α
δ(P, Q),
δ(W + BPB> , W + BQB> ) ≤
α+β
Recall that
Σk+1|k = AΣk|k A> + Q.
Also, from [1, p. 139], we have
Γk|k = Γk|k−1 + U.
Clearly, similar relations hold for Σ?k|l and Γ?k|l . Then, it
follows from Proposition 3-3 that,
δ Σk+1|k , Σ?k+1|k =δ AΣk|k A> + Q, AΣ?k|k A> + Q
≤υ̃1 δ Σk|k , Σ?k|k ,
(31)
δ Γk|k , Γ?k|k =δ Γk|k−1 + U, Γ?k|k−1 + U
≤υ̃2 δ Γk|k−1 , Γ?k|k−1 ,
(32)
where α = max{kBPB> k, kBQB> k} and β =
σmin (W).
4. If P > Q, then kP − Qk ≤ eδ(P,Q) − 1 kQk .
The main result of this section is given in Theorem 5.
Its proof requires the following technical result.
?
?
Lemma 4. Let Γk|l = Σ−1
k|l and Γk|l = Σk|l . Then
−1
Σk|k , Σ?k|k
≤
σ,
Γk|k , Γ?k|k
≤
ω,
with
υ̃1 =
and
δ Σk|k , Σ?k|k ≤
δ Γk|k , Γ?k|k ≤
υk δ P, P? ,
υk δ P, P? ,
(21)
ω
=
o
n
max kPk , P? , Σ̄ ,
−1
max P−1 , P? , Σ̄−1 ,
υ2
=
=
υ1 υ1 ,
υ2 =
ω
ω + U −1
−1
,
σ kAk2
σ kAk +
2
−1
Q−1
(23)
,
−1
Q−1
and
υ̃2 =
ω̃
ω̃ + U −1
−1
.
with υ̃ = υ̃1 υ̃2 . Finally, the above implies that Σ̄? = Σ̄.
Hence, the parameters σ̃ and ω̃ given in (25)-(26) are
equivalent to σ and ω in (22)-(23), respectively, and the
result follows.
(22)
with P? denoting the diagonal matrix formed by the
block diagonal entries of the matrix P,
υ
σ̃ kAk +
2
It then follows from (31)-(32) and Proposition 3-2, that
δ Σk|k , Σ?k|k ≤ υ̃k δ P, P? ,
δ Γk|k , Γ?k|k ≤ υ̃k δ P, P? .
(20)
where
σ =
σ̃ kAk2
We now introduce the main result of the section, stating a bound on Σk|k − Σ?k|k .
Theorem 5. Let Σ̃k|l = Σk|l − Σ?k|l and Γ̃k|l = Γk|l − Γ?k|l .
Then
(24)
U = C > R−1C,
Σ̃k|k ≤ κσυk
where
and Σ̄ = limk→∞ Σk|k .
4
and
Γ̃k|k ≤ κωυk ,
?
κ = eδ(P,P ) − 1.
Lemma 7. Let
Proof. Using (22)-(23), together with (20)-(21), Proposition 3-4 and Lemma 11, we obtain
k
?
Σ̃k|k ≤ eυ δ(P,P ) − 1 Σk|k ≤ κυk Σk|k
Γ̃k|k
>
∆k = E x̃k|k−1 x̃k|k−1
.
(34)
∆k ≤ Hk ∆k−1 Hk> + λυk I,
(35)
Then
≤
κσυk ,
≤
κυk Γk|k ≤ κωυk .
where I is the identity matrix, υ is defined in (24), and
!
q
λ , sup ζ + 2 ζ kHk k2 k∆k−1 k < ∞,
(36)
?
5.2. Convergence of x̂k|k
to x̂k|k
In this subsection, we study the convergence of state
?
estimate x̂k|k
, obtained through the distributed method,
and that of the centralized one x̂k|k . Moreover, we derive
?
a bound on the error x̂k|k - x̂k|k
. We start by introducing a
number of lemmas which are instrumental for establishing our main results.
k∈N
with
?
Lemma 6. Let x̃k|l = x̂k|l − x̂k|l
. Then
ζ
=
α
=
p
(α + β) + 2 αβ,
κ2 ω2 σ2 kAk2 σ kAk2 + kQk ,
β
=
κ ω σ kAk .
2
2
3
(37)
2
Hk
=
A I − Σk|k U ,
Proof. We split the argument in steps:
Step 1) From Lemmas 6 and 12
E ξk ξk>
≤ E ak a>k + E bk b>k
r
E ak a>k
E bk b>k .
+ 2
ξk
=
ak + bk ,
=
?
AΣk|k Γ̃k|k x̂k|k−1
,
Now, using Lemma 4,
ak
bk
=
?
.
AΣ̃k|k Γ?k|k x̂k|k
x̃k+1|k = Hk x̃k|k−1 + ξk .
(33)
where
E ak a>k ≤ kAk2 Σk|k
2
Γ̃k|k
2
?>
?
x̂k|k−1
E x̂k|k−1
≤ κ2 ω2 σ2 kAk2 Σ?k|k−1 υ2k ≤ αυ2k ,
?
?
and γ̃k|l = γk|l −
= Γ?k|l x̂k|l
Proof. Let γk|l = Γk|l x̂k|l , γk|l
?
γk|l . We can easily obtain
and
x̃k+1|k = A x̃k|k .
E bk b>k
Also, from [1, p. 140], we obtain
Then
γ̃k|k = γ̃k|k−1 .
Then it is easy to check that
2
2
? ?>
x̂k|k
E x̂k|k
≤
kAk2 Γ?k|k
≤
κ2 ω2 σ2 kAk2 Σ?k|k υ2k ≤ βυ2k .
Σ̃k|k
E ξk ξk> ≤ ζυ2k .
Step 2) From (33) and Lemma 12, we have
?
?
= Σk|k γk|k − Σ?k|k γk|k
x̃k|k = x̂k|k − x̂k|k
∆k ≤ Hk ∆k−1 Hk > + E ξk ξk>
r
?
= Σk|k γ̃k|k + Σ̃k|k γk|k
,
+2
and
kHk ∆k−1 Hk k E ξk> ξk I ≤ Fk (∆k−1 ),
with
?
γ̃k|k−1 = γk|k−1 − γk|k−1
?
?
= Γk|k−1 x̂k|k−1 −Γ?k|k−1 x̂k|k−1
= Γk|k−1 x̃k|k−1 +Γ̃k|k−1 x̂k|k−1
.
Fk (X) =
We then have
Hk XHk>
!
q
2
+ ζυ + 2 ζ kHk k kXk Iυk .
k
Clearly, if A > B then Fk (A) > Fk (B). Also, there
¯ < ∆,
¯ for all k ≥ k̄.
clearly exists k̄ and ∆¯ such that Fk (∆)
Hence, limk→∞ ∆(k) < ∞, and the result follows.
?
x̃k+1|k = AΣk|k γ̃k|k−1 + AΣ̃k|k γk|k
= AΣk|k Γk|k−1 x̃k|k−1 + ξk
= AΣk|k Γk|k − U x̃k|k−1 + ξk = Hk x̃k|k−1 + ξk .
The following result states a family of upper bounds
on the norm of the covariance matrix of x̃k|l .
5
Let dk = kDk k. Then
Theorem 8. Consider ∆k as defined in (34). Let Hk =
Vk Jk Vk −1 and H̄ = V̄ J¯V̄ −1 be the Jordan decompositions of Hk and H̄, respectively. Then for every > 1,
there exists k ∈ N such that
dk ≤ λ
A =
λψ φ
,
ψ − υ
B =
υl Πl,k Π>l,k ≤ λφ
hk = φ ψk
d(z) = h(z)u(z) =
=
φ
=
m
=
H̄ = lim Hk ,
k→∞
2(k −1)
m
2
−1 2
,
V̄
V̄
ρ H̄
n
o
max 1, kH1 k , · · · , Hk −1 .
and
uk = λυk .
Taking z-transform we get
and
ψ
υl ψk−l = (hk ∗ uk ) ,
with
λυφ
.
υ − ψ
ρ H̄ ,
k
X
l=1
l=1
k∆k k ≤ A ψk + B υk ,
where
k
X
(38)
λφ
1 − ψ z−1 1 − υz−1
A
B
=
+
.
1 − ψ z−1 1 − υz−1
Hence,
dk = A ψk + B υk ,
and the result follows from the definition of dk and (39).
Proof. We split the argument in steps:
Step 1) Let
Theorem 8 states that the covariance of the difference
?
between x̂k|k−1
and x̂k|k−1 is bounded by two exponential terms. The term B υk is due to the convergence
of the Kalman gain Kk? to Kk , while the term A ψk is
due to the convergence of the states given by the system dynamics. In order to use this result to show the
?
asymptotic convergence of x̂k|k−1
to x̂k|k−1 , we need that
υ < 1 and ψ < 1, for some > 0. While it is clear
from (24) that the former is true, guaranteeing the latter
is not that straightforward. The following proposition
addresses this issue.
Dk = Hk Dk−1 Hk > + λIυk .
with D1 = 0, From (35), and since ∆1 = D1 = 0, it
follows that
∆ k ≤ Dk .
(39)
Step 2) Let
Πl,k
=
Hk−1 Hk−2 · · · Hl
=
Vk−1 Jk−1 Vk−1 −1 · · · Vl Jl Vl −1 .
From (38), there exists k ∈ N such that, for all k ≥ k ,
√
V̄ ,
kVk k ≤
√
−1
Vk
V̄ −1 ,
≤
√
,
Vk+1 −1 Vk ≤
√
ρ H̄ .
kJk k ≤
Proposition 9. If theh pair [A,
i C] is completely detectable and the pair A, Q1/2 is completely stabiliz
able, then ρ H̄ < 1, where ρ(H̄) denotes the spectral
radius of matrix H̄.
Proof. Let Kk? = diag Kk(i) : i = 1, · · · , N . From Theorem 5,
lim Kk = lim Kk? , K̄.
Then, for all k ≥ l,
Πl,k
≤ mk −l kVk−1 k kJk−1 k Vk −1 Vk−1 ×
k→∞
· · · × Vk +1 −1 Vk Jk Vk
k−k
√
≤
V̄ V̄ −1 mk −l ρ H̄
m k −l k−l
√
−1
=
V̄ V̄
ρ H̄
ρ H̄
k−l
p
≤
φ ψ 2 .
Dk+1 = λ
x̂k+1|k
=
?
x̂k+1|k
=
A (I − Kk C) x̂k|k−1 + (AKk + L) yk ,
?
A I − Kk?C x̂k|k−1
+ AKk? + L yk .
Hence, if we had that Kk = Kk? = K̄, for all k ∈ N, then
x̃k+1|k = A I − K̄C x̃k|k−1 .
Step 3) We have
k
X
k→∞
Now,
However, under the same assumption, according to
Lemma 6, x̃k+1|k = H̄ x̃k|k−1 . Hence,
H̄ = A I − K̄C .
υl Πl,k Π>l,k .
l=1
6
0.07
S
5
S
4
S
0.06
1
0.05
S
3
S
x
~kjk!1
0.04
2
0.03
0.02
0.01
Figure 1: Communication graph of the five subsystems S i , i =
1, 2, . . . , 5.
0
-0.01
-0.02
0
1
2
3
4
5
6
7
k
i.e., H̄ equals the matrix that determines the asymptotic
dynamics of the centralized Kalman filter’s estimation
error. Then, in view of the model (4)-(5), the result follows from [1, S 4.4].
Figure 2: Difference between the estimated states obtained via both
filtering schemes.
6
5.3. The case when the initial covariance is block diagonal
#10 -17
4
It turns out that, when the initial covariance matrix
has a block diagonal structure both estimation methods
are completely identical. This is summarized in the following corollary.
x
~kjk!1
2
0
-2
10.inConsider
TexPointCorollary
fonts used
EMF. the network of subsystems (1)(2).
If
the
matrix
is block
thenbox.:
the disRead the TexPoint manualPbefore
youdiagonal,
delete this
AA
-4
-6
tributed Kalman filter scheme (15)-(19) produces, for
each i, the same estimate as the centralized Kalman filter (9)-(13).
0
1
2
3
4
5
6
7
k
Figure 3: Difference between the estimated states obtained via both
filtering schemes, when P is block-diagonal.
Proof. Recall that matrices A, Q, C and R are all block
diagonal. It then follows from (10) that, if Σk|k is block
diagonal, so is Σk+1|k . One can easily check from (12)
and (13) that the same holds for Kk and Σk|k if Σk|k−1
is block diagonal. Since Σ1|0 = P is block diagonal, it
follows that the matrices Σk|k−1 and Σk|k remain block
diagonal for all k. Now, it is straightforward to verify
that (9)-(13) become equivalent to (15)-(19), when Σk|k
and Σk|k−1 are block diagonal. Hence, the distributed
and centralized Kalman filters produce the same estimate and the result follows.
centralized one. To this end, we set the initial conditions
?
= 0.
for both schemes to be the same, i.e., x̂0|0 = x̂0|0
6. Simulations
In the first simulation, the initial covariance matrix is
chosen by randomly generating a positive-definite matrix using P = LL> + 0 I5 , where 0 = 0.1 and the entries L ∈ R5×5 are drawn from the uniform distribution
U(0, 1). Fig. 2 shows the time evolution of the entries
?
of the difference x̃k|k−1 = x̂k|k−1
− x̂k|k−1 between the es?
timation outcome x̂k|k−1 of the distributed Kalman filter
and that x̂k|k−1 of the centralized one. We see that how
this difference asymptotically converges to zero.
In this section, we present numerical results
to compare the performance of the proposed distributed scheme (15)-(19) and the optimal centralized
scheme (9)-(13). We assume that each subsystem is a
single integrator with a pole at 0.2. The communication
graph is undirected, as in Fig. 1, and the nonzero values
of L(i, j) are set to 0.3. Furthermore, vk ∼ N (0, 0.1I5 )
and wk ∼ N (0, 0.1I5 ).
We now compare the state estimation produced by the
proposed distributed scheme with that produced by the
In the second simulation we evaluate the same difference in the case where the initial covariance matrix is
block-diagonal. To this end, we choose P = 1 I5 , where
1 is a random scaler drawn from the uniform distribution U(0, 1). The time evolution entries of the difference x̃k|k−1 are shown in Fig. 3. We see that these differences are very negligible, only due to numerical errors. This confirms our claim that the estimates of both
Kalman filter schemes are the same when the matrix P
is block-diagonal.
7
7. Conclusion
[10] Z. Lin, L. Wang, Z. Han, and M. Fu. A graph laplacian approach
to coordinate-free formation stabilization for directed networks.
IEEE Transactions Automatic Control, 61(5):1269–1280, 2016.
[11] C. G. Lopes and A. H. Ali. Diffusion least-mean squares
over adaptive networks: Formulation and performance analysis.
IEEE T. Signal Proces, 56(7):3122–3136, 2008.
[12] D. Marelli and M. Fu. Distributed weighted linear least squares
estimation with fast convergence in large-scale systems. Automatica, 51:27–39, 2015.
[13] R. Olfati-Saber. Distributed kalman filter with embedded consensus filters. In IEEE Conference on Decision and Control,
pages 8179–8184, 2005.
[14] R. Olfati-Saber. Kalman-consensus filter : Optimality, stability,
and performance. In IEEE Conference on Decision and Control,
pages 7036–7042, 2009.
[15] A. Ribeiro, I. Schizas, S. Roumeliotis, and G. Giannakis.
Kalman filtering in wireless sensor networks. IEEE Control Systems, 30(2):66–86, 2010.
[16] M. Subbotin and R. Smith. Design of distributed decentralized
estimators for formations with fixed and stochastic communication topologies. Automatica, 45(11):2491 – 2501, 2009.
[17] T. Sui, D. Marelli, and M. Fu.
Accuracy analysis for distributed weighted least-squares estimation in finite steps and loopy networks.
submitted to Automatica.
URL http://www.cifasis-conicet.gov.ar/
marelli/DWLS_accuracy.pdf.
[18] A. Teixeira, I. Shames, H. Sandberg, and K. H. Johansson. A
secure control framework for resource-limited adversaries. Automatica, 51:135–148, 2015.
[19] V. Ugrinovskii. Distributed robust filtering with consensus of
estimates. Automatica, 47(1):1 – 13, 2011.
[20] V. Ugrinovskii. Distributed robust estimation over randomly
switching networks using consensus. Automatica, 49(1):160 –
168, 2013.
[21] M. Zamani and V. Ugrinovskii. Minimum-energy distributed
filtering. In IEEE Conference on Decision and Control, pages
3370–3375, 2014.
[22] M. Zamani, U. Helmke, and B. D. O. Anderson. Zeros of networked systems with time-invariant interconnections,. Automatica, pages 97–105, 2015.
[23] W. Zhang, M. S. Branicky, and S. M. Phillips. Stability of networked control systems. IEEE Control Systems, 21(1):84–99,
2001.
[24] R. Zheng, Z. Lin, M. Fu, and D. Sun. Distributed control for uniform circumnavigation of ring-coupled unicycles. Automatica,
53:23–29, 2015.
We studied the distributed Kalman filter problem in
a network of linear time-invariant subsystems. We proposed a distributed Kalman filter scheme, which uses
only local measurements, and we studied the extent to
which this scheme approximates the centralized (i.e.,
optimal) Kalman filter. It turned out that the covariance
matrix associated with the initial value of the state vector plays an important role. We showed that if this matrix is block diagonal, the proposed distributed scheme
is optimal. Moreover, if that condition is dropped, the
estimation error covariances, and the associated estimates, obtained through these two approaches approximate each other exponentially fast. We also established
proper bounds on error between estimates and its covariance matrix.
Appendix A. Some lemmas
Lemma 11. [17, Lemma 25] For every x ∈ R and 0 ≤
y ≤ 1, e xy − 1 ≤ (e x − 1) y.
"
#
A B>
Lemma 12. [17, Lemma 26] If
≥ 0, then
B C
√
kBk ≤ kAk kCk.
References
[1] B .D. O. Anderson and J. B. Moore. Optimal Filtering. Englewood Cliffs, NJ: Prentice-Hall, 1979.
[2] R. R. Bitmead, M. R. Gevers, I. R. Petersen, and R. J. Kaye.
Monotonicity and stabilizability-properties of solutions of the
riccati difference equation: Propositions, lemmas, theorems, fallacious conjectures and counterexamples. Systems & Control
Letters, 5(5):309–315, 1985.
[3] P. Bougerol. Kalman filtering with random coefficients and contractions. SIAM Journal on Control and Optimization, 31(4):
942–959, 1993.
[4] A. J. Conejo, S. De la Torre, and M. Canas. An optimization
approach to multiarea state estimation. IEEE Transactions on
Power Systems, 22(1):213–221, 2007.
[5] A. Gómez-Expósito, A. De la Villa Jaén, C. Gómez-Quiles, Patricia P. Rousseaux, and T. Van Cutsem. A taxonomy of multiarea state estimation methods. Electric Power Systems Research,
81(4):1060–1069, 2011.
[6] S. Kar, J. M. F Moura, and K. Ramanan. Distributed parameter
estimation in sensor networks: Nonlinear observation models
and imperfect communication. IEEE Transactions on Information Theory, 58(6):3575–3605, 2012.
[7] U. A. Khan and J. M. F. Moura. Distributing the kalman filter for
large-scale systems. IEEE Transactions on Signal Processing,
56(10):4919–4935, 2008.
[8] Z. Lin, L. Wang, Z. Han, and M.Fu. Distributed formation
control of multi-agent systems using complex Laplacian. IEEE
Transactions Automatic Control, 59(7):1765–1777, 2014.
[9] Z. Lin, L. Wang, Z. Chen, M. Fu, and Z. Han. Necessary
and sufficient graphical conditions for affine formation control. IEEE Transactions Automatic Control, 61(10):2877–2891,
2016.
8
| 3cs.SY
|
Coordinated trajectory tracking of multiple vertical take-off
and landing UAVs ⋆
arXiv:1711.03261v1 [cs.SY] 9 Nov 2017
Yao Zou a, Ziyang Meng a,
a
Department of Precision Instrument, Tsinghua University, Beijing 100084, P. R. China.
Abstract
This paper investigates the coordinated trajectory tracking problem of multiple vertical takeooff and landing (VTOL) unmanned
aerial vehicles (UAVs). The case of unidirectional information flow is considered and the objective is to drive all the follower
VTOL UAVs to accurately track the trajectory of the leader. Firstly, a novel distributed estimator is developed for each
VTOL UAV to obtain the leader’s desired information asymptotically. With the outputs of the estimators, the solution to the
coordinated trajectory tracking problem of multiple VTOL UAVs is transformed to individually solving the tracking problem
of each VTOL UAV. Due to the under-actuated nature of the VTOL UAV, a hierarchical framework is introduced for each
VTOL UAV such that a command force and an applied torque are exploited in sequence, then the position tracking to the
estimated desired position and the attitude tracking to the command attitude are achieved. Moreover, an auxiliary system
with proper parameters is implemented to guarantee the singularity-free command attitude extraction and to obviate the
use of the unavailable desired information. The stability analysis and simulations effectively validate the achievement of the
coordinated trajectory tracking of multiple VTOL UAVs with the proposed control approach.
Key words: Unmanned aerial vehicle (UAV); Coordinated trajectory tracking; Directed graph; Distributed estimators
1
Introduction
The past few decades have witnessed a rapid development in the formation control of unmanned aerial vehicles (UAVs). Replacing a single monolithic UAV with a
formation of multiple micro ones can effectively improve
efficiency without costly expense [1]. More recently, the
vertical takeoff and landing (VTOL) UAV, as a representative UAV, has received increasing interest, due
to its capacities of hovering and low-speed/low-altitude
flight [2]. Additionally, the VTOL UAV is a canonical nonlinear system with under-actuation property [3],
which raises a lot of technical problems for control theory research. Therefore, the formation control of multiple VTOL UAVs deserves intensive studies.
⋆ This paper was not presented at any IFAC meeting. This
work has been supported in part by National Natural Science
Foundation of China under Grant 61503249 and Beijing Municipal Natural Science Foundation under Grant 4173075,
and the National Key Research and Development Program
of China under Grant 2016YFB0500900/2.
Corresponding author: Ziyang Meng.
Email addresses: [email protected]
(Yao Zou), [email protected]
(Ziyang Meng).
Preprint submitted to Automatica
Generally, the study of formation control problem is categorized into leaderless and leader-follower formation
control problem. The leaderless formation requires its
members to simply reach a prescribed pattern [4]. For
example, a distributed control algorithm is proposed in
[5] such that the formation of VTOL UAVs with an identical velocity was achieved. The special case with communication delay was also studied [6] for the leaderless
formation objective and corresponding control solutions
were proposed. Another formation protocol was developed in [7] to realize a time-varying formation of VTOL
UAVs without a leader, and the obtained theoretical results were verified with practical experiments.
Compared with the leaderless formation, the objective
of the leader-follower formation is that followers reach
an agreement with the desired information associated
with a leader while forming the prescribed pattern [8].
This may lead the formation to complete some intricate
missions, where the leader is responsible for performing the desired trajectory of the formation and it is
delivered via the communication network between the
leader and followers. Although [9,10,11] proposed control approaches to achieve the leader-follower formation
of VTOL UAVs, the leader’s desired information was
required to be available to all the followers. In prac-
10 November 2017
control approach project and the stability analysis. Section 4 performs some simulations to verify the theoretical results. And section 5 draws final conclusions.
tice, due to limited information exchange and communication constraints, the leader’s desired information
is only accessible to a portion of the followers. To
achieve the leader-follower formation under restricted
communication networks, distributed algorithms were
implemented with a local information exchange mechanism [12,13,14,15,16,17]. Using backstepping and filtering strategies, a distributed control algorithm was
developed in [18] to realize the asymptotically stable
leader-follower formation of VTOL UAVs. A distributed
formation and reconfiguration control approach is designed in [19] to accomplish the leader-follower formation without inter-vehicle collisions. With feedback
linearization technique, [20] proposed a leader-follower
formation protocol for VTOL UAVs, which ensured
their heading synchronization as well. [21] presented
a distributed control strategy over a switched graph
and derived necessary and sufficient conditions on the
time-varying leader-follower formation of VTOL UAVs.
However, the network graphs among the followers used
in [18,19,20,21] are undirected, which means that each
pair of the followers interacts bidirectionally. This undirected graph condition is quite restrictive, which, due
to communication constraints, can hardly be met in
practical applications.
Notations. Rm×n denotes the m × n Euclidean space.
Given a vector x = [x1 , x2 , · · · , xn ]T , define sgn(x)
=
Pn
T
[sgn(x1 ), sgn(x
),
·
·
·
,
sgn(x
)]
,
and
kxk
=
|x
2
n
1
i|
i=1
√
and kxk = xT x are its 1-norm and 2-norm. Given
a square matrix A = [aij ] ∈ Rn×n , define λmin (A) and
λmax (A)qas its minimum and maximum eigenvalues, and
Pn Pn
2
kAk =
i=1
j=1 aij is its F-norm. In is an n × n
identity matrix and 1n is an n-dimensional vector with
all entries being one. Furthermore, given a vector x =
[x1 , x2 , x3 ]T , superscript × represents a transformation
from x to a skew-symmetric matrix:
0
−x3 x2
x× =
x3 0 −x1 .
−x2 x1 0
2
This paper proposes a coordinated trajectory tracking
control approach for multiple VTOL UAVs with local
information exchange, where the desired trajectory information is described by a leader. The network graph
among the follower VTOL UAVs is assumed to be directed. This effectively relaxes the restrictive assumption that the graph is symmetric. By applying a novel
distributed estimator, the leader’s desired information
is accurately estimated for each follower VTOL UAV.
Based on the hierarchial framework, a command force
and an applied torque are synthesized for each VTOL
UAV such that the coordinated trajectory tracking is
achieved for a group of VTOL UAVs. Compared with
the aforementioned work, the main contributions of this
paper are three-fold. First, in contrast to the work in
[5,6,7], where only a prescribed pattern is formed with
a constant velocity, the leader-follower tracking of multiple VTOL UAVs is achieved by introducing a novel
distributed estimator. Second, the coordinated tracking
is achieved with weak connectivity, where the network
graph among the followers is directed, rather than the
limited undirected one used in [18,19,20,21]. Third, instead of solely discussing the position loop [7,21], a complete VTOL UAV system is studied based on a hierarchical framework, where an auxiliary system is proposed
to ensure the non-singular command attitude extraction
and to avoid the use of the unavailable desired information.
2.1
Background
Problem statement
Suppose that there are n follower VTOL UAVs in a team,
which are labeled by V = {1, 2, . . . , n}. Each UAV is
a six-dof (degree of freedom) rigid body and operates
in two reference frames: inertia frame I = {OI xI yI zI }
which is attached to the earth and body frame B =
{OB xB yB zB } which is fixed to the fuselage. To establish
the model of the UAVs, rotation matrix Ri ∈ SO(3) ,
{R ∈ R3×3 | det R = 1, RT R = RRT = I3 } and unit
quaternion Qi = [σi , qiT ]T ∈ Q , {Q ∈ R × R3 | σ 2 +
q T q = 1} are applied to represent the attitude of each
UAV. In terms of Euler formula [22], an explicit relation
between these two attitude representations is derived as
Ri (Qi ) = (σi2 − qiT qi )I3 + 2qi qiT − 2σi qi× .
(1)
Based on Euler-Newton formulae, the kinematics and
dynamics of the i-th VTOL UAV are given by
ṗi = vi ,
Ti
v̇i = −gê3 +
Ri (Qi )ê3 ,
mi
1
Q̇i = Gi (Qi )ωi ,
2
Ji ω̇i = −ωi× Ji ωi + τi ,
(2)
(3)
(4)
(5)
where pi = [pxi , pyi , pzi ]T and vi = [vix , viy , viz ]T denote
the position and velocity of the center of gravity of the
UAV in frame I, respectively, mi is the total mass, g
is the local gravitational acceleration, ê3 , [0, 0, 1]T ,
The remaining sections are arranged as follows. Section
2 describes the problem to be solved and provides some
useful preliminaries. Section 3 states the main results
in detail, including the distributed estimator design, the
2
Ti denotes the applied thrust along ê3 , Qi = [σi , qiT ]T
and Ri (Qi ) are the unit quaternion and rotation matrix, Gi (Qi ) = [−qi , σi I3 − qi× ]T , ωi = [ωix , ωiy , ωiz ]T
denotes the angular velocity of the UAV in frame B,
Ji = diag{Jix , Jiy , Jiz } is the inertial matrix with respect
to frame B, and τi denotes the applied torque in frame B.
di0 > 0 if node i is accessible
to #the leader and di0 = 0
"
0 01×n
otherwise; and L̄ ,
, where M = [mij ] ,
−d0 M
L + diag{d10 , d20 , · · · , dn0 }.
Assumption 2.2 The leader-follower graph Gn+1 has a
directed spanning tree with the leader being the root.
In addition to n followers, there is a leader, labeled by
0, to represent the global desired information including
the desired position pr and its derivatives. The control
objective is to design applied thrust Ti and torque τi
for each follower VTOL UAV described by (2)-(5) such
that all the followers track the leader while maintaining
a prescribed formation pattern. More specifically, given
a desired position offset δi , the objective is to guarantee
that
Some important properties associated with matrix M
are given in Lemma 2.1 [23].
Lemma 2.1 Under Assumption 2.2, M is a nonsingular M-matrix with the properies that all its eigenvalues have positive real parts, and there exists a positive definite diagonal matrix Θ = diag{θ1 , θ2 , · · · , θn } such that
Ξ = MT Θ+ΘM is strictly diagonally dominant and positive definite, where [1/θ1 , 1/θ2 , · · · , 1/θn ]T = M−1 1n .
lim (pi (t)−pr (t)) = δi , lim (vi (t)− ṗr (t)) = 0, ∀i ∈ V.
t→∞
t→∞
(6)
Due to communication constraints, the leader’s desired
information is only available to a subset of the followers and the followers only have access to their neighbors’ information. To solve such a coordinated tracking
problem via local information exchange, distributed algorithms are implemented. Moreover, it follows from (6)
that limt→∞ (pi (t) − pj (t)) = δij , where δij = δi − δj ,
∀i, j ∈ V. This means that the followers form a pattern determined by δi while tracking the leader. Therefore, a proper position offset δi is essential such that the
proposed algorithm ensures followers’ convergence to a
well-defined and unique formation.
2.3
Consider the vector differential equation
ẋ = f (x, t),
(7)
where f : Rn × R → Rn is measurable and essentially
locally bounded.
In what follows, the definitions of Filippov solution, generalized gradient and regular function are given according to [24,25,26].
Assumption 2.1 The desired position pr and its
(3)
derivatives ṗr , p̈r and pr are bounded.
2.2
Filippov solution and non-smooth analysis
Definition 2.1 (Filippov solution) A vector function x(t) is called a solution of (7) on [t0 , t1 ], if x(t)
is absolutely continuous on [t0 , t1 ] and for almost all
t ∈ [t0 , t1 ], ẋ ∈ K[f ](x, t), where
Graph theory
K[f ](x, t) =
Communication topology among UAVs is described by a
graph Gn , (V, E), which is composed of a node set V ,
{1, 2, · · · , n} and an edge set E ⊆ V × V. For a directed
graph, (i, j) ∈ E means that the information of node i is
accessible to node j, but not conversely. All neighbours
of node i are included in set Ni = {j ∈ V | (j, i) ∈ E}.
A path from node i to node j is a sequence of edges.
\ \
ρ>0 µN =0
cof (B(x, ρ) − N, t),
T
µN =0 denotes the intersection over all sets N of
Lebesgue measure zero, co(·) denotes the vector convex
closure, and B(x, ρ) denotes the open ball of radius ρ
centered at x.
For a follower graph Gn , its adjacent matrix D = [dij ] ∈
Rn×n is defined such that dij > 0 if (j, i) ∈ E and dij =
0 otherwise, and the associated nonsymmetric Laplan×n
cian
is defined such that lii =
Pn matrix L = [lij ] ∈ R
j=1,j6=i dij and lij = −dij for j 6= i. For a leaderfollower graph Gn+1 , {V̄, Ē} (leader is labeled as 0)
with V̄ = {0, 1, · · · , n} and Ē ⊆ V̄ × V̄, we define D̄ ∈
R(n+1)×(n+1) and L̄ ∈ R(n+1)×(n+1) as its adjacent matrix and
Laplacian matrix. Specifically,
" nonsymmetric
#
0 01×n
D̄ ,
, where d0 = [d10 , d20 , · · · , dn0 ]T and
d0 D
Definition 2.2 (Generalized gradient) For a locally
Lipschitz function V : Rn × R → R, its generalized gradient at (x, t) is defined as
∂V (x, t) = co{lim ▽V (x, t) | (xi , ti ) → (x, t), (xi , ti ) ∈
/ ΩV },
where ΩV is the set of measure zero where the gradient of V is not defined. Furthermore, the generalized derivative of V along system (7) is defined as
"
#
T
K[f
](x,
t)
˙
.
Ṽ , φ∈∂V φT
1
3
Definition 2.3 (Regular) f (x, t) : Rn × R is called
regular if
(1) for all ν ≥ 0, the usual one-sided directional derivative f ′ (x; ν) exists;
(2) for all ν ≥ 0, f ′ (x; ν) = f o (x; ν), where the generalized directional derivative f o (x; ν) is defined as
f o (x; ν) = lim sup
y→x t↓0
command attitude extraction. For i ∈ V, a distributed
estimator is proposed as follows:
p̂˙i =v̂i − kp
v̂˙ i =âi − kv
f (y + tν) − f (y)
.
t
Lemma 2.2 Let system (7) be essentially locally
bounded and 0 ∈ K[f ](x, t) in a region Rn × [0, ∞).
Suppose that f (0, t) is uniformly bounded for all t ≥ 0.
Let V : Rn × [0, ∞) → R be locally Lipschitz in t
and regular such that W1 (x) ≤ V (t, x) ≤ W2 (x) and
Ṽ˙ (x, t) ≤ −W (x), where W (x) and W (x) are continu-
(8a)
dij (v̂i − v̂j ),
(8b)
j=0
j=0
(8c)
where p̂0 = pr , v̂0 = ṗr and â0 = p̈r are specified,
dij is the (i, j)-th entry of the adjacent matrix D associated with the follower graph Gn , kp , kv , ka and la
are positive parameters, and Γi = diag{µxi , µyi , µzi } with
µki = tanh(γ̇ik )γ̇ik and Γ̄i = diag{µ̄xi , µ̄yi , µ̄zi } with µ̄ki =
1 − tanh2 (γik ) for k = x, y, z. Next, define the estimation
errors p̄i = p̂i −pr , v̄i = v̂i − ṗr and āi = âi − p̈r for i ∈ V.
It then follows from (8) that their dynamics satisfy
2
that ϕ ≤ −W , ∀ϕ ∈ Ṽ˙ . Then, all Filippov solutions of
system (7) are bounded and satisfy limt→∞ W (x(t)) = 0.
Main results
p̄˙ i =v̄i − kp
Due to the under-actuated nature of the VTOL UAV, a
hierarchical strategy is applied to solve the coordinated
trajectory tracking problem of multiple VTOL UAV systems. First, a distributed estimator using local information interaction is designed for each follower UAV to estimate the leader’s desired information. Then, the coordinated trajectory tracking problem of multiple VTOL
UAVs is transformed into the asymptotic stability problem of each individual error system. Next, a command
force and an applied torque are exploited for each UAV
to asymptotically stabilize the position and attitude error systems, respectively. Finally, the stability of each
error system is analyzed.
3.1
dij (p̂i − p̂j ),
j=0
ous positive definite functions, W (x) is a continuous positive semi-definite function, and Ṽ˙ (x, t) ≤ −W (x) means
3
j=0
n
X
γ̈i = − la γ̇i + 2Γi γ̇i
n
n
X
ka −1 X
− Γ̄i
dij (â˙ i − â˙ j )
dij (âi −âj )+
kγ
j=0
j=0
n
n
X
X
dij (â˙ i − â˙ j ) ,
+sgn
dij (âi − âj ) +
The Lyapunov stability criterion for non-smooth systems is given in Lemma 2.2 [27].
1
n
X
v̄˙ i =āi − kv
n
X
j=1
n
X
mij p̄j ,
(9a)
mij v̄j ,
(9b)
j=1
¨i =kγ Γ̄i γ̈i − 2kγ Γ̄i Γi γ̇i − p(4)
ā
r
!
n
n
X
X
= − la kγ Γ̄i γ̇i − ka
mij āj +
mij ā˙ j
j=1
− ka sgn
n
X
mij āj +
mij ā˙ j
j=1
j=1
= − la ā˙ i − ka
j=1
n
X
n
X
mij āj +
− ka sgn
Distributed estimator design
j=1
Since the leader’s desired information including the desired position pr and its derivatives is not available to all
the followers, a distributed estimator is firstly designed
for each VTOL UAV to estimate them.
mij āj +
− p(4)
r
mij ā˙ j
j=1
j=1
n
X
n
X
!
n
X
j=1
mij ā˙ j
!
!
(9c)
+ Np ,
where mij denotes the (i, j)-th entry of M defined in
(4)
(3)
Section 2, and Np = la pr − pr is bounded according
to Assumption 2.1. Equivalently, the error dynamics (9)
can be rewritten as
For i ∈ V, we define p̂i , v̂i and âi = kγ tanh(γi ) as the
estimations of pr , ṗr and p̈r , respectively, where γi is an
auxiliary variable and parameter kγ ≥ supt≥0 p̈r (t). As
will be shown subsequently, the definition of âi using the
hyperbolic tangent function enables the control parameters to be chosen explicitly in case of singularity in the
p̄˙ =v̄ − kp (M ⊗ I3 )p̄,
v̄˙ =ā − kv (M ⊗ I3 )v̄,
˙
¨ = − la ā˙ − ka (M ⊗ I3 )(ā + ā)
ā
˙ + 1n ⊗ Np ,
− ka sgn ((M ⊗ I3 )(ā + ā))
4
(10a)
(10b)
(10c)
According to Definition 2.2, the generalized derivative
of Ls along (11) satisfies
where p̄, v̄ and ā are the column stack vectors of p̄i , v̄i
and āi , respectively, and operator ⊗ denotes the kronecker
define a sliding surface si =
Pn product. Moreover,
Pn
la j=1 mij āj + j=1 mij ā˙ j for i ∈ V, and correspond˙ It
ingly, its column stack vector s = (M ⊗ I3 )(la ā + ā).
follows from (10c) that the dynamics of s satisfies
ṡ = (M ⊗ I3 ) (−ka s − ka sgn(s) + 1n ⊗ Np ) .
\
L̃˙ s = (φ + s)T (Θ ⊗ I3 )K[(M ⊗ I3 )(−ka s − ka sgn(s)
φ∈∂Ls
+ 1n ⊗ Np )],
\
= (φ + s)T (ΘM⊗ I3 )(−ka s− ka ∂ksk1 + 1n ⊗ Np ),
(11)
φ∈∂Ls
(15)
Theorem 3.1 indicates that the developed distributed
estimator (8) with appropriate parameters enables the
estimation errors p̄i , v̄i and āi for each VTOL UAV to
converge to zero asymptotically.
sji ∈ R+
{1}
j
where ∂|si | = {−1} sji ∈ R− , ∀i ∈ V, j = x, y, z,
[−1, 1] sj = 0
i
and the calculation of K is applied using the same argument given in [24].
Theorem 3.1 Under Assumptions 2.1 and 2.2, if the
estimator parameters kp , kv , la and ka are chosen based
on
kΘk2
,
λmin (Ξ)2
kp λmin (Ξ)kΘk2
la >
,
λmin (MT ΘM)(kp kv λmin (Ξ)2 − kΘk2)
√
2 nkΘMkN̄p
ka >
,
λmin (Ξ)
kp kv >
If L̃˙ s 6= ∅, suppose ϕ ∈
/ L̃˙ s , then we know that ∀φ ∈
T
∂ksk1, ϕ = (φ + s) (ΘM ⊗ I3 )(−ka s − ka δ + 1n ⊗ Np )
for some δ ∈ ∂ksk1 . Choose φ = arg minδ∈∂ksk1 (δ +
T
Θ
s)T ( ΘM+M
⊗ I3 )(δ + s). According to [24], for all
2
δ ∈ ∂ksk1 , we have that
(12)
(13)
(14)
(φ + s)T (ΘM ⊗ I3 )(δ + s)
ΘM + MT Θ
T
⊗ I3 (φ + s)
≥(φ + s)
2
1
= (φ + s)T (Ξ ⊗ I3 )(φ + s).
2
where N̄p = supt≥0 kNp (t)k, and Θ and Ξ are given in
Lemma 2.1, the distributed estimator (8) ensures that
limt→∞ p̄i (t) = 0, limt→∞ v̄i (t) = 0 and limt→∞ āi (t) = 0,
∀i ∈ V.
Proof: The proof is divided into two parts: first, the
sliding surface si , i ∈ V is proven in Proposition 3.1 to
converge to zero asymptotically; then, the final result is
shown in Proposition 3.2.
It then follows that L̃˙ s further satisfies
ka
L̃˙ s ≤ − (φ + s)T (Ξ ⊗I3 )(φ + s) +(φ + s)T (1n ⊗Np )
2
√
ka λmin (Ξ)
≤−
kφ + sk2 + nkφ + skkΘMkN̄p
2
√
ka λmin (Ξ)
kφ + sk − nkΘMkN̄p kφ + sk.
=−
2
(16)
Proposition 3.1 Under Assumptions 2.1 and 2.2, if the
estimator parameter ka satisfies (14), the distributed estimator (8) guarantees that limt→∞ si (t) = 0, ∀i ∈ V.
Proof: Obviously, system (11) is non-smooth; thereafter,
the solution of (11) is studied in the sense of Filippov
and the non-smooth framework is applied. The stability
of system (11) is to be proven based on Lemma 2.2.
Note that if φ = 0, then s = 0 and kφ+sk = 0, and if φ 6=
0, then kφ + sk ≥ 1. Hence, if the estimator parameter
ka√ satisfies (14), there exists a constant k̄a satisfying
2 nkΘMkN̄p
(Ξ)
≤ k̄a < ka such that −( k̄a λmin
kφ + sk −
2
√ λmin (Ξ)
nkΘMkN̄p)kφ + sk ≤ 0. Therefore, it follows that
Pn
We first propose a Lyapunov function Ls = i=1 θi ksi k1+
sT (Θ ⊗ I3 )s, where θi is the i-th diagonal entry of Θ.
Note that Ls is non-smooth but regular [24]. It can be
derived that Ls is bounded by
(ka − k̄a )λmin (Ξ)
L̃˙ s ≤ −
kφ + sk2 .
2
s
W1 (s) ≤ L ≤ W2 (s),
(17)
In addition, each entry of s has the same sign as its
counterpart in sgn(s), and thus, it follows that kφ+sk ≥
ksk. We finally have that
√
where W1 (s) = λmin (Θ)ksk2 and W2 (s) = nλmax (Θ)
· (ksk1 + ksk2 ). In terms of Lemma 2.2, the stable result
can be deduced if only the generalized derivative of Ls
satisfies L̃˙ s ≤ W (s), where W (s) is a continuous positive
semi-definite function.
(ka − k̄a )λmin (Ξ)
ksk2 = −W (s).
L̃˙ s ≤ −
2
5
(18)
this case, L̇e further satisfies
Since W (s) ≥ 0 has been ensured, it follows that
Rt
W (s(τ ))dτ is bounded for t ≥ 0, which implies that s
0
is bounded. Hence, it follows from (11) that ṡ is bounded,
which implies that s is uniformly continuous in t. This
means that W (s) is uniformly continuous in t. Based on
Lemma 2.2, we can conclude that limt→∞ W (s(t)) = 0,
which further implies that limt→∞ s(t) = 0.
L̇e ≤ − λmin (Ω)kzk2 + kΘMkkzkksk
√
≤ − ϑ1 Le + ϑ2 ksk Le ,
where ϑ1 =
(19)
1
Le = p̄T (Θ ⊗ I3 )p̄ + v̄ T (Θ⊗I3 )v̄ + āT (MT ΘM ⊗ I3 )ā.
2
λ1 kzk2 ≤ Le ≤ λ2 kzk2,
where λ1 = min{λmin (Θ), 21 λmin (MT ΘM)} and λ2 =
max{λmax (Θ), 12 λmax (MT ΘM)}. The derivative of Le
along (10) and (19) satisfies
L̇e = − p̄T ((MT Θ + ΘM) ⊗ I3 )p̄ + 2p̄T (Θ ⊗ I3 )v̄
3.2
Since the leader’s desired information has been estimated via the distributed estimator (8) for each follower
VTOL UAV, the remaining problem is to transform the
coordinated trajectory tracking problem into the simultaneous tracking problem for the decoupled VTOL UAV
group. This is verified as follows.
+ āT (MT Θ ⊗ I3 )(−la (M ⊗ I3 )ā + s)
= − kp p̄T (Ξ ⊗ I3 )p̄ + 2p̄T (Θ ⊗ I3 )v̄ − kv v̄ T (Ξ ⊗ I3 )v̄
+ 2v̄ T (Θ ⊗ I3 )ā − la āT (MT ΘM ⊗ I3 )ā
+ āT (MT Θ ⊗ I3 )s
≤ − kp λmin (Ξ)kp̄k2 + 2kΘkkp̄kkv̄k − kv λmin (Ξ)kv̄k2
+ 2kΘkkv̄kkāk − la λmin (MT ΘM)kāk2
+ kΘMkkākksk
≤ − z1T Ωz1 + kΘMkkakksk,
Define the position error pei = pi − pr − δi and the velocity error vie = vi − ṗr for i ∈ V. Using the estimations p̂i and v̂i obtained from the distributed estimator
(8), we rewrite pei and vie as pei = pi − p̂i − δi + p̄i and
vie = vi − v̂i + v̄i . Since Theorem 3.1 has shown that
limt→∞ p̄i (t) = 0 and limt→∞ v̄i (t) = 0, ∀i ∈ V, the
coordinated tracking control objective (6) can be transformed into the following simultaneous tracking objective:
(20)
where z1 = [kp̄k, kv̄k, kāk]T and
Ω=
−kΘk
0
Problem transformation
T
− v̄ ((M Θ + ΘM) ⊗ I3 )v̄ + 2v̄ (Θ ⊗ I3 )ā
kp λmin (Ξ)
(22)
Remark 3.1 According to (8), singularity may occur in
the distributed estimator when some diagonal entry of
Γ̄i equals to zero, and this corresponds to the case where
some entry of the auxiliary variable γi tends to infinity. Theorem 3.1 has shown that, with a bounded initial
value, the estimation error āi for each UAV is bounded
all the time. This implies that Γ̄i is always positive definite. Consequently, no singularity is introduced in the
developed distributed estimator (8).
It is bounded by
and kzk = kz1 k has
When V e = 0, it can be shown that D+ V e ≤ ϑ2 ksk.
Thus, it follows that D+ V e satisfies (22) all the time
[28]. For system ẏ = −ϑ1 y + ϑ2 ksk with respect to
y ∈ [0, ∞), it can be proven in terms of input-to-state
stability theory [28] that limt→∞ y(t) = 0 given the fact
that limt→∞ s(t) = 0. According to Comparison Principal [28], it follows that limt→∞ V e (t) = 0, which further implies that limt→∞ p̄(t) = 0, limt→∞ v̄(t) = 0 and
limt→∞ ā(t) = 0.
Define z = [p̄T , v̄ T , āT ]T and assign a Lyapunov function
T
kΘMk
√
,
λ1√
e
V̇ e ≤ −ϑ1 V e + ϑ2 ksk.
Proof: Consider the definition of the sliding surface s,
then the dynamics of the estimator error ā satisfies
T
ϑ2 =
been applied. Next, take V = 2Le . When Le 6= 0, it
follows from (21) that the derivative of V e satisfies
Proposition 3.2 Under Assumptions 2.1 and 2.2, if
the estimator parameters kp , kv and la satisfy (12) and
(13), then limt→∞ si (t) = 0 is sufficient to ensure that
limt→∞ p̄i (t) = 0, limt→∞ v̄i (t) = 0 and limt→∞ āi (t) =
0, ∀i ∈ V.
(M ⊗ I3 )ā˙ = −la (M ⊗ I3 )ā + s.
λmin (Ω)
,
λ2
(21)
−kΘk
kv λmin (Ξ)
−kΘk
0
−kΘk
−la λmin (MT ΘM)
.
lim (pi (t) − p̂i (t)) = δi , lim (vi (t) − v̂i (t)) = 0, ∀i ∈ V.
t→∞
t→∞
We next define p̄ei = pi − p̂i − δi and v̄ie = vi − v̂i for
i ∈ V. It follows from (2), (3) and (8) that their dynamics
If the estimator parameters kp , kv and la are chosen
based on (12) and (13), then Ω is positive definite. In
6
Based on the above discussions, by introducing the distributed estimator (8), the coordinated trajectory tracking problem for multiple VTOL UAV systems (2)-(5) can
be transformed into the simultaneous asymptotic stability problem for each error system (23), (27) and (28).
Lemma 3.2 summarizes this point.
satisfy
p̄˙ ei =v̄ie +
n
X
mij p̄j ,
(23a)
j=1
v̄˙ ie =ui − gê3 − âi +
n
X
mij v̄j +
j=1
Ti
(Ri (Qi )− Ri (Qci ))ê3 ,
mi
Lemma 3.2 Consider the i-th error system (23),
(27) and (28). If a command force ui and an applied
torque τi can be developed such that limt→∞ p̄ei (t) = 0,
limt→∞ v̄ie (t) = 0, limt→∞ qie (t) = 0 and limt→∞ ωie (t) =
0, the coordinated trajectory tracking of multiple VTOL
UAV systems (2)-(5) is achieved in the sense of (6).
(23b)
Ti
Ri (Qci )ê3 is the command force with Qci
where ui = m
i
being the command unit quaternion. Moreover, once
the command force ui can be determined, in view of
kRi (Qci )ê3 k = 1, the applied thrust Ti is derived as
Ti = mi kui k, ∀i ∈ V.
3.3
(24)
In this subsection, a command force ui for each VTOL
UAV will be synthesized. The main difficulties here are
that the command force ui should comply with the nonsingular condition (25) and the desired position pr and
its derivatives are not available in the command force
ui and the subsequent applied torque τi due to limited
communication.
Now that the control strategy is based on a hierarchical
framework, the command unit quaternion Qci , as the attitude tracking objective for each VTOL UAV, should
be extracted from the command force ui . Based on minimal rotation principle, a viable extraction algorithm is
proposed in Lemma 3.1 [5].
To address the above dilemmas, we introduce the virtual
position error p̃ei = p̄ei − ηi and the virtual velocity error
ṽie = v̄ie − η̇i for i ∈ V, where ηi is an auxiliary variable. It
follows from (23) that the dynamics of p̃ei and ṽie satisfy
Lemma 3.1 For i ∈ V, if the command force ui =
[uxi , uyi , uzi ]T satisfies the non-singular condition:
ui ∈
/ U , {u ∈ R3 | u = [0, 0, uz ]T , uz ≤ 0},
the command unit quaternion
tracted as
σic =
s
1 g − uzi
+
, qc =
2
2kui k i
Qci
=
[σic , (qic )T ]T
uyi
1
−ux
i
2kui kσic
0
.
(25)
p̃˙ei = ṽie + ~1 ,
ṽ˙ ie = ui − η̈i − gê3 − âi + ~2 ,
is ex-
1
Gi (Qei )ωie ,
2
(29a)
(29b)
Pn
Pn
where ~1 =
j=1 mij p̄j and ~2 =
j=1 mij v̄j +
Ti
c
mi (Ri (Qi ) − Ri (Qi ))ê3 .
(26)
Lemma 3.3 Consider the i-th virtual position error
system (29). If a command force ui can be synthesized such that limt→∞ p̃ei (t) = 0, limt→∞ ṽie (t) =
0, limt→∞ ηi (t) = 0 and limt→∞ η̇i (t) = 0, then
limt→∞ p̄ei (t) = 0 and limt→∞ v̄ie (t) = 0 are achieved.
Next, define the attitude error Qei = [σie , (qie )T ]T =
(Qci )−1 ⊙ Qi for i ∈ V, where operator ⊙ is the
unit quaternion product. According to [29], Qei =
[±1, 0, 0, 0]T corresponds to the extract attitude tracking. The dynamics of Qei satisfies
Q̇ei =
Command force development
To guarantee the condition in Lemma 3.3, for i ∈ V, we
propose the following command force:
ui = gê3 + âi − kη (tanh(ηi + η̇i ) + tanh(η̇i )),
(30)
(27)
and introduce a dynamic system with respect to the auxiliary variable ηi :
where ωie = ωi − Ri (Qei )T ωic is the angular velocity error
with ωic being the command angular velocity. Please refer
to [29] for the derivations of ωic and its derivative ω̇ic . In
addition, it follows from (5) and Ṙi (Qei ) = Ri (Qei )(ωie )×
that the dynamics of ωie satisfies
η̈i = −kη (tanh(ηi + η̇i ) + tanh(η̇i )) + lp p̃ei + lv ṽie ,
(31)
Ji ω̇ie = − ωi× Ji ωi + τi + Ji [(ωie )× Ri (Qei )T ωic
−
Ri (Qei )T ω̇ic ].
where kη , lp and lv are positive control parameters. Substituting (30) and (31) into (29) yields
p̃˙ei = ṽie + ~1 ,
ṽ˙ e = − lp p̃ei − lv ṽie + ~2 .
(28)
7
(32a)
(32b)
A proper control parameter kη should be chosen such
that the non-singular condition (25) is met. Specifically,
kη <
1
(g − kγ ).
2
functions of the derivatives of the command force ui .
Their expressions are presented as follows:
u̇i =kγ Γ̄i γ̇i − kη [Di η̇i + (Di + Si )η̈i ],
üi = − 2kγ Γ̄i Γi γ̇i + kγ Γ̄i γ̈i − kη [D̄i η̇i + (D̄i + Di + S̄i )η̈i
(33)
(3)
In such a case, the third row of the command force ui
satisfies
+ (Di + Si )ηi ],
where Γi and Γ̄i have been specified below (8),
Di = diag{ǫxi , ǫyi , ǫzi } with ǫji = 1 − tanh2 (ηij + η̇ij ),
Si = diag{νix , νiy , νiz } with νij = 1 − tanh2 (η̇ij ),
D̄i = {ǭxi , ǭyi , ǭzi } with ǭji = −2 tanh(ηij + η̇ij )(1 −
tanh2 (ηij + η̇ij ))(ηij + η̇ij ) and S̄i = {ν̄ix , ν̄iy , ν̄iz } with
ν̄ij = −2 tanh(η̇ij )(1 − tanh2 (η̇ij ))η̇ij , for j = x, y, z, and
uzi =g + âzi − kη (tanh(ηiz + η̇iz ) + tanh(η̇iz ))
≥g − kγ − 2kη > 0,
where âzi = kγ tanh(γiz ) and the property that
| tanh(·)| < 1 have been applied. To this end, kη satisfying (33) is sufficient to guarantee that the developed
command force ui in (30) for each UAV strictly satisfies
the non-singular condition (25).
(3)
ηi
Remark 3.2 By defining âi = kγ tanh(γi ) in the distributed estimator (8) and introducing the auxiliary dynamics (31), the developed command force ui for i ∈ V is
equipped with a saturation property. Based on this property, the choice of the control parameter kη is independent on any estimator state.
From the above derivations, it it trivial to see that the
desired information is not used in the developed applied
torque τi for the UAV without accessibility to the leader.
3.5
Remark 3.3 It follows from (24), (30) and kâi k ≤
√
3kγ that the resulting applied thrust Ti is bounded by
√
√
Ti ≤ mi (g + 2 3kη + 3kγ ), ∀i ∈ V,
(34)
Theorem 3.2 Consider n follower VTOL UAV systems
(2)-(5) with Assumptions 2.1 and 2.2. The synthesized
command force ui in (30) and applied torque τi in (36)
guarantee the coordinated trajectory tracking of multiple
VTOL UAVs in the sense of (6).
Applied torque development
Proof: Theorem 3.1 has shown that, for i ∈ V, the
distributed estimator developed in (8) enables the estimation errors p̄i and v̄i to converge to zero asymptotically. Based on this, it follows from Lemmas 3.2 and
3.3 that the coordinated trajectory tracking objective is
achieved, if the following results are guaranteed by the
synthesized command force ui and applied torque τi :
Th2.i) limt→∞ qie (t) = 0 and limt→∞ ωie (t) = 0, ∀i ∈ V,
Th2.ii) limt→∞ p̃ei (t) = 0 and limt→∞ ṽie (t) = 0, ∀i ∈ V,
Th2.iii) limt→∞ ηi (t) = 0 and limt→∞ η̇i (t) = 0, ∀i ∈ V.
They will be proven in Propositions 3.3-3.5, respectively.
Define a sliding surface ri = lq qie + ωie for i ∈ V, where
lq > 0. From (27) and (28), the dynamics of ri satisfies
lq
Ji ṙi = Ji (σie I3 + (qie )× )ωie − ωi× Ji ωi + τi
2
+ Ji [(ωie )× Ri (Qei )T ωic − Ri (Qei )T ω̇ic ].
(35)
We propose an applied torque τi for each UAV as follows:
lq
Ji (σie I3 + (qie )× )ωie + ωi× Ji ωi
2
− Ji [(ωie )× Ri (Qei )T ωic − Ri (Qei )T ω̇ic ],
τi = − kq ri −
(36)
Proposition 3.3 Consider the attitude error system
(27) and (28). The developed applied torque τi in (36)
guarantees that limt→∞ qie (t) = 0 and limt→∞ ωie (t) = 0,
∀i ∈ V.
where kq > 0. Substituting (36) into (35) yields
Ji ṙi = −kq ri .
Stability analysis
Theorem 3.2 summarizes the final stability result associated with the coordinated trajectory tracking of multiple VTOL UAV systems (2)-(5) controlled by the developed applied thrust and torque.
which means that each Ti is upper bounded by a constant
associated with the individual mass mi and the specified
parameters kη and kγ .
3.4
= − kη [Di η̇i + (Di + Si )η¨i ] + lp p̃˙ ei + lv ṽ˙ ie .
(37)
Proof: It follows from (37) that, for i ∈ V, the developed applied torque τi enables the sliding surface ri
to converge to zero asymptotically. Then, assign a nonnegative function yi = 21 [kqie k2 + (1 − σie )2 ] = 1 − σie ≤ 2
It follows from (36) that, the command angular velocity
ωic and its derivative ω̇ic are necessary to determine each
applied torque τi . According to [29], ωic and ω̇ic are the
8
for i ∈ V. With the definition of ri , the derivative of yi
along (27) satisfies
It is trivial to verify that
1
1
(k tanh(ηi + η̇i )k2 + k tanh(η̇i )k2 + kη̇i k2 ).
2
kη
(39)
The derivative of Lηi along (31) satisfies
Lηi ≥
1
lq
1
ẏi = (qie )T ωie = − kqie k2 + (qie )T ri
2
2
2
lq e 2 1
2
≤ − kqi k + kri k
4
4
lq
1
= (−2yi + yi2 ) + kri k2 .
4
4
L̇ηi = tanh(ηi + η̇i )T η̇i + [tanh(ηi + η̇i ) + tanh(η̇i )
1
+ η̇i ]T [−kη (tanh(ηi + η̇i ) + tanh(η̇i )) + εi ]
kη
= − kη k tanh(ηi + η̇i ) + tanh(η̇i )k2 − η̇iT tanh(η̇i )
1
+ [tanh(ηi + η̇i ) + tanh(η̇i ) + η̇i ]T εi
(40)
kη
≤ − kη k tanh(ηi + η̇i ) + tanh(η̇i )k2 − η̇iT tanh(η̇i )
q
+ 2 Lηi kεi k
q
≤2 Lηi ρ(kεi (0)k, 0).
(41)
l
It can be shown that system ẏ = 4q (−2y + y 2 ) with
respect to y ∈ [0, 2] is asymptotically stable. For system
l
ẏ = 4q (−2y + y 2 ) + 41 kri k2 with respect to y ∈ [0, 2], by
using input-to-state stability theory [28], it follows that
limt→∞ y(t) = 0 given the fact limt→∞ ri (t) = 0. According to Comparison Principal [28], limt→∞ yi (t) = 0
is obtained, that is, limt→∞ qie (t) = 0, which, together with limt→∞ ri (t) = 0, further implies that
limt→∞ ωie (t) = 0, ∀i ∈ V.
Proposition 3.4 Consider the virtual position error system (29) with Assumptions 2.1 and 2.2. If
limt→∞ qie (t) = 0, limt→∞ p̄i (t) = 0 and limt→∞ v̄i (t) =
0 are achieved, the developed command force ui in (30)
guarantees that limt→∞ p̃ei (t) = 0 and limt→∞ ṽie (t) = 0,
∀i ∈ V.
Integrating both sides of (41), we obtain that
q
q
Lηi (t) − Lηi (0) ≤ ρ(kεi (0)k, 0)t, ∀t ≥ 0,
which indicates that Lηi cannot escape to infinity in finite
time. In addition, it follows from (40) that L̇ηi satisfies
e
Proof: It follows from (1) that (Ri (Qei ) − I3 )ê3 = ϕ×
i qi ,
ey ex
e T
c
where ϕi = [−qi , qi , σi ] . In terms of kRi (Qi )k = 1,
kQei k = 1 and (34), it follows from limt→∞ qie (t) = 0
Ti
Ti
that each m
(Ri (Qi )−Ri (Qci ))ê3 = m
Ri (Qci )(Ri (Qei )−
i
i
I3 )ê3 converges to zero asymptotically. This, together
with limt→∞ p̄i (t) = 0 and limt→∞ v̄i (t) = 0, guarantees
that the perturbation items ~1 and ~2 in the virtual position error system (32) converge to zero asymptotically.
Furthermore, it can be shown that system
L̇ηi ≤ − tanh(η̄i )T Λ tanh(η̄i )
1
+ kηi + η̇i k + (1 + )kη̇i k ρ(kεi (0)k, t)
kη
≤ − min{1, kη }kD tanh(η̄i )k2 + c2 kη̄i kρ(kεi (0)k, t)
≤ − c1 k tanh(η̄i )k2 + c2 kη̄i kρ(kεi (0)k, t),
(43)
where η̄i =
p̃˙ei = ṽie ,
ṽ˙ e = − lp p̃ei − lv ṽie ,
0
tanh(χ)T dχ+
Z
0
η̇i
tanh(χ)T dχ+
#
, c1 =
Λ = I3 ⊗
"
kη
kη
kη kη + 1
#
,D=
√
3− 5
2
c (χ )2 ∆
tanh(∆ )
L̇ηi ≤ −c1 (χη )2 kη̄i k2 + c2 kη̄i kρ(kεi (t1 )k, t − t1 ). (44)
Proof: Denote εi = lp p̃ei + lv ṽie for i ∈ V. It follows
from limt→∞ p̃ei (t) = 0 and limt→∞ ṽie (t) = 0 that
limt→∞ εi (t) = 0. To this end, there exists a KL-class
function ρ(kεi (0)k, t) such that kεi (t)k ≤ ρ(kεi (0)k, t).
For i ∈ V, the following Lyapunov function is proposed:
ηi +η̇i
11
+ η̇iT , η̇iT ]T ,
η
t1 ) ≤ 1 cη2 η for t ∈ [t1 , ∞), where χη <
is a
∆η
constant. In particular, k tanh(η̄i )k ≤ χη kη̄i k. Thus, for
t ≥ t1 , L̇ηi satisfies
Proposition 3.5 Consider the auxiliary system (31). If
limt→∞ p̃ei (t) = 0 and limt→∞ ṽie (t) = 0 are achieved,
then limt→∞ ηi (t) = 0 and limt→∞ η̇i (t) = 0, ∀i ∈ V.
Z
"
[ηiT
min{1, kη } and c2 = 1+ k1η . Since
12
Lηi cannot escape in finite time, there exist t1 and ∆η
such that kη̄i (t)k ≤ ∆η for t ∈ [0, t1 ] and ρ(kεi (t1 )k, t −
I3 ⊗
is asymptotically stable. Thus, it follows from inputto-state stability theory [28] that limt→∞ p̃ei (t) = 0
and limt→∞ ṽie (t) = 0, ∀i ∈ V given the fact that
limt→∞ ~(t) = 0, where ~ = [~T1 , ~T2 ]T .
Lηi =
(42)
This implies that L̇ηi is negative outside the set
Zi = η̄i ∈ R6 | kη̄i k ≤
1
kη̇i k2 .
2kη
c2
ρ(kεi (t1 )k, t − t1 ) .
c1 (χη )2
Thus, η̄i is bounded and ultimately converges to the
set Zi . In view of limt→∞ ρ(kεi (t1 )k, t − t1 ) = 0, it
9
100
80
←80 sec
pz (m)
60
←60 sec
p1
p2
p3
p4
pr
←40 sec
40
20
←20 sec
0
←0 sec
−20
5
0
py (m)
0
−5
10
5
px (m)
Fig. 2. Snapshots of coordinated trajectory tracking of
VTOL UAVs.
Fig. 1. Leader-follower graph Gn+1 .
follows that limt→∞ η̄i (t) = 0, which implies that
limt→∞ ηi (t) = 0 and limt→∞ η̇i (t) = 0, ∀i ∈ V.
pex (m)
2
Since Th2.i)-Th2.iii) have been proven, it can be concluded that the coordinated trajectory tracking of multiple VTOL UAVs is achieved in the sense of (6).
0
−2
pey (m)
−4
20
Simulations
In this section, simulations are performed to verify the
proposed distributed control approach on a formation
of four VTOL UAVs described by (2)-(5). The inertial
parameters are assumed to be identical: mi = 0.85(kg)
and Ji = diag{4.856, 4.856, 8.801} × 10−2 (kgm2 ), i =
1, 2, 3, 4. The leader-follower graph Gn+1 is illustrated
in Fig. 1, where each arrow denotes the corresponding
information flow. Furthermore, define dij = 1 if follower j is accessible to follower i, and dij = 0, otherwise, for i, j = 0, 1, · · · , 4. The desired trajectory is
described as pr (t) = [5 cos(0.2t), 5 sin(0.2t), t]T (m), and
the desired position offsets of the followers relative to
the leader are δ1 = [2, 2, 0]T (m), δ2 = [2, −2, 0]T (m),
δ3 = [−2, −2, 0]T (m) and δ4 = [−2, 2, 0]T (m), respectively. This indicates that the desired formation pattern is a square. The distributed estimator states of
each follower UAV are initialized as zeros. The follower
UAVs are initially situated at p1 (0) = [5, 3, −1]T (m),
p2 (0) = [9, −4, 1]T (m), p3 (0) = [4, −2, −3]T (m) and
p4 (0) = [−1, 4, −2]T (m) and their initial attitudes are
Qi (0) = [1, 0, 0, 0]T , i = 1, 2, 3, 4. The estimator and
control parameters are chosen as follows: kγ = 0.5, kp =
kv = 8 based on (12), la = 12 based on (13), ka = 4
based on (14), lp = lv = kη = 4 and lq = kq = 16. The
simulation results are illustrated in Figs. 2-4.
pe1
5
10
15
20
25
30
35
40
pe2
0
pe3
−2
10
pez (m)
4
−5
5
10
15
5
10
15
20
25
30
35
40
20
25
30
35
40
0
pe4
−1
−2
−3
0
time (sec)
vex (m/sec)
Fig. 3. Position error of follower VTOL UAVs.
4
2
0
vey (m/sec)
0
vez (m/sec)
v1e
−2
5
10
15
20
25
30
35
40
2
v2e
0
v3e
−2
0
2
5
10
15
20
25
30
35
40
v4e
1
0
−1
0
5
10
15
20
time (sec)
25
30
35
40
Fig. 4. Velocity error of follower VTOL UAVs.
that each error component converges to zero asymptotically. Consequently, the simulation results validate
that the proposed distributed control approach effectively guarantees the coordinated trajectory tracking of
multiple VTOL UAVs in the sense of (6).
Fig. 2 exhibits the evolution of the VTOL UAV formation with respect to the leader in the three-dimensional
space, where the formation is depicted every twenty seconds. It can be seen that the follower UAVs reach the prescribed square pattern while tracking the leader. Figs. 3
and 4 describe the position and velocity errors of the follower UAVs with respect to the leader. It can be observed
5
Conclusion
A distributed control strategy is proposed in this paper
to achieve the coordinated trajectory tracking of multiple VTOL UAVs with local information exchange. The
10
connectivity of the network graph is weak in the sense
that we only require the graph to contain a directed
spanning tree. A novel distributed estimator is firstly
designed for each VTOL UAV to obtain the leader’s desired information asymptotically. Then, under the hierarchical framework, a command force and an applied
torque are exploited for each VTOL UAV to fulfill the
accurate tracking to the desired information asymptotically. In addition, an auxiliary system is introduced in
the control development to avoid the non-singular command attitude extraction and the use of the unavailable
desired information. Simulations are carried out to validate the theoretical results.
[14] Z. Li, X. Liu, W. Ren, and L. Xie, “Distributed tracking
control for linear multiagent systems with a leader of bounded
unknown input,” IEEE Transactions on Automatic Control,
vol. 58, no. 2, pp. 518–523, 2013.
References
[18] J. Ghommam, L. F. Luque-Vega, B. Castillo-Toledo, and
M. Saad, “Three-dimensional distributed tracking control
for multiple quadrotor helicopters,” Journal of The Franklin
Institute, vol. 353, no. 10, pp. 2344–2372, 2016.
[15] W. Yu, G. Chen, and M. Cao, “Distributed leader-follower
flocking control for multi-agent dynamical systems with timevarying velocities,” Systems & Control Letters, vol. 59, no. 9,
pp. 543–552, 2010.
[16] J. Qin, C. Yu, and B. Anderson, “On leaderless and leaderfollowing consensus for interacting clusters of second-order
multi-agent systems,” Automatica, vol. 74, pp. 214–221, 2016.
[17] T. Yang, Z. Meng, G. Shi, Y. Hong, and K. H. Johansson,
“Network synchronization with nonlinear dynamics and
switching interactions,” IEEE Transactions on Automatic
Control, vol. 61, no. 10, pp. 3103–3108, 2016.
[1] F. Giulietti, L. Pollini, and M. Innocenti, “Autonomous
formation flight,” IEEE Control Systems Magazine, vol. 20,
no. 6, pp. 34–44, 2000.
[19] F. Liao, R. Teo, J. Wang, X. Dong, F. Lin, and K. Peng,
“Distributed formation and reconfiguration control of vtol
uavs,” IEEE Transactions on Control Systems Technology,
vol. 25, no. 1, pp. 270–277, 2017.
[2] M. Hua, T. Hamel, P. Morin, and C. Samson, “Introduction
to feedback control of underactuated vtol vehicles: A review
of basic control design ideas and principles,” IEEE Control
Systems Magazine, vol. 22, no. 1, pp. 61–75, 2013.
[20] A. Mahnmood and Y. Kim, “Leader-following formation
control of quadcopters with heading synchronization,”
Aerospace Science and Technology, vol. 47, pp. 68–74, 2015.
[3] Z. Zuo, “Trajectory tracking control design with commandfiltered compensation for a quadrotor,” IET Control Theory
and Applications, vol. 4, no. 11, pp. 2343–2355, 2010.
[21] X. Dong, Y. Zhou, Z. Ren, and Y. Zhong, “Timevarying formation tracking for second-order multi-agent
systems subjected to switching topologies with application
to quadrotor formation flying,” IEEE Transactions on
Industrial Electronics, vol. DOI: 10.1109/TIE.2016.2593656,
2016.
[4] Y. Zhang and Y. Tian, “Consentability and protocol design
of multi-agent systems with stochastic switching topology,”
Automatica, vol. 45, no. 5, pp. 1195–1201, 2009.
[5] A. Abdessameud and A. Tayebi, “Formation control of vtol
uavs,” in 48th IEEE Conference on Decision and Control,
Shanghai, P. R. China, 2009, pp. 3454–3459.
[22] M. Shuster, “A survey of attitude representations,” The
Journal of the Astronautical Sciences, vol. 41, no. 4, pp. 439–
517, 1993.
[6] A. Abdessameud, I. G. Polushin, and A. Tayebi, “Motion
coordination of thrust-propelled underactuated vehicles with
intermittent and delayed communications,” Systems &
Control Letters, vol. 79, pp. 15–22, 2015.
[23] Z. Qu, Cooperative control of dynamical systems: applications
to autonomous vehicles. Berlin: Springer, 2009.
[7] X. Dong, B. Yu, Z. Shi, and Y. Zhong, “Time-varying
formation control for unmanned aerial vehicles: Theories
and experiments,” IEEE Transactions on Control Systems
Technology, vol. 23, no. 1, pp. 340–348, 2015.
[24] B. Paden and S. Sastry, “A calculus for computing
filipovs differential inclusion with application to the variable
structure control of robot manipulators,” IEEE Transactions
on Circuits and Systems, vol. CAS-34, no. 1, pp. 73–82, 1987.
[8] Y. Hong, G. Chen, and L. Bushnell, “Distributed observers
design for leader-following control of multi-agent networks,”
Automatica, vol. 44, no. 3, pp. 846–850, 2008.
[25] D. Shevitz and B. Paden, “Lyapunov stability theory of
nonsmooth systems,” IEEE Transactions on Automatic
Control, vol. 39, no. 9, pp. 1910–1914, 1994.
[9] B. Yun, B. Chen, K. Lum, and T. Lee, “Design and
implementation of a leader-follower cooperative control
system for unmanned helicopters,” Journal of Control Theory
and Applications, vol. 8, no. 1, pp. 61–68, 2010.
[26] F. Clarke, Optimization and nonsmooth analysis. New York:
Wiley, 1983.
[27] N. Fischer, R. Kamalapurkar, and W. Dixon, “Lasalleyoshizawa corollaries for nonsmooth systems,” IEEE
Transactions on Automatic Control, vol. 58, no. 9, pp. 2333–
2338, 2013.
[10] D. Mercado, R. Castro, and R. Lozano, “Quadrotors flight
formation control using a leader-follower,” in European
Control Conference, Zurich, Switzerland, 2013, pp. 3858–
3863.
[28] H. Khalil, Nonlinear Systems, Third Edition.
Prentice-Hall, 2002.
[11] D. Lee, “Distributed backstepping control of multiple thrustpropelled vehicles on a balanced graph,” Automatica, vol. 48,
no. 11, pp. 2971–2977, 2012.
New Jersey:
[29] Y. Zou, “Nonlinear robust adaptive hierarchical sliding mode
control approach for quadrotors,” International Journal of
Robust and Nonlinear Control, vol. DOI: 10.1002/rnc.3607,
2016.
[12] A. Loria, J. Dasdemir, and N. A. Jarquin, “Leader-follower
formation and tracking control of mobile robots along straight
paths,” IEEE Transactions on Control Systems Technology,
vol. 24, no. 2, pp. 727–732, 2016.
[13] G. Wen, Y. Zhao, Z. Duan, and W. Yu, “Containment of
higher-order mmlti-leader multi-agent systems: A dynamic
output approach,” IEEE Transactions on Automatic Control,
vol. 61, no. 4, pp. 1135–1140, 2016.
11
| 3cs.SY
|
A meshless, integration-free, and boundary-only RBF technique
W. Chen* and M. Tanaka
Department of Mechanical System Engineering, Shinshu University, Wakasato 4-17-1, Nagano City,
Nagano, JAPAN (Present email:* [email protected] )
Abstract
Based on the radial basis function (RBF), non-singular general solution and dual reciprocity method
(DRM), this paper presents an inherently meshless, integration-free, boundary-only RBF collocation
techniques for numerical solution of various partial differential equation systems. The basic ideas
behind this methodology are very mathematically simple. In this study, the RBFs are employed to
approximate the inhomogeneous terms via the DRM, while non-singular general solution leads to a
boundary-only RBF formulation for homogenous solution. The present scheme is named as the
boundary knot method (BKM) to differentiate it from the other numerical techniques. In particular,
due to the use of nonsingular general solutions rather than singular fundamental solutions, the BKM is
different from the method of fundamental solution in that the former does no require the artificial
boundary and results in the symmetric system equations under certain conditions. The efficiency and
utility of this new technique are validated through a number of typical numerical examples.
Completeness concern of the BKM due to the only use of non-singular part of complete fundamental
solution is also discussed.
Key words: boundary knot method, dual reciprocity method, BEM, method of fundamental solution,
radial basis function, nonsingular general solution
1
1. Introduction
It has long been claimed that the boundary element method (BEM) is a viable alternative to the
domain-type finite element method (FEM) and finite difference method (FDM) due to its advantages
in dimensional reducibility and suitability to infinite domain problems. However, nowadays the FEM
and FDM still dominate science and engineering computations. The major bottlenecks in performing
the BEM analysis have long been its weakness in handling inhomogeneous terms such as
time-dependent and nonlinear problems. The recent introduction of the dual reciprocity BEM
(DRBEM) by Nardini and Brebbia [1] greatly eases these inefficiencies. Notwithstanding, as was
pointed out in [2], the method is still more mathematically complicated and requires strenuous manual
effort compared with the FEM and FDM. In particular, the handling of singular integration is not easy
to non-expert users and often computationally expensive. The use of the low order approximation in
the BEM also slows convergence. More importantly, just like the FEM, surface mesh or re-mesh in
the BEM requires costly computation, especially for moving boundary and nonlinear problems. The
method of fundamental solution (MFS) is shown an emerging technique to alleviate these drawbacks
and is getting increasing attraction especially due to some recent works by Golberg, Chen and
Muleskov [2-5]. The MFS affords advantages of integration-free, spectral convergence, and meshless.
However, the use of artificial boundary outside physical domain has been a major limitation of the
MFS, which may cause severe ill-conditioning of the resulting equations, especially for complex
boundary geometry [6,7]. These inefficiencies of the MFS motivate us to find an alternative technique,
which keeps its merits but removes its shortcomings undermining its attractiveness.
Recently, Golberg et al. [4] established the DRBEM on the firm mathematical theory of the radial
basis function (RBF). The DRBEM can be regarded as a two-step methodology. In terms of dual
reciprocity method (DRM), the RBF is applied at first to approximate the particular solution of
2
inhomogeneous terms, and then the standard BEM is used to discretize the remaining homogeneous
equation. Chen et al [3] and Golberg et al [2,4] extended this RBF approximation of particular
solution to the MFS, which greatly enhances its applicability. In fact, the MFS itself can also be
considered a special RBF collocation approach, where the fundamental solution of the governing
equation is taken as the radial function. On the other hand, the domain-type RBF collocation is also
now under intense study since Kansa’s pioneer work [8]. The major charming of the RBF-type
techniques is its meshless inherence. The construction of mesh in high dimension is not a trivial work.
Unlike the recently developed meshless FEM with the moving least square, the RBF approach is truly
cheap meshless technique without any difficulty applying boundary conditions [9, 10]. The RBF is
therefore an essential component in this study to construct a viable numerical technique.
Kamiya et al. [11] and Chen et al. [12] pointed out that the multiple reciprocity BEM (MRM) solution
of the Helmholtz problems with the Laplacian plus its high-order terms is in fact to employ only the
singular real part of complete complex fundamental solution of the Helmholtz operator. Power [13]
simply indicated that the use of either the real or imaginary part of the Helmholtz Green’s
representation formula could formulate interior Helmholtz problems. This study extends these ideas to
general problems such as Laplace and convection-diffusion problems by a combined use of the
nonsingular general solution, the dual reciprocity method, and RBF. This mixed technique is named as
the boundary knot method (BKM) [14] due to its essential meshless property, namely, the BKM does
not need any discretization grids for any dimension problems and only uses knot points. The inherent
inefficiency of the MFS due to the use of the fictitious boundary is alleviated in the BKM, which leads
to tremendous improvement in computational efficiency and produces the symmetric matrix structure
under certain conditions. It is noted that the BKM dose not involve integration operation due to the
use of the collocation technique. Just like the MFS, the method is very simple to implement. The
3
nonsingular general solution for multidimensional problems can be understood the nonsingular part of
complete fundamental solution of various operators. The preliminary numerical studies of this paper
show that the BKM is a promising technique in terms of efficiency, accuracy, and simplicity. We also
use the BKM with the response knot-dependent nonsingular general solutions to solve the varying
parameter problems successfully.
This paper is organized as follows. Section 2 involves the procedure of the DRM and RBF
approximation to particular solution. In section 3, we introduce nonsingular general solution and
derive the analogization equations of the BKM. Numerical results are provided and discussed in
section 4 to establish the validity and accuracy of the BKM. Completeness concern of the BKM is
discussed in section 5. Finally, section 6 concludes some remarks based on the reported results. In
appendix, we give the nonsingular general solution of some 2D, 3D steady and time-dependent
operators.
2. RBF approximation to particular solution
Like the DRBEM and MFS, the BKM can be viewed as a two-step numerical scheme, namely, DRM
and RBF approximation to particular solution and the evaluation of homogeneous solution. The latter
is the emphasis of this paper. The former has been well developed now [2-4]. For the sake of
completeness, we here outline the basic methodology to approximate particular solution. Let us
consider the differential equation
L {u( x )}= f ( x ), x ∈Ω
(1)
with boundary conditions
u(x ) = b1 (x ) ,
x ⊂ Γu ,
(2)
4
∂ u( x )
= b2 (x ) ,
∂n
x ⊂ ΓT ,
(3)
where L is a differential operator, f(x) is a known forcing function. n is the unit outward normal.
d
x ∈ R , d is the dimension of geometry domain, which is bounded by a piece-wise smooth boundary
Γ = Γu + ΓT . In order to facilitate discussion, it is assumed here that operator L includes Laplace
operator, namely,
L{u} = ∇ 2 u + L1 {u}.
(4)
We should point out that this assumption is not necessary [15]. Eq. (1) can be restated as
∇ 2 u + u = f (x ) + u − L1 {u}.
(5)
The solution of the above equation (5) can be expressed as
u = v + up ,
(6)
where v and up are the general and particular solutions, respectively. The latter satisfies equation
∇ 2 up + u p = f (x ) + u − L1 {u},
(7)
but does not necessarily satisfy boundary conditions (2) and (3). v is the homogeneous solution of the
Helmholtz equation
2
∇ v +v =0,
x ∈Ω ,
(8)
v (x ) = b1 (x ) − up , x ⊂ Γu ,
∂u ( x )
∂v(x )
= b2 ( x ) − p ,
∂n
∂n
(9)
x ⊂ ΓT .
(10)
The first step in the BKM is to evaluate the particular solution up by the DRM and RBF. After this, Eq.
(8)-(10) can be solved by the boundary RBF methodology using non-singular general solution
proposed in later section 3.
Unless the right side of Eq. (7) is rather simple, it is practically impossible to get analytical particular
solution in general cases. In addition, even if the analytical solutions for some problems are available,
5
their forms are usually too complicated to use in practice. Therefore, we prefer to approximate these
inhomogeneous terms numerically. The DRM with the RBF is a very promising approach for this task
[1-5], which analogizes the particular solution by the use of a series of approximate particular solution
at all specified nodes. The right side of Eq. (7) is approximated by the RBF approach, namely,
f (x ) + u − L1 {u} ≅
∑α φ ( x − x
N+L
j
j
j=1
)+ ψ (x ),
(11)
where α j are the unknown coefficients. N and L are respectively the numbers of knots on boundary
and domain.
represents the Euclidean norm, φ (
)
is the RBF. Additional polynomial term ψ(x)
is required to assure nonsingularity of interpolation matrix if the RBF is conditionally positive definite
such as multiquadratics (MQ) and thin plate spline (TPS) [8,16]. For example, in the 2D case with
linear polynomial restrains we have
f (x ) + u − L1 {u} ≅
N+L
∑α φ (r )+ α
j
j
j=1
N + L +1
x + α N + L +2 y + α N + L +3 ,
(12)
where rj = x − x j . The corresponding side conditions are given by
N+L
N+L
N+L
j =1
j =1
j =1
∑α j =
∑α j x j =
∑α y
j
j
=0.
(13)
By forcing Eq. (12) to exactly satisfy Eqs. (7) and (13) at all nodes, we can get a set of simultaneous
equations to uniquely determine the unknown coefficients α j . In this procedure, we need to evaluate
the approximate particular solutions in terms of the RBF φ . The standard approach is that φ in Eq.
(11) is first selected, and then corresponding approximate particular solutions are determined by
analytically integrating differential operator. The advantage of this method is that it is a
mathematically reliable technique. However, this methodology easily performs only for simple
operators and RBFs. Recently Muleskov et al. [5] made a substantial advance to discover the analytic
approximate particular solutions for Helmholtz-type operators using the polyharmonic splines. But the
analytical approximate particular solutions for general cases such as the MQ and other differential
operators are not yet available now due to great difficulty involved. Another scheme evaluating
6
approximate solution is a reverse approach [17,18]. Namely, the approximate particular solution is at
first chosen, and then we can evaluate the corresponding φ by simply substituting the specified
particular solution into certain operator of interest. It is a very difficult task to mathematically prove
under what conditions this approach is reliable, although it seems to work well so far for many
problems [17-19]. This scheme is in fact equivalent to the approximation of particular solution using
the Kansa’s method [8,9]. In this study, we use this scheme in terms of the MQ. The chosen
approximate particular solution is
ϕ (rj )= (rj2 + c 2j ) ,
32
(14)
where cj is the shape parameter. The corresponding MQ-like radial function is
φ (rj )= 6(rj2 + c 2j )+
32
3r 2
+ (rj2 + c 2j ) ,
2
2
rj + c j
(15)
Finally, we can get particular solutions at any point by weighted summation of approximate particular
solutions at all nodes with coefficients α j . For more details on the procedure see [1-5].
3. Non-singular general solution and boundary knot method
One may think that the placement of source points outside domain in the MFS is to avoid the
singularities of fundamental solutions. However, we found through numerical experiments that even if
all source and response points were placed differently on physical boundary to circumvent the
singularities, the MFS solutions were still degraded severely. In the MFS, the more distant the source
points are located from physical boundary, the more accurate MFS solutions are obtained [2].
However, unfortunately the resulting equations can become extremely ill-conditioned which in some
cases deteriorates the solution [2,6,7].
To illustrate the basic idea of the boundary collocation using nonsingular general solution, we take the
7
2D Helmholtz operator as an illustrative example, which is the simplest among various
often-encountered operators having nonsingular general solution. The Laplace operator has not
nonsingular general solution. For the other nonsingular general solutions see appendix.
The 2D homogeneous Helmholtz equation (8) has two general solutions, namely,
v (r) = c1 J 0 ( r) + c2 Y0 (r ) ,
(16)
where J0(r) and Y0(r) are the zero-order Bessel functions of the first and second kinds, respectively. In
the standard BEM and MFS, the Hankel function
H(r ) = J0 (r) + iY0 (r)
(17)
is applied as the fundamental solution. It is noted that Y0(r) encounters logarithm singularity, which
causes the major difficulty in applying the BEM and MFS. Many special techniques have been
developed to solve or circumvent this singular trouble.
The present BKM scheme discards the singular general solution Y0(r) and only use J0(r) as the radial
function to collocate the boundary condition Eqs. (9) and (10). It is noted that J0(r) exactly satisfies
the Helmholtz equation and we can therefore get a boundary-only collocation scheme. Unlike the
MFS, all collocation knots are placed only on physical boundary and can be used as either source or
response points.
Let
{xk }k =1
N
denote a set of nodes on the physical boundary, the homogeneous solution v(x) of Eq. (8)
is approximated in a standard collocation fashion
N
v (x ) = ∑ β k J0 (rk ) ,
(18)
k =1
where rk = x − xk . k is the index of source points. N is the number of boundary knots. βk are the
8
desired coefficients. Collocating Eqs. (9) and (10) in terms of Eq. (18), we have
N
∑ β J (r ) = b (x ) − u (x ) ,
k =1
k
N
∑β
k =1
k
0
ik
1
i
p
(19)
i
∂J 0 (rjk )
∂u (x )
= b2 (x j ) − p j ,
∂n
∂n
(20)
where i and j indicate Dirichlet and Neumann boundary response knots, respectively. If internal nodes
are used, we need to constitute another set of supplement equations
N
∑ β J (r ) = u
k =1
k
0
lk
l
− u p (xl ) ,
l = 1, K , L ,
(21)
where l indicates the internal response knots and L is the number of interior points. Now we get total
N+L simultaneous algebraic equations. It is stressed that the use of interior points is not always
necessary in the BKM as in the DRBEM [15,17,20]. The term “boundary-only” is used here in the
sense as in the DRBEM and MFS that only boundary knots are required, although internal knots can
improve solution accuracy in some cases.
Before proceeding the numerical experiments, we consider to choose the radial basis function. In
general, RBFs are globally defined basis functions and lead to a dense matrix, which becomes highly
ill-conditioned if very smooth radial basis functions are used with large number of interpolation nodes
[16]. This causes severe stability problems and computationally inefficiency for large size problem. A
number of approaches have been proposed to remedy this problem such as domain decomposition and
compactly supported RBFs (CS-RBFs). The latter is recently developed by Wendland [21], Wu [22]
and Schaback [23]. Golberg et al. [2], Wong et al. [24] and Chen et al [25] respectively applied the
CS-RBFs to the MFS, Kansa’s method and DRBEM successfully. However, in this study we will not
use the CS-RBFs to focus on the illustration of the basic idea of the BKM with globally-supported
RBFs.
9
The MQ [26], TPS [27] and linear RBF [1] are most widely used globally-defined RBFs now. Among
them, it is well known that the MQ ranks the best in accuracy [28] and is the only available RBF with
desirable merit of spectral convergence [4]. However, its accuracy is greatly influenced by the shape
parameter [29,30]. So far the optimal determination of shape parameter is still an open research topic.
Despite this problem, the MQ is still most widely used in the RBF solution of various differential
systems. For numerical example of the Laplace equation shown in the next section 4.2, the linear and
generalized TPS RBFs show evidently slower convergence rate than the MQ. For example, the
average relative error is 0.91% for the Linear RBF with 11 knots, 0.39% for the TPS with 11 knots,
and 0.023% for the MQ (shape parameter 2) with 9 knots. We even got 0.5% relative average error by
the MQ with only 3 knots. To simplify the presentation, this paper only uses the MQ with the DRM
although the use of the TPS is also attractive in many cases.
4. Numerical results and discussions
In this paper, all numerical examples unless otherwise specified are taken from [20]. The geometry of
test problem is all an ellipse featured with semi-major axis of length 2 and semi-minor axis of length 1.
These examples are chosen since their analytical and numerical solutions are obtainable to compare.
More complicated problems can be handled in the same BKM fashion without any extra difficulty.
The zero order Bessel and modified Bessel functions of the first kind are evaluated via short
subroutines given in [31]. The 2D Cartesian co-ordinates (x,y) system is used as in [20].
4.1. Helmholtz equation
The 2D homogeneous Helmholtz equation is given by
2
∇ u+u = 0
(22)
with inhomogeneous boundary condition
10
u = sin x .
(23)
It is obvious that Eq. (23) is also a particular solution of Eq. (22). Numerical results by the present
BKM is displayed in Table 1 together with those by the DRBEM for comparison.
The numbers in the brackets of Table 1 mean the total nodes used. It is found that the present BKM
converges very quickly. This demonstrates that the BKM enjoys the super-convergent property as in
the other types of collocation methods [32]. The BKM solutions using 7 nodes are adequately accurate.
In stark contrast, the DRBEM with 16 boundary and 17 interior points [20] produced relatively less
accurate solution due to the use of the Laplacian and the low order of BEM convergence ratio. Please
note that in this case there is no particular solution to be approximated by using the RBF and DRM in
the BKM.
4.2. Laplace equation
Readers may argue that it is somehow unfair to choose the homogeneous Helmholtz equation to
compare the BKM and DRBEM. The latter used the Laplace fundamental solution in the previous
example. In the following we will further justify the superconvergence of the BKM through a
comparison with the BEM for Laplace equation
2
∇ u=0
(24)
with boundary condition
u = x + y.
(25)
Eq. (25) is easily found to be a particular solution of Eq. (24). This homogeneous problem is typically
well suited to be handled by the standard BEM technique. In contrast, there is inhomogeneous term in
the BKM formulation to apply the nonsingular general solution of the Helmholtz operator. Namely, Eq.
(24) is rewritten as
11
2
∇ u + u = u,
(26)
where the right inhomogeneous term u is approximated by the DRM as shown in the section 2. The
numerical results are displayed in Table 2 where the BEM solutions come from [20].
The MQ shape parameter c is set 25 for both 3 and 5 boundary knots in the BKM. It is observed that
the BKM solutions are not sensitive to the parameter c. It is seen from Table 2 that the BKM results
using 3 boundary nodes achieve the accuracy of four significant digits and are far more accurate than
the BEM solution using 16 boundary nodes. This striking accuracy of the BKM again validates its
spectral convergence. In this case only boundary points are employed to approximate the particular
solution by the DRM and RBF. It is noted that the coefficient matrices of the BEM and BKM are both
fully populated. Unlike the BEM, however, the BKM yields symmetric coefficient matrix for all
self-adjoint operators with one type of boundary conditions. This Laplace problem is a persuasive
example to verify high accuracy and efficiency of the BKM vis-a-vis the BEM.
4.3. Convection-diffusion problems
The FEM and FDM encounter some difficulty to produce accurate solution to the systems involving
the first order derivative of convection term. Special care need be taken to handle this problem with
these methods. It is claimed that the BEM does not suffer similar accuracy problem. In particular, the
DRBEM was said to be very suitable for this type problem [17,20]. Let us consider the convection
diffusion equation
∇ 2 u = − ∂u ∂x ,
(27)
which is given in [20] to test the DRBEM. The boundary condition is stated as
−x
u=e ,
(28)
which also constitutes a particular solution of this problem. By adding u on dual sides of Eq. (27), we
12
have
∇ 2 u + u = u − ∂u ∂x .
(29)
The results by both the DRBEM and BKM are listed in Table 3.
The MQ shape parameter is chosen 4. The BKM employed 7 boundary knots and 8 or 11 internal
knots. In contrast, the DRBEM [20] used 16 boundary and 17 inner nodes. It is stressed that unlike the
previous examples, in this case the use of the interior points can improve the solution accuracy
evidently. This is due to the fact that the governing equation has convection domain-dominant solution.
Only by using boundary nodes, the present BKM with the Helmholtz non-singular solution and the
DRBEM with the Laplacian [20] can not well capture convection effects of the system equation. It is
found from Table 3 that both the BKM and DRBEM achieve the salient accurate solutions with inner
nodes. The BKM outperforms the DRBEM in computational efficiency due to the super-convergent
features of the MQ interpolation and global BKM collocation.
Further consider equation
∇ 2 u = − ∂u ∂x − ∂u ∂y
(30)
with boundary conditions
−x
−y
u= e +e ,
(31)
which is also a particular solution of Eq. (30). The numerical results are summarized in Table 4.
In the BKM, the MQ shape parameter is taken 5.5. We employed 7 boundary knots and 8 or 11 inner
points in the BKM compared with 16 boundary nodes and 17 inner points in the DRBEM [20]. The
BKM worked equally well in this case as in the previous ones. It is seen from Table 4 that the BKM
with fewer points produced almost the same accurate solutions as the DRBEM. Considering the
13
extremely mathematical simplicity and easy-to-use advantages of the BKM, the method is superior to
the DRBEM in this problem.
4.4. Varying-parameter Helmholtz problem
Consider the varying-parameter Helmholtz equation
2
∇ u−
2
u=0
x2
(32)
with inhomogeneous boundary condition
u = −2 / x .
(33)
Eq. (33) is also a particular solution of Eq. (32). This problem is the simplified Berger
convection-diffusion problem given in [20]. Note that the origin of the Cartesian co-ordinates system
is dislocated to the node (3,0) to circumvent singularity at x=0. The response knot-dependent
nonsingular general solution of varying parameter equation (32) is given by
2
u(rik , xi ) = I0
rik ,
xi
(34)
where Io is the zero order modified Bessel function of the first kind. i and k respectively index the
response and source nodes. In terms of the BKM, the problem can be analogized by
N
∑ α I (r
k =1
k 0
ik
2 xi )= −2 / xi .
(35)
Note that only the boundary nodes are used in Eq. (35). After evaluating the coefficients α, we can
easily evaluate the value of u at any inner node p by
N
up = ∑α k I0 (rpk 2 x p ).
(36)
k =1
Table 5 lists the BKM results against the DRBEM solutions. The BKM average relative errors under
N=9, 13, 15 are respectively 9.7e-3, 8.1e-3 and 7.6e-3, which numerically demonstrates its
convergence. The accuracies of the BKM and DRBEM solutions are comparable. It is noted that the
14
DRBEM used 33 nodes (16 inner and 17 boundary knots) in this case, while the BKM only employed
much less boundary knots. The accuracy and efficiency of this BKM scheme are very encouraging.
Note that the present BKM representation differs from the previous ones in that we here use the
response point-dependent nonsingular general functions. Similarly, we can easily constitute response
node-dependent fundamental solutions. Thus, the essential idea behind this work may be extended to
the BEM and DRBEM solution of varying parameter problems. For example, unlike the DRBEM
scheme for varying velocity convection-diffusion problems given in Ref. [17], the variable
convection-diffusion fundamental solutions with response node-dependent velocity parameters may
be employed to the BEM or the DRBEM formulations, which may be especially attractive for high
Peclet number problems.
5. Completeness concern
One of major potential concerns of the BKM is its completeness due to the fact that the BKM employs
only the non-singular part of fundamental solutions of differential operators. This incompleteness may
limit its utility. Although the given numerical experiments favor the method, now we can not
theoretically ascertain of the general applicability of the BKM. On the other hand, Kamiya and Andoh
[11] validated that the similar incompleteness occurs in the multiple reciprocity BEM using the
Laplacian for Helmholtz operators. Namely, if the Laplace fundamental solution plus its higher-order
terms are used in the MRM for the Helmholtz problems as in its usual form, we actually employ only
the singular part of Helmholtz operator fundamental solution. Although the MRM performed well in
many numerical experiments, it is mathematically incomplete. It is interesting to note that the BKM
and MRM respectively employ the nonsingular and singular parts of the complete complex
fundamental solution. It should also be stressed that although Power [13] simply indicated that the
singular or nonsingular parts of Green representation can formulate the interior Helmholtz problems,
15
no any related numerical and theoretical results are available from the published reports.
Kamiya and Andoh [11] also pointed out that the MRM formulation with the Laplacian can not satisfy
the well-known Sommerfeld radiation conditions at infinity. Chen et al. [12] addressed the issues
relating to spurious eigenvalues of the MRM with the Laplacian. Some literatures referred to in [12]
also discussed the issues applying MRM with the Laplacian to problems with degenerate boundary
conditions. These concerns of the MRM raise some cares concerning the applicability of the BKM
which implements the nonsingular part of fundamental solution compared to the MRM using the
singular part. Power [13] discussed the incompleteness issue of the MRM for the Brinkman equation
and indicated that using one part of complex fundamental solution of Helmholtz operator may fail to
the exterior Helmholtz problems. However, now we can not justify whether or not the BKM works for
exterior problems since the method differs the MRM in using the DRM approximation of particular
solutions. Dai [33] successfully applied the dual reciprocity BEM with the Laplacian to waves
propagating problems in an infinite or semi-infinite region. It is worth pointing out that the Laplace
fundamental solution used in the DRBEM also does not satisfy the Sommerfeld radiation condition.
Unlike the MRM, the BKM and DRBEM do not employ the higher-order fundamental solutions to
approximate the particular solution. Our next work will investigate if the BKM with the DRM can
analyze the unbounded domain problems.
In fact, all existing numerical techniques encounter some limits. The BKM is not exceptional. Power
[13] pointed out that the incompleteness in the MRM is problem-dependent. Therefore, the essential
issue relating to the concerns of the BKM completeness is under what conditions the method works
reliably and efficiently.
16
6. Concluding remarks
The present BKM can be regarded one kind of the Trefftz method [35] where the trial function is
required to satisfy governing equation. The BKM distinguishes from the other Trefftz techniques such
as the MFS in that we employ nonsingular general solution. The shortcomings of the MFS using
fictitious boundary are eliminated in the BKM. The term “BKM” can be interpreted as a boundary
modeling technique combining the DRM, RBF, and nonsingular general solution. In conclusion, the
presented BKM is inherently possesses some desirable numerical merits which include meshless,
boundary-only, integration-free, and mathematical simplicity. The implementation of the method is
remarkably easy. The remaining two concerns of the BKM are the possible incompleteness in solving
some types of problems due to the only use of nonsingular general solution and solvability of the
Kansa’s method for finding the particular solutions. More numerical experiments to test the BKM will
be beneficial. This paper can be regarded a starting point of a series of works.
Appendix
By its very basis, it is straightforward to extend the BKM to the nonlinear, three-dimensional,
time-dependent partial differential systems. The following lists the nonsingular general solutions of
some important steady and transient differential operators.
For the 3D Helmholtz-like operators
∇ u± λ u =0,
2
2
(A1)
we respectively have the nonsingular general solution
u* = A
sin(λr )
r
(A2)
and
17
u* = A
sinh(λr )
,
r
(A3)
where sinh denotes the hyperbolic function, A is constant. For the 2D biharmonic operator
∇ w −λ w =0,
4
2
(A4)
we have the nonsingular general solution
w = A1 J0 (λ r) + A2 I0 (λ r ).
*
(A5)
The non-singular general solution of the 3D biharmonic operator is given by
w * = A1
sin(λ r)
sinh(λr )
+ A2
.
r
r
(A6)
For the 3D time-dependent heat and diffusion equation
∆u =
1 ∂u
,
k ∂t
(A7)
we have the nonsingular general solution
u (r, t, t k ) = Ae
*
−k (t −t k )
sin(r)
.
r
(A8)
Furthermore, consider 3D transient wave equation
1 ∂ 2u
∆u = 2 2 ,
c ∂t
(A9)
we get the general solution
sin (r)
A
u* (r,t,t k ) = A1 cos(c(t − tk ))+ 2 sin (c(t − t k ))
.
r
c
(A10)
The 2D nonsingular general solutions of transient problems can be easily derived in a similar fashion.
Two standard techniques handling time derivatives are the time-stepping integrators and the model
analysis. The former involves some difficult issues relating to the stability and accuracy, while the
latter is not very applicable for many cases such as shock. The BKM using time-dependent
nonsingular solutions may circumvent these drawbacks. The difficulty implementing such BKM
schemes may lie in how to satisfy the inharmonic initial conditions inside domain as in the
time-dependent BEM. On the other hand, the time-dependent nonsingular general solutions may be
18
directly applied in the domain-type RBF collocation schemes such as the Kansa’s method. It is worth
pointing out that the analogous method proposed by Katsikadelis et al. [15] may be combined with the
BKM to handle the differential systems which do not include Laplace or biharmonic operators.
Acknowledgements:
Some valuable comments from a referee reshaped this paper into its present form. The authors also
express grateful acknowledgement of helpful discussions with M. Golberg. Y.C. Hon and A.H.D.
Cheng. The first author was supported as a JSPS fellow by the Japan Society of Promotion of Science.
References:
1. D. Nardini and C.A. Brebbia, A new approach to free vibration analysis using boundary elements.
Applied Mathematical Modeling, 7 157-162 (1983).
2. M.A. Golberg and C.S. Chen, The method of fundamental solutions for potential, Helmholtz and
diffusion problems. In Boundary Integral Methods - Numerical and Mathematical Aspects, (Edited
by M.A. Golberg), pp. 103-176, Computational Mechanics Publications, (1998).
3. C.S. Chen, The method of potential for nonlinear thermal explosion, Commun Numer. Methods
Engng, 11 675-681 (1995).
4. M.A. Golberg, C.S. Chen, H. Bowman and H. Power, Some comments on the use of radial basis
functions in the dual reciprocity method, Comput. Mech. 21 141-148 (1998).
5. A.S. Muleskov, M.A. Golberg and C.S. Chen, Particular solutions of Helmholtz-type operators
using higher order polyharmonic splines, Comput. Mech. 23 411-419 (1999).
6. T. Kitagawa, On the numerical stability of the method of fundamental solutions applied to the
Dirichlet problem. Japan Journal of Applied Mathematics, 35 507-518, (1988).
7. T. Kitagawa, Asymptotic stability of the fundamental solution method. Journal of Computational
19
and Applied Mathematics, 38 263-269 (1991).
8. E.J. Kansa, Multiquadrics: A scattered data approximation scheme with applications to
computational fluid-dynamics. Comput. Math. Appl. 19 147-161 (1990).
9. E.J. Kansa and Y.C. Hon, Circumventing the ill-conditioning problem with multiquadric radial
basis functions: applications to elliptic partial differential equations. Comput. Math. Appls. 39
123-137 (2000).
10.Y.C. Hon and X.Z. Mao, A radial basis function method for solving options pricing model.
Financial Engineering, 81(1) 31-49 (1999).
11.N. Kamiya and E. Andoh, A note on multiple reciprocity integral formulation for the Helmholtz
equation, Commun. Numer. Methods Engng., 9 9-13 (1993).
12.J.T. Chen, C.X. Huang, and K.H. Chen, Determination of spurious eigenvalues and multiplicities
of true eigenvalues using the real-part dual BEM, Comput. Mech., 24 41-51 (1999).
13.H. Power, On the completeness of the multiple reciprocity series approximation, Commun. Numer.
Methods Engng., 11 665-674 (1995).
14.W. Chen and M. Tanaka, New advances in dual reciprocity and boundary-only RBF methods, In
Proceeding of BEM technique conference,(Edited by M. Tanaka), Vol. 10, pp. 17-22, Tokyo, Japan,
(2000).
15.J.T. Katsikadelis and M.S.Nerantzaki, The boundary element method for nonlinear problems,
Engineering Analysis with Boundary Element, 23 365-273 (1999).
16.M. Zerroukat, H. Power and C.S. Chen, A numerical method for heat transfer problems using
collocation and radial basis function, Inter. J. Numer. Method Engng. 42 1263-1278 (1998).
17.L.C. Wrobel and D.B. DeFigueiredo, A dual reciprocity boundary element formulation for
convection-diffusion problems with variable velocity fields, Engng. Analysis with BEM, 8(6)
312-319 (1991).
20
18.N.A. Schclar, Anisotropic Analysis Using Boundary Elements, Comput. Mech. Publ., Southampton,
1994.
19.M. Kogl and L. Gaul, Dual reciprocity boundary element method for three-dimensional problems
of dynamic piezoelectricity. In Boundary Elements XXI, 537-548, Southampton, 1999.
20.P.W. Partridge, C.A. Brebbia and L.W. Wrobel, The Dual Reciprocity Boundary Element Method,
Comput. Mech. Publ., Southampton, UK (1992).
21.H. Wendland, Piecewise polynomial, positive definite and compactly supported radial function of
minimal degree, Adv. Comput. Math. 4 389-396 (1995).
22.Z. Wu, Multivariate compactly supported positive definite radial functions, Adv. Comput. Math. 4
283-292 (1995).
23.R. Schaback, Creating surfaces from scattered data using radial basis function. In Mathematical
Methods for Curves and Surfaces, (Edited by M. Dahlen et al.), pp. 477-496, Vanderbilt University
Press, Nashville, (1995).
24.S.M. Wong, Y.C. Hon and M.A. Golberg, Compactly supported radial basis functions for the
shallow water equations, Appl. Math. Comput. (to appear).
25.C.S. Chen, C.A. Brebbia and H. Power, Boundary element methods using compactly supported
radial basis functions, Commun. Numer. Meth. Engng. 15 137-150 (1999).
26.R.L. Hardy, Multiquadratic equations for topography and other irregular surfaces, J. Geophys. Res.,
176 1905-1915 (1971).
27.J. Duchon, Interpolation des fonctions de deux variables suivant le principe de la flexion des
plaques minces, RAIRO Analyse Numeriques, 10 5-12 (1976).
28.R. Franke, Scattered data interpolation: tests of some methods, Math. Comput. 48 181-200 (1982).
29.Y.C. Hon and X.Z. Mao, An efficient numerical scheme for Burgers' equation, Appl. Math. Comput.
95(1) 37-50 (1998).
21
30.R.E. Carlson and T.A. Foley, The parameter R2 in multiquadratic interpolation, Comput. Math.
Appl. 21 29-42 (1991).
31.W.H. Press, S.A. Teukolsky, W.T. Vetterling and B.P. Flannery, Numerical Recipes in Fortran,
Cambridge University Press (1992).
32.C. Canuto, M.Y. Hussaini, A. Quarteroni and T.A. Zang, Spectral Methods in Fluid Dynamics,
Springer, Berlin (1988).
33.D.N. Dai, An improved boundary element formulation for wave propagation problems, Engng.
Analysis with Boundary Element, 10 277-281 (1992).
34.J. Duchon, Splines minimizing rotation invariant semi-norms in Sobolov spaces, in: Constructive
Theory of Functions of Several Variables, Springer-Verlag, Berlin, 1976.
35.R. Piltner, Recent development in the Trefftz method for finite element and boundary element
application. Advances in Engineering Software, 2 107-115 (1995).
22
Table 1. Results for Helmholtz problem
x
y
Exact
DRBEM (33)
BKM (7)
BKM (11)
1.5
0.0
0.997
0.994
0.999
0.997
1.2
-0.35
0.932
0.928
0.931
0.932
0.6
-0.45
0.565
0.562
0.557
0.565
0.0
0.0
0.0
0.0
0.0
0.0
0.9
0.0
0.783
0.780
0.779
0.783
0.3
0.0
0.296
0.294
0.289
0.296
0.0
0.0
0.0
0.0
0.0
0.0
Table 2. Results for Laplace problem
x
y
Exact
BEM (16)
BKM (3)
BKM (5)
1.5
0.0
1.500
1.507
1.500
1.500
1.2
-0.35
0.850
0.857
0.850
0.850
0.6
-0.45
0.150
0.154
0.150
0.150
0.0
0.0
-0.450
-0.451
-0.450
-0.450
0.9
0.0
0.900
0.913
0.900
0.900
0.3
0.0
0.300
0.304
0.300
0.300
0.0
0.0
0.0
0.0
0.0
0.0
Table 3. Results for ∇ 2 u = − ∂u ∂x
x
y
Exact
DRBEM (33)
BKM(15)
BKM (18)
1.5
0.0
0.223
0.229
0.229
0.224
1.2
-0.35
0.301
0.307
0301
0.305
0.0
-0.45
1.000
1.003
1.010
1.000
-0.6
-0.45
1.822
1.819
1.822
1.818
-1.5
0.0
4.482
4.489
4.484
4.477
0.3
0.0
0.741
0.745
0.744
0.743
-0.3
0.0
1.350
1.348
1.353
1.354
0.0
0.0
1.000
1.002
1.003
1.004
23
Table 4. Results for ∇ 2 u = − ∂u ∂x − ∂u ∂y
x
y
Exact
DRBEM (33)
BKM(15)
BKM (18)
1.5
0.0
1.223
1.231
1.225
1.224
1.2
-0.35
1.720
1.714
1.725
1.723
0.0
-0.45
2.568
2.557
2.546
2.551
-0.6
-0.45
3.390
3.378
3.403
3.405
-1.5
0.0
5.482
5.485
5.490
5.491
0.3
0.0
1.741
1.731
1.729
1.731
-0.3
0.0
2.350
2.335
2.349
2.350
0.0
0.0
2.000
1.989
1.992
1.993
Table 5. Relative errors for varying parameter Helmholtz problem
x
4.5
4.2
3.6
3.0
2.4
1.8
1.5
3.9
3.3
3.0
2.7
2.1
y
0.0
-0.35
-0.45
-0.45
-0.45
-0.35
0.0
0.0
0.0
0.0
0.0
0.0
DRBEM(33)
2.3e-3
2.1e-3
5.4e-3
4.5e-3
1.2e-3
9.0e-4
*
3.9e-3
3.3e-3
4.5e-3
2.7e-3
3.2e-3
BKM (9)
3.3e-3
4.1e-3
6.8e-3
1.1e-2
1.4e-2
5.2e-3
9.4e-3
7.0e-3
1.1e-2
1.3e-2
1.5e-2
1.6e-2
BKM (15)
2.6e-3
3.3e-3
4.7e-3
4.4e-3
9.1e-4
1.7e-2
3.4e-2
5.3e-3
6.3e-3
5.6e-3
3.4e-3
8.8e-3
24
| 5cs.CE
|
arXiv:1710.11471v2 [cs.PF] 21 Nov 2017
Distributed Server Allocation for Content Delivery
Networks
Sarath Pattathil, Vivek S. Borkar and Gaurav S. Kasbekar
∗
Abstract
We propose a dynamic formulation of file-sharing networks in terms
of an average cost Markov decision process with constraints. By analyzing a Whittle-like relaxation thereof, we propose an index policy in
the spirit of Whittle and compare it by simulations with other natural
heuristics.
1
Introduction
Recently, Content Delivery Networks (CDNs), which distribute content (e.g.,
video and audio files, webpages) using a network of server clusters situated at multiple geographically distributed locations, have been extensively
deployed in the Internet by content providers themselves (e.g., Google) as
well as by third-party CDNs that distribute content on behalf of multiple
content providers (e.g., Akamai’s CDN distributes Netflix and Hulu content) [18], [14]. The delay incurred in downloading content to an end user is
often significantly lower when a CDN is used compared to the case where all
content is downloaded from a single centralized host, since the server clusters
of a CDN are located close to end users [18], [14].
In this paper, we consider a server cluster, which contains M ≥ 2 servers
and is part of a CDN. The server cluster stores N large files (e.g., videos).
There is a high demand for each file, and hence each file is replicated at
mutiple servers within the cluster. Requests for the N files from end users
∗
Sarath Pattathil, Vivek S. Borkar and Gaurav Kasbekar are with the Department of
Electrical Engineering, Indian Institute of Technology Bombay, Powai, Mumbai 400076,
India. Email: [email protected], [email protected], [email protected].
The work of VSB was supported in part by a J. C. Bose Fellowship and a grant for ‘Approximation of High Dimensional Optimization and Control Problems’ from the Department
of Science and Technology, Government of India.
1
or from smaller server clusters arrive at the server cluster from time to time.
Now, there are two approaches to serving the file requests [26, 28]: (i) Single
Server Allocation: Each file request is served by a single server [26, 28].
(ii) Resource Pooling: Each file request is simultaneously served by multiple
servers– in particular, different chunks of the file are served by different
servers in parallel [26, 28]. Resource pooling has been found to outperform
single server allocation in prior studies [26, 27, 28] and hence, in this paper,
we assume that resource pooling is used. Also, we allow multiple files to be
simultaneously downloaded from a given server. At any time instant, the
sum of the rates at which a server j transmits different files is constrained to
be at most µj . Requests for different files are stored in different queues, and
there is a cost for storing a request in a queue. Let ξ ij (t) be the rate at which
server j transmits file i at time t. We consider the problem of determining
the rates ξ ij (t) for each i, j and t so as to minimize the total storage cost.
We formulate this problem as a Markov Decision Process (MDP) [11]. We
show that this problem is Whittle-like indexable [33] and use this result to
propose a Whittle-like scheme [33] that can be implemented in a distributed
manner 1 . We evaluate the performance of our scheme using simulations and
show that it outperforms several natural heuristics for the problem such as
Balanced Fair Allocation, Uniform Allocation, Weighted Allocation, Random
Allocation and Max-Weight Allocation.
We now review related prior literature. In [27], performance of Content
Delivery Networks is evaluated in a static framework. This work also studies
the tradeoffs between delay for each packet vs the energy used etc. The polymatroid structure of the service capacity in this model is exploited to get an
expression for mean file transfer delay that is experienced by incoming file
requests. Performance of dynamic server allocation strategies, such as random server allocation or allocation of least loaded server, are also explored.
We use the model of [27] for CDN, but go a step further by looking at a fully
dynamic optimization problem as an MDP.
In [28], a centralized content delivery systems with collocated servers is
studied. Files are replicated in these servers and these serve as a pooled resource which cater to file requests. The article shows how dynamic server capacity allocation outperforms simple load balancing strategies such as those
which assign the least loaded server, or assign the servers at random. The
article also goes on to study file placement strategies that improve the utility
1
We use the phrase ‘Whittle-like’ instead of just Whittle because the scheme introduced
in this paper, although is in the same spirit of Whittle’s original paper, is not exactly the
same.
2
of the system.
Several works including [16, 20, 34, 17] look at large-scale content delivery networks, focusing on placement of content in the servers. Of these,
[16] also studies the greedy method of server allocation and its efficiency
under various regimes of server storage capacities, and under what content
placement strategy it would be efficient. Article [34] studies strategies for
scheduling after the content placement stage, and proposes an algorithm,
called the Fair Sharing with Bounded Degree (FSBD), for server allocation.
In [29] muliclass queuing systems are studied with different arrival rates.
The service rates are constrained to be in a symmetric polymatroid region.
Large scale systems with a growing number of service classes are studied and
several asymptotic results regarding fairness and mean delays are obtained.
Multi-server models are studied in [30] with each server connected to
multiple file types and each file type stored in multiple servers, thereby creating a bipartite graph. This article focuses on the scaling regime where the
number of servers goes to infinity. It is shown that even if the average degree dn << n := the number of servers, an asymptotically vanishing queuing
delay can be obtained. These results are based on a centralized scheduling
strategy.
In [3], multi-server queues are studied with an arbitrary compatibility
graph between jobs and servers. The paper designs a scheduling algorithm
which achieves balanced fair sharing of the servers. Several parameters are
analyzed using this policy by drawing a parallel between the state of the
system at any time to that of a Whittle network.
However, none of the above papers [28, 16, 20, 34, 17, 29, 30, 3] show
Whittle indexability of the respective resource allocation problems they address. In the present work we take an alternative approach which considers
a dynamic optimization or control problem that can be interpreted as a
problem of scheduling restless bandits. We analyze it in the framework laid
down by Whittle for deriving a heuristic index policy [33]. To the best of
our knowledge, this paper is the first to show Whittle-like indexability of
the server allocation problem, in the setting of a CDN server cluster that
uses resource pooling, with the objective of minimization of the total file
request storage cost. The fact that this problem is Whittle-like indexable
allows us to decouple the original average cost MDP, which is difficult to
solve directly, into more tractable separate control problems for individual
file types. The decoupling leads to an efficient algorithm based on computation of Whittle-like indices, which outperforms several natural heuristics for
the problem.
The Whittle index heuristic has been successfully applied to various re3
source allocation problems including: crawling for ephemeral content [5],
congestion control [6], UAV routing [23], sensor scheduling [22], routing in
clusters [21], opportunistic scheduling [8], inventory routing [2], cloud computing [9] etc. General applications to resource allocation problems can be
found in [15]. Book length treatments of restless bandits can be found in
[13] and [25].
The rest of the paper is structured as follows: In section 2, we discuss
the model which is used and formulate the problem as a Markov Decision
Process (MDP). Section 3 shows various structural properties of the value
function of the MDP formulated in section 2. In section 4, we prove that
the problem of server allocation in the resource pooling setting is in fact
indexable and provide a scheme to compute this index. Section 5 discusses
other heuristics for server allocation and presents numerical comparisons of
the proposed index policy with some other heuristics. We conclude the paper
with a brief discussion in Section 6.
We conclude this section with a brief introduction to the Whittle index
[33]. Let X i (t), t ≥ 0, 1 ≤ i ≤ N , be N Markov chains, each with two
modes of operations: active and passive, with associated transition kernels
p1 (·|·), p0 (·|·) resp. Let r1 (X i (t)), r0 (X i (t)) be instantaneous rewards for the
ith user in the respective modes with r1 (·) ≥ r0 (·). The goal is to schedule
active/passive modes so as to maximize the total expected average reward
T −1
1 XX
lim
E[rν j (t) (X j (t))]
T →∞ T
t=0
j
where ν j (t)P= 1 if jth process is active at time t and 0 if not, under the
constraint j ν j (t) ≤ M ∀t, i.e., at most M processes are active at each
time instant. This hard constraint makes the problem difficult to solve (see
[24]). So following Whittle, one relaxes the constraint to
T −1
1 XX
E[ν j (t)] ≤ M.
T →∞ T
lim
t=0
j
This makes it a problem with separable cost and constraints which, given the
Lagrange multiplier λ, decouples into N individual problems with reward for
passivity changed to λ + r0 (·). The problem is Whittle indexable if under
optimal policy, the set of passive states increases monotonically from empty
set to full state space as λ varies from −∞ to +∞. If so, the Whittle index
for a given state can be defined as the value of λ for which both modes (active
and passive) are equally desirable. The index policy is then to compute these
4
for the current state profile, sort them in decreasing order, and render active
the top M processes, the rest passive. The decoupling implies O(N ) growth
of state space as opposed to the original problem, for which it is exponential
in N . Further, the processes are coupled only through a simple index policy.
The latter is known to be asymptotically optimal as N ↑ ∞ [31]. However,
no convenient general analytic bound on optimality gap seems available.
2
Model and problem formulation
Consider a server cluster that contains multiple servers, each of which stores
one or more files. We represent this system using a bipartite graph 2 G =
(F ∪ S; E) where F is a set of N files, S is a set of M servers, E is the set of
edges, and each edge e ∈ E connecting a file i ∈ F and server j ∈ S implies
that a copy of file i is replicated at server j (see Figure 1).
Figure 1: The model
For j ∈ S, Fj denotes the set of files that are stored in server j. Similarly,
for i ∈ F , Si denotes the set of servers that store file i. Requests for file
type i ∈ F arrive to the server cluster according to an independent Poisson
process with rate Λi . We assume that the job (requested file) sizes have
an exponential distribution. Let ξ ij (t) denote the rate at which server j
transmits file type i at time t. Then the capacity constraint at each server
2
Recall that a graph G = (V, E) is said to be bipartite if its node set V can be partitioned into two sets F and S such that every edge in E is between a node in F and a
node in S [32].
5
can be expressed as:
X
ξ ij (t) ≤ µj
∀t ≥ 0, j ∈ {1, 2, · · · , M }
(1)
i∈Fj
where µj is the maximum permissible rate of transmission from server j. We
make the blanket stability assumption.
Stability Assumption: Let F denote the set of files. For any A ⊂ F , let
SA = ∪i∈A Si . Then for stability of all queues, we assume:
X
X
Λi ≤
µj
∀A ⊂ F
(2)
i∈A
j∈SA
Let f i (x) be the cost for storing x jobs in the queue i. We assume f i (·) to
be an increasing convex function for i = 1, 2, · · · , N . Our aim is to minimize
the long run average cost, given by
1
lim sup
T →∞ T
T
Z
E[
0
N
X
f i (X i (t))]dt,
(3)
i=1
where X i (t) is the length of queue i at time t. Following the classic paper of
Whittle [33], we relax the per stage constraint (1) to an averaged constraint
1
lim sup
T →∞ T
T
Z
0
X
E[ξ ij (t)] ≤ µj
∀j ∈ {1, 2, · · · , M }.
(4)
i∈Fj
This is now a constrained Markov decision process [7] that leads to the
unconstrained problem with cost
Z T
N
X
X
X
1
lim sup
E
f i (X i (t)) +
λ̂j (µj −
ξ ij (t)) dt,
(5)
T
T →∞
0
i=1
j
i∈Fj
where λ̂j is the Lagrange multiplier associated with the j th relaxed constraint
lim sup
T →∞
1
T
Z
T
X
E[
ξ ij (t)] ≤ µj .
0
i
6
Given the values of the Lagrange multipliers λ̂j , this optimization problem
decouples into separate control problems for individual file types, with the
cost function for file type i given by:
X
ci (xi , λ̂) = f i (xi ) +
λ̂j (µj − ξ ij (t)).
(6)
j∈Si
where λ̂ = [λˆ1 , λˆ2 , · · · ]. The average cost dynamic programming (DP) equation for this MDP for file type i is given by [11]:
X
V i (n)Qi (n|m, ξ ij , j ∈ Si ) = 0
(7)
min ci (m, λ̂) − β i +
ξ ij ,j∈Si
n
where β i is the optimal cost for file type i and Qi is its rate matrix given by:
for z > 0,
Qi (z + 1|z, ξ ij , j ∈ Si ) = Λi ,
X
ξ ij (t),
Qi (z − 1|z, ξ ij , j ∈ Si ) =
(8)
(9)
j∈Si
X
Qi (z|z, ξ ij , j ∈ Si ) = − Λi +
ξ ij (t) ,
(10)
j∈Si
for z = 0,
Qi (1|0, ξ ij , j ∈ Si ) = Λi ,
(11)
Qi (0|0, ξ ij , j ∈ Si ) = −Λi ,
(12)
and V i (·) is the value function. Substituting the values of Qi back in the
DP equation and dropping the isolated superscript i for ease of notation, we
have3 :
X
min c(m, λ̂) − β + V (m + 1)Λ + V (m − 1)
ξ ij (t)
ξ ij ,j∈Si
j∈Si
− V (m) Λ +
X
ξ ij (t)
= 0,
(13)
j∈Si
3
Note that when the queue of file type i is empty, no server needs to provide any service
to that particular file type.
7
equivalently,
min
ξ ij ,j∈Si
f (m) +
X
λ̂j (µj − ξ ij (t)) − β + V (m + 1)Λ
j∈Si
+ V (m − 1)
X
ij
ξ (t) − V (m) Λ +
j∈Si
X
ij
ξ (t)
= 0.
(14)
j∈Si
The expression which is to be minimized is linear in ξ ij (t), j ∈ Si and each
ξ ij (t) has the capacity constraint which restricts the values of ξ ij (t) to be
≤ µj i.e. ξ ij (t) ∈ [0, µj ]. This combined with the fact that the objective is
linear ensures that the minimum is attained at a corner where each server is
either serving at full capacity or at zero capacity, i.e., ξ ij (t) = 0 or ξ ij (t) = µj
for all j ∈ Si .
Following [33], we propose an index policy for the problem of server
allocation to file types. Let us define λk to be the marginal utility of allowing
server k to serve at 0 rate, when all other servers containing the file type are
already serving at their full capacity. Since all other servers are serving at
full rate, we have that for all servers other than server k
ξ ij (t) = µj
∀ j ∈ Si , j 6= k
(15)
Adding V (m) on both sides of equation (14) we get:
X
V (m) = min f (m) +
λ̂j (µj − ξ ij (t))
ξ ij ,j∈Si
j∈Si
− β + V (m + 1)Λi + V (m − 1)
X
ξ ij (t)
j∈Si
+ V (m) 1 − Λi +
X
ξ ij (t)
(16)
j∈Si
If we scale all transition rates by a fixed multiplicative factor, it is tantamount
to time scaling which will scale the average cost, but not affect the optimal
policy. Hence without loss of generality, we can assume that the arrival and
service rates are such that the coefficients of V (·) that appear in equation
(16) are between 0 and 1 and can be interpreted as transition probabilities
of a discrete time Markov chain as explained below. Let λk = λ̂k µk . On
8
substituting ξ ij (t) = µj ∀ j ∈ Si , j 6= k, we have:
X
V (m) = f (m) − β + min λk +
p1 (n|m)V (n),
X
p2 (n|m)V (n) .
(17)
Here p1 (·|·) is the transition probability when the server does not serve this
file type and is given by (m > 0):
p1 (m + 1|m) = Λ,
X
p1 (m|m) = 1 − Λ +
µj ,
j∈Si ,j6=k
X
p1 (m − 1|m) =
µj ,
(18)
j∈Si ,j6=k
and p2 (·|·) is the transition probability when the server serves this file type
and is given by (m > 0):
p2 (m + 1|m) = Λ,
p2 (m|m) = 1 − Λ +
X
µj ,
j∈Si
p2 (m − 1|m) =
X
µj .
(19)
j∈Si
For m = 0, the transition probabilities p1 (·|·) and p2 (·|·) are the same and
are given by (for i = 1, 2):
pi (1|0) = Λ
pi (0|0) = 1 − Λ
In the next section, we prove some structural properties of the value
function.
3
Structural Properties of the Value function
This section closely follows the proof techniques of [1], [9] and [10].
Lemma 3.1. V (·) is increasing in the number of files.
9
Proof. The proof follows along the same lines as Lemma 5 in [10]. The main
idea is to consider two processes {Xt } and {Xt0 } on a common probability
space with different initial conditions x < x0 resp., but common arrival and
control processes. Then the result follows by pathwise dominance: Xt0 ≥ Xt
a.s. ∀t, combined with monotonicity of the cost, first for infinite horizon
discounted cost problem and then, by passing through the vanishing discount
limit, for average cost.
Lemma 3.2. V (·) has the property of increasing differences, i.e., for z > 0
and x > y.
V (x + z) − V (x) ≥ V (y + z) − V (y).
Proof. The proof follows along similar lines as Lemma 6 in [10] and Theorem
4 in [1], but with several crucial differences. The argument uses induction
on the finite horizon discounted problem. Let Vn (·) denote the α discounted
n−step problem. We embed the state space into the positive real line, R+ .
Consider the case when x > 0. First, we rearrange the terms to get the
following. Here u is the optimal control for state x.
X
Vn (x) = f (x)+αVn−1 (x + 1)Λ + αVn−1 (x)(1 − Λ −
µi )
i
+ αVn−1 (x − 1)
X
µi + (1 − u)λ
i6=k
+ (1 − u)αµk Vn−1 (x) + uαµk Vn−1 (x − 1)
(20)
We have that V0 (x) = f (x) which is convex. Assume that Vn−1 is convex.
Consider two states x1 and x2 with optimal control u1 and u2 . Without loss
of generality, assume that x1 > x2 . In this part, we assume x2 > 0.
Vn (x1 ) + Vn (x2 ) = f (x1 ) + f (x2 )
X
X
+ αVn−1 (x1 )(1 − Λ −
µi ) + αVn−1 (x2 )(1 − Λ −
µi )
i
i
+ αVn−1 (x1 + 1)Λ + αVn−1 (x2 + 1)Λ
X
X
+ αVn−1 (x1 − 1)
µi + αVn−1 (x2 − 1)
µi
i6=k
i6=k
+ (1 − u1 )λ + (1 − u1 )αµk Vn−1 (x1 ) + u1 αµk Vn−1 (x1 − 1)
+ (1 − u2 )λ + (1 − u2 )αµk Vn−1 (x2 ) + u2 αµk Vn−1 (x2 − 1).
10
Consider two separate cases depending on the values of u1 , u2 .
Case 1: u1 = u2
Vn (x1 ) + Vn (x2 )
x1 + x2
∗1
≥ 2f
2
X
x1 + x2
(1 − Λ −
µi )
+ 2αVn−1
2
i
X
x1 + x2
x1 + x2
+ 2αVn−1
+ 1 Λ + 2αVn−1
−1
µi
2
2
i6=k
u1 + u2
x1 + x2
u1 + u2
+2 1−
αµk Vn−1
+ 2λ 1 −
2
2
2
u1 + u2
x1 + x2
+2
αµk Vn−1
−1
2
2
X
x1 + x2
x1 + x2
+ 2αVn−1
(1 − Λ −
µi )
2
2
i
X
x1 + x2
x1 + x2
+ 1 Λ + 2αVn−1
−1
µi
+ 2αVn−1
2
2
i6=k
x1 + x2
+ 2λ(1 − u3 ) + 2(1 − u3 )αµk Vn−1
2
x1 + x2
+ 2u3 αµk Vn−1
−1
2
x1 + x2
= 2Vn
.
2
≥∗2 2f
(21)
2
Here u3 is the optimal control when the state has x1 +x
files. Inequality ∗1
2
follows from the convexity of f (·) and Vn−1 (·). Inequality ∗2 follows from
the definition of the optimal control u3 .
Case 2: u1 6= u2 :
11
Consider the case u2 = 0, u1 = 1 (The other case is similar)
Vn (x1 ) + Vn (x2 )
x1 + x2
∗1
≥ 2f
2
X
x1 + x2
+ 2αVn−1
(1 − Λ −
µi )
2
i
X
x1 + x2
x1 + x2
+ 2αVn−1
+ 1 Λ + 2αVn−1
−1
µi
2
2
i6=k
+ λ + αµk Vn−1 (x2 ) + αµk Vn−1 (x1 − 1)
X
x1 + x2
x1 + x2
+ 2λ + 2αVn−1
(1 − Λ −
µi )
= 2f
2
2
i
X
x1 + x2
x1 + x2
+ 2αVn−1
µi
+ 1 Λ + 2αVn−1
−1
2
2
i6=k
1
1
1
+ 2αµk Vn−1 (x2 ) + Vn−1 (x1 − 1)
+ 2λ 1 −
2
2
2
X
x1 + x2
x1 + x2
∗2
≥ 2f
+ 2αVn−1
(1 − Λ −
µi )
2
2
i
X
x1 + x2
x1 + x2
+ 2αVn−1
+ 1 Λ + 2αVn−1
−1
µi
2
2
i6=k
1
x1 + x2
1
+ 2αµk Vn−1
+ 2λ 1 −
2
2
2
1
x1 + x2
+ Vn−1
−1
2
2
X
x1 + x2
x1 + x2
+ 2αVn−1
(1 − Λ −
µi )
≥ 2f
2
2
i
X
x1 + x2
x1 + x2
+ 2αVn−1
+ 1 Λ + 2αVn−1
−1
µi
2
2
i6=k
x1 + x2
+ 2λ(1 − u3 ) + 2(1 − u3 )αµk Vn−1
2
x1 + x2
+ 2u3 αµk Vn−1
−1
2
∗3
12
= 2Vn
x1 + x2
.
2
(22)
2
Here u3 is the optimal control when the state has x1 +x
files. Inequali2
ties ∗1, ∗2 follow from the convexity of f (·) and Vn−1 (·) (we use the fact that
convexity imples increasing differences, i.e., f (x+a)−f (x) ≥ f (y +a)−f (y)
for x > y, a > 0). Inequality ∗3 follows from the definition of the optimal
control u3 .
Now, consider the case where x1 > x2 = 0. We have:
Vn (0) = f (0) + (1 − u)λ + (1 − Λ)Vn−1 (0) + ΛVn−1 (1)
(23)
From this equation, we see that u = 1 if λ > 0 and u = 0 otherwise. We
rearrange equation (23) as:
X
Vn (0) = f (0) + (1 − u)λ + (1 − Λ −
µi )Vn−1 (0)
i
+
X
µi Vn−1 (0) + µk Vn−1 (0) + ΛVn−1 (1).
(24)
i6=k
We have:
Vn (x1 ) + Vn (0) = f (x1 ) + f (0)
X
X
+ αVn−1 (x1 )(1 − Λ −
µi ) + αVn−1 (0)(1 − Λ −
µi )
i
i
+ αVn−1 (x1 + 1)Λ + αVn−1 (1)Λ
X
X
+ αVn−1 (x1 − 1)
µi + αVn−1 (0)
µi
i6=k
i6=k
+ (1 − u1 )λ + (1 − u1 )αµk Vn−1 (x1 ) + u1 αµk Vn−1 (x1 − 1)
+ (1 − u2 )λ + αµk Vn−1 (0)
x1
,
≥∗1 2Vn
2
(25)
where ∗1 is derived using convexity and by following similar arguments as in
the case when x2 > 0
This proves convexity of the finite horizon problem. Convexity is preserved under pointwise convergence, so it follows for the infinite horizon
discounted problem by letting the time horizon go to infinity, and then for
the average cost problem by the ‘vanishing discount’ argument as in [1].
13
Convexity implies increasing differences. Therefore V (·) has increasing differences. The function restricted to the non-negative integers will retain this
property, thereby proving the lemma.
Lemma 3.3. The optimal policy is a threshold policy i.e. ∃ x∗ such that if
x > x∗ , the server serves at full capacity, otherwise the server does not serve
this file type.
Proof. In order to prove this, we show that the function:
X
X
g(x) =
p2 (y|x)V (y) −
p1 (y|x)V (y)
is decreasing. On simplifying this expression, we get:
g(x) = µk (V (x − 1) − V (x))
(26)
which is a decreasing function in x by Lemma 3.2. This implies that the
optimal policy is a threshold policy.
In order to specify the threshold uniquely for each λ, we take it to be the
smallest value of the state x for which it is equally desirable to be active or
passive.
4
Whittle Indexability
We next prove a Whittle-like indexability result in the spirit of [33]. We use
the phrase ‘Whittle-like’ because our problem formulation differs from that
of [33], though it builds upon it. By abuse of terminology, we shall still use
the terms Whittle indexability, Whittle index, etc. for our variants thereof.
Let π k denote the stationary probability distribution when the threshold
is k. That is, if the number of jobs is ≤ k, then the server does not transmit,
and if number of jobs is > k, then the server transmits at full rate. We have
the following lemma.
Pk
k
Lemma 4.1.
i=0 π (i) is non-decreasing with k.
P
Proof. Let µ̂ = j∈Si ;j6=k µj . The Markov chain formed with a threshold of
k is shown in Figure 2. This is a time reversible Markov chain with stationary
probabilities given by:
i
Λ
k
k
π (i) = π (0)
if i ≤ k,
µ̂
k
i−k
Λ
Λ
k
k
π (i) = π (0)
if i > k,
(27)
µ̂
µk + µ̂
14
where π k (0) is the stationary probability of state 0.
Figure 2: Markov Chain
From this, we see that
k
X
(Λ
)k −1
µ̂
π k (i) =
i=0
(Λ
)−1
µ̂
(Λ
)k −1
µ̂
Λ
( µ̂ )−1
(28)
µ̂+µk
k
+ (Λ
µ̂ ) ( µ̂+µk −Λ )
which is a non-decreasing function of k, proving the claim.
Theorem 4.1. This problem is Whittle indexable, i.e., the set of passive
states decreases monotonically from the whole state space to the empty set φ
as λ ↑ ∞.
Proof. The proof is along the lines of Theorem 1 in [10]. It has been reproduced for the sake of completeness.
The optimal average cost of the problem is given by
X
X
β(λ) = inf {
f (k)π(k) + λ
π(k)}
k
k∈B
where π the stationary distribution and B the set of passive states. β(λ)
is the infimum of this over all admissible policies, which by Lemma 3.3 is
the same as the infimum over all threshold policies. Hence it is concave
non-decreasing with slope < 1. By Lemma 3.3, we can write this as:
β(λ) =
∞
X
x(λ)
f (k)π
x(λ)
(k) + λ
k=0
X
k=0
15
π x(λ) (k)
where x(λ) is the optimal threshold under λ. By the envelope theorem
(Theorem 1, [19]), the derivative of this function with respect to λ is given
by
x(λ)
X
π x(λ) (k).
k=0
Since β(λ) is a concave function, its derivative has to be a non-increasing
function of λ, i.e.,
x(λ)
X
π x(λ) (k) is non-increasing with λ.
k=0
P
But, from Lemma 4.1, we know that kj=0 π k (j) is a non-decreasing function
of k, where k is the threshold. Then x(λ) is a non-increasing function of λ.
The set of passive states for λ is given by [0, x(λ)]. Since x(λ) is a nonincreasing function of λ, we have that the set of passive states monotonically
decreases to φ as λ ↑ ∞. This implies Whittle indexability.
4.1
Proposed Policy
We propose the following heuristic policy inspired by [33]:
Our decision epochs are the time instances when there is some change
in the system, i.e., either an arrival or a departure occurs. For each server
j ∈ S, the index computed for each file type connected to this server is known.
The file type which has the smallest index is chosen and the server serves it
at full rate. Each time there is either an arrival into a file type or there is
a job completion, the new indices are sent to the server which then decides
which queue to serve.
4.2
Computation of the Whittle index
The Whittle index λ(x) (when the number of jobs is x) is given by the
following set of equations:
Vλn (y) = f (y) + λn (x) + Ep1 [Vλn (y)] − βλn
if y < x,
(29)
Vλn (y) = f (y) + Ep2 [Vλn (y)] − βλn
if y ≥ x,
(30)
Vλn (0) = 0
(31)
λn+1 (x) = λn + η(Ep2 [Vλn (x)] − Ep1 [Vλn (x)] − λn )
16
(32)
Here η is a small step size (taken to be 0.01). Epi [·] denotes taking expectation with respect to the probability distribution pi , i = 1, 2 (as defined in
equations (18) and (19)). Since η is small, we can view Equation 32 as an
approximate discretization to the ODE
λ̇(t) = F (λ(t)) − λ(t)
(33)
where
F (λ) := Ep2 [Vλ (x)] − Ep1 [Vλ (x)].
This is a scalar ODE and the function F is continuous decreasing with equilibrium given by that value of λ for which
λ = Ep2 [Vλ (x)] − Ep1 [Vλ (x)],
(34)
i.e., the Whittle index at state x. As a well-posed scalar ODE with bounded
trajectories, it converges to its equilibrium. The convergence of iteration
(32) to a neighborhood of this equilibrium then follows from Theorem 1 of
[12] by standard arguments. The linear system (29), (30) and (31) is solved
as a subroutine by a suitable linear system solver.
5
Simulations
In this section, we report simulations to compare the performance of the
proposed Whittle index policy with other natural heuristic policies given as
follows:
• Balanced Fair Allocation: This is a centralized scheme for allocating server capacities. See [4] for more details.
• Uniform Allocation: At each instant in time, each server splits its
rate equally to all the files that it contains.
• Weighted Allocation: The server rates are split according to prescribed weights proportional to the arrival rates into the different file
types.
• Random Allocation: The decision epochs are the same. At each
instant, for each server, a file type is chosen randomly and the server
serves this at full capacity.
• Max-Weight Allocation: Each server serves at full capacity to that
file type which has the most number of jobs at any instant.
17
We first compare the Whittle policy with the Optimal scheme and the
balanced fairness allocation scheme for the network shown in Figure 3. The
results of the simulations are shown in Figure 4. The values of parameters are
as follows: Λ1 = 0.2; f 1 (x) = 13x; Λ2 = 0.1; f 2 (x) = 10x; µ1 = 0.2; µ2 = 0.2.
The optimal value for the network in figure 3 is computed using the following
set of iterative equations:
Vn+1 (x1 + x2 ) = f 1 (x1 ) + f 2 (x2 ) − Vn (0, 0)
+ min
Ei [Vn (·)|x1 , x2 ]
i∈{1,2,3,4}
un+1 = argmin Ei [Vn (·)|x1 , x2 ]
i
Here, the control u denotes the following: u = 1 denotes server 1, 2 serve file
1. u = 2 denotes server 1 serves file 1 and server 2 serves file 2. u = 3 denotes
server 1 serves file 2 and server 2 serves file 1. u = 4 denotes server 1, 2 serve
file 2. Ei [·] denotes expectation with respect to the probability distribution
under control i. Figure 4 shows that the proposed Whittle index policy
outperforms the balanced fairness allocation policy.
Figure 3: The Network that we use for simulations
The second network that we consider is shown in Figure 5. The parameters in this simulation are as follows: Λ1 = 0.1; f 1 (x) = 10x; Λ2 =
0.2; f 2 (x) = 20x; Λ3 = 0.1; f 3 (x) = 10x; µ1 = 0.2; µ2 = 0.3.
Figure 6 shows the Whittle indices assigned by file types 1 and 2 to server
1 and Figure 7 shows the Whittle indices assigned by file type 1 to the two
servers.
18
Figure 4: Comparison of Whittle policy, True Optimal, and Balanced Fairness scheme
Figure 5: The Network that we use for simulations
19
Figure 6: Whittle index assigned by different files to the same server
Figure 7: Whittle index assigned by the same file to different servers
Figures 8 and 9 compare the performance of the Whittle index policy with
that of the Uniform Allocation, Random Allocation, Weighted Allocation
and Max-Weight Allocation policies that were described at the beginning
of this section 4 . We can see that the Whittle index based policy performs
better than the other methods of server allocation.
4
We have separated these figures to enable better comparison. This is because the
performance of the Uniform and Random allocation policies is much worse than that of
the other policies.
20
Figure 8: Comparison of Whittle policy with Uniform and Random policies
Figure 9: Comparison of Whittle policy with Weighted and Max Weight
policies
Figure 10 shows simulation results for the model with 10 file types and 10
servers such that file type i is stored in servers i, i + 1(mod10). The parameters used are as follows: Λi = 0.2, f i (x) = 15x, µi = 0.2 for i = 1, 4, 7, 10.
21
Λi = 0.3, f i (x) = 20x, µi = 0.3 for i = 2, 5, 8. Λi = 0.1, f i (x) = 10x, µi = 0.2
for i = 3, 6, 9. Figure 10 shows that the proposed Whittle index policy outperforms the Weighted Allocation and Max-Weight Allocation policies.
Figure 10: Comparison of Whittle policy with Weighted and Max Weight
policies
6
Conclusions and Future Work
We have proved Whittle-like indexability of the problem of server allocation
in resource pooling networks. The allocation of servers using the Whittlelike scheme can be implemented in a distributed manner. The next step
would be to extend this work to more general file types and possibly more
complicated network topologies.
References
[1] Agarwal, M., Borkar, V. S. and Karandikar, A., “Structural properties of
optimal transmission policies over a randomly varying channel." IEEE
Transactions on Automatic Control 53.6, pp. 1476-1491, 2008.
22
[2] Archibald, T. W.; Black, D. P. and Glazebrook, Kevin D., “Indexability
and index heuristics for a simple class of inventory routing problems."
Operations Research 57.2, pp. 314-326, 2009.
[3] Bonald, T. and Comte, C., “Balanced fair resource sharing in computer
clusters." Performance Evaluation 116 pp. 70-83, 2017.
[4] Bonald, T. and Proutiere, A. “Insensitive bandwidth sharing in data
networks." Queueing systems 44.1 pp. 69-100, 2003.
[5] Avrachenkov, K. and Borkar, V. S., “Whittle index policy for crawling
ephemeral content." IEEE Trans. on Control of Network Systems, 2016
(available online).
[6] Avrachenkov, K. ,Borkar, V. S. and Pattathil, S., “Controlling G-AIMD
using Index Policy." 56th IEEE Conference on Decision and Control,
Melbourne, Dec. 12-15, 2017.
[7] Borkar V. S., “Convex Analytic Methods in Markov Decision Processes’.
In Feinberg E. A., Shwartz A. (eds), Handbook of Markov Decision
Processes, Springer, Boston, MA, 2002
[8] Borkar, V. S., Kasbekar, G. S., Pattathil, S. and Shetty, P. Y., “Opportunistic scheduling in restless bandits." IEEE Trans. on Control of
Network Systems, 2017 (available online).
[9] Borkar, V. S. , Ravikumar, K. and Saboo, K. , “An index policy for
dynamic pricing in cloud computing under price commitments." Applicationes Mathematicae, 2017 (available online)
[10] Borkar, V. S. and Pattathil, S. , “Whittle indexability in egalitarian processor sharing systems." Annals of Operations Research, 2017 (available
online).
[11] Guo, X. and Hérnandez-Lerma, O. , Continuous-time Markov Decision
Processes: Theory and Applications, Springer Verlag, Berlin-Heidelberg,
2009.
[12] Hirsch, Morris W. “Convergent activation dynamics in continuous time
networks." Neural networks 2.5, 331-349, 1989
[13] Jacko, P. Dynamic priority allocation in restless bandit models, Lambert
Academic Publishing, 2010.
23
[14] Kurose, J.F. and Ross, K.W. Computer Networking: A Top-Down Approach, Addison Wesley, Sixth Edition, 2012.
[15] Larranaga, M. , Ayesta, U. and Verloop, I. M., “Dynamic control
of birth-and-death restless bandits: application to resource-allocation
problems." IEEE/ACM Trans. Networking, 24.6, pp. 3812-3825, 2016.
[16] Leconte, M., Lelarge, M. and Massoulie, L. , “Bipartite graph structures
for efficient balancing of heterogeneous loads", Proc. ACM Sigmetrics/
Performance, pp. 41-52, 2012.
[17] Leconte, M., Lelarge, M. and Massoulie, L., “Designing adaptive replication schemes in distributed content delivery networks", Proc. 27th
IEEE International Teletraffic Congress (ITC), pp. 28-36, 2015.
[18] Leighton, T. “Improving Performance on the Internet”, Communications
of the ACM, Vol. 52, No. 2, Feb. 2009, pp. 44-51.
[19] Milgrom, P. and Segal, I., “Envelope theorems for arbitrary choice sets."
Econometrica 70.2, pp. 583-601, 2002.
[20] Moharir, S., Ghaderi, J., Sanghavi, S. and Shakkottai, S., “Serving content with unknown demand: the high-dimensional regime", Proc. of
ACM Sigmetrics, pp. 435-447, 2014.
[21] Ninõ-Mora, J. , “Admission and routing of soft real-time jobs to multiclusters: Design and comparison of index policies." Computers & Operations Research 39.12, pp. 3431-3444, 2012.
[22] Nino-Mora, J. and Villar, S. S., “Sensor scheduling for hunting elusive
hiding targets via Whittle’s restless bandit index policy." 5th International Conference on Network Games, Control and Optimization (NetGCooP), Paris, Oct. 12-14, 2011.
[23] Ny, J.L., Dahleh, M. and Feron, E. , “Multi-UAV dynamic routing
with partial observations using restless bandit allocation indices." Proc.
American Control Conference, Seattle, June 11-13, 2008, pp. 4220-4225,
2008.
[24] Papadimitriou, C. H. and Tsitsiklis, J. N. , “The complexity of optimal queuing network control", Mathematics of Operations Research
24.2 (1999), 293-305.
[25] Ruiz-Hernandez, D. , Indexable Restless Bandits, VDM Verlag, 2008.
24
[26] Shah, V. , Centralized content delivery infrastructure exploiting resource pools: Performance models and asymptotics. Ph.D.
Dissertation, Department of Electrical and Computer Engineering, University of Texas at Austin, 2015, available at
https://repositories.lib.utexas.edu/handle/2152/31419.
[27] Shah, V. and de Veciana, G. ,“Performance evaluation and asymptotics
for content delivery networks." Proc. IEEE INFOCOM, Atlanta, May
1-4, 2014, pp. 2607-2615, 2014.
[28] Shah, V. and de Veciana, G. , “High-performance centralized content
delivery infrastructure: models and asymptotics", IEEE/ ACM Transactions on Networking, 23.5, pp. 1674-1687, Oct. 2015.
[29] Shah, V. and de Veciana, D., “Impact of fairness and heterogeneity
on delays in large-scale centralized content delivery systems", Queuing
Systems, 83.3-4, pp. 361-397, 2016.
[30] Tsitsiklis, J. and Xu, K. , “Flexible queuing architectures", Operations
Research, 65.5, pp. 1398-1413, 2017.
[31] Weber, Richard R., and Weiss,G., “On an index policy for restless bandits." Journal of Applied Probability 27.3, pp. 637-648, 1990.
[32] D. West, Introduction to Graph Theory, Prentice Hall, Second Edition,
2000.
[33] Whittle, P. , “Restless bandits: activity allocation in a changing world",
Journal of Applied Probability Vol. 25: A Celebration of Applied Probability, pp. 287-298, 1988.
[34] Zhou, Y., Fu, T. and Chiu, D. “A unifying model and analysis of P2P
VoD replication and scheduling", IEEE/ ACM Transactions on Networking, 23.4, pp. 1163-1175, 2015.
25
| 3cs.SY
|
Subsets and Splits
No saved queries yet
Save your SQL queries to embed, download, and access them later. Queries will appear here once saved.